Methods in

Molecular Biology 2104

Shuzhao Li Editor

Computational
Methods and
Data Analysis for
Metabolomics
METHODS IN MOLECULAR BIOLOGY

Series Editor
John M. Walker
School of Life and Medical Sciences
University of Hertfordshire
Hatfield, Hertfordshire, UK

For further volumes:

http://www.springer.com/series/7651
For over 35 years, biological scientists have come to rely on the research protocols and
methodologies in the critically acclaimed Methods in Molecular Biology series. The series was
the first to introduce the step-by-step protocols approach that has become the standard in all
biomedical protocol publishing. Each protocol is provided in readily-reproducible step-by-
step fashion, opening with an introductory overview, a list of the materials and reagents
needed to complete the experiment, and followed by a detailed procedure that is supported
with a helpful notes section offering tips and tricks of the trade as well as troubleshooting
advice. These hallmark features were introduced by series editor Dr. John Walker and
constitute the key ingredient in each and every volume of the Methods in Molecular Biology
series. Tested and trusted, comprehensive and reliable, all protocols from the series are
indexed in PubMed.
Computational Methods and Data
Analysis for Metabolomics

Edited by

Shuzhao Li
Department of Medicine, Emory University School of Medicine, Atlanta, GA, USA
Editor
Shuzhao Li
Department of Medicine
Emory University School of Medicine
Atlanta, GA, USA

ISSN 1064-3745 ISSN 1940-6029 (electronic)

Methods in Molecular Biology
ISBN 978-1-0716-0238-6 ISBN 978-1-0716-0239-3 (eBook)
https://doi.org/10.1007/978-1-0716-0239-3

© Springer Science+Business Media, LLC, part of Springer Nature 2020

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction
on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation,
computer software, or by similar or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply,
even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations
and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book are believed to
be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty,
express or implied, with respect to the material contained herein or for any errors or omissions that may have been made.
The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

This Humana imprint is published by the registered company Springer Science+Business Media, LLC part of Springer
Nature.
The registered company address is: 233 Spring Street, New York, NY 10013, U.S.A.
Preface

Metabolomics is the new biochemistry. It reinvigorates the old discipline by new data at a
large scale: simultaneous measurement of thousands of chemicals in biological samples.
Many of these chemicals are beyond the known metabolic intermediates. This new informa-
tion fills an important gap between the interactions of genome and environment, thus
conferring enormous potential for improving human health. The metabolomics data also
overlap significantly with the exposome, which aims to quantify all environmental exposures.
The explosive growth of metabolomics creates a large gap in training on metabolomics data
analysis. This book shall provide a comprehensive guide to scientists, engineers, and students
that employ metabolomics in their work, with an emphasis on the understanding and
interpretation of the data.
The book is organized as follows. Chapter 1 provides an overview of the field, and the
following chapters are presented in four sections: data processing for major experimental
platforms (Chapters 2–7), databases and metabolite annotation (Chapters 8–13), major
techniques in data analysis (Chapters 14–19), and biomedical applications (Chapters
20–23).
While it is not possible to cover all the databases and software tools, we aim to have
representations of each major topic and give readers a foundation to work in this field. It is
critical to note that the scientific landscape keeps evolving and tools keep changing. There-
fore, it is more important to understand the rationale and principles than to replicate the
protocols. This book is supplemented by example data and code at GitHub (https://
metabolomics-data.github.io), which can be continuously updated by the community.
I would like to express my gratitude to the metabolomics group at Emory University,
especially Dean Jones, Tianwei Yu, Young-Mi Go, Youngja Park, Karan Uppal, Douglas
Walker, and Gary Miller. Their intellectual input and friendship made my scientific journey
truly rewarding.

Atlanta, GA, USA Shuzhao Li

v
Contents

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
Contributors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
1 Overview of Experimental Methods and Study Design in Metabolomics,
and Statistical and Pathway Considerations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Stephen Barnes
2 Metabolomics Data Processing Using XCMS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
Xavier Domingo-Almenara and Gary Siuzdak
3 Metabolomics Data Preprocessing Using ADAP and MZmine 2 . . . . . . . . . . . . . . 25
Xiuxia Du, Aleksandr Smirnov, Tomáš Pluskal, Wei Jia,
and Susan Sumner
4 Metabolomics Data Processing Using OpenMS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
Marc Rurik, Oliver Alka, Fabian Aicheler, and Oliver Kohlbacher
5 Analysis of NMR Metabolomics Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
Wimal Pathmasiri, Kristine Kay, Susan McRitchie, and Susan Sumner
6 Key Concepts Surrounding Studies of Stable Isotope-Resolved
Metabolomics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
Stephen F. Previs and Daniel P. Downes
7 Extracting Biological Insight from Untargeted Lipidomics Data . . . . . . . . . . . . . . 121
Jennifer E. Kyle
8 Overview of Tandem Mass Spectral and Metabolite Databases
for Metabolite Identification in Metabolomics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
Zhangtao Yi and Zheng-Jiang Zhu
9 METLIN: A Tandem Mass Spectral Library of Standards . . . . . . . . . . . . . . . . . . . . 149
J. Rafael Montenegro-Burke, Carlos Guijas, and Gary Siuzdak
10 Metabolomic Data Exploration and Analysis with the Human
Metabolome Database. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
David S. Wishart
11 De Novo Molecular Formula Annotation and Structure Elucidation
Using SIRIUS 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
Marcus Ludwig, Markus Fleischauer, Kai Dührkop,
Martin A. Hoffmann, and Sebastian Böcker
12 Annotation of Specialized Metabolites from High-Throughput
and High-Resolution Mass Spectrometry Metabolomics . . . . . . . . . . . . . . . . . . . . . 209
Thomas Naake, Emmanuel Gaquerel, and Alisdair R. Fernie
13 Feature-Based Molecular Networking for Metabolite Annotation . . . . . . . . . . . . . 227
Vanessa V. Phelan
14 A Bioinformatics Primer to Data Science, with Examples for Metabolomics . . . . 245
W. Stephen Pittard, Cecilia “Keeko” Villaveces, and Shuzhao Li
15 The Essential Toolbox of Data Science: Python, R, Git, and Docker . . . . . . . . . . 265
W. Stephen Pittard and Shuzhao Li

vii
viii Contents

16 Predictive Modeling for Metabolomics Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313

Tusharkanti Ghosh, Weiming Zhang, Debashis Ghosh,
and Katerina Kechris
17 Using MetaboAnalyst 4.0 for Metabolomics Data Analysis,
Interpretation, and Integration with Other Omics Data . . . . . . . . . . . . . . . . . . . . . 337
Jasmine Chong and Jianguo Xia
18 Using Genome-Scale Metabolic Networks for Analysis, Visualization,
and Integration of Targeted Metabolomics Data. . . . . . . . . . . . . . . . . . . . . . . . . . . . 361
Jake P. N. Hattwell, Janna Hastings, Olivia Casanueva,
Horst Joachim Schirra, and Michael Witting
19 Pathway Analysis for Targeted and Untargeted Metabolomics . . . . . . . . . . . . . . . . 387
Alla Karnovsky and Shuzhao Li
20 Application of Metabolomics to Renal and Cardiometabolic Diseases . . . . . . . . . 401
Casey M. Rebholz and Eugene P. Rhee
21 Using the IDEOM Workflow for LCMS-Based Metabolomics
Studies of Drug Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419
Anubhav Srivastava and Darren J. Creek
22 Analyzing Metabolomics Data for Environmental Health and
Exposome Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447
Yuping Cai, Ana K Rosen Vollmar, and Caroline Helen Johnson
23 Network-Based Approaches for Multi-omics Integration. . . . . . . . . . . . . . . . . . . . . 469
Guangyan Zhou, Shuzhao Li, and Jianguo Xia

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 489
Contributors

FABIAN AICHELER • Applied Bioinformatics Group, University of Tübingen, Tübingen,

Germany
OLIVER ALKA • Applied Bioinformatics Group, University of Tübingen, Tübingen, Germany
STEPHEN BARNES • Department of Pharmacology & Toxicology and Targeted Metabolomics
and Proteomics Laboratory, University of Alabama at Birmingham, Birmingham, AL,
USA
SEBASTIAN BÖCKER • Chair for Bioinformatics, Friedrich-Schiller University, Jena, Germany
YUPING CAI • Department of Environmental Health Sciences, Yale School of Public Health,
New Haven, CT, USA
OLIVIA CASANUEVA • Department of Epigenetics, Babraham Institute, Cambridge, UK
JASMINE CHONG • Institute of Parasitology, McGill University, Montreal, QC, Canada
DARREN J. CREEK • Drug Delivery, Disposition and Dynamics, Monash Institute of
Pharmaceutical Sciences, Monash University, Parkville, VIC, Australia
XAVIER DOMINGO-ALMENARA • Scripps Center for Metabolomics, The Scripps Research
Institute, La Jolla, CA, USA
DANIEL P. DOWNES • Department of Chemistry, Merck & Co., Inc., Kenilworth, NJ, USA
XIUXIA DU • Department of Bioinformatics and Genomics, University of North Carolina at
Charlotte, Charlotte, NC, USA
KAI DÜHRKOP • Chair for Bioinformatics, Friedrich-Schiller University, Jena, Germany
ALISDAIR R. FERNIE • Central Metabolism, Max Planck Institute of Molecular Plant
Physiology, Potsdam-Golm, Germany
MARKUS FLEISCHAUER • Chair for Bioinformatics, Friedrich-Schiller University, Jena,
Germany
EMMANUEL GAQUEREL • Institute of Plant Molecular Biology, University of Strasbourg,
Strasbourg, France; Centre for Organismal Studies, University of Heidelberg, Heidelberg,
Germany
DEBASHIS GHOSH • Department of Biostatistics and Informatics, Colorado School of Public
Health, University of Colorado Anschutz Medical Campus, Aurora, CO, USA
TUSHARKANTI GHOSH • Department of Biostatistics and Informatics, Colorado School of
Public Health, University of Colorado Anschutz Medical Campus, Aurora, CO, USA
CARLOS GUIJAS • Scripps Center for Metabolomics, The Scripps Research Institute, La Jolla,
CA, USA
JANNA HASTINGS • Department of Epigenetics, Babraham Institute, Cambridge, UK
JAKE P. N. HATTWELL • Centre for Advanced Imaging, The University of Queensland,
Brisbane, QLD, Australia
MARTIN A. HOFFMANN • Chair for Bioinformatics, Friedrich-Schiller University, Jena,
Germany; International Max Planck Research School “Exploration of Ecological
Interactions with Molecular and Chemical Techniques”, Max Planck Institute for
Chemical Ecology, Jena, Germany
WEI JIA • University of Hawaii Cancer Center, Honolulu, HI, USA
CAROLINE HELEN JOHNSON • Department of Environmental Health Sciences, Yale School of
Public Health, New Haven, CT, USA

ix
x Contributors

ALLA KARNOVSKY • Department of Computational Medicine and Bioinformatics, University

of Michigan, Ann Arbor, MI, USA
KRISTINE KAY • Department of Nutrition, School of Public Health, NIH Eastern Regional
Comprehensive Metabolomics Resource Core (ERCMRC), Nutrition Research Institute,
University of North Carolina at Chapel Hill, Kannapolis, NC, USA
KATERINA KECHRIS • Department of Biostatistics and Informatics, Colorado School of Public
Health, University of Colorado Anschutz Medical Campus, Aurora, CO, USA
OLIVER KOHLBACHER • Applied Bioinformatics Group, University of Tübingen, Tübingen,
Germany; Biomolecular Interactions, Max Planck Institute for Developmental Biology,
Tübingen, Germany; Quantitative Biology Center, University of Tübingen, Tübingen,
Germany; Institute for Translational Bioinformatics, University Hospital Tübingen,
Tübingen, Germany; Institute for Bioinformatics and Medical Informatics, University of
Tübingen, Tübingen, Germany
JENNIFER E. KYLE • Biological Sciences Division, Pacific Northwest National Laboratory,
Richland, WA, USA
SHUZHAO LI • Department of Medicine, Emory University School of Medicine, Atlanta, GA,
USA
MARCUS LUDWIG • Chair for Bioinformatics, Friedrich-Schiller University, Jena, Germany
SUSAN MCRITCHIE • Department of Nutrition, School of Public Health, NIH Eastern
Regional Comprehensive Metabolomics Resource Core (ERCMRC), Nutrition Research
Institute, University of North Carolina at Chapel Hill, Kannapolis, NC, USA
J. RAFAEL MONTENEGRO-BURKE • Scripps Center for Metabolomics, The Scripps Research
Institute, La Jolla, CA, USA
THOMAS NAAKE • Central Metabolism, Max Planck Institute of Molecular Plant Physiology,
Potsdam-Golm, Germany
WIMAL PATHMASIRI • Department of Nutrition, School of Public Health, NIH Eastern
Regional Comprehensive Metabolomics Resource Core (ERCMRC), Nutrition Research
Institute, University of North Carolina at Chapel Hill, Kannapolis, NC, USA
VANESSA V. PHELAN • Department of Pharmaceutical Sciences, Skaggs School of Pharmacy
and Pharmaceutical Sciences, University of Colorado, Aurora, CO, USA
W. STEPHEN PITTARD • Department of Biostatistics and Bioinformatics, Rollins School of
Public Health, Emory University, Atlanta, GA, USA
TOMÁŠ PLUSKAL • Whitehead Institute for Biomedical Research, Cambridge, MA, USA
STEPHEN F. PREVIS • Department of Chemistry, Merck & Co., Inc., Kenilworth, NJ, USA
CASEY M. REBHOLZ • Welch Center for Prevention, Epidemiology and Clinical Research,
Baltimore, MD, USA; Department of Epidemiology, Johns Hopkins Bloomberg School of
Public Health, Baltimore, MD, USA
EUGENE P. RHEE • Massachusetts General Hospital, Richard B. Simches Research Center,
Boston, MA, USA
ANA K ROSEN VOLLMAR • Department of Environmental Health Sciences, Yale School of
Public Health, New Haven, CT, USA
MARC RURIK • Applied Bioinformatics Group, University of Tübingen, Tübingen, Germany
HORST JOACHIM SCHIRRA • Centre for Advanced Imaging, The University of Queensland,
Brisbane, QLD, Australia
GARY SIUZDAK • Scripps Center for Metabolomics, The Scripps Research Institute, La Jolla,
CA, USA; Department of Molecular and Computational Biology, The Scripps Research
Institute, La Jolla, CA, USA
 Contributors xi

ALEKSANDR SMIRNOV • Department of Bioinformatics and Genomics, University of North

Carolina at Charlotte, Charlotte, NC, USA
ANUBHAV SRIVASTAVA • Drug Delivery, Disposition and Dynamics, Monash Institute of
Pharmaceutical Sciences, Monash University, Parkville, VIC, Australia
SUSAN SUMNER • Department of Nutrition, School of Public Health, NIH Eastern Regional
Comprehensive Metabolomics Resource Core (ERCMRC), Nutrition Research Institute,
University of North Carolina at Chapel Hill, Kannapolis, NC, USA
CECILIA “KEEKO” VILLAVECES • Department of Mathematics, University of Georgia, Athens,
GA, USA
DAVID S. WISHART • Department of Computing Science, University of Alberta, Edmonton,
AB, Canada; Department of Biological Sciences, University of Alberta, Edmonton, AB,
Canada
MICHAEL WITTING • Research Unit Analytical BioGeoChemistry, Helmholtz Zentrum
München, Neuherberg, Germany; Analytical Food Chemistry, Technical University of
Munich, Freising, Germany
JIANGUO XIA • Institute of Parasitology, McGill University, Montreal, QC, Canada;
Department of Animal Science, McGill University, Montreal, QC, Canada; Department
of Microbiology and Immunology, McGill University, Montreal, QC, Canada; Department
of Human Genetics, McGill University, Montreal, QC, Canada
ZHANGTAO YI • Interdisciplinary Research Center on Biology and Chemistry, Shanghai
Institute of Organic Chemistry, Chinese Academy of Sciences, Shanghai, People’s Republic
of China
WEIMING ZHANG • Department of Biostatistics and Informatics, Colorado School of Public
Health, University of Colorado Anschutz Medical Campus, Aurora, CO, USA
GUANGYAN ZHOU • Institute of Parasitology, McGill University, Montreal, QC, Canada
ZHENG-JIANG ZHU • Interdisciplinary Research Center on Biology and Chemistry, Shanghai
Institute of Organic Chemistry, Chinese Academy of Sciences, Shanghai, People’s Republic
of China
 Chapter 1

Overview of Experimental Methods and Study Design

in Metabolomics, and Statistical and Pathway
Considerations
Stephen Barnes

Abstract
Metabolomics has become a powerful tool in biological and clinical investigations. This chapter reviews the
technological basis of metabolomics and the considerations in answering biomedical questions. The work-
flow of metabolomics is explained in the sequence of data processing, quality control, metabolite annota-
tion, statistical analysis, pathway analysis, and multi-omics integration. Reproducibility in both sample
analysis and data analysis is key to the scientific progress, and the recommendation is made on reporting
standards in publications. This chapter explains the technical aspects of metabolomics in the context of
systems biology and applications to human health.

Key words Metabolomics, GC-MS, LC-MS, NMR, Precision medicine, Systems medicine, Annota-
tion, Recommendation

1 Introduction

1.1 The Age of Omics In biomedical science, the late 1980s saw a great change in the
and Precision forms and the scale of research data. Instead of cloning individual
Medicine cDNAs that encoded the open reading frames of individual genes,
the NIH funded sequencing of the whole human genome. The data
were collected without recourse to a hypothesis; instead, the
human genome project was intended to create a national (and
later international) resource of genes and genomes. As a result,
the opportunity arose to engineer placing representatives of the
open reading frames of genes onto small glass chips (microarrays),
thus allowing for the “whole” transcriptome to be interrogated in a
single experiment. Since then, in studying the transcriptome, the
need for selected DNA sequences has been removed, and instead,
with further engineering, direct sequencing of the RNA transcripts
(RNA-Seq) has occurred.

Shuzhao Li (ed.), Computational Methods and Data Analysis for Metabolomics, Methods in Molecular Biology, vol. 2104,
https://doi.org/10.1007/978-1-0716-0239-3_1, © Springer Science+Business Media, LLC, part of Springer Nature 2020

1
2 Stephen Barnes

In parallel to the advances in gene sequencing, the field of

protein identification was also making quick strides. Instead of
chemically sequencing a protein, in the late 1980s, results from
cDNA cloning of the open reading frames were used to create the
entire amino acid sequence of proteins. With the introduction of
matrix-assisted laser desorption ionization (MALDI) combined
with time-of-flight (TOF) mass analysis, peptide mass fingerprint-
ing allowed for even faster identification of individual proteins [1];
indeed, using bioinformatics, even proteins (with amino acid
sequences generated from genome sequencing data) that had
never been isolated before could be identified. The turning point
regarding the generation of massive datasets came with the intro-
duction of electrospray ionization (ESI) that coupled liquid chro-
matography with mass spectrometry (LC-MS) [2]. Further
improvements involved scaling chromatographic flow rates down
to nl/min to increase sensitivity [3]. With the use of ion traps, these
assays generated large datasets consisting of information on
20,000–50,000 peptides per analysis. It was immediately obvious
that it was no longer possible to interpret peptide MS/MS spectra
by hand. Computer interrogation of libraries built from multiple
sources, in particular from genome sequencing, began. Later, the
use of reverse protein databases much improved the quality of the
peptide/protein matches [4–7].
In a step relevant to metabolomics, two types of mass spec-
trometer, a hybrid quadrupole time-of-flight mass analyzer
(QTOF) and a quadrupole-Orbitrap, came into use with the ability
to collect high mass accurate (1–3 ppm or better) MS spectra of
peptides eluting from LC columns and MS/MS spectra of selected
peptides. This further increased the number of peptides observed in
an analysis and hence the size of the collected datasets. The empha-
sis in proteomics until quite recently was qualitative, that is, discov-
ering the presence of proteins with rudimentary quantitative
analysis. This was somewhat overcome with the use of isotopically
labeled reagents (iTRAQ and TMT). The introduction of Sequen-
tial Windows-MS of All the Theoretical Mass Spectra (SWATH-
MS) [8] and Parallel Reaction Monitoring (PRM) [9] has led to the
acquisition of complete MS1 and MS2 digital libraries and much
improved quantification of detected peptides.
The rapid progress in genomics and proteomics has changed
most biomedical disciplines, and translated into clinical practice.
Genetic variants now guide the diagnosis and treatment of diseases.
The quest of functional genomics (i.e., understanding gene func-
tions) is, however, far from complete [10]. The contributions to
human diseases by the genome and by the environment are yet to
be quantified and understood, but key to the concept of precision
medicine [11, 12]. Metabolomics, which shares some degree of
analytical chemistry with proteomics, is critical to fill this gap
between genome and environment [13].
 Overview of Metabolomics 3

1.2 Technologies Metabolomics is different from earlier studies of metabolism

for Metabolomics (1935–1950) in that it is a method based on chromatography,
mass spectrometry, and NMR to discover all chemical forms in a
biological sample. Initially, the study of endogenous metabolites
from known metabolic pathways and of specific exogenous com-
pounds (known drugs and natural and man-made toxins) was
principally centered on gas liquid chromatography (GC) and later
in combination with mass spectrometry (GC-MS). With the intro-
duction of ESI and atmospheric pressure chemical ionization
(APCI), a much wider group of metabolites and other compounds
found in biological systems could now be studied.

1.2.1 GC-MS GC-MS is a targeted approach to identify and quantify compounds

in a biological sample. While GC-MS has extensive libraries of
spectral and retention time data, it is nonetheless generally limited
to compounds with masses of 400 Da or less and which once
derivatized are thermally stable. Since electron impact ionization
dissociates the volatile form of the metabolite to create ions, only
product ion spectra are collected for each “metabolite.” This has
placed an emphasis on reproducible, chromatographic separation
that was made possible by the commercial introduction of open
tubular, wall-coated, capillary GC in the 1970s.

1.2.2 LC-MS The transition in metabolomics to LC-MS because of the introduc-

tion of ESI and APCI enabled a much wider group of compounds
in the metabolome to be detected as positively and negatively
charged molecular ions formed in each interface. The molecular
ions in turn could be selectively isolated (by a quadrupole mass
filter), passed into a chamber for collision with neutral gas to
generate their product ion (MS/MS) spectra. Initially, the vast
number of molecular ions hinted at the possibility of discovery of
many new metabolic pathways. However, the soft ionization of the
ESI interface leads to appearance of many adducts (Na+, NH4+, and
K+ in positive ion spectra and formate, acetate and chloride in
negative ion spectra) as well as isotopic (13C) species, and multiply
charged and dimeric species. Therefore, a single metabolite typi-
cally may be represented by many (5–10 or more) different ion
features. The number observed is a function of the amount of a
metabolite. As a further issue, although most metabolites remain as
intact molecular ions, some such as β-glucuronides will undergo
in-source decay giving rise to ions of glucuronic acid (m/z 176 and
113) as well as neutral loss ions (M-176). In most cases, all these
multiple ion forms of a metabolites will coelute.

1.2.3 CE-MS Significant progress has been made to capillary electrophoresis—

MS (CE-MS) in the past few years, by improving the interfacing
between CE and ESI. Due to the selectivity of electrophoresis,
4 Stephen Barnes

CE-MS can be fast and work well with small volumes [14]. As is
occurring in transcriptomics and proteomics, there is a strong
interest in studying the metabolome of single cells. However, the
fluid coming from a single cell without dilution is in the nL or even
pL range. CESI-MS is suited to these volumes and since it also
generates a very sharp peak for each metabolite, it has allowed for
the analysis of over 100 metabolites from a frog embryo cell [15].

1.2.4 Ion Mobility Besides using GC, LC or CE to chromatographically separate

metabolites and MS, MS/MS and even MSn to distinguish them
from one another, another dimension of separation that is rapidly
being introduced is ion mobility which comes in several flavors.
FAIMS (high-field asymmetric waveform ion mobility spectrome-
try) and differential mobility operate in the ionization interface.
Ions move through a small gap and are subjected to an oscillating,
orthogonal and asymmetric, electric field. By applying a particular
compensating voltage, individual ions can be coaxed to pass into
the mass spectrometer. Further metabolite separation can be
achieved by including low levels of metal ions in mobile phase or
by resolvating the ions with isopropanol or other volatile solvents.
The other classes of ion mobility occur inside the mass spectrome-
ter and allow for the calculation of the collision cross section of a
metabolite ion. This can be an absolute value of a metabolite ion,
independent of the instrument used for its measurement or lab
where it is done.

1.2.5 NMR The application of NMR to metabolomics came as a consequence

of two innovations—the introduction of superconducting magnets
and pulse sequences. The much higher field strengths (for proton
resonances from 90 MHz for iron magnets to first 400 MHz and
now up to 1000 MHz) significantly improved the resolution of
individual metabolite resonances, allowing direct quantitative anal-
ysis. Prior to the interest in metabolomics, samples had to be
carefully dried to exclude water since it was in present in many
orders higher amounts to the metabolites. The introduction of
WEFT and 1D-NOESY pulse sequences suppressed the water pro-
ton signal, thereby allowing analysis of biofluids as is.

1.3 What Is The metabolome in a human biological fluid or even in the culture
the Metabolome? of a single cell is not fully predicted by the genes in the genome of
that species. Since humans can only synthesize from simpler pre-
cursors of ten of the twenty amino acids that make up human
proteins, they and other forms of life therefore have to eat (or be
fed) foods coming from other genomes. Some foods, such as fruits
and vegetables, have genomes and hence secondary metabolic path-
ways, that are completely unrelated to those in humans. In addition
to amino acids, these other sources may provide critical small
molecules that are vital for life (vitamins), many of which are
 Overview of Metabolomics 5

enzyme cofactors. Furthermore, humans and other species are

polygenic due to the presence of vast populations of microbial
species, particularly in the gastrointestinal tract. These microbial
communities carry out other pathways not represented in mamma-
lian or plant genomes. Therefore, pathways are not two dimen-
sional as in a single species. Instead, they are polydimensional and as
such are not (yet) represented in published pathway maps or
databases.
Environmental exposures and their metabolites are also present
in humans. With increased sensitivity, some of these molecules can
be directly detected and measured in metabolomics assays. Besides,
the endogenous molecules in the metabolomics data can be used to
assess biological effects from the exposures. Thus, exposome is a
highly active area of metabolomics research (as discussed in
Chapter 22).
As a subset of metabolomics, lipidomics is often discussed
separately because lipids require their own sample preparation,
analytical protocol and data processing. Within each lipid class,
there are many variations of different length of carbon chains and
number of double bonds. This leads to particular patterns in the
mass spectrometry data, which can be leveraged in data processing.
Chapter 7 by Dr. Kyle gives a practical guide on analyzing these
lipidomics data.

2 Challenges in Interrogating the “Metabolome”

The diverse nature of the metabolome generates very large and

complex datasets requiring statistical analysis, metabolite identifi-
cation and verification, and pathway analysis procedures, as
described in the chapters of this book. Although, integrated, com-
mercial software tools exist, the metabolomics community benefit
more from free and open source tools. Because commercial soft-
ware is unlikely to keep up with the pace of scientific innovation,
adaptability to the changing of specific needs and modifying source
code is key to scientific investigations.
Before these methods can be invoked, collected data require
preprocessing. For NMR (without preliminary chromatography),
the chemical shifts of the NMR resonances are first referenced to
0.00 ppm with trimethylsilyl-2,2,3,3-deuterated-propionic acid
added as an internal standard. The data can be analyzed by binning,
dividing the NMR spectrum into small windows. However, due to
pH differences in individual samples, this may lead to false metabo-
lite comparisons. To overcome this, spectra are adjusted for the
effects of pH. The methods of data preprocessing and data analysis
strategies used in NMR based metabolomics analysis are described
in Chapter 5 by Dr. Wimal Pathmasiri and his colleagues.
6 Stephen Barnes

GC-MS and LC-MS data both contain information about ion

features—mass-to-charge (m/z) values, retention times, and inten-
sities. In the case of LC-MS data, besides precursor ion data (MS1),
there are accompanying MS/MS (MS2) data. The latter may be on
selected precursor ions, or be comprehensive, as in SWATH-MS.
MS1 data can be examined in order to identify ion features. Since
many of the ions are from known, abundant metabolites, their true
m/z values can be used to calibrate the other observed feature
masses. A further adjustment is made for variation in retention
times before statistical analysis. Although this can largely be dealt
with by using reproducible chromatographic methods, there are
alignment tools to do this. The data preprocessing for MS meta-
bolomics is described in relation to three popular tools, XCMS
(Drs. Xavier Domingo-Almenara and Gary Siuzdak, Chapter 2),
MzMine (Dr. Xiuxia Du et al., Chapter 3), and OpenMS
(Dr. Kohlbacher and colleagues, Chapter 4).
The greatest challenge in metabolomics is identifying the
metabolite features. Although the mass of a molecular ion can be
measured with accuracies in the sub-ppm range, sufficient to obtain
its empirical formula, this does not allow for its identification since
there are usually multiple metabolites with identical formulae. In
some cases, the metabolites can be resolved chromatographically;
however, validating their identification may depend on having an
authentic standard which is frequently not the case.
Chapter 11 describes the state of the art in computational
prediction of chemical formula and structures by Dr. Böcker and
colleagues. Initially, metabolite databases such as METLIN
(Chapter 9) and the Human Metabolite Database (HMDB,
Chapter 10) accumulated MS1 data from multiple sources. For
some metabolites, there was accompanying MS/MS data. How-
ever, the MS/MS data were often carried out under collision con-
ditions that were not equivalent to those used by an individual
metabolomics investigator, particularly if the mass analyzer used
was different (low resolution/mass accuracy ion trap or quadrupole
analyzer, versus a high resolution/mass accuracy time-of-flight,
orbitrap or ion cyclotron resonance analyzer). This led to the
Siuzdak group to generate MS/MS data under differing voltage
collisions (e.g., 0, 10, 20, or 40 V). However, creating these spectra
on more than 100,000 metabolites is a substantial task. To over-
come this mountain of work, there has been a strong interest in
generating predicted MS/MS spectra as well as using high-
resolution fragment ion libraries to suggest partial structures of a
metabolite ion. An overview of tandem mass spectral and metabo-
lite databases is given by Yi and Zhu in Chapter 8. Different
strategies of metabolite annotation based on the concept of
“molecular networking” are illustrated by Dr. Naake et al. in
Chapter 12 on plant metabolites, and by Dr. Phelan in
Chapter 13 on microbial metabolites.
 Overview of Metabolomics 7

The level of metabolite identification dictates the strategy of

data analysis, which is explained in detail later. For the reproduc-
ibility of science, a set of standards for reporting the findings is
necessary.

3 Metabolomics Reporting Recommendation

From the Metabolomics Standards Initiative (MSI) by the Meta-

bolomics Society, a minimal reporting standard for metabolomics
experiments was published to guide how a metabolomics study and
its biological system should be described in a reproducible way
[16, 17]. The leading data repositories (Metabolomics Workbench,
https://www.metabolomicsworkbench.org; MetaboLights,
https://www.ebi.ac.uk/metabolights/) now enforce a submission
standard similar to those.
Part of the initiative produced the definition of four levels of
reporting metabolite identification in scientific literature [18]. The
four levels are level 1 for identified metabolites, level 2 for putatively
annotated compounds, level 3 for putatively characterized com-
pound classes, and level 4 for unknown compounds. Schymanski
et al. [19] proposed an updated, popular level system as follows:
level 1 for confirmed structure by reference standard, level 2 for
probable structure, level 3 for tentative candidate, level 4 for
unequivocal molecular formula, and level 5 for exact mass. The
support data requirement is defined for each level (e.g., experimen-
tal MS and MSn data for level 3). Schrimpe-Rutledge et al. ([20]
further proposed that orthogonal information, e.g. IM-MS and
NMR data, can be used to achieve level 2 and level 3 identification).
For scientific publications, authors are recommended to share their
data in a major repository, and report metabolite identification level
according to one of these above references.
Besides reproducibility in sample analysis, we are reminded that
reproducibility in data analysis is as important. The use of compu-
tational notebooks, version control and Docker containers, as
described in Chapters 14 and 15, is among best practices. The
recommendation of TRIPOD (Transparent reporting of a multi-
variable prediction model for individual prognosis or diagnosis) is
described in Chapter 16.

4 Data Analysis

Once detected and aligned (both NMR and GC-MS/LC-MS)

processed data can undergo statistical analysis. Chapters 14, 15,
and 16 introduce major skills that are employed in statistics and
data science, using examples on metabolomics data analysis, includ-
ing quality control and data visualization. The most commonly
8 Stephen Barnes

used, online statistical software is MetaboAnalyst (http://www.

MetaboAnalyst.ca) which is described in Chapter 17. MetaboAna-
lyst contains routines for many univariate and multivariate statistical
methods, hierarchal clustering, heatmaps and vector support mod-
ule [21]. A strong feature of MetaboAnalyst is that it allows the
investigator to download a record of all steps taken to analyze a set
of data. Besides the statistical routines, MetaboAnalyst has many
other components including pathway analysis.
While metabolite identification is required to perform conven-
tional pathway and network analysis, an innovative solution for
pathway without the formal identification of metabolites is Mum-
michog [22]. It takes all possible interpretations of an ion feature,
including grouping the features with adducts and isotopic forms.
The ions are separated into two groups based on their statistical
importance (defined by the user). Ions from the nonsignificant
group, equivalent in number to those in the statistically significant
group, are selected at random and mapped to known metabolite
pathways. This repeated 100 or more times to estimate nonsignifi-
cant association with each pathway. Then the significant group of
metabolites are mapped to the same pathways and pathway
enhancement detected. Mummichog also links individual metabo-
lites together based on their chemistry, thereby creating chemical
networks. The network overlaps with the concept of genome scale
metabolic models, which are covered in depth by Dr. Witting and
colleagues in Chapter 18.

5 From Metabolomics to Systems Medicine

The applications of metabolomics are broad and transformative.

Both untargeted and targeted assays are used to dissect intracellular
molecular mechanisms. Even more powerful is the combination
with stable isotope tracing, which can be used to identify the
pathway of action, and to discover new pathways (Chapter 6).
Arguably, metabolomics is the most exciting innovation in epide-
miology since the infusion of genomics. Hundreds of papers have
been published using metabolomics in population studies, often via
the approach of MWAS: metabolome-wide association studies.
Chapter 20 by Drs. Rebholz and Rhee gives a practical discourse
on using metabolomics for biomarkers and clinical investigations.
Mass spectrometry has always been important to pharmacology.
Drs. Srivastava and Creek describe a detailed example of using
metabolomics to track drug mechanisms (Chapter 21). As men-
tioned earlier, the mass spectrometry data can capture environmen-
tal exposures and their biological responses. Chapter 22 by
Caroline Johnson and colleagues describes the concepts and
approaches of exposome, the ambitious effort to investigation the
total environmental exposures.
 Overview of Metabolomics 9

In the three decades following the launch of human genome

project, the technologies of -omics have become widespread and
more economical. Multi-omics studies, for example, employing
transcriptomics, proteomics, and metabolomics in matched sam-
ples, have become feasible. With mass spectrometry and careful
sample preparation, one can even perform proteomics, metabolo-
mics, and lipidomics from a single sample [23]. Given that each
-omics technology measures only a slice of real biological complex-
ity, the integration of multiple -omics data will produce more
complete understanding of the biology. Because of the high-
throughput nature of the data, a single data type is prone to false
positives. The intersection of multiple -omics data is therefore
effective to identify specific mechanisms. These efforts on data
integration often depend on particular biological models; for exam-
ple, proteins and mRNA transcripts are both linked to gene models
in genome annotation. When the measurements are based on dif-
ferent biological matrices, data- or network-based integration is
more useful, as discussed in Chapter 23.
Personal genomics is now in clinical practice. Metabolomics
holds the promise of molecular phenotyping and is indispensable
to our quest of precision medicine [13, 24]. As metabolite panels
are routinely used in blood tests, it is possible in the foreseeable
future that comprehensive metabolomics screen will become part of
routine healthcare. We are optimistic that the methodologies pre-
sented in this book will help us get to this future.

References
1. Hillenkamp F, Karas M (1990) Mass spectrom- 5. Craig R, Cortens JC, Fenyo D, Beavis RC
etry of peptides and proteins by matrix-assisted (2006) Using annotated peptide mass spec-
ultraviolet laser desorption/ionization. Meth- trum libraries for protein identification. J Pro-
ods Enzymol 193:280–295 teome Res 5(8):1843–1849. https://doi.org/
2. Want EJ, Cravatt BF, Siuzdak G (2005) The 10.1021/pr0602085
expanding role of mass spectrometry in metab- 6. Huttlin EL, Hegeman AD, Harms AC, Suss-
olite profiling and characterization. Chembio- man MR (2007) Prediction of error associated
chem 6(11):1941–1951. https://doi.org/10. with false-positive rate determination for pep-
1002/cbic.200500151 tide identification in large-scale proteomics
3. Barnes S, Benton HP, Casazza K, Cooper SJ, experiments using a combined reverse and for-
Cui X, Du X, Engler J, Kabarowski JH, Li S, ward peptide sequence database strategy. J Pro-
Pathmasiri W, Prasain JK, Renfrow MB, Tiwari teome Res 6(1):392–398. https://doi.org/10.
HK (2016) Training in metabolomics 1021/pr0603194
research. I. Designing the experiment, collect- 7. Nesvizhskii AI (2007) Protein identification by
ing and extracting samples and generating tandem mass spectrometry and sequence data-
metabolomics data. J Mass Spectrom 51 base searching. Methods Mol Biol
(7):461–475. https://doi.org/10.1002/jms. 367:87–119. https://doi.org/10.1385/1-
3782 59745-275-0:87
4. Cox J, Mann M (2008) MaxQuant enables 8. Anjo SI, Santa C, Manadas B (2017) SWATH-
high peptide identification rates, individualized MS as a tool for biomarker discovery: from
p.p.b.-range mass accuracies and proteome- basic research to clinical applications. Proteo-
wide protein quantification. Nat Biotechnol mics 17(3–4). https://doi.org/10.1002/
26(12):1367–1372. https://doi.org/10. pmic.201600278
1038/nbt.1511
10 Stephen Barnes

9. Peterson AC, Russell JD, Bailey DJ, Westphall reporting standards for metabolite annotation
MS, Coon JJ (2012) Parallel reaction monitor- and identification in metabolomic studies.
ing for high resolution and high mass accuracy Gigascience 2(1):13. https://doi.org/10.
quantitative, targeted proteomics. Mol Cell 1186/2047-217X-2-13
Proteomics 11(11):1475–1488. https://doi. 18. Sumner LW, Amberg A, Barrett D, Beale MH,
org/10.1074/mcp.O112.020131 Beger R, Daykin CA, Fan TW, Fiehn O,
10. Rinschen MM, Ivanisevic J, Giera M, Siuzdak Goodacre R, Griffin JL, Hankemeier T,
G (2019) Identification of bioactive metabo- Hardy N, Harnly J, Higashi R, Kopka J, Lane
lites using activity metabolomics. Nat Rev Mol AN, Lindon JC, Marriott P, Nicholls AW, Reily
Cell Biol 20(6):353–367. https://doi.org/10. MD, Thaden JJ, Viant MR (2007) Proposed
1038/s41580-019-0108-4 minimum reporting standards for chemical
11. Collins FS (2004) The case for a US prospec- analysis Chemical Analysis Working Group
tive cohort study of genes and environment. (CAWG) Metabolomics Standards Initiative
Nature 429(6990):475–477. https://doi. (MSI). Metabolomics 3(3):211–221. https://
org/10.1038/nature02628 doi.org/10.1007/s11306-007-0082-2
12. Manrai AK, Cui Y, Bushel PR, Hall M, 19. Schymanski EL, Jeon J, Gulde R, Fenner K,
Karakitsios S, Mattingly CJ, Ritchie M, Ruff M, Singer HP, Hollender J (2014) Iden-
Schmitt C, Sarigiannis DA, Thomas DC, tifying small molecules via high resolution mass
Wishart D, Balshaw DM, Patel CJ (2017) spectrometry: communicating confidence.
Informatics and data analytics to support Environ Sci Technol 48(4):2097–2098.
exposome-based discovery for public health. https://doi.org/10.1021/es5002105
Annu Rev Public Health 38:279–294. 20. Schrimpe-Rutledge AC, Codreanu SG, Sher-
https://doi.org/10.1146/annurev- rod SD, McLean JA (2016) Untargeted meta-
publhealth-082516-012737 bolomics strategies-challenges and emerging
13. Li S, Cirillo P, Hu X, Tran V, Krigbaum N, directions. J Am Soc Mass Spectrom 27
Yu S, Jones DP, Cohn B (2019) Understanding (12):1897–1905. https://doi.org/10.1007/
mixed environmental exposures using metabo- s13361-016-1469-y
lomics via a hierarchical community network 21. Chong J, Soufan O, Li C, Caraus I, Li S,
model in a cohort of California women in Bourque G, Wishart DS, Xia J (2018) Meta-
1960’s. Reprod Toxicol. pii: S0890-6238(18) boAnalyst 4.0: towards more transparent and
30603-8. https://doi.org/10.1016/j. integrative metabolomics analysis. Nucleic
reprotox.2019.06.013 Acids Res 46(W1):W486–W494. https://doi.
14. Stolz A, Jooss K, Hocker O, Romer J, org/10.1093/nar/gky310
Schlecht J, Neususs C (2019) Recent advances 22. Li S, Park Y, Duraisingham S, Strobel FH,
in capillary electrophoresis-mass spectrometry: Khan N, Soltow QA, Jones DP, Pulendran B
instrumentation, methodology and applica- (2013) Predicting network activity from high
tions. Electrophoresis 40(1):79–112. https:// throughput metabolomics. PLoS Comput Biol
doi.org/10.1002/elps.201800331 9(7):e1003123. https://doi.org/10.1371/
15. Onjiko RM, Portero EP, Moody SA, Nemes P journal.pcbi.1003123
(2017) In situ microprobe single-cell capillary 23. Nakayasu ES, Nicora CD, Sims AC, Burnum-
electrophoresis mass spectrometry: metabolic Johnson KE, Kim YM, Kyle JE, Matzke MM,
reorganization in single differentiating cells in Shukla AK, Chu RK, Schepmoes AA, Jacobs
the live vertebrate (Xenopus laevis) embryo. JM, Baric RS, Webb-Robertson BJ, Smith
Anal Chem 89(13):7069–7076. https://doi. RD, Metz TO (2016) MPLEx: a robust and
org/10.1021/acs.analchem.7b00880 universal protocol for single-sample integrative
16. Members MSIB, Sansone SA, Fan T, proteomic, metabolomic, and lipidomic ana-
Goodacre R, Griffin JL, Hardy NW, lyses. mSystems 1(3). https://doi.org/10.
Kaddurah-Daouk R, Kristal BS, Lindon J, 1128/mSystems.00043-16
Mendes P, Morrison N, Nikolau B, 24. Guo L, Milburn MV, Ryals JA, Lonergan SC,
Robertson D, Sumner LW, Taylor C, van der Mitchell MW, Wulff JE, Alexander DC, Evans
Werf M, van Ommen B, Fiehn O (2007) The AM, Bridgewater B, Miller L, Gonzalez-Garay
metabolomics standards initiative. Nat Bio- ML, Caskey CT (2015) Plasma metabolomic
technol 25(8):846–848. https://doi.org/10. profiles enhance precision medicine for volun-
1038/nbt0807-846b teers of normal health. Proc Natl Acad Sci U S
17. Salek RM, Steinbeck C, Viant MR, A 112(35):E4901–E4910. https://doi.org/
Goodacre R, Dunn WB (2013) The role of 10.1073/pnas.1508425112
 Chapter 2

Metabolomics Data Processing Using XCMS

Xavier Domingo-Almenara and Gary Siuzdak

Abstract
XCMS is one of the most used software for liquid chromatography–mass spectrometry (LC-MS) data
processing and it exists both as an R package and as a cloud-based platform known as XCMS Online. In this
chapter, we first overview the nature of LC-MS data to contextualize the need for data processing software.
Next, we describe the algorithms used by XCMS and the role that the different user-defined parameters play
in the data processing. Finally, we describe the extended capabilities of XCMS Online.

Key words XCMS, Liquid chromatography, Mass spectrometry, Metabolomics, Data processing

1 Introduction

Untargeted or global metabolomics aims at quantifying as many

metabolites as possible in biological samples [1]. Specifically, liquid
chromatography–electrospray ionization mass spectrometry
(LC-ESI-MS or simply LC-MS) is one of the most used technolo-
gies for that purpose, due to its easy sample preparation process and
wide metabolite coverage. LC-MS, as we will further explain in the
next sections, generates large and complex datasets comprising
thousands of chromatographic peaks [2]. Computational tools are
therefore needed to process this data and convert it into interpret-
able information.
The untargeted computational data processing workflow was
popularized in 2006 by the papers describing two of the most
widely used tools in untargeted metabolomics: XCMS [3] and
mzMine [4, 5]. The untargeted metabolomics data processing
workflow typically consists of a peak-picking process followed by a
peak alignment which ultimately yields a set of features, that is, a
peak or a set of peaks across samples with a unique m/z and
retention time. The aim of this process is to transform raw data
into a matrix containing the list of observed features with their
relative peak areas or intensities for each sample in a given experi-
ment. This list is known as the feature list, and it is used to compare

Shuzhao Li (ed.), Computational Methods and Data Analysis for Metabolomics, Methods in Molecular Biology, vol. 2104,
https://doi.org/10.1007/978-1-0716-0239-3_2, © Springer Science+Business Media, LLC, part of Springer Nature 2020

11
12 Xavier Domingo-Almenara and Gary Siuzdak

the peak area (and thus relative concentration) between the same
ion peak across samples, allowing us to find statistically significant
dysregulated peaks for a given phenotype.
The processes that usually follows peak-picking and alignment
are statistical analysis - to find dysregulated peaks in a given
phenotype-, or metabolite annotation. Computational metabolite
annotation aims at providing chemical information on the observed
features [6]. This annotation usually consists in determining what
features stem from the same and each metabolite, determining the
nature of features (e.g., determine if a feature is a protonated/
deprotonated species, an adduct, an in-source fragment, an isotope,
a dimer, etc.), and providing with putative metabolite identifica-
tions. As an example, the simplest annotation process consists in
using the m/z of the observed features and match them against the
database to determine their potential identity. This is considered,
however, a very weak annotation, and as we will discuss further in
this chapter, more advanced techniques based on in-source frag-
ment annotation can be used.
In this chapter, we describe XCMS, a computational workflow
that integrates peak picking with alignment. It is available as an R
package and also as a cloud-based resource. The latter encompasses
a comprehensive suite of tools through an easy-to-use visual inter-
face that allows for data to be shared with the community or
collaborators and has a higher computational performance. It also
integrates tools and modules that enable systems biology analysis
and advanced metabolite annotation through its integration with
the METLIN spectral library (Fig. 1). A version of XCMS Online
for targeted metabolomics is also available, called XCMS-MRM.

Fig. 1 XCMS Online overview. Through an interactive user interface, XCMS Online allows for data sharing, data
streaming, advanced metabolite annotation, and systems biology analysis and multi-omics integration
 XCMS Data Processing 13

2 Overview of LC-MS Data

When samples are injected into an LC-MS system, molecules are

separated on the chromatographic column (LC). At the end of the
column, the MS scans at a given scan rate (scans per second)
yielding a signal that depends on the molecules being detected.
This process generates a three-dimensional set of chromatographic
peaks, each with specific retention time (RT), m/z, and intensity, as
a result of the underlying molecules in the analyzed sample.
In LC-ESI-MS, each detected molecule generates multiple
chromatographic peaks. In most cases, a single molecule will gen-
erate a protonated/deprotonated species when using common
untargeted metabolomics chromatographic methods. In addition,
isotope, adducts and in-source fragments will also be observed.
Particularly, in-source fragmentation, a common phenomenon in
LC-ESI-MS, leads to the observation of (in-source) fragments or
neutral losses. Some of these neutral losses are known as common
neutral losses (e.g., -H2O, -NH3, -HCOOH), as they are observed
across a wide range of biological organic molecules. Some other
peaks including dimers or multimers are less frequently observed.
For the reasons mentioned above, LC-MS generates large and
complex datasets comprising thousands of chromatographic peaks.
Papers using untargeted metabolomics usually report the number
of features observed in their experiments. However, there is still not
a consensus on how many metabolites are approximately observed
in routine metabolomics analyses. Through computational annota-
tion algorithms, some studies have reported between 1000 meta-
bolites in Escherichia coli [7] to 2000 metabolites in Saccharomyces
cerevisiae [8]. How to effectively attribute all the peaks to bona fide
metabolites is an active research area.
In that sense, peak or metabolite annotation in LC-MS aims at
(1) grouping all features stemming from the same metabolite and
(2) assigning putative metabolite identities to observed features. It
is also worth mentioning that this term is also employed to refer to
the annotation of fragments and metabolites in tandem MS
(LC-MS/MS).

3 XCMS Workflow and Algorithms

As mentioned before, XCMS is a computational tool that integrates

peak picking with alignment to generate a list of features from raw
MS data. Peak picking is a computational procedure used to detect
peaks in MS data and integrate their area. Historically, two types of
peak picking algorithms have been the most popular: the match
filter algorithm and its advanced variation for high-resolution mass
spectrometry data known as the centWave algorithm. Natively,
14 Xavier Domingo-Almenara and Gary Siuzdak

XCMS has the option of using either the original match filter
algorithm or the centWave. In this section, we will describe these
two types of algorithms.
When peaks in raw MS data are detected across multiple sam-
ples, the same peak stemming from a given ion will ideally appear
across the different samples with different intensities. Due to the
high accuracy of modern mass spectrometers such as TOF and
Orbitrap instruments, m/z values across samples tend to have var-
iations as low as 5 and 1 ppm, respectively. On the other hand, the
retention time will have larger variations across samples. Therefore,
peaks from the same ion detected across different samples need to
be aligned and grouped, that is, their area needs to be assigned to
the same row in a data matrix so they can be quantitatively com-
pared across different samples. Specifically, in XCMS, this peak
alignment is performed by a two-step process known as peak
grouping and retention time alignment. These two processes will
be described in the following sections.
After peaks are detected and aligned, statistics can be used to
find features showing statistically significant changes among peak
areas among phenotypes. However, after peak-picking and align-
ment, the existence of missing values is a frequent scenario, typically
affecting more than 80% of the detected features [9]. This means
that around 80% of the detected features will have missing values
for certain samples, that is, zero peak area or intensity. This occurs
when for a given group of samples, the ion peak is below the
detection limit. In other cases, the peak is observed above the
detection limit but the peak picking algorithm has failed in detect-
ing it due to low peak intensity or because the peak is masked by
noise or other coeluting peaks. Having missing values reduce the
power of statistical tests and analysis and can lead to biased results
[9]. This is of special importance if multivariate analyses are applied,
as missing values will bias the results. To tackle this problem there
exist different strategies known as missing value imputation strate-
gies [10]. Among these strategies, a commonly accepted approach
is known as filling peaks, where “missing” peaks are searched again
in the raw data. XCMS uses a fill peaks strategy to “fill” the
“missing” peaks in the feature list and this strategy will be detailed
in the following sections.
The following sections explain in detail the different steps of
the XCMS workflow (Fig. 2) which consist of (1) peak picking,
(2) peak grouping and retention time alignment, and (3) fill peaks.

3.1 Peak Picking: Peak picking aims at finding the chromatographic peaks stemming
Operation from molecules eluting from the chromatographic column. Histor-
and Algorithms ically, these have been based on filtering algorithms, which filter the
signal to clear it from noise and thus distinguish the underlying
peaks from the raw signal. There are many different peak picking
algorithms that have been designed, some of which are targeted
 XCMS Data Processing 15

Fig. 2 XCMS general workflow. The workflow consists of peak detection,

grouping of peaks into features, alignment, missing peaks filling, and the
statistical analysis of features across classes

specifically at dealing with certain types of data. In LC-MS data the

peak detectors need to process the complex three-dimensional
LC-MS data to find these peaks. Some algorithms such as the
“centWave” algorithm need the data to be reduced beforehand.
This reduction consists of converting the data, which is typically
acquired in profile mode, into centroid mode data [11]. This can
dramatically reduce the data size and make the processing of the
data far simpler. While most vendors have their own centroiding
algorithms, open source alternatives can also be used. However,
users should be careful if the mass accuracy is retained with these
algorithms.
Specifically, in XCMS, peak picking can be performed via two
algorithms: Matched filter and centWave.

3.1.1 Matched Filter It is the original algorithm of detection of XCMS [3]. To detect
peaks, the algorithm first performs a “binning” procedure consist-
ing in slicing the data into bins of 0.1 m/z units (defined by the
parameter step). For each bin, the algorithm then detects any peak
that is above the signal-to-noise threshold as defined by the S/N
ratio cutoff parameter (default 10). This peak detection is per-
formed by leveraging the typical Gaussian-like shape of chro-
matographic peaks to detect them. That means that if these data
points fit well into a second derivative Gaussian (essentially a nor-
mal distribution) they are detected as a peak. The width of this
Gaussian is defined by the full width at half maximum (FWHM)
parameter, a value in seconds. The lower the value of this parame-
ter, the more likely is to detect false positive peaks (noise detected as
a peak). To make sure that peaks are not split between two m/z bins
the algorithm combines pairs of consecutive m/z bins. This algo-
rithm was originally designed for data acquired by low resolution
instruments such as single quadrupole mass spectrometers, where
the highest mass accuracy was around 0.1 Da. For current high-
resolution mass spectrometers (HRMS), the centWave algorithm is
recommended (described below). However, if the MatchedFilter is
to be applied to HRMS data, the binning value should be signifi-
cantly lowered. The output of the algorithm is a peak list where
each m/z, RT (and the deviations) and integrated peak intensity
(peak area) are given.
16 Xavier Domingo-Almenara and Gary Siuzdak

3.1.2 centWave Published in 2008 [12], this algorithm is a high resolution and
consequently high mass accuracy peak detection algorithm. The
“centWave” algorithm consists also of two steps. The first step
consists in a dynamic binning that finds potential areas containing
peaks, known as region of interest (ROI). The second step per-
forms the peak detection within these ROIs using a highly sensitive
wavelet filter (Fig. 3a). This algorithm looks for ROI consisting of
data points that show low m/z deviations and increase and then
decrease in intensity (Fig. 3a). The main parameters that control
this behavior are the “ppm” parameter that selects the m/z span of
the ROIs and the peak width which is measured in seconds and
dictates how long the peak should be in chromatographic time.
Once a ROI fulfilling these requirements has been discovered it is
analyzed by the wavelet filter. The wavelet that is used is a Mexican
hat wavelet that models the peak shape and allows for the selection
of multiple closely eluting peaks within this ROI (Fig. 3b). The

Fig. 3 The centWave peak detection overview. First, regions of interest (ROI) are detected. For each detected
ROI, a wavelet filter is applied and peaks within the ROI are detected
 XCMS Data Processing 17

scale or height of the peak is modulated until the best fit is achieved.
If the fitting parameters are not satisfied, then the ROI is rejected as
a peak. The output of this algorithm is the same as in the Match-
edFilter and consists of the m/z, RT (deviations of each) and the
integrated intensities.

3.2 Peak Alignment: In XCMS, what is considered as peak alignment is performed by a

Peak Grouping two-step approach consisting of peak grouping and alignment
and Retention Time across samples. The order of these two steps (which one will be
Alignment applied first) depends on which algorithm will be used for the
alignment.
When the original “peakGroups” alignment method is used,
peaks must be first grouped into features. This is done by an
algorithm named “peakDensity,” which uses a similar approach as
in the “matchedFilter” algorithm described above, but on the pro-
cessed peak list instead of the raw data. Using thin m/z slices
(defined by the “mzwid” parameter), the algorithm finds groups
of peaks across samples that cluster around a certain RT. This is to
find “well-behaved” peak groups (WBPG) that will be subsequently
used to align the rest of the chromatogram yielding a higher
accuracy in overall peak grouping. To find these WBPG, the algo-
rithm uses a kernel density filter that pinpoints areas with groups of
“well-behaved” peaks. The filter can be manipulated by the “bw”
parameter or bandwidth which is defined in seconds. For HPLC
data, 30 s is a common default value. WBPG need to be scattered
across the retention time to allow for a good nonlinear correction.
Using the median retention time of each WBPG, an alignment
profile can be made by a regression technique called LOcally Esti-
mated Scatter-plot Smoother (LOESS). This technique is a non-
parametric technique that fits a smoothed line through each
chromatogram and compares these smoothed lines across chroma-
tograms to finally correct the retention time across samples.
An alternative to the nonlinear retention time alignment algo-
rithm is the “obiwarp” algorithm. The obiwarp does not require
this pregrouping stage and can directly align peaks across samples.
The “obiwarp” algorithm was implemented into XCMS by Prince
et al. [13]. This algorithm uses a technique called dynamic pro-
gramming to find the best alignment between two chromatograms.
The algorithm works in a pairwise fashion where it first finds a
median retention profile across multiple chromatograms, and
then finds the path to iteratively align all the chromatograms to
that median profile. This algorithm can be computationally
demanding and time consuming.
Regardless of the retention time alignment algorithm used, the
next step is always to group the peaks. By grouping peaks, the
algorithm outputs features, defined as a single or a set of peaks
across samples with a unique RT and m/z. The above “peakDen-
sity” algorithm can be used again to complete this task. Now that
18 Xavier Domingo-Almenara and Gary Siuzdak

the peaks have been aligned the “bw” parameter can be greatly
reduced. As an additional robustness filter, there is a parameter
called “minfrac.” This parameter states that for any one class
(e.g., KO for knockout, WT for wild type), there must be at least
a certain fraction of peaks present in that feature to be considered as
a valid feature. For example, if we have 6 samples in a KO class and
6 samples in a WT class, then for a particular m/z and RT we need
to have at least 3 peaks present in either class if “minfrac” is set at
50%. It should be not that it is “any one class” and not both classes.
Therefore, if KO has 3 peaks and WT has 0, it is still a valid feature.
This parameter can be good for increasing the robustness when we
need to seek peaks in all samples.

3.3 Fill Peaks As commented before, two main causes can lead to peaks being
missed. The first is that the peak has not been correctly detected or
aligned by the algorithm, or the peak is in under the detection limit.
The fill peaks will effectively resolve and find missing peaks when
the problem is due to the algorithm (the peak exist in the data but it
was not correctly detected or aligned), leveraging the information
from other samples where the peak has been detected (RT and m/
z). In the second case, where the peak is under the detection limit,
the fill peaks step will use the background noise in the area where
the peak should be expected to determine the missing peak value.
Because there may be biological reasons that a peak is missing in
one class of samples not the others, the user can define the number
of peaks that need to be found in a single class.

4 XCMS R Package

As commented before, XCMS comes as an open-source R package.

The open source nature of XCMS allows it to be updated and
improved by the community.
The R version of XCMS is a command line driven interface that
lets users interact with the data directly on their systems via a
collection of commands. As part of this chapter we have included
workflows for both the original and alternative pipelines.
The updated XCMS (named XCMS 3 to avoid confusion from
XCMS2 [14]) uses new R capabilities, packages and objects to
reduce memory requirements. For instance, now XCMS 3 reduces
memory requirements by not loading entire raw data files into
memory. Instead, it only loads selections of what is currently
being handled. Additionally, the new objects also allow users to
save their object files and see what methods have been called on the
data, stored in the history of the object.
 XCMS Data Processing 19

5 XCMS Online: Extended Capabilities

XCMS Online, which was first introduced in 2012 [15], allows

users to have a cloud based and graphical system to use XCMS.
The added benefits of this are that local hardware is no longer
required to process data and data can be easily shared between
users. Users have a selection of simple options to set themselves
on the data processing path. Moreover, it allows users to perform
different processing types (job types) depending on their experi-
ment. These job types are outlined below:
Single job—Either a single file or single class of data.
Pairwise job—The classic knockout vs. wild-type experiment where
a simple two class system is needed. Statistical choices can be
parametric or nonparametric with both unpaired samples and
paired sample types.
Multigroup job—A very dynamic job type, useful for time series
where many classes are going to be compared. Statistical
choices are ANOVA and Kruskal-Wallis, both with post hoc
analysis.
In addition, XCMS Online allows data to be streamed directly
to the cloud for processing. It allows for advanced metabolite
annotation based on in-source fragment matching, and it allows
for systems biology analysis and multi-omics integration (Fig. 1).
More recently, XCMS-MRM Online, the targeted counterpart of
XCMS, was released [16]. The following section gives an overview
of these extended capabilities.

5.1 Data Streaming Data streaming emulates the streaming video platforms where users
with XCMS can directly watch movies without the need to wait until the movie
has been completely downloaded. In the case of LC-MS data, we
typically have to wait until all samples have been acquired and
upload all the samples to the XCMS cloud-based system. However,
XCMS Online allows LC-MS data to be streamed [17]. By instal-
ling a program on their instrument computer and login to XCMS
Online, users can stream their data directly to the cloud and start
processing the data while the data of the rest of samples are being
acquired. The parameters for the job can be chosen beforehand.
The streaming process dramatically reduces the overall dead time
from acquisition to processing.

5.2 MISA: METLIN- XCMS Online also integrates an algorithm for advanced metabolite
Guided In-Source annotation, known as the METLIN-guided in-source annotation
Annotation (MISA) algorithm [18]. This algorithm aims at using the informa-
tion from in-source fragments (ISF) naturally occurring in MS1
data to provide a robust peak and metabolite annotation. We
20 Xavier Domingo-Almenara and Gary Siuzdak

observed that more than 80% of molecules in METLIN readily

dissociate into multiple fragments at low collision energies
[6]. This implies that a considerable fraction of features usually
observed in untargeted experiments are ISF. Existing tools for
metabolite annotation typically take into account only common
neutral losses as a result of in-source fragmentation. These common
neutral losses are not specific for each metabolite. On the other
hand, ISF provide more specific information on the metabolite due
to low-energy fragmentation.
In-source fragments observed in LC-MS are similar to those
observed in low-energy collision-induced dissociation (CID) tan-
dem MS (MS/MS) spectra. In that sense, low-energy spectra in
METLIN can be used to guide this annotation and detect the
presence of specific ISF from metabolites with experimental
MS/MS spectra in METLIN (Fig. 4). The current version of

Fig. 4 In-source fragment annotation examples. LC-MS experimental peaks

(black) and low energy MS/MS spectra extracted from METLIN (color,
negatively rotated), for the metabolites tyrosine and phosphocholine.
Protonated ion/precursor is noted (asterisk)
 XCMS Data Processing 21

METLIN now contains experimental MS/MS spectra for more

than 500,000 molecular standards in both positive and negative
mode and at different collision energies, including experimental
spectra acquired at low-energy (0 and 10 eV).
MISA is applied after data is processed by XCMS Online. MISA
searches for features corresponding to protonated/deprotonated
or adduct ion species from available MS/MS spectra in METLIN.
Next, MISA searches for the corresponding ISF of those molecules.
If both the protonated/deprotonated ion species and one or more
ISF are found, MISA assigns the feature with a putative identity. In
addition, it provides two scores to assess the likelihood of that
annotation being correct. More information on both MISA and
the scores can be found in the original study [18].

5.3 XCMS-Guided A system levels analysis can provide important insights to under-
Systems Biology stand biochemical mechanisms. XCMS online enables metabolo-
mics data to be projected onto metabolic pathways and integrate it
with transcriptomics and proteomics data [19].
To project quantitative metabolic data onto metabolic net-
works, first, metabolites need to be identified. Metabolite identifi-
cation is a process that heavily relies on manual labor and expert
curation. XCMS Online uses an automated predictive pathway
analysis method, developed by Li et al. and known as Mummichog
[20], that bypasses metabolite identification and instead uses bio-
chemical information to annotate features and project them onto
metabolic pathways. Scientist can then interpret the data and for-
mulate biochemical mechanisms and hypothesis, and then confirm
these identifications via tandem MS analysis. More details of this
algorithm can be found in the original study [20] and Chapter 19
of this book.
In that sense, in XCMS Online-guided systems biology, users
can directly map their results into metabolic networks, without the
need to transfer data among different applications. Users can
upload gene and protein data to overlay it within the pathways.
Results are shown as a table and also by an interactive Pathway
Cloud plot. The Pathway Cloud plot shows dysregulated pathways,
ordering them by overlap percentage to other omics data and
statistical significance. To the date, there are over 7600 metabolic
models available for pathway analysis from BioCyc4 v19.5–20.0.

5.4 XCMS-MRM Small molecule quantification is performed using triple quadrupole

MS configured for multiple reaction monitoring (MRM)
[21]. MRM is considered the gold standard for targeted quantita-
tive analysis due to its high sensitivity and specificity. In order to
process MRM data, XCMS-MRM was designed and it constitutes
the XCMS Online counterpart for processing data from targeted
metabolomics assays [16]. XCMS-MRM recognizes raw data files
in any vendor format, providing automatic transition peak
22 Xavier Domingo-Almenara and Gary Siuzdak

detection, area integration and alignment, minimizing false peak

integration, and reducing data analysis time. Absolute concentra-
tions can be acquired through stable-isotope dilution, external
calibration and standard addition methods. XCMS-MRM quanti-
tative results include quality control indicators to assess accuracy
and specificity, and limits of detection. It also assesses statistical
significance for metabolite concentration changes. The statistical
graphics and result visualization are unique features of XCMS-
MRM.

6 Conclusions

LC-MS is a widely used metabolomics technology, and XCMS

enables it by meeting the demands of this ever-growing scientific
community. In this chapter, we have described the underlying
algorithms of XCMS, and the XCMS Online platform. XCMS is a
software program for LC-MS data processing, implemented as an R
package, widely used by the community and actively maintained.
XCMS Online is a cloud-based resource and extends its capabilities
to a full suite for metabolomics data analysis. Hosted on high-
performance dedicated servers at the Scripps Research Institute,
XCMS Online offers outstanding processing speed that is now
easily accessed by thousands of users. Together, with a freely avail-
able R package and a cloud-based data processing service, XCMS
facilitates data analysis and result sharing across all vendor
platforms.

Acknowledgments

This research was partially funded by the US National Institutes of

Health grants R35 GM130385, P30 MH062261, P01 DA026146
and U01 CA235493; and by Ecosystems and Networks Integrated
with Genes and Molecular Assemblies (ENIGMA), a Scientific
Focus Area Program at Lawrence Berkeley National Laboratory
for the U.S. Department of Energy, Office of Science, Office of
Biological and Environmental Research, under contract number
DE-AC02-05CH11231. This research benefited from the use of
credits from the National Institutes of Health (NIH) Cloud Credits
Model Pilot, a component of the NIH Big Data to Knowledge
(BD2K) program.
 XCMS Data Processing 23

References
1. Patti GJ, Yanes O, Siuzdak G (2012) Metabo- 10. Runmin W, Wang J, Su M, Jia E, Chen S,
lomics: the apogee of the omic triology. Nat Chen T, Ni Y (2018) Missing value imputation
Rev Mol Cell Biol 13(4):263–269. https:// approach for mass spectrometry-based meta-
doi.org/10.1038/nrm3314 bolomics data. Sci Rep 8. https://doi.org/10.
2. Vinaixa M, Samino S, Saez I, Duran J, Guino- 1038/s41598-017-19120-0
vart JJ, Yanes O (2012) A guideline to univari- 11. Vereyken L, Dillen L, Vreeken RJ, Cuyckens F
ate statistical analysis for LC/MS-based (2019) High-resolution mass spectrometry
untargeted metabolomics-derived data. Meta quantification: impact of differences in data
2(4):775–795. https://doi.org/10.3390/ processing of centroid and continuum data. J
metabo2040775 Am Soc Mass Spectrom 30(2):203–212.
3. Smith CA, Want EJ, O’Maille G, Abagyan R, https://doi.org/10.1007/s13361-018-2101-
Siuzdak G (2006) XCMS: processing mass 0
spectrometry data for metabolite profiling 12. Tautenhahn R, Böttcher C, Neumann S
using nonlinear peak alignment, matching, (2008) Highly sensitive feature detection for
and identification. Anal Chem 78 high resolution LC/MS. BMC Bioinformatics
(3):779–787. https://doi.org/10.1021/ 9:504. https://doi.org/10.1186/1471-2105-
ac051437y 9-504
4. Mikko K, Miettinen J, Oresic M (2006) 13. Prince JT, Marcotte EM (2006) Chro-
MZmine: toolbox for processing and visualiza- matographic alignment of ESI-LC-MS proteo-
tion of mass spectrometry based molecular pro- mics data sets by ordered bijective interpolated
file data. Bioinformatics 22(5):634–636. warping. Anal Chem 78(17):6140–6152.
https://doi.org/10.1093/bioinformatics/ https://doi.org/10.1021/ac0605344
btk039. 14. Benton HP, Wong DM, Trauger SA, Siuzdak G
5. Tomás P, Castillo S, Villar-Briones A, Oresic M (2008) XCMS2: processing tandem mass spec-
(2010) MZmine 2: modular framework for trometry data for metabolite identification and
processing, visualizing, and analyzing mass structural characterization. Anal Chem 80
spectrometry-based molecular profile data. (16):6382–6389. https://doi.org/10.1021/
BMC Bioinformatics 11:395. https://doi. ac800795f
org/10.1186/1471-2105-11-395 15. Tautenhahn R, Patti GJ, Rinehart D, Siuzdak
6. Domingo-Almenara X, Montenegro-Burke JR, G (2012) XCMS online: a web-based platform
Benton PH, Siuzdak G (2018) Annotation: a to process untargeted metabolomic data. Anal
computational solution for streamlining meta- Chem 84(11):5035–5039. https://doi.org/
bolomics analysis. Anal Chem 90(1):480–489. 10.1021/ac300698c
https://doi.org/10.1021/acs.analchem. 16. Domingo-Almenara X, Montenegro-Burke JR,
7b03929 Ivanisevic J, Thomas A, Sidibé J, Teav T, Guijas
7. Mahieu NG, Patti GJ (2017) Systems-level C et al (2018) XCMS-MRM and METLIN-
annotation of a metabolomics data set reduces MRM: a cloud library and public resource for
25 000 features to fewer than 1000 unique targeted analysis of small molecules. Nat Meth-
metabolites. Anal Chem 89 ods 15(9):681–684. https://doi.org/10.
(19):10397–10406. https://doi.org/10. 1038/s41592-018-0110-3
1021/acs.analchem.7b02380 17. Montenegro-Burke JR, Aisporna AE, Benton
8. Lin W, Xing X, Chen L, Yang L, Su X, HP, Rinehart D, Fang M, Huan T, Warth B
Rabitz H, Lu W, Rabinowitz JD (2019) Peak et al (2017) data streaming for metabolomics:
annotation and verification engine for untar- accelerating data processing and analysis from
geted LC–MS metabolomics. Anal Chem 91 days to minutes. Anal Chem 89
(3):1838–1846. https://doi.org/10.1021/ (2):1254–1259. https://doi.org/10.1021/
acs.analchem.8b03132 acs.analchem.6b03890
9. Trinh DK, Wahl S, Raffler J, Molnos S, 18. Domingo-Almenara X, Montenegro-Burke JR,
Laimighofer M, Adamski J, Suhre K et al Guijas C, Majumder EL-W, Benton HP, Siuz-
(2018) Characterization of missing values in dak G (2019) Autonomous METLIN-guided
untargeted MS-based metabolomics data and in-source fragment annotation for untargeted
evaluation of missing data handling strategies. metabolomics. Anal Chem 91(5):3246–3253.
Metabolomics 14(10). https://doi.org/10. https://doi.org/10.1021/acs.analchem.
1007/s11306-018-1420-2 8b03126
24 Xavier Domingo-Almenara and Gary Siuzdak

19. Tao H, Forsberg EM, Rinehart D, Johnson throughput metabolomics. PLoS Comput Biol
CH, Ivanisevic J, Paul Benton H, Fang M 9(7). https://doi.org/10.1371/journal.pcbi.
et al (2017) Systems biology guided by 1003123
XCMS online metabolomics. Nat Methods 14 21. Vinzenz L, Picotti P, Domon B, Aebersold R
(5):461–462. https://doi.org/10.1038/ (2008) Selected reaction monitoring for quan-
nmeth.4260 titative proteomics: a tutorial. Mol Syst Biol
20. Li S, Park Y, Duraisingham S, Strobel FH, 4:222. https://doi.org/10.1038/msb.2008.
Khan N, Soltow QA, Jones DP, Pulendran B 61
(2013) Predicting network activity from high
 Chapter 3

Metabolomics Data Preprocessing Using ADAP and MZmine 2

Xiuxia Du, Aleksandr Smirnov, Tomáš Pluskal, Wei Jia,
and Susan Sumner

Abstract
The informatics pipeline for making sense of untargeted LC–MS or GC–MS data starts with preprocessing
the raw data. Results from data preprocessing undergo statistical analysis and subsequently mapped to
metabolic pathways for placing untargeted metabolomics data in the biological context. ADAP is a suite of
computational algorithms that has been developed specifically for preprocessing LC–MS and GC–MS data.
It consists of two separate computational workflows that extract compound-relevant information from raw
LC–MS and GC–MS data, respectively. Computational steps include construction of extracted ion chro-
matograms, detection of chromatographic peaks, spectral deconvolution, and alignment. The two work-
flows have been incorporated into the cross-platform and graphical MZmine 2 framework and ADAP-
specific graphical user interfaces have been developed for using ADAP with ease. This chapter summarizes
the algorithmic principles underlying key steps in the two workflows and illustrates how to apply ADAP to
preprocess LC–MS and GC–MS data.

Key words ADAP, MZmine 2, Metabolomics, LC–MS, GC–MS, Data preprocessing, Peak picking,
Alignment, Spectral deconvolution, Visualization

1 Introduction

Untargeted metabolomics, detection and relative quantitation of

ideally all metabolites in a biological system, has become a powerful
discovery tool in many scientific disciplines. It has benefited greatly
from advances in mass spectrometry (MS) and chromatography. As
a result, liquid chromatography (LC) and gas chromatography
(GC) coupled to mass spectrometry (MS) have become primary
analytical platforms for untargeted metabolomics.
The informatics pipeline for making sense of the resulting LC–
MS and GC–MS data involves preprocessing of the raw mass spec-
tral data to detect chemical species, assignment of specific metabo-
lites to these species, and integration of these metabolites into a
coherent and physiologically meaningful integrated multi-omics
framework that can yield a holistic understanding of the biological

Shuzhao Li (ed.), Computational Methods and Data Analysis for Metabolomics, Methods in Molecular Biology, vol. 2104,
https://doi.org/10.1007/978-1-0716-0239-3_3, © Springer Science+Business Media, LLC, part of Springer Nature 2020

25
26 Xiuxia Du et al.

system (Fig. 1). As the first step of this informatics pipeline, data
preprocessing is critical for the success of a metabolomics study
because preprocessing errors can propagate downstream into spu-
rious or missing compound identifications and cause misinterpreta-
tion of the metabolome. Data preprocessing workflows (Fig. 2) in
open-source software tools generally consist of four sequential
steps after masses have been detected from profile mass spectra
(i.e., converting mass spectra from profile to centroid format).
These four steps are construction of extracted ion chromatograms
(EIC), detection of chromatographic peaks from EICs, peak
grouping for LC–MS data or spectral deconvolution for GC–MS
data, and alignment. ADAP (Automated Data Analysis Pipeline) is
one such open-source workflow that has been developed for pre-
processing both GC–MS and LC–MS data and incorporated into
the MZmine 2 framework [1].
In sections below, we first briefly describe the evolution of
ADAP and subsequently focus on describing how to carry out
LC–MS and GC–MS preprocessing workflows using primarily
ADAP modules in the MZmine 2 framework. These modules
include EIC construction, chromatographic peak detection, spec-
tral deconvolution, and alignment. To facilitate describing how to
specify relevant parameters, we provide a brief summary of the
algorithmic principles underlying these ADAP modules. To make
the workflow complete and this chapter self-contained, we provide
information on how to carry out other essential tasks using
non-ADAP modules in MZmine 2. These tasks include: (1) import
raw LC–MS and GC–MS data files into MZmine 2 and inspect
them, (2) detect masses from profile mass spectra (step 1 in Fig. 2),
(3) group chromatographic peaks (step 4 in Fig. 2) detected in LC–
MS data, and (4) export preprocessing results for downstream
statistical analysis, metabolite identification, and -omics data
integration.

raw GC-MS or data statistical -omics data integration

LC-MS data preprocessing analysis and metabolite identification

Fig. 1 Informatics pipeline for making sense of untargeted GC–MS and LC–MS
metabolomics data

1 2 3 4 5 6
detect masses from construct detect alignment / database
GC-MS mass spectra EICs chromatographic peaks
deconvolution
correspondence search

detect masses from construct detect alignment / database

LC-MS mass spectra EICs chromatographic peaks
annotation
correspondence search

Fig. 2 General computational workflows for preprocessing GC–MS and LC–MS data in existing open-source
software tools. EIC stands for extracted ion chromatogram
 Metabolomics Data Preprocessing Using ADAP and MZmine 2 27

2 Evolution of ADAP

Development of ADAP started in 2009. The first version of

ADAP was developed by Jiang et al. for fully automated preproces-
sing of raw GC–MS untargeted metabolomics data [2]. It was
comprised of a suite of algorithms for steps 2–5 of the preproces-
sing workflow for GC–MS data (Fig. 2). As a critical step in this
ADAP-GC workflow, spectral deconvolution has undergone signif-
icant improvements over the years made by Ni et al. [3, 4] for
ADAP-GC 3.0 and ADAP-GC 3.0 and Smirnov et al. [5] for
ADAP-GC 3.2.
The year of 2016 witnessed research and development efforts
by Myers et al. to equip ADAP with the capability to preprocess
high mass resolution LC–MS untargeted metabolomics data while
addressing the high rate of false positive peaks that had been
reported [6, 7]. Toward this end, ADAP algorithms were devel-
oped for constructing extracted ion chromatograms (EICs) and for
detecting chromatographic peaks from EICs. Following peak
detection, alignment methods in MZmine 2 can be used to correct
the retention time shift from sample to sample for a complete
preprocessing workflow for LC–MS data.
All of the aforementioned ADAP algorithms were written in
Java and have been incorporated into the open-source and graphi-
cal MZmine 2 framework. Furthermore, specific user-friendly
graphical user interfaces (GUI) have been developed to facilitate
users with using the ADAP modules within MZmine 2.

3 Install MZmine 2

MZmine 2 can be downloaded at http://mzmine.github.io/down

load.html. To start MZmine 2, users should unzip the downloaded
file and then open MZmine 2 by running the following script files
according to the operating system of your computer.
1. startMZmine_MacOSX.command
2. startMZmine_Windows.bat
3. startMZmine_Linux.sh

4 Preprocessing Workflow for LC–MS Data

4.1 Import and Raw data files are imported into MZmine 2 using the Raw data
Inspection of Raw Data methods drop-down menu as shown in Fig. 3. Acceptable file
Files formats include mzXML and netCDF. One of the greatest
strengths of MZmine 2 lies in the rich built-in visualization func-
tions that allow users to inspect the raw data, which greatly
28 Xiuxia Du et al.

Fig. 3 Import raw data into MZmine 2. (a) Raw data import is in the Raw
data methods drop-down menu. (b) Imported raw data files are listed under
Raw data files on the left

facilitates users with understanding the data and making informed

decisions to specify preprocessing parameters. Herein we demon-
strate only the visualization capabilities that can inform data pre-
processing. Readers are advised to explore the other visualization
capabilities in MZmine 2.
 Metabolomics Data Preprocessing Using ADAP and MZmine 2 29

Display of Raw Mass Spectra: Figure 4 shows the MZmine 2 cap-

abilities to display raw spectra and the spectra meta data that
includes spectra level (MS1 or MS2), acquisition time, type (p repre-
sents profile or c represents centroid), and polarity (+ represents
positive and - represents negative).

Fig. 4 Capabilities of MZmine 2 that allow users to inspect the raw spectra. (a)
List of spectra in a raw data file and the spectra meta data. (b) Double click on
any spectrum opens up a separate window displaying it
30 Xiuxia Du et al.

Display of Chromatograms: Base peak chromatograms (BPC)

and total ion chromatograms (TIC) can reveal retention time shift
among the data files and the approximate amount of retention time
correction that is needed via alignment. Figure 5 displays the BPCs
of 12 data files.
Display of m/z and Retention Time: The 2D visualizer in the
Visualization drop-down menu can provide an overview of the
ion species that should be detected by the peak picking algorithms
from the entire data file (Fig. 6). Each ion species is characterized by
a unique pair of m/z (y-axis) and retention time (x-axis).

4.2 Detect Masses The mass detection step detects mass centroids from profile mass
from Profile Mass spectra. MZmine 2 provides five centroiding methods that include
Spectra Centroid, Exact mass, Local maxima, Recursive thresh-
old, and Wavelet transform. The Centroid mass detector is
for spectra that have been centroided and the other four detectors
are for profile mass spectra only. The Exact mass detector is
suitable for high-resolution MS data, such as provided by FTMS
instruments. The Local maxima mass detector simply detects all
local maxima within a spectrum, except those signals below the
specified noise level. The Recursive threshold mass detector is
suitable for data that has too much noise for the Exact mass
detector to be used. The Wavelet transform mass detector is
suitable for both high-resolution and low-resolution data. It uses
the Ricker wavelet (also called Mexican Hat wavelet) and carry out
a continuous wavelet transform (CWT) of the continuous profile
spectra.
This Wavelet transform mass detector provides a sensitive
and robust way to detect masses (Fig. 7) and we describe it in more
detail herein. It requires users to set three parameters: noise level,
scale level, and wavelet window size. Noise level specifies the mini-
mum intensity level for a data point to be considered part of a
spectrum. All data points below this intensity level are ignored.
Scale level is the scale factor that either dilates or compresses the
wavelet signal. When it is small (e.g., below 10), the Ricker wavelet
is more contracted which in turn results in more noisy peaks being
detected. Wavelet window size ( % ) is the size of the window
used to calculate the wavelet signal. When the size of the window is
small, more noisy peaks can be detected. Among the three para-
meters, scale level, in particular, can have a large impact on mass
detection.
When the scale level is small, a significant number of very
narrow noise peaks can be detected. They are passed to the
subsequent EIC construction and can form false EIC peaks. As
the scale level increases, the number of detected noise peaks
decreases. However, a larger scale level could cause a noticeable
shift in the centroid m/z values. Figure 8c, d depicts the m/z values
 Metabolomics Data Preprocessing Using ADAP and MZmine 2 31

Fig. 5 Inspect BPCs. (a, b) Display BPCs by using the TIC/XIC visual-
izer in the visualization drop-down menu. (c) BPCs of 12 data files
32 Xiuxia Du et al.

Fig. 6 2D visualization of a raw data file. (a, b) 2D visualizer can be

accessed via the Visualization drop-down menu. (c) Ion species are
displayed via the 2D visualization
 Metabolomics Data Preprocessing Using ADAP and MZmine 2 33

Fig. 7 Mass detection in MZmine 2. (a) The Mass detection method can be accessed via the Raw data
methods draw-down menu. (b) Wavelet transform is one of the mass detection methods. Click the
button pointed to by the red arrow would open the window in (c) for specifying parameters (c) User-defined
parameters for the wavelet transform method. Check Show preview opens up the preview
window. Effect of parameter changes is displayed almost immediately, which greatly facilitates specifying
parameters. The inset shows the profile mass peak in blue and the detected mass centroid in red
34 Xiuxia Du et al.

detected from consecutive scans when scale levels are set at 5 and
15, respectively. Compared to the m/z values detected at scale level
equal to 5, most of the m/z values detected at scale level 15 are
larger. When the final representative m/z for a chromatographic
peak is calculated as the weighted average of all of the centroid m/z
values along the EIC as shown in Fig. 8b, the difference in the final
representative m/z values between using scale level¼5 and scale
level¼15 is 19 ppm. This difference in the mass values is big
enough to cause different compounds to be eventually identified.
Regardless of which of the mass detectors is used, the results of
mass detection for a particular profile mass spectrum can be
accessed by clicking masses under the profile mass spectrum
(Fig. 9). It is relevant to note that mass detection can also be carried
out by using msConvert that is part of ProteoWizard [8]. msConvert
detects masses by either using a CWT-based method or calling

(A) (B) (C) (D)

1e5 1e5
1.6 1.6 566.36
566.36
1.2 1.2 566.35
566.35
intensity

intensity

m/z

m/z
0.8 0.8
566.34
566.34
0.4 0.4
566.33 566.33
0.0 0.0
564 565 566 567 568 569 570 571 7.90 7.95 8.00 8.05 8.10 8.15 7.9 8.0 8.1 8.2 7.9 8.0 8.1 8.2
m/z rt rt rt

Fig. 8 Differences in the resulting mass values caused by different scale levels when using the wavelet
transform-based mass detection in MZmine 2. (a) One of the consecutive mass spectra from which the mass
indicated by red arrow is to be detected. (b) The EIC of the mass. The mass values of the blue dots along the
elution profile are depicted in (c) and (d) with scale level being 5 and 15, respectively

Fig. 9 Examine mass detection results for a particular profile mass spectrum.
Vertical lines in green indicate mass values that have been detected
 Metabolomics Data Preprocessing Using ADAP and MZmine 2 35

functions provided by vendors of mass spectrometers. The resulting

centroid data can be imported into MZmine 2 for data
preprocessing.

4.3 Construct EICs In untargeted metabolomics, the masses of ion species that have
by ADAP been detected by a mass analyzer are unknown prior to data pre-
processing. It is up to the step of EIC construction to determine.
With mass centroids detected from profile mass spectra, construc-
tion of EICs can begin. Figure 10 shows how to carry out this step
using ADAP. ADAP examines all of the data points in the entire
data file and works from the largest intensity data point down to the
smallest. As a result, a list of ions is produced that have been
detected by the mass analyzer over a continuous retention time
period. This approach in constructing EICs is in contrast to the
EIC construction process in other open-source software tools such
as XCMS where EICs are built chronologically in retention time.
The advantage of starting an EIC from the highest intensity point
among all of the data points belonging to this EIC is that the
reference mass for the EIC has the highest possible mass measure-
ment accuracy. This is particularly important for TOF-type mass
analyzers whose mass measurement accuracy tends to be higher for
more intense signals.

Construction of EICs by ADAP requires that the following

four parameters be specified:
(a) Min group size in number of scans. In the entire chromatogram
there must be at least this number of sequential scans having
points above the Group intensity threshold set by the user.
(b) Group intensity threshold. See above
(c) Min highest intensity. There must be at least one point in the
chromatogram that has an intensity greater than or equal to
this value.
(d) m/z tolerance. Maximum m/z difference of data points in
consecutive scans in order to be connected to the same
chromatogram.
As a result of the EIC construction, a list of EICs is produced
for each data file (Fig. 10c). Each EIC can be examined by double
clicking it and opening up a window as shown in Fig. 11.

4.4 Detect After EICs have been constructed, ADAP detects chromatographic
Chromatographic peaks from each of these EICs using the continuous wavelet trans-
Peaks by ADAP form (CWT) that is similar to what the wavelet transform mass
detector uses. Specifically, wavelet coefficients are first calculated as
the inner product between the EIC and the Ricker wavelets at
different wavelet scales and locations. Subsequently, peak location
36 Xiuxia Du et al.

Fig. 10 Construction of EICs in ADAP. (a) EIC construction is achieved by using the ADAP Chromatogram
builder method. (b) Parameters relevant to this method. (c) A list of EICs is produced for each data file
 Metabolomics Data Preprocessing Using ADAP and MZmine 2 37

Fig. 11 Examine EICs. (a) Select a particular EIC. (b) Double click the selected EIC opens this window. Select
Chromatogram and click Show opens the EIC in (c) for visual examination
38 Xiuxia Du et al.

and boundaries are determined through ridgeline detection and

simple local minima search. Finally, peak boundaries are adjusted
using a local minima search. This boundary adjustment is necessary
because the rough estimates for the left and right boundary based
on ridgeline detection are symmetric, i.e., having the same distance
from the peak location.
This ADAP peak detection method is accessed via Chromato-
gram deconvolution in the Peak list methods drop-down
menu (Fig. 12a). To choose the parameters appropriately, we
strongly recommend that users check the Show preview box.
The preview function allows a user to see the effect of parameter
changes immediately on peak detection for a chosen EIC. Any EIC
from any of the data files can be chosen using the Peak list and
Chromatogram drop-down menu (Fig. 12b). The following six
parameters need to be specified:
(a) SNR Threshold. Signal-to-noise threshold to filter out noise
peaks. For details about how SNR is calculated, we refer read-
ers to the publication by Myers et al. [9].
(b) Min feature height. The smallest intensity a peak can have
and be considered a real feature.
(c) Coefficient/area threshold. The best coefficient (larg-
est inner product of wavelet with peak in ridgeline) divided by
the area under the curve of the feature.
(d) Peak duration range. The acceptable range of peak widths.
Peaks with widths outside this range will be rejected.
(e) RT wavelet range. The range of wavelet scales used to build
matrix of coefficients. Scales are expressed as RT values (min-
utes) and correspond to the range of wavelet scales that will be
applied to the chromatogram. Choose a range that is similar to
the range of peak widths expected to be found from the data.

4.5 Alignment Alignment intends to identify corresponding peaks across samples.

MZmine 2 provides four alignment algorithms: Join aligner,
RANSAC aligner, Hierarchical aligner ( GC) , and ADAP
Aligner. The first two algorithms, Join aligner and RANSAC
aligner, are for aligning LC–MS data and the latter two, Hier-
archical aligner ( GC) and ADAP Aligner, are for aligning
GC–MS data. Both of the two algorithms for aligning LC–MS data
achieve alignment by finding chromatographic peaks that have
similar m/z and retention time. Figure 13 shows how to perform
alignment using the RANSAC aligner in MZmine 2. Aligned peaks
can be examined and exported (Fig. 14). The exported peak list can
be used for univariate and multivariate statistical analysis for deter-
mining the significant metabolites between phenotypes and train-
ing a predictive model for predicting phenotypes.
 Metabolomics Data Preprocessing Using ADAP and MZmine 2 39

Fig. 12 Detection of chromatographic peaks using ADAP. (a) Detection of

chromatographic peaks from EICs by ADAP can be accessed via Chromato-
gram deconvolution in the Peak list methods drop-down menu.
(b) Specification of parameters is facilitated by the Show preview function
40 Xiuxia Du et al.

Fig. 13 Alignment. (a) Alignment methods can be accessed through the Peak
list methods drop-down menu. (b) It is strongly recommended to use the
preview function for specifying parameters. (c) After alignment, an Aligned
peak list is produced and can be exported
 Metabolomics Data Preprocessing Using ADAP and MZmine 2 41

Fig. 14 Visualization and export of the aligned peak list. (a) Double clicking the
Aligned peak list opens up the list of peaks for visual examination. (b)
The aligned peak list can be exported in .csv, MetaboAnalyst, or other format

5 Preprocessing Workflow for GC–MS Data

As shown in Fig. 2, the preprocessing workflows for both LC–MS

and GC–MS data contain the steps of mass detection, EIC con-
struction, and detection of EIC peaks. The corresponding methods
and the procedures that have been described above for LC–MS data
preprocessing can be used for GC–MS data preprocessing as well.
However, the GC–MS workflow contains a step called spectral
deconvolution that is unique. This stems from the fact that the
commonly used electron ionization used in GC–MS analysis frag-
ments molecular ions into product ions in the ionization source.
When compounds are not resolved chromatographically, product
ions from different molecular ions co-exist in the same mass spec-
trum. In order to eventually identify/annotate the compounds that
correspond to each molecular ion, spectral deconvolution needs to
be performed to produce a pure mass spectrum of product ions and
the molecular ion for the compound. Spectral deconvolution is
especially necessary for low mass resolution GC–MS data that is
still commonly acquired.
42 Xiuxia Du et al.

In addition to the unique spectral deconvolution in GC–MS

preprocessing, the ADAP-GC preprocessing workflow features an
alignment algorithm that is compound-based, rather than peak-
based. Specifically, the ADAP-GC alignment algorithm looks for
similar compounds across samples based on spectral similarity and
proximity in retention time. This is very different from the RAN-
SAC alignment algorithm and other peak-based algorithms that
align chromatographic peaks only. If n-alkanes was added into the
samples and therefore retention index of compounds can be calcu-
lated, alignment of compounds should take advantage of the reten-
tion index information, but ADAP-GC is currently not equipped
with this capability yet.

5.1 Spectral The most recent version of the ADAP-GC spectral deconvolution
Deconvolution algorithm is 3.2 [5]. The algorithm starts with automated determi-
nation of deconvolution windows. For each deconvolution win-
dow, a sequence of four computational steps is carried out
including: (1) two rounds of hierarchical clustering for estimating
the number of compounds in the window, (2) selection of the
sharpest and unique chromatographic peaks as the model peaks,
(3) construction of pure mass spectrum for each compound, and
(4) correction of splitting issues. Figure 15 shows how to access
ADAP-GC 3.2 in MZmine 2 and lists the user-defined parameters.
Similar to ADAP peak detection described earlier, it is strongly
recommended that users use the Show preview function to make
informed decisions about the parameters (Fig. 15b). After spectral
deconvolution completes, a list of pure mass spectra is produced for
each data file (Fig. 16).

5.2 Alignment GC–MS samples are aligned by finding the same compounds across
the data files based on spectral similarity and retention time prox-
imity. Specifically, a score is calculated as follows to measure the
likelihood that two spectra, c1 and c2, correspond to the same
compound:
Scoreðc 1 , c 2 Þ ¼ wS time ðc 1 , c 2 Þþð1  wÞS spec ðc 1 , c 2 Þ, ð1Þ

where Stime is the retention time proximity between c1 and c2 and

Sspec is the spectrum similarity between c1 and c2. w is a weighting
factor specifying the relative importance of Stime and Sspec. Sspec is
calculated as the normalized dot product between c1 and c2. Fig-
ure 17 shows how to use the alignment method. The following
parameters need to be specified.
1. Min confidence: minimum fraction of samples where aligned
components must be present. It takes values between 0.0 and
1.0.
 Metabolomics Data Preprocessing Using ADAP and MZmine 2 43

Fig. 15 Spectral deconvolution. (a) ADAP-GC 3.2 spectral deconvolution can be

accessed via Peak list methods ! Spectral deconvolution
! Hierarchical Clustering. (b) User-defined parameters and the
preview function that helps with specifying these parameters. (c) After spectral
deconvolution, a list of pure mass spectra is produced for each data file
44 Xiuxia Du et al.

Fig. 16 Examine spectral deconvolution results. (a) Expand the list of mass spectra that has been produced for
each data file. (b) Each mass spectrum can be examined by double clicking it to open up a window. Select the
data file and Mass spectrum and click Show to open the window in (c). The constructed pure spectrum is
shown in green in the context of the raw spectrum shown in blue
 Metabolomics Data Preprocessing Using ADAP and MZmine 2 45

Fig. 17 GC alignment. (a) The Alignment method can be accessed in the drop-down menu of Peak
list methods ! Alignment ! ADAP Aligner. (b) Specify parameters whose meaning can be
found in the Help file. (c) The aligner produces the Aligned peak list
46 Xiuxia Du et al.

Fig. 18 Export GC–MS spectra. (a) Select the Aligned peak list and
specify the export file name. (b) An example of the exported spectra produced by
the spectral deconvolution. (c) Exported GC–MS spectra are stored in an .msp
file
 Metabolomics Data Preprocessing Using ADAP and MZmine 2 47

2. Retention time tolerance: maximum retention time difference

between aligned compounds in different samples.
3. m/z tolerance: maximum m/z difference to consider two m/z
values in two spectra as the same. This is used for determining
the quantitation mass for a particular compound. This mass is
defined as the most frequent mass across all of the spectra for
this compound.
4. Score threshold: minimum score as calculated in Eq. 1 to con-
sider c1 and c2 to correspond to the same compound. It takes
values between 0.0 and 1.0. The default value is 0.75.
5. Score weight: w in Eq. 1 and takes values between 0.0 and 1.0.
The default value is 0.1.
6. Retention time similarity: Stime in Eq. 1 as the difference in
retention time.

5.3 Export of GC–MS The pure mass spectra that the spectral deconvolution step has
Preprocessing Results constructed can be exported in .msp or .mgf format for matching
the spectra against spectral libraries for compound identification or
annotation. Figure 18 shows the procedure. The resulting .msp file
can be directly imported to the NIST MS Search software tool for
compound identification or annotation.

6 Conclusions

ADAP is a suite of computational algorithms and the associated

graphical user interface for preprocessing untargeted LC–MS and
GC–MS metabolomics data. Incorporation of these algorithms into
the prevalent MZmine 2 takes advantage of the rich visualization
capabilities in MZmine and benefits users of MZmine 2.

Acknowledgements

We thank the USA National Science Foundation award 1262416

and National Institutes of Health/National Cancer Institute grant
U01CA235507 for funding the research and development
of ADAP.

References
1. Pluskal T, Castillo S, Villar-Briones A, Oresic M TOF-MS metabonomics studies. J Proteome
(2010) MZmine 2: modular framework for pro- Res 9(11):5974–5981
cessing, visualizing, and analyzing mass 3. Ni Y, Qiu Y, Jiang W, Suttlemyre K, Su M,
spectrometry-based molecular profile data. Zhang W, Jia W, Du X (2012) ADAP-GC 2.0:
BMC Bioinf 11:395 deconvolution of coeluting metabolites from
2. Jiang W, Qiu Y, Ni Y, Su M, Jia W, Du X (2010) GC/TOF-MS data for metabolomics studies.
An automated data analysis pipeline for GC- Anal Chem 84(15):6619–6629
48 Xiuxia Du et al.

4. Ni Y, Su M, Qiu Y, Jia W, Du X (2016) ADAP- 8. Chambers MC, Maclean B, Burke R, Amodei D,

GC 3.0: improved peak detection and deconvo- Ruderman DL, Neumann S, Gatto L, Fischer B,
lution of co-eluting metabolites from GC/TOF- Pratt B, Egertson J, Hoff K, Kessner D, Tasman
MS data for metabolomics studies. Anal Chem N, Shulman N, Frewen B, Baker TA, Brusniak
88(17):8802–8811 MY, Paulse C, Creasy D, Flashner L, Kani K,
5. Smirnov A, Jia W, Walker DI, Jones DP, Du X Moulding C, Seymour SL, Nuwaysir LM,
(2018) ADAP-GC 3.2: graphical software tool Lefebvre B, Kuhlmann F, Roark J, Rainer P,
for efficient spectral deconvolution of gas Detlev S, Hemenway T, Huhmer A, Langridge
chromatography-high-resolution mass spec- J, Connolly B, Chadick T, Holly K, Eckels J,
trometry metabolomics data. J Proteome Res Deutsch EW, Moritz RL, Katz JE, Agus DB,
17(1):470–478 MacCoss M, Tabb DL, Mallick P (2012) A
6. Coble JB, Fraga CG (2014) Comparative evalu- cross-platform toolkit for mass spectrometry
ation of preprocessing freeware on chromatog- and proteomics. Nat Biotechnol 30(10):918–
raphy/mass spectrometry data for signature 920
discovery. J Chromatogr A 1358:155–164 9. Myers OD, Sumner SJ, Li S, Barnes S, Du X
7. Rafiei A, Sleno L (2015) Comparison of peak- (2017) One step forward for reducing false pos-
picking workflows for untargeted liquid itive and false negative compound identifications
chromatography/high-resolution mass spec- from mass spectrometry metabolomics data:
trometry metabolomics data analysis. Rapid new algorithms for constructing extracted ion
Commun Mass Spectrom 29(1):119–127 chromatograms and detecting chromatographic
peaks. Anal Chem 89(17):8696–8703
 Chapter 4

Metabolomics Data Processing Using OpenMS

Marc Rurik, Oliver Alka, Fabian Aicheler, and Oliver Kohlbacher

Abstract
This chapter describes the open-source tool suite OpenMS. OpenMS contains more than 180 tools which
can be combined to build complex and flexible data-processing workflows. The broad range of functionality
and the interoperability of these tools enable complex, complete, and reproducible data analysis workflows
in computational proteomics and metabolomics. We introduce the key concepts of OpenMS and illustrate
its capabilities with a complete workflow for the analysis of untargeted metabolomics data, including
metabolite quantification and identification.

Key words OpenMS, Data analysis, Workflows, Reproducible science, Metabolomics

1 Introduction

1.1 OpenMS Mass spectrometry workflows can be highly complex, due to many
different separation techniques, acquisition strategies, quantification,
and identification methods. As a solution, we present OpenMS, an
open-source C++ library for LC-MS data management that offers a
wide range of algorithms for proteomics and metabolomics data
analysis. It is available under the permissive 3-clause BSD license
for Windows, macOS, and Linux (www.openms.de) [1, 2]. OpenMS
is composed of a set of individual tools that can be combined in
a highly flexible manner allowing its application from standard opera-
tions to newly developed methods. The software is interesting
for both users and developers since the library can be used as a
foundation for tool and algorithm development (see Note 1). Fast
scripting and prototyping is supported via Python bindings (pyO-
penMS) [3]. Command line tools are available for high-throughput
processing, which is especially useful on cluster or cloud infrastruc-
tures. Additionally, OpenMS is integrated into the workflow engine
Konstanz Information Miner (KNIME), allowing for the visual
construction of data analysis workflows [4, 5] (see Note 2). In the
following, we will present the application of OpenMS and KNIME
for the quantification and identification of metabolomics data.

Shuzhao Li (ed.), Computational Methods and Data Analysis for Metabolomics, Methods in Molecular Biology, vol. 2104,
https://doi.org/10.1007/978-1-0716-0239-3_4, © Springer Science+Business Media, LLC, part of Springer Nature 2020

49
50 Marc Rurik et al.

1.2 Data Preparation A multitude of vendor-specific, proprietary file formats exist for
LC-MS data. In order to process this data with OpenMS it needs
to be converted to the open HUPO-PSI standard format mzML
[6]. Data conversion from the vendor-specific formats to mzML
can be performed using the free software MSConvertGUI included
in ProteoWizard [7]. Additionally, if the data was acquired in
profile mode, centroiding is necessary as almost all OpenMS algo-
rithms require data to be in centroid mode. This step can be
performed during the data conversion by using the peakPicking
filter provided by ProteoWizard. Alternatively, the OpenMS tool
PeakPickerHiRes can be used as the first step of the analysis
pipeline.

1.3 Data Data visualization is a common starting point for many studies. It
Visualization Using allows additional insight into the data, can act as a quality control
TOPPView step and help with the optimization of the tools’ parameter values.
In addition, visualization can be used to inspect intermediate or
final results throughout the data analysis process. To this end,
OpenMS provides TOPPView [8], a graphical application for the
interactive visualization of raw data and analysis results. It supports
the visualization of individual spectra and entire LC-MS maps in
2D and 3D representations. Results of different OpenMS tools
such as the location and extent of detected features (see Subheading
2.1) can be inspected in conjunction with the raw data. An example
for this is shown in Fig. 1.

Fig. 1 LC-MS sample visualized in TOPPView in 2D and 3D with color-coded intensities. Two features that
were detected with FeatureFinderMetabo (see Subheading 2.1) are shown as rectangles representing their
location and extent
 Metabolomics Data Processing Using OpenMS 51

During the development of a new analysis pipeline it is often

useful to assess and tune various parameters to best fit the experi-
mental and instrument setup. To assist with this, many OpenMS
tools can be executed directly from within TOPPView allowing for
the immediate visualization of their results with different parameter
choices.

1.4 Workflow KNIME (https://www.knime.com) is an open-source graphical

Construction Using workflow engine that allows for the user-friendly construction of
KNIME complex data analysis workflows. A workflow is a series of compu-
tational steps that describe exactly how the input data is processed.
Due to its comprehensive set of data analysis tools and its openness
to extensions, KNIME is well-suited to construct complete compu-
tational workflows. These workflows describe all tools that make up
the data analysis and, most importantly, also all parameters
required. Such workflows are an essential step towards reproducible
analysis of metabolomics data. Finalized workflows can easily be
shared with others, e.g., in the supplemental material of a publica-
tion, allowing them to reproduce the results.
In KNIME, nodes represent the individual data-processing
steps of the workflow. A node uses ports to read input data and
write output data. The data is transported along the edges that
connect the input and output ports of a series of nodes. An example
is shown in Fig. 2.
Another advantage of KNIME is that its community of users
provide a large collection of plugins which can be integrated into the
workflow. This makes it possible to add all commonly performed
downstream analysis steps including statistical hypothesis testing,
data visualization, machine learning, and cheminformatics methods
to the same workflow. In addition, scripting nodes for R, Python
and other languages have proven to be useful. Example workflows
using OpenMS in conjunction with these plugins can be down-
loaded from https://www.openms.de/workflows (see Note 3).

Fig. 2 A short KNIME workflow that detects and quantifies metabolites in a single LC-MS sample and exports
the results to an Excel file. Note that some nodes process entire files (square ports) while other nodes process
Excel-like tables (triangle ports)
52 Marc Rurik et al.

2 Untargeted Quantification of Metabolites

In this section, we will demonstrate how to construct an automated

workflow for the analysis of untargeted LC-MS metabolomics data.
The first step is the quantification of putative metabolites in each
input file. Afterwards, metabolites that are observed with different
adducts are grouped together. Finally, we correct for potential
retention time shifts and link the information across all input files.
This results in a list of, at this point, unidentified metabolites that
are characterized by their m/z, charge, retention time, and inten-
sity across all input files. In Subheading 3, different tools for
compound identification are added to the workflow.

2.1 Feature FeatureFinderMetabo [9] is a robust algorithm to detect and quan-

Detection tify metabolite features in an LC-MS sample. The term feature
refers here to all signals caused by an analyte in a specific charge
state and adduct situation. The signals corresponding to a feature
do contain the full chromatographic profile in the retention time
dimension and all isotopic traces in the m/z dimension. However,
due to different adduct species, multiple features can belong to the
same metabolite (this is addressed in Subheading 2.2).
Feature detection can be divided into four steps: (a) mass trace
detection, (b) peak separation, (c) hypothesis generation, and
(d) feature assembly.
Mass trace detection starts at the peaks with the highest inten-
sity which serve as potential seeds. It extends them in both direc-
tions along the retention time dimension. The mass error is
expected to be heteroscedastic, which means that it will be larger
at lower intensities. To account for this, the estimations that decide
whether additional peaks are added to a mass trace are iteratively
refined to make sure that the low intensity ends are properly
detected (Fig. 3a).
Metabolites that have a sufficiently similar m/z and elute in
close proximity can be located on a single uninterrupted mass trace.
FeatureFinderMetabo detects these cases based on the elution
profile and splits the feature at the local minimum between both
elution profiles (Fig. 3b).
Next, mass traces that likely originate from the same metabolite
are assembled into a feature. To decide whether an additional mass
trace should be added to a feature, it has to satisfy three criteria:
(1) the mass trace co-elutes within a user-defined RT range, (2) the
mass trace has the correct m/z distance within a user-defined mass
accuracy and range of possible charge states, and (3) the observed
isotope pattern is likely to be caused by a metabolite, i.e., only
certain intensity ratios between the isotopic traces are plausible.
In proteomics, the averagine model is widely used to identify
plausible isotopic traces, but this model is not applicable here
since metabolites differ widely in their elemental composition. To
 Metabolomics Data Processing Using OpenMS 53

Fig. 3 FeatureFinderMetabo feature detection process: (a) Mass trace detection starts at the peaks of highest
intensity (marked by a blue square) and extends in both directions. (b) Split the mass trace at the local
minimum if it contains more than one compound. (c) Test whether co-eluting mass traces are isotopic traces
of the same compound. (d) Assemble highest scoring feature hypotheses. The final features are characterized
by centroid m/z, retention time, charge, and intensity. A visualization of a feature in TOPPView is shown in
Fig. 1

address this, distinct sum formulas between 1 and 1000 Da were

used to train a support vector machine (SVM) to distinguish
between likely and unlikely isotope abundance ratios (Fig. 3c).
All possible feature hypotheses are scored based on their agree-
ment with the model and the best-scoring hypotheses are assem-
bled into features. The resulting features are characterized by their
centroid retention time, m/z, charge, and intensity, where the
feature intensity corresponds to the area under the monoisotopic
trace (Fig. 3d).

2.2 Adduct Grouping The previous feature finding step aims to detect and group all
isotopic traces of a compound ion into one feature represented by
centroid m/z, retention time, and intensity. However, a compound
might be detected multiple times in the same sample, with different
adduct species or common charge neutral modifications.
54 Marc Rurik et al.

Successfully grouping different ion species of a compound

allows to leverage knowledge across ions and infer information
missing from some of the features. Besides determining the neutral
mass, ion adducts and charges can be annotated to individual
features, facilitating computations in subsequent sample alignment
and linking steps.
For metabolomics, the MetaboliteAdductDecharger provides
this functionality. Inspired by similar algorithms used in peptide
decharging [10], this tool aggregates features inside an elution
window mainly by mass shifts and feature charges that are consis-
tent with assumed adduct hypotheses.
Given user-defined adduct probabilities and charge ranges, a
combinatorial table of pairwise adduct complexes and resulting
mass shifts is considered to create connected graphs of co-eluting
features. For each such connected group, the feature adducts
corresponding to the observed mass shifts are then assigned via an
integer linear programming (ILP) approach. That is, we seek to
resolve all conflicts arising from multiple explanations for observed
mass shifts. The optimal solution is chosen by maximizing overall
ion probabilities.

2.3 Map Alignment After feature detection, an optional but oftentimes appropriate
processing step is to perform chromatographic alignment of the
samples with each other. Here, the goal is to compensate for
commonly occurring constant or linear chromatography-related
distortions between samples.
This can be done with the MapAlignerPoseClustering tool,
which reduces feature elution time differences between sample
maps. To achieve this, the method aims to find affine retention
time transformations between samples that minimize the overall
time shifts of corresponding features. As a consequence,
subsequent aggregation of respective compound ion features
found in multiple measurements is more straightforward. To search
the solution space of possible time shifts and scaling distortions
(i.e., compression or extension of a samples chromatography) effi-
ciently, pose clustering is performed [11]. The method aligns the
features of multiple maps by using one as the reference to which the
other maps are aligned, akin to matching star charts in astronomy.
It determines the optimal linear transformation for each pair of
maps and transforms the feature retention times accordingly (see
Fig. 4a, b).

2.4 Feature Linking The FeatureLinkerUnlabeledQT algorithm is one of the OpenMS

methods to group features belonging to the same compound ion
across multiple samples. Using a QT clustering approach, features
are matched based on m/z and retention time [12]. By considering
all possible pairs inside user-defined m/z and retention difference
 Metabolomics Data Processing Using OpenMS 55

Fig. 4 Map alignment and feature linking are used to combine information across
multiple samples. (a) Features have been quantified in all samples individually
using FeatureFinderMetabo. (b) MapAlignerPoseClustering applies a linear
transformation to address potential retention time shifts. (c) The retention
time-aligned features can then be grouped using FeatureLinkerUnlabeledQT
(see Subheading 2.4)

tolerances, the tool can resolve variations in the local chromatogra-

phy not addressed by the previous affine retention time transforma-
tions, like differing elution orders.
In case a previous alignment step was performed, retention
time tolerances of the search window used for linking can be
tightened. As a consequence, the number of pairs of features across
samples having to be considered for correspondence is reduced.
Besides computational advantages, this also reduces the likelihood
of spurious matches.
Resulting linked features are collected as so-called consensus
features. Each consensus feature is represented by a centroid which
integrates the m/z, retention time, and charge of its member
features. A consensus feature further allows the comparison of ion
intensities across maps (see Fig. 5).
56 Marc Rurik et al.

Fig. 5 FeatureLinkerUnlabeledQT takes several feature maps and stores the

corresponding features in a consensus map. Here, feature A and feature B are
linked across three different maps

3 Basic Metabolite Identification

The previous section discussed how to detect and quantify metab-

olite features characterized by retention time, m/z, charge, and
intensity. Identifying the underlying compounds remains one of
the most challenging problems in metabolomics. This section
introduces three computational approaches implemented in
OpenMS to arrive at putative metabolite identifications.

3.1 Accurate Mass The foundation of the AccurateMassSearch (AMS) tool is a data-
Search base of known metabolites and their exact neutral masses. OpenMS
provides access to the Human Metabolome Database (HMDB)
[13], but generally any compound database can be used. In addi-
tion to such a database, a list of potential adducts has to be consid-
ered to arrive at the neutral masses of the compounds. OpenMS
provides two adduct lists for positive and negative polarity, which
can be edited depending on the experimental setup, e.g., which
buffers were used. Ideally this step should use the same set of
adducts as during adduct grouping.

3.2 Spectral Library While accurate mass search can be used to obtain a general idea
Search which compounds may be present in the sample, it will often
provide multiple putative identifications. To arrive at more confi-
dent identifications, it is necessary to fragment the compounds and
 Metabolomics Data Processing Using OpenMS 57

Fig. 6 KNIME workflow to identify metabolites by searching all tandem mass spectra against a spectral library,
in this case MassBank

compare them with reference spectra. Spectral libraries are a fast

and reliable way to refine compound identifications, but are limited
by the number of available compounds in the database.
OpenMS provides the tool MetaboliteSpectralMatcher to
search all tandem mass spectra against a spectral library (see
Fig. 6). The candidate spectra are scored against the database
spectra using a modified version of the Hyperscore described by
Fenyö and Beavis [14]
!
X
n
HSðS query , S database Þ¼ log I i  I 0i þ log n! ð1Þ
i¼0

I 0i
where Ii and are the intensities of all n matched peaks. Metabo-
liteSpectralMatcher can use any public or in-house spectral library
in the mzML format. By default, OpenMS provides MassBank, the
most comprehensive publicly available spectral library [15].

3.3 De Novo For de novo identification, SIRIUS [16, 17] and CSI:FingerID
Identification [18] are integrated in the OpenMS tool SiriusAdapter (see Fig. 7).
A preprocessing step is performed using the OpenMS framework to
optimize and convert the input data to a SIRIUS compatible for-
mat. SiriusAdapter can then perform identification of features at
MS2 level, i.e., molecular formulas are assigned de novo. Subse-
quently, the feature can be searched in a molecular structure data-
base. SiriusAdapter uses information from both MS and MS/MS
spectra (mzML) and an optional featureXML. The tool provides
different modes, depending on the input and output configuration.
Preprocessing can be applied if additional feature information is
present, which is used to map all MS2 spectra to their
corresponding features. SIRIUS will internally merge and jointly
process all MS2 spectra allocated to the same feature. To reduce the
feature space further, a mass trace filter (number of isotopic traces)
can be applied. Additionally, adduct information can be provided
using a featureXML processed by the MetaboliteAdductDecharger
or AccurateMassSearch. Depending on the workflow, SiriusAdapter
can be used either for preprocessing only (output port SIRIUS.ms)
or full data processing (output port SIRIUS.mzTab), see Fig. 7.
58 Marc Rurik et al.

Fig. 7 Workflow for metabolite quantification using FeatureFinderMetabo and de novo identification using
SiriusAdapter

4 Quantification and Identification Workflow

An example OpenMS workflow for the analysis of an untargeted

metabolomics study consisting of any number of samples is shown
in Fig. 8. Note that metabolite detection and adduct grouping are
performed separately for each file. For this reason, these nodes are
enclosed by the nodes ZipLoopStart and ZipLoopEnd, which will
automatically iterate over all input files, process them and collect
the results. The alignment and aggregation of all detected features
across all samples are then performed by MapAlignerPoseCluster-
ing and FeatureLinkerUnlabeledQT as described in Subheadings
2.3 and 2.4. In this workflow, metabolite identification is per-
formed by querying HMDB using AccurateMassSearch. Putative
identifications can be refined with spectral library search or de novo
identifications (see Subheadings 3.2 and 3.3). Finally, MzTabRea-
der converts the data to a table-based file format that can be read
using Excel and similar software.

5 Notes

OpenMS is in constant development to improve its algorithm

library, tools, and usability. For further information about metabo-
lomics data analysis with OpenMS and KNIME, we recommend
the following sources:
 Metabolomics Data Processing Using OpenMS 59

Fig. 8 Basic OpenMS workflow for metabolite quantification and identification in KNIME

1. The full OpenMS source code is available in our GitHub

repository https://github.com/OpenMS/OpenMS. We are
always open to contributions, feel free to file issues on GitHub
or use the chat system Gitter at https://gitter.im/OpenMS/
OpenMS to get in contact with the developers.
2. Installers for OpenMS and the workflow engine KNIME are
available at
l OpenMS: https://www.openms.de/download/openms-
binaries
l KNIME: https://www.knime.org/downloads/overview
We provide detailed instructions on how to install the
OpenMS plugin in KNIME at https://www.openms.de/get
ting-started/creating-workflows.
3. Visit https://www.openms.de for general information and
news about OpenMS. A detailed step-by-step guide can be
obtained from https://www.openms.de/tutorials. In addition,
we provide ready-to-use example workflows for common tasks
at https://www.openms.de/workflows. Please note the impor-
tance to adapt the parameters of the algorithms to your experi-
mental and instrument setup. You can use the TOPPView to
assess the impact of different parameter choices.
60 Marc Rurik et al.

References
1. Röst HL, Sachsenberg T, Aiche S, Bielow C, charge ladders generated in ESI-MS. J Prote-
Weisser H, Aicheler F, Andreotti S, Ehrlich ome Res 9(5):2688–2695
H-c, Gutenbrunner P, Kenar E, Liang X, 11. Lange E, Gropl C, Schulz-Trieglaff O,
Nahnsen S, Nilse L, Pfeuffer J, Leinenbach A, Huber C, Reinert K (2007) A
Rosenberger G, Rurik M, Schmitt U, Veit J, geometric approach for the alignment of liquid
Walzer M, Wojnar D, Wolski WE, Schilling O, chromatography mass spectrometry data. Bio-
Choudhary JS, Malmström L, Aebersold R, informatics 23(13):i273–i281
Reinert K, Kohlbacher O (2016) OpenMS: a 12. Weisser H, Nahnsen S, Grossmann J, Nilse L,
flexible open-source software platform for mass Quandt A, Brauer H, Sturm M, Kenar E,
spectrometry data analysis. Nat Methods Kohlbacher O, Aebersold R, Malmström L
13:741–748 (2013) An automated pipeline for high-
2. Pfeuffer J, Sachsenberg T, Alka O, Walzer M, throughput label-free quantitative proteomics.
Fillbrunn A, Nilse L, Schilling O, Reinert K, J Proteome Res 12:1628–1644
Kohlbacher O (2017) OpenMS - a platform for 13. Wishart DS, Feunang YD, Marcu A, Guo AC,
reproducible analysis of mass spectrometry Liang K, Vázquez-Fresno R, Sajed T,
data. J Biotechnol 261(February):142–148 Johnson D, Li C, Karu N, Sayeeda Z, Lo E,
3. Röst HL, Schmitt U, Aebersold R, Malmström Assempour N, Berjanskii M, Singhal S,
L (2014) pyOpenMS: a Python-based interface Arndt D, Liang Y, Badran H, Grant J, Serra-
to the OpenMS mass-spectrometry algorithm Cayuela A, Liu Y, Mandal R, Neveu V, Pon A,
library. Proteomics 14(1):74–77 Knox C, Wilson M, Manach C, Scalbert A
4. Berthold MR, Cebron N, Dill F, Gabriel TR, (2018) HMDB 4.0: the human metabolome
Kötter T, Meinl T, Ohl P, Sieb C, Thiel K, database for 2018. Nucleic Acids Res 46(D1):
Wiswedel B (2007) KNIME: the Konstanz D608–D617
Information Miner. In: Studies in classification, 14. Fenyö D, Beavis RC (2003) A method for
data analysis, and knowledge organization assessing the statistical significance of mass
(GfKL 2007). Springer, Berlin spectrometry-based protein identifications
5. Fillbrunn A, Dietz C, Pfeuffer J, Rahn R, using general scoring schemes. Anal Chem 75
Landrum GA, Berthold MR (2017) KNIME (4):768–774
for reproducible cross-domain analysis of life 15. Horai H, Arita M, Kanaya S, Nihei Y, Ikeda T,
science data. J Biotechnol 261 Suwa K, Ojima Y, Tanaka K, Tanaka S,
(February):149–156 Aoshima K, Oda Y, Kakazu Y, Kusano M,
6. Martens L, Chambers M, Sturm M, Kessner D, Tohge T, Matsuda F, Sawada Y, Hirai MY,
Levander F, Shofstahl J, Tang WH, Römpp A, Nakanishi H, Ikeda K, Akimoto N, Maoka T,
Neumann S, Pizarro AD et al (2011) mzML— Takahashi H, Ara T, Sakurai N, Suzuki H,
a community standard for mass spectrometry Shibata D, Neumann S, Iida T, Tanaka K,
data. Mol Cell Proteomics 10(1): Funatsu K, Matsuura F, Soga T, Taguchi R,
R110.000133 Saito K, Nishioka T (2010) MassBank: a public
7. Kessner D, Chambers M, Burke R, Agus D, repository for sharing mass spectral data for life
Mallick P (2008) ProteoWizard: open source sciences. J Mass Spectrom 45(7):703–714
software for rapid proteomics tools develop- 16. Böcker S, Letzel MC, Lipták Z, Pervukhin A
ment. Bioinformatics 24(21):2534–2536 (2009) SIRIUS: decomposing isotope patterns
8. Sturm M, Kohlbacher O (2009) TOPPView: for metabolite identification. Bioinformatics 25
an open-source viewer for mass spectrometry (2):218–224
data. J Proteome Res 8(7):3760–3763 17. Böcker S, Dührkop K (2016) Fragmentation
9. Kenar E, Franken H, Forcisi S, Wörmann K, trees reloaded. J Cheminform 8(1):1–26
H€aring H-U, Lehmann R, Schmitt-Kopplin P, 18. Dührkop K, Shen H, Meusel M, Rousu J,
Zell A, Kohlbacher O (2014) Automated label- Böcker S (2015) Searching molecular structure
free quantification of metabolites from liquid databases with tandem mass spectra using CSI:
chromatography–mass spectrometry data. Mol FingerID. Proc Natl Acad Sci
Cell Proteomics 13:348–359 112:12580–12585
10. Bielow C, Ruzek S, Huber CG, Reinert K
(2010) Optimal decharging and clustering of
 Chapter 5

Analysis of NMR Metabolomics Data

Wimal Pathmasiri, Kristine Kay, Susan McRitchie, and Susan Sumner

Abstract
In this chapter, we summarize data preprocessing and data analysis strategies used for analysis of NMR data
for metabolomics studies. Metabolomics consists of the analysis of the low molecular weight compounds in
cells, tissues, or biological fluids, and has been used to reveal biomarkers for early disease detection and
diagnosis, to monitor interventions, and to provide information on pathway perturbations to inform
mechanisms and identifying targets. Metabolic profiling (also termed metabotyping) involves the analysis
of hundreds to thousands of molecules using mainly state-of-the-art mass spectrometry (MS) and nuclear
magnetic resonance (NMR) spectroscopy technologies. While NMR is less sensitive than mass spectrome-
try, NMR does provide a wealth of complex and information rich metabolite data. NMR data together with
the use of conventional statistics, modeling methods, and bioinformatics tools reveals biomarker and
mechanistic information. A typical NMR spectrum, with up to 64k data points, of a complex biological
fluid or an extract of cells and tissues consists of thousands of sharp signals that are mainly derived from
small molecules. In addition, a number of advanced NMR spectroscopic methods are available for extract-
ing information on high molecular weight compounds such as lipids or lipoproteins. There are numerous
data preprocessing, data reduction, and analysis methods developed and evolving in the field of NMR
metabolomics. Our goal is to provide an extensive summary of NMR data preprocessing and analysis
strategies by providing examples and open source and commercially available analysis software and bioin-
formatics tools.

Key words NMR, Metabolomics, Metabotyping, Quality control, Preprocessing, Multivariate data
analysis

1 Introduction

1.1 Definitions Nuclear magnetic resonance (NMR) spectroscopy is a quantitative,

robust, highly reproducible, and nondestructive analytical tech-
nique, and it is widely used in metabolomics analyses. NMR spec-
troscopy relies on the magnetic properties of the nucleus of an atom
[1]. When placed in a strong magnetic field, these nuclei resonate at
characteristic frequencies in the radio frequency range (e.g.,
500 MHz for 1H at 11.7 T magnetic field, Table 1). NMR spec-
troscopy has long been used as an analytical tool in studying the
chemical structure, conformation, and molecular dynamics of both
small molecules and macromolecules and more details about NMR

Shuzhao Li (ed.), Computational Methods and Data Analysis for Metabolomics, Methods in Molecular Biology, vol. 2104,
https://doi.org/10.1007/978-1-0716-0239-3_5, © Springer Science+Business Media, LLC, part of Springer Nature 2020

61
62 Wimal Pathmasiri et al.

Table 1
NMR frequencies and natural abundance of some useful nuclei in an 11.7 T magnetic field strength
(values adopted from Bruker NMR frequency table)

Nucleus Basic frequency (MHz) Natural abundance (%)

1
H 500 100
13
C 125.7 1.07
15
N 50 0.36
31
P 202.5 100

spectroscopy are described in detail in textbooks [1] and literature.

For easy understanding, some important features in the NMR
spectroscopy and NMR metabolomics analysis are defined below.
NMR chemical shift and J-coupling. Variations in the NMR fre-
quency of an atom occur from neighboring atoms or attached
functional groups within the molecular structure (chemical
microenvironment), and these provide detailed information
about the structure and conformation of the molecule. The
variation in resonance is called the chemical shift and is typically
reported in parts per million (ppm) to enable comparison of
results between spectrometers of different frequency. Parts per
million is the frequency of the resonance in Hertz divided by
the spectrometry frequency. Splitting patterns that arise in the
NMR spectrum (also referred to as J-coupling) are used
together with chemical shift for identification of chemical
structure. A variety of NMR pulse sequences are used to reveal
the chemical shift and J-coupling information and extract
structural information through detection of various nuclei
(1H, 13C, 15N, 31P).
Metabolome, metabolites, and metabotype. The metabolome consists
of low molecular weight chemical compounds called metabo-
lites that are present in biological specimens. These metabolites
can either be produced as a result of endogenous cellular
metabolism (endogenous metabolites) or as a result of expo-
sure (environmental chemicals, food ingestion, drug intake), or
derived from microbial metabolism.
Stable isotope resolved metabolomics. Some atoms can exist in multi-
ple forms called isotopes. These isotopes are used for investi-
gating (tracing) the fate of endogenous metabolites between
metabolic pathways. Carbon 13 (13C) is an example of one
stable isotope used in NMR metabolomics studies. Naturally
occurring compounds contain only 1.1% 13C, and thus, high
concentrations of these compounds or long NMR run times are
needed for signal detection. Using compounds that are
 Analysis of NMR Metabolomics Data 63

enriched in 13C enables the performance of experiments that

trace the 13C, to reveal the metabolism of the enriched mole-
cules which have signal intensities above those of the endoge-
nous background.

1.2 Applications Cells, tissues, and biological fluids are rich in low molecular weight
of Metabolomics metabolites, and metabolomics involves the analysis of these low
Analysis Methods molecular weight metabolites [2–5]. Common biological fluids
used in metabolomics include serum, plasma, saliva, urine, cerebro-
spinal fluid, feces, and exhaled breath. Unlike other omics (geno-
mics, transcriptomics, and proteomics) technologies,
metabolomics gives the most functional information about the
system since metabolites are the final end products of genomic,
transcriptomic, and/or proteomic perturbations [6] within a com-
plex network of biological pathways. Metabolomics studies are
designed to reveal the pattern of signals that increase or decrease
as a function of health and wellness, or response to treatment.
Through determining signals that associate with health and disease
states, metabolites can be identified and mapped to biochemical
pathways to provide insights into biological mechanisms.
In the past decade, advances in technology have enabled the
application of metabolomics in a variety of diverse research areas
that cover basic, biomedical, and clinical sciences to measure these
metabolites in readily accessible biological fluids, cells, and tissues
to correlate back with the responses at the cellular or organ level.
The terms metabolomics, metabonomics, metabolic profiling, met-
abolic phenotyping, and metabotyping are widely used inter-
changeably by the research community to basically the same
technique and procedures [7, 8]. The technologies that can be
used to detect and measure these metabolites are very diverse.
NMR spectroscopy and mass spectrometry (often coupled with
chromatography) are the two main analytical platforms in metabo-
typing whereas other analytical techniques such as liquid chroma-
tography coupled electrochemical detection (for detecting neuro
transmitters [9] or capillary electrophoresis methods [10]) can be
used in the targeted analysis. Advances in high-throughput, high-
resolution analytical technologies and data analysis platforms have
allowed for metabotyping in large-scale population-based Metabo-
lome Wide Association Studies (MWAS) and metabotype quantita-
tive trait locus (mQTL) mapping studies by using both NMR
Spectroscopy [11–13] and mass spectrometry [14–16]. These
methods are more useful than Genome Wide Association Studies
(GWAS) and QTL mapping for discovering biomarkers for disease
risk because metabolic perturbations are the most downstream
products of all of the other omics activities. Furthermore, integrat-
ing of MWAS and mQTL with GWAS and QTL mapping studies
improves biomarker identification.
64 Wimal Pathmasiri et al.

A number of metabolome/phenome centers [17–21] have

been established to provide services to the growing demand for
metabotyping in large scale population-based studies. A metabo-
typing study can be conducted in the form of an untargeted or a
targeted analysis method. Untargeted methods are typically used
for discovery and hypothesis generation. In the untargeted work-
flow, signals that correspond to the metabolome (endogenous and
exogenous molecules) are detected, and the data is subjected to
subsequent statistical analysis, modeling, and bioinformatics ana-
lyses. Signals that are responsible for the differentiation of study
phenotypes (e.g., difference in case vs control, difference in
high vs. low dose) are then identified. On the other hand, targeted
methods are more often used to test hypotheses. In this workflow, a
set of analytes are selected to test a hypothesis, for example, about
the disease state or response to treatment. The targeted metabolo-
mics analysis involves the use of metabolite standards (including
isotopically labeled compounds) and calibration curves. Statistical
analysis and other modeling approaches are performed on the
analytes once these metabolites are quantified. The findings from
any untargeted study must be confirmed or validated by using an
appropriate targeted metabolomics analysis method.

2 NMR Metabolomics

2.1 NMR In a typical 1H NMR experiment, a radiofrequency pulse in a kHz

Spectroscopy range is applied using a transmitter coil in the NMR probe to the
sample (in a tube) placed in the external magnetic field. Then the
individual nuclei are in Boltzmann equilibrium aligned (low-energy
state) or antialigned (high-energy state) with respect to the applied
magnetic field. Once the RF pulse is applied, the Boltzmann equi-
librium is perturbed and the nuclei in the low energy state become
excited into the high-energy state. The nuclei are then relaxed back
to the original equilibrium status by a mechanism called relaxation
(nuclei–nuclei and nuclei–matrix) and a decaying signal is recorded
in the receiver in the form of a free induction decay (FID), which is
a time domain spectrum. The recorded NMR signal is a collection
of decaying sine waves (sinusoids) and this analog electronic signal
is digitized to create a free induction decay (FID). The FID is
Fourier transformed to convert the time domain spectrum into a
frequency domain spectrum (Fig. 1). A window function is also
applied (apodization) during the Fourier Transform process to
improve the NMR spectrum. It has been an active area of NMR
research to study the chemical shift and J-coupling patterns of
molecules with varying functional moieties and molecular confor-
mation and there is a vast number of literature data (including
NMR tables from instrument vendors) available that assists in
structure elucidation purposes. A number of databases [22–26]
 Analysis of NMR Metabolomics Data 65

Fig. 1 NMR signal is measured as a FID and then Fourier transformed to obtain the frequency domain
spectrum (example shown here is for a 1H NMR spectrum of a cholesterol derivative)

are available for searching NMR spectra of molecules (HMDB,

BMRB, Birmingham Metabolite Database, NMRShift DB2) and
many software tools have been developed to predict NMR spectra
based on NMR properties (ChemDraw, ACD, MNova). A variety
of instrumental methods called pulse programs have been devel-
oped (and are still being developed) to collect one-dimensional
(1D), two-dimensional (2D), and multidimensional NMR spectra,
and these experiments assist in elucidating structure of the mole-
cules. NMR spectroscopy is a powerful tool for elucidating struc-
tural details of both small molecules (e.g., identification of natural
products and pharmaceutical compounds) and macromolecules
(e.g., proteins and nucleic acids), and their molecular complexes
and their dynamics. The resonance frequency is proportional to the
applied external magnetic field so that nuclei resonate at higher
radio frequencies on instruments with higher magnetic fields (e.g.,
500 MHz at 11.7 T and 950 MHz at 22.3 T for 1H). A higher
resolution is created by a larger dispersion of the peaks in the NMR
spectrum. Most NMR metabolomics studies have been conducted
with 600 and 700 MHz field strengths, while 950 MHz have also
been employed [27, 28].

2.2 NMR NMR spectroscopy and chromatography coupled mass spectrome-

Metabolomics try are the two main spectroscopic methods used in the metabolo-
Analysis mics investigations. Both platforms are highly suitable for
metabolomics analysis, are complimentary, and have their own
analytical strengths and weaknesses. This chapter focuses on
NMR spectroscopy which can detect a wide variety of chemical
classes (alcohols, amines, amino acids, aromatics, bile acids, carbox-
ylic acids (including fatty acids), ketones, lipids, nucleosides,
nucleotides, phenols, sulfur-containing compounds, and sugars)
including compounds which are difficult to ionize on mass spec-
trometry systems.
The metabolites in a complex biological specimen (urine,
serum, plasma, saliva, feces, etc.) can be measured simultaneously
and provide a metabolic fingerprint of the system. We recover
information on metabolic changes related to the metabotype by
applying standard statistical techniques, chemometrics,
66 Wimal Pathmasiri et al.

bioinformatics, and other data analysis and mining methods on

spectral data. In addition, stable isotope resolved metabolomics
[29] is used to extract metabolic flux information [30] in specific
metabolic pathways. Cellular metabolism is dynamic and is a com-
plex system of metabolic networks. Therefore, metabolic flux anal-
ysis [31] using stable isotopes facilitates determining the metabolic
reaction rates (fluxes) for better understanding of the cellular
mechanisms at systems biology level. In addition, stable isotope-
based tracer experiments can be used to analyze the fate of certain
metabolite(s) in metabolic pathways (e.g., metabolites of drug
molecules or an external environmental compound).
The high-resolution NMR spectroscopy (500 MHz and above)
technique is a quantitative, robust, highly reproducible, and non-
destructive technique [32] for analysis of biological fluids, as well as
extracts of tissues and cells. NMR typically does not require chro-
matographic separation of metabolites from the complex mixture
(biological fluid or tissue/cell extract), while mass spectrometry
methods typically require chromatographic liquid—or gas—sepa-
ration and is destructive to the sample.
Sample contamination and carryover is a complex issue in mass
spectrometry studies but is not problematic in NMR studies since
there is no physical contact between the sample and the spectrome-
ter. Furthermore, routine automatic data acquisition methods cou-
pled with automatic sample changers are available which
accommodate samples in 96 tube format in NMR tube racks.
Here samples are stored under refrigerated conditions and equili-
brated to the desired temperature before and after inserting the
tube into the magnet. This is especially helpful in acquiring data
sequentially for batches of samples in large-scale metabolomics
studies. Typically, it requires a few minutes to acquire a 1H NMR
spectrum, and about 15–20 min to acquire a routinely selected set
of NMR spectra (1D 1H and 2D-JRES) for a single sample.
In longitudinal larger-scale epidemiological or clinical metabo-
typing studies in particular, it is not possible to acquire data for all
samples simultaneously. Such analyses require multiday and multi-
batch sample preparation and data acquisitions. Therefore, the
longitudinal stability of NMR spectroscopy makes it ideal for
high-throughput data acquisition needs for such studies
[33]. The data acquired in multiple batches (with extensive quality
control procedures described in Subheading 3) can be combined
for analysis, including workflows for processing multicohort untar-
geted metabolomics [34].
For NMR metabolomics analysis, samples are typically prepared
and transferred to a 5 mm NMR tube, while lower diameter
(1.7–3 mm) tubes are available if limited in sample volume, and
NMR flow cells or microprobes are also available to accommodate
samples with very low volumes (e.g., 10 μl). There are techniques
such as hyperpolarized NMR methods [22, 35, 36] that can
 Analysis of NMR Metabolomics Data 67

increase the sensitivity to detect atoms of low natural abundance

(such as 13C and 15N) thus enabling the use of 13C and 15N NMR
(1D and 2D, for example) spectroscopy for detection of metabo-
lites in NMR metabolomics analysis in a short period of time.
Additionally, there are hyphenated NMR systems that are coupled
with liquid chromatography (LC-NMR) or solid phase extraction
technique (SPE-NMR) in cases where a reduced complexity of the
spectra is needed. Combined LC-NMR-MS or LC-SPE-NMR- MS
systems have also been developed, thus enabling recovery of exact
molecular weight information of metabolites in addition to the
NMR spectra for improved identification and quantification of
metabolites in the profiling of complex biological samples [22].

2.3 NMR An example workflow for broad spectrum (or untargeted) metabo-
Metabolomics lomics data using NMR is shown in Fig. 2. This workflow includes
Workflow study design, sample collection, sample selection, sample randomi-
zation, sample preparation, data acquisition, data preprocessing,
and data analysis.

2.3.1 Study Design Metabolomics investigations have been conducted that have
revealed that genetics (e.g., race, gender, polymorphisms), nutri-
tion (e.g., diet, nutrients, weight, gut microbiome), mental health
(e.g., stress, depression, cognition, behaviors), and exposures (e.g.,
tobacco use, pollution, drugs, medications, personal care products)

Fig. 2 Schematic diagram explaining the NMR metabolomics workflow. One advantage of NMR metabolomics
is that both untargeted and targeted data analysis can be performed using the same NMR spectrum. In this
workflow, metabolites are identified after using statistics and modeling approaches to reveal signals that are
important to defining the study phenotype. Quality Controls (underlined) are included at each step of the
pipeline
68 Wimal Pathmasiri et al.

all influence the metabotype [37]. Therefore, it is critical to take

these factors into consideration when designing the sample collec-
tion for the study to ensure that the appropriate metadata is col-
lected for all subjects, and that the study is properly powered. In
addition, when selecting samples from a repository, these factors
should be considered in the power calculations in order to select
samples for subjects for each of the phenotypic groups, to match
the subjects between phenotypic groups for these factors, and to
determine other uncontrollable confounding factors for consider-
ation. Otherwise, subsequent statistical analysis methods will fail to
achieve the desired goals of the study.

2.3.2 Sample Selection The number of samples needed for the study will depend on the
and Randomization variables mentioned above, as well as the strength of the phenotype.
More samples are needed when the study involves weaker pheno-
types (e.g., control vs. mild anxiety), whereas smaller number of
samples are sufficient for situations with stronger phenotypes (e.g.,
healthy vs. cancer cases) to extract desirable information. For a
validation or a confirmation study, a sufficient number of samples
that can be determined by power calculations such as multivariate
power calculations [38] must be used to achieve statistically signifi-
cant results. Selection of biospecimens derived from a different
cohort is preferable for such validation studies. Samples must be
randomized for both the sample preparation and analysis sequence
in the data acquisition in order to avoid any analytical bias.

2.3.3 Biospecimen It is also necessary to consider other factors such as the type of
Availability anticoagulant (or preservative, if any), storage condition, and num-
ber of freeze–thaw cycles. Biospecimens that have undergone dif-
ferent collection procedure, a different number of freeze-thaw
cycles, or different storage temperatures (
20  C,
80  C) can
give rise to different profiles. Ideally, samples with no freeze–thaw
cycles and stored at
80  C are best suited for metabolomics
analysis. This can be achieved by preparing multiple aliquots at
the sample collection stage, prior to freezing. The most important
thing is to keep all conditions of the samples the same within a
particular metabolomics study. The type of anticoagulant used for
samples can affect the NMR spectrum. Serum is best suited for
NMR metabolomics, while heparinized plasma (preferable over
EDTA or citrated plasma) can also be used. EDTA (and its Ca2+,
and Mg2+ complexes) or citrate in the biospecimens results in
signals that hinder or obscure the signals of some endogenous
metabolite signals and therefore such peaks must be removed
from further analysis. Therefore, valuable information is lost if
EDTA or citrate is used as an anticoagulant in samples. However,
it was found that useful biochemical could still be effectively recov-
ered using samples collected with these anticoagulants [39]. Once a
 Analysis of NMR Metabolomics Data 69

biomarker metabolite or a set of metabolites is discovered using an

untargeted metabolomics analysis method, it is important to con-
firm or validate the findings by designing and applying an appro-
priate targeted analysis method by even using samples generated in
a different cohort. Subsequently, blinded studies can also be
designed for biomarker validation purposes.

2.3.4 Biological Sample Human subject metabolomics investigations using NMR have gen-
Types erally used urine, serum, or plasma, while saliva and stool samples
are becoming more commonly utilized. In addition to these avail-
able biological matrices, a wide variety of other sample types such as
cerebrospinal fluid, seminal fluid, bronchoalveolar lavage fluid,
amniotic fluid, seminal fluid, and extracts of organ tissue have
been used in NMR metabolomics. On the other hand, any available
biological matrices (e.g., urine, serum or plasma, feces, tissue,
organ, and other biological samples) can be used for studies involv-
ing animal models.

2.3.5 Sample Collection Standard established protocols are needed for the sample collection
and Storage and storage [40] of biological samples. Consistency in sample
collection and storage is an important consideration [41]. For
biomarker discovery and validation, metabolomics data analysis
using combined sample sets derived from longitudinal or multicen-
ter multicohort population studies is important for increasing the
statistical power and, also, is challenging due to inconsistencies in
the sample collection in the cohorts and storage conditions in the
biobanks. There are ongoing efforts to integrate metabolomics
data obtained from samples in multiple biobanks [42]. While it is
not always possible to obtain biospecimens from different cohorts
that have been collected and stored in identical manner, it is impor-
tant to have access to the protocols regarding the collection and
handling of the samples in order to understand differences that may
arise in the profiles between cohorts.

2.3.6 Sample NMR metabolomics is typically performed on samples (biofluids,

Preparation extracts of biofluids, cells, or tissue) in solution. There are a number
of standard protocols designed for metabotyping in both small and
large-scale studies and published in literature for preparing
biological samples for NMR spectroscopy [32, 43]. There are
other methods available for analyzing metabolites in solid samples
such as intact tissue biopsy samples using high-resolution magic
angle spinning NMR [44, 45] or using in vivo magnetic resonance
spectroscopy [46] to record metabolite profiles of organs or tissues
(in magnetic resonance imaging). Regardless of the method uti-
lized, standard established protocols for the sample preparation
should be used and documented, and method development work
should be undertaken for new biological matrices using representa-
tive samples before handling the real study samples.
70 Wimal Pathmasiri et al.

1. Minimal sample preparation method

Biological samples such as blood (serum and plasma),
urine, cerebrospinal fluid (CSF), and saliva can be prepared by
mixing an NMR buffer into each sample aliquot. This method
has the advantage of preserving the sample so that it can be
analyzed by other assays following the NMR metabolomics
analysis. NMR buffer typically contains a phosphate buffer
(pH ¼ 7.4), an internal standard (trimethyl-silyl-[2,2,3,3-d4]
propionic acid (TSP) or 4,4-dimethyl-4-silapentane-1-sulfonic
acid (DSS-d6), chemical shift indicator), sodium azide (NaN3,
to prevent bacterial growth), and D2O (for deuterium locking
of the spectrometer).
2. Extraction sample preparation method
Samples generated from tissues, feces, or blood (serum or
plasma) can be prepared by homogenization or extracting with
a solvent (organic or aqueous or mixtures), drying, and recon-
stitution in an NMR buffer [27, 47–49]. Ultrafiltration tech-
niques using molecular weight cut-off filters can also be
employed to remove macromolecules from biofluid samples
such as serum, plasma, cerebrospinal fluid, and breast milk
before mixing with NMR buffer solution [50, 51].

2.3.7 Data Acquisition The data acquisition methods selected depend on the study ques-
tion. Typically, NMR metabolomics uses 1D 1H NMR spectro-
scopic methods because of the high sensitivity of 1H (and
simplicity) in contrast to very low sensitivity of 13C or 15N because
of the very low natural abundance. Serum and plasma samples
contain macromolecules such as proteins, lipids, and lipoproteins.
Atoms in these macromolecules relax faster and therefore the sig-
nals deriving from these molecules become broadened making it
harder to analyze small molecule metabolites.
A number of different NMR pulse sequences are commonly
used in 1H NMR metabotyping of biological samples to address
these issues to improve the spectral quality of the small molecules.
The four most commonly used sequences in routine metabolomics
data acquisition from blood samples are NOESY presaturation
(presat), Carr-Purcell-Meiboom-Gill (CPMG), J-Resolved, and
Diffusion-edited method [43].
1. NOESY presat: The NOESY presat pulse sequence is the first
increment of a NOESY pulse sequence [24] (RD-gz(1)-90 -t-
90 -tm-gz(2)-90 -ACQ, where RD is the relaxation delay, t is a
short delay between pulses, tm is the mixing time, 90 is the
duration of the radio frequency pulse, gz(1) and gz(2) are the
pulse field gradients, and ACQ is the data acquisition time).
2. Carr-Purcell-Meiboom-Gill (CPMG) spin-echo sequence: The
CPMG spin-echo sequence uses a CPMG presat pulse
 Analysis of NMR Metabolomics Data 71

sequence which is a relaxation editing method applied to atten-

uate signals derived from macromolecules (faster relaxing
molecules) in order to analyze small molecule metabolites.
The CPMG pulse program (RD-90 -(t-180 -t)n-ACQ, where
in addition to definitions above, 180 is a 180 RF pulse, t is
spin-echo delay, n is the number of loops) is a relaxation-
editing pulse sequence.
3. 2D J-resolved sequence: 1H J-resolved experiment provides
J-coupling information and aids in metabolite identification
purposes. In addition, the use of skyline or sum projections of
2D J-resolved spectra [52, 53] have shown to be useful in high
throughput binning approaches to reduce spectral overlaps.
The J-Resolved pulse program [36] has the sequence RD-90-

-t1–180 -t1-ACQ, where t1 is an incremented time period.
4. 1D diffusion edited (DOSY): 1D DOSY is used to attenuate
signals of faster moving small molecules in order to observe
slower moving macromolecules such as lipoproteins and lipids.
DOSY is useful in characterizing lipoprotein fractions in blood
based on diffusion coefficients [54]. The Diffusion-edited
NMR spectra use a pulse sequence with bipolar gradients
(RD-90 -G1–180 -G1–90 -G2-T-90 -G1–180 -G1–90-

-G2-t-90 -ACQ, where RD is a relaxation delay, 90 is a 90
RF pulse, G1 is the pulsed-field gradient that is applied to allow
editing, 180 is a 180 RF pulse, G2 is a spoil gradient applied
to remove unwanted magnetization components.
The diffusion delay D is the time during which the mole-
cules are allowed to diffuse, the period (90 -G1–180 -G1–90-

-G2-T-); t is a delay to allow the longitudinal eddy currents
caused within the sample to decay [36].
For urine, NOESY-presat, and J-resolved spectra are used fre-
quently for data acquisition. Additional selective 1D (e.g., selective
1D-TOCSY), 2D and multidimensional spectra (TOCSY, HSQC,
HMBC, HSQC-TOCSY) that are needed for metabolite identifica-
tion can be acquired using a subset of selected samples, or a pooled
sample. The 2D 1H-13C HSQC experiment is particularly useful for
identification and quantification of metabolites [55]. Other experi-
ments such as 31P NMR can be used to study molecules with
phosphate groups such as energy metabolites (ADP, AMP, and
ATP). 13C or 15N enrichment is used in stable isotope resolved
metabolomics studies to investigate metabolites derived from the
enriched isotopic compound over that of background from endog-
enous sources.
72 Wimal Pathmasiri et al.

3 Quality Control (QC) Methods

Quality Control (QC) starts with the development of an analytical

and data analysis plan that is consistent with the metabolomics
workflow and the goals of the investigation. This plan is reviewed
by all investigators involved in handling samples, data acquisition,
and data analysis. In the metabolomics laboratory, standard
operating procedures and quality control checklists are utilized for
each process: (a) documentation and recording, (b) biospecimen
receipt, inventory, and storage, (c) biospecimen preparation,
(d) instrument tuning and calibration, and adequacy of standards
pre- and poststudy sample analysis, (e) data acquisition of study
samples, and (f) assessment of quality control standards during
study sample data acquisition. A standard laboratory procedure
for following, reporting, and documenting all steps in the workflow
should be established and followed by all laboratory personnel.
Types of quality control standards can vary between labora-
tories based on the methods being employed. The use of reference
material and the use of quality control pools are optimal.
1. Reference material
For large scale studies, longitudinal studies, or studies that
will involve multiple sites conducting metabolites analysis, it is
important to establish a reference material (Fig. 3a). For exam-
ple, creating a large pool of human urine, or plasma/serum,
that is aliquoted and provided to all participating laboratories
can be an ideal way to monitor the metabolomics signatures
longitudinally and between sites. These reference materials are
interspersed with the randomized study samples in each batch,
and the signals/peaks assessed for variability within batch runs,
comparisons between batches, or in longitudinal analysis. Such
reference material facilitates data filtration and merging of data
acquired from multiple batches or between laboratories. This
external pooled sample can be prepared in-house or purchased.
For example, a large number of volunteers contribute biospeci-
mens to a bulk pool, and aliquots are made and stored frozen
for subsequent use. The samples must have the same freeze–
thaw cycles throughout the entire studies in order to use in the
subsequent QC and data analysis procedures. Standard refer-
ence material (e.g., NIST SRM1950 [56]) can also be used as
an external reference standard. Typically, external reference
material samples are prepared identical to the randomized
study samples and interspersed in each batch [43].
2. Study pools
Study samples can also be pooled to create a quality control
within each batch (Fig. 3b, c). The use of QC study pools
[57, 58] not only enables assessment of data quality but also
assists in metabolite identification. In studies where sufficient
 Analysis of NMR Metabolomics Data 73

Fig. 3 Schematic diagram showing preparation of QC samples. (a) Preparation of external pooled QC reference
samples. Biological fluids (serum, plasma, or urine) can be collected from volunteers to make a bulk pooled
QC reference material, which is aliquoted and stored in desirable volumes at
80  C. (b) Preparation of total
study pooled QC samples. For small studies, an equal amount from each sample in a study or a randomly
selected subset of samples is transferred into a container for pooling. Contents are thoroughly mixed and
aliquots of the pooled QC are prepared. For large studies, study pooled QC samples can be prepared by
randomly selecting a subset of samples for pooling and aliquots of these study pooled samples are used for
each of the batch of samples. In addition, batch pooled samples can be prepared by pooling equal amounts
from all samples within the batch. (c) Preparation of both phenotypic and combined pooled QC samples using
the study biospecimens

amount of each biospecimen is available, phenotypic pools and

a total study pool can be prepared. For large-scale metabotyp-
ing studies that involve multiple batches of samples, it is neces-
sary to include QC pools to assess both intra- and inter-batch
variation within a specific study. In studies such as longitudinal
studies, it is also impractical to prepare a pooled QC samples by
using all samples in the study. In such cases, a random selection
of samples representing the whole study can be used. For
example, the entire study sample set can be randomized into
batches, and the first batch can be used to prepare a pooled QC
sample at the beginning of the study, which is aliquoted in
sufficient amounts, and used across the whole study.

3.1 QC of Data NMR spectrometer performance and the field homogeneity is eval-
Acquisition uated routinely, using NMR reference standards (e.g., by optimiz-
ing for best line shape and line width), to ensure high quality of data
acquisition. In addition, NMR spectra are visually inspected for
74 Wimal Pathmasiri et al.

Water Water

0.9 0.9

0.8 0.8
Normalized Intensity

Normalized Intensity
0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
8.5 8.0 7.5 7.0 6.5 6.0 5.5 5.0 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 8.5 8.0 7.5 7.0 6.5 6.0 5.5 5.0 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5
Chemical Shift (ppm) Chemical Shift (ppm)
Poor water suppression Good water suppression

Fig. 4 Effect of water suppression on the quality of the NMR spectrum of serum. Signal intensities of
metabolites are very low in the spectrum with the poor water suppression

water suppression, line shape, phase, and baseline. Water signal in

aqueous samples needs to be suppressed as metabolites are in
micromolar (μM) to millimolar (mM) concentration whereas the
concentration of water is ~55 M (i.e., 110 M 1H). Otherwise HDO
signal will dominate the spectrum making it difficult to analyze or
identify comparatively small metabolite peaks (Fig. 4). Narrow
linewidths of NMR signals increases the peak resolution, especially
in complex spectra with overlapped peak areas. Typically, a line-
width of 1.0  0.5 Hz (without line broadening factor) at half
height of the singlet peak of the chemical shift indicator is desirable.
In addition, the spectral lines should be Lorentzian in shape. If the
water suppression is not sufficient or the line shape and width
are not to the specifications, then the spectra needs to be rerun or
discarded from the analysis.

3.2 Analytical The quality of normalized spectral data is evaluated using the
Reproducibility pooled QC samples. The QC pools are technical replicates and
there should be minimal variation among these samples. An initial
evaluation of data can be carried out using a nonsupervised analysis
method such as principal component analysis (PCA) and other
unsupervised methods such as intraclass correlation [59], or hierar-
chical clustering trees [33] can also be used to assess the analytical
reproducibility. In PCA, a tight clustering of pooled QC samples
and dispersion of study samples [60, 61] indicates that the analyti-
cal variation is minimal compared to biological variation between
study samples. If the QC samples were created from the study
samples, then the QC samples in the PCA plot should be centered
among the sample group from which the QC was created. If
external pools were used then the location of the QC pools in
relation to the study samples does not provide any additional
 Analysis of NMR Metabolomics Data 75

information. QC samples can also be used for applying data filtra-

tion methods to obtain the overall high quality of data. Typically,
coefficient of variation (determined as %CV) within a variable is
assessed using QC samples, and the data with higher %CV (>30%,
for example) are excluded from further data analysis of the entire
dataset.

4 NMR Metabolomics Data Analysis

4.1 NMR Data The position of NMR signals of observed atoms (e.g., 1H) within a
particular molecule are measured as chemical shifts in parts per
million (ppm). The 1H NMR spectrum of alanine is shown is
Fig. 5. The alanine molecule has three carbon atoms with attached
protons with different functional groups (Fig. 5). The 1H NMR
spectrum of alanine in the solvent D2O has two distinct signals: one
for the methyl group (signal A), and one for the H attached to the
C-2 carbon (signal B). 1H atoms in the NH2 group and OH group
usually exchange with deuterium in the solvent and are not observ-
able on the NMR time scale (millisecond). The 1H signal splits into
coupling patterns that are dependent on the neighboring atoms

Fig. 5 NMR Spectrum of alanine showing chemical shifts and J- coupling of CH3 and CH proton groups. The
chemical structure of alanine is shown (in the box) above the NMR spectrum. Protons in the methyl group
(A) are split into 2 lines by a neighboring proton (B) and the CH is split into 4 lines by neighboring methyl
protons (A). Proportion of peak intensities of signal A: B is 3: 1
76 Wimal Pathmasiri et al.

Fig. 6 A 700 MHz 1H NMR spectrum of a pooled urine sample with example identified metabolites.
4,4-dimethyl-4-silapentane-1-sulfonic acid (DSS), is used as an internal standard for chemical shift referen-
cing and relative concentration determination

and their conformations within the molecule (J-coupling). In the

spectrum of alanine, the methyl signal is split into a doublet (two
lines, J coupling ¼ ~7 Hz) by the neighboring CH proton, and the
CH signal is split into a quartet (four lines, J coupling ¼ ~7 Hz) by
the neighboring CH3 protons. The intensities of each signal are
proportional to the number of 1H atoms underlying each signal
(CH3:CH is 3:1). The chemical shifts and splitting pattern (J-
coupling) enables the deconvolution of peaks in the NMR spec-
trum and is used in metabolite identification. Biological samples are
complex mixtures and may contain hundreds of molecules detect-
able by NMR belonging to various compound classes with varying
concentrations that results in a large dynamic range. Some of these
signals are characteristic metabolite peaks, while others are indistin-
guishable due to overlap, and others represent unknowns. A typical
1
H spectrum of human urine with some metabolite annotations is
shown in Fig. 6. To assist with identification of metabolites in
biological samples, multidimensional methods or peak deconvolu-
tion can be incorporated; some of these techniques are described
later in the chapter.

4.2 Data As described earlier, NMR data are recorded as FIDs and are then
Preprocessing Fourier transformed into spectra and this is performed at the
instrument level. Before the Fourier transformation step of the
 Analysis of NMR Metabolomics Data 77

FID, zero filling and apodization is applied to increase the quality of

the spectrum. In the zero filling step, the size of the dataset is
artificially increased by adding zeros to the FID. It increases the
digital resolution of the peaks without affecting the chemical shift
or J-coupling. During apodization, a window function is applied to
the FID (after zero filling) in the form of an exponential multiplier
with a line broadening factor. Exponential multiplication increases
the signal-to-noise ratio of the spectrum and the line broadening
factor reduces the resolution. Therefore, selecting the optimum
parameters for apodization is a tradeoff between the resolution
and the signal-to-noise ratio. The resultant spectral signals are
Lorentzian in shape for a spectrum recorded in a homogeneous
magnetic field. There can be phase errors in the spectrum due to the
time delay between the NMR excitation pulse and the switching on
the receiver. Phase correction is applied to obtain a spectrum in the
absorption mode. The next step is to correct baseline offsets by
subtracting a low order polynomial fit using software. Finally, the
origin of the chemical shift scale is calibrated to the position of the
internal standard signal or other standard peak. One of the com-
plications in using NMR-based metabolic profiles is the presence of
very small but significant variations in the chemical shift of metab-
olite peaks between spectra. NMR chemical shift depends on the
microenvironment within the molecule and subtle changes in pH
or ionic strength among samples can result in variations in peak
positions. A peak alignment algorithm must be used in such cases.
A number of algorithms have been reported in literature, for exam-
ple, recursive segment-wise peak alignment (RSPA) [62], interval-
correlation-shifting (icoshift) [63], hierarchical cluster-based peak
alignment (CluPA), [64], correlation optimized warping (COW)
[65], dynamic time warping (DTW) [65], and recursive unrefer-
enced alignment of spectra (RUNAS) [66]. An example of peak
alignment using CluPA algorithm applied in NMRProcFlow using
urine samples is shown in Fig. 7. Although these methods improve
peak alignment, they each suffer slight limitations because of chal-
lenges associated with complexity and signal overlap in NMR spec-
tra. However, it is useful in situations where certain select peaks
need to be aligned.

4.2.1 Phenotypic After phase and baseline correction, and chemical shift referencing,
Anchoring NMR data is transferred to a data format accessible for further
statistical analysis and modeling. An NMR spectrum is a data matrix
consisting of chemical shift and peak intensity information. A single
spectrum usually contains up to 65,536 data points. There are two
methods for handling NMR data: (1) Using raw data (full resolu-
tion) or binned data (data reduction); (2) Peak deconvolution,
identification, and relative quantification of metabolites. Full-
resolution data analysis is computationally intensive and generally
requires high-performance computing clusters. In such cases, the
78 Wimal Pathmasiri et al.

700000
600000
Before
500000

400000

300000
200000

100000
0
3 2.9 2.8 2.7 2.6 2.5 2.4 2.3
700000
600000
After
500000

400000

300000
200000

100000
0
3 2.9 2.8 2.7 2.6 2.5 2.4 2.3

Fig. 7 NMR spectra of urine samples aligned by CluPA algorithm using NMRProcFlow software

raw data can directly be used for modeling. Otherwise, data must
be preprocessed using a process called binning. Both of these
processes offer high-throughput data analysis. Some artifacts and
unwanted spectral regions must be removed prior to further analy-
sis of either raw or processed spectral data. These include residual
water signal, internal standards, and applied chemicals or medica-
tion and their metabolites used as treatments in the study. The
NMR peaks in the spectra can be deconvoluted into metabolites
and quantified. Despite efforts toward automation of metabolite
identification, fitting, and quantification of metabolite levels in 1D
1
H NMR spectra, this process is somewhat hampered by signal
overlap, high dynamic range of metabolites, small shifts in chemical
shift between spectra for the same metabolites, and complex
splitting patterns (multiplicity) of the peaks of a metabolite, and
level of noise. Another disadvantage of automation is that unknown
metabolites cannot be identified, since only known metabolite
peaks can be modeled in automating.

4.2.2 Data Analysis In order to recover meaningful information on metabolic biomar-

Using Raw NMR Data kers using NMR data, data preprocessing steps and application of a
variety of bioinformatics or statistical analysis methods are neces-
sary. NMR data can be analyzed by using full resolution spectra or
binning approach with reduced dimensionality. In either of the
methods, NMR data is typically organized in a matrix in such a
way that the samples (observations) are in rows and variables are in
columns. In the full resolution data analysis, data are directly
imported into analysis software of choice. Either FIDs or raw Four-
ier transformed data can be imported into software and an array of
data can be prepared using software codes. The RBNMR [67] code
 Analysis of NMR Metabolomics Data 79

written in Matlab is a useful tool that can prepare a data structure

that includes metadata. Appropriate Matlab codes must be used
(in-house or already coded) to extract the data from such data
structures into further analysis. In-house or publicly available
codes written in other languages (R, Python, Java, C++) can also
be used for this purpose. It is also possible to import raw FIDs
(before Fourier transformation) directly into software and auto-
mate data processing and preprocessing steps by using in-house or a
variety of publicly available tools or algorithms. These steps include
Fourier transformation, phase and baseline correction, chemical
shift referencing, peak alignment, removal of unwanted regions,
and normalization and scaling. Once generated, a variety of data
analytic methods can be applied on the resultant data matrix to
recover biomarker information from full resolution NMR data
(methods outlined or described in other sections).
In a Bayesian-based method called CRAFT [68] (Complete
Reduction to Amplitude Frequency Table), which is a time domain
approach, FIDs are directly converted into a frequency-amplitude
table. The workflow in the CRAFT consists of a two-step process
[68]. The first step involves digitally filtering and down sampling of
data to obtain several sub-FIDs. During the second part of the
CRAFT process, these sub-FIDs are modeled as sums of decaying
sinusoids using the Bayesian probability theory. Authors report that
this approach is better than the analysis of Fourier transformed
frequency domain spectra for identifying changes in metabolite
signals in complex spectra with signal overlaps, such as biological
samples [68].

4.2.3 NMR Binning Binning or bucketing implements data reduction of NMR metabo-
lomics data. In the conventional binning approach, the spectrum is
divided into evenly spaced windows called bins or buckets and the
peak area in each bin is integrated. The bucket width is typically
0.01 or 0.04 ppm (some uses 0.001 ppm bin width). Binning can
also reduce the effect of variations in peak positions in spectra to
some extent. The main disadvantage of conventional binning is that
parts of metabolite peaks can fall into the neighboring bins or
buckets. Several methods have been proposed to overcome this:
intelligent bucketing [69, 70], adaptive binning [71], and opti-
mized bucketing [72]. In the intelligent binning, the bin bound-
aries are loosened by a specified percentage (e.g., 50% looseness)
allowing peaks to fall within the same bin. Adaptive binning cor-
rects for variation in chemical shifts through utilization of a wavelet
transform. The optimized bucketing algorithm generates an aver-
age spectrum of all data and defines bin boundaries using that
spectrum. The NMR binning of projections of 2D J-resolved spec-
tra can be performed by using the JBA algorithm [73] employed in
the R-based MWASTools [74] software.
80 Wimal Pathmasiri et al.

4.2.4 NMR Data Data normalization is one of the most important steps in metabo-
Normalization lomics analysis, and the goal is to remove any variation in total
amount of material between samples. Specifically, it is important for
analyzing urine samples where concentration can have changes in
orders of magnitude. Similarly, it is important to normalize data
when analyzing cell or tissue extracts where it can be difficult to
control analyte levels due to factors such as sampling or extraction
efficiency. These changes would be disproportionate to the relevant
biological or biochemical variations attributed to metabotype, thus
hindering extracting valuable biomarker information. Despite
many methods in the literature for data normalization, no single
approach is optimum for any given study. The data must be pro-
cessed appropriately according to the study and type of samples and
data. The normalization of data is typically performed row-wise
(on observations). Commonly used normalization methods
include constant sum normalization, probabilistic normalization,
cubic spline normalization, normalizing to a standard peak, and
normalizing to the total volume collected, specific gravity, or to the
total mass of sample used. In the constant sum normalization, each
integral (or intensity) is divided by the sum of all integrals or
intensities and typically multiplied by a constant number such as
1000. Probabilistic quotient normalization (PQN) [75] is based on
the calculation of a most probable dilution factor by looking at the
distribution of the quotients of the amplitudes of a test spectrum by
those of a reference spectrum. It was shown that the PQN method
is more robust and more accurate than the widespread total sum
normalization using experimental spectra of a complete metabolo-
mics study as well as simulated spectra. It should be noted that the
choice of a standard metabolite should be carefully selected and it
should not have biological variation due to metabotype. For exam-
ple, though normalization to the creatinine peak in urine samples is
common in clinical studies, creatinine would not be a suitable
choice where it is expected to change in patients (e.g., in kidney
disease).

4.2.5 Centering, Scaling Appropriate data pretreatment could include centering, scaling or
and Transformation of Data transformation, methods which are performed column-wise
(on variables) in a table of data [76]. Centering (mean centering)
helps to adjust for the skew between high and low abundance
metabolites in order to emphasize the relevant variation (covari-
ance) between biological samples. Centering is done by centering
the mean of each column (variable) to zero and subtracting the
value of the variable from the mean of each column. Scaling of data
is a method that adjusts for the differences in magnitude between
metabolites by converting the data into concentration change rela-
tive to a scaling factor [76].
Data analysis without this preprocessing step could lead to loss
of information arising from only slight changes in metabolite levels
 Analysis of NMR Metabolomics Data 81

in the metabolic pathways. Metabolites with higher concentration

levels could become dominant and hinder small but significant
biological variations in the metabotype. Scaling methods fall
under one of two subclasses where the scaling factor is either (1) a
data dispersion measure (e.g., standard deviation) or (2) a size
measure (e.g., mean, median) [76]. Examples of the data dispersion
method which use spread as the scaling factor are: auto scaling;
Pareto Scaling; Variable stability (VAST) scaling [77]; range scaling.
In auto scaling (also called unit variance (UV)), the standard devia-
tion is used as the scaling factor. Because in this method all meta-
bolites become equally important, one potential problem is the
artificial inflation of small values. VAST scaling builds upon auto-
scaling but aims to focus on stable metabolites by using both
standard deviation and the coefficient of variation as scaling factors.
Pareto scaling uses the square root of standard deviation as scaling
factors, and can reduce the dominant effect of high metabolite
concentrations while minimizing potential inflation of low concen-
tration metabolite data. In the range scaling method, the range
between minimum and maximum value is used [76]. Examples of
size measure scaling (with a magnitude-based scaling factor) are:
level scaling and median scaling. Level scaling uses mean concen-
tration as the scaling factor, and converts metabolite concentration
changes into changes in percent concentration. Level scaling is an
effective approach when large relative changes are expected but
typically median scaling is a the more robust size measure scaling
approach. Transformations are used to make skewed distributions
of data more symmetric. Log and power transformation are some
examples. Since normalization and scaling have different purposes,
it is possible to use a combination of these methods.

4.3 Data Analysis As outlined earlier, the NMR metabolomics data can be analyzed in
Methods either raw or preprocessed forms. A vast number of statistical
analysis and modeling approaches are available in the literature
and the number is rapidly increasing. Most of these analyses are
performed by using Matlab (or Mathematica), R, Java, or Python
codes. Some methods have focused on integrating the whole NMR
metabolomics pipeline (from importing, preprocessing, analyzing,
and graphical visualization to identifying biomarker metabolites)
while others have been developed for addressing specific data pro-
cessing, modeling, or identifying needs.

4.3.1 Multivariate Data Metabolomics analysis involves data in high dimensional data
Analysis matrices: typically, samples in rows and variables in columns. The
use of multivariate data analysis methods [78] is common in analyz-
ing metabolomics data [79]. Multivariate data analysis is a
projection-based method that uses eigenvalues and a number of
variations have been reported in literature [78, 80, 81]. The non-
supervised multivariate method, principal component analysis
82 Wimal Pathmasiri et al.

(PCA), is a technique that can be used to observe pattern, trends, or

outliers in a data set. Quality control of data at the sample level can
also be assessed using PCA (QC pooled samples must be tightly
clustered in PCA [60, 61]). Supervised analysis methods such as
PLS-DA and OPLS-DA can be used to identify discriminatory
metabolites or features that are perturbed in the biological system.
The OPLS-DA method is used to filter out unwanted variation
within data that are not due to biochemical/biological changes in
the metabotype. The scores plot typically depicts the subjects and
the trends in PCA or group separation in PLS-DA and OPLS-DA
and the loading plot shows the variables that are responsible for the
trend or group separation. The models must be cross-validated in
order to avoid overfitting. Another important feature in the super-
vised multivariate data analysis (PLS-DA and OPLS-DA) is the
Variable Influence on Projections (VIP) plot, which generates a
list of variables (metabolites or features) that are responsible for
the discrimination of the study phenotypes. The coefficient plots
also provide biomarker information in the OPLS-DA which can be
used to generate statistical correlation spectroscopic plots such as
STOCSY (described later). Most of the methods in literature have
used codes developed in Matlab or other similar programming
languages or commercially available software packages such as
SIMCA. Since metabotype is dependent on many factors such as
genetics, diet, disease status, and environment, there can be situa-
tions where confounding factors can affect the outcome of model-
ing using multivariate data. Posma and coworkers reported that
Covariate-Adjusted Projection to Latent Structures (CA-PLS) [82]
approach can be used in such situations where confounding is an
issue.

4.3.2 Statistical The development of statistical correlation spectroscopic tools was

Correlation Spectroscopy based on the covariance in peak intensities across multiple indepen-
dent samples to generate associations between signals that arise
from the same molecular structure [8]. Statistical total correlation
spectroscopy (STOCSY) and an extended collection of software
tools based on STOCSY using statistical correlation spectroscopy
has been introduced and some select tools are described below.
Statistical total correlation spectroscopy (STOCSY) [83] is a
method that assists biomarker identification in NMR metabolomics
and displays the correlating peaks in the NMR spectra. It differs
from most standard 2D correlation spectroscopy (such as TOCSY)
with the advantage that it correlates intramolecular peaks and also
can show correlations of peaks that belong to other metabolites in
the same pathway (positively or negatively) [8, 83] that are per-
turbed by biological processes. A combination of OPLS-DA
(supervised multivariate data analysis) and STOCSY can provide
information about discriminatory metabolite peaks and their
 Analysis of NMR Metabolomics Data 83

identification as potential biomarkers in metabolic pathways. In this

method, a driver peak is used to identify and assign other correlated
peaks into metabolites. Subset optimization by reference matching
(STORM) [84] is a tool that uses multivariate data analysis for
information recovery. It is capable of selecting subsets of spectra
that have information about discrimination between the study
groups. Cluster analysis statistical spectroscopy (CLASSY) is
another method that combines qualitative metabolic profiling and
quantitative changes in the biological matrix using a local-global
correlation clustering Scheme [84]. It can visualize graphically in a
high throughput way. Blaise and coworkers have reported an algo-
rithm called statistical recoupling of variables (SRV) [85] and it
utilizes covariance/correlation ratios of consecutive variables in the
NMR spectra before significant testing. It can extract information
about statistically significant metabolites in the perturbed
biological systems. The recoupled-statistical total correlation spec-
troscopy (RSTOCSY) is another useful tool that combines SRV
with STOCSY [86]. A further development with orthogonal signal
filtering method has been applied in orthogonal filtered recoupled-
STOCSY (OR-STOCSY) [86] to find a list of pairs of metabolites
that experience correlated perturbations. Statistical homogeneous
cluster spectroscopy (SHOCSY) [87] is another statistical
approach that was developed to remove unwanted variation within
a biological class in the metabolomics analysis. The algorithm in
SHOCSY is shown to be capable of identifying a subset of spectra
by categorizing into clusters of spectra that have similar spectral
features (similar biochemical composition) improving the predict-
ability in modeling.

4.3.3 Metabolite A 1H NMR spectrum of a biological matrix is complex and consists

Identification of thousands of signals deriving from both endogenous and exoge-
nous metabolites. Therefore, a successful combination of multivar-
iate and univariate data analysis approaches can provide biomarker
spectral features in the 1H NMR spectra. The statistical correlation
spectroscopy allows for recovering important information on
marker metabolites by reducing the complexity in the spectra. In
addition to the 1D NMR spectra, two or multidimensional NMR
methods can provide valuable information for the identification of
these features as metabolites. The application of a combination of
1D, 2D (1H J-resolved, 2D DQF-COSY, TOCSY, HSQC, and
HMBC), and multidimensional NMR spectroscopic methods
[88–90] to obtain spectra on selected samples (pooled QC sample,
for example) aids in identification of metabolites in biological sam-
ples. In addition, phenotype-specific metabolite information
(abundant in one phenotype while minimally present or absent in
another) may also be recovered if spectral data are available for
representative phenotypic pooled samples. The metabolite peaks
84 Wimal Pathmasiri et al.

in 1D NMR or 2D NMR can be used to search online databases

such as Human Metabolome Database (HMDB), Biological Mag-
netic Resonance Data Bank (BMRB), and Birmingham Metabolite
Database. The DQF-COSY (1H-1H) and TOCSY (1H-1H or
13
C-13C) experiments are designed to obtain homonuclear cou-
pling information within a molecule. The TOCSY experiment pro-
vide information about the entire 1H-1H spin system that are
coupled in a molecule. The 2D HSQC method gives direct con-
nectivity of atoms in the molecule (e.g., 1H-13C or 1H-15N)
whereas HMBC can provide information about long-range bonds
between atoms within the molecule. Another useful NMR experi-
ment is a hybrid HSQC and TOCSY experiment, 2D 1H-13C
HSQC-TOCSY, for example. The 2D HSQC-TOCSY experiment
can provide two types of information: direct 1H-13C correlations
and relayed correlations connecting each protonated carbon with
all 1H nuclei belonging to the same spin system. 2D INADE-
QUATE experiment provides 13C-13C direct bond information
enhancing the elucidation of carbon backbone structure of com-
pounds. 3D HCCH-TOCSY and 3D HCCH-COSY experiments
are some examples for multidimensional experiments that aid
metabolite identification. Use of 2D, hybrid, and multidimensional
NMR spectra (alone or combined) facilitate identification of known
and unknown metabolites found in complex mixtures without the
need of any physical separation of metabolites [90]. A number of
software-based approaches for metabolite identification is available
in literature for metabolite identification. Complex Mixture Analy-
sis by NMR (COLMAR) [91] is a publicly available web server,
which has a collection of tools deployed in web servers [89] for
identification of metabolites using spectra collected as 1D (1H,
13
C), 2D (TOCSY, HSQC), and multidimensional (HSQC-
TOCSY) spectra and by incorporating searches in databases such
as BMRB and the Carbon TOCSY NMR Metabolomics Database
(TOCCATA) [92].
The ratio analysis nuclear magnetic resonance spectroscopy
(RANSY) [93], a covariance-based method reported in literature,
and uses ratios of peak height or integrals to identify all the peaks of
a specific metabolite in NMR spectra of a biological samples. It is
based on the principle that all the peak ratios of a molecule are fixed
and proportional to the number of protons underlying each peak.
The peak ratios are divided by covariance in this method to gener-
ate an individual spectrum of a metabolite.
Use of smart isotope tags in NMR spectroscopy is another
method that enables identification of metabolites in complex
biological samples. In this approach, 15N-labeled ethanolamine or
cholamine [94, 95] is used in a derivatization reaction to tag the
15
N label into carboxyl-containing metabolites. A 15N and 1H
chemical shift library of metabolites has been developed using 2D
 Analysis of NMR Metabolomics Data 85

1
H-15N HSQC spectra. The smart tagging is performed on
biological sample, 1H-15N HSQC spectra are recorded on the
samples, and 15N and 1H chemical shifts observed for cross peaks
in the spectra are compared to those in the library.
1D 13C spectroscopy (proton decoupled) has many advantages
for a metabolomics study, including a large spectral dispersion,
narrow singlets at natural abundance, and a direct measure of the
backbone structures of metabolites compared to 1H NMR spectra.
However, it suffers from the very low sensitivity due to low natural
abundance of 13C isotope (~1.1%) combined with a decreased
gyromagnetic ratio γ (one quarter of that of 1H; energy transition
is proportional to γ3, and sensitivity is decreased by 64 times) unless
the metabolomics experiments are performed by using 13C enrich-
ment experiments. Another disadvantage of this experiment is the
inability to detect quaternary carbon atoms (carbon atoms with no
protons attached). On the other hand, an INADEQUATE (incred-
ible natural abundance double quantum transfer experiment)
experiment can provide the complete carbon skeleton structure of
a metabolite. This type of experiment is very helpful when 13C
isotope enrichment is performed prior to analysis. A semiauto-
mated algorithm was developed by Clendinen and coworkers
[96], INADEQUATE Network Analysis (INETA), where untar-
geted analysis can be performed and metabolites can be identified
by using INADEQUATE spectral data and an in silico database
constructed using BMRB database.
There are efforts to deconvolute the signals in NMR spectra
into metabolites for the identification of metabolites. Another
advantage of NMR spectroscopy is that the same spectrum can be
used for both untargeted and semitargeted analysis. Since NMR
spectroscopy is a quantitative technique, the use of an internal
standard with known concentration allows for quantitation of
metabolites [97]. This is a relative or semiquantitative method
because there is no calibration curve employed to obtain absolute
concentration. For example, Chenomx is a commercial package
that contains a library of metabolites that has been modeled to
account for chemical shift variation due to pH. It should also be
noted that the spectra should be acquired to account for the total
relaxation of 1H atoms of the metabolites in the biological sample,
in order to allow for identification and semiquantification of meta-
bolites in the sample. It is possible to apply this method to profile a
selected set of metabolites in a batch profiling mode. The experi-
ence of the analyst or prior knowledge of metabolites signals plays a
critical role in this approach to a successful completion of identifi-
cation and semiquantification using this method. On the other
hand, Röhnisch et al. [98] have recently reported automated quan-
tification algorithm (AQuA) for targeted quantification of metabo-
lites that can account for signal overlaps of metabolites in complex
86 Wimal Pathmasiri et al.

spectra in a high throughput fashion. It utilizes spectral data

extracted from a library consisting of one standard calibration
spectrum for each metabolite. The authors have tested this algo-
rithm, compared with manual fitting using Chenomx software, and
concluded that this approach produced accurate results with high
efficiency. There are open source methods such as BATMAN and
Bayesil that has been reported in literature for automation efforts
for the identification and quantification of metabolites. BATMAN
[99, 100] is an R- package that models NMR resonances on the
basis of a user-controllable set of templates, each of which specifies
the chemical shifts, J-couplings and relative peak intensities for a
single metabolite. It produces outputs to include relative concen-
tration for named metabolites together with associated Bayesian
uncertainty estimates, and the fit of the remainder of the spectrum
using wavelets. An improvement to this approach has been
reported [101], where spectral ordering (sorting) is combined
with peak deconvolution using BATMAN software. BAYESIL
[102] (http://bayesil.ca/) performs several NMR processing
steps and then matches spectra against a reference metabolite
library using a probabilistic graphical model that approximates the
most probable metabolic profile. It should be kept in mind that
metabolites in signal overlapped regions of an 1H NMR spectra
make deconvolution and quantification difficult for complex
biological samples. The use of 2D 1H-13C HSQC is helpful since
it enables signal dispersion. A method has been developed by using
the fast maximum likelihood reconstruction (FMLR) algorithm
[55] to identify and quantify metabolites using 2D 1H-13C
HSQC spectra acquired in about 15 min. The software package
rNMR [103] provides a simple GUI-based method for visualizing,
identifying, and quantifying metabolites across multiple one- or
two-dimensional NMR spectra.

4.3.4 Workflows There are a number of open source workflows, and web-based
and Web-Based Analysis programs available for NMR metabolomics analysis. In addition,
Programs standalone programs and web-based servers have been proposed in
literature for metabolomics data analysis. Some of these methods
are briefly described in this section with relevant literature and links.
Automics [104] is an integrated metabolomics analysis platform
that has functionality to each step in a NMR metabolomics work-
flow including processing spectra, nonsupervised and supervised
multivariate data analysis and graphical visualization. KIMBLE
(KNIME-based Integrated MetaBoLomics Environment [105] is
a KNIME workflow management system based NMR metabolo-
mics workflow. It is a platform that is integrated with algorithms
and software tools necessary for untargeted and targeted analysis. It
is self-documenting and combines data, algorithms, and software in
one file [105], allowing for reuse of the exact workflow parameters
 Analysis of NMR Metabolomics Data 87

in future. Even the workflow for a single project can be saved with
all tools and parameters as a single virtual machine. These capabil-
ities make KIMBLE a more robust platform for reproducibly using
the NMR metabolomics workflow. Metaboanalyst is a web-based
metabolomics platform [106–109] for uploading metabolomics
data (both NMR and MS) and data analysis including pathway
analysis. It has a number of modules with various preprocessing,
statistical, multivariate analysis, identification, pathway analysis, and
integrated omics data analysis tools. A standalone package for local
installation is available on Metaboanalyst web page. Metabolomics
Univariate and Multivariate Statistical Analysis (MUMA) [110] is
an R-based pipeline that guides the user in the data analysis (uni-
variate and multivariate data analysis) and interpretation. It also
provides additional tools specifically designed to help the user in
the interpretation of NMR data, such as STOCSY and RANSY.
MVAPACK [111] is an open source package written in GNU
Octave programming language to process metabolomics data
from FIDs to modeling. It has a collection of tools including
those needed for traditional NMR processing (apodization, zero-
filling, Fourier transformation, manual and automatic phase cor-
rection, region of interest selection, and peak picking, integration,
and referencing), data preprocessing (alignment, binning, normali-
zation, and scaling), and multivariate data analysis (nonsupervised
and supervised), and model validation tools. NMRProcFlow
[112] (https://nmrprocflow.org/) is a web-based work flow to
process 1D NMR spectra interactively with visualization tools.
The workflow can be accessed through the web or can be installed
locally as a virtual machine. It allows the user to import spectra and
perform baseline and phase correction, chemical shift calibration,
peak alignment, NMR binning (after removing unwanted regions
including baseline noise), and merging of metadata. The workflow
process can be saved and used again similarly for other datasets. A
cloud-based computing framework for NMR metabolomics analy-
sis using Hadoop streaming with Matlab [113] has been reported
indicating the possibility of using the cloud-based computing
power for NMR metabolomics applications. Pathomx [114] is
another workflow-based tool developed using Python for proces-
sing, analyzing, visualizing, and exporting metabolomics data. The
R-based package speaq 2.0 [115] contains a workflow that includes
peak alignment using CluPA algorithm, peak picking, data imputa-
tion, and peak table generation. The output that speaq produces is
compatible for using other R-based statistical and multivariate data
analysis packages. Workflow4Metabolomics (W4M) [116] is an
open source and Galaxy-based web platform (https://
workflow4metabolomics.org/) that has algorithms for data pre-
processing, statistical analysis, and annotation. It is a virtual
research environment based on high performance computing
88 Wimal Pathmasiri et al.

environment (660 cores and 100 TB). It has modules for analyzing
both MS and NMR data. Users can access the web version or the
whole framework and computing tools can be installed locally as a
virtual machine. ASCIS (Automatic Statistical Identification in
Complex Spectra) [117], is another R-based software package
that comprises of a workflow with algorithms developed for auto-
mated analyzing of NMR spectra. The workflow includes raw NMR
data import, baseline processing, peak alignment, NMR binning,
metabolite identification and quantification. In addition, ASICS
workflow includes tools for multivariate data analysis (PCA and
OPLS-DA), identification of NMR bins or metabolites that dis-
criminate phenotypes, and some other statistical analysis.

4.3.5 Tools In molecular epidemiological studies, genome-wide association

for Metabotyping studies (GWAS) are conducted to discover the genetic associations
in Population-Based with the disease risk in the populations. With the technical advances
Association Studies in spectroscopy, Metabolome Wide Association Studies
(MWAS) can be conducted to reveal metabolite associations with
the disease risk. The use of metabotyping in epidemiologic studies
can facilitate the identification of disease-risk biomarker metabo-
lites that can be used for developing diagnostic tools or tools for
monitoring treatment interventions. The potential application of
NMR metabotyping in MWAS studies was demonstrated in litera-
ture [52, 118, 119]. The 1D 1H NMR spectra [118] or 1D
projections (skyline or sum) of 2D 1H J-resolved spectra (pJRES)
[52] can also be utilized in the MWAS approach. An R-based
MWASTools (https://rdrr.io/bioc/MWASTools/) [74] package
was introduced by Rodriguez-Martinez and coworkers and it pro-
vides a complete pipeline to perform metabolome-wide association
studies. It has many tools such as quality control analysis, data
filtering, different MWAS association models, model validation,
visualization, NMR metabolite identification using STOCSY algo-
rithm, and biological interpretation using pathway mapping. Simi-
larly, metabotype quantitative trait locus (mQTL) mapping has
emerged as a new tool that uses quantitative variation of 1H
NMR spectra [12, 13] between biological samples. The 1D projec-
tions of J-resolved spectra (pJRES) have been used [52] in this
approach as pJRES spectra have been shown to demonstrate
enhanced peak dispersion, efficient attenuation of lipid resonances,
better metabolite identification capacities, and analytical
reproducibility.

4.3.6 Metabolic Pathway Metabotyping studies can result in identification of certain meta-
Analysis bolites, which have undergone changes in levels in the biological
samples. Then the next step is to interpret these findings to answer
the biological questions mechanistically by exploring the metabolic
pathways, which have been perturbed due to the disease
 Analysis of NMR Metabolomics Data 89

phenotype, or treatment intervention, for example. Metabolic

pathways are complex and inter-connected system of genes, tran-
scripts, proteins, and metabolites. In order to query pathways,
software tools are needed and a number of pathway analysis tools
have been developed and curated by researchers and commercial
vendors. A few useful software tools are discussed here. Metaboa-
nalyst has tools for metabolic pathway analysis and modules for
integrative analysis for metabolomics, metagenomics, and tran-
scriptomics data. It has algorithms for over-representation analysis,
metabolite set enrichment analysis, and pathway topology analysis.
A built-in reference metabolome is included in the pathway analysis
and the user has the option using own in-house reference metabo-
lome if available. Metscape [120] is a tool for interactive explora-
tion and visualization of experimental metabolomics and gene
expression data in the context of metabolic networks. A prominent
feature is the ability to enter gene expression data and to examine
them in the context of metabolic networks. Metscape plug-in can
be used using Cytoscape software (https://cytoscape.org/).
Integrated pathway-level analysis (IMPaLA) [121], a web-based
tool, allows for joint pathway analysis using transcriptomics or
proteomics and metabolomics data. Overrepresentation or enrich-
ment analysis can be performed with user-specified lists of metabo-
lites and genes using over 3000 preannotated pathways from
11 databases. MetaCore (https://clarivate.com/products/meta
core/, Clarivate Analytics, PA, USA) is a commercially available
web-based software, and it has many features including pathway
enrichment analysis, network building, and integrated pathway
analysis.

4.3.7 Metabolomics Data It is important to make data and metadata including experimental
Repositories data of metabolomics studies available to the research community
for various reasons. It helps a wider community of researchers and
bioinformaticians to obtain data for testing their software tools
using publicly available data. It also facilitates reproducibility of
conducting metabolomics studies by the availability of standardized
experimental protocols. Therefore, establishing metabolomics data
repositories and data coordinating centers are important and the
Metabolomics Workbench [21] in the USA and MetaboLights in
the Europe are two such data repositories. The Metabolomics
Workbench serves as a national and international repository for
metabolomics data and metadata and provides analysis tools and
access to metabolite standards, protocols, tutorials, training, and
more functionalities. On the other hand, Metabolights [122] is a
database for metabolomics experiments, experimental data, and
metadata for a variety of species, techniques and it covers metabo-
lite structures and their reference spectra as well as information on
their biological roles, locations, and concentrations.
90 Wimal Pathmasiri et al.

5 Summary

In this chapter, we have described the major steps in a typical NMR

metabolomics workflow with the focus on the available data analysis
(computational and bioinformatics) tools and key important points
to consider when designing an analysis method. The high-
resolution NMR spectroscopy is a powerful analytical technique
with higher reproducibility for metabolic phenotyping studies.
Because of its robustness, NMR based metabotyping is particularly
suitable for large-scale population-based metabolomics studies that
involves thousands of samples and data acquisition using multiple
batches of samples. Along with the advances in the NMR technol-
ogy, the computational power (hardware, clusters, and cloud com-
puting) has grown and the development of statistical and
bioinformatics tools is also on the rise. Hence, the potential use
of NMR spectroscopy for large-scale metabotyping studies
enabling discovery of biomarkers of disease risk in populations in
parallel to the genome wide association studies (GWAS) has also
become increased. The methods described in this chapter can
equally be applicable for metabolomics studies involving plants,
microbes, and other model organisms.

References

1. Jacobsen NE (2007) NMR spectroscopy Physiol 227(8):2975–2981. https://doi.

explained : simplified theory, applications org/10.1002/jcp.24002
and examples for organic chemistry and struc- 7. Ryan D, Robards K (2005) Metabolomics:
tural biology. Wiley, Hoboken, NJ the greatest omics of them all? Anal Chem
2. Nicholson JK, Connelly J, Lindon JC, 24:285–293
Holmes E (2002) Metabonomics: a platform 8. Robinette SL, Lindon JC, Nicholson JK
for studying drug toxicity and gene function. (2013) Statistical spectroscopic tools for bio-
Nat Rev Drug Discov 1:153–161 marker discovery and systems medicine. Anal
3. Kaddurah-Daouk R, Kristal BS, Weinshil- Chem 85(11):5297–5303. https://doi.org/
boum RM (2008) Metabolomics: a global 10.1021/ac4007254
biochemical approach to drug response and 9. Bird SS, Sheldon DP, Gathungu RM,
disease. Annu Rev Pharmacol Toxicol Vouros P, Kautz R, Matson WR, Kristal BS
48:653–683. https://doi.org/10.1146/ (2012) Structural characterization of plasma
annurev.pharmtox.48.113006.094715 metabolites detected via LC-electrochemical
4. Stewart DA, Dhungana S, Clark RF, Pathma- coulometric array using LC-UV fractionation,
siri WW, McRitchie SL, Sumner SJ (2015) MS, and NMR. Anal Chem 84
Omics technologies used in systems biology. (22):9889–9898. https://doi.org/10.1021/
In: Fry R (ed) Systems biology in toxicology ac302278u
and environmental health, 1st edn. Academic, 10. Sasaki K, Sagawa H, Suzuki M, Yamamoto H,
Waltham, MA, pp 57–84 Tomita M, Soga T, Ohashi Y (2018) A meta-
5. Sumner SCJ, Pathmasiri W, Carlson JE, bolomics platform by capillary electrophoresis
McRitchie SL, Fennell TR (2018) Metabolo- coupled with a high-resolution mass spec-
mics. In: Smart R, Hodgeson E (eds) Molec- trometry for plasma analysis. Anal Chem 91
ular and biochemical toxicology. Wiley, (2):1295–1301. https://doi.org/10.1021/
Hoboken, NJ acs.analchem.8b02994
6. Johnson CH, Gonzalez FJ (2012) Challenges 11. Bictash M, Ebbels TM, Chan Q, Loo RL, Yap
and opportunities of metabolomics. J Cell IK, Brown IJ, de Iorio M, Daviglus ML,
Holmes E, Stamler J, Nicholson JK, Elliott P
 Analysis of NMR Metabolomics Data 91

(2010) Opening up the "Black Box": meta- 19. Phenome Center Birmingham. https://www.
bolic phenotyping and metabolome-wide birmingham.ac.uk/research/activity/
association studies in epidemiology. J Clin phenome-centre/index.aspx. Accessed
Epidemiol 63(9):970–979. https://doi.org/ February 2019
10.1016/j.jclinepi.2009.10.001 20. Australian National Phenome Center.
12. Hedjazi L, Gauguier D, Zalloua PA, Nichol- https://www.wahtn.org/enabling-pla
son JK, Dumas ME, Cazier JB (2015) mQTL. tforms/australian-national-phenome-centre/
NMR: an integrated suite for genetic mapping . Accessed February 2019
of quantitative variations of (1)H NMR-based 21. Sud M, Fahy E, Cotter D, Azam K, Vadivelu I,
metabolic profiles. Anal Chem 87 Burant C, Edison A, Fiehn O, Higashi R, Nair
(8):4377–4384. https://doi.org/10.1021/ KS, Sumner S, Subramaniam S (2016) Meta-
acs.analchem.5b00145 bolomics Workbench: an international reposi-
13. Cazier JB, Kaisaki PJ, Argoud K, Blaise BJ, tory for metabolomics data and metadata,
Veselkov K, Ebbels TM, Tsang T, Wang Y, metabolite standards, protocols, tutorials and
Bihoreau MT, Mitchell SC, Holmes EC, Lin- training, and analysis tools. Nucleic Acids Res
don JC, Scott J, Nicholson JK, Dumas ME, 44(D1):D463–D470. https://doi.org/10.
Gauguier D (2012) Untargeted metabolome 1093/nar/gkv1042
quantitative trait locus mapping associates 22. Markley JL, Bruschweiler R, Edison AS, Egh-
variation in urine glycerate to mutant glyce- balnia HR, Powers R, Raftery D, Wishart DS
rate kinase. J Proteome Res 11(2):631–642. (2017) The future of NMR-based metabolo-
https://doi.org/10.1021/pr200566t mics. Curr Opin Biotechnol 43:34–40.
14. Gibson G, Gieger C, Geistlinger L, https://doi.org/10.1016/j.copbio.2016.08.
Altmaier E, Hrabé de Angelis M, 001
Kronenberg F, Meitinger T, Mewes H-W, 23. Ludwig C, Easton JM, Lodi A, Tiziani S,
Wichmann HE, Weinberger KM, Adamski J, Manzoor SE, Southam AD, Byrne JJ, Bishop
Illig T, Suhre K (2008) Genetics meets meta- LM, He S, Arvanitis TN, Günther UL, Viant
bolomics: a genome-wide association study of MR (2011) Birmingham metabolite library: a
metabolite profiles in human serum. PLoS publicly accessible database of 1-D 1H and
Genet 4(11):e1000282. https://doi.org/10. 2-D 1H J-resolved NMR spectra of authentic
1371/journal.pgen.1000282 metabolite standards (BML-NMR). Metabo-
15. Sekula P, Goek ON, Quaye L, Barrios C, lomics 8(1):8–18. https://doi.org/10.1007/
Levey AS, Romisch-Margl W, Menni C, s11306-011-0347-7
Yet I, Gieger C, Inker LA, Adamski J, 24. Wishart DS (2008) Quantitative metabolo-
Gronwald W, Illig T, Dettmer K, mics using NMR. TrAC Trends Anal Chem
Krumsiek J, Oefner PJ, Valdes AM, 27(3):228–237. https://doi.org/10.1016/j.
Meisinger C, Coresh J, Spector TD, Mohney trac.2007.12.001
RP, Suhre K, Kastenmuller G, Kottgen A 25. Wishart DS, Tzur D, Knox C, Eisner R, Guo
(2016) A metabolome-wide association AC, Young N, Cheng D, Jewell K, Arndt D,
study of kidney function and disease in the Sawhney S, Fung C, Nikolai L, Lewis M,
general population. J Am Soc Nephrol 27 Coutouly MA, Forsythe I, Tang P,
(4):1175–1188. https://doi.org/10.1681/ Shrivastava S, Jeroncic K, Stothard P,
ASN.2014111099 Amegbey G, Block D, Hau DD, Wagner J,
16. Kraus WE, Muoio DM, Stevens R, Craig D, Miniaci J, Clements M, Gebremedhin M,
Bain JR, Grass E, Haynes C, Kwee L, Qin X, Guo N, Zhang Y, Duggan GE, Macinnis
Slentz DH, Krupp D, Muehlbauer M, Hauser GD, Weljie AM, Dowlatabadi R,
ER, Gregory SG, Newgard CB, Shah SH Bamforth F, Clive D, Greiner R, Li L,
(2015) Metabolomic quantitative trait loci Marrie T, Sykes BD, Vogel HJ, Querengesser
(mQTL) mapping implicates the ubiquitin L (2007) HMDB: the human metabolome
proteasome system in cardiovascular disease database. Nucleic Acids Res 35(Database
pathogenesis. PLoS Genet 11(11): issue):D521–D526. https://doi.org/10.
e1005553. https://doi.org/10.1371/jour 1093/nar/gkl923
nal.pgen.1005553 26. Kuhn S, Schlorer NE (2015) Facilitating qual-
17. MRC-NIHR National Phenome Center. ity control for spectra assignments of small
https://www.imperial.ac.uk/phenome-cen organic molecules: nmrshiftdb2—a free
tre. Accessed February 2019 in-house NMR database with integrated
18. Clinical Phenotyping Centre. http://www. LIMS for academic service laboratories.
imperial.ac.uk/clinical-phenotyping-centre/. Magn Reson Chem 53(8):582–589. https://
Accessed February 2019 doi.org/10.1002/mrc.4263
92 Wimal Pathmasiri et al.

27. Laine JE, Bailey KA, Olshan AF, Smeester L, 35. Bornet A, Maucourt M, Deborde C, Jacob D,
Drobna Z, Styblo M, Douillet C, Garcia- Milani J, Vuichoud B, Ji X, Dumez JN,
Vargas G, Rubio-Andrade M, Pathmasiri W, Moing A, Bodenhausen G, Jannin S, Girau-
McRitchie S, Sumner SJ, Fry RC (2017) Neo- deau P (2016) Highly repeatable dissolution
natal metabolomic profiles related to prenatal dynamic nuclear polarization for heteronuc-
arsenic exposure. Environ Sci Technol 51 lear NMR metabolomics. Anal Chem 88
(1):625–633. https://doi.org/10.1021/acs. (12):6179–6183. https://doi.org/10.1021/
est.6b04374 acs.analchem.6b01094
28. Szabo DT, Pathmasiri W, Sumner S, Birn- 36. Dumez JN, Milani J, Vuichoud B, Bornet A,
baum LS (2017) Serum metabolomic profiles Lalande-Martin J, Tea I, Yon M,
in neonatal mice following oral brominated Maucourt M, Deborde C, Moing A,
flame retardant exposures to hexabromocy- Frydman L, Bodenhausen G, Jannin S, Girau-
clododecane (HBCD) alpha, gamma, and deau P (2015) Hyperpolarized NMR of plant
commercial mixture. Environ Health Perspect and cancer cell extracts at natural abundance.
125(4):651–659. https://doi.org/10.1289/ Analyst 140(17):5860–5863. https://doi.
EHP242 org/10.1039/c5an01203a
29. Fan TW, Lane AN (2011) NMR-based stable 37. Johnson CH, Patterson AD, Idle JR, Gonza-
isotope resolved metabolomics in systems bio- lez FJ (2012) Xenobiotic metabolomics:
chemistry. J Biomol NMR 49(3–4):267–280. major impact on the metabolome. Annu Rev
https://doi.org/10.1007/s10858-011- Pharmacol Toxicol 52:37–56. https://doi.
9484-6 org/10.1146/annurev-pharmtox-010611-
30. Creek DJ, Chokkathukalam A, Jankevics A, 134748
Burgess KE, Breitling R, Barrett MP (2012) 38. Blaise BJ, Correia G, Tin A, Young JH, Verg-
Stable isotope-assisted metabolomics for naud AC, Lewis M, Pearce JT, Elliott P,
network-wide metabolic pathway elucidation. Nicholson JK, Holmes E, Ebbels TM (2016)
Anal Chem 84(20):8442–8447. https://doi. Power analysis and sample size determination
org/10.1021/ac3018795 in metabolic phenotyping. Anal Chem 88
31. Zamboni N, Fendt S-M, Rühl M, Sauer U (10):5179–5188. https://doi.org/10.1021/
(2009) 13C-based metabolic flux analysis. acs.analchem.6b00188
Nat Protoc 4(6):878–892. https://doi.org/ 39. Barton RH, Waterman D, Bonner FW,
10.1038/nprot.2009.58 Holmes E, Clarke R, Procardis C, Nicholson
32. Beckonert O, Keun HC, Ebbels TMD, JK, Lindon JC (2010) The influence of EDTA
Bundy J, Holmes E, Lindon JC, Nicholson and citrate anticoagulant addition to human
JK (2007) Metabolic profiling, metabolomic plasma on information recovery from
and metabonomic procedures for NMR spec- NMR-based metabolic profiling studies. Mol
troscopy of urine, plasma, serum and tissue BioSyst 6(1):215–224. https://doi.org/10.
extracts. Nat Protoc 2(11):2692–2703. 1039/b907021d
https://doi.org/10.1038/nprot.2007.376 40. Bernini P, Bertini I, Luchinat C, Nincheri P,
33. Dumas M-E, Maibaum EC, Teague C, Staderini S, Turano P (2011) Standard
Ueshima H, Zhou B, Lindon JC, Nicholson operating procedures for pre-analytical
JK, Stamler J, Elliott P, Queenie HE (2006) handling of blood and urine for metabolomic
Assessment of analytical reproducibility of 1H studies and biobanks. J Biomol NMR 49
NMR spectroscopy based metabonomics for (3–4):231–243. https://doi.org/10.1007/
large-scale epidemiological research: the s10858-011-9489-1
INTERMAP study. Anal Chem 41. Haid M, Muschet C, Wahl S, Romisch-Margl-
78:2199–2208 W, Prehn C, Moller G, Adamski J (2018)
34. Karaman I, Ferreira DL, Boulange CL, Long-term stability of human plasma metabo-
Kaluarachchi MR, Herrington D, Dona AC, lites during storage at
80 degrees C. J Pro-
Castagne R, Moayyeri A, Lehne B, Loh M, de teome Res 17(1):203–211. https://doi.org/
Vries PS, Dehghan A, Franco OH, 10.1021/acs.jproteome.7b00518
Hofman A, Evangelou E, Tzoulaki I, 42. Dane AD, Hendriks MM, Reijmers TH,
Elliott P, Lindon JC, Ebbels TM (2016) Harms AC, Troost J, Vreeken RJ, Boomsma
Workflow for integrated processing of multi- DI, van Duijn CM, Slagboom EP, Hankeme-
cohort untargeted (1)H NMR metabolomics ier T (2014) Integrating metabolomics
data in large-scale metabolic epidemiology. J profiling measurements across multiple bio-
Proteome Res 15(12):4188–4194. https:// banks. Anal Chem 86(9):4110–4114.
doi.org/10.1021/acs.jproteome.6b00125 https://doi.org/10.1021/ac404191a
 Analysis of NMR Metabolomics Data 93

43. Dona AC, Jimenez B, Schafer H, Humpfer E, Bamforth F, Greiner R, McManus B, New-
Spraul M, Lewis MR, Pearce JT, Holmes E, man JW, Goodfriend T, Wishart DS (2011)
Lindon JC, Nicholson JK (2014) Precision The human serum metabolome. PLoS One 6
high-throughput proton NMR spectroscopy (2):e16957. https://doi.org/10.1371/jour
of human urine, serum, and plasma for large- nal.pone.0016957
scale metabolic phenotyping. Anal Chem 86 51. Smilowitz JT, O’Sullivan A, Barile D, German
(19):9887–9894. https://doi.org/10.1021/ JB, Lonnerdal B, Slupsky CM (2013) The
ac5025039 human milk metabolome reveals diverse oli-
44. Beckonert O, Coen M, Keun HC, Wang Y, gosaccharide profiles. J Nutr 143
Ebbels TMD, Holmes E, Lindon JC, Nichol- (11):1709–1718. https://doi.org/10.3945/
son JK (2010) High-resolution magic-angle- jn.113.178772
spinning NMR spectroscopy for metabolic 52. Rodriguez-Martinez A, Posma JM, Ayala R,
profiling of intact tissues. Nat Protoc 5 Harvey N, Jimenez B, Neves AL, Lindon JC,
(6):1019–1032. https://doi.org/10.1038/ Sonomura K, Sato TA, Matsuda F, Zalloua P,
nprot.2010.45 Gauguier D, Nicholson JK, Dumas ME
45. Wong A, Jimenez B, Li X, Holmes E, Nichol- (2017) J-resolved (1)H NMR
son JK, Lindon JC, Sakellariou D (2012) 1D-projections for large-scale metabolic phe-
Evaluation of high resolution magic-angle notyping studies: application to blood plasma
coil spinning NMR spectroscopy for meta- analysis. Anal Chem 89(21):11405–11412.
bolic profiling of nanoliter tissue biopsies. https://doi.org/10.1021/acs.analchem.
Anal Chem 84(8):3843–3848. https://doi. 7b02374
org/10.1021/ac300153k 53. Fonville JM, Maher AD, Coen M, Holmes E,
46. Gillies RJ, Morse DL (2005) In vivo magnetic Lindon oC, Nicholson JK (2010) Evaluation
resonance spectroscopy in cancer. Annu Rev of full-resolution J-resolved 1H NMR projec-
Biomed Eng 7:287–326. https://doi.org/ tions of biofluids for metabonomics informa-
10.1146/annurev.bioeng.7.060804.100411 tion retrieval and biomarker identification.
47. Stewart DA, Winnike JH, McRitchie SL, Anal Chem 82:1811–1821
Clark RF, Pathmasiri WW, Sumner SJ (2016) 54. Liu M, Tang H, Nicholson JK, Lindon JC
Metabolomics analysis of hormone- (2002) Use of1H NMR-determined diffusion
responsive and triple-negative breast cancer coefficients to characterize lipoprotein frac-
cell responses to paclitaxel identify key meta- tions in human blood plasma. Magn Reson
bolic differences. J Proteome Res 15 Chem 40(13):S83–S88. https://doi.org/10.
(9):3225–3240. https://doi.org/10.1021/ 1002/mrc.1121
acs.jproteome.6b00430 55. Chylla RA, Hu K, Ellinger JJ, Markley JL
48. Livanos AE, Greiner TU, Vangay P, (2011) Deconvolution of two-dimensional
Pathmasiri W, Stewart D, McRitchie S, Li H, NMR spectra by fast maximum likelihood
Chung J, Sohn J, Kim S, Gao Z, Barber C, reconstruction: application to quantitative
Kim J, Ng S, Rogers AB, Sumner S, Zhang metabolomics. Anal Chem 83
XS, Cadwell K, Knights D, Alekseyenko A, (12):4871–4880. https://doi.org/10.1021/
Backhed F, Blaser MJ (2016) Antibiotic- ac200536b
mediated gut microbiome perturbation accel- 56. Phinney KW, Ballihaut G, Bedner M, Benford
erates development of type 1 diabetes in mice. BS, Camara JE, Christopher SJ, Davis WC,
Nat Microbiol 1(11):16140. https://doi. Dodder NG, Eppe G, Lang BE, Long SE,
org/10.1038/nmicrobiol.2016.140 Lowenthal MS, McGaw EA, Murphy KE,
49. Loeser RF, Pathmasiri W, Sumner SJ, Nelson BC, Prendergast JL, Reiner JL, Rim-
McRitchie S, Beavers D, Saxena P, Nicklas mer CA, Sander LC, Schantz MM, Sharpless
BJ, Jordan J, Guermazi A, Hunter DJ, Mess- KE, Sniegoski LT, Tai SS, Thomas JB, Vetter
ier SP (2016) Association of urinary metabo- TW, Welch MJ, Wise SA, Wood LJ, Guthrie
lites with radiographic progression of knee WF, Hagwood CR, Leigh SD, Yen JH, Zhang
osteoarthritis in overweight and obese adults: NF, Chaudhary-Webb M, Chen H, Fazili Z,
an exploratory study. Osteoarthr Cartil 24 LaVoie DJ, McCoy LF, Momin SS,
(8):1479–1486. https://doi.org/10.1016/j. Paladugula N, Pendergrast EC, Pfeiffer CM,
joca.2016.03.011 Powers CD, Rabinowitz D, Rybak ME,
50. Psychogios N, Hau DD, Peng J, Guo AC, Schleicher RL, Toombs BM, Xu M,
Mandal R, Bouatra S, Sinelnikov I, Zhang M, Castle AL (2013) Development of
Krishnamurthy R, Eisner R, Gautam B, a Standard Reference Material for metabolo-
Young N, Xia J, Knox C, Dong E, Huang P, mics research. Anal Chem 85
Hollander Z, Pedersen TL, Smith SR,
94 Wimal Pathmasiri et al.

(24):11732–11738. https://doi.org/10. 20(5):198–208. https://doi.org/10.1002/

1021/ac402689t cem.991
57. Dunn WB, Broadhurst D, Begley P, Zelena E, 66. Alonso A, Rodriguez MA, Vinaixa M,
Francis-McIntyre S, Anderson N, Brown M, Tortosa R, Correig X, Julia A, Marsal S
Knowles JD, Halsall A, Haselden JN, Nicholls (2014) Focus: a robust workflow for
AW, Wilson ID, Kell DB, Goodacre R, one-dimensional NMR spectral analysis. Anal
Human Serum Metabolome C (2011) Proce- Chem 86(2):1160–1169. https://doi.org/
dures for large-scale metabolic profiling of 10.1021/ac403110u
serum and plasma using gas chromatography 67. RBNMR. https://www.mathworks.com/
and liquid chromatography coupled to mass matlabcentral/fileexchange/40332-rbnmr.
spectrometry. Nat Protoc 6(7):1060–1083. Accessed February 2019
https://doi.org/10.1038/nprot.2011.335 68. Krishnamurthy K (2013) CRAFT (complete
58. Gika HG, A G, Theodoridis EM, Wilson ID reduction to amplitude frequency table)—
(2012) A QC approach to the determination robust and time-efficient Bayesian approach
of day-to-day reproducibility and robustness for quantitative mixture analysis by NMR.
of LC–MS methods for global metabolite Magn Reson Chem 51(12):821–829.
profiling in metabonomics/metabolomics. https://doi.org/10.1002/mrc.4022
Bioanalysis 4(18):2239–2247 69. Intellegent bucketing: Part 1. https://www.
59. Townsend MK, Clish CB, Kraft P, Wu C, acdlabs.com/download/publ/2004/
Souza AL, Deik AA, Tworoger SS, Wolpin enc04/intelbucket.pdf. Accessed February
BM (2013) Reproducibility of metabolomic 2019
profiles among men and women in 2 large 70. Intellegent bucketing: Part 2. https://www.
cohort studies. Clin Chem 59 acdlabs.com/download/publ/2004/
(11):1657–1667. https://doi.org/10.1373/ intelbucket2.pdf. Accessed February 2019
clinchem.2012.199133
71. Davis RA, Charlton AJ, Godward J, Jones SA,
60. Masson P, Spagou K, Nicholson JK, Want EJ Harrison M, Wilson JC (2007) Adaptive bin-
(2011) Technical and biological variation in ning: an improved binning method for meta-
UPLC-MS-based untargeted metabolic bolomics data using the undecimated wavelet
profiling of liver extracts: application in an transform. Chemom Intell Lab Syst 85
experimental toxicity study on galactosamine. (1):144–154. https://doi.org/10.1016/j.
Anal Chem 83(3):1116–1123. https://doi. chemolab.2006.08.014
org/10.1021/ac103011b
72. Sousa SAA, Magalhães A, Ferreira MMC
61. Chan EC, Pasikanti KK, Nicholson JK (2011) (2013) Optimized bucketing for NMR spec-
Global urinary metabolic profiling procedures tra: three case studies. Chemom Intell Lab
using gas chromatography-mass spectrome- Syst 122:93–102. https://doi.org/10.
try. Nat Protoc 6(10):1483–1499. https:// 1016/j.chemolab.2013.01.006
doi.org/10.1038/nprot.2011.375
73. Rodriguez-Martinez A, Ayala R, Posma JM,
62. Veselkov KA, Lindon JC, Ebbels TMD, Harvey N, Jimenez B, Sonomura K, Sato TA,
Crockford D, Volynkin VV, Holmes E, Davies Matsuda F, Zalloua P, Gauguier D, Nicholson
DB, Nicholson JK (2009) Recursive segment- JK, Dumas ME (2018) pJRES Binning Algo-
wise peak alignment of biological 1H NMR rithm (JBA): a new method to facilitate the
spectra for improved metabolic biomarker recovery of metabolic information from
recovery. Anal Chem 81:56–66 pJRES 1H NMR spectra. Bioinformatics.
63. Savorani F, Tomasi G, Engelsen SB (2010) https://doi.org/10.1093/bioinformatics/
icoshift: a versatile tool for the rapid align- bty837
ment of 1D NMR spectra. J Magn Reson 74. Rodriguez-Martinez A, Posma JM, Ayala R,
202(2):190–202. https://doi.org/10.1016/ Neves AL, Anwar M, Petretto E, Emanueli C,
j.jmr.2009.11.012 Gauguier D, Nicholson JK, Dumas ME
64. Vu TN, Valkenborg D, Smets K, Verwaest (2018) MWASTools: an R/bioconductor
KA, Dommisse R, Lemière F, Verschoren A, package for metabolome-wide association
Goethals B, Laukens K (2011) An integrated studies. Bioinformatics 34(5):890–892.
workflow for robust alignment and simplified https://doi.org/10.1093/bioinformatics/
quantitative analysis of NMR spectrometry btx477
data. BMC Bioinformatics 12:405 75. Dieterle F, Ross A, Schlotterbeck G, Senn H
65. Larsen FH, van den Berg F, Engelsen SB (2006) Probabilistic quotient normalization
(2006) An exploratory chemometric study as robust method to account for dilution of
of1H NMR spectra of table wines. J Chemom complex biological mixtures. Application in
 Analysis of NMR Metabolomics Data 95

1H NMR metabonomics. Anal Chem biomarker structural information from 1H

78:4281–4290 NMR spectra of biofluids. Anal Chem 84
76. van den Berg RA, Hoefsloot HC, Westerhuis (24):10694–10701. https://doi.org/10.
JA, Smilde AK, van der Werf MJ (2006) Cen- 1021/ac302360v
tering, scaling, and transformations: improv- 85. Blaise BJ, Shintu L, Bnd E, Emsley L, Dumas
ing the biological information content of M-E, Toulhoat P (2009) Statistical recou-
metabolomics data. BMC Genomics 7:142. pling prior to significance testing in nuclear
https://doi.org/10.1186/1471-2164-7- magnetic resonance based metabonomics.
142 Anal Chem 81:6242–6251
77. Keun HC, Ebbels TMD, Antti H, Bollard 86. Blaise BJ, Navratil V, Emsley L, Toulhoat P
ME, Beckonert O, Holmes E, Lindon JC, (2011) Orthogonal filtered recoupled-
Nicholson JK (2003) Improved analysis of STOCSY to extract metabolic networks asso-
multivariate data by variable stability scaling: ciated with minor perturbations from NMR
application to NMR-based metabolic spectroscopy. J Proteome Res 10
profiling. Anal Chim Acta 490 (9):4342–4348. https://doi.org/10.1021/
(1–2):265–276. https://doi.org/10.1016/ pr200489n
s0003-2670(03)00094-1 87. Zou X, Holmes E, Nicholson JK, Loo RL
78. Eriksson L, Byrne T, Johansson E, Trygg J, (2014) Statistical HOmogeneous Cluster
Vikström C (2013) Multi-and megavariate SpectroscopY (SHOCSY): an optimized sta-
data analysis basic principles and applications. tistical approach for clustering of (1)H NMR
Umetrics Academy, Umeå spectral data to reduce interference and
79. Johan T, Holmes E, Lundstedt T (2007) enhance robust biomarkers selection. Anal
Chemometrics in metabonomics. J Proteome Chem 86(11):5308–5315. https://doi.org/
Res 6:469–479 10.1021/ac500161k
80. Bylesjö M, Rantalainen M, Cloarec O, 88. Dona AC, Kyriakides M, Scott F, Shephard
Nicholson JK, Holmes E, Trygg J (2006) EA, Varshavi D, Veselkov K, Everett JR
OPLS discriminant analysis: combining the (2016) A guide to the identification of meta-
strengths of PLS-DA and SIMCA classifica- bolites in NMR-based metabonomics/meta-
tion. J Chemom 20(8–10):341–351. bolomics experiments. Comput Struct
https://doi.org/10.1002/cem.1006 Biotechnol J 14:135–153. https://doi.org/
81. Bylesjo M, Rantalainen M, Nicholson JK, 10.1016/j.csbj.2016.02.005
Holmes E, Trygg J (2008) K-OPLS package: 89. Bingol K (2018) Recent advances in targeted
kernel-based orthogonal projections to latent and untargeted metabolomics by NMR and
structures for prediction and interpretation in MS/NMR methods. High Throughput 7(2).
feature space. BMC Bioinformatics 9:106. https://doi.org/10.3390/ht7020009
https://doi.org/10.1186/1471-2105-9- 90. Bingol K, Bruschweiler R (2017) Knowns and
106 unknowns in metabolomics identified by mul-
82. Posma JM, Garcia-Perez I, Ebbels TMD, Lin- tidimensional NMR and hybrid MS/NMR
don JC, Stamler J, Elliott P, Holmes E, methods. Curr Opin Biotechnol 43:17–24.
Nicholson JK (2018) Optimized phenotypic https://doi.org/10.1016/j.copbio.2016.07.
biomarker discovery and confounder elimina- 006
tion via covariate-adjusted projection to latent 91. Robinette SL, Zhang F, Brüschweiler-Li L,
structures from metabolic spectroscopy data. Brüschweiler R (2008) R web server based
J Proteome Res 17(4):1586–1595. https:// complex mixture analysis by NMR. Anal
doi.org/10.1021/acs.jproteome.7b00879 Chem 80:3606–3611
83. Cloarec O, Dumas ME, Craig A, Barton RH, 92. Bingol K, Zhang F, Bruschweiler-Li L,
Trygg J, Hudson J, Blancher C, Gauguier D, Bruschweiler R (2012) TOCCATA: a custo-
Lindon JC, Holmes E, Nicholson J (2005) mized carbon total correlation spectroscopy
Statistical total correlation spectroscopy: an NMR metabolomics database. Anal Chem
exploratory approach for latent biomarker 84(21):9395–9401. https://doi.org/10.
identification from metabolic 1H NMR data 1021/ac302197e
sets. Anal Chem 77(5):1282–1289. https:// 93. Wei S, Zhang J, Liu L, Ye T, Gowda GA,
doi.org/10.1021/ac048630x Tayyari F, Raftery D (2011) Ratio analysis
84. Posma JM, Garcia-Perez I, De Iorio M, Lin- nuclear magnetic resonance spectroscopy for
don JC, Elliott P, Holmes E, Ebbels TM, selective metabolite identification in complex
Nicholson JK (2012) Subset optimization by samples. Anal Chem 83(20):7616–7623.
reference matching (STORM): an optimized https://doi.org/10.1021/ac201625f
statistical approach for recovery of metabolic
96 Wimal Pathmasiri et al.

94. Ye T, Mo H, Shanaiah N, Nagana Gowda GA, Magn Reson Chem 47(Suppl 1):S123–S126.
Zhang S, Raftery D (2009) Chemoselective https://doi.org/10.1002/mrc.2526
15N tag for sensitive and high-resolution 104. Wang T, Shao K, Chu Q, Ren Y, Mu Y, Qu L,
nuclear magnetic resonance profiling of the He J, Jin C, Xia B (2009) Automics: an
carboxyl-containing metabolome. Anal integrated platform for NMR-based metabo-
Chem 81:4882–4888 nomics spectral processing and data analysis.
95. Tayyari F, Gowda GA, Gu H, Raftery D BMC Bioinformatics 10:83. https://doi.org/
(2013) 15N-cholamine—a smart isotope tag 10.1186/1471-2105-10-83
for combining NMR- and MS-based metabo- 105. Verhoeven A, Giera M, Mayboroda OA
lite profiling. Anal Chem 85(18):8715–8721. (2018) KIMBLE: a versatile visual NMR
https://doi.org/10.1021/ac401712a metabolomics workbench in KNIME. Anal
96. Clendinen CS, Pasquel C, Ajredini R, Edison Chim Acta 1044:66–76
AS (2015) (13)C NMR metabolomics: 106. Chong J, Soufan O, Li C, Caraus I, Li S,
INADEQUATE network analysis. Anal Bourque G, Wishart DS, Xia J (2018) Meta-
Chem 87(11):5698–5706. https://doi.org/ boAnalyst 4.0: towards more transparent and
10.1021/acs.analchem.5b00867 integrative metabolomics analysis. Nucleic
97. Weljie AM, Newton J, Mercier P, Carlson E, Acids Res 46(W1):W486–W494. https://
Slupsky CM (2006) Targeted profiling: quan- doi.org/10.1093/nar/gky310
titative analysis of 1H NMR metabolomics 107. Xia J, Psychogios N, Young N, Wishart DS
data. Anal Chem 78(13):4430–4442. (2009) MetaboAnalyst: a web server for
https://doi.org/10.1021/ac060209g metabolomic data analysis and interpretation.
98. Rohnisch HE, Eriksson J, Mullner E, Nucleic Acids Res 37(Web Server issue):
Agback P, Sandstrom C, Moazzami AA W652–W660. https://doi.org/10.1093/
(2018) AQuA: an automated quantification nar/gkp356
algorithm for high-throughput NMR-based 108. Xia J, Wishart DS (2011) Web-based infer-
metabolomics and its application in human ence of biological patterns, functions and
plasma. Anal Chem 90(3):2095–2102. pathways from metabolomic data using Meta-
https://doi.org/10.1021/acs.analchem. boAnalyst. Nat Protoc 6(6):743–760.
7b04324 https://doi.org/10.1038/nprot.2011.319
99. Hao J, Astle W, De Iorio M, Ebbels TM 109. Metaboanalyst. https://www.metaboanalyst.
(2012) BATMAN—an R package for the ca/MetaboAnalyst/faces/home.xhtml.
automated quantification of metabolites Accessed February 2019
from nuclear magnetic resonance spectra 110. Gaude E, Chignola F, Spiliotopoulos D,
using a Bayesian model. Bioinformatics 28 Spitaleri A, Ghitti M, Garcia-Manteiga M,
(15):2088–2090. https://doi.org/10.1093/ Mari S, Musco G (2013) muma, An R pack-
bioinformatics/bts308 age for metabolomics univariate and multivar-
100. Hao J, Liebeke M, Astle W, De Iorio M, iate statistical analysis. Curr Metabolomics 1
Bundy JG, Ebbels TM (2014) Bayesian (2):180–189. https://doi.org/10.2174/
deconvolution and quantification of metabo- 2213235x11301020005
lites in complex 1D NMR spectra using BAT- 111. Worley B, Powers R (2014) MVAPACK: a
MAN. Nat Protoc 9(6):1416–1427. https:// complete data handling package for NMR
doi.org/10.1038/nprot.2014.090 metabolomics. ACS Chem Biol 9
101. Liebeke M, Hao J, Ebbels TM, Bundy JG (5):1138–1144. https://doi.org/10.1021/
(2013) Combining spectral ordering with cb4008937
peak fitting for one-dimensional NMR quan- 112. Jacob D, Deborde C, Lefebvre M,
titative metabolomics. Anal Chem 85 Maucourt M, Moing A (2017) NMRProc-
(9):4605–4612. https://doi.org/10.1021/ Flow: a graphical and interactive tool dedi-
ac400237w cated to 1D spectra processing for
102. Ravanbakhsh S, Liu P, Bjorndahl TC, NMR-based metabolomics. Metabolomics
Mandal R, Grant JR, Wilson M, Eisner R, 13(4):36. https://doi.org/10.1007/
Sinelnikov I, Hu X, Luchinat C, Greiner R, s11306-017-1178-y
Wishart DS (2015) Accurate, fully-automated 113. Gunaratna K, Anderson P, Ranabahu A, Sheth
NMR spectral profiling for metabolomics. A (2010) A study in hadoop streaming with
PLoS One 10(5):e0124219. https://doi. matlab for NMR data processing. Paper pre-
org/10.1371/journal.pone.0124219 sented at the 2010 IEEE second international
103. Lewis IA, Schommer SC, Markley JL (2009) conference on cloud computing technology
rNMR: open source software for identifying and science.
and quantifying metabolites in NMR spectra.
 Analysis of NMR Metabolomics Data 97

114. Fitzpatrick MA, McGrath CM, Young SP MR, Spagou K, Dona AC, Evangelos V,
(2014) Pathomx: an interactive workflow- Tracy R, Greenland P, Lindon JC,
based tool for the analysis of metabolomic Herrington D, Ebbels TMD, Elliott P,
data. BMC Bioinformatics 15(1):396 Tzoulaki I, Chadeau-Hyam M (2017)
115. Beirnaert C, Meysman P, Vu TN, Improving visualization and interpretation of
Hermans N, Apers S, Pieters L, Covaci A, metabolome-wide association studies: an
Laukens K (2018) speaq 2.0: a complete application in a population-based cohort
workflow for high-throughput 1D NMR using untargeted (1)H NMR metabolic
spectra processing and quantification. PLoS profiling. J Proteome Res 16
Comput Biol 14(3):e1006018. https://doi. (10):3623–3633. https://doi.org/10.1021/
org/10.1371/journal.pcbi.1006018 acs.jproteome.7b00344
116. Giacomoni F, Le Corguille G, Monsoor M, 120. Karnovsky A, Weymouth T, Hull T, Tarcea
Landi M, Pericard P, Petera M, Duperier C, VG, Scardoni G, Laudanna C, Sartor MA,
Tremblay-Franco M, Martin JF, Jacob D, Stringer KA, Jagadish HV, Burant C,
Goulitquer S, Thevenot EA, Caron C Athey B, Omenn GS (2012) Metscape 2 bio-
(2015) Workflow4Metabolomics: a collabo- informatics tool for the analysis and visualiza-
rative research infrastructure for computa- tion of metabolomics and gene expression
tional metabolomics. Bioinformatics 31 data. Bioinformatics 28(3):373–380.
(9):1493–1495. https://doi.org/10.1093/ https://doi.org/10.1093/bioinformatics/
bioinformatics/btu813 btr661
117. Lefort G, Liaubet L, Canlet C, Tardivel P, 121. Kamburov A, Cavill R, Ebbels TMD,
Pere MC, Quesnel H, Paris A, Iannuccelli N, Herwig R, Keun HC (2011) Integrated
Vialaneix N, Servien R (2019) ASICS: an R pathway-level analysis of transcriptomics and
package for a whole analysis workflow of 1D metabolomics data with IMPaLA. Bioinfor-
1H NMR spectra. Bioinformatics. https:// matics 27(20):2917–2918. https://doi.org/
doi.org/10.1093/bioinformatics/btz248 10.1093/bioinformatics/btr499
118. Chadeau-Hyam M, Ebbels TMD, Brown IJ, 122. Haug K, Salek RM, Conesa P, Hastings J, de
Chan Q, Stamler J, Huang CC, Daviglus ML, Matos P, Rijnbeek M, Mahendraker T,
Ueshima H, Zhao L, Holmes E, Nicholson Williams M, Neumann S, Rocca-Serra P,
JK, Elliott P, Iorio MD (2010) Metabolic Maguire E, Gonzalez-Beltran A, Sansone
profiling and the metabolome-wide associa- SA, Griffin JL, Steinbeck C (2013) Metabo-
tion study: significance level for biomarker Lights—an open-access general-purpose
identification. J Proteome Res 9:4620–4627 repository for metabolomics studies and asso-
119. Castagne R, Boulange CL, Karaman I, ciated meta-data. Nucleic Acids Res 41(Data-
Campanella G, Santos Ferreira DL, Kaluar- base issue):D781–D786. https://doi.org/
achchi MR, Lehne B, Moayyeri A, Lewis 10.1093/nar/gks1004
 Chapter 6

Key Concepts Surrounding Studies of Stable

Isotope-Resolved Metabolomics
Stephen F. Previs and Daniel P. Downes

Abstract
“Omics”-based analyses are widely used in numerous areas of research, advances in instrumentation (both
hardware and software) allow investigators to collect a wealth of data and therein characterize metabolic
systems. Although analyses generally examine differences in absolute or relative (fold-) changes in concen-
trations, the ability to extract mechanistic insight would benefit from the use of isotopic tracers. Herein, we
discuss important concepts that should be considered when stable isotope tracers are used to capture
biochemical flux. Special attention is placed on in vivo systems, however, many of the general ideas have
immediate impact on studies in cellular models or isolated-perfused tissues. While it is somewhat trivial to
administer labeled precursor molecules and measure the enrichment of downstream products, the ability to
make correct interpretations can be challenging. We will outline several critical factors that may influence
choices when developing and/or applying a stable isotope tracer method. For example, is there a “best”
tracer for a given study? How do I administer a tracer? When do I collect my sample(s)? While these
questions may seem straightforward, we will present scenarios that can have dramatic effects on conclusions
surrounding apparent rates of metabolic activity.

Key words Isotope, Metabolomics, Flux, Pathway, Metabolic activity

1 Introduction: Tracer Studies Can Yield an Unambiguous View of Pathway Activity

The levels of a given metabolite are often a critical factor for

differentiating healthy and disease states. For example, the concen-
tration of circulating glucose in a fasted subject serves as a key
diagnostic of diabetes. In other cases, intracellular metabolites
(e.g., acyl-CoA, diacylglycerol or ceramide concentrations) could
play a pivotal role in driving a disease phenotype [1–3]. Therefore,
one could argue that attention should be focused on developing
methods to readily quantify the concentration of any/all analytes.
While this is certainly a valid perspective, the ability to measure
pathway activity should also be closely positioned, especially since a
biochemical flux rate(s) must change before a pool can change
[4]. In fact, measurements of flux become more critical when one

Shuzhao Li (ed.), Computational Methods and Data Analysis for Metabolomics, Methods in Molecular Biology, vol. 2104,
https://doi.org/10.1007/978-1-0716-0239-3_6, © Springer Science+Business Media, LLC, part of Springer Nature 2020

99
100 Stephen F. Previs and Daniel P. Downes

Fig. 1 Metabolic flux can change independent of pool size. Sprague-Dawley rats (n ¼ 15, 253  12 g,
mean  sem) were fed a standard rodent diet on a 12-h light–dark cycle (dark between 6 PM and 6 AM). An
intraperitoneal bolus of 2H-water (15 μL 99% 2H-water per gram body weight) was given to a subgroup of
5 rats at 6 AM and samples of plasma and mixed leg muscle were collected at 10 AM, a second group was
then given a bolus of 2H-water and samples collected at 2 PM, a final group was given a bolus of 2H-water at
2 PM and samples were collected at 6 PM. Protein synthesis was estimated from measurements of the 2H-
labeling of plasma water and alanine (isolated from total muscle protein, n ¼ 5 per group, Eq. 5) [7]. Although
muscle mass did not change, different metabolic flux (protein synthesis) was observed at different times after
the feeding cycle ended (Asterisk represents p < 0.05 using a 2-tailed t-test, assuming equal variance; this
study was completed using an IACUC approved protocol)

aims to examine large pools that have seemingly low metabolic

activity and/or when more than one pathway may contribute to
the appearance of a product [5, 6].
We consider the following example to emphasize the novelty of
studying pathway flux in a case where there is a large pool with a
relatively low metabolic activity (Fig. 1). Sprague-Dawley rats
weighing ~250 g were maintained on a 12-h feeding schedule
(6 PM to 6 AM), following the end of a feeding cycle animals
were subdivided into thee groups. The first group was given an
intraperitoneal bolus of 2H-water at 6 AM and samples of mixed
skeletal muscle were collected at 10 AM, at that time a second
group was given a bolus of 2H-water and samples were collected
at 2 PM, the final group was given a bolus of 2H-water at 2 PM and
the samples were collected at 6 PM. Although changes in muscle
mass could not be observed over these intervals, the incorporation
of 2H-alanine into total mixed muscle protein was dramatically
different (Fig. 1b). The inability to detect differences in the pool
size reflect the fact that the overall traffic flow (i.e., protein synthe-
sis) only represents a small fraction of the total amount of protein,
that is, less than 1% is made per hour (Fig. 1c); however, the tracer
analyses clearly reveal differences in biochemical flux at different
periods after the feeding. The literature contains the necessary
details for readers who may be interested in using 2H-water to
study protein synthesis [7–12].
The fact that tracers can detect changes in metabolic activity
before one can discern any change in concentration or pool size is
novel but should be interpreted with some degree of caution too
[13]. For example, the study outlined above only considered the
 Key Concepts Surrounding Studies of Stable Isotope-Resolved Metabolomics 101

input of amino acids, we do not have a measure of protein degra-

dation. Consequently, it is not possible to draw definitive conclu-
sions regarding true long-term outcomes surrounding muscle
protein homeostasis if we only examine one side of the biology. In
cases where one can observe a change in the pool size while using
tracers to probe the synthesis of a product, it should be possible to
calculate the degradation (i.e., change in pool size ¼ synthesis  deg-
radation) [5, 14].
A second area where tracers can make an impact concerns
examples where a product may be derived from multiple sources.
Namely, although increased glucose production has been recog-
nized as an important factor in diabetes, the contribution of specific
organs (i.e., liver, kidney, and intestine) has been debated for many
years [15–17]. Likewise, the importance of targeting glycogenoly-
sis vs gluconeogenesis has generated considerable discussion [18–
20]. Fortunately, tracers can be used to gain novel insight regarding
integrative, whole-body metabolic biochemistry, including the
pathophysiology surrounding disease states. We will further
develop these types of scenarios by using triglyceride synthesis as a
model problem. This example will draw out lesser appreciated
factors that one must contend with when developing and/or apply-
ing a stable isotope tracer protocol.

2 Strategies to Enable Study Designs

We will consider several key questions related to the experimental

design. In our opinion it is not possible to give specific protocol
recommendations since conditions and models will vary, however,
there are a few critical decision points that influence virtually all
tracer studies. Our goal is to outline these steps using practical
examples that underscore the impact of various choices.

2.1 What Is the This question is obviously broad, so we will try to provide an
“Best” Tracer? answer in the context of studying lipid flux, more specifically tri-
glyceride synthesis. Perhaps a first line of thinking is to use a labeled
fatty acid, one is then immediately faced with subsequent questions
around which fatty acid and how to formulate the tracer? When
thinking about “which fatty acid” one may consider 13C- vs 2H-
labeled forms, as this could certainly be important from an analyti-
cal perspective, but the question is more complicated in the eyes of
a biologist since palmitate and oleate may not be handled in the
same way [21, 22]. Therefore, choosing one fatty acid over another
could lead to different outcomes in the flux calculations under
conditions where fatty acids are not utilized in proportion to their
availability [6, 21, 22].
A less biased tracer might be labeled glycerol or glucose, these
precursors are expected to provide the backbone on which fatty
acids are esterified [23]. Again, the choice of how to label the
102 Stephen F. Previs and Daniel P. Downes

precursor and what product to measure could influence the data

interpretation. Patterson and colleagues clearly demonstrated the
formation of multiple glycerol species following the acute adminis-
tration of [2H5]glycerol. Although the injected glycerol tracer was
5 amu greater than that of the endogenous glycerol (M5 vs M0,
respectively), the M5 triglyceride-glycerol only represented ~30%
of the total labeled triglycerides that were synthesized [23]. In fact,
they observed the appearance of triglyceride-glycerol species with
1, 2, 3 and 4 2H (M1, M2, M3 and M4, respectively). While this
type of data is consistent with known biochemical pathways, the
implications of “isotope scrambling” needs to be considered when
one thinks about modeling the metabolic flux (discussed later). For
example, if we expected that since the precursor was M5 we only
need to measure products that contain 5 2H, could there be con-
sequences if we ignore the M1, M2, M3, and M4 mass
isotopomers?
Indeed, other tracer approaches could also be considered. For
example, one could administer a tracer that labels fatty acids during
de novo synthesis (i.e., 13C-acetate or 2H-water) or during (re)
esterification (i.e., 18O-water) [24]. Consequently, the ability to
define a truly “best” tracer can be difficult. Perhaps the most
important factor here has little to do with the tracer protocol per
se but rather the biological question that is to be addressed.
Finally, investigators should consider whether the tracer could
perturb the metabolic flux that is under investigation [25]. In cases
where 13C-glucose is used to probe lipid flux, one might hypothe-
size that there would be little metabolic alteration if we increased
the pool size ~10% (which is a reasonable expectation if the goal is
to reach ~10% enrichment in the glucose precursor). Presumably a
relatively minor perturbation like this would have minimal impact
on downstream processes. In contrast, if we chose to administer
13
C-acetate we would have to contend with the delivery of extra
Na+ (to avoid the acid load) and we would have to make assump-
tions about flux via a “new” pathway. For example, under most
conditions one expects negligible flux of acetate [25, 26]. Therefore,
even though 13C-acetate can be incorporated into fatty acids
[27, 28], one may question whether the physiological state approx-
imates true endogenous conditions.

2.2 How Do I To build on the examples that are outlined above we need to
Administer the Tracer? appreciate other challenges of the experimental design, namely,
the tracer needs to enter the system that is under investigation. If
we imagine in vivo studies, we are immediately presented with
several options and challenges. Although most of the tracers
noted earlier are water soluble, it is not easy to simply dissolve a
labeled fatty acid in saline prior to administration. Fatty acid tracers
are typically bound to albumin or solubilized with Intralipid [29],
when an intra-vascular route of administration is used, or they can
 Key Concepts Surrounding Studies of Stable Isotope-Resolved Metabolomics 103

Fig. 2 Theoretical labeling profiles for different routes of tracer administration in vivo. If we consider a simple
one compartment model (Panel a), we can expect distinct labeling profiles depending on how a tracer is
administered (Panel b). For simplicity, we have assumed that one would target ~10% enrichment of a
circulating pool. In the case where one contrasts the intravenous (IV) vs intraperitoneal (IP) or oral bolus
methods, we have assumed that the same dose of tracer is administered. For purposes of illustration, we
expect that the maximum labeling achieved with the IP or oral bolus will never reach that when the same dose
of tracer is given as an IV bolus; when using the IP or oral bolus the temporal rise to maximum enrichment is
impacted by the balance between absorption and metabolism kinetics. A constant infusion will be associated
with a rise to steady-state labeling, whereas a primed-infusion could lead to an “immediate” steady state if
the system if well described

be mixed with food and then administered (consumed) orally

[22, 30, 31].
The route of administration is associated with more caveats,
namely, if we obtain intravenous access should we use a bolus
injection or an infusion [32, 33]? Likewise, if we choose an oral
route of administration is it better to mix the tracer with food or
can we add isotopes to the drinking water [34]? Again, there is not
really a right or wrong answer, the approach will likely be driven by
the question. We should also note that many tracer studies use a
hybrid technique, that is, the primed-infusion (Fig. 2). Although
intuition may lead one to think that primed-infusions imply that we
are talking about intravenous studies, in fact, the use of labeled-
water also falls under this category; one can give an intraperitoneal
bolus followed by addition of the tracer to the drinking water,
therein achieving a “square-wave” profile of the precursor. The
ability to deliver tracers for long periods of time allows for studies
of analytes with slow rates of turnover [6, 11].

2.3 How Can I Get a An important factor to consider when deciding on the approach for
Metabolic Rate from administering a tracer is the type of kinetic data that one intends to
My Tracer Data? obtain. For example, studying the overall kinetics of a product may
require a different strategy as compared to dissecting the source
(s) of that product. We will consider the kinetics of triglyceride in
adipose tissue as a model problem, however, we will first review
some general concepts concerning a simpler system and then
104 Stephen F. Previs and Daniel P. Downes

introduce the key ideas surrounding the use of precursor ! product

labeling profiles in studies of more complicated problems.
Figure 2a contains an example of a single pool, production
(input, or rate of appearance) is usually thought to follow zero-
order kinetics and removal (outflow or rate of disappearance) is
generally assumed to be first-order [35]. Figure 2b demonstrates
four scenarios that can be encountered when tracers are used to
estimate a metabolic rate, for example, glucose turnover in the
plasma [35]. The plots represent the outcomes that one expects if
labeled glucose (e.g., [U-13C6]) is given as an intravenous bolus
(open circles), an oral or intraperitoneal bolus (solid circles), a
constant infusion (diamonds) or a primed-infusion (squares). We
will briefly review how each approach can be used to estimate the
kinetics of a given analyte.
First, if we assume there is rapid mixing of the tracer in a single
pool, then, following the administration of an intravenous bolus of
the tracer (open circles, Fig. 2b), we can determine the glucose flux
from the pool size and the fractional turnover. The former is
determined from the initial tracer dilution and the later from the
tracer dilution measured over time. Specifically, one can estimate
the pool size using the equation:
tracer dose
pool size ¼ ð1Þ
initial enrichment
where the units of tracer dose (mg, μmol, etc.) reflect the units that
define the pool size. The term “initial enrichment” can be esti-
mated by extrapolating the tracer dilution to the time ¼ 0 point or,
if the system mixes very rapidly, one may be able to simply measure
the enrichment shortly after injecting the tracer. If we consider the
problem of glucose kinetics, one typically assumes a single (well-
mixed) compartment and therefore the decrease in enrichment
(or tracer dilution) can be described using the equation:
enrichment ðt Þ ¼ initial enrichment  e kt ð2Þ
where “k” represents the fractional turnover. In this scenario there
are different approaches for determining the tracer dilution. When
several samples are collected one can fit the decrease in enrichment
using nonlinear regression, or, in cases where two samples are
collected, one can transform the y-axis to “ln scale” and determine
the slope using linear regression [35, 36].
Second, in cases where an intraperitoneal (or oral) bolus is used
one can estimate the kinetics by modeling the entire curve or by
using a limited region of the data (i.e., after the peak has been
reached). Again, if we assume that the analyte exists in a single, well-
mixed, pool we can determine the flux using the equation:
Key Concepts Surrounding Studies of Stable Isotope-Resolved Metabolomics 105

FD 0 ½glucosekabs kdil
rate of appearance ¼ ð3Þ
kabs c dil abs dil
0  c0 k

where F is the bioavailability of the tracer, D0 is the dose of the

tracer administered at time 0, and [glucose] is the mean blood
glucose pool throughout the experiment. The terms kabs and kdil
are the kinetic turnover rates for absorbance and tracer dilution
(commonly referred to as elimination), respectively, and c abs
0 and c 0
dil

are the back calculated concentrations of the tracer at time 0 for the
absorbance and dilution (elimination) curves, respectively
[33]. When tracers are given via an IP or oral bolus one can estimate
the metabolic rate using different approaches. For example, van
Dijk et al. proposed an elegant application of this logic in studies of
glucose kinetics in rodent models, in which they considered the
entire labeling profile [33]. We performed comparable studies but
only modeled the tail of the tracer dilution [32, 37].
When either a constant infusion (triangles) or a primed-
infusion (squares) is used, we can estimate the turnover from the
steady-state enrichment (i.e., the dilution rate of the tracer) accord-
ing to the equation:
rate of appearance ¼ tracer infusion rate
 
enrichment of the infusate
 1
enrichment of the analyte pool
ð4Þ
where the units that define the “rate of appearance” are the same as
those used to define the “tracer infusion rate.” We should note that
this approach is often used in studies of circulating analytes; there-
fore, the “enrichment of the analyte pool” represents the enrich-
ment of some endogenous molecule in the plasma. In theory, the
use of a primed-infusion allows one to reach a steady-state enrich-
ment of the analyte pool in a shorter amount of time as compared
to when a constant infusion is used. That said, there are examples
where achieving the correct balance between priming doses and
infusion rates can require some development, otherwise investiga-
tors run a risk of making misleading conclusions [35, 38, 39].
The types of scenarios that were just described have been widely
used to quantify the kinetics in cases where a pool turns over with
considerable frequency over the duration of an experiment (circu-
lating glucose is an excellent model problem). However, there are
other cases where a pool may experience very limited synthesis
(e.g., mixed muscle protein, Fig. 1). We will consider how tracers
can be used to estimate kinetics in these cases as well. This next
scenario emphasizes a key point, in that, we may not be able to
directly administer a labeled form of the product. Consequently,
one will typically rely on the use of a precursor-product labeling
106 Stephen F. Previs and Daniel P. Downes

ratio. The general rationale is that if a precursor can be maintained

at a stable enrichment then the enrichment of the product will
increase and reach a steady-state enrichment that is equal to that
of the precursor [13]; although this may seem straightforward
there are some important caveats to this logic which will be high-
lighted [13, 40].
In cases where a precursor-product labeling ratio is used to
quantify the kinetics one can estimate the fractional synthesis of a
product using the equation:
initial change in the enrichment of the product
fraction newly made product ¼
enrichment of the precursor  time
ð5Þ
This equation applies when the precursor rapidly (“instanta-
neously”) reaches a steady-state enrichment and one can measure
a pseudolinear increase in the enrichment of the product
[41, 42]. In cases where one collects multiple samples it is possible
to model the change in enrichment of the product using the
equation:
product enrichmentðt Þ ¼ steady  state product enrichment


 1  e kt ð6Þ
As with Eq. 5, Eq. 6 requires that the precursor enrichment is
stable, preferably reflecting a “square wave” over the period in
which the product enrichment is being sampled [41]. In cases
where the precursor enrichment may vary, the same logic will
apply but adjustments need to be made in the calculations [43, 44].
Now let us consider how we might estimate triglyceride kinet-
ics, for example, in adipose tissue. In this scenario it is not possible
to directly administer a labeled triglyceride and measure its enrich-
ment over time in adipose tissue (triglycerides are normally hydro-
lyzed by lipoprotein lipase and not directly removed by adipocytes)
[45]. Therefore, we need to consider how to measure the conver-
sion of a labeled precursor into a labeled product, (e.g., [13C6]
glucose ! [13C3]glycerol-3-phosphate ! [13C3]triglyceride-glyc-
erol). For simplicity we will assume what it expected when either a
single IV bolus or a primed-infusion is used (Fig. 3), as well, we will
assume that one would measure the enrichment of the triglyceride-
glycerol backbone. It is certainly possible that fatty acids will be
labeled too, in fact, one can find major discrepancies between the
kinetics of triglyceride-glycerol vs triglyceride-fatty acids. Although
these discrepancies are especially obvious when the structural com-
position of the pool is being remodeled we will ignore fatty acid
labeling for now [6].
When the precursor is given as a bolus, the product enrichment
will rise and then fall over time (Fig. 3a). Obviously, the magnitude
 Key Concepts Surrounding Studies of Stable Isotope-Resolved Metabolomics 107

Fig. 3 Theoretical precursor and product labeling profiles for different routes of tracer administration in vivo.
Panel a considers a case where two subjects are given the same IV bolus of a labeled precursor (e.g., [U-13C6]
glucose) and the temporal dilution is identical between the subjects (dashed line), however, the product (e.g.,
triglyceride) is turning over at ~0.04 or 0.80 pool/h, solid or open symbols, respectively. Panel b considers the
same scenario, however, the precursor is given as a primed-infusion (dashed line); again, the product is
turning over at ~0.04 or 0.80 pool/h, solid or open symbols, respectively

and duration of the change in product enrichment will be driven by

how much precursor is administered, the half-life of the precursor
and the turnover of the product [44]. In contrast, if we give a
primed-infusion of the precursor we expect to see a steady rise in
the enrichment of the product (Fig. 3b). Note that in both Fig. 3a,
b, we assume that the product represents a single pool modeled by a
single exponential. We will now compare these modes of tracer
administration and include the labeling profiles in cases where
products may turnover at ~0.04 or ~0.8 pools/h (we have arbi-
trarily set the glucose precursor to be ~10% enriched). When a
bolus is administered it is possible to draw different conclusions
regarding the relative activity of the respective pools. For example,
although there is a 20-fold difference between the turnover of the
respective product pools, if we collected a sample at a later point in
time (e.g. at ~ 12 hours), we might erroneously conclude that the
pools have comparable activity, or that the slower pool is turned
over more rapidly than the faster pool (Fig. 3a). In contrast, the use
of a primed-infusion would allow us to ascribe the correct relative
activities of the pools regardless of when we collected a sample,
however, a drawback is that the true difference between the activ-
ities of the respective pools is grossly underestimated if we assumed
that a single sample could provide a definitive answer regarding the
kinetics (Fig. 3b). Namely, using a sample collected at 8, 16 or 24 h
(open circles, Fig. 3b) and applying Eq. 5 we would conclude that
the fractional synthesis is ~0.13, 0.06, or 0.04 pools/h, respectively
[41]; none of these values reflect the true turnover (~0.80 pool/h),
therein underscoring the impact of indiscriminate sampling.
108 Stephen F. Previs and Daniel P. Downes

2.4 How Many Building on the example shown in Fig. 3 raises obvious questions
Samples Do I Need and around how the number of samples and the timing of their collec-
When Should They Be tion will impact our conclusions [46]. In our experience, even if we
Collected? could collect multiple samples the use of a single tracer bolus is not
practical for studying a problem such as triglyceride flux in adipose
tissue. The pool size is generally quite large, and the turnover is
relatively slow. We should note that the time scale used here is
exaggerated even for rodent studies [5, 6]. This does not mean
that the logic surrounding a bolus administration of a tracer is
flawed, in fact, this approach is used in studies of plasma lipoprotein
kinetics [23, 47]. However, if a single bolus is administered we are
somewhat obligated to collect multiple samples to determine if we
are on the up-slope or the down-slope of a product labeling curve.
Note that for the rapidly turning over product in Fig. 3a, the
enrichment is comparable at ~1 h and ~5 h. If we had only collected
a sample at ~5 h, we could draw appropriate conclusions regarding
the relative differences between the kinetics of the two product
pools. The open circles clearly demonstrate substantially more
enrichment as compared to the closed circles, consistent with the
faster relative turnover (Fig. 3a). Unfortunately, we would be far
from an accurate estimate of the true turnover, for example, the
open circles are on the descending phase of the enrichment curve
by 5 h and therefore we have grossly underestimated the synthesis
in that condition.
Those conclusions are in strong contrast with the example in
Fig. 3b, where the enrichment continues to increase until it reaches
a plateau, making it possible to consider using a single time point to
estimate the kinetics [42]. Such a statement would be appealing to
the analytical lab since it would mean fewer samples to process and
less data to acquire, which implies a level of resource sparing.
However, it also has the potential drawback that the longer we
wait to collect our samples the less accurate will be our estimation of
the true differences between the two product pools [42]. Although
it is tempting to think that a single data point can be used to infer
the kinetics (e.g., Eq. 5), several factors must align to ensure the
validity of that logic. Given the complexity of some biological
problems it is perhaps better to consider more extensive sampling
schemes. At the very least, pilot studies should evaluate critical
assumptions noted here. Those preliminary data may suggest rea-
sonable options to simplify the experimental design in follow up
studies.

These questions are of central importance in virtually all studies

where tracers are administered in the context of studying precur-
sor ! product relationships, as such, they deserve special attention.
In addition, addressing these questions should underscore the
importance of tracer administration modes and sampling schemes
on the data interpretation. A discussion of the topic is best
 Key Concepts Surrounding Studies of Stable Isotope-Resolved Metabolomics 109

2.5 How Do We Know exemplified if we build off Fig. 3b. We will add some complexity to
Whether the reflect true events surrounding metabolic biochemistry and inte-
Intracellular Dilution of grative physiology.
a Precursor Is Creating The example in Fig. 3b shows a case where our substrate (e.g.,
13
a Problem and How C-glucose) is the only precursor for the glycerol backbone in
Can We Adjust the triglyceride; therefore, we expect that we will eventually reach a
Calculations of steady-state enrichment of triglyceride-glycerol and that its enrich-
Metabolic Flux? ment will equal that of the precursor. In the example, the labeled
glucose precursor pool is ~10% enriched, consequently, the product
should reach the same level if there are no other sources of the
glycerol backbone. However, several studies have demonstrated the
existence of alternative sources of triglyceride-glycerol [5, 48–
51]. For example, although the adipose tissue is thought to have
little or no glycerokinase, it has been hypothesized that glycero-
neogenesis could occur (i.e., the conversion of lactate, pyruvate,
etc. to triose-phosphates) [52]. Let us further evolve the reaction
sequence in our metabolic pathway to see what happens if glucose is
not the only source of triglyceride-glycerol.
Figure 4a outlines a scenario that is supported by experimental
data. If we assume that 13C-glucose is given as a primed-infusion,
and that the precursor is ~10% enriched we expect that the enrich-
ment of triglyceride-glycerol will reach that of the precursor
(as shown in Fig. 4c, open triangles). However, if we found com-
parable precursor enrichment in two conditions but different prod-
uct enrichment (i.e., we collected samples from the respective
groups at any of the time points noted in Fig. 4c) can we confi-
dently state that the metabolic flux of the product is different
between the respective groups? The answer is “no.” Although the
examples represented in Fig. 4a, b demonstrate a case where two
groups are maintained at the same glucose enrichment (~10%) and
the product enrichment is approximately twofold different for the
respective curves. We do not know whether some factor other than
triglyceride synthesis is altered. In fact, the simulation shown here
reflects differences in the precursor dilution and not in the triglyc-
eride kinetics. The triangles (Fig. 4c) represent the enrichment of
triglyceride-glycerol, in each case the glucose is ~10% enriched and
the half-life of triglyceride is ~ 3.4 h (or ~0.2 pool/h). However,
the open triangles reflect the outcome if 13C-glucose was the sole
precursor (Fig. 4a), whereas the solid triangles reflect the outcome
if there was another source of triglyceride-glycerol (lactate, pyru-
vate, etc.), that enters the cell “cold” (Fig. 4b). Since this second
source of the glycerol backbone is not labeled the pathway is
invisible, this underscores a key assumption, in that, the labeled
precursor that we administer may not be the true precursor. For
example, if we collected a single sample at ~6 h and we applied Eq. 5
we would conclude that the turnover was ~0.11 vs ~0.06 pools/h,
open vs. solid triangles, respectively. This is in strong contrast to the
true value of ~0.20 pools/h in both cases.
110 Stephen F. Previs and Daniel P. Downes

Fig. 4 Product labeling can be impacted by biochemical rates and sources of substrates. One can consider
using a primed-infusion of [U-13C6]glucose in studies of triglyceride kinetics. For simplicity, we assume that
13
C- will only be incorporated into the glycerol-backbone (in reality, 13C- can be incorporated into the fatty
acids as well). If glucose is the only carbon source (Panel a), the 13C- enrichment of triglyceride-glycerol will
reach a steady-state (Panel c, open triangles), which is equal to that of the 13C-glucose being infused (Panel c,
dashed line). In theory, one could estimate triglyceride turnover by collecting a single sample during the
pseudolinear phase of the triglyceride labeling and comparing that against the 13C-enrichment of the infused
glucose (see Eq. 5). In contrast, if glucose and pyruvate, for example, each made similar contributions to the
triose-phosphate (and glycerol-phosphate) pool (Panel b), we could observe very different absolute changes in
the triglyceride enrichment (Panel c, solid triangles). In this case, we would draw erroneous conclusions by
comparing the enrichment of the product with that of the infused glucose (see Eq. 5). The true precursor (i.e.,
glycerol-3-phosphate) receives equal input of carbon from labeled (e.g., glucose) and unlabeled (e.g.,
pyruvate) sources. If we sampled at several time points (e.g., 2, 6, 12, and 24 h) and applied Eq. 6 we
would determine that the triglyceride turnover is the same between the groups (i.e., the correct conclusion).
Note that the correct stoichiometry to generate the curve represented by the closed triangles in Panel c is 1
glucose and 2 pyruvate

It is important to note that the sampling scheme can help

dissect these types of problems. For example, if we ran the experi-
ment described above, and we determined that the glucose enrich-
ment was the same between two groups, but the enrichment of
triglyceride-glycerol was different (i.e., we collected a sample at
~6 h) we might be inclined to conclude that triglyceride synthesis
Key Concepts Surrounding Studies of Stable Isotope-Resolved Metabolomics 111

was different between the groups (as we just noted). However, this
would assume that the 13C-enrichment of glucose reflects the true
precursor labeling for both groups. On the other hand, if we were
able to collect several samples across a reasonable time range (e.g.,
~2, 6, 12, and 24 h) we would immediately recognize that the
fractional turnover of the triglyceride is comparable between the
two groups, that is, if we fit the enrichment curves using Eq. 6 we
would see that it takes the same amount of time to reach a plateau in
both groups (Fig. 4c). We would also note that the plateau labeling
is different between the groups, which could be explained by the
entry of “cold” carbon; the dilution of our labeled glucose precur-
sor could be “proved” by measuring the enrichment of an interme-
diate at or beyond the mixing point. That is, if we sampled the
triose-phosphates or glycerol-3-phosphate we would see that the
enrichment is 50% lower than that of the glucose. Future studies
could then be designed with this new knowledge in mind. For
example, we could sample the triose-phosphates or glycerol-3-
phosphate enrichment and set that value as the true precursor
labeling, we could then consider using a single time point (e.g.,
measure triglyceride-glycerol at a 6 h time point and Eq. 5) and
therein draw reliable conclusions regarding triglyceride kinetics
[42]. This could make some aspects of our studies simpler since
we could perhaps then collect and measure fewer samples, and the
experimental time might be shortened (e.g., from 24 h to 6 h).
A critical take-home message is that although we may observe
the incorporation of a labeled precursor into a product we should
question the potential of any dilution in the steps between the
precursor and the product. The example shown here might seem
extreme since there are several reactions between glucose and
triglyceride-glycerol and readers may think that if we use a more
direct precursor that this problem would not arise. That is not
entirely the case – studies of protein synthesis often administer a
labeled amino acid and measure its incorporation into a protein
(s) [35, 53]. Readers may recognize that there are fewer steps in
that precursor ! product relationship, that is, free amino
acids ! tRNA-bound amino acids ! protein-bound amino acids.
However, another problem is observed: the enrichment of the
amino acid that is administered can undergo substantial intracellu-
lar dilution since cells continuously degrade proteins and recycle
amino acids. In fact, the enrichment of free amino acids has been
shown to experience ~40% dilution upon entry into a cell [54].
This raises a question regarding the novelty of using 2H- or
18
O-water to study metabolic flux [24]. Briefly, water readily moves
across membranes and its enrichment is virtually identical in all
body compartments shortly after it is administered [55, 56]. For
example, we have given ~5–20 μL of 2H-water per gram body
weight to rodents (via an intraperitoneal bolus) and observed
90% distribution within ~20 min (unpublished observations).
Since labeled water rapidly enters cells we should expect a homoge-
neous labeling. However, we then need to assume that the entry of
112 Stephen F. Previs and Daniel P. Downes

labeled water into intermediates that are used in the synthesis of

end-products is faster than the synthesis of the end-products [57],
that is, the generation of 2H-labeled triose-phosphates should be
much faster than their incorporation into triglyceride-glycerol. A
critical caveat centers on whether it is acceptable to measure the
enrichment of the entire triglyceride-glycerol or whether we should
devise a strategy to measure the enrichment of specific hydrogens
[5, 6, 43, 48, 58].

3 Impact of “Secondary” Tracers

Our discussion has considered somewhat clean systems, in which

the precursor has only a single route through which it can be
incorporated into a product. This is rarely the case when we con-
duct in vivo experiments, as one might expect since labeled pre-
cursors can traverse pathways in multiple tissues and then mix again
in the plasma before undergoing incorporation into a product of
interest [59]. We will now outline some of the problems that can
arise around data interpretation in cases where secondary tracers are
produced, this can impact seemingly “clean” in vitro models just as
severely as it can impact more complex in vivo models. There are
good examples where investigators have recognized how tracer
recycling could impact the interpretation of stable isotope-resolved
studies in cases where acute tracer perturbations are used to probe
tissue- and pathway-specific activity in vivo [60, 61].

3.1 Our Primary We previously incubated hepatocytes with ~150 μM [2H5]glycerol

Tracer Undergoes and measured the 2H-enrichment of glyceride-glycerol at various
Rearrangement Before times. Figure 5 demonstrates the generation of different labeled
Reaching Its Final species which can be explained if [2H5]glycerol undergoes conver-
Destination sion to triose-phosphates faster than incorporation into
triglyceride-glycerol. These data are consistent with the literature
[23] and imply that we could miss aspects of triglyceride-flux by
ignoring the generation of alternate precursors. For example, in this
case the primary glycerol tracer was M5 but the product
triglyceride-glycerol contained all of the mass isotopomers. As
eluded to earlier, the observation is not an artifact of using an
in vitro model, since this occurs in humans [23] and is consistent
with known biochemical reactions (Fig. 5a). Therefore, one should
be aware of the potential for isotopic rearrangements and/or
scrambling of the labeling localization.

3.2 Cross Talk of If we aim to study metabolic activity in vivo we must recognize that
Tracers Between our primary tracer can generate other labeled precursors. Not only
Tissues can this occur in the form of different mass or positional isomers of
the original precursor (shown in Fig. 5) but one can also expect new
sources of label altogether. An example of secondary tracers can be
 Key Concepts Surrounding Studies of Stable Isotope-Resolved Metabolomics 113

Fig. 5 Impact of side compartments on isotope scrambling and potential errors estimating biochemical flux.
The biochemical model outlines potential reactions in the conversion of [2H5]glycerol to triglyceride-glyceride
(Panel a). Although there are two steps in the pathway (i.e., glycerol ! glycerol-3-phosphate ! triglyceride),
the pool of glycerol-3-phosphate is also linked to the pool of triose-phosphates. Consequently, the relative
rates of flux (i.e., glycerol-3-phosphate ! triglyceride vs glycerol-3-phosphate $ triose-phosphates) will
impact the mass isotopomer distribution profile of triglyceride-glycerol (Panel a). Hepatocytes were isolated
from rats and incubated immediately with ~150 μM [2H5]glycerol (M5, >95% enriched), triglycerides were
isolated at different times and the appearance of various isotopically labeled species (M1 ! M5) was
measured (Panel b, unpublished data from previous studies [76]). Note that solid and open squares represent
2
H and 1H, respectively

seen in studies of triglyceride flux in adipose tissue, where investi-

gators used labeled water to capture total glyceride flux and carbon-
labeled glucose to differentiate the sources of triglyceride-glycerol
(although in some studies the glucose tracer contained radioactive
14
C- atoms the impact would be the same if it had contained 13C-
glucose) [49, 50]. In these types of studies, we should appreciate
the fact that carbon-labeled glucose is consumed by virtually all
cells and we should expect the generation of carbon-labeled by-pro-
ducts (e.g., carbon-labeled lactate and pyruvate) (Fig. 6a)
[62]. Therefore, it is not advisable to conclude that the carbon
labeling of triglyceride-glycerol can differentiate the contribution
of glucose vs. nonglucose sources of triglyceride-glycerol.
Although one may consider that 3-carbon intermediates will enter
and traverse the TCA cycle (shown as red symbols, Fig. 6a) and
therefore one could make corrections for tracer recycling by mea-
suring the M1 ! M3 species, the scheme has oversimplified the
physiological problem. In fact, the carbon-labeled lactate generated
by the red blood cells, for example, could be removed by the liver or
kidney, enter/traverse the TCA cycle, and appear again in the blood
as carbon-labeled glucose [59], in which case, one would still
observe a scrambled isotope pattern in the triglyceride-glycerol;
however, the M1 ! M3 species could now also come from glucose.
This type of tissue cross talk can dramatically confound the
114 Stephen F. Previs and Daniel P. Downes

Fig. 6 Integrative physiology and tissue-induced tracer cross talk. The interpretation of tracer studies requires
special attention when there is tissue cross talk. Although labeled-glucose can be used to quantify triglyceride
synthesis, one should exercise caution when choosing the tracers if the goal is to sort out the contribution of
specific pathways. As shown in Fig. 4, sampling of glycerol-phosphate (or triose-phosphates) may be
necessary to study true rates of triglyceride turnover. One could imply that comparing the enrichment of
glycerol-phosphate with that of glucose could also yield information on the source(s) of triglyceride-glycerol.
Panel A demonstrates that it is possible to “recycle” label from one tissue to another. In this case, the infusion
of [U-13C6]glucose can lead to the generation of 13C-pyruvate in an adipocyte (via glycolytic flux/transfer of
13
C- from the red blood cell, RBC). The solid red circles represent 13C-lactate/pyruvate that traverses the
Krebs cycle prior to conversion into triglyceride-glycerol; those reactions would lead to the generation of
different mass isotopomers. In contrast, Panel B demonstrates that [6,6-2H2]glucose should not be impacted
by this type of tissue cross talk, 2H should only appear in triglyceride-glycerol is there is direct conversion of
glucose. The 2H-labeling of pyruvate could be affected by keto-enol tautomerism, transamination and/or by
traversing the TCA cycle before it moves to glycerol-3-phosphate

interpretation in cases where one aimed to differentiate the sources

of an end-product.
In contrast, the outcome is expected to be different if one uses
[6,6-2H]glucose (Fig. 6b). This tracer will undergo the same
peripheral metabolism as carbon-labeled glucose however the iso-
tope is lost during those reactions. Consequently, we should expect
 Key Concepts Surrounding Studies of Stable Isotope-Resolved Metabolomics 115

that the appearance of labeled triglyceride-glycerol from [6,6-2H]

glucose will accurately reflect the direct contribution of glucose [5].
The generation of secondary tracers via tissue cross talk has
been known for many years and can present problems for studies
that rely on in vivo models [25, 59, 63]. Not only can this be seen in
examples where we look for precursor ! product conversions, as
shown above, but we can also see an inverse type of problem when
we look at substrate oxidation [64]. For example, suppose we
aimed to measure the conversion of glucose to CO2, that is, the
movement of a key metabolic substrate to a waste product. Again,
intuition may lead us to think that administering a 13C-glucose
tracer would be desirable and that all we would need to do is
measure the 13C-enrichment of CO2. In fact, this is not so simple
since 13C- can escape the pathway via exchange reactions at key
nodes. This was clearly demonstrated by Krebs and Katz in classical
studies that used radiolabeled tracers [59, 63] and elegantly
demonstrated in more recent studies by Wolfe and colleagues
using stable isotopes in humans [64–66]. Just as we might expect,
correction strategies have been proposed to circumvent some of the
problems that are associated with the oxidation of 13C-tracers [65–
67]. As an alternative, investigators have proposed the use of 2H-
labeled tracers, where measures of substrate oxidation via the pro-
duction of 2H-water would be less influenced by exchange reac-
tions [68–70].

4 Notes

Presumably readers may now start to develop a better appreciation

of key areas on which to focus attention when designing and
executing tracer-based studies of metabolic flux. Although there
have been many advances in the hardware and software as related to
mass spectrometry and nuclear magnetic resonance spectroscopy,
problems surrounding the physiology of integrative systems still
require attention. It is our experience that investigators can now
readily collect and process large sets of data, however, we should
recognize the many caveats and pitfalls can influence our ability to
reliably interpret the biochemical flux or metabolic activity. Fortu-
nately, there is a wealth of information, which can help us avoid
some of the problems outlined here [35, 44, 53, 71].
We should also note that our discussion has been biased toward
a consideration of enrichment, that is, the ratio of labeled to total
species of a given analyte. Although there are different views on
how to express stable isotope labeling [72–75] it is necessary to
emphasize a key aspect of tracer studies when the goal is to measure
a synthetic process vs a degradative process [13]. The examples
outlined here have focused attention on events surrounding the
synthesis of new material. For example, we gave 13C-glucose that
116 Stephen F. Previs and Daniel P. Downes

was ~10% enriched and we noted that our product, that is,
triglyceride-glycerol, could reach 10% enrichment if glucose was
the sole substrate being used in the synthesis. In cases where we aim
to measure the breakdown can we extract information from the
decrease in enrichment? For example, if one thinks about classical
pulse-chase experiments, intuition might lead one to consider the
decrease in enrichment being caused by a clearance or degradation
process. In fact, this is not true [13]. The decrease in enrichment
that one observes during a “chase,” that is, once the precursor
input has been stopped, will reflect the dilution of a tracer (or,
stated another way, the synthesis of that product from a cold
source). There seems to be some confusion in the literature regard-
ing these matters, and this is certainly one area where stable
isotope-based experiments can deviate from classical radiolabeled
tracer studies [13].
In our experience it is not possible to give strict conditions for
designing a tracer study, as the exact details will depend on many
variables. For example, we would likely design different protocols if
we were interested in studying triglyceride synthesis in adipose
tissue vs liver in a mouse model or if we were interested in triglyc-
eride synthesis and secretion in plasma in a rodent model vs a
human subject. This is in strong contrast to purely analytical meth-
ods, that is, lipidomic analyses that might be used to measure the
abundance of a triglyceride. In those cases, the same (or similar
enough) approach can be used to quantify the abundance and
isotopic labeling of triglycerides regardless of their source. Classical
methods (e.g., the Folch or Bligh and Dyer extraction) can be used
to isolate triglycerides from plasma, liver or adipose tissue samples
(regardless of the model). Therefore, when one aims to quantify the
abundance and/or isotope labeling of an analyte it is possible to
devise more concrete steps or rules to ensure consistency of the
data.
Hopefully we have provided some general guidelines on which
to begin designing tracer experiments. Investigators should appre-
ciate the fact that protocols which are valid in one type of study may
require substantial editing for use in another type of study, even if it
appears that we have the same goal in the respective studies. In our
experience tracer studies require that one respect a handful of rules
but there is also an opportunity for creativity. Everyone will recog-
nize a basic rule learned in elementary school, that is, mixing blue
and yellow yields green; however, artists regularly adjust the mix-
ture to achieve the “best” green for a given painting. In our
opinion, the effective application of tracer methods requires that
one balance some strict principles against some imagination since
biological problems vary in their complexity, including the con-
straints that are imposed when examining different models.
 Key Concepts Surrounding Studies of Stable Isotope-Resolved Metabolomics 117

Acknowledgments

SF Previs thanks Dr Richard Higashi (University of Kentucky) for

interesting discussions which influenced this work, the central ideas
and flow of this manuscript evolved as we shared notes regarding
our respective talks for a joint presentation at a conference.

References
1. Samuel VT, Liu ZX, Qu X, Elder BD, Bilz S, skeletal muscle protein synthetic rates over
Befroy D, Romanelli AJ, Shulman GI (2004) increasing periods of label incorporation. J
Mechanism of hepatic insulin resistance in Appl Physiol 118(6):655–661
non-alcoholic fatty liver disease. J Biol Chem 11. Rachdaoui N, Austin L, Kramer E, Previs MJ,
279(31):32345–32353 Anderson VE, Kasumov T, Previs SF (2009)
2. Erion DM, Shulman GI (2010) Diacylglycerol- Measuring proteome dynamics in vivo. Mol
mediated insulin resistance. Nat Med 16 Cell Proteomics 8(12):2653–2663
(4):400–402 12. Busch R, Kim YK, Neese RA, Schade-Serin V,
3. Coen PM, Goodpaster BH (2012) Role of Collins M, Awada M, Gardner JL, Beysen C,
intramyocelluar lipids in human health. Trends Marino ME, Misell LM et al (2006) Measure-
Endocrin Metab 23(8):391–398 ment of protein turnover rates by heavy water
4. Turner SM, Hellerstein MK (2005) Emerging labeling of nonessential amino acids. Biochim
applications of kinetic biomarkers in preclinical Biophys Acta 1760(5):730–744
and clinical drug development. Curr Opin 13. Daurio NA, Wang SP, Chen Y, Zhou H, McLa-
Drug Discov Devel 8(1):115–126 ren DG, Roddy TP, Johns DG, Milot D,
5. Bederman IR, Foy S, Chandramouli V, Alexan- Kasumov T, Erion MD et al (2017) Enhancing
der JC, Previs SF (2009) Triglyceride synthesis studies of pharmacodynamic mechanisms via
in epididymal adipose tissue contribution of measurements of metabolic flux: fundamental
glucose and non-glucose carbon sources. J concepts and guiding principles for using stable
Biol Chem 284(10):6101–6108 isotope tracers. J Pharmacol Exp Ther 363
6. Brunengraber DZ, Mccabe BJ, Kasumov T, (1):80–91
Alexander JC, Chandramouli V, Previs SF 14. Samarel AM (1991) In vivo measurements of
(2003) Influence of diet on the modeling of protein turnover during muscle growth and
adipose tissue triglycerides during growth. Am atrophy. FASEB J 5(7):2020–2028
J Physiol Endocrinol Metab 285(4): 15. DeFronzo RA, Ferrannini E (1987) Regula-
E917–E925 tion of hepatic glucose metabolism in humans.
7. Gasier HG, Fluckey JD, Previs SF (2010) The Diabetes Metab Rev 3(2):415–459
application of (H2O)-H-2 to measure skeletal 16. DeFronzo RA, Ferrannini E, Hendler R,
muscle protein synthesis. Nutr Metab 7:31 Wahren J, Felig P (1978) Influence of hyper-
8. Previs SF, Fatica R, Chandramouli V, Alexan- insulinemia, hyperglycemia, and the route of
der JC, Brunengraber H, Landau BR (2004) glucose administration on splanchnic glucose
Quantifying rates of protein synthesis in exchange. Proc Natl Acad Sci U S A 75
humans by use of (H2O)-H-2: application to (10):5173–5177
patients with end-stage renal disease. Am J 17. Previs SF, Brunengraber DZ, Brunengraber H
Physiol Endocrinol Metab 286(4):E665–E672 (2009) Is there glucose production outside of
9. Wilkinson DJ, Franchi MV, Brook MS, Narici the liver and kidney? Annu Rev Nutr 29:43–57
MV, Williams JP, Mitchell WK, Szewczyk NJ, 18. Hundal RS, Krssak M, Dufour S, Laurent D,
Greenhaff PL, Atherton PJ, Smith K (2014) A Lebon V, Chandramouli V, Inzucchi SE, Schu-
validation of the application of D2O stable mann WC, Petersen KF, Landau BR et al
isotope tracer techniques for monitoring day- (2000) Mechanism by which metformin
to-day changes in muscle protein subfraction reduces glucose production in type 2 diabetes.
synthesis in humans. Am J Physiol Endocrinol Diabetes 49(12):2063–2069
Metab 306(5):E571–E579 19. Shulman GI, Landau BR (1992) Pathways of
10. Miller BF, Wolff CA, Peelor FF III, Shipman glycogen repletion. Physiol Rev 72
PD, Hamilton KL (2015) Modeling the con- (4):1019–1035
tribution of individual proteins to mixed
118 Stephen F. Previs and Daniel P. Downes

20. Chung ST, Chacko SK, Sunehag AL, Hay- 30. Barrows BR, Timlin MT, Parks EJ (2005) Spill-
mond MW (2015) Measurements of gluconeo- over of dietary fatty acids and use of serum
genesis and glycogenolysis: a methodological nonesterified fatty acids for the synthesis of
review. Diabetes 64(12):3996–4010 VLDL-triacylglycerol under two different
21. Bessesen DH, Vensor SH, Jackman MR (2000) feeding regimens. Diabetes 54(9):2668–2673
Trafficking of dietary oleic, linolenic, and stea- 31. Verhoeven NM, Schor DSM, Previs SF,
ric acids in fasted or fed lean rats. Am J Physiol Brunengraber H, Jakobs C (1997) Stable iso-
Endocrinol Metab 278(6):E1124–E1132 tope studies of phytanic acid alpha-oxidation:
22. Romanski SA, Nelson RM, Jensen MD (2000) in vivo production of formic acid. Eur J Pediatr
Meal fatty acid uptake in human adipose tissue: 156:S83–S87
technical and experimental design issues. Am J 32. Wang SP, Zhou D, Yao Z, Satapati S, Chen Y,
Physiol Endocrinol Metab 279(2):E447–E454 Daurio NA, Petrov A, Shen X, Metzger D, Yin
23. Patterson BW, Mittendorfer B, Elias N, W et al (2016) Quantifying rates of glucose
Satyanarayana R, Klein S (2002) Use of stable production in vivo following an intraperitoneal
isotopically labeled tracers to measure very low tracer bolus. Am J Physiol Endocrinol Metab
density lipoprotein-triglyceride turnover. J 311(6):E911–E921
Lipid Res 43(2):223–233 33. van Dijk TH, Laskewitz AJ, Grefhorst A, Boer
24. Previs SF, McLaren DG, Wang SP, Stout SJ, TS, Bloks VW, Kuipers F, Groen AK, Reijn-
Zhou H, Herath K, Shah V, Miller PL, goud DJ (2013) A novel approach to monitor
Wilsie L, Castro-Perez J et al (2014) New glucose metabolism using stable isotopically
methodologies for studying lipid synthesis and labelled glucose in longitudinal studies in
turnover: looking backwards to enable moving mice. Lab Anim 47(2):79–88
forwards. Biochim Biophys Acta 1842 34. Sun RC, Fan TW, Deng P, Higashi RM, Lane
(3):402–413 AN, Le AT, Scott TL, Sun Q, Warmoes MO,
25. Previs SF, Kelley DE (2015) Tracer-based Yang Y (2017) Noninvasive liquid diet delivery
assessments of hepatic anaplerotic and TCA of stable isotopes into mouse models for deep
cycle flux: practicality, stoichiometry, and metabolic network tracing. Nat Commun 8
hidden assumptions. Am J Physiol Endocrinol (1):1646
Metab 309(8):E727–E735 35. Wolfe RR, Chinkes DL (2005) Isotope tracers
26. Befroy DE, Perry RJ, Jain N, Dufour S, Cline in metabolic research: principles and practice of
GW, Trimmer JK, Brosnan J, Rothman DL, kinetic analyses. Wiley-Liss, Hoboken, NJ
Petersen KF, Shulman GI (2014) Direct assess- 36. Shipley RA, Clark RE (1972) Tracer methods
ment of hepatic mitochondrial oxidative and for in vivo kinetics. Theory and applications.
anaplerotic fluxes in humans using dynamic Academic, New York
(13)C magnetic resonance spectroscopy. Nat 37. Wang SP, Satapati S, Daurio NA, Kelley DE,
Med 20(1):98–102 Previs SF (2017) Reply to letter to the editor:
27. Hellerstein MK, Christiansen M, Kaempfer S, “The art of quantifying glucose metabolism”.
Kletke C, Wu K, Reid JS, Mulligan K, Heller- Am J Physiol Endocrinol Metab 313(2):
stein NS, Shackleton CHL (1991) Measure- E259–E261
ment of denovo hepatic lipogenesis in humans 38. Matthews DE, Downey RS (1984) Measure-
using stable isotopes. J Clin Investig 87 ment of urea kinetics in humans: a validation of
(5):1841–1852 stable isotope tracer methods. Am J Phys 246
28. Beysen C, Ruddy M, Stoch A, Mixson L, (6 Pt 1):E519–E527
Rosko K, Riiff T, Turner SM, Hellerstein MK, 39. Ostlund RE Jr, Matthews DE (1993) [13C]
Murphy EJ (2018) Dose-dependent quantita- cholesterol as a tracer for studies of cholesterol
tive effects of acute fructose administration on metabolism in humans. J Lipid Res 34
hepatic de novo lipogenesis in healthy humans. (10):1825–1831
Am J Physiol Endocrinol Metab 315(1): 40. Zhou H, Wang SP, Herath K, Kasumov T,
E126–E132 Sadygov RG, Previs SF, Kelley DE (2015)
29. McLaren DG, He T, Wang SP, Mendoza V, Tracer-based estimates of protein flux in cases
Rosa R, Gagen K, Bhat G, Herath K, Miller of incomplete product renewal: evidence and
PL, Stribling S et al (2011) The use of stable- implications of heterogeneity in collagen turn-
isotopically labeled oleic acid to interrogate over. Am J Physiol Endocrinol Metab 309(2):
lipid assembly in vivo: assessing pharmacologi- E115–E121
cal effects in preclinical species. J Lipid Res 52 41. Foster DM, Barrett PH, Toffolo G, Beltz WF,
(6):1150–1161 Cobelli C (1993) Estimating the fractional syn-
thetic rate of plasma apolipoproteins and lipids
 Key Concepts Surrounding Studies of Stable Isotope-Resolved Metabolomics 119

from stable isotope data. J Lipid Res 34 carbohydrate-free diet. Horm Metab Res 27
(12):2193–2205 (7):310–313
42. Daurio NA, Wang Y, Chen Y, Zhou H, 52. Ballard FJ, Hanson RW, Leveille GA (1967)
Carballo-Jane E, Mane J, Rodriguez CG, Phosphoenolpyruvate carboxykinase and the
Zafian P, Houghton A, Addona G et al synthesis of glyceride-glycerol from pyruvate
(2019) Spatial and temporal studies of meta- in adipose tissue. J Biol Chem 242
bolic activity: contrasting biochemical kinetics (11):2746–2750
in tissues and pathways during fasted and fed 53. Waterlow JC (2006) Protein turnover. CABI,
states. Am J Physiol Endocrinol Metab 316(6): Oxfordshire
E1105–E1117 54. Lichtenstein AH, Cohn JS, Hachey DL, Millar
43. Bederman IR, Dufner DA, Alexander JC, Pre- JS, Ordovas JM, Schaefer EJ (1990) Compari-
vis SF (2006) Novel application of the “doubly son of deuterated leucine, valine, and lysine in
labeled” water method: measuring CO2 pro- the measurement of human apolipoprotein A-I
duction and the tissue-specific dynamics of and B-100 kinetics. J Lipid Res 31
lipid and protein in vivo. Am J Physiol Endo- (9):1693–1701
crinol Metab 290(5):E1048–E1056 55. Mccabe BJ, Bederman IR, Croniger CM, Mill-
44. Steele R (1971) Tracer probes in steady-state ward CA, Norment CJ, Previs SF (2006)
systems. Springfield. Charles C Thomas, Reproducibility of gas chromatography-niass
Illinois spectrometry measurements of H-2 labeling
45. Frayn KN, Coppack SW, Fielding BA, Hum- of water: application for measuring body com-
phreys SM (1995) Coordinated regulation of position in mice. Anal Biochem 350
hormone-sensitive lipase and lipoprotein lipase (2):171–176
in human adipose tissue in vivo: implications 56. Annegers J (1954) Total body water in rats and
for the control of fat storage and fat mobiliza- in mice. Proc Soc Exp Biol Med 87
tion. Adv Enzym Regul 35:163–178 (2):454–456
46. Previs SF, Herath K, Castro-Perez J, Mahsut A, 57. Brook MS, Wilkinson DJ, Atherton PJ, Smith
Zhou H, McLaren DG, Shah V, Rohm RJ, K (2017) Recent developments in deuterium
Stout SJ, Zhong W et al (2015) Effect of oxide tracer approaches to measure rates of
error propagation in stable isotope tracer stud- substrate turnover: implications for protein,
ies: an approach for estimating impact on lipid, and nucleic acid research. Curr Opin
apparent biochemical flux. Methods Enzymol Clin Nutr Metab Care 20(5):375–381
561:331–358 58. Strawford A, Antelo F, Christiansen M, Heller-
47. Melish J, Le NA, Ginsberg H, Steinberg D, stein MK (2004) Adipose tissue triglyceride
Brown WV (1980) Dissociation of apoprotein turnover, de novo lipogenesis, and cell prolifer-
B and triglyceride production in very-low-den- ation in humans measured with 2H2O. Am J
sity lipoproteins. Am J Phys 239(5): Physiol Endocrinol Metab 286(4):E577–E588
E354–E362 59. Krebs HA, Hems R, Weidemann MJ, Speake
48. Chen JL, Peacock E, Samady W, Turner SM, RN (1966) The fate of isotopic carbon in kid-
Neese RA, Hellerstein MK, Murphy EJ (2005) ney cortex synthesizing glucose from lactate.
Physiologic and pharmacologic factors influen- Biochem J 101(1):242–249
cing glyceroneogenic contribution to triacyl- 60. Kowalski GM, De Souza DP, Burch ML,
glyceride glycerol measured by mass Hamley S, Kloehn J, Selathurai A, Tull D,
isotopomer distribution analysis. J Biol Chem O’Callaghan S, McConville MJ, Bruce CR
280(27):25396–25402 (2015) Application of dynamic metabolomics
49. Nye CK, Hanson RW, Kalhan SC (2008) Gly- to examine in vivo skeletal muscle glucose
ceroneogenesis is the dominant pathway for metabolism in the chronically high-fat fed
triglyceride glycerol synthesis in vivo in the mouse. Biochem Biophys Res Commun 462
rat. J Biol Chem 283(41):27565–27574 (1):27–32
50. Botion LM, Brito MN, Brito NA, Brito SRC, 61. Kowalski GM, De Souza DP, Risis S, Burch
Kettelhut IC, Migliorini RH (1998) Glucose ML, Hamley S, Kloehn J, Selathurai A,
contribution to in vivo synthesis of glyceride- Lee-Young RS, Tull D, O’Callaghan S et al
glycerol and fatty acids in rats adapted to a (2015) In vivo cardiac glucose metabolism in
high-protein, carbohydrate-free diet. Metabo- the high-fat fed mouse: comparison of
lism 47(10):1217–1221 euglycemic-hyperinsulinemic clamp derived
51. Botion LM, Kettelhut IC, Migliorini RH measures of glucose uptake with a dynamic
(1995) Increased adipose-tissue glyceroneo- metabolomic flux profiling approach. Biochem
genesis in rats adapted to a high-protein, Biophys Res Commun 463(4):818–824
120 Stephen F. Previs and Daniel P. Downes

62. Landau BR, Wahren J, Ekberg K, Previs SF, acids for the measurement of dietary fat oxida-
Yang DW, Brunengraber H (1998) Limitations tion during physical activity. J Lipid Res 45
in estimating gluconeogenesis and Cori cycling (12):2339–2344
from mass isotopomer distributions using 70. Votruba SB, Zeddun SM, Schoeller DA (2001)
[U-C-13(6)]glucose. Am J Physiol 37(5): Validation of deuterium labeled fatty acids for
E954–E961 the measurement of dietary fat oxidation: a
63. Katz J, Chaikoff IL (1955) Synthesis via the method for measuring fat-oxidation in free-
Krebs’ cycle in the utilization of acetate by rat living subjects. Int J Obes Relat Metab Disord
liver slices. Biochim Biophys Acta 18 25(8):1240–1245
(1):87–101 71. Landau BR, Wahren J (1992) Nonproductive
64. Sidossis LS, Coggan AR, Gastaldelli A, Wolfe exchanges: the use of isotopes gone astray.
RR (1995) Pathway of free fatty acid oxidation Metabolism 41(5):457–459
in human subjects. Implications for tracer stud- 72. Ramakrishnan R (2006) Studying apolipopro-
ies. J Clin Invest 95(1):278–284 tein turnover with stable isotope tracers: cor-
65. Sidossis LS, Coggan AR, Gastaldelli A, Wolfe rect analysis is by modeling enrichments. J
RR (1995) A new correction factor for use in Lipid Res 47(12):2738–2753
tracer estimations of plasma fatty acid oxida- 73. Cobelli C, Toffolo G, Foster DM (1992)
tion. Am J Phys 269(4 Pt 1):E649–E656 Tracer-to-tracee ratio for analysis of stable iso-
66. Wolfe RR, Jahoor F (1990) Recovery of labeled tope tracer data: link with radioactive kinetic
CO2 during the infusion of C-1- vs C-2- formalism. Am J Phys 262(6 Pt 1):E968–E975
labeled acetate: implications for tracer studies 74. Chinkes DL, Aarsland A, Rosenblatt J, Wolfe
of substrate oxidation. Am J Clin Nutr 51 RR (1996) Comparison of mass isotopomer
(2):248–252 dilution methods used to compute VLDL pro-
67. Toth MJ, MacCoss MJ, Poehlman ET, Mat- duction in vivo. Am J Phys 271(2 Pt 1):
thews DE (2001) Recovery of (13)CO E373–E383
(2) from infused [1-(13)C]leucine and 75. Kharroubi AT, Masterson TM, Aldaghlas TA,
[1,2-(13)C(2)]leucine in healthy humans. Am Kennedy KA, Kelleher JK (1992) Isotopomer
J Physiol Endocrinol Metab 281(2): spectral analysis of triglyceride fatty acid syn-
E233–E241 thesis in 3T3-L1 cells. Am J Phys 263(4 Pt 1):
68. Beysen C, Murphy EJ, McLaughlin T, Riiff T, E667–E675
Lamendola C, Turner HC, Awada M, Turner 76. Previs SF, Hallowell PT, Neimanis KD,
SM, Reaven G, Hellerstein MK (2007) Whole- David F, Brunengraber H (1998) Limitations
body glycolysis measured by the deuterated- of the mass isotopomer distribution analysis of
glucose disposal test correlates highly with glucose to study gluconeogenesis—heteroge-
insulin resistance in vivo. Diabetes Care 30 neity of glucose labeling in incubated hepato-
(5):1143–1149 cytes. J Biol Chem 273(27):16853–16859
69. Raman A, Blanc S, Adams A, Schoeller DA
(2004) Validation of deuterium-labeled fatty
 Chapter 7

Extracting Biological Insight from Untargeted

Lipidomics Data
Jennifer E. Kyle

Abstract
Lipidomics data generated using untargeted mass spectrometry techniques can offer great biological insight
to metabolic status and disease diagnoses. As the community’s ability to conduct large-scale studies with
deep coverage of the lipidome expands, approaches to analyzing untargeted data and extracting biological
insight are needed. Currently, the function of most individual lipids are not known; however, meaningful
biological information can be extracted. Here, I will describe a step-by-step approach to identify patterns
and trends in untargeted mass spectrometry lipidomics data to assist users in extracting information leading
to a greater understanding of biological systems.

Key words Lipidomics, Lipidome, Untargeted, Mass spectrometry, LipidMaps, Blood plasma

1 Introduction

Lipidomics is the study of lipids within a cell or organisms to

understand the function and structure of biological systems. Lipids
are highly diverse biomolecules that have many roles critical to
living systems, including acting as membrane structural compo-
nents, signaling molecules, and energy sources [1]. The field of
lipidomics has been increasingly recognized in the past few years [2]
due to the advances of mass spectrometry techniques, and
biological insight gained through lipidomics measurements. Lipids
have been shown to be important in cellular homeostasis [3, 4],
signatures of disease [5–7], and affected by inborne errors in
metabolism [8]. Even with these advancements the exact function
of most individual lipids is not known. For untargeted lipidomics
analyses, a few hundred individual lipids may be identified within a
biological matrix [9, 10]. This leads to many questions such as how
the cell utilizes lipids for their critical functions [11] and how
cellular homeostasis may be affected by alterations in the lipid
profile. The interpretation of lipids is currently focused on known

Shuzhao Li (ed.), Computational Methods and Data Analysis for Metabolomics, Methods in Molecular Biology, vol. 2104,
https://doi.org/10.1007/978-1-0716-0239-3_7, © Springer Science+Business Media, LLC, part of Springer Nature 2020

121
122 Jennifer E. Kyle

roles of the lipid components (e.g., fatty acids) or the type of lipids
(e.g., triglyceride) identified.
Multiple review papers on sample collection to storage proce-
dures, lipid extraction from biological matrices, mass spectrometry
technologies and methods, appropriate QCs as well as software are
available [2, 12–16] and will not be covered here. In this chapter, I
will provide an approach to exploring lipidomics data at the global
level (i.e., all lipids identified in the biological study) to help guide
the understanding and interpretation of the lipidome. The integra-
tion of other “omics” data, such as proteomics and transcriptomics,
can aid in interpretation of lipidomics data as it enables mechanistic
support for lipidomics observations.

2 Lipid Classification and Annotation

Lipids are highly structural diverse biomolecules mainly due to the

various combinations of fatty acids, fatty acid linkages, and head
groups. Despite this great diversity, lipids have two basic biochemi-
cal building blocks, ketoacyl or isoprene [17]. LipidMaps has
played an important role in the field of lipidomics through standar-
dizing the classification and annotation scheme of lipid nomencla-
ture. LipidMaps has categorized lipids into eight main categories
[18], three of which (sphingolipids (SL), glycerophospholipids
(GP), and glycerolipids (GL)) are routinely identified and charac-
terized using total lipid extraction methods [19–21] and untar-
geted mass spectrometry for human health studies. Each lipid
category is further divided using a subclassification hierarchy
including lipid main classes and lipid subclasses [18]. This chapter
will focus on the lipid categories SL, GP, and GL.
Most of the lipids that are SL, GP, and GL contain a common
name that utilizes a specific annotation Scheme ZZ(X1:Y1/X2:Y2):
ZZ(X1:Y1/X2:Y2) ! example PC(16:0/20:4)
ZZ ¼ lipid class (e.g., PC)
X1 ¼ number of carbons in chain 1 (e.g., 16)
Y1 ¼ number of double bonds in chain 1 (e.g., 0)
X2 ¼ number of carbons in chain 2 (e.g., 20)
Y2 ¼ number of double bonds in chain 2 (e.g., 4)
This scheme enables the lipid common name to be used for
identifying patterns and trends in untargeted lipidomics data. Addi-
tional annotations within the parentheses can provide further sub-
class differentiation; for example, a “P” in PE(P-16:0/20:4) reveals
the PE lipid is plasmalogen (or alkenyl chain), and absence of a
chain as 0:0 in PC(0:0_16:0) reveals the lipid is a monoacylglycer-
ophosphocholine (LPC) or lysoPC.
Recently, the annotation scheme has become more refined in
that the “/” now designates the sn position of the chains is known,
whereas “_” has been suggested to indicate the sn position is not
 Analyzing Untargeted Lipidomics 123

known [22]. This change is slowly becoming incorporated in the

lipidomics literature.
The LipidMaps annotation scheme has not been universally
adopted as lipid annotation prior to LipidMaps continues to be
used (e.g., PtdPC vs. PC). However, the abbreviations of lipids are
similar and therefore can still be parsed as detailed below.

3 Parsing the Lipid Common Name for Lipidomics Pattern Discovery

One of the difficulties with interpreting lipidomics data is that the

specific functions for most lipids are not known nor is there an
inclusive database for those with known roles. Biological knowl-
edge of lipids exists at the subclass level [23] and the fatty acid level
(e.g., arachidonic acid (AA), saturated fatty acids, etc.), but knowl-
edge of the intact or complete lipid itself is sparse [24, 25] com-
pared to the number of unique lipids in biological systems
[17]. Since the role of individual components of lipids is better
understood, the common name of most SLs, GPs, and GLs can be
used to gather information to begin initial steps toward biological
interpretation of lipidomics data (Fig. 1).
Recently, a tool of Lipid Mini-On (Mining and Ontology) was
developed to parse the lipid common names into ontology bins
corresponding to its components (e.g., classification, chain compo-
sition, chain characteristics) [26]. The tool further performs
enrichment analysis on those components and assigns a significance
value. More details are provided in Subheading 4.2 step 4.

Fig. 1 Classification information and specific chain details can be gathered from the lipid common name. For
example, PC(16:_20:4), the PC reveals the type of lipid (i.e., classification scheme) while the fatty acids (FAs)
can be parsed at the fatty level (16:0 and 20:4), the FA characteristics (e.g., 16 carbons long (C16)), and also
by the total number of carbons and double bonds of the chains (i.e., 36 carbons (C36) and 4 double bonds
(DB 4)). The FAs themselves include a saturated FA (SFA) and polyunsaturated FA (PUFA) and the chain lengths
show they are both long chain FAs (LCFAs). A second example using Cer(d18:0/24:1) can be similarly parsed.
MUFA monounsaturated FA, VLCFA very long chain FA
124 Jennifer E. Kyle

4 Approach to Pattern Discovery and Understanding the Lipidome

Using the following approach, you will see that lipids tend of have
patterns based on their classification, both at the inter- and intra-
subclass level as well as chain composition or chain characteristics.
Here I present my method using two lipidomics studies, cell-sorted
human lung tissue [27] and human plasma from donors with Ebola
virus disease [28]. This approach applies to lipidomics studies were
confident identifications have been made and statistical analysis has
provided results on differences (e.g., log2 fold change or Zscore)
between groups (e.g., controls and diseased samples). This
approach involves:
1. Understanding the lipids within the sample biological matrix
(Subheading 4.1).
2. Identifying signatures and patterns in lipidomics data (Sub-
heading 4.2).
3. Integrating ‘omics data (Subheading 4.3).

4.1 Understanding Understanding how lipids influence membranes and where lipids
the Lipids Within are localized at the cellular level or within biofluids will aid in the
the Sample Biological interpretation of the lipidome. Several excellent reviews on the
Matrix cellular lipid subclass distribution and fatty acid characteristics are
available [4, 11, 29, 30]. These reviews detail the influence of the
phospholipid head groups and fatty acid composition on mem-
brane composition, fluidity, width, and charge. For example, the
small head group of PE lipids induces a negative membrane curva-
ture, and unsaturated lipids lead to more fluid membranes. Very few
lipid classes in human or mammalian systems are organelle specific.
Exceptions include cardiolipin which is located mostly in the inner
mitochondrial membrane and bis(monoacylglycero)phosphate
(BMP; also known as monoacylglycerophosphomonoradylglycerol
(LBPA)) located in late endosome and lysosomes.
Less information is available on the localization of lipids in
biological fluids. Biofluids are particularly complicated due to
both endogenous and exogenous sources of lipids. Lipids have
been identified in many biofluids [31]; however, given their hydro-
phobic nature they are, for the most part, present in low abun-
dance. In plasma and CSF, lipids are primarily associated with
lipoproteins [32–34]. In CSF, lipids are also associated with
brain-derived nanoparticles [35]. In urine, the presence of lipids
has been suggested to be due to urea facilitating the dissolution of
the lipids in urine [36].
Most lipidomics studies using biofluids are conducted using
blood plasma where lipids are abundant. Lipids in blood plasma are
associated with lipoprotein membranes (e.g., PC and sphingomye-
lin (SM)), lipoprotein cargo (e.g., triglycerides (TGs) and choles-
terol ester)), and proteins (e.g., lyso-PC). A handful of publications
 Analyzing Untargeted Lipidomics 125

Fig. 2 Summary of lipid subclass distribution per lipoprotein type in human

plasma. Lipoproteins contain an outer lipid membrane and transport lipids
(lipoprotein cargo), which make up most of the lipids identified in blood plasma.
These subclass associations may be helpful in interpreting the source of lipid
subclass trends in lipidomics data for human plasma

have characterized the lipid profile of the different types of lipopro-

teins [37–41] revealing which lipid subclass is highest in a particular
lipoprotein (Fig. 2). For example, ceramides (Cer) and sphingo-
myelins (SM) are most abundant in low-density lipoproteins (LDL)
and diacylglycerophosphoethanolamine (PE) in high-density lipo-
proteins (HDL) [37, 40, 41]. These same studies also reveal that
phospholipids and sphingolipids are present in all lipoproteins, but
the abundance of the lipids varies. Associating this knowledge with
a plasma can be difficult, in particular due to the multiple factors
that can influence the lipid abundance including diet [42], gender
[43], and circadian rhythms [44]. In addition to lipoproteins,
human bodily fluids also contain extracellular vesicles, which are
thought to contribute to physiology and pathology, increasing or
decreasing in abundance in diseased states [45] and possibly con-
tributing to lipidome signature.

4.2 Identifying To understand the lipidome, one must identify the trends, patterns,
Signatures and unique signatures. Organizing the lipid identifications with
and Patterns associated abundance values in the following manner will allow
in Lipidomics Data for observations to be gathered at the global level as well as inter-
and intrasubclass levels. Perform the steps 1–4 with both the global
results (i.e., results from all of the identified lipids) as well as the
results from your query of interest (e.g., statistically significant
lipids)
Step 1. Organize the statistical results file based on the LipidMaps
classification scheme. This will highlight global trends as
well as trends across lipid categories and subclasses
(Fig. 3a).
126 Jennifer E. Kyle

Fig. 3 Organization of lipidome of human Ebola virus disease. (a) The lipids are organized by category
(glycerolipids and glycerophospholipids), subclass, and then intrasubclass by the total number of chain carbons
and double bonds. (b) Detailed intrasubclass organization of TGs. MG monoacylglyceride, DG diacylglyceride, TG
triacylglyceride, LPC monoacylglycerophosphocholine, oxPC oxidized diacylglycerophosphocholine, PC diacyl-
glycerophosphocholine, LPE monoacylglycerophosphoethanolamine, PE diacylglycerophosphoethanolamine,
PEP plasmalogen PE, PG diacylglycerophosphosphoglycerol OR bis(monoacylglycerol)phosphate, LPI PI mono-
acylglycerophosphoinositol, PI diacylglycerophosphoinositol, PS diacylglycerophosphoserine. Data used to gen-
erate the figure is from [28] (Fig. 2) from the original Supplemental Table S2 of publication [28]
 Analyzing Untargeted Lipidomics 127

(a) Lipid category

e.g., Glycerophospholipids,
(b) Lipid main class
e.g., Glycerophosphocholines,
(c) Lipid subclass
e.g., monoacylglycerophosphocholine (LPC), dia-
cylglycerophosphocholine (PC), alkyl-
acylglycerophosphocholine (PCO).
Step 2. Organize within each subclass by the sum of the total
number of chain carbons (SumC) and then the sum of
the total number of chain double bonds (SumDB). This
will highlight intrasubclass trends (Fig. 3b).
(a) This organization largely results in shorter chained
carbons with no or low number of double bonds at
the top and longer chained polyunsaturated chains at
the bottom.
(b) LipidSplit can calculate SumC and SumDB values
and is freely available (https://github.com/PNNL-
Comp-Mass-Spec/LIPID-Split)
Note: For sphingolipids and ether linked phospho-
lipids also sort alphabetically as intrasubclass trends
are not as common as with phospholipids and glycer-
olipids or based on the long chain base or ether lipid
chain. For example, organizing ceramides alphabeti-
cally groups the long chain base by sphinganine (d18:0)
then sphingosine (d18:1) and plasmalogens will orga-
nize by P-16:0 then P-18:0.
Step 3. Examine the fatty acid composition at both the intersub-
class and intrasubclass level. Lipids with particular fatty
acids, commonly polyunsaturated fatty acids (PUFA),
may have a different pattern than the other lipids. In the
example presented in Fig. 4, all PI lipids containing 20:4
tend to increase, whereas other lipids in that subclass tend
to decrease.
Step 4. Perform enrichment analysis. Lipid Mini-On performs
enrichment analyses using the lipid common name anno-
tation and associated classification to highlight ontology
terms within a given list of lipids [26]. The tool examines
the data as outlined above such that it determines if lipids
are enriched in a statistically significant way based on their
classification (e.g., lipid category, main class, and sub-
class) and also inter- and intraclassification by the chain
characteristics (e.g., fatty acids themselves and fatty acid
length and number of double bonds) (see Fig. 1). The
query list is compared to the list of all lipids identified in
128 Jennifer E. Kyle

Fig. 4 Intrasubclass differences based on the fatty acid composition. PI lipids containing the fatty acid 20:4
were high (positive Zscore) in the endothelial cells of lung tissues of three 20-month-old donors [27], whereas
the rest of the PIs were low (negative Zscore). Data used to generate this figure is from [27] from the original
Supplemental Table S4 [27]

the associated study. Each classifier (or ontology term) is

assigned a p-value, enabling the identification of global
trends in the lipidome (Table 1). The tool also displays
the classifiers as nodes in a network along with the asso-
ciated lipids such that the user can see the exact lipids in
the enriched ontology term (Fig. 5).

4.3 Integrating The interaction of lipids with other cellular components, including
‘Omics Data proteins, genes, and metabolites, can enable a greater understand-
ing of the mechanisms behind the observations in your lipidomics
data generated in steps 1–4 above. LipidMaps protein database
contains thousands of entries across ten model organisms that
contain protein and gene identifiers for proteins and genes asso-
ciated with lipid metabolism and homeostasis. For humans, there
are 2273 entries with over 1100 unique entrez gene and gene
symbols, uniprot identifications, and protein entries. To identify
proteins or genes that may have a role in the lipid signatures noted
in Subheading 4.2, perform the following steps with your asso-
ciated proteomics (and/or transcriptomics) results.
 Analyzing Untargeted Lipidomics 129

Table 1
Lipid category and subclass enrichment in Ebola virus disease [28], using Fisher’s exact test with p-
value < 0.05

EVD—
increased in
fatalities vs
controls
FDR.
Test. Count. Count. %. %. p- q- Fold.
performed Classifier query universe query universe value value change
Category Glycerolipid 43/ 94/379 37.4 24.8 0.012 0.037 1.5
115
Sub class DG( 16/ 19/379 13.9 5.0 0.003 0.008 2.8
115
Sub class PE( 20/ 27/379 17.4 7.1 0.002 0.008 2.4
115
Sub class PG( 17/ 20/379 14.8 5.3 0.002 0.008 2.8
115
Total chain Glycerophospholipid 18/ 42/212 35.3 19.8 0.025 0.225 1.8
carbon by all with a total number of 51
chain carbon of 36
Total chain PC with a total number 1/1 3/83 100.0 3.6 0.048 1.000 27.7
carbon by all of chain carbon of 32
Specific chains PC( with the chain 16:0 2/2 15/110 100.0 13.6 0.022 1.000 7.3
by all
Total number SM(d with a total 2/3 2/28 66.7 7.1 0.037 0.074 9.3
of DB by all number of chain
unsaturation of 0

Step 1. Download LipidMaps protein database (http://www.

lipidmaps.org/data/proteome/download.php).
Step 2. Select the entries for the model organism of interest.
e.g., Homo sapiens,
Step 3. Choose the molecular identifier that aligns with your
‘omics results and check for (and remove, if appropriate)
duplicate identifier values. Also, copy the gene_name
column, as this provides some initial details on the func-
tion of the identifier.
e.g., protein_entry,
Step 4. Copy the molecular identifier chosen in step 3 and deter-
mine which lipid associated entries were detected in your
proteomics data.
130 Jennifer E. Kyle

Fig. 5 Lipid enrichment network result performed using Lipid Mini-On. The network shows statistically
significant ontology terms (diamond shapes) and the individual lipids associated with those terms were
found to be enriched in human Ebola virus disease fatalities versus controls. The network was produced from
an enrichment test using a Fisher’s Exact test with a p-value cutoff of <0.05. The query list contained lipids
that were statistically elevated in fatalities versus the controls . Data used for this analysis is from [28] from
the original Supplemental Table S2 of publication [28]

(a) If using Excel, highlight both the LipidMaps column

and the column with your data and use the highlight
“duplicate values” function. Then organize your data
columns by cell color or font to group lipid asso-
ciated proteins (Fig. 6).
(i) At this step you will have a list of all lipid-associated
proteins (and/or genes) in your ‘omics data.
(b) To obtain information on the protein(s) or gene
(s) of interest, search for the identifier in protein or
gene databases such as Uniprot (https://www.
uniprot.org/) or GeneCards (https://www.
genecards.org/).
(c) Examine the proteins with functions related to the
lipidomics results from Subheading 4.2 (Fig. 7)
Step 5. To associate the proteins (or genes) to pathways, copy the
list of lipid associated proteins (genes) into DAVID
 Analyzing Untargeted Lipidomics 131

Fig. 6 Identifying lipid-associated proteins in proteomics data. Proteomics data from a human lung study [27]
was compared against the LipidMaps protein database using the protein_entry identifier. Matching protei-
n_entry identifiers were highlighted in green. The gene_name was also tracked for the protein (gene)
descriptor. The proteomics data shows lipid-associated proteins identified in three donors (D1, D2, and D3)
of sorted endothelial cells (END), epithelial cells (EPI), mesenchymal cells (MES), and immune cells (MIC) from
the lung tissue. Proteins that are higher in the associated cell type are in red, and those that are down are in
blue. Black cells indicate that the protein was not identified in that cell type for the associated donor.
Proteomics data presented in this figure is from [27] from the original Supplemental Table S2 of
publication [27]

[46, 47] or Pathview [48, 49] (other software tools and

codes are available [16, 50]).
(a) DAVID—proteins/genes associated with pathways.
Below is one example:
(i) Go to https://david.ncifcrf.gov/
(ii) Click on “Functional Annotation”,
(iii) Copy and paste list of lipid-associated proteins,
(iv) Select identifier used in step iii,
(v) Select “Gene List” as list type,
(vi) Click “Submit List”,
(vii) Check species is correct (Homo sapiens is
default),
(viii) Click on “Pathways”,
(ix) Explore options provided,
132 Jennifer E. Kyle

Fig. 7 Integrating lipidomics, proteomics, and transcriptomics data to understand the lung lipidome. Lipido-
mics analysis of cell-sorted lung tissue from 3 20-month-old female donors revealed intrasubclass trends for
PG and TG lipids [27]. For the PGs, lipids with shorter retention times elevated in immune (MIC) cells and were
found to be BMP lipids [27], which are known to be associated with late endosomes and lysosomes.
 Analyzing Untargeted Lipidomics 133

(b) Pathview—visualize proteins (or genes) with abun-

dance information on pathways
(i) Go to https://pathview.uncc.edu/analysis
(ii) Upload gene data by clicking on “choose file”,
l Pathview offers examples of data format and
analyses to the right of the file upload and
selection area.
(iii) Select the options and pathways aligning with
your data (e.g., sphingolipid metabolism),
(iv) Click “Submit”,
(v) Review results and download as needed.

5 Summary

In this chapter, I discussed and provided an approach of using the

lipid common name to gather information and understanding of
lipidomics data. In addition, steps for the integration of other
‘omics data to assist with the interpretation of the patterns identi-
fied in the lipidomics data was provided. Given most individual
lipids do not have a known specific function and the lack of path-
ways or networks for specific lipids, the goal of this chapter is to
provide the reader with a step-by-step approach to analyze their
lipidomics data for extracting biological information leading to
interpretation.
Step 1. Organize the statistical results file based on LipidMaps
classification scheme.
Step 2. Organize within each subclass by the sum of the total
number of chain carbons (SumC) and then the sum of
the total number of chain double bonds (SumDB).
Step 3. Examine the fatty acid composition at both the intersub-
class and intrasubclass level.
Step 4. Perform enrichment analysis.


Fig. 7 (continued) Proteomics supported the presence of BMPs in MIC cells, as lysosomal acid lipase, a
protein associated with lysosomes, was only identified in MIC cells. Protein abundance is LFQ intensity. For
TGs, organization of the lipids as shown in Subheading 4.2 step 2 revealed that TGs with shorter total chain
carbons and no or low number of total double bonds showed elevated levels in mesenchymal cells (MES),
whereas longer chained, more polyunsaturated fatty acids showed elevated levels in the MIC cells. Examining
transcriptomics data revealed that lipoprotein lipase, which hydrolyzes TGs, was expressed the greatest in
MES cells, and perilipin-2, which is associated with lipid droplets, was the greatest in MIC cells. Transcript
values are log 2 abundances. END endothelial cells, EPI epithelial cells, three lung tissue donors (D1, D2, and
D3). Lipidomics data presented in this figure is modified from [27] from the original Supplemental Table S4 of
publication [28]. Protein and transcript graphs were modified from the supplemental Fig. S1 and S2 from the
publication [27]
134 Jennifer E. Kyle

Step 5. Integrate other ‘omics data.

6 Data Deposition

All mass spectrometry datasets generated from the human Ebola

virus disease study were previous published [51] and deposited in
Mass Spectrometry Interactive Virtual Environment (MassIVE) at
the University of California at San Diego under the code
MSV000080129. The cell-sorted lung lipidome data was previ-
ously published [27] and the data deposited at MassIVE under
the code MSV000081973.

Acknowledgments

I would like thank Geremy Clair and Ernesto S. Nakayasu for their
comments and careful review of the manuscript. This work was
supported by an administrative supplement to grant
U19AI106772, provided by the National Institute of Allergy and
Infectious Diseases (NIAID), National Institute of Health (NIH)
(USA). Research conducted on the lung samples were supported by
grant HL122703 from the National Heart Lung Blood Institute of
NIH.

References
1. Gross RW, Han X (2011) Lipidomics at the altered membrane lipid metabolism in breast
interface of structure and function in systems cancer progression. Cancer Res 71
biology. Chem Biol 18(3):284–291. https:// (9):3236–3245. https://doi.org/10.1158/
doi.org/10.1016/j.chembiol.2011.01.014 0008-5472.can-10-3894
2. Rustam YH, Reid GE (2018) Analytical chal- 6. Lydic TA, Goo YH (2018) Lipidomics unveils
lenges and recent advances in mass spectrome- the complexity of the lipidome in metabolic.
try based lipidomics. Anal Chem 90 diseases 7(1):4. https://doi.org/10.1186/
(1):374–397. https://doi.org/10.1021/acs. s40169-018-0182-9
analchem.7b04836 7. Zhao YY, Miao H, Cheng XL, Wei F (2015)
3. Agmon E, Stockwell BR (2017) Lipid homeo- Lipidomics: novel insight into the biochemical
stasis and regulated cell death. Curr Opin mechanism of lipid metabolism and
Chem Biol 39:83–89. https://doi.org/10. dysregulation-associated disease. Chem Biol
1016/j.cbpa.2017.06.002 Interact 240:220–238. https://doi.org/10.
4. Holthuis JC, Menon AK (2014) Lipid land- 1016/j.cbi.2015.09.005
scapes and pipelines in membrane homeostasis. 8. Lamari F, Mochel F, Saudubray JM (2015) An
Nature 510(7503):48–57. https://doi.org/ overview of inborn errors of complex lipid bio-
10.1038/nature13474 synthesis and remodelling. J Inherit Metab Dis
5. Hilvo M, Denkert C, Lehtinen L, Muller B, 38(1):3–18. https://doi.org/10.1007/
Brockmoller S, Seppanen-Laakso T, s10545-014-9764-x
Budczies J, Bucher E, Yetukuri L, Castillo S, 9. Dautel SE, Kyle JE, Clair G, Sontag RL, Weitz
Berg E, Nygren H, Sysi-Aho M, Griffin JL, KK, Shukla AK, Nguyen SN, Kim YM, Zink
Fiehn O, Loibl S, Richter-Ehrenstein C, EM, Luders T, Frevert CW, Gharib SA,
Radke C, Hyotylainen T, Kallioniemi O, Laskin J, Carson JP, Metz TO, Corley RA,
Iljin K, Oresic M (2011) Novel theranostic Ansong C (2017) Lipidomics reveals dramatic
opportunities offered by characterization of lipid compositional changes in the maturing
 Analyzing Untargeted Lipidomics 135

postnatal lung. Sci Rep 7:40555. https://doi. purification of total lipides from animal tissues.
org/10.1038/srep40555 J Biol Chem 226(1):497–509
10. Quehenberger O, Armando AM, Brown AH, 20. Bligh EG, Dyer WJ (1959) A rapid method of
Milne SB, Myers DS, Merrill AH, total lipid extraction and purification. Can J
Bandyopadhyay S, Jones KN, Kelly S, Shaner Biochem Physiol 37(8):911–917. https://doi.
RL, Sullards CM, Wang E, Murphy RC, Bark- org/10.1139/o59-099
ley RM, Leiker TJ, Raetz CR, Guan Z, Laird 21. Matyash V, Liebisch G, Kurzchalia TV,
GM, Six DA, Russell DW, McDonald JG, Shevchenko A, Schwudke D (2008) Lipid
Subramaniam S, Fahy E, Dennis EA (2010) extraction by methyl-tert-butyl ether for high-
Lipidomics reveals a remarkable diversity of throughput lipidomics. J Lipid Res 49
lipids in human plasma. J Lipid Res 51 (5):1137–1146. https://doi.org/10.1194/jlr.
(11):3299–3305. https://doi.org/10.1194/ D700041-JLR200
jlr.M009449 22. Liebisch G, Vizcaino JA, Kofeler H,
11. van Meer G, de Kroon AI (2011) Lipid map of Trotzmuller M, Griffiths WJ, Schmitz G,
the mammalian cell. J Cell Sci 124(Pt 1):5–8. Spener F, Wakelam MJ (2013) Shorthand
https://doi.org/10.1242/jcs.071233 notation for lipid structures derived from
12. Hu T, Zhang JL (2018) Mass-spectrometry- mass spectrometry. J Lipid Res 54
based lipidomics. J Sep Sci 41(1):351–372. (6):1523–1530. https://doi.org/10.1194/jlr.
https://doi.org/10.1002/jssc.201700709 M033506
13. Hyotylainen T, Oresic M (2015) Optimizing 23. Han X (2016) Lipidomics for studying metab-
the lipidomics workflow for clinical studies— olism. Nat Rev Endocrinol 12(11):668–679.
practical considerations. Anal Bioanal Chem https://doi.org/10.1038/nrendo.2016.98
407(17):4973–4993. https://doi.org/10. 24. Grosch S, Schiffmann S, Geisslinger G (2012)
1007/s00216-015-8633-2 Chain length-specific properties of ceramides.
14. Hyotylainen T, Oresic M (2016) Bioanalytical Prog Lipid Res 51(1):50–62. https://doi.org/
techniques in nontargeted clinical lipidomics. 10.1016/j.plipres.2011.11.001
Bioanalysis 8(4):351–364. https://doi.org/ 25. Veldhuizen R, Nag K, Orgeig S, Possmayer F
10.4155/bio.15.244 (1998) The role of lipids in pulmonary surfac-
15. Kyle JE, Crowell KL, Casey CP, Fujimoto GM, tant. Biochim Biophys Acta 1408
Kim S, Dautel SE, Smith RD, Payne SH, Metz (2–3):90–108
TO (2017) LIQUID: an-open source software 26. Clair G, Reehl S Stratton KG, Monroe ME,
for identifying lipids in LC-MS/MS-based lipi- Tfaily MM, Ansong C, Kyle JE (2019) Lipid
domics data. Bioinformatics 33 Mini-On: mining and ontology tool for enrich-
(11):1744–1746. https://doi.org/10.1093/ ment analysis of lipidomic data. Bioinformatics
bioinformatics/btx046 35(2):4507–4508. https://doi.org/10.1093/
16. Misra BB, Mohapatra S (2019) Tools and bioinformatics/btz250
resources for metabolomics research commu- 27. Kyle JE, Clair G, Bandyopadhyay G, Misra RS,
nity: a 2017-2018 update. Electrophoresis 40 Zink EM, Bloodsworth KJ, Shukla AK, Du Y,
(2):227–246. https://doi.org/10.1002/elps. Lillis J, Myers JR (2018) Cell type-resolved
201800428 human lung lipidome reveals cellular coopera-
17. Fahy E, Subramaniam S, Murphy RC, tion in lung function. Sci Rep 8(1):13455.
Nishijima M, Raetz CR, Shimizu T, Spener F, https://doi.org/10.1038/s41598-018-
van Meer G, Wakelam MJ, Dennis EA (2009) 31640-x
Update of the LIPID MAPS comprehensive 28. Kyle JE, Burnum-Johnson KE (2019) Plasma
classification system for lipids. J Lipid Res 50 lipidome reveals critical illness and recovery
(Suppl):S9–S14. https://doi.org/10.1194/jlr. from human Ebola virus disease. Proc Natl
R800095-JLR200 Acad Sci U S A 116(9):3919–3928. https://
18. Fahy E, Subramaniam S, Brown HA, Glass CK, doi.org/10.1073/pnas.1815356116
Merrill AH Jr, Murphy RC, Raetz CR, Russell 29. Harayama T, Riezman H (2018) Understand-
DW, Seyama Y, Shaw W, Shimizu T, Spener F, ing the diversity of membrane lipid composi-
van Meer G, VanNieuwenhze MS, White SH, tion. Nat Rev Mol Cell Biol 19(5):281–296.
Witztum JL, Dennis EA (2005) A comprehen- https://doi.org/10.1038/nrm.2017.138
sive classification system for lipids. J Lipid Res 30. van Meer G, Voelker DR, Feigenson GW
46(5):839–861. https://doi.org/10.1194/jlr. (2008) Membrane lipids: where they are and
E400004-JLR200 how they behave. Nat Rev Mol Cell Biol 9
19. Folch J, Lees M, Sloane Stanley GH (1957) A (2):112–124. https://doi.org/10.1038/
simple method for the isolation and nrm2330
136 Jennifer E. Kyle

31. Ghosh A, Nishtala K (2017) Biofluid lipidome: https://doi.org/10.1016/j.chemphyslip.

a source for potential diagnostic biomarkers. 2015.05.005
Clin Transl Med 6(1):22. https://doi.org/10. 41. Wiesner P, Leidl K, Boettcher A, Schmitz G,
1186/s40169-017-0152-7 Liebisch G (2009) Lipid profiling of FPLC-
32. Borghini I, Barja F, Pometta D, James RW separated lipoprotein fractions by electrospray
(1995) Characterization of subpopulations of ionization tandem mass spectrometry. J Lipid
lipoprotein particles isolated from human cere- Res 50(3):574–585. https://doi.org/10.
brospinal fluid. Biochim Biophys Acta 1255 1194/jlr.D800028-JLR200
(2):192–200 42. Hodson L, Skeaff CM, Fielding BA (2008)
33. Koch S, Donarski N, Goetze K, Kreckel M, Fatty acid composition of adipose tissue and
Stuerenburg HJ, Buhmann C, Beisiegel U blood in humans and its use as a biomarker of
(2001) Characterization of four lipoprotein dietary intake. Prog Lipid Res 47(5):348–380.
classes in human cerebrospinal fluid. J Lipid https://doi.org/10.1016/j.plipres.2008.03.
Res 42(7):1143–1151 003
34. Mahley RW (2016) Central nervous system 43. Sales S, Graessler J, Ciucci S, Al-Atrib R,
lipoproteins: ApoE and regulation of choles- Vihervaara T, Schuhmann K, Kauhanen D,
terol metabolism. Arterioscler Thromb Vasc Sysi-Aho M, Bornstein SR, Bickle M, Cannis-
Biol 36(7):1305–1315. https://doi.org/10. traci CV, Ekroos K, Shevchenko A (2016) Gen-
1161/atvbaha.116.307023 der, contraceptives and individual metabolic
35. Harrington MG, Fonteh AN, Oborina E, predisposition shape a healthy plasma lipidome.
Liao P, Cowan RP, McComb G, Chavez JN, Sci Rep 6:27710. https://doi.org/10.1038/
Rush J, Biringer RG, Huhmer AF (2009) The srep27710
morphology and biochemistry of nanostruc- 44. Chua EC, Shui G, Lee IT, Lau P, Tan LC, Yeo
tures provide evidence for synthesis and signal- SC, Lam BD, Bulchand S, Summers SA,
ing functions in human cerebrospinal fluid. Puvanendran K, Rozen SG, Wenk MR, Gooley
Cerebrospinal Fluid Res 6(10). https://doi. JJ (2013) Extensive diversity in circadian regu-
org/10.1186/1743-8454-6-10 lation of plasma lipids and evidence for differ-
36. Bouatra S, Aziat F, Mandal R, Guo AC, Wilson ent circadian metabolic phenotypes in humans.
MR, Knox C, Bjorndahl TC, Krishnamurthy R, Proc Natl Acad Sci U S A 110
Saleem F, Liu P, Dame ZT, Poelzer J, Huynh J, (35):14468–14473. https://doi.org/10.
Yallou FS, Psychogios N, Dong E, Bogumil R, 1073/pnas.1222647110
Roehring C, Wishart DS (2013) The human 45. Yuana Y, Sturk A, Nieuwland R (2013) Extra-
urine metabolome. PLoS One 8(9):e73076. cellular vesicles in physiological and pathologi-
https://doi.org/10.1371/journal.pone. cal conditions. Blood Rev 27(1):31–39.
0073076 https://doi.org/10.1016/j.blre.2012.12.002
37. Dashti M, Kulik W, Hoek F, Veerman EC, 46. Huang d W, Sherman BT, Lempicki RA (2009)
Peppelenbosch MP, Rezaee F (2011) A phos- Systematic and integrative analysis of large gene
pholipidomic analysis of all defined human lists using DAVID bioinformatics resources.
plasma lipoproteins. Sci Rep 1:139. https:// Nat Protoc 4(1):44–57. https://doi.org/10.
doi.org/10.1038/srep00139 1038/nprot.2008.211
38. Kim SH, Yang JS, Lee JC, Lee JY, Lee JY, 47. Huang da W, Sherman BT, Lempicki RA
Kim E, Moon MH (2018) Lipidomic altera- (2009) Bioinformatics enrichment tools:
tions in lipoproteins of patients with mild cog- paths toward the comprehensive functional
nitive impairment and Alzheimer’s disease by analysis of large gene lists. Nucleic Acids Res
asymmetrical flow field-flow fractionation and 37(1):1–13. https://doi.org/10.1093/nar/
nanoflow ultrahigh performance liquid gkn923
chromatography-tandem mass spectrometry. J 48. Luo W, Brouwer C (2013) Pathview: an
Chromatogr A 1568:91–100. https://doi. R/Bioconductor package for pathway-based
org/10.1016/j.chroma.2018.07.018 data integration and visualization. Bioinfor-
39. Kontush A, Lhomme M, Chapman MJ (2013) matics 29(14):1830–1831. https://doi.org/
Unraveling the complexities of the HDL lipi- 10.1093/bioinformatics/btt285
dome. J Lipid Res 54(11):2950–2963. 49. Luo W, Pant G, Bhavnasi YK, Blanchard SG Jr,
https://doi.org/10.1194/jlr.R036095 Brouwer C (2017) Pathview Web: user friendly
40. Serna J, Garcia-Seisdedos D, Alcazar A, Lasun- pathway visualization and data integration.
cion MA, Busto R, Pastor O (2015) Quantita- Nucleic Acids Res 45(W1):W501–w508.
tive lipidomic analysis of plasma and plasma https://doi.org/10.1093/nar/gkx372
lipoproteins using MALDI-TOF mass spec-
trometry. Chem Phys Lipids 189:7–18.
 Analyzing Untargeted Lipidomics 137

50. Pedersen HK, Forslund SK, Walters KB, Watanabe T, Fukuyama S,

Gudmundsdottir V, Petersen AO, Yamashita M, Jacobs JM, Kim YM, Casey CP,
Hildebrand F, Hyotylainen T, Nielsen T Stratton KG, Webb-Robertson BM, Gritsenko
(2018) A computational framework to inte- MA, Monroe ME, Weitz KK, Shukla AK,
grate high-throughput ‘-omics’ datasets for Tian M, Neumann G, Reed JL, van Bakel H,
the identification of potential mechanistic Metz TO, Smith RD, Waters KM, N’Jai A,
links. Nat Protoc 13(12):2781–2800. Sahr F, Kawaoka Y (2017) Multi-platform
https://doi.org/10.1038/s41596-018-0064- ‘Omics analysis of human Ebola virus disease
z pathogenesis. Cell Host Microbe 22
51. Eisfeld AJ, Halfmann PJ, Wendler JP, Kyle JE, (6):817–829.e818. https://doi.org/10.
Burnum-Johnson KE, Peralta Z, Maemura T, 1016/j.chom.2017.10.011
 Chapter 8

Overview of Tandem Mass Spectral and Metabolite

Databases for Metabolite Identification in Metabolomics
Zhangtao Yi and Zheng-Jiang Zhu

Abstract
Liquid chromatography–mass spectrometry (LC-MS) is one of the most popular technologies in metabo-
lomics. The large-scale and unambiguous identification of metabolite structures remains a challenging task
in LC-MS based metabolomics. Tandem mass spectral databases provide experimental and in silico MS/MS
spectra to facilitate the identification of both known and unknown metabolites, which has become a gold
standard method in metabolomics. In addition, metabolite knowledge databases offer valuable biological
and pathway information of metabolites. In this chapter, we have briefly reviewed the most common and
important tandem mass spectral and metabolite databases, and illustrated how they could be used for
metabolite identification.

Key words Metabolite identification, Metabolomics, Tandem mass spectrum, Metabolite database

1 Introduction

Metabolomics comprehensively measures metabolites in a

biological system to provide mechanistic insights of biological and
disease-related events. With technology advances in mass spectrom-
etry, hundreds to thousands of metabolites can now be quantita-
tively profiled from biological samples. Liquid chromatography–
mass spectrometry (LC-MS) is one of the most popular technolo-
gies in metabolomics due to its high sensitivity, selectivity, and
throughput. However, the major bottleneck of metabolomics has
been the challenge of determining the structural identities of the
MS peaks found to be dysregulated in the metabolomic profiling
[1, 2]. The use of accurate mass for identifying metabolites in
LC-MS based metabolomics is highly challenging. The problem is
that there are many possible molecular formulas within the toler-
ance, and numerous structural isomers for each molecular formula.
The alternative approach which matches the accurate precursor
mass (MS1) and tandem mass spectrum (MS/MS or MS2) with
those from the standard spectral library to identify metabolites has

Shuzhao Li (ed.), Computational Methods and Data Analysis for Metabolomics, Methods in Molecular Biology, vol. 2104,
https://doi.org/10.1007/978-1-0716-0239-3_8, © Springer Science+Business Media, LLC, part of Springer Nature 2020

139
140 Zhangtao Yi and Zheng-Jiang Zhu

Table 1
Statistics of tandem mass spectral databases

Compounds Number of In silico Free Free

Database with spectra Spectra spectra Search download
NIST 17 13,808 574,826
$ $
METLIN >200,000 NA √ √

MoNA ~75,000 261,917 √ √ √
mzCloud > 8000 ~2,000,000 a

√

HMDB 2265 22,247 b

√ √
LipidBlast 119,200 212,516 √ – √
a
The number of recalibrated spectra in mzCloud
b
Total experimental LC-MS/MS spectra in HMDB

become the “gold” standard method [1]. This method requires the
availability of standard MS/MS spectra. In past decades, many
efforts have been made to expand the existing tandem mass spectral
databases such as METLIN [3], NIST [4], and MassBank [5]
(Table 1). Clearly, all mass spectral databases are hindered by the
lack of chemical standards for many cellular metabolites, and suffer
from having uncharacterized spectral variations across different
LC-MS instruments, acquisition condition, and laboratories.
Other efforts have also been made to theoretically predict the
MS/MS spectra in silico [6, 7]. Limited by the high diversity of
chemical structures of metabolites and the relatively small size of
training dataset, the accuracy for the in silico prediction approach
still requires a substantial improvement. Molecular and metabolic
pathways can also be utilized for metabolite identification by
mapping dysregulated metabolic features into the metabolic net-
work, such as Mummichog [8] and PIUMet [9]. In this approach,
metabolic pathway databases are generally required.
Many metabolomics-related databases have been constructed
and are either freely accessible or commercially available to provide
chemical structures, physiochemical properties, spectral informa-
tion, biological functions, and pathway information. In this chap-
ter, we have provided an overview of current databases for
metabolomics, especially the supportive databases for metabolite
identification (Table 1), including tandem mass spectral databases
like NIST, METLIN, MassBank, and mzCloud; metabolite knowl-
edge databases such as HMDB; and metabolic pathway databases
such as KEGG [10], MetaCyc [11] and Reactome [12]. However,
common chemical databases such as PubChem and ChemSpider
are not covered in this chapter.
 Tandem Mass Spectral and Metabolite Databases 141

2 METLIN

METLIN was first developed in 2003 by Scripps Center for Meta-

bolomics led by Prof. Gary Siuzdak [13] and is one of the largest
tandem mass spectral databases for metabolites. METLIN now
covers more than 300,000 compounds (data from Chapter 10),
including metabolites, drugs, toxicants, peptides and so
on. Metadata in METLIN includes compound name, structure,
formula, mass, CAS number, KEGG ID, HMDB ID, and Pub-
Chem ID. The commercial availabilities of compounds are also
provided. Most importantly, over 200,000 compounds have either
experimental or in silico-predicted tandem mass spectra. Among
them, experimental tandem mass spectra were acquired from
authentic chemical standards. All experimental tandem mass spectra
were collected under a strict experimental protocol on high-
resolution Agilent QTOF instruments (6500 series). MS2 spectra
were acquired using four collision energy levels (0, 10, 20, and
40 eV) on both positive and negative modes. The MS/MS spectra
at 0 eV represent the in-source fragmentation, and facilitate the
annotation of in-source fragments in untargeted metabolomics. All
MS2 spectra in METLIN were curated and recalibrated. The possi-
ble chemical structures were also assigned to product ion using the
bioinformatic tool MetFrag. Originally, METLIN only contains the
experimental tandem mass spectra, but recently it also incorporated
the in silico-predicted MS2 spectra. The in silico prediction of MS2
spectra were performed by inputoutput kernel regression
(IOKR)-based machine learning algorithm [3]. The experimental
MS2 spectra in METLIN were used for training the machine
learning algorithm. METLIN is freely available through the web
server (https://metlin.scripps.edu) with a user-friendly interface.
METLIN web server has provided multiple searching capabilities
including single, batch, precursor ion, neutral loss, accurate mass,
and fragment searches. MS/MS spectral match in METLIN was
based on X-Rank similarity algorithm, which provides the robust
identification result when matching experimental spectra from dif-
ferent instrument platforms. The only drawback of METLIN is that
the MS2 spectra are not downloadable. For a more detailed descrip-
tion of METLIN, please refer to Chapter 10.

3 NIST 17 Tandem Mass Spectral Libraries

NIST mass spectral libraries are developed by the NIST mass spec-
trometry data center led by Dr. Stephen E. Stein (Gaithersburg,
MD, USA). The NIST spectral libraries aim to facilitate chemical
compound identification in GC- and LC-MS by providing standard
mass spectra. These libraries were originally developed as GC-MS
142 Zhangtao Yi and Zheng-Jiang Zhu

spectral libraries (by electron ionization) and later incorporated

LC-MS based tandem mass spectra. NIST Tandem mass spectral
library were acquired using a range of fragmentation conditions
including positive and negative modes from ion-trap and collision
cells such as QTOF, Orbitrap, and QqQ. The tandem mass spectra
include both low-resolution ones from QqQ and ion trap, and
high-resolution ones from QTOF and Orbitrap, and cover a wide
range of compounds including small molecules and peptides, so
NIST libraries are not specific for metabolites. The NIST tandem
mass spectral libraries are updated every 3 years, and NIST17 is the
latest version (https://chemdata.nist.gov/dokuwiki/doku.php?
id¼chemdata:nist17).
NIST 17 tandem spectral libraries contain a total of 13,808
compounds (118,082 precursor ions, and 574,826 tandem spec-
tra), and 1904 biologically relevant peptides in an additional library.
Each compound entry in the library contains compound name,
formula, CAS number, InChIKey, MS/MS spectra, data acquisi-
tion condition, and so on. MS/MS spectra were collected from
multiple collision energy levels over a wide range of instrument
platforms, mainly from Thermo LTQ Orbitrap Elite, Agilent
QTOF 6530, Micromass Quattro Micro QqQ, Thermo Finnigan
LTQ Orbitrap Velos, and so on. For LTQ Orbitrap instruments,
spectra from both HCD and IT/ion trap collision type were
provided. NIST 17 also includes tandem mass spectra for multiple
common ion adducts, such as [M+H]+, [M+H-H2O]+, [M+Na]+,
and [M+NH4]+ in positive mode and [M-H] and [M-H-H2O]
in negative mode.
For compound identification, NIST provides the MS search
software (now version 2.3) to search the libraries. With the search
software, one can browse or search against the libraries by exact
mass, formula, precursor or product ion mass, CAS number, and so
on. However, for batch query, one has to use the other software MS
PepSearch. Mass spectra in libraries can also be individually
exported as msp format, and used in other software tools for
metabolite identification. Notably, NIST 17 has also included a
new search method, hybrid similarity search (HSS) [14]. HSS is
very useful to annotate new chemical structures that do not have
tandem mass spectra in the libraries, because it allows matching the
experimental spectrum to a library spectrum with a neutral loss
using the m/z differences between their precursor ions. In sum-
mary, NIST 17 tandem mass spectral libraries feature a broad
coverage of small molecules, and MS2 spectra from various types
of ion adducts with different collision energy levels and instru-
ments. However, NIST 17 is not freely available, and it lacks
biological information or links to metabolite databases for metabo-
lomics research. NIST 17 should be used as a tool to elucidate
compound structures.
 Tandem Mass Spectral and Metabolite Databases 143

4 MassBank and MoNA

MassBank was originally designed as an open-source and public

mass spectral repository to collect and store freely available tandem
mass spectra. MassBank web server was first developed in Japan [5]
(2010, http://www.massbank.jp) and later as a mirror in Europe
(2012, https://massbank.eu/MassBank). Tandem mass spectra in
MassBank were contributed from different research groups with
author names tagged to each entry. MassBank allows the users to
freely download and upload mass spectra. MassBank now encom-
passes a total of experimental mass spectra of around 40,000 CID
spectra, 13,000 EI spectra and other types of spectra. CID spectra
were mostly acquired from TOF, QqQ, and ion trap instruments.
In 2013, an R package RMassBank [15] was developed as a recali-
bration tool for uploaded spectra.
Recently, MoNA (MassBank of North America) has been
developed by Prof. Oliver Fiehn Lab in University of California-
Davis, and incorporated all spectra in MassBank. In addition,
MoNA also collected tandem mass spectra from other databases
such as ReSpect (http://spectra.psc.riken.jp), GNPS (https://
gnps.ucsd.edu), and MetaboBASE (from Bruker), together with
Fiehn lab in-house libraries such as Fiehn HILIC, and Vaniya/
Fiehn natural products library. In silico spectral libraries for lipids,
such as LipidBlast, were also included in MoNA. Therefore, MoNA
is currently the most comprehensive tandem spectral database in
the world. Now, MoNA has 122,171 experimental spectra, of
which 96,996 spectra are specified as LC-MS spectra, and
139,746 in silico spectra. MoNA covered MS2 spectra for common
ion adducts obtained from both QTOF and Orbitrap instruments.
Each entry in MoNA contains the submitter, mass spectrum, com-
pound name, downloadable 2D structure, classification provided
by ClassyFire [16], InChIKey, InChI, SMILES, and metadata
including instrument, chromatographic condition, and so on.
MoNA web server allows both “quick search” by exact mass
and “similarity search” by uploading an MS/MS spectrum (e.g., in
an MGF file). When one is browsing the mass spectrum in MoNA,
it also allows for searching “similar spectra” in the databases. All
spectra from MoNA are freely downloaded as either JSON or msp
format. Distribution of most spectra in MoNA is under license of
Creative Commons by Attribution (CC-BY). The downloaded
spectra can be imported into other software tools such as NIST
MS search software and MS-DIAL for metabolite identification.
Each spectrum in MoNA is given an identifier—SPLASH [17],
which is a hashed string that provides an unambiguous, database-
independent identifier. The use of SPLASH allows for the quick
and accurate query and cross-reference of the spectra across differ-
ent databases.
144 Zhangtao Yi and Zheng-Jiang Zhu

5 mzCloud

mzCloud (https://www.mzcloud.org) is a commercial tandem

mass spectral database hosted by HighChem and has a focus on
spectra acquired only from Orbitrap instruments. Both MS/MS
and multistage MSn spectra were acquired using chemical standards
at various collision energies using both collision-induced dissocia-
tion (CID) and higher-energy collisional dissociation (HCD). Cur-
rently, mzCloud includes 23,800 spectra trees covering 209,012
precursors from over 8000 compounds. All raw tandem mass spec-
tra were postprocessed by filtering and recalibration to improve the
spectral quality. Fragment ions in tandem mass spectra were anno-
tated through applying the general fragmentation rules and frag-
mentation mechanisms published in peer-reviewed journals,
manual evaluation, and ab initio quantum chemical computation.
mzCloud can be freely accessed through the web server. It supports
MS/MS spectral similarity match in single or batch mode for
metabolite identification. For compounds not included in the data-
base, mzCloud enables the de novo metabolite structural elucida-
tion via identifying the possible substructures in the query tandem
mass spectrum through comparing the product ion spectra of
structurally related compounds. However, all tandem mass spectra
in mzCloud cannot be downloaded, and all spectra are exclusively
from Orbitrap instruments.

6 HMDB

HMDB (human metabolome database, http://www.hmdb.ca) was

developed by the Canada metabolomics innovation center led by
Prof. David Wishart. HMDB is the most comprehensive and Homo
sapiens-only metabolite knowledge database. HMDB is now
updated to version 4.0 [18] and covers a total of 114,100 metabo-
lites. Each metabolite entry includes massive metadata information
such as name, formula, MDL Molfile, InChI, SMILES, chemical
taxonomy, physiological data, normal and abnormal concentra-
tions, related genomic and enzymatic information, pathway infor-
mation, links to other databases such as PubChem, ChemSpider,
KEGG, BioCyc, and related spectral characterizations. Only a small
portion of metabolites in HMDB have provided the tandem mass
spectra. Specifically, there are 22,247 experimental tandem spectra
from 2265 compounds in HMDB. Most of these were imported
from MassBank and other spectral databases. In HMDB 4.0
release, in silico-predicted MS2 spectra obtained from CFM-ID
[19] were also incorporated. In addition to tandem mass spectra,
HMDB also includes NMR spectra for metabolite identification.
The HMDB provides many different search functions, including
 Tandem Mass Spectral and Metabolite Databases 145

ChemQuery structure search, molecular weight search, and

LC-MS/MS search. For a more detailed description of HMDB,
please refer to Chapter 11. All information in HMDB can be freely
downloaded.

7 LIPID MAPS and LipidBlast

LIPID MAPS is a lipid knowledge database developed by The

LIPID MAPS consortium [20, 21]. It includes several different
databases including structure database (LMSD), gene/proteome
database (LMPD), in silico structure database (LMISSD), and
database of computationally generated bulk lipid species
(COMP_DB). Among them, LMSD is the most important lipid
database with a total of 43,403 biologically relevant lipids from
eight categories, namely, fatty acyls (FA), glycerolipids (GL), gly-
cerophospholipids (GP), sphingolipids (SP), sterol lipids (ST), pre-
nol lipids (PR), saccharolipids (SL), and polyketides (PK). Each
category also includes various subclassification hierarchies. Each
lipid in LMSD has a unique LIPID MAPS identification num-
ber—“LM_ID” identifier [22]. One could search the lipid class
via classification-based, text/ontology-based, and structure-based
searches. LMISSD is an in silico-generated lipid structure databases
through the theoretical and computational expanding of head-
groups and chains from different lipid classes accounting for a
total number of lipids >one million. Each entry in LMSD and
LMISSD has a LIPID MAPS ID, MDL Molfile, SMILES, InChI,
InChIKey, exact mass, formula, systematic name, and a set of
abbreviations for sum composition, chain composition, and exact
structure. All of this data can be freely downloaded as structure-
data file (SDF) format.
For some lipids in LMSD, experimental tandem mass spectra
were collected using chemical standards. However, these spectra are
only provided as images and cannot be used for spectral similarity
match. The experimental spectra for lipids are not as important as
metabolites, since fragmentation of lipids are well studied and easier
to be predicted and simulated. There are many computational tools
to generate in silico-predicted MS2 spectra for lipids such as Lipid-
Blast [23] and in silico prediction tools in LIPID MAPS. LipidBlast
is one of the largest stand-alone in silico tandem spectral databases
for lipid identification, which covers a total of 212,516 predicted
spectra of 119,200 compounds from 26 lipid compound classes.
Spectra from common ion adducts such as [M+H]+, [M+Na]+, [M
+NH4]+,and [M-H] are covered. LipidBlast can be freely down-
loaded (https://fiehnlab.ucdavis.edu/projects/LipidBlast). Other
databases such as MoNA and LMSD have incorporated the spectra
in LipidBlast. LipidBlast can be imported to other MS/MS search
tool such as NIST MS Search and NIST MS PepSearch for lipid
identification.
146 Zhangtao Yi and Zheng-Jiang Zhu

8 Metabolic Pathway Databases

The metabolic pathway database contains information about meta-

bolic pathways, metabolic reactions and related genes and enzymes
[24]. All of the information is important to interpret the biological
functions of metabolites that are found to be dysregulated in meta-
bolomics experiment. In addition, several important software tools
such as Mummichog [8], PIUMet [9], and MetDNA [25] also use
the metabolic pathway and metabolic reactions for metabolite
identification.
Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway
database (https://www.genome.jp/kegg/) is one of the most
curated and comprehensive pathway databases, which contains a
collection of metabolic pathway maps representing the knowledge
on the molecular interaction, reaction, and relation networks for
metabolism. Currently, KEGG includes a total of 530 pathway
maps from 491 eukaryotes and 5379 prokaryotes, composed of
18,505 metabolites and other small compounds and 11,146 reac-
tions. Massive genomic and enzymatic information are also
incorporated. Every metabolite in KEGG has a KEGG ID and
contains name, classification, reaction, pathway, reference, and so
on. All data can be freely searched in KEGG web server or via
KEGG API (https://www.kegg.jp/kegg/rest/).
MetaCyc (https://MetaCyc.org) is a curated metabolic path-
way database with 2666 metabolic pathways from 2960 different
organisms. It contains experimental data about chemical com-
pounds, reactions, enzymes, and metabolic pathways curated
from more than 54,000 publications, making it the largest collec-
tion and reference database of metabolic pathways. Data in Meta-
Cyc can be searched in web server by keyword, ontology, identifier,
and so on. All data can be freely downloaded for local analysis or
imported into other database. Each metabolite entry has InChI,
InChIKey, SMILES, standard Gibbs free energy, reaction, enzy-
matic regulation information, and references.
Reactome (https://reactome.org/) is a free, open-source path-
way database and provides tools for the visualization, interpretation
and analysis of pathway knowledge. It contains around 20,000
pathways from different species covering around 80,000 reactions.
Genomic and enzymatic information are also included. Reactome
provides a pathway browser in which the pathway can be freely
browsed in an expandable hierarchy or browsed by typing a key-
word. Each pathway entry contains compounds and proteins
involved. All data and software tools in Reactome are freely avail-
able for download.
Despite the increase of pathway knowledge, there are still some
missing pathways and metabolites. To complement the gap, in
silico metabolite databases such as MINEs [26] and
 Tandem Mass Spectral and Metabolite Databases 147

MyCompoudID [27] provide computationally generated meta-

bolic network and reaction information derived from known meta-
bolites and biochemical reactions. This enlarged pathway
information can further expand the metabolite knowledge data-
bases and facilitate annotating new metabolites.

Acknowledgments

The work has been supported by National Key R&D Program of

China (2018YFA0800902), National Natural Science Foundation
of China (Grants 21575151) and Chinese Academy of Sciences
Major Facility-based Open Research Program. Z.-J. Z. is supported
by Thousand Youth Talents Program.

References

1. Vinaixa M, Schymanski EL, Neumann S, 8. Li S, Park Y, Duraisingham S, Strobel FH,

Navarro M, Salek RM, Yanes O (2016) Mass Khan N, Soltow QA, Jones DP, Pulendran B
spectral databases for LC/MS- and GC/MS- (2013) Predicting network activity from high
based metabolomics: state of the field and throughput metabolomics. PLoS Comput Biol
future prospects. Trends Anal Chem 78:23–35 9:e1003123
2. Domingo-Almenara X, Montenegro-Burke JR, 9. Pirhaji L, Milani P, Leidl M, Curran T, Avila-
Benton HP, Siuzdak G (2018) Annotation: a Pacheco J, Clish CB, White FM, Saghatelian A,
computational solution for streamlining meta- Fraenkel E (2016) Revealing disease-associated
bolomics analysis. Anal Chem 90:480–489 pathways by network integration of untargeted
3. Guijas C, Montenegro-Burke JR, Domingo- metabolomics. Nat Methods 13:770–776
Almenara X, Palermo A, Warth B, 10. Kanehisa M, Furumichi M, Tanabe M, Sato Y,
Hermann G, Koellensperger G, Huan T, Morishima K (2016) KEGG: new perspectives
Uritboonthai W, Aisporna AE et al (2018) on genomes, pathways, diseases and drugs.
METLIN: a technology platform for identify- Nucleic Acids Res 45:D353–D361
ing knowns and unknowns. Anal Chem 11. Caspi R, Billington R, Fulcher CA, Keseler IM,
90:3156–3164 Kothari A, Krummenacker M, Latendresse M,
4. Yang X, Neta P, Liang Y, Stein SE (2017) Midford PE, Ong Q, Ong WK et al (2018) The
Extending a comprehensive reference tandem MetaCyc database of metabolic pathways and
mass spectral library for more reliable metabo- enzymes. Nucleic Acids Res 46:D633–D639
lite identification. 65th Annual ASMS confer- 12. Fabregat A, Jupe S, Matthews L,
ence on mass spectrometry and allied topics, Sidiropoulos K, Gillespie M, Garapati P,
Indianapolis, Indiana, June 4–8, 2017 Haw R, Jassal B, Korninger F, May B (2017)
5. Horai H, Arita M, Kanaya S, Nihei Y, Ikeda T, The reactome pathway knowledgebase.
Suwa K, Ojima Y, Tanaka K, Tanaka S, Nucleic Acids Res 46:D649–D655
Aoshima K (2010) MassBank: a public reposi- 13. Smith CA, O’Maille G, Want EJ, Qin C, Trau-
tory for sharing mass spectral data for life ger SA, Brandon TR, Custodio DE,
sciences. J Mass Spectrom 45:703–714 Abagyan R, Siuzdak G (2005) METLIN: a
6. Kind T, Tsugawa H, Cajka T, Ma Y, Lai Z, metabolite mass spectral database. Ther Drug
Mehta SS, Wohlgemuth G, Barupal DK, Sho- Monit 27:747–751
walter MR, Arita M et al (2018) Identification 14. Burke MC, Mirokhin YA, Tchekhovskoi DV,
of small molecules using accurate mass Markey SP, Heidbrink Thompson J, Larkin C,
MS/MS search. Mass Spectrom Rev Stein SE (2017) The hybrid search: a mass
37:513–532 spectral library search method for discovery of
7. Blazenovic I, Kind T, Ji J, Fiehn O (2018) modifications in proteomics. J Proteome Res
Software tools and approaches for compound 16:1924–1935
identification of LC-MS/MS data in metabo- 15. Stravs MA, Schymanski EL, Singer HP, Hol-
lomics. Meta 8:31 lender J (2013) Automatic recalibration and
148 Zhangtao Yi and Zheng-Jiang Zhu

processing of tandem mass spectra using for- 22. Fahy E, Subramaniam S, Murphy RC,
mula annotation. J Mass Spectrom 48:89–99 Nishijima M, Raetz CR, Shimizu T, Spener F,
16. Feunang YD, Eisner R, Knox C, Chepelev L, van Meer G, Wakelam MJ, Dennis EA (2009)
Hastings J, Owen G, Fahy E, Steinbeck C, Update of the LIPID MAPS comprehensive
Subramanian S, Bolton E (2016) ClassyFire: classification system for lipids. J Lipid Res 50:
automated chemical classification with a com- S9–S14
prehensive, computable taxonomy. J Chem 23. Kind T, Liu KH, Lee DY, DeFelice B, Meissen
8:61 JK, Fiehn O (2013) LipidBlast in silico tandem
17. Wohlgemuth G, Mehta SS, Mejia RF, mass spectrometry database for lipid identifica-
Neumann S, Pedrosa D, Pluskal T, Schymanski tion. Nat Methods 10:755–758
EL, Willighagen EL, Wilson M, Wishart DS 24. Fiehn O, Barupal DK, Kind T (2011) Extend-
et al (2016) SPLASH, a hashed identifier for ing biochemical databases by metabolomic
mass spectra. Nat Biotechnol 34:1099–1101 surveys. J Biol Chem 286:23637–23643
18. Wishart DS, Feunang YD, Marcu A, Guo AC, 25. Shen X, Wang R, Xiong X, Yin Y, Cai Y, Ma Z,
Liang K, Vázquez-Fresno R, Sajed T, Liu N, Zhu Z-J (2019) Metabolic reaction
Johnson D, Li C, Karu N et al (2018) network-based recursive metabolite annotation
HMDB 4.0: the human metabolome database for untargeted metabolomics. Nat Commun
for 2018. Nucleic Acids Res 46:D608–D617 10:1516
19. Allen ‘F, Greiner R, Wishart D (2014) Com- 26. Jeffryes JG, Colastani RL, Elbadawi-Sidhu M,
petitive fragmentation modeling of ESI-MS/ Kind T, Niehaus TD, Broadbelt LJ, Hanson
MS spectra for putative metabolite identifica- AD, Fiehn O, Tyo KE, Henry CS (2015)
tion. Metabolomics 11:98–110 MINEs: open access databases of computation-
20. Sud M, Fahy E, Cotter D, Brown A, Dennis ally predicted enzyme promiscuity products for
EA, Glass CK, Merrill AH Jr, Murphy RC, untargeted metabolomics. J Chem 7:44
Raetz CR, Russell DW (2006) LMSD: lipid 27. Huan T, Tang C, Li R, Shi Y, Lin G, Li L
maps structure database. Nucleic Acids Res (2015) MyCompoundID MS/MS search:
35:D527–D532 metabolite identification using a library of pre-
21. Fahy E, Sud M, Cotter D, Subramaniam S dicted fragment-ion-spectra of 383,830 possi-
(2007) LIPID MAPS online tools for lipid ble human metabolites. Anal Chem
research. Nucleic Acids Res 35:W606–W612 87:10619–10626
 Chapter 9

METLIN: A Tandem Mass Spectral Library of Standards

J. Rafael Montenegro-Burke, Carlos Guijas, and Gary Siuzdak

Abstract
Untargeted mass spectrometry metabolomics studies rely on accurate databases for the identification of
metabolic features. Leveraging unique fragmentation patterns as well as characteristic dissociation routes
allows for structural information to be gained for specific metabolites and molecular classes, respectively.
Here we describe the evolution of METLIN as a resource for small molecule analysis as well as the tools
(e.g., Fragment Similarity Search and Neutral Loss Search) used to query the database and their workflows
for the identification of molecular entities. Additionally, we will discuss the functionalities of isoMETLIN, a
database of isotopic metabolites, and the latest addition to the METLIN family, METLIN-MRM, which
facilitates the analysis of quantitative mass spectrometry data generated with triple quadrupole
instrumentation.

Key words Untargeted metabolomics, Spectral database, MS/MS spectra, Metabolite identification

1 Introduction

As the interest of the scientific community fell on endogenous

metabolites almost 20 years ago due to their implications in diag-
nostics and biomarker discovery, technical challenges had to be
overcome in order to use mass spectrometry or metabolomics for
these biomedical purposes. Particularly, chromatographic data pro-
cessing and metabolite characterization/identification of untar-
geted metabolomic experiments represented the biggest obstacles
for researchers worldwide [1]. While the development of chro-
matographic data processing tools is covered in this book, this
chapter will focus on the metabolite database METLIN and its
associated tools, from inception with a few hundred metabolites
to the most current version with over ~600,000 compounds
(November 2019).
METLIN was developed to assist metabolomics research across
a broad range of disciplines, especially to facilitate the identification

J. Rafael Montenegro-Burke and Carlos Guijas contributed equally to this work.

Shuzhao Li (ed.), Computational Methods and Data Analysis for Metabolomics, Methods in Molecular Biology, vol. 2104,
https://doi.org/10.1007/978-1-0716-0239-3_9, © Springer Science+Business Media, LLC, part of Springer Nature 2020

149
150 J. Rafael Montenegro-Burke et al.

of metabolites in LC-MS based metabolomics and bridge the gap

to other “omics” sciences such as genomics and proteomics
[1]. The molecular or macromolecular identification process to
assign names and annotations to genes and proteins from genomics
and proteomics experiments has been made possible not only due
to great efforts in their respective communities but also due to the
predicative sequence of nucleosides and amino acids. On the other
hand, the large chemical diversity found in the metabolome as well
as the vast number of different molecular entities limits the predic-
tion of fragmentation patterns (MS/MS) in MS-based experiments
thereby limiting the ability to make identifications [2].
METLIN was first developed in 2003 and contained a few
hundred metabolites and MS/MS spectra [1]. However, its growth
in terms of the number of metabolites and MS/MS spectra but also
the incorporation of new tools has been continuous. For example,
after it was made publicly available in 2005, METLIN grew from a
few hundred to more than 10,000 metabolites with their respective
MS/MS spectra by 2012 [3]. Additionally, tools to facilitate and
automate the identification of known and unknown metabolites
have been integrated [4–6].
Concomitantly, other academic, public, and private entities
have developed metabolite/small molecule databases, which
could be classified into two categories: (1) pathway-centric and
(2) compound-centric [7]. METLIN belongs to the latter as it
contains chemical structures, spectral profiles and its tools are
designed to facilitate metabolite identification.
Similar databases containing MS/MS spectra include Mass-
Bank, HMDB, GNPS, MoNA, LIPID MAPS, NIST 14, and
mzCloud [8–12]. For a more detailed comparison between these
databases we direct the reader to the work of Vinaixa et al.
[2]. Briefly some of the main differences are data collection and
curation processes, instrumentation used to generate data, standard
reference materials, types of molecules, source and origin of mole-
cule, and accessibility. In METLIN, MS/MS spectra have been
acquired at the Scripps Center for Metabolomics following strict
protocols and only using standard reference material that is either
commercially available or has been synthesized. Early on in the
development process, it was decided against the use of complex
biological samples for reference fragmentation spectra in order to
avoid interfering molecules and provide the best spectral quality.
This is unlike other databases where MS/MS spectra have been
acquired in different laboratories under different conditions and
instrumentation. Over the years, METLIN has also incorporated
metabolites from a wide range of molecular classes and biological or
chemical origins. Therefore, the broad spectrum of small molecules
available in METLIN include endogenous metabolites such as
lipids, amino acids, nucleotides and carbohydrates; toxicants have
recently been included to assist exposome research [13]. Further,
 METLIN: A Metabolite Mass Spectral Database 151

drugs and drug secondary metabolites can be found in METLIN as

well, as the use of untargeted metabolomics in drug development
has gained traction in recent years [14].

2 Tools for the Identification of Known and Unknown Metabolites

2.1 Basic Search In a traditional untargeted metabolomics workflow, dysregulated

Engines features (i.e., specific m/z values with their respective retention time
values) will be searched against a database for putative metabolite
identifications based on accurate mass. METLIN provides three
search options: (1) Simple, (2) Batch, and (3) Advanced Search
(Fig. 1). Simple Search allows the search of both m/z values and
neutral masses within a selected mass tolerance. Several adducts in
both positive and negative polarities can be selected. Batch Search
has the same functions and capabilities as Simple Search, however,
several m/z values can be searched simultaneously, facilitating the
annotation of adducts and common losses that stem from the same
metabolite. Similarly, ions with a different molecular origin can be
easily distinguished and linked to other putative m/z values with
this search feature. Advanced Search, on the other hand, allows the
user to search based on other metabolite information such as
molecular formula, metabolite names, SMILES and KEGG, CAS
and MID numbers. These searches provide a list of metabolite
names (or molecular formulas) that could potentially correspond
to the dysregulated feature of interest. However, given the low
elemental diversity in biomolecules (C, H, N, O, P, and S), this
list can contain tens to hundreds of putative metabolite
identifications.

2.2 Identification In order to reduce the list of putative identifications obtained from
of Known Metabolites: m/z value searching, experimental MS/MS spectra are compared
MS/MS Spectrum against spectral libraries in terms of fragmentation patterns (m/z of
Match Search fragments and their intensities) (Fig. 1). MS/MS Spectrum Match
Search is a tool that allows the autonomous identification of meta-
bolites. Here, users upload the fragmentation profile as a table of
m/z values and intensities, and enter the mass of the precursor with
a specific tolerance, collision energy, and polarity. This tool then
searches, compares, and scores the similarity of the experimental
spectra with the reference spectra in the library for all metabolites
within the selected mass tolerance, relying on a modified X-Rank
similarity algorithm [15]. Without a doubt the ~300,000 molecu-
lar standards with MS/MS spectra are METLIN’s most valuable
contribution to metabolomics research. The fragmentation spectra
for these ~600,000 molecules have been acquired at four collision
energies (0, 10, 20, and 40 V) in both positive- and negative-ion
mode. While low collision energy spectra were previously not
extensively used in the identification process, they have been
152 J. Rafael Montenegro-Burke et al.

Fig. 1 METLIN search functions for small molecule identification. (a) Simple and Batch Search allow users to
search small molecules against a database of 1 million compounds based on both m/z values and neutral
masses within a selected mass tolerance. Advanced Search allows searches based on metabolite information
such as molecular formula, metabolite names, SMILES and KEGG, and CAS and MID numbers. (b) With the
MS/MS Spectrum Match Search, experimental and library MS/MS spectra can be searched, matched, and
scored in an automatic way. (c) Fragment Similarity Search and Neutral Loss Search aid the identification of
metabolites or chemical structures by searching m/z values of the fragments or neutral losses, respectively,
regardless of the precursor mass. Analytical Chemistry 2018, 90, 3156–3164. (Figure 1, with permission from
ACS and RightsLink)

recently used as a second layer in metabolite annotation algorithms

and data deconvolution [16]. The combination of METLIN’s large
spectral library and annotation tools significantly simplifies data
analysis and improves the confidence of putative identifications.

2.3 Identification Interestingly, spectral libraries serve a dual purpose in the identifi-
of Unknown cation/characterization of metabolic features. As mentioned above,
Metabolites the main use of these libraries is to compare experimental spectra
with reference spectra for the purpose of identification (up to level
2 according to the Metabolomics Standards Initiative)
[2, 17]. However, given the large number of metabolites and the
broad range of chemistries, no library is complete despite consider-
able efforts dedicated to increasing their populations. Furthermore,
the number of metabolites in nature still is a subject of debate but
estimates in the million range are not hyperbolized. Such numbers
dwarf the ~20,000 genes and proteins (without taking
 METLIN: A Metabolite Mass Spectral Database 153

posttranslational modifications into account). Thus, the second

purpose of spectral libraries is to aid the identification/characteri-
zation of known metabolites without MS/MS spectra available and
unknown metabolites, whose structures have not been previously
described in any library or resource. For this particular purpose,
two tools, Fragment Similarity Search and Neutral Loss Search,
have been developed and continually refined over the last 11 years
[4]. These tools leverage the large number of spectral information
and the dissociation routes similarities between compounds with
related structures and chemical moieties.

2.3.1 Fragment Similarity The Fragment Similarity Search algorithm was originally imple-
and Neutral Loss Search mented into METLIN to find chemical similarities between the
desired unknown features and the known molecules available in the
library based on the experimental MS/MS acquired by the user and
the over 4 million high-resolution MS/MS spectra in the METLIN
library. Specifically, the Fragment Similarity Search algorithm relies
on a shared peak count method and facilitates the identification of
the molecule of interest or molecular class by prioritizing molecules
with a larger chemical fragment overlap [4].
The Neutral Loss Search algorithm was designed as a comple-
mentary tool to Fragment Similarity Search. While in Fragment
Similarity Search shared fragments (i.e., ions with the same m/z
value) provide structural information of the ion of interest, similar
structures with different molecular formulas, ergo different masses
and m/z values, would generate fragments with different masses
and no similarity would be determined. However, in Neutral Loss
Search, mass differences between precursor ions and their frag-
ments ions can provide structural information based on common
“leaving groups” in their respective dissociation routes. Both these
tools leverage the vast number of carefully curated MS/MS spectra
from a wide range of small molecular entities in the METLIN
library to facilitate the identification of not only known molecules
with no available MS/MS spectra, but also of unknown molecules
which have not been previously described in any form.
Here, we provide an example using these tools to identify
unknown metabolites detected in a murine macrophage cell line
(RAW264.7). The analysis was performed using an I-class UPLC
system coupled to a Synapt G2-Si mass spectrometer (Waters Corp.
Milford, MA). After the unsuccessful identification of a metabolic
feature using accurate mass and MS/MS spectra matching proce-
dures described in Subheadings 2.1 and 2.2 respectively, Fragment
Similarity Search and Neutral Loss Search tools were employed to
gain structural information and therefore clues to molecular iden-
tity in the following steps:
154 J. Rafael Montenegro-Burke et al.

Fig. 2 Fragment Similarity Search facilitates the identification of unknown metabolites where no MS/MS
spectral data are available. The fragments of an unknown metabolite were searched against METLIN and all of
the four fragments were found to match with methionine sulfoxide. The comparison between experimental
and library MS/MS spectra implies high structural similarities

1. Selected fragments from the murine macrophage metabolite

MS/MS spectrum were searched against the METLIN
MS/MS spectral database using Fragment Similarity Search
tool. High intensity fragments are more likely to provide struc-
tural information of the metabolic feature, and in some cases,
these can be very specific to the metabolite of interest. On the
other hand, lower mass fragments are commonly shared with
many molecules, making it harder to gain useful structural
information from their analysis. Additionally, larger errors in
the mass accuracy are inherent for low mass fragments given the
mass accuracy definition [18]. In this case, we selected the
fragments 149.02, 102.06, 74.02, and 56.05 from the
unknown feature of interest for analysis, which resulted in
many hits to molecules in METLIN. However, Fragment Sim-
ilarity Search orders the results based on the number of match-
ing fragments. Methionine sulfoxide (Met sulfoxide) matched
4 of the 4 fragments searched, hinting at a metabolite contain-
ing such a chemical structure or moieties (Fig. 2).
2. In order to gain more structural information, other fragments
(higher mass) were searched as described above. The higher
mass fragment searches did not yield any results that matched a
particular compound or molecular class.
3. Assuming that Met sulfoxide with a monoisotopic neutral mass
of 165.05 is part of the molecule of interest and we can observe
a prominent fragment with the mass 166.05, the next step is to
use the mass of the molecule to identify what chemical struc-
tures could constitute the rest of the molecule. After calculat-
ing the mass difference between the precursor ion and the
potential methionine sulfoxide fragment
(295.10–166.05 ¼ 129.04), we use the Neutral Loss Search
 METLIN: A Metabolite Mass Spectral Database 155

tool and search for 129.04 within a selected ppm window.

Most of the Neutral Loss Search results consist of molecules
that contain glutamic acid (Glu), suggesting Glu is a second
piece of the unknown metabolite.
4. After using Fragment Similarity Search and Neutral Loss
Search, we have gathered some information about the possible
chemical structures of the metabolite of interest. We then
deduced how the two pieces might be connected. A plausible
link between methionine sulfoxide and glutamic acid is an
amide bond, which would form the dipeptide Glu Met
sulfoxide.
5. Advanced Search can be used to search for such a peptide in the
METLIN database by name and molecular formula. Unfortu-
nately, this peptide is not available in the library. However, a
similar dipeptide containing Glu and methionine (Met) is avail-
able with MS/MS spectra. While the only difference between
the structures is the oxidation of the sulfur atom in methionine,
such a small modification can have considerable differences in
the dissociation routes and those differences should be
considered.
6. The MS/MS spectra between the metabolite of interest (puta-
tively Glu Met sulfoxide) and of the dipeptide Glu Met are
compared next to try to further characterize the metabolite
identity. Several fragments with a mass difference of an oxygen
atom (15.99) are observed and were attributed to the oxida-
tion of the sulfur atom in methionine. An additional feature of
METLIN is the MetFrag algorithm developed by S. Neumann
and coworkers, which indicates putative structures for the dif-
ferent fragment ions in the MS/MS library [19]. Based on
MetFrag’s analysis, we saw that several of the predicted struc-
tures that have a mass difference of 15.99 actually correspond
to fragments that contain the thiol moiety. This further
increases the confidence in the putative identification of the
unknown feature as Glu Met sulfoxide (Fig. 3).

3 METLIN Family

3.1 isoMETLIN In parallel to the development of METLIN, other related databases

have been released relying in the growing number of analytical
standards available in METLIN to reach new goals in the field of
metabolomics. isoMETLIN was implemented in 2014 as a database
for isotope-based metabolomics [20]. With an analogous interface
to METLIN, isoMETLIN provides accurate mass of all computed
isotopologs accrued in METLIN, compounds with a different
number of isotope-labeled atoms and consequently, different m/z
values. isoMETLIN Search includes the most commonly used
156 J. Rafael Montenegro-Burke et al.

Fig. 3 Neutral loss search aides the identification of unknown metabolites based on mass differences between
precursor ions and their fragments ions (i.e., common “leaving groups”). A Neutral Loss Search of 129.04
(295.10–166.05 ¼ 129.04) yielded 168 results, where ~70% contain a glutamic acid moiety. Based on the
masses of Met sulfoxide and Glu, a dipeptide is a likely option. Such a peptide is not available in the library;
however, a similar compound Glu Met contains MS/MS spectra for comparison. Several fragments, which
contain the thiol moiety, have a mass difference of 15.99 (monoisotopic mass of oxygen atom) corresponding
to the oxidation of sulfur in methionine

Fig. 4 isoMETLIN Simple Search menu allows searching the m/z of isotopologs within a selected mass
tolerance. Type of isotopic labeling, ion charge, and ion adducts can also be selected. This search renders a
list of all possible metabolites taking all possible isotopic combinations into account. Additionally, MS/MS
spectra can be accessed when available

stable isotopes in labelling experiments, such as 13C, 15N, 2H, and

18
O. Even though isotopologs can be discerned by accurate mass
measurements, the further analysis of their MS/MS spectra is nec-
essary to determine the position of the isotopic label within the
same isotopologs, a pivotal feature for the investigation of meta-
bolic pathways [21, 22]. To accomplish this, isoMETLIN also
incorporates the MS/MS spectra for hundreds of isotopomers
(same isotopolog with different location of labeled atoms) to help
in tracing the isotopic label, providing a vast amount of information
about the de novo synthesis of metabolites in certain pathways
(Fig. 4). Although the principal application of isoMETLIN’s
 METLIN: A Metabolite Mass Spectral Database 157

Fig. 5 Systematic generation of MS/MS spectra of uniformly labeled metabolites in isoMETLIN. Pichia pastoris
was grown in unlabeled and 13C-labeled glucose, producing uniformly labeled metabolites after several
generations. Unlabeled and labeled metabolite extracts were analyzed by high-resolution untargeted meta-
bolomics to generate pairs of unlabeled and labeled putative metabolites. Finally, the MS/MS spectra of
identified pairs was incorporated into isoMETLIN after a careful curation

capabilities is the analysis of metabolic fluxes, which constitutes an

emerging “omics” discipline by itself, other uses of isoMETLIN
include isotope dilution quantitative metabolomics and identifica-
tion of compounds, as is shown below.

3.1.1 Untargeted Similar to METLIN, isoMETLIN fragmentation spectra was initi-

Generation of MS/MS Data ally acquired on qToF instruments at different collision energies,
Using Isotope-Labeled using authentic isotope-labeled standards. However, a major short-
Microorganisms coming in populating a spectral database with MS/MS spectra of
metabolite isotopologs is that the number of isotopomers increases
with molecular weight (number of atoms) and that most isotopo-
mers are not commercially available. To address this limitation, a
novel approach using uniformly labeled microorganisms was
recently developed [23]. Briefly, two metabolite extracts were gen-
erated by growing Pichia pastoris in 12C-glucose- and 13C-glucose-
containing media (Fig. 5a). Before the acquisition of the fragmen-
tation spectra, the labeling efficiency of the yeast metabolites was
verified to be above 99%. These two unlabeled and uniformly
labeled metabolite extracts were analyzed by different LC-MS plat-
forms in an untargeted way. The unlabeled metabolites were
grouped with all their isotopologs containing the isotopic trace
and the MS/MS spectra of hundreds of fully labeled metabolites
was generated (Fig. 5). It is worth mentioning that the reduced
cost and time efforts are crucial advantages of this systematic work-
flow of generating MS/MS spectra for several hundreds of isotopic-
labeled metabolites.

3.1.2 Identification An additional application of fully labeled MS/MS spectra is the gain
of Metabolites Using of structural information for metabolite identification. By lever-
Isotopes aging the mass differences between analogous fragments of isoto-
pologs, the number of C atoms in each fragment can be
158 J. Rafael Montenegro-Burke et al.

Fig. 6 (a) The MS/MS of an unknown metabolite is matched with the MS/MS of the Pichia pastoris extract
unlabeled compound. To consider the Pichia pastoris pair of compounds for identification, the retention time of
the unknown metabolite should match under the same analytical conditions. The number of carbons of each
fragment can be used to determine the structures of both parent and fragment ions. In this example,
pseudouridine was identified. (b) Proposed parallel analysis of microorganisms labeled with other stable
isotopes. MS/MS spectra of fully labeled known metabolites can be incorporated into isoMETLIN, whereas
MS/MS spectra of unknown metabolites can be used for their identification as described in (a)

determined. This information is of great interest for the identifica-

tion of metabolites, whose MS/MS is not available, as it is the case
in the identification of pseudouridine (Fig. 6a). Furthermore, two
additional examples of identified metabolites using uniformly
labeled Pichia pastoris can be found in the work of Guijas et al. [23].
Currently, these spectra are being incorporated into isoME-
TLIN to aid researchers in the process of identifying of metabolites
and the incorporation of analogous data generated by the analysis
of microorganisms uniformly labeled with other isotopes, such as
15
N and 34S is a future goal. Not only will this allow the addition of
new MS/MS spectra of uniformly labeled metabolites containing
these two atoms, but it will also generate new layers of MS/MS
spectra that can be complementary to the information provided by
the 13C-labeled metabolites for the identification of metabolites
(Fig. 6b).
 METLIN: A Metabolite Mass Spectral Database 159

Fig. 7 METLIN-MRM ensembles three types of transitions: (1) transitions experi-

mentally optimized by standard materials with QqQ via an established protocol,
(2) transitions computationally (comp.) optimized by using METLIN’s MS/MS
spectra, and (3) and public repository transitions. Experimentally and computa-
tionally optimized transitions were optimized for sensitivity and selectivity,
respectively. METLIN-MRM also serves as a public repository (PR), in which
the community can populate the library with new transitions

3.2 METLIN-MRM The METLIN-MRM library consists of a small-molecule transi-

tions for multiple-reaction monitoring (MRM) for more than
15,500 unique small molecules. It was developed to streamline
absolute quantitation, which typically is accomplished with triple-
quadrupole (QqQ) mass spectrometers configured to monitor a
particular set of precursor–product ion transitions. However, in
order to determine the transitions for the different molecules of
interest, each target molecule must be optimized with pure stan-
dard materials. In the library, three different types of transitions are
available: (1) traditional experimentally optimized transitions,
(2) computationally optimized experimental transitions, and
(3) public repository transitions (Fig. 7). Experimentally optimized
transitions were acquired for more than 1000 molecules in both
positive and negative mode by following established protocols
[24]. These small molecule transitions were optimized for the
highest intensity to achieve low limits of detection.
In addition to experimentally acquired data, transitions for
more than 14,000 and 4700 molecules in positive and negative
mode, respectively, were computationally optimized by using the
METLIN spectral library [23] (acquired at different CE on a qToF
instrument) by ranking empirical MS/MS fragments according to
their selectivity (uniqueness of a product fragment for a given
molecule). The developed ranking algorithm compares the
MS/MS spectra of compounds with precursors within a 0.7 Da
160 J. Rafael Montenegro-Burke et al.

window and selects fragments with the best selectivity without

compromising sensitivity. This strategy enables high-throughput
quantitation analysis as transitions are no longer required to be
optimized with standard reference materials and, by the same
token, minimizes errors caused by interfering molecules as transi-
tions less likely to be masked are selected. For more detailed infor-
mation about the algorithms and the selection process, we refer the
readers to the main published work [25].
Lastly, METLIN-MRM also serves as a public repository, in
which community members can upload transitions. Submitted lists
of transitions are assigned with a unique accession number that can
be used as a reference for publications. This process ultimately
facilitates the deposition of transitions used in experiments and
scientific literature into a standardized and searchable database,
thereby increasing the traceability and reproducibility of experi-
ments and the reuse/sharing of optimized transitions. Currently,
3300 transitions for more than 1500 small molecules are available
from peer-reviewed publications with their corresponding original
source (doi).

4 METLIN Population

METLIN contains experimental tandem mass spectrometry data

on over 600,000 molecular standards and also has structures on
over a million small molecules corresponding to a wide range of
molecular classes. Over the years, small molecular entities have been
incorporated into the library without bias for a particular class or
type of compounds. Endogenous metabolites spanning the three-

Fig. 8 Experimental electrospray ionization tandem mass spectrometry data has

been generated on over ~300,000 molecular standards and incorporated into
METLIN. This data was generated in both positive and negative ionization modes
and at multiple collision energies (0, 10, 20, and 40 V). The rapid increase in the
rate of METLIN growth is the culmination of multiple analytical and informatic
challenges being overcome in 2018. Analytical Chemistry 2018,
90, 13128–13129. (Figure 1, with permission from ACS and RightsLink)
 METLIN: A Metabolite Mass Spectral Database 161

domain system (archaea, bacteria, and eukarya), modified metabo-

lites, synthetic drugs and toxicants. The growth of the MS/MS
database has been exponential in the last year, growing from
14,000 in 2017 to over 600,000 in November 2019 in both
positive- and negative-ion mode at multiple collision energies
(Fig. 8) [26]. A growth curve we anticipate will continue into the
future.

5 Conclusion

In this chapter, METLIN’s evolution since its beginnings, its tools

for metabolite identification, and future developments have been
discussed. It has adapted and developed its tools to not only facili-
tate the identification of known compounds (with or without
MS/MS spectra) but also the discovery of unknown compounds.
In the coming years, hundreds of thousands of small molecules will
be characterized using the same strict MS/MS protocols and data
curation. Further, given the ubiquitous coupling of preionization
techniques and mass spectrometry instrumentation for untargeted
metabolomics, tools for retention time prediction are currently
being developed. Ultimately, it is safe to say that METLIN’s goal
from its inception of facilitating metabolomics research has not
changed, although its library and tools have continuously evolve
to meet the changing needs of modern metabolomics research as
well as other chemical entities.

Acknowledgments

This research was partially funded by National Institutes of Health

grants R35 GM130385, P30 MH062261, P01 DA026146 and
U01 CA235493; and by Ecosystems and Networks Integrated with
Genes and Molecular Assemblies (ENIGMA), a Scientific Focus
Area Program at Lawrence Berkeley National Laboratory for the
U.S. Department of Energy, Office of Science, Office of Biological
and Environmental Research, under contract number DE-AC02-
05CH11231. This research benefited from the use of credits from
the National Institutes of Health (NIH) Cloud Credits Model
Pilot, a component of the NIH Big Data to Knowledge (BD2K)
program.
162 J. Rafael Montenegro-Burke et al.

References
1. Smith CA, Maille GO, Want EJ, Qin C, Trau- sciences. J Mass Spectrom 45(7):703–714.
ger SA, Brandon TR, Custodio DE, https://doi.org/10.1002/jms.1777
Abagyan R, Siuzdak G (2005) METLIN: a 9. Wishart DS, Tzur D, Knox C, Eisner R, Guo
metabolite mass spectral database. Ther Drug AC, Young N, Cheng D, Jewell K, Arndt D,
Monit 27(6):747–751. https://doi.org/10. Sawhney S, Fung C, Nikolai L, Lewis M, Cou-
1097/01.ftd.0000179845.53213.39 touly M-A, Forsythe I, Tang P, Shrivastava S,
2. Vinaixa M, Schymanski EL, Neumann S, Jeroncic K, Stothard P, Amegbey G, Block D,
Navarro M, Salek RM, Yanes O (2016) Mass Hau DD, Wagner J, Miniaci J, Clements M,
spectral databases for LC/MS- and GC/MS- Gebremedhin M, Guo N, Zhang Y, Duggan
based metabolomics: state of the field and GE, MacInnis GD, Weljie AM, Dowlatabadi R,
future prospects. TrAC Trends Anal Chem Bamforth F, Clive D, Greiner R, Li L,
78:23–35. https://doi.org/10.1016/j.trac. Marrie T, Sykes BD, Vogel HJ, Querengesser
2015.09.005 L (2007) HMDB: the human metabolome
3. Tautenhahn R, Cho K, Uritboonthai W, database. Nucleic Acids Res 35(suppl_1):
Zhu Z, Patti GJ, Siuzdak G (2012) An acceler- D521–D526. https://doi.org/10.1093/nar/
ated workflow for untargeted metabolomics gkl923
using the METLIN database. Nat Biotechnol 10. Wang M, Carver JJ, Phelan VV, Sanchez LM,
30:826. https://doi.org/10.1038/nbt.2348 Garg N, Peng Y, Nguyen DD, Watrous J,
4. Benton HP, Wong DM, Trauger SA, Siuzdak G Kapono CA, Luzzatto-Knaan T, Porto C,
(2008) XCMS2: processing tandem mass spec- Bouslimani A, Melnik AV, Meehan MJ, Liu
trometry data for metabolite identification and W-T, Crüsemann M, Boudreau PD,
structural characterization. Anal Chem Esquenazi E, Sandoval-Calderón M, Kersten
80:6382–6389. https://doi.org/10.1021/ RD, Pace LA, Quinn RA, Duncan KR, Hsu
ac800795f C-C, Floros DJ, Gavilan RG, Kleigrewe K,
5. Benton HP, Ivanisevic J, Mahieu NG, Kurczy Northen T, Dutton RJ, Parrot D, Carlson
ME, Johnson CH, Franco L, Rinehart D, EE, Aigle B, Michelsen CF, Jelsbak L,
Valentine E, Gowda H, Ubhi BK, Sohlenkamp C, Pevzner P, Edlund A,
Tautenhahn R, Gieschen A, Fields MW, Patti McLean J, Piel J, Murphy BT, Gerwick L,
GJ, Siuzdak G (2015) Autonomous metabolo- Liaw C-C, Yang Y-L, Humpf H-U,
mics for rapid metabolite identification in Maansson M, Keyzers RA, Sims AC, Johnson
global profiling. Anal Chem 87(2):884–891. AR, Sidebottom AM, Sedio BE, Klitgaard A,
https://doi.org/10.1021/ac5025649 Larson CB, Boya PCA, Torres-Mendoza D,
Gonzalez DJ, Silva DB, Marques LM, Demar-
6. Montenegro-Burke JR, Phommavongsay T, que DP, Pociute E, O’Neill EC, Briand E, Hel-
Aisporna AE, Huan T, Rinehart D, frich EJN, Granatosky EA, Glukhov E,
Forsberg E, Poole FL, Thorgersen MP, Ryffel F, Houson H, Mohimani H, Kharbush
Adams MWW, Krantz G, Fields MW, Northen JJ, Zeng Y, Vorholt JA, Kurita KL,
TR, Robbins PD, Niedernhofer LJ, Lairson L, Charusanti P, McPhail KL, Nielsen KF,
Benton HP, Siuzdak G (2016) Smartphone Vuong L, Elfeki M, Traxler MF, Engene N,
analytics: mobilizing the lab into the cloud for Koyama N, Vining OB, Baric R, Silva RR,
omic-scale analyses. Anal Chem 88 Mascuch SJ, Tomasi S, Jenkins S, Macherla V,
(19):9753–9758. https://doi.org/10.1021/ Hoffman T, Agarwal V, Williams PG, Dai J,
acs.analchem.6b02676 Neupane R, Gurr J, Rodrı́guez AMC,
7. Fiehn O, Barupal DK, Kind T (2011) Extend- Lamsa A, Zhang C, Dorrestein K, Duggan
ing biochemical databases by metabolomic BM, Almaliti J, Allard P-M, Phapale P, Nothias
surveys. J Biol Chem 286(27):23637–23643. L-F, Alexandrov T, Litaudon M, Wolfender
https://doi.org/10.1074/jbc.R110.173617 J-L, Kyle JE, Metz TO, Peryea T, Nguyen
8. Horai H, Arita M, Kanaya S, Nihei Y, Ikeda T, D-T, VanLeer D, Shinn P, Jadhav A,
Suwa K, Ojima Y, Tanaka K, Tanaka S, Müller R, Waters KM, Shi W, Liu X, Zhang L,
Aoshima K, Oda Y, Kakazu Y, Kusano M, Knight R, Jensen PR, Palsson BØ, Pogliano K,
Tohge T, Matsuda F, Sawada Y, Hirai MY, Linington RG, Gutiérrez M, Lopes NP, Ger-
Nakanishi H, Ikeda K, Akimoto N, Maoka T, wick WH, Moore BS, Dorrestein PC, Bandeira
Takahashi H, Ara T, Sakurai N, Suzuki H, N (2016) Sharing and community curation of
Shibata D, Neumann S, Iida T, Tanaka K, mass spectrometry data with global natural
Funatsu K, Matsuura F, Soga T, Taguchi R, products social molecular networking. Nat
Saito K, Nishioka T (2010) MassBank: a public Biotechnol 34:828. https://doi.org/10.
repository for sharing mass spectral data for life 1038/nbt.3597
 METLIN: A Metabolite Mass Spectral Database 163

11. Tsugawa H, Kind T, Nakabayashi R, (11):1821–1835. https://doi.org/10.1016/j.

Yukihira D, Tanaka W, Cajka T, Saito K, jasms.2010.06.006
Fiehn O, Arita M (2016) Hydrogen rearrange- 19. Wolf S, Schmidt S, Müller-Hannemann M,
ment rules: computational MS/MS fragmenta- Neumann S (2010) In silico fragmentation for
tion and structure elucidation using computer assisted identification of metabolite
MS-FINDER software. Anal Chem 88 mass spectra. BMC Bioinformatics 11(1):148.
(16):7946–7958. https://doi.org/10.1021/ https://doi.org/10.1186/1471-2105-11-
acs.analchem.6b00770 148
12. Sud M, Fahy E, Cotter D, Brown A, Dennis 20. Cho K, Mahieu N, Ivanisevic J,
EA, Glass CK, Merrill AH Jr, Murphy RC, Uritboonthai W, Chen Y Jr, Siuzdak G, Patti
Raetz CRH, Russell DW, Subramaniam S GJ (2014) isoMETLIN: a database for isotope-
(2006) LMSD: LIPID MAPS structure data- based metabolomics. Anal Chem 86
base. Nucleic Acids Res 35(suppl_1): (19):9358–9361. https://doi.org/10.1021/
D527–D532. https://doi.org/10.1093/nar/ ac5029177
gkl838 21. Kurczy ME, Forsberg EM, Thorgersen MP,
13. Warth B, Spangler S, Fang M, Johnson C, Poole FL, Benton HP, Ivanisevic J, Tran ML,
Forsberg E, Granados A, Domingo-Almenara- Wall JD, Elias DA, Adams MWW, Siuzdak G
X, Huan T, Rinehart D, Montenegro-Burke (2016) Global isotope metabolomics reveals
JR, Hilmers B, Aisporna AE, Hoang L, adaptive strategies for nitrogen assimilation.
Uritboonthai W, Benton HP, Richardson S, ACS Chem Biol 11(6):1677–1685. https://
Williams A, Siuzdak G (2017) Exposome- doi.org/10.1021/acschembio.6b00082
scale investigations guided by global metabo- 22. Badur MG, Metallo CM (2018) Reverse engi-
lomics, pathway analysis, and cognitive com- neering the cancer metabolic network using
puting. Anal Chem 89(21):11505–11513. in flux analysis to understand drivers of human
press disease. Metab Eng 45:95–108. https://doi.
14. Wishart DS (2016) Emerging applications of org/10.1016/j.ymben.2017.11.013
metabolomics in drug discovery and precision 23. Guijas C, Montenegro-Burke JR, Domingo-
medicine. Nat Rev Drug Discov 15:473. Almenara X, Palermo A, Warth B,
https://doi.org/10.1038/nrd.2016.32 Hermann G, Koellensperger G, Huan T,
15. Mylonas R, Mauron Y, Masselot A, Binz P-A, Uritboonthai W, Aisporna AE, Wolan DW,
Budin N, Fathi M, Viette V, Hochstrasser DF, Spilker ME, Benton HP, Siuzdak G (2018)
Lisacek F (2009) X-rank: a robust algorithm METLIN: a technology platform for identify-
for small molecule identification using tandem ing knowns and unknowns. Anal Chem 90
mass spectrometry. Anal Chem 81 (5):3156–3164. https://doi.org/10.1021/
(18):7604–7610. https://doi.org/10.1021/ acs.analchem.7b04424
ac900954d 24. Lange V, Picotti P, Domon B, Aebersold R
16. Domingo-Almenara X, Montenegro-Burke JR, (2008) Selected reaction monitoring for quan-
Guijas C, Majumder ELW, Benton HP, Siuz- titative proteomics: a tutorial. Mol Syst Biol 4
dak G (2019) Autonomous METLIN-guided (1):222. https://doi.org/10.1038/msb.
in-source fragment annotation for untargeted 2008.61
metabolomics. Anal Chem 91(5):3246–3253. 25. Domingo-Almenara X, Montenegro-Burke JR,
https://doi.org/10.1021/acs.analchem. Ivanisevic J, Thomas A, Sidibé J, Teav T,
8b03126 Guijas C, Aisporna AE, Rinehart D, Hoang L,
17. Sumner LW, Amberg A, Barrett D, Beale MH, Nordström A, Gómez-Romero M, Whiley L,
Beger R, Daykin CA, Fan TW-M, Fiehn O, Lewis MR, Nicholson JK, Benton HP, Siuzdak
Goodacre R, Griffin JL, Hankemeier T, G (2018) XCMS-MRM and METLIN-MRM:
Hardy N, Harnly J, Higashi R, Kopka J, Lane a cloud library and public resource for targeted
AN, Lindon JC, Marriott P, Nicholls AW, Reily analysis of small molecules. Nat Methods 15
MD, Thaden JJ, Viant MR (2007) Proposed (9):681–684. https://doi.org/10.1038/
minimum reporting standards for chemical anal- s41592-018-0110-3
ysis. Metabolomics 3(3):211–221. https://doi. 26. Guijas C, Siuzdak G (2018) Reply to comment
org/10.1007/s11306-007-0082-2 on METLIN: a technology platform for iden-
18. Brenton AG, Godfrey AR (2010) Accurate tifying knowns and unknowns. Anal Chem 90
mass measurement: terminology and treatment (21):13128–13129. https://doi.org/10.
of data. J Am Soc Mass Spectrom 21 1021/acs.analchem.8b04081
 Chapter 10

Metabolomic Data Exploration and Analysis with the Human

Metabolome Database
David S. Wishart

Abstract
The Human Metabolome Database (HMDB) is a comprehensive, online, digital database designed to
support the analysis and interpretation of metabolomic data acquired from human and/or mammalian
metabolomic studies. This chapter covers three methods or protocols pertinent to using the HMDB:
(1) understanding the general layout of the HMDB; (2) exploring the contents of a typical HMDB
“MetaboCard”; and (3) an example of how HMDB can be used in a metabolomics study on human
glioblastoma.

Key words Database, Metabolomics, Human, Disease, Data analysis

1 Introduction

The Human Metabolome Database (HMDB) is the primary vehicle

for relaying knowledge collected from the Human Metabolome
Project (HMP) to the public. The HMP was launched in 2005
[1] as a multi-institutional project based at the University of
Alberta in Canada. Now entering its 15th year of operation, the
HMP has been using advanced metabolomic techniques combined
with computer-aided literature/text mining to compile as much
information about the human metabolome as possible. Over the
past 15 years the HMP has experimentally characterized a number
of human biofluids and excreta, including the metabolome of
human cerebrospinal fluid [2], human serum [3], human urine
[4], human saliva [5], and human feces [6]. This information
along with other information compiled from computer-aided text
mining and manual curation has been used to assemble a much
more extensive collection of human metabolites for many more
tissues and biofluids. This information is periodically released to
the public in the form an online database, known as the HMDB,
which is located at www.hmdb.ca [7, 8].

Shuzhao Li (ed.), Computational Methods and Data Analysis for Metabolomics, Methods in Molecular Biology, vol. 2104,
https://doi.org/10.1007/978-1-0716-0239-3_10, © Springer Science+Business Media, LLC, part of Springer Nature 2020

165
166 David S. Wishart

Simply stated, the HMDB is an open-access, web-enabled

database that contains detailed information about essentially all
known human metabolites. This detailed information includes
fully referenced data about their biological roles, physiological
concentrations, disease associations, chemical reactions, metabolic
pathways, and reference mass spectrometry (MS) or nuclear mag-
netic resonance (NMR) spectra. Over the past 12 years [7], the
HMDB has grown by a factor of nearly 50X and has rapidly evolved
to meet the changing needs of both the metabolomics and clinical
research communities. The most recent release of the HMDB
(HMDB 4.0) contains 114,100 metabolites that are grouped into
4 major classes: (1) Detected and quantified, (2) Detected but not
quantified, (3) Expected, and (4) Predicted [8]. “Detected” meta-
bolites are those with measured concentrations or experimental
confirmation of their existence in human biofluids, cells, or tissues.
“Expected” metabolites are those with a known structure for which
biochemical pathways are known and where human intake/expo-
sure is frequent, but the compound has yet to be detected in the
body. “Predicted” compounds represent a small number of highly
feasible and biologically reasonable metabolites that have been
generated through computational or “in silico” biotransformations
using the program known as BioTransformer [9].
The HMDB is widely used by the metabolomics community
and receives >5 million web hits each year. It is frequently accessed
by researchers wishing to learn more about the structure, nomen-
clature, functions, or pathways of identified metabolites in both
human and mammalian metabolomic studies. The HMDB is also
used to obtain referential concentrations for medical diagnostic
work, to identify compounds through spectral comparisons, and
to learn more about the pathways or processes that involve meta-
bolites identified in metabolomic experiments. Even though the
HMDB explicitly includes the word “human” in its title, almost all
of the metabolites and much of the information about these meta-
bolites can be readily used when studying the metabolome of other
mammalian models (mouse, rat, dog, cat, cow, etc.).
This chapter provides an overview of the HMDB layout, its
contents, and an example of how to use it in a metabolomic study.

2 Materials

The HMDB is an online, web-accessible database. It is reachable via

any modern computer or web-enabled device (smart phone or
tablet) equipped with a reasonably up-to-date web browser (Goo-
gle Chrome, Windows Internet Explorer, Apple Safari, Mozilla
Firefox, or Opera) and a moderately high-resolution screen.
 Human Metabolome Database 167

3 Methods

Four methods are covered here: (1) exploring the general layout of
the HMDB; (2) exploring the contents of a typical HMDB “Meta-
boCard”; and (3) an example of how HMDB can be used in
metabolomics study on human glioblastoma. All activities can be
done with almost any internet-compatible device although a larger
screen (18 cm or larger) is preferable.

3.1 Exploring the 1. Open a preferred web browser and enter the following URL
HMDB Layout (web address): http://www.hmdb.ca.
2. The HMDB home page should be visible (Fig. 1).
3. On the top right is a Search box with a pull-down menu (with
options for metabolites, diseases, pathways, proteins, reactions)
and a blue search button. Users may enter text into the search
box to search any of the categories in the pull-down menu.
Enter the name “Alanine” into the search box and press the
search button (or press the return key). A “Search Results”
page should appear listing more than 100 matches to Alanine
with the most likely match appearing at the top of the list
(Fig. 2). See Note 1 regarding a more detailed explanation of
what is seen in this page.
4. Press the “back” button on your browser or the “HMDB” icon
(top left of the Search Results page) to return to the HMDB
home page.

Fig. 1 A screenshot of the HMDB home page

168 David S. Wishart

Fig. 2 A screenshot of the “Search Results” page for a search using Alanine as a text entry

5. Use a mouse or computer track pad to select “Browse” on the

HMDB home page menu. A list of 14 browsing options will be
displayed (Fig. 3). These are described in more detail in Note
2. Selecting any of these options will generate a specially for-
matted page that allows users to freely browse the data in each
of these categories.
6. User a mouse or computer track pad to select “Search” on the
HMDB home page menu. A list of 11 search or query func-
tions will be displayed (Fig. 4). These are described in more
detail in Note 3. Selecting any of these options will generate a
specially formatted page that allows users to search the HMDB
according to the selected search type.
7. Use a mouse or computer track pad to select the “Downloads”
option on the HMDB home page menu. Selecting this will
generate a formatted, scrollable list of different files (for differ-
ent versions of the HMDB) that can be downloaded to a local
computer. These include protein and gene sequences
(in FASTA format), chemical structures (in SDF format),
metabolite/protein data (in XML format) and MS/NMR
spectra (in XML format). The size of each file is displayed on
the right. Clicking the dark blue “Download” button initiates
the file download to your computer or device.
 Human Metabolome Database 169

Fig. 3 A screenshot of the HMDB home page with the “Browse” menu displayed

Fig. 4 A screenshot of the HMDB home page with the “Search” menu displayed
170 David S. Wishart

8. Press the “back” button on your browser or the “HMDB” icon

(top left of the Search Results page) to return to the HMDB
home page.
9. User a mouse or computer track pad to select “About” on the
HMDB home page menu. A list of 9 useful links concerning
the HMDB will be displayed. These are described in more
detail in Note 4. The function of the remaining links on the
HMDB home page are either self-evident or redundant and
will not be discussed here. Users are encouraged to click on the
menu options found under “Search” and “Browse” to further
explore each of the searching and browsing functions in
the HMDB.
10. This completes the HMDB overview.

3.2 Exploring the 1. Go to the HMDB home page (http://www.hmdb.ca) and use
Content of the HMDB your mouse or track pad to select “Browse” on the HMDB
MetaboCard home page menu.
2. Select “Metabolites” from the pull-down menu (it is at the top
of the pull-down list).
3. Scroll down the page, move the mouse pointer to the name of
the first compound in the table (1-methylhistidine) and click
on it. Users may alternately click on the tan-colored HMDB
identifier (HMDB0000001) as this will yield the same result.
4. A “MetaboCard” for 1-Methylhistidine should appear that
shows the Record Information along with some of the Metab-
olite identification data (Fig. 5). Additional details about the
MetaboCard concept are given in Note 5.
5. Use a mouse or computer track pad to scroll down the page to
view more of the Metabolite Identification data (see Note 6).
6. Use a mouse or computer track pad to scroll down the page to
view more of the Chemical Taxonomy data (Fig. 6). Additional
information about this section is given in Note 7.
7. Use a mouse or computer track pad to scroll down the page to
view more of the Ontology data (Fig. 7). Additional informa-
tion about this section is given in Note 8.
8. Use a mouse or computer track pad to scroll down the page to
view more of the Physical Property data (Fig. 8).
9. Use a mouse or computer track pad to scroll down the page to
view more of the Spectral data and Biological Property data
(Fig. 9). Additional information about these sections is given in
Note 9.
10. Use a mouse or computer track pad to scroll down the page to
view data on the Normal Concentrations, Abnormal
 Human Metabolome Database 171

Fig. 5 A screenshot of the upper portion of the 1-Methylhistidine MetaboCard

Fig. 6 A screenshot of the Chemical Taxonomy data field from the 1-Methylhistidine MetaboCard
172 David S. Wishart

Fig. 7 A screenshot of the Ontology data field from the 1-Methylhistidine MetaboCard

Fig. 8 A screenshot of the Physical Property data field from the 1-Methylhistidine MetaboCard
 Human Metabolome Database 173

Fig. 9 A screenshot of the Spectral data and Biological Property data fields from the 1-Methylhistidine
MetaboCard

Concentrations, and Associated Disorders and Diseases

(Fig. 10). More information about these sections is given in
Note 10.
11. Use a mouse or computer track pad to scroll down the page to
view External Links and References (Fig. 11).
12. Use a mouse or computer track pad to scroll down the page to
view the Enzymes page (Fig. 12). More information about the
enzymes is given in Note 11.
13. This completes the HMDB MetaboCard overview.

3.3 Using the HMDB 1. Go to the HMDB home page (http://www.hmdb.ca) and use
for Disease Studies your mouse or track pad to select “Search” on the HMDB
(Glioblastoma) home page menu.
2. The drop-down menu will display several search options
including structure, sequence, text, spectral and molecular
weight searches. For this example select the “LC-MS Search”
option (Fig. 13).
3. As explained in Note 12, the following m/z values were found
to be significantly altered in cerebrospinal fluid samples
between patients with glioblastoma (brain cancer) and healthy
controls: 149.0444, 117.0183, 119.0339, 91.0388,
147.0763, 193.0341, 308.0911. Enter these numbers into
the “Query Masses” box as shown in Fig. 14.
174 David S. Wishart

Fig. 10 A screenshot of the Normal and Abnormal Concentrations data fields from the 1-Methylhistidine
MetaboCard

Fig. 11 A screenshot of the External Links and References data fields from the 1-Methylhistidine MetaboCard
 Human Metabolome Database 175

Fig. 12 A screenshot of the Enzymes and Transporters page from the 1-Methylhistidine MetaboCard

Fig. 13 A screenshot of the “LC-MS Search” page from the HMDB

176 David S. Wishart

Fig. 14 A screenshot of the “LC-MS Search” page with the 149.0444, 117.0183, 119.0339, 91.0388,
147.0763, 193.0341, 308.0911 m/z values entered into the “Query Masses” box

4. Ensure the Ion Mode is marked positive and the molecular

weight tolerance is set to 0.001 Da. Choose the adduct type
to be “M + H”. Press the “Search” button.
5. After a few seconds, a list of the top scoring hits will appear as
shown in Fig. 15. See Note 13 for a more detailed explanation
of how to interpret the resulting table.
6. From the list of top matches, click on any HMDB hyperlink
under the “Compound” column on the left. For example, click
on the HMDB0000606 MetaboCard link. This opens up the
HMDB MetaboCard for D-2-hydroxyglutarate (Fig. 16).
7. Read through the information on D-2-hydroxyglutarate to
learn more about its known links to cancer and its role as an
oncometabolite. Click on the corresponding pathway links or
pathway thumbnail images in the MetaboCard to learn more
about the metabolism and mechanism of action of this onco-
metabolite (see Note 14).
8. Repeat the above process for the remaining identified metabo-
lites including: fumarate, succinate, lactate, glutamine, isoci-
trate, and glutathione. See Note 15 for more detailed
information about how these metabolites were identified.
9. This concludes the section on using the HMDB for disease
studies.
 Human Metabolome Database 177

Fig. 15 A screenshot showing the results of the “LC-MS Search” using the seven entered masses or m/z
values

Fig. 16 A screenshot of the MetaboCard for D-2-hydroxyglutarate

178 David S. Wishart

4 Notes

1. The Search Results page has several components. At the top left
of the page are “change page” or page selector icons that allow
users to quickly select and display different result pages con-
sisting of 25 metabolites per page. Users can select a page by
number or they may select the next page or the last page of
results. Below the page selector panel is a greyed-in area for
filtering or limiting the displayed results. Users can filter by
metabolite status (detect, expected, predicted) or by biospeci-
men type (blood, urine, saliva, etc.), which refers to the biofluid
or tissue location(s) where the metabolite is found. Using a
mouse or track-pad, users can click on one of the desired check
boxes after which they can press on the “Apply Filter” button
on the far right. Applying the filter generates a new “Search
Results” page with the results filtered according to the user
selection. Below the filter selection box is a scrollable that dis-
plays the hits from the search. On the left side is a tan-colored
button that displays the compound’s HMDB identifier. Click-
ing on this button will display the compound’s corresponding
MetaboCard. The number below each MetaboCard button is
the CAS (Chemical Abstract Services) identifier, if available. To
the right of the MetaboCard button (in the same metabolite
row) is the common name and the IUPAC name of the com-
pound along with the text in which the word “Alanine” has
matched to the text in the corresponding MetaboCard (high-
lighted in yellow). The biofluid location in which the metabo-
lite has been found is indicated in the dark grey boxes at the top
of each metabolite row. On the far right of each metabolite row
is a 2D structure of the metabolite drawn in an electronically
neutral format.
2. The HMDB has 14 different browsing options. The most
frequently chosen option is the “Metabolite” category, which
is listed at the top of the pull-down menu. When selected, the
metabolite browse option allows users to interactively view,
scroll through and filter metabolite data. The data is presented
in a structured table that provides synoptic information (name,
structure, chemical formula, biofluid source) about the
110,000+ metabolites in the HMDB. The filtering option
allows users to select metabolites based on their status (quanti-
fied, expected, predicted, etc.), biospecimen type (urine,
blood, etc.), general origin (microbial, food, endogenous,
etc.), and subcellular location. Each metabolite is linked to a
specific “MetaboCard,” which is described in more detail in
Subheading 3.2. Selecting the “Diseases” browsing option
allows users to view a structured, scrollable table that lists
(age and gender-specific) metabolite concentration data for
 Human Metabolome Database 179

>650 different diseases along with their published or online

references. Likewise, selecting the “Pathways” browsing option
allows users to view a structured, scrollable table containing
data on nearly 50,000 metabolic pathways. The pathway table
contains expandable thumbnail images along with hyperlinks
to the metabolites in each pathway. Other browsing categories
offered in HMDB’s Browse menu allow users to view metabo-
lites via their biospecimen or biofluid of origin (“Biospeci-
mens”), by their chemical class, as determined via ClassyFire
[10], (“Class”), by the 5700+ known enzymes or protein
transporters that act on human metabolites (“Proteins”), or
by the 18,000+ enzymatic reactions in the HMDB (“Reac-
tions”). Other browsing options support the display of
known metabolites (along with their references) that vary
with body mass index or BMI (“BMI Metabolomics”), with
age (“Age Metabolomics”), with gender or sex (“Gender
Metabolomics”), diurnal cycle (“Diurnal Metabolomics”),
drug intake (“Pharmaco Metabolomics”), and single nucleo-
tide polymorphism or SNP type (“Geno Metabolomics”).
3. The HMDB has 11 different search options. The most fre-
quently chosen option is the “ChemQuery Structure Search”
which is listed at the top of the pull-down menu. This search
function allows users to draw a structure using the MarvinS-
ketch applet from ChemAxon and to search for similar struc-
tures throughout the HMDB’s collection of 110,000+
structures. The applet also allows users to paste in a SMILES
[11] or InChI [12] identifier directly into the drawing canvas
to automatically generate the corresponding query structure.
These structure searches may be done via structure similarity
scores or by substructure matching and the results may be
filtered by molecular weight, number of results or metabolite
status. Other searches that are supported include a simple
“Molecular Weight Search,” a “Text Search” that supports
Boolean (AND, OR, NOT) text queries, a BLAST-based
“Sequence Search” (to look for protein or gene sequence
similarities among HMDB’s 5000+ protein/gene sequences)
and a wide variety of spectral searches (MS and NMR). The MS
(LC-MS, LC-MS/MS, LC-MS CMM, and GC-MS) searches
allow users to put in lists of m/z peaks, mass tolerances and to
select various adducts (for LC-MS data). These queries are then
searched against HMDB’s collection of 350,000 MS spectra.
The NMR searches allow users to paste in lists of chemical shifts
(for 1D or 2D NMR spectra) and to search against HMDB’s
collection of 4000 1H and 13C NMR spectra. Another impor-
tant search offering is the HMDB’s “Advanced Search” (for
metabolites or concentrations). This search option allows users
to selectively search for numbers, numeric ranges or text from
180 David S. Wishart

more than 30 HMDB data fields using a variety of conditions

(matches, does not match, starts with, etc.). Effectively,
“Advanced Search” functions as a user-friendly structured
query language (SQL) search tool.
4. The “About” pull-down menu offers 10 options that describe
additional details about the HMDB. The “About the HMDB”
option provides a short description of the current version of the
HMDB. The “Release Notes” contains detailed information
about each HMDB release over the past 12 years while the
“Citing the HMDB” provides detailed publication information
about how to properly cite the database. The “Statistics”
option lists very extensive and up-to-date information regard-
ing the numbers of compounds, names, synonyms, spectra,
proteins, genes, reactions, diseases, and so on. in the HMDB.
This information is updated regularly. The “Data Sources”
provides information about how the HMDB was assembled
and where the information was obtained while the “Other
Databases” offers short descriptions and hyperlinks to a num-
ber of popular, alternative metabolomics databases. Perhaps
the most useful link in the “About” menu is the “Help/Tuto-
rial” option. This links users to the HMDB tutorial video, a
10-min video (also available through YouTube) that provides a
nice introduction to the HMDB and how to use its many
browse, search and query options.
5. Most of the viewable data in the HMDB is contained in series
of synoptic summary tables called “MetaboCards” or metabo-
lism information cards. Each MetaboCard, which is associated
with a specific metabolite, contains >130 data fields with
approximately two-thirds of the data fields associated with
chemical or physicochemical data and the other one third
associated with biological or biomedical data. The data fields
in each MetaboCard are grouped into 14 distinctive categories
which include: (1) record information, (2) metabolite identifi-
cation, (3) chemical taxonomy, (4) chemical ontology, (5) phys-
ical properties, (6) spectra, (7) biological properties,
(8) normal concentrations, (9) abnormal concentrations,
(10) associated disorders, (11) external links, (12) references,
(13) enzymes, and (14) transporters.
6. The Metabolite Identification data field in the HMDB contains
13 different pieces of information that can be used to identify
or characterize a metabolite. These include common names,
the IUPAC name, synonyms, structure images (2D and 3D),
structure files, the chemical formula, the molecular weight, the
SMILES string, and InChI identifiers. The most useful piece of
information in the identification field is the metabolite descrip-
tion. This typically provides a short 100–200 word description
of the metabolite and its biological role(s) or origin. Every
 Human Metabolome Database 181

metabolite in the HMDB has a hand-written metabolite

description.
7. The Chemical Taxonomy data field contains 10 pieces of infor-
mation that describe each metabolite’s structural features and
their relationships to each other. The HMDB’s taxonomy is
automatically generated via the program called ClassyFire
[10]. ClassyFire takes a chemical structure, analyzes it and
then classifies it into a fully defined, hierarchical taxonomy.
The ClassyFire chemical taxonomy (like the Linnean taxonomy
in biology) includes a chemical kingdom, superclass, class, and
subclass as well as the chemical parents and substituents. Each
descriptor in the ClassyFire taxonomy has a formal, written
definition and many of these definitions are accessible via
hyperlinks in the HMDB. The point of a structural taxonomy
is to enable consistent chemical classifications and consistent
chemical descriptions across labs and across publications. The
ClassyFire chemical taxonomy has been adopted by most of the
world’s major chemical databases, including PubChem [13],
LIPID MAPS [14], ChEBI [15], and others.
8. The Ontology data field contains four subfields that describe
each metabolite’s function in the HMDB through a structured
language and a standard set of definitions. An ontology is a set
of concepts in a subject area that shows their properties and the
relations between them. The point of an ontology is to enable
consistent descriptions and to assist in computational text
mining. HMDB’s functional ontology is modeled after the
Gene Ontology or GO [16] that is widely used in molecular
biology and genetics. The main categories used in HMDB’s
Ontology cover (1) Physiological effects; (2) Disposition
(source or origin); (3) Process (involvement biological path-
ways as well as natural or industrial processes); and (4) Role
(industrial or biological application).
9. The Spectra data field contains a scrollable list of all the
GC-MS, LC-MS, LC-MS/MS and NMR spectra collected for
each metabolite in the HMDB. The spectra listed in this field
include both experimentally observed as well as predicted spec-
tra. Each spectrum has a short description of the instrument or
collection conditions and a link to enable facile viewing of the
selected spectrum with one of two interactive, online spectral
viewers (JSpectraViewer is the most commonly used). The
Biological Properties field contains four subfields that display
detailed information about the known cellular locations of the
metabolite, the known biofluids or excreta in which the metab-
olite is found, the tissue in which the metabolite is found as well
as the biological pathways that use, produce or process the
metabolite. The pathways subfield shows the name of the
182 David S. Wishart

pathway as well as clickable thumbnail views of the pathways as

made available by either by KEGG [17] or SMPDB [18].
10. Many MetaboCards in the HMDB have relatively comprehen-
sive information on Normal Concentration and Abnormal
Concentrations. These two data fields contain specific informa-
tion about metabolite concentrations in different biofluids or
excreta. Information about age, sex, and (if abnormal) the
associated condition is presented along with links to the
PubMed references and any additional notes. The diseases or
conditions identified in the Abnormal Concentrations field are
further elaborated in the Associated Disorders and Diseases
data field, which includes disease references and, if appropriate,
links to the Online Mendelian Inheritance of Man (OMIM)
database [19].
11. The last data field in the HMDB displays the enzymes and/or
transporters that are associated with each metabolite. These
data fields or data boxes are framed with a green border to
make them visibly distinct from each metabolite’s Metabo-
Card. Each of these enzyme/protein boxes contains the pro-
tein name (which is hyperlinked) along with brief descriptions
about the general function, specific function, gene name, Uni-
Prot ID, estimated molecular weight, and known reactions.
Clicking on the protein name (or the “Enzyme Details” box on
the right) will open up a richly detailed “Protein Card” con-
taining dozens of pieces of information about the protein, its
amino acid and DNA sequence, its functions, pathways, reac-
tions, GO classification, and various measurable gene or pro-
tein properties.
12. The example chosen here is fictitious but is based on published
information from several metabolomic studies on glioblastoma
(a form of brain cancer). Assume that a high resolution (Orbi-
Trap or QTOF) MS experiment has been completed on a set of
30 cerebrospinal fluid (CSF) samples collected from 30 patients
with glioblastoma and another set of 30 CSF samples from
30 patients with no cancer symptoms (but with suspected
meningitis, spinal fractures, etc.) that required the mandatory
collection of CSF. Using the positive ion mode, peaks with the
following m/z values were found to be significantly increased
in the cancer patients: 149.0444, 117.0183, 119.0339,
91.0388 while the peaks with the following m/z values were
found to be significantly decreased in the cancer patients rela-
tive to the non-cancer “controls”: 147.0763, 193.0341,
308.0911. Our task is to use the HMDB to find out what
these compounds are, how they might be involved in glioblas-
toma and to identify which biochemical or signaling pathways
they may be associated with.
 Human Metabolome Database 183

13. The table that displays the mass hits includes 7 subtables with a
subtable of mass matches for each of the 7 query masses. The
first subtable shows matches for the m/z 149.0444 query, the
second subtable shows mass matches for the m/z 117.0183
query, and so on. In some cases up to 10–12 metabolites with
identical masses or m/z values (isobaric compounds) will be
displayed. Also shown in each table or subtable is the name of
the compound, the molecular formula, the monoisotopic mass
of the parent compound, the adduct type (selected by the
user), the adduct mass, and the difference between the query
mass and the matching mass (given as Delta, in parts per
million or ppm). One of the challenges in working with only
m/z data (or only LC-MS data) is that multiple candidates or
multiple metabolites can match to a given m/z value. Usually
additional information, such as retention time, expected abun-
dance in the given biofluid, additional MS/MS spectra or
selected information regarding biological context must be
used to decide which unique compound(s) are most likely
matching to the observed MS spectra.
14. Reading through the MetaboCard for D-2-hydroxyglutarate
will reveal that it has been implicated as an oncometabolite as
well as a key metabolite in a rare inborn of metabolism called D-
2-hydroxyglutaricaciduria. Scrolling down the MetaboCard
will show more details regarding D-2-hydroxyglutarate’s struc-
ture, synonyms, chemical classifications, concentrations in dif-
ferent biofluids, pathways, and the enzymes/transporters that
bind it or act on it.
15. While multiple “hits” are possible with the mass list provided in
this example, the results of this particular LC-MS study on
glioblastoma CSF should ideally show that D-2-hydroxygluta-
rate, fumarate, succinate, and lactate have increased in concen-
tration in cancer patients, while glutamine, isocitrate, and
glutathione have decreased in cancer patients. The metabolites
that increased in concentration are known oncometabolites
and appear to catalyze further mutations and epigenetic
changes (D-2-hydroxyglutarate, fumarate) or contribute to
immunosuppression, metastasis, inflammation, and other
well-known hallmarks of cancer (succinate and lactate). This
simple example was chosen for didactic purposes. Identifying
compounds using only m/z data is often risky and leads to
multiple mass-matching redundancies. If MS/MS data were
available it would have been possible to search for similar
MS/MS spectra via HMDB’s “MS/MS Search” option.
Finding matching MS/MS spectra (either observed or pre-
dicted) to any query MS/MS spectrum provides important
supporting evidence regarding the actual identity of the com-
pound or compounds of interest.
184 David S. Wishart

Acknowledgments

This work was supported by Genome Alberta (a division of

Genome Canada), The Canadian Institutes of Health Research
(CIHR), Western Economic Diversification (WED), and the
Canada Foundation for Innovation (CFI).

References
1. Wishart DS (2007) Proteomics and the human with a comprehensive, computable taxonomy.
metabolome project. Expert Rev Proteomics J Cheminform 8:61
4:333–335 11. Weininger D (1988) SMILES, a chemical lan-
2. Mandal R, Guo AC, Chaudhary KK, Liu P, guage and information system. 1. Introduction
Yallou FS, Dong E et al (2012) Multi-platform to methodology and encoding rules. J. Chem
characterization of the human cerebrospinal Inf Comput Sci 28:31–36
fluid metabolome: a comprehensive and quan- 12. Heller S, McNaught A, Stein S,
titative update. Genome Med 4:38 Tchekhovskoi D, Pletnev I (2013) InChI—
3. Psychogios N, Hau DD, Peng J, Guo AC, the worldwide chemical structure identifier
Mandal R, Bouatra S et al (2011) The human standard. J Chem 5:7
serum metabolome. PLoS One 6:e16957 13. Kim S, Thiessen PA, Bolton EE, Chen J, Fu G,
4. Bouatra S, Aziat F, Mandal R, Guo AC, Wilson Gindulyte A et al (2016) PubChem substance
MR, Knox C et al (2013) The human urine and compound databases. Nucleic Acids Res
metabolome. PLoS One 8:e73076 44(Database issue):D1202–D1213
5. Dame ZT, Aziat F, Mandal R, 14. Sud M, Fahy E, Cotter D, Brown A, Dennis
Krishnamurthy R, Bouatra S, Borzouie S et al EA, Glass CK et al (2007) LMSD: LIPID
(2015) The human saliva metabolome. Meta- MAPS structure database. Nucleic Acids Res
bolomics 11:1864–1883 35(Database issue):D527–D532
6. Karu N, Deng L, Slae M, Guo AC, Sajed T, 15. Hastings J, Owen G, Dekker A, Ennis M,
Huynh H et al (2018) A review on human fecal Kale N, Muthukrishnan V et al (2016) ChEBI
metabolomics: methods, applications and the in 2016: improved services and an expanding
human fecal metabolome database. Anal Chim collection of metabolites. Nucleic Acids Res 44
Acta 1030:1–24 (Database issue):D1214–D1219
7. Wishart DS, Tzur D, Knox C, Eisner R, Guo 16. Ashburner M, Ball CA, Blake JA, Botstein D,
AC, Young N et al (2007) HMDB: the human Butler H, Cherry JM et al (2000) Gene ontol-
Metabolome database. Nucleic Acids Res 35 ogy: tool for the unification of biology. The
(Database issue):D521–D526 gene ontology consortium. Nat Genet
8. Wishart DS, Feunang YD, Marcu A, Guo AC, 25:25–29
Liang K, Vázquez-Fresno R et al (2018) 17. Kanehisa M, Furumichi M, Tanabe M, Sato Y,
HMDB 4.0: the human metabolome database Morishima K (2017) KEGG: new perspectives
for 2018. Nucleic Acids Res 46(D1): on genomes, pathways, diseases and drugs.
D608–D617 Nucleic Acids Res 45(Database issue):
9. Djoumbou-Feunang Y, Fiamoncini J, Gil-de- D353–D361
la-Fuente A, Greiner R, Manach C, Wishart DS 18. Jewison T, Su Y, Disfany FM, Liang Y, Knox C,
(2019) BioTransformer: a comprehensive Maciejewski A et al (2014) SMPDB 2.0: big
computational tool for small molecule metab- improvements to the small molecule pathway
olism prediction and metabolite identification. database. Nucleic Acids Res 42(Database
J Cheminform 11:2 issue):D478–D484
10. Djoumbou Feunang Y, Eisner R, Knox C, 19. Amberger JS, Bocchini CA, Scott AF, Hamosh
Chepelev L, Hastings J, Owen G et al (2016) A (2019) OMIM.org: leveraging knowledge
ClassyFire: automated chemical classification across phenotype-gene relationships. Nucleic
Acids Res 47(Database):D1038–D1043
 Chapter 11

De Novo Molecular Formula Annotation and Structure

Elucidation Using SIRIUS 4
Marcus Ludwig, Markus Fleischauer, Kai Dührkop, Martin A. Hoffmann,
and Sebastian Böcker

Abstract
SIRIUS 4 is the best-in-class computational tool for metabolite identification from high-resolution tandem
mass spectrometry data. It offers de novo molecular formula annotation with outstanding accuracy. When
searching fragmentation spectra in a structure database, it reaches over 70% correct identifications. A
predicted fingerprint, which indicates the presence or absence of thousands of molecular properties, helps
to deduce information about the compound of interest even if it is not contained in any structure database.
Here, we present best practices and describe how to leverage the full potential of SIRIUS 4, how to
incorporate it into your own workflow, and how it adds value to the analysis of mass spectrometry data
beyond spectral library search.

Key words Metabolomics, LC–MS/MS, Annotation, Molecular formula, Structure prediction, SIR-
IUS, Metabolite identification

1 Introduction

Comprehensive identification of small molecules is one of the most

urgent needs in metabolomics and related fields such as in natural
products research, biomarker discovery, and environmental science.
Yet, this task remains highly challenging. Liquid-chromatography
tandem mass spectrometry (LC–MS/MS) is one of the most prom-
inent analytical techniques to identify biomolecules. The mere mass
of a compound is not sufficient to determine the correct molecular
formula, let alone its structure. Tandem mass spectrometry pro-
vides additional information but is non-trivial to interpret. Usually,
metabolite identification is performed by searching fragmentation
spectra in a spectral library [15, 36, 41, 44]. However, spectral
libraries are—and always will be—highly incomplete. This repre-
sents a major obstacle, particularly for secondary metabolism anal-
ysis. During the last years multiple tools were developed for
searching in structure databases which are orders of magnitudes

Shuzhao Li (ed.), Computational Methods and Data Analysis for Metabolomics, Methods in Molecular Biology, vol. 2104,
https://doi.org/10.1007/978-1-0716-0239-3_11, © Springer Science+Business Media, LLC, part of Springer Nature 2020

185
186 Marcus Ludwig et al.

larger compared to spectral libraries; this includes CFM-ID [1],

DEREPLICATOR+ [25], MAGMa [30], MetFrag [32, 45],
MIDAS [40], MS-FINDER [37], and CSI:FingerID [7].
Currently, the best performing tool for this task is CSI:Fin-
gerID, successor of FingerID [13]. It is part of SIRIUS 4 [9], a
software for metabolite identification from high-resolution frag-
mentation spectra. SIRIUS started off as a method for de novo
molecular formula identification, but now integrates CSI:FingerID
to offer combined molecular formula annotation and structure
database search. SIRIUS performs metabolite identification in a
two-step approach: Firstly, the molecular formula of the query
compound is determined via isotope pattern analysis and fragmen-
tation trees. Second, SIRIUS uses CSI:FingerID to predict a
molecular fingerprint from the given spectrum and fragmentation
tree. This predicted fingerprint can be searched against a structure
database to identify the most likely candidate. Searching CASMI
2016 [33] positive ion mode spectra in a database of 0.5 million
structures of biological interest resulted in 74.0% correct identifica-
tions [9]. When searching in PubChem [20], which contains many
millions of structures, CSI:FingerID still achieves an identification
rate of 39.4% (74.8% in the top 10). These rates were reached
without using meta-information such as citation frequencies or
production volumes; using such meta-information can be very
harmful in practice [2].
Whereas spectral library search will only allow a “peek through
the keyhole,” SIRIUS enables untargeted identification to draw a
more complete picture of a metabolic system [5]. It is understood
that not every existing biomolecule is or will be contained in
structure databases. But even for these instances SIRIUS offers
valuable insight by providing a predicted molecular fingerprint to
assist de novo structure elucidation and by searching in databases of
hypothetical structures such as the in silico generated MINE data-
bases [17]. Comprehensive compound identification is not a luxury
but an indispensable step to answer biological questions. Com-
pared to spectral library search SIRIUS offers highly increased
coverage; compared to searching compounds only by mass it offers
tremendously improved accuracy. Here, we present how to use
SIRIUS to systematically annotate your compounds, and provide
insight on common practices, judging the results and necessary
prerequisites of your data.

2 What Data Can Be Processed by SIRIUS?

SIRIUS processes high-resolution, high mass accuracy fragmenta-

tion spectra, but also uses first stage of mass spectrometry (MS1)
data. The statistical model of SIRIUS and the machine learning
model of CSI:FingerID were trained on tandem mass spectra
 Metabolite Identification Using SIRIUS 187

(MS/MS) created by collision-induced dissociation (CID), as com-

monly applied in LC–MS/MS experiments. Most of the training
compounds were ionized by electrospray ionization (ESI). How-
ever, it has been reported that SIRIUS is also able to analyze
compounds from GC–MS data which has been acquired using the
soft ionization method dopant-assisted atmospheric pressure
chemical ionization (dAPCI) and subsequently fragmenting ions
by CID [22]. At present, SIRIUS only handles single-charged
compounds.

3 Preprocessing

SIRIUS is specialized in metabolite identification and relies on

other tools for proper preprocessing. Input spectra must be in
centroid mode (peak picked). Besides, further preprocessing of
the data is highly beneficial for good results. Open source software
exists for feature finding, to group isotope peaks of each com-
pound, estimate adducts, and reject all MS/MS which cannot be
assigned to a proper feature in the MS1. OpenMS [31] and
MZmine 2 [27] both provide export functions tailored to the
needs for SIRIUS.
It is beyond the scope of this chapter to go into the details of
the different preprocessing steps, but see Chapter 4 in this book for
details on OpenMS processing. Unfortunately, we cannot propose
optimal parameters, since these depend on the data. A metabolo-
mics OpenMS workflow to preprocess data for SIRIUS may use the
following OpenMS tools: FeatureFinderMetabo, Metaboli-
teAdductDecharger, and SiriusAdapter. The SiriusAdap-
ter can be used either to directly run SIRIUS or to export .ms-files
for SIRIUS to import.
SIRIUS benefits from the following preprocessing steps:
l A reasonably averaged MS1 is more accurate than using a single
MS1 spectrum. Determining the masses and intensities of the
compound’s isotope pattern using the chromatographic peaks
can reduce errors.
l When measuring multiple MS/MS spectra of the same com-
pound, in particular at different collision energies, it is beneficial
to analyze a merged spectrum rather than the individual spectra.
Fragmentation spectra can be grouped by their corresponding
MS1 feature. SIRIUS will merge all grouped spectra. This is
preferred over directly providing a merged spectrum as input
for SIRIUS.
l MS/MS spectra which cannot be assigned to any MS1 feature
should be rejected; these spectra are likely of bad quality.
188 Marcus Ludwig et al.

l MS/MS spectra with low total intensity or very few signal peaks
should be rejected. Usually it is difficult to confidently identify
the corresponding compounds.
It is usually not necessary to preprocess fragmentation spectra
by removing “noise peaks” or recalibrating masses; such preproces-
sing can substantially worsen results, as signal peaks may be
removed or masses shifted into the wrong direction. SIRIUS can
decide for itself which of the peaks in the spectrum are noise, but it
cannot recover the masses of accidentally removed signal peaks. To
this end, be cautious when using intensity thresholds. If the data is
noisy and necessitates “noise peak” removal, use a low intensity
threshold to remove as few signal peaks as possible. Furthermore,
we propose to use a low MS1 intensity threshold and not-to-
restrictive parameters for feature detection. A high number of
spurious features might pose a problem for MS1-only analysis.
But here, we concentrate on metabolite identification based on
fragmentation spectra, and spurious features can easily be recog-
nized because these will not produce significant signal peaks within
the fragmentation spectrum. Using liberal parameters will help to
detect more low intensity isotope peaks and include them into the
compound’s isotope pattern.
Instrumental setup has huge impact on spectrum quality and
some setups might be more suitable for structure elucidation with
computational tools. See Tip 1 for more information.

Tip 1: Spectra Quality

High quality spectra are indispensable to obtain good com-

pound annotations. Spectra of high quality possess many signal
peaks with intensities considerably above the noise level and mass
errors of less than 10 ppm. On the other hand, few high-intensity
signal peaks and mass errors of over 15 ppm indicate a spectrum of
bad quality. It is understood that some molecules produce few
fragments. But the information content of a spectrum increases
with the number of (non-noise) peaks; identifying a compound
from one peak is mere guessing. A proper instrumental setup can
facilitate peak-rich spectra. Instead of using a single collision
energy, spectra should be measured at multiple energies and
merged. Alternatively, a ramped collision energy can be used to
cover a large range of energies. In both cases, we expect to see more
fragmentation peaks and, hence, better results.
Broad isolation windows favor chimeric spectra, being com-
posed of fragments from more than one compound. Such chimeric
spectra will interfere with fragmentation tree computation and also
complicate the identification of structures via CSI:FingerID. In
addition, broad isolation windows will result in isotope patterns
 Metabolite Identification Using SIRIUS 189

for all fragments. Selecting only the monoisotopic peak for frag-
mentation makes it easier to interpret the fragmentation spectrum.
SIRIUS provides an option to account for isotopes in the fragmen-
tation spectrum, but this assumes that the isolation window is
broad and isotope patterns of fragments are undisturbed. Unfortu-
nately, filtering is imperfect in practice: An isolation window of
width, say, 3 Da may select 100% of the monoisotopic peak, 80%
or the first and 50% of the second isotope peak. This will distort the
isotope patterns of fragments in a non-trivial way. At present,
SIRIUS cannot deal with distorted fragment isotopes patterns.
Compound identification benefits from choosing an instru-
mental setup which minimizes chimeric spectra, and favors peak-
rich and low noise fragmentation spectra.

4 Metabolite Identification

SIRIUS identifies metabolites in two steps: namely molecular for-

mula annotation and searching in a structure database. Both steps
can be performed on a complete dataset using a single command;
but users are advised to manually validate all results, including
intermediate results. Here, we will explain the usage of SIRIUS
step-by-step. For the sake of a more vivid description we will refer
to the graphical user interface (GUI) of SIRIUS. All computations
can be performed via the command-line interface (CLI), using the
GUI as a mere visualization tool for final results (see Subheading 5).
An overview of the SIRIUS GUI is displayed in Fig. 1. The
analysis starts with importing the data; this is done via the import
dialog or drag-and-drop. SIRIUS imports spectra from .csv, .ms, or
.mgf files. Imported compounds are displayed in the compound list
located in the left panel. To find specific compounds, use the search
field above the panel. Start computations by clicking the Compute
All button or by selecting a set of compounds and using the context
menu (right-click). If only a single compound is selected, additional
parameters can be specified such as the known molecular formula.

4.1 Molecular SIRIUS finds the most likely molecular formula by considering all
Formula Annotation possible molecular formulas, and is able to annotate biomolecules
with a molecular formula missing from any database. Necessary
parameters for SIRIUS are:
Elements Set of considered elements. Some elements
can be auto-detected if an isotope pattern is
given (see Tip 2).
ppm Allowed mass deviation in ppm. This is the
maximum value a molecular formula expla-
nation is allowed to deviate from the peaks’
measured mass. Molecular formulas with a
190 Marcus Ludwig et al.

Fig. 1 The SIRIUS Overview tab displays the spectrum and fragmentation trees of the top molecular formula
candidates. The best candidate C24H38O3 is selected; the corresponding explained spectrum and fragmenta-
tion tree are shown. The left panel contains a searchable list of all compounds; selected compounds are
highlighted. The data and results of the first selected compound are displayed in all the views to the right of
the compound list. The upper panel provides functionalities to import spectra, save and load workspaces,
export result tables, start computations, and display their status in the jobs panel. The SIRIUS overview tab
displays various scores for each molecular formula candidate and can be sorted accordingly

higher mass error are ignored. Note that for

all peaks below 200 Da an absolute error is
assumed which corresponds to the specified
deviation in ppm at 200 Da.
Considered ion types Set of considered ion types. For details see
Tip 4.
Candidates Number of candidates to be displayed. Frag-
mentation trees are computed for all molec-
ular formula candidates using the Critical
Path3 heuristic from [8]. The top
k fragmentation trees are recomputed using
an exact algorithm; here, k corresponds to
the number of displayed candidates plus 10.
Hence, a larger number of displayed candi-
dates increases running times.

Depending on the dataset, anticipated elements and ion types

can be selected. Select a reasonable set of elements. The mass
deviation is the maximum allowed deviation. Spectra measured on
 Metabolite Identification Using SIRIUS 191

an instrument with advertised sub-ppm mass accuracy might still

have much larger mass deviation (e.g., if not properly calibrated or
because of bad peak picking). More restrictive parameters, in par-
ticular for the allowed elements, can make computations substan-
tially faster. Never select all uncommon elements at once. This will
lead to a combinatorial explosion of potential molecular formulas;
running times will increase dramatically; the number of correct
molecular formula annotations will decrease. SIRIUS provides
scoring profiles for Q-TOF and Orbitrap, which mainly change
some background parameters. In case you are unsure if your data
really has the instrument’s advertised accuracy, use the default
profile and set your allowed mass deviation accordingly.
Fragmentation trees are computed from a merged spectrum
combining all input fragmentation spectra. Isotope pattern analysis
is performed on a merged MS1 spectrum or using the isotope
pattern provided by a preprocessing tool. A fragmentation spec-
trum which possesses peaks broadly distributed across the whole
mass range presents more information to SIRIUS than a spectrum
composed of either low or high mass peaks only.

4.1.1 Judging Results Molecular formula annotation results are displayed in the Sirius
Overview tab (see Fig. 1). Candidates are ranked by the sum of
isotope pattern and fragmentation tree score (see Tip 2 on isotopes
and Tip 3 on fragmentation trees). Colored bars for each score ease
comparison between candidates. Each candidate molecular formula
has an adduct. At this stage, this is an ion type; after structure
database search with CSI:FingerID this adduct corresponds to an
adduct type (compare Figs. 1 and 3 and see Tip 4).
The displayed attributes are:
Score Overall score by which candi-
dates are ranked. This is the
sum of isotope and tree score.
Isotope score Similarity score comparing the
measured isotope pattern with
the theoretical pattern for each
candidate molecular formula.
Usually, a score close to zero or
low in comparison to the remain-
ing candidates indicates an incor-
rect molecular formula, or at
least an annotation of low confi-
dence. Besides being the incor-
rect candidate, this might
indicate improper data quality
such as high intensity deviation
or a low number of detected iso-
tope peaks. The scored isotope
192 Marcus Ludwig et al.

pattern is highlighted in the

merged MS1 and can be assessed
via the Spectrum view tab.
Tree score Score of the computed
fragmentation tree.
Explained peaks The number of peaks in the spec-
trum which can be explained by
the fragmentation tree. A high
number of unexplained peaks
indicates an incorrect annota-
tion, a noisy spectrum, or two
compounds being fragmented
simultaneously.
Total explained intensity Summed relative intensity of all
explainable peaks. Values of 95%
or higher indicate good quality;
for values below 80%, results
should be interpreted with care.
Median absolute mass deviation The median absolute mass devia-
tion of explained peaks in ppm.
Low deviations are clearly
desirable.

Selecting a molecular formula candidate displays the

corresponding fragmentation tree and spectrum in which explained
peaks are highlighted. The merged MS1 spectrum displays the
selected isotope pattern. Mass errors of each fragment are shown
to spot unlikely explanations; the displayed fragmentation tree can
be colored accordingly. The user can inspect fragmentation tree
annotations in varying degree of detail; individual fragments may
support or contradict a particular molecular formula candidate.
The user may decide by manual validation how well a candidate is
supported.

Tip 2: Isotope Pattern and Element Detection

CAUTION: If no isotope pattern is provided and compounds
are expected to contains elements beside CHNOPS, we
strongly recommend to restrict molecular formulas to those
from a molecular structure database. Do not select all uncom-
mon elements for molecular formula annotation with SIR-
IUS. This will lead to a combinatorial explosion of potential
molecular formulas; running times will increase dramatically.

Isotope patterns offer valuable information about elemental

composition. The presence of uncommon elements that result in
characteristic isotope pattern changes can be automatically detected
[24]. Detectable elements are sulfur, chlorine, bromine, boron, and
selenium. When detected, SIRIUS adds these elements to the
 Metabolite Identification Using SIRIUS 193

default set of elements CHNOP to determine the molecular for-

mula. A predictor for silicon is disabled by default, as it results in a
relatively large number of false positive predictions; the silicon
isotope pattern is not “special” enough to permit a reliable auto-
detection. In contrast to [24], the current version of SIRIUS uses a
deep neural network for auto-detection of elements. Automated
detection can be enabled or disabled via the compute dialog. Not
considering elements which are extremely unlikely, substantially
improves running times and may slightly improve results [24]. SIR-
IUS may still choose a molecular formula which does not contain an
element with positive auto-detection, just as it might choose a
molecular formula which does not contain any other enabled ele-
ment. The final score of each molecular formula candidate is a
combination of the fragmentation tree score and the isotope pat-
tern score.

Tip 3: Fragmentation Trees

A fragmentation tree annotates peaks in the fragmentation

spectrum with molecular formulas and identifies likely losses
between the fragments—similar to “fragmentation diagrams” cre-
ated by experts. The calculated tree must not be understood as
ground truth but can be used to derive information about the
measured compound’s fragmentation [29]. Fragmentation trees
are also used to identify the molecular formula of an unknown
compound. For every molecular formula candidate of the precursor
ion, a separate fragmentation tree is computed which best explains
the spectrum, as evaluated by a Maximum A Posteriori estimator
[3]. This estimation takes into account information such as mass
deviations, intensities, common losses, and loss sizes. The overall
best-scoring fragmentation tree corresponds to the most likely
molecular formula explanation. In addition, CSI:FingerID uses
the fragmentation tree to predict the compound’s molecular
fingerprint.
A simplified example of a fragmentation tree is presented in
Fig. 2. A fragmentation tree is computed from the fragmentation
spectrum given the (candidate) molecular formula of the precursor
ion. Initially, a fragmentation graph is constructed in the following
way: For every fragment peak, all possible molecular formula expla-
nations are computed. These explanations must be subformulas of
the precursor molecular formula—a fragment only loses, but never
gains new atoms. Every such molecular formula is a node in the
graph. Nodes are connected by an edge if one node is a subformula
of another node—this represents a potential loss. Using combina-
torial optimization, the best-scoring fragmentation tree is com-
puted which explains every peak at most once. Unexplained peaks
are considered noise.
194 Marcus Ludwig et al.

Fig. 2 Example of a fragmentation tree computed from a fragmentation graph in (a), given the spectrum in
(b). The molecular formula of the neutral precursor is assumed to be C9H12NO2. Molecular formulas are
computed for all fragment peaks and serve as the nodes of the graph; nodes with the same color indicate
molecular formulas corresponding to the same peaks. Nodes are connected by edges if one node is a
subformula of another, thereby creating the fragmentation graph. A fragmentation tree is a connected
subgraph which explains each color (peak) at most once and has no cycles. The best-scoring fragmentation
tree, corresponding to a Maximum A Posteriori estimator, is computed by combinatorial optimization. The
optimal fragmentation tree is indicated by solid lines; nodes which are not used are grayed out. These
computations are repeated for each molecular formula candidate explaining the precursor mass, and the best
such fragmentation tree is reported

Tip 4: Ion and Adduct Types

SIRIUS differentiates between ion types and adduct types.

Default ion types for positive ion mode spectra are protonation,
sodium, and potassium; default ion types for negative ion
mode spectra are deprotonation and chlorine. Adduct types can be
seen as sub-types of an ion type. For example, the ion type protonation
includes adduct types “intrinsically charged” ([M]+), “protonated“
([M + H]+), “protonated with water loss” ([M H2O + H]+), and
“ammonium group” ([M + NH4]+).
Adduct types cannot be determined from the fragmentation
spectrum—the fragments [C4H6O2 + NH4]+ and [C4H9NO2 +
H]+ result in the exact same peak; and so will [C5H7]+ and
[C5H8O H2O + H]+. That is why SIRIUS considers ion types,
not adduct types, during the molecular formula annotation step.
Multiple adduct types of the determined ion type can be considered
for structure database search with CSI:FingerID (see Figs. 3 and 4).
When a specific ion type plus adduct type is provided by the user, it
will be used during all computation steps. Users can specify addi-
tional ion and adduct types within the GUI or by modifying the
config file.
 Metabolite Identification Using SIRIUS 195

Fig. 3 Additional candidates are added to the SIRIUS Overview tab after searching with CSI:FingerID in a
structure database considering adduct types [M + H]+, [M + NH4]+ and [M H2O + H]+. Molecular formulas
C24H40O4 and C24H38O3 differ by an in-source loss of H2O and are not distinguishable by MS/MS since in both
cases, the ion [C24H38O3 + H]+ is fragmented; hence, both have identical score. (The same holds for the pairs
C22H33N2O2 vs. C22H36N3O2 and C18H39N4O2P vs. C18H36N3O2P). Displayed is the resolved fragmentation tree
for [C24H40O4 H2O + H]+, where an H2O loss has been added to its top

Tip 5: Molecular Fingerprint

A molecular fingerprint is a binary vector of fixed length where

each position corresponds to a specific molecular property; for
example, position # 393 may encode the presence or absence of a
benzene ring as a substructure. In general, a “1” indicates this
specific substructure is present in the molecule, a “0” indicates it
is not. There exist several types of fingerprints, such as PubChem
CACTVS fingerprints,1 Klekota–Roth fingerprints [21], and
MACCS fingerprints. Given a molecular structure, the
corresponding fingerprint can be deterministically computed.
Unfortunately, different structures can have the same molecular
fingerprint.
Molecular fingerprints can be used to perform similarity search
in a structure database. A common way to compare molecular

1
ftp://ftp.ncbi.nlm.nih.gov/pubchem/specifications/pubchem_fingerprints.pdf
196 Marcus Ludwig et al.

Fig. 4 The CSI:FingerID Details tab displays structure candidates for a selected molecular formula. The
highlighted molecular property, which is predicted to be present in the query, is contained in the top 2 hits.
Candidates are sorted by their score which is displayed on the right-hand side. Numbers in percent indicate
the Tanimoto similarity between the predicted fingerprint and the fingerprint of each candidate. Candidates
can be filtered by database, SMARTS string and XlogP value

structures using fingerprints is the Tanimoto similarity, also known

as Jaccard index. Identical fingerprints produce a similarity of
1, whereas two structures not sharing a single molecular property
have a Tanimoto of 0. Clearly, the similarity value depends on the
choice of fingerprint type.
CSI:FingerID predicts a variety of molecular properties from
several fingerprint types; only those molecular properties were
selected which could also be predicted in evaluations. Given a
spectrum and corresponding fragmentation tree, CSI:FingerID
predicts a probabilistic fingerprint, see Subheading 4.3. This pre-
dicted fingerprint is compared to the deterministic fingerprints
from a structure database to find the best match. The CSI:FingerID
Overview tab also displays, for every structure candidate, the Tani-
moto similarity against the predicted fingerprint. However, CSI:
FingerID uses a different scoring function to rank candidates,
which results in a larger number of correct identifications [7, 23].

4.2 Searching After the molecular formula has been identified, the compound is
in Structure Databases searched in a structure database. Firstly, a molecular fingerprint of
the query (see Tip 5) is predicted from the spectrum and fragmen-
tation tree. Next, this predicted fingerprint is compared to (and
 Metabolite Identification Using SIRIUS 197

scored against) fingerprints of structures in a database, to find the

best matching structure. It must be understood that the molecular
fingerprints of the candidate structures are fixed, known and inde-
pendent of our tools.
To predict the molecular fingerprint, we have to know the
molecular formula, ion type, and adduct type of the query. By
default, not only the top-scoring molecular formula but multiple
high-scoring molecular formula candidates are considered, apply-
ing a soft score threshold: All molecular formula candidates with a
score above 0.75 of the optimal score are considered. To this end,
we iterate over all possible combinations of molecular formula
candidate and adduct type. The ion type of the query is determined
by the molecular formula candidate; but various adducts types can
be specified to search the database, see Tip 4 on ion and adduct
types. When searching in the database, candidate structures must
match the estimated molecular formula of the neutral molecule.
Fragmentation trees of different adduct types differ as, say, a neutral
loss is added to the top of the tree. These trees have exactly the same
score. For each molecular formula and adduct type with candidate
structures in the database, the resolved fragmentation tree is dis-
played in the SIRIUS Overview tab, see Fig. 3.
Scored structure candidates are displayed in the CSI:FingerID
Overview tab. The CSI:FingerID Details tab allows to examine the
scored structures in more detail for each molecular formula and
adduct type separately (see Fig. 4).
As a default, users should search compounds in the PubChem
database, and filter results to the biocompound structure database
or a subset thereof (see Tip 6). You may accept those query identi-
fications for which there is a high-scoring structure candidate in the
restricted database; potentially, this is even the highest-scoring
candidate for all of PubChem. For those cases where no reasonable
candidate was found in the biocompound structure database, and
for cases where the best PubChem candidate scores substantially
better than the best biocompound candidate, you can extend your
search space to all of PubChem. Obviously, it makes much sense to
integrate biochemical background knowledge at this point: This may
be information about the organism the sample was taken from, or
information about the biochemical preparation of the sample. Such
meta-information is not integrated into SIRIUS and CSI:Fin-
gerID, as this integration is highly non-trivial; but it is straightfor-
ward how to integrate the information manually.

4.2.1 Judging Results Users should check if the best structure candidate agrees with the
best molecular formula candidate. Sometimes, CSI:FingerID deci-
des that, based on its machine learning model and the given candi-
date structures, a structure with a different molecular formula
better agrees with the data. Users should verify if the selected
structure database does not contain any structures for the best-
198 Marcus Ludwig et al.

scoring molecular formula candidate; this can be an indication that

the selected database is too restrictive. Besides, check if the correct
adduct type has not been selected for database search.
CSI:FingerID ranks structure candidates by a logarithmic pos-
terior probability [23], so that scores are negative numbers and
zero is the optimum. Additionally, the predicted Tanimoto similar-
ity is displayed. Since this is based on the predicted probabilistic
fingerprint, this similarity usually underestimates the Tanimoto
similarity between the true fingerprints. Candidates can be filtered
by database, XlogP values [38, 39] predicted using the Chemistry
Development Kit [35, 43], or a specific SMARTS string. Structures
are linked to database entries; clicking on the database icon opens
the appropriate website. One CSI:FingerID candidate structure
may link to several “3D structures” in a database, as CSI:FingerID
ignores stereochemistry in its computations. The number of
PubMed citations2 is also displayed in the CSI:FingerID Overview
tab. This value can contribute valuable information for the identifi-
cation, for example, as a sanity check. But on startup, these values
must not be used to filter results: Doing so, we ignore the actual
experimental data and potentially make our decisions based solely
on prior knowledge [2].
The example in Fig. 4 shows two top-scoring structure candi-
dates. Both are structurally very similar and consequently, also have
similar scores. The user may decide which structure is more likely,
based on background knowledge about the sample. Comparing
the, say, top 5 hits may also help to get an idea about a “core”
structure which CSI:FingerID predicts to be present. Blue and red
squares next to each candidate molecular structure represent its
molecular properties. Blue properties are predicted to be present
by CSI:FingerID and also present in the candidate; red properties
are predicted to be absent but are present in the candidate. The size
of the square represents the quality (F1 score, harmonic mean of
precision and recall) of the predictor, as determined beforehand in
cross validation; but a large F1 score does not guarantee that the
prediction is correct for this query. In contrast, the saturation of the
color indicates how sure CSI:FingerID is about the property, for
this query. One specific property—a carboxyl group attached to a
carbon chain—has been highlighted in Fig. 4; it is present in the
predicted fingerprint and in the first two candidates. A score close
to zero and many blue squares usually indicates a confident identi-
fication—in this example, CSI:FingerID is very certain that the
correct structure is at least very similar to the top hit. Even in case
the best structure candidate is not correct, it is often structurally
similar to the correct one and can help to elucidate the structure or
answer the underlying biological question. Be warned that CSI:

2
https://www.ncbi.nlm.nih.gov/pubmed
 Metabolite Identification Using SIRIUS 199

FingerID scores between different query compounds are usually

not comparable; be cautious when using this score to differentiate
between true and bogus identifications.
As explained in Subheading 4.1.1, users can also examine the
fragmentation tree to decide how well a candidate is supported: For
example, are specific side chains supported by fragments, losses or
even fragmentation cascade in the fragmentation tree?

Tip 6: Some Notes on Database Size

CSI:FingerID correctly identifies 39.4% of CASMI 2016 posi-

tive ion mode spectra when searching in PubChem (in a structure-
disjoint cross-validation setup). Searching in PubChem is difficult
because it contains many millions of structures. If the search is
performed in a database with 0.5 million structures of biological
interest, correct identifications increase to 74.0% [9]. To further
increase identification rates, we might even be more restrictive and
search in HMDB [44] or ChEBI [12]. Limiting CSI:FingerID
search to the same structures which are contained in spectral
libraries will even result in identification rates comparable to spec-
tral library search! Does this mean it is advisable to search in a
database with as few structures as possible? Clearly not! Results
will look great in evaluation as long as all reference structures are
contained in the restricted database. But in application, many com-
pounds will be absent from the database, meaning you cannot find
them at all.
Furthermore, there are—often ignored—side effects of search-
ing in small databases. Firstly, the measured data becomes less
important. You can easily identify a compound from one peak if
you limit the candidate list to a few structures. Unfortunately,
doing so does not increases the identification’s confidence. It
merely means that one candidate better matches the data compared
with the other candidates, always assuming the correct structure is
present in the candidate list. Second, incorrect identifications can
be hard to spot, because they still “make sense”: If all candidates in
our database are frequently cited structures, then any identification
(including the incorrect ones) will be a frequently cited structure
and, hence, “reasonable.”
Clearly, there is a trade-off between small and large databases.
In a small database, many relevant biomolecules are missing. On the
other hand, searching in PubChem decreases the number of correct
identifications even though many PubChem structures are very
unlikely to be actual biomolecules. CSI:FingerID provides a bio-
compound database with 0.5 million structures of biological inter-
est, containing structures from ChEBI [12], KNApSAcK [34],
HMDB [44], KEGG [18], HSDB [10], MaConDa [42], BioCyc
200 Marcus Ludwig et al.

[4], UNDP [11], a biological subset of ZINC [16], GNPS [41],

MassBank [15], and MeSH-annotated PubChem compounds
[20, 26]. In application, it is reasonable to search in this biocom-
pound database, which is much smaller than PubChem, but still
much more diverse than spectral libraries. For those queries where
we find no reasonable explanation in the biocompound database,
we can then consider the PubChem candidates.

4.3 Beyond Structure It is understood that certain query biomolecules are not contained
Database Search in any structure database. But even for such difficult instances,
SIRIUS and CSI:FingerID can assist in structural elucidation.
Recall that the SIRIUS molecular formula annotation step (Sub-
heading 4.1) is done de novo. Hence, molecular formulas can be
determined even for “novel compounds” absent from any structure
database. Even if a structure is not contained in the structure
databases, CSI:FingerID may find a very similar structure. Further-
more, CSI:FingerID allows the user to search in custom databases
which may contain hypothetical structures, to identify “novel
compounds.”
But one key feature sets CSI:FingerID apart from other
computational tools for structure elucidation: Predicting the
molecular fingerprint of the query compound does not require
any molecular structure database! The fingerprint is predicted
from fragmentation spectrum and tree, and contains information
about thousands of molecular properties. From that, we may draw
conclusions what kind of substructures the query compound con-
tains; and this information may be sufficient to decide if it is worth
to further investigate the examined compound.

4.3.1 Judging Results The predicted fingerprint is displayed in the Predicted Fingerprint
tab, see Fig. 5. Most molecular properties are described by
SMARTS (SMiles ARbitrary Target Specification) strings.3
SMARTS allows a flexible encoding of substructures; for example,
a property might be described as “a methyl group bound to a
hetero atom.” Since SMARTS strings are usually hard to visualize,
SIRIUS displays a set of example structures from the training data
that have a particular molecular property.
A posterior probability is predicted for every molecular prop-
erty. Estimates close to 1 indicate the property is likely being
present in the query compound, whereas estimates close to 0 indi-
cate it is not. But be careful: Since CSI:FingerID predicts thousands
of properties, even some “rather certain predictions” must be
wrong. A 98% chance of being present also corresponds to 2%
chance of being absent; if 1000 molecular properties are predicted
at this level of certainty, then 20 predictions are wrong. Also be

3
http://www.daylight.com/dayhtml/doc/theory/theory.smarts.html
 Metabolite Identification Using SIRIUS 201

Fig. 5 The Predicted Fingerprint tab displays a predicted molecular fingerprint for each molecular formula
candidate. The molecular fingerprint is predicted independently from any database. It can help deducing
structural information on the compound even if the compound is not present in any structure database.
Highlighted is a property mainly consisting of a ring. Training examples are displayed at the bottom. As shown,
the oxygen is not a mandatory part of the substructure. The posterior probability of each property is also
visualized as a color bar, to allow the user to swiftly distinguish properties predicted being present and absent.
Green bars going to the right encode presence, red bars going the left encode absence

reminded that these probabilities are estimates. To provide addi-

tional information on the quality of a prediction, the F1 score—a
measure of the predictor quality—is displayed. The F1 score is the
harmonic mean of precision (fraction of correct yes-predictions
among all yes-predictions) and recall (fraction of correct
yes-predictions among all yes-instances). A high F1 score indicates
a good predictor, and 1.0 is the optimum. There is no general rule
on what is a “good” F1 score; as a rule of thumb for this decision,
one may assume that the F1 score equals precision and recall. Since
many properties are rare and only present in few structures, the
number of positive training examples is another indicator for the
generalizability of the predictor. To help the user to concentrate on
the most promising predictions, properties can be sorted by poste-
rior probability, F1 score, or the number of atoms. The last option
is useful to consider only larger, presumably more informative
substructures.
202 Marcus Ludwig et al.

5 Using SIRIUS in Automated Workflows

SIRIUS offers a powerful command-line interface (CLI) which

allows for a flexible integration of SIRIUS into automated work-
flows. Technically speaking, the SIRIUS GUI is a visualization of
the CLI functionality. Therefore, every task that can be done via the
graphical user interface, can also be executed using the CLI.
Corresponding to the two step approach in the GUI, the CLI
provides self-contained sub tools for molecular formula identifica-
tion and structure elucidation with separate parameter sets.
Furthermore, CLI and GUI share the same input and output
formats. Both, CLI and GUI store the computed results in the
SIRIUS project-space (see Fig. 6) which in turn can also be an input
for the GUI or the CLI. This allows the user to review results in the
GUI that have been computed with an automated workflow using
the CLI.

5.1 The SIRIUS The SIRIUS project-space is a standardized directory structure that
Project-Space is organized in a three hierarchy levels, namely the project level, the
compound level, and the method level (see Fig. 6 for details).
On the project level, each compound corresponds to one
sub-directory (compound level) storing the input data, parameters,
and results of the different analysis methods. These data is continu-
ously written to the project-space, so that it represents the actual
progress of a SIRIUS analysis. Further, the .progress file gives an
overview about the progress of the ongoing analysis. On the com-
pound level, each method provided by SIRIUS stores its results in its
own sub-directory (method level). This allows the user to redo one
analysis step without having to recompute the intermediate results
it depends on. Further, SIRIUS is able to transfer intermediate
results to a new project-space, so different parameters can easily
be evaluated without having to recompute intermediate results.
Since a project-space can be imported into the GUI, the user is
able to judge intermediate results using the GUI before executing
further analysis steps. Project-spaces can be read and written as an
uncompressed directory or a compressed zip archive when using
the .sirius file extension.
In addition to the method level results, the project-space con-
tains summaries of these results on the project level and the com-
pound level. These summaries are in csv format to provide easy
access to the results for further downstream analysis, data sharing,
and data visualization. The summaries are not imported into SIR-
IUS but are (re-)created based on the actual results every time a
project-space is exported.
 Metabolite Identification Using SIRIUS 203

Fig. 6 The SIRIUS project-space is a standardized directory structure that stores results, summarized results,
input data, parameters, and version information of a SIRIUS analysis. It is organized on three levels, namely
the project level (dashed line), the compound level (dashed-and-dotted line), and the method level (solid line).
The compound level contains sub-directories (blue) for each compound, summaries (green) about the whole
dataset, and additional information (red) about the version of SIRIUS that created the output. The compound
level contains a sub-directory for each method that was applied to the compound as well as the summaries of
these methods results. Further, it contains additional information, such as the input data and the parameters
used for the computations. On the method level, SIRIUS stores the results of a specific method for a given
compound (grey)

5.2 Standardized The project-space is a SIRIUS-specific format that allows the user
Project-Space to access all results and analysis details, but may not be optimal for
Summary sharing this data with third party tools or data archives. For this
with mzTab-M purpose, SIRIUS provides an analysis report (report.mztab) in
the standardized mzTab-M format [14]. All results summarized in
this report are linked to the results in the corresponding SIRIUS
project-space, allowing the user to share summarized results using
mzTab-M without losing the connection to the detailed results
provided in the project-space. Furthermore, SIRIUS passes meta-
information such as scan numbers and identifiers of the input data
into this analysis report. This allows for an easy combination of the
SIRIUS results with the results of other analyses such as MS1-based
quantification.
204 Marcus Ludwig et al.

6 Custom Databases

Users may define their own structure databases to search in. These
“custom databases” can be created via GUI and CLI. In the GUI,
the Databases button opens a dialogue listing existing databases.
New ones can be created with one click. Structures are imported
by inserting structure descriptors (InChI or SMILES) into the
import field; one structure per line. Custom databases are useful in
case the users has a limited set of structures of interest. When
screening for pollutants or drugs, a list of suspected structures can
be collected in advance.
When searching with CSI:FingerID it does not matter if the
structures in the database are known biomolecules or if these are
hypothetical structures, which have not yet been discovered in any
organism. Clearly, it is not reasonable to search in an arbitrarily
large database. Databases of hypothetical structures have to be
compiled with care to avoid combinatorial explosion. Available
tools are BioTransformer [6] and the in silico generated MINE
databases [17]. Currently, there exist MINE extensions for Ecocyc
[19], YMDB [28], and KEGG [18]. But in principle, any existing
structure database can be extended by such methods. Say, you are
interested in finding new bile acids. A database of hypothetical bile
acids can be created by applying biotransformations to known bile
acids. This new database can then be searched with CSI:FingerID
to find new bile acids synthesized by the investigated organisms.

7 Conclusion

To leverage the full potential of metabolomics, we need to over-

come the limitations of spectral library search. This chapter pre-
sented concepts behind SIRIUS and CSI:FingerID, best-in-class
computational tools for metabolite identification from high-
resolution tandem mass spectra. We stress that computational
tools currently cannot replace experts, but are meant to assist
them. As a consequence, users must not accept identifications
blindly but verify them properly. Here, we gave some advice on
how this can be done.
SIRIUS ships with a command line tool which makes it easy to
run computations on compute clusters and properly integrate it
into automated workflows. Popular mass spectrometry data proces-
sing tools can create input files for SIRIUS, and SIRIUS outputs
results in the standardized mzTab-M format to facilitate integra-
tion. The metabolomics community benefits from new computa-
tional tools, but tool development also benefits from the
communities’ input and more public training spectra. Finally,
method development is an ongoing process, and SIRIUS is evol-
ving to further improve metabolite identification.
 Metabolite Identification Using SIRIUS 205

References
1. Allen F, Greiner R, Wishart D (2015) Compet- Rousu J, Böcker S (2019) Sirius 4: a rapid tool
itive fragmentation modeling of ESI-MS/MS for turning tandem mass spectra into metabo-
spectra for putative metabolite identification. lite structure information. Nat Methods.
Metabolomics 11(1):98–110. https://doi. https://doi.org/10.1038/s41592-019-0344-
org/10.1007/s11306-014-0676-4 8
2. Böcker S (2017) Searching molecular structure 10. Fonger GC, Hakkinen P, Jordan S, Publicker S
databases using tandem MS data: are we there (2014) The National Library of Medicine’s
yet? Curr Opin Chem Biol 36:1–6. https:// (NLM) Hazardous Substances Data Bank
doi.org/10.1016/j.cbpa.2016.12.010. (HSDB): background, recent enhancements
https://authors.elsevier.com/a/1UF- and future plans. Toxicology 325:209–216.
u4sz6LvFfY https://doi.org/10.1016/j.tox.2014.09.003
3. Böcker S, Dührkop K (2016) Fragmentation 11. Gu J, Gui Y, Chen L, Yuan G, Lu HZ, Xu X
trees reloaded. J Cheminform 8:5. https://doi. (2013) Use of natural products as chemical
org/10.1186/s13321-016-0116-8. http:// library for drug discovery and network phar-
www.jcheminf.com/content/8/1/5 macology. PLoS One 8(4):1–10
4. Caspi R, Altman T, Billington R, Dreher K, 12. Hastings J, Owen G, Dekker A, Ennis M,
Foerster H, Fulcher CA, Holland TA, Keseler Kale N, Muthukrishnan V, Turner S,
IM, Kothari A, Kubo A, Krummenacker M, Swainston N, Mendes P, Steinbeck C (2016)
Latendresse M, Mueller LA, Ong Q, Paley S, ChEBI in 2016: improved services and an
Subhraveti P, Weaver, DS, Weerasinghe D, expanding collection of metabolites. Nucleic
Zhang P, Karp PD (2014) The MetaCyc data- Acids Res 44(D1):D1214–D1219. https://
base of metabolic pathways and enzymes and doi.org/10.1093/nar/gkv1031. http://
the BioCyc collection of pathway/genome europepmc.org/articles/PMC4702775
databases. Nucleic Acids Res 42(D1): 13. Heinonen M, Shen H, Zamboni N, Rousu J
D459–D471. https://doi.org/10.1093/nar/ (2012) Metabolite identification and molecular
gkt1103. http://nar.oxfordjournals.org/con fingerprint prediction via machine learning.
tent/42/D1/D459.abstract Bioinformatics 28(18):2333–2341. https://
5. da Silva RR, Dorrestein PC, Quinn RA (2015) doi.org/10.1093/bioinformatics/bts437
Illuminating the dark matter in metabolomics. 14. Hoffmann N, Rein J, Sachsenberg TT,
Proc Natl Acad Sci U S A 112 Hartler J, Haug K, Mayer G, Alka O,
(41):12549–12550. https://doi.org/10. Dayalan S, Pearce JTM, Rocca-Serra P et al
1073/pnas.1516878112 (2019) mzTab-M: a data standard for sharing
6. Djoumbou-Feunang Y, Fiamoncini J, Gil-de-la quantitative results in mass spectrometry meta-
Fuente A, Greiner R, Manach C, Wishart DS bolomics. Anal Chem 91(5):3302–3310.
(2019) BioTransformer: a comprehensive https://doi.org/10.1021/acs.analchem.
computational tool for small molecule metab- 8b04310
olism prediction and metabolite identification. 15. Horai H, Arita M, Kanaya S, Nihei Y, Ikeda T,
J Cheminf 11(1):2 Suwa K, Ojima Y, Tanaka K, Tanaka S,
7. Dührkop K, Shen H, Meusel M, Rousu J, Aoshima K, Oda Y, Kakazu Y, Kusano M,
Böcker S (2015) Searching molecular structure Tohge T, Matsuda F, Sawada Y, Hirai MY,
databases with tandem mass spectra using CSI: Nakanishi H, Ikeda K, Akimoto N, Maoka T,
FingerID. Proc Natl Acad Sci U S A 112 Takahashi H, Ara T, Sakurai N, Suzuki H,
(41):12580–12585. https://doi.org/10. Shibata D, Neumann S, Iida T, Tanaka K,
1073/pnas.1509788112 Funatsu K, Matsuura F, Soga T, Taguchi R,
8. Dührkop K, Lataretu MA, White WTJ, Böcker Saito K, Nishioka T (2010) MassBank: a public
S (2018) Heuristic algorithms for the maxi- repository for sharing mass spectral data for life
mum colorful subtree problem. In: Proceed- sciences. J Mass Spectrom 45(7):703–714.
ings of workshop on algorithms in https://doi.org/10.1002/jms.1777
bioinformatics (WABI 2018). Leibniz interna- 16. Irwin JJ, Sterling T, Mysinger MM, Bolstad
tional proceedings in informatics (LIPIcs), vol ES, Coleman RG (2012) ZINC: a free tool to
113. Schloss Dagstuhl–Leibniz-Zentrum fuer discover chemistry for biology. J Chem Inf
Informatik, Dagstuhl, pp 23:1–23:14. https:// Model 52(7):1757–1768
doi.org/10.4230/LIPIcs.WABI.2018.23. 17. Jeffryes JG, Colastani RL, Elbadawi-Sidhu M,
http://drops.dagstuhl.de/opus/volltexte/ Kind T, Niehaus TD, Broadbelt LJ, Hanson
2018/9325 AD, Fiehn O, Tyo KEJ, Henry CS (2015)
9. Dührkop K, Fleischauer M, Ludwig M, Akse- MINEs: open access databases of computation-
nov AA, Melnik AV, Meusel M, Dorrestein PC, ally predicted enzyme promiscuity products for
206 Marcus Ludwig et al.

untargeted metabolomics. J Cheminform 7:44. in the organization of knowledge. Kluwer Aca-

https://doi.org/10.1186/s13321-015-0087- demic Publishers, Dordrecht, pp 171–184.
1 http://www.nlm.nih.gov/mesh/meshrels.
18. Kanehisa M, Sato Y, Kawashima M, html
Furumichi M, Tanabe M (2016) KEGG as a 27. Pluskal T, Castillo S, Villar-Briones A, Oresic
reference resource for gene and protein anno- M (2010) MZmine 2: modular framework for
tation. Nucleic Acids Res 44(D1):D457–D462 processing, visualizing, and analyzing mass
19. Keseler IM, Mackie A, Santos-Zavaleta A, spectrometry-based molecular profile data.
Billington R, Bonavides-Martı́nez C, Caspi R, BMC Bioinf 11:395. https://doi.org/10.
Fulcher C, Gama-Castro S, Kothari A, 1186/1471-2105-11-395
Krummenacker M, Latendresse M, Muñiz- 28. Ramirez-Gaona M, Marcu A, Pon A, Guo AC,
Rascado L, Ong Q, Paley S, Peralta-Gil M, Sajed T, Wishart NA, Karu N, Djoumbou
Subhraveti P, Velázquez-Ramı́rez DA, Feunang Y, Arndt D, Wishart DS (2017)
Weaver D, Collado-Vides J, Paulsen I, Karp YMDB 2.0: a significantly expanded version
PD (2017) The EcoCyc database: reflecting of the yeast metabolome database. Nucleic
new knowledge about Escherichia coli k-12. Acids Res 45:D440–D445
Nucleic Acids Res 45:D543–D550 29. Rasche F, Svatoš A, Maddula RK, Böttcher C,
20. Kim S, Thiessen PA, Bolton EE, Chen J, Fu G, Böcker S (2011) Computing fragmentation
Gindulyte A, Han L, He J, He S, Shoemaker trees from tandem mass spectrometry data.
BA, Wang J, Yu B, Zhang J, Bryant SH (2016) Anal Chem 83(4):1243–1251. https://doi.
PubChem substance and compound databases. org/10.1021/ac101825k
Nucleic Acids Res 44:D1202–D1213. https:// 30. Ridder L, van der Hooft JJJ, Verhoeven S, de
doi.org/10.1093/nar/gkv951 Vos RCH, Bino RJ, Vervoort J (2013) Auto-
21. Klekota J, Roth FP (2008) Chemical substruc- matic chemical structure annotation of an
tures that enrich for biological activity. Bioin- LC-MS(n) based metabolic profile from green
formatics 24(21):2518–2525. https://doi. tea. Anal Chem 85(12):6033–6040. https://
org/10.1093/bioinformatics/btn479 doi.org/10.1021/ac400861a
22. Larson EA, Hutchinson CP, Lee YJ (2018) Gas 31. Röst HL, Sachsenberg T, Aiche S, Bielow C,
chromatography-tandem mass spectrometry of Weisser H, Aicheler F, Andreotti S, Ehrlich
lignin pyrolyzates with dopant-assisted atmo- HC, Gutenbrunner P, Kenar E, Liang X,
spheric pressure chemical ionization and Nahnsen S, Nilse L, Pfeuffer J,
molecular structure search with CSI:FingerID. Rosenberger G, Rurik M, Schmitt U, Veit J,
J Am Soc Mass Spectrom 29(9):1908–1918. Walzer M, Wojnar D, Wolski WE, Schilling O,
https://doi.org/10.1007/s13361-018-2001- Choudhary JS, Malmström L, Aebersold R,
3 Reinert K, Kohlbacher O (2016) OpenMS: a
23. Ludwig M, Dührkop K, Böcker S (2018) flexible open-source software platform for mass
Bayesian networks for mass spectrometric spectrometry data analysis. Nat Methods 13
metabolite identification via molecular finger- (9):741–748. https://doi.org/10.1038/
prints. Bioinformatics 34(13):i333–i340. nmeth.3959
https://doi.org/10.1093/bioinformatics/ 32. Ruttkies C, Schymanski EL, Wolf S, Hollender J,
bty245. Proceedings of Intelligent Systems for Neumann S (2016) MetFrag relaunched: incor-
Molecular Biology (ISMB 2018) porating strategies beyond in silico fragmenta-
24. Meusel M, Hufsky F, Panter F, Krug D, tion. J Cheminform 8:3. https://doi.org/10.
Müller R, Böcker S (2016) Predicting the pres- 1186/s13321-016-0115-9
ence of uncommon elements in unknown bio- 33. Schymanski EL, Ruttkies C, Krauss M,
molecules from isotope patterns. Anal Chem Brouard C, Kind T, Dührkop K, Allen FR,
88(15):7556–7566. https://doi.org/10. Vaniya A, Verdegem D, Böcker S, Rousu J,
1021/acs.analchem.6b01015 Shen H, Tsugawa H, Sajed T, Fiehn O,
25. Mohimani H, Gurevich A, Shlemov A, Ghesquière B, Neumann S (2017) Critical assess-
Mikheenko A, Korobeynikov A, Cao L, ment of small molecule identification 2016: auto-
Shcherbin E, Nothias LF, Dorrestein PC, mated methods. J Cheminf 9:22. https://doi.
Pevzner PA (2018) Dereplication of microbial org/10.1186/s13321-017-0207-1
metabolites through database search of mass 34. Shinbo Y, Nakamura Y, Altaf-Ul-Amin M,
spectra. Nat Commun 9(1):4035. https:// Asahi H, Kurokawa K, Arita M, Saito K,
doi.org/10.1038/s41467-018-06082-8 Ohta D, Shibata D, Kanaya S (2006) KNAp-
26. Nelson SJ, Johnston WD, Humphreys BL SAcK: a comprehensive species-metabolite
(2001) Relationships in medical subject head- relationship database. In: Saito K, Dixon RA,
ings. In: Bean CA, Green R (eds) Relationships Willmitzer L (eds) Plant metabolomics.
 Metabolite Identification Using SIRIUS 207

Biotechnology in agriculture and forestry, vol 41. Wang M et al (2016) Sharing and community
57. Springer, Berlin, pp 165–181 curation of mass spectrometry data with Global
35. Steinbeck C, Han Y, Kuhn S, Horlacher O, Natural Products Social molecular networking.
Luttmann E, Willighagen E (2003) The Nat Biotechnol 34(8):828–837. https://doi.
Chemistry Development Kit (CDK): an open- org/10.1038/nbt.3597
source Java library for chemo- and bioinfor- 42. Weber RJM, Li E, Bruty J, He S, Viant MR
matics. J Chem Inf Comput Sci 43:493–500 (2012) MaConDa: a publicly accessible mass
36. Tautenhahn R, Cho K, Uritboonthai W, spectrometry contaminants database. Bioinfor-
Zhu Z, Patti GJ, Siuzdak G (2012) An acceler- matics 28(21):2856–2857. https://doi.org/
ated workflow for untargeted metabolomics 10.1093/bioinformatics/bts527
using the METLIN database. Nat Biotechnol 43. Willighagen EL, Mayfield JW, Alvarsson J,
30(9):826–828. https://doi.org/10.1038/ Berg A, Carlsson L, Jeliazkova N, Kuhn S,
nbt.2348 Pluskal T, Rojas-Chertó M, Spjuth O,
37. Tsugawa H, Kind T, Nakabayashi R, Torrance G, Evelo CT, Guha R, Steinbeck C
Yukihira D, Tanaka W, Cajka T, Saito K, (2017) The Chemistry Development Kit
Fiehn O, Arita M (2016) Hydrogen rearrange- (CDK) v2.0: atom typing, depiction, molecular
ment rules: computational ms/ms fragmenta- formulas, and substructure searching. J Che-
tion and structure elucidation using minf 9(1):33. http://dx.doi.org/10.1186/
MS-FINDER software. Anal Chem 88 s13321-017-0220-4
(16):7946–7958. https://doi.org/10.1021/ 44. Wishart DS, Feunang YD, Marcu A, Guo AC,
acs.analchem.6b00770 Liang K, Vázquez-Fresno R, Sajed T,
38. Wang R, Fu Y, Lai L (1997) A new atom- Johnson D, Li C, Karu N, Sayeeda Z, Lo E,
additive method for calculating partition coef- Assempour N, Berjanskii M, Singhal S,
ficients. J Chem Inf Comput Sci 37 Arndt D, Liang Y, Badran H, Grant J, Serra-
(3):615–621. https://doi.org/10.1021/ Cayuela A, Liu Y, Mandal R, Neveu V, Pon A,
ci960169p Knox C, Wilson M, Manach C, Scalbert A
39. Wang R, Gao Y, Lai L (2000) Calculating par- (2018) HMDB 4.0: the human metabolome
tition coefficient by atom-additive method. database for 2018. Nucleic Acids Res 46(D1):
Perspect Drug Discov Des 19(1):47–66. D608–D617. http://dx.doi.org/10.1093/
https://doi.org/10.1023/A:1008763405023 nar/gkx1089
40. Wang Y, Kora G, Bowen BP, Pan C (2014) 45. Wolf S, Schmidt S, Müller-Hannemann M,
MIDAS: a database-searching algorithm for Neumann S (2010) In silico fragmentation for
metabolite identification in metabolomics. computer assisted identification of metabolite
Anal Chem 86(19):9496–9503. https://doi. mass spectra. BMC Bioinf 11:148. https://doi.
org/10.1021/ac5014783 org/10.1186/1471-2105-11-148
 Chapter 12

Annotation of Specialized Metabolites from High-

Throughput and High-Resolution Mass Spectrometry
Metabolomics
Thomas Naake, Emmanuel Gaquerel, and Alisdair R. Fernie

Abstract
High-throughput mass spectrometry (MS) metabolomics profiling of highly complex samples allows the
comprehensive detection of hundreds to thousands of metabolites under a given condition and point in
time and produces information-rich data sets on known and unknown metabolites. One of the main
challenges is the identification and annotation of metabolites from these complex data sets since the number
of authentic standards available for specialized metabolites is far lower than an account for the number of
mass spectral features. Previously, we reported two novel tools, MetNet and MetCirc, for putative
annotation and structural prediction on unknown metabolites using known metabolites as baits. MetNet
employs differences between m/z values of MS1 features, which correspond to metabolic transformations,
and statistical associations, while MetCirc uses MS/MS features as input and calculates similarity scores
of aligned spectra between features to guide the annotation of metabolites. Here, we showcase the use of
MetNet and MetCirc to putatively annotate metabolites and provide detailed instructions as to how
those can be used. While our case studies are from plants, the tools find equal utility in studies on bacterial,
fungal, or mammalian xenobiotic samples.

Key words Annotation, Plant metabolite, Specialized metabolite, Unknown metabolite, Metabolic
modification, Molecular networking

1 Introduction

Based on estimations, the plant kingdom contains between

100,000 and 1 million metabolites [1, 2] with any species contain-
ing more than 5000 metabolites [3]. While the core central metab-
olism is comparable to nonplant species, a wide array of specialized
(formerly known as secondary) metabolites contributes to the high
number of predicted metabolites in plants. As summarized in
Alseekh and Fernie [4], it is estimated that plants are able to
synthesize about 27,000 different alkaloids [5–7], 10,000 different
phenolics [6–8] and 16,000 different terpenoid compounds
[6, 7]. Given this, one crucial aspect of plant metabolomics lies in

Shuzhao Li (ed.), Computational Methods and Data Analysis for Metabolomics, Methods in Molecular Biology, vol. 2104,
https://doi.org/10.1007/978-1-0716-0239-3_12, © Springer Science+Business Media, LLC, part of Springer Nature 2020

209
210 Thomas Naake et al.

the de novo annotation of natural products requiring software

solutions that are not based on previously reported metabolites
(cf. [9]), that is, to select metabolite candidates for further identifi-
cation in an unbiased way.
High-throughput mass spectrometry offers an unbiased
approach to capture in a comprehensive manner the metabolic
state of cells and tissues under given experimental conditions. For
highly complex biological matrices (e.g., from plant tissues),
hundreds to thousands of metabolites are typically detected within
a liquid chromatography (LC) mass spectrometry (MS) profile [10]
imposing difficulties in accurately processing this high number of
metabolites [4]. While for gas chromatography (GC)-MS extensive
database entries for primary metabolites exist (e.g., [11]), this is less
common for LC-MS data [10]. Often LC-MS data contain a high
proportion of unknown metabolites and it is a critical challenge in
metabolic profiling to select metabolic features for further identifi-
cation. Often the metabolites measured by LC-MS belong to the
class of specialized metabolites. These metabolites typically display
a highly diverse chemistry [12], exhibit a high dynamic range
depending on environmental conditions [3] and on enzymatic
capabilities of the species or ecotype (e.g., [13]). They, further-
more, usually display strong compartmentalization at the tissue,
cellular, and subcellular levels [14], requiring diverse and deliberate
sampling and methods which intensifies the analysis of these highly
complex samples.
Metabolic profiles can be acquired either in MS1 or in MS2
mode. While MS1 profiling allows for rapid throughput of many
samples, MS2 captures structural information on metabolic fea-
tures by fragmentation. For MS2, two modes of data acquisition
exist: data-dependent acquisition (DDA) and data-independent
acquisition (DIA). While DDA selects a certain number of precur-
sor ions for fragmentation, DIA uses all ions (or ions in a certain m/
z range) for fragmentation and thus allows to comprehensively
capture fragmentation data on detected metabolic features.
Different software solutions exist to reconstruct metabolic net-
works and/or putatively annotate metabolites (reviewed in [10]).
Starting from MS1 data, typically peak tables are annotated for
isotopes, fragments and adducts (e.g., by CAMERA [15] that inter-
acts directly with the processed peak data from xcms or AStream
[16] that reduces redundancy between various chemical forms in
one compounds and removed noise artifacts and potential sample
contamination). Several software solutions exist to putatively anno-
tate metabolic features within metabolomics datasets: xMSanno-
tator [17] employs metabolic pathway associations by combining
data-driven approaches (using intensity profiles, retention time,
mass defect, isotope, and adduct characteristics) and chemical,
metabolic, and environmental databases. PlantMat [18] matches
features of the dataset against combinatorial enumerations of
 Plant Metabolite Annotation 211

predefined aglycones and decorating building blocks (e.g., glycosyl

subunits). Another tool, mummichog [19], uses network analysis
for pathway enrichment, prediction of functional activity and puta-
tive metabolite identification. While these tools are valuable when it
comes to analysis of well-defined metabolomes with a good cover-
age of metabolites (xMSannotator), known aglycones (PlantMat)
or availability of genome-scale metabolic models (mummichog),
other software tools are needed to putatively annotate de novo
metabolites from MS1 data.
For MS2 data, Mass2Motifs [9, 20] applies Latent Dirichlet
Allocation to metabolomics data sets to assign conserved fragments
and losses of metabolites that reflect chemical substructures. This
approach allows to group molecules based on shared substructures
regardless of spectral similarity. Another approach is MetFamily, to
detect regulated metabolite families by hierarchical cluster analysis
of MS2 spectra and linking these clusters to MS1 data [21].
We previously created two novel open-source software tools,
MetNet [22] and MetCirc [23], for dedicated putative annotation
of metabolic features acquired by MS1 and MS2 (DDA and DIA
alike) mass spectrometry. Both MetNet and MetCirc use molecu-
lar networking to guide metabolite annotation. Network construc-
tion from high-resolution MS1 data was pioneered by Breitling
et al. [24] who proposed to use pair-wise differences of metabolic
features for network construction. Similarly, MetNet uses the
information-rich output of MS1 metabolic profiling to link features
based on structural information and quantitative information (i.e.,
intensity values). In brief, MetNet matches all pair-wise differences
of metabolic features against a table of known metabolic transfor-
mations that typically occur or are specifically searched for (e.g.,
glycosylation or hydroxylation). Furthermore, MetNet creates net-
works from quantitative data by statistically associating metabolic
features to each other. Metabolites that share a specific pathway are
normally coregulated and show a statistical association (e.g., [25–
27]). The two results are combined to a consensus network that
contains both information from structural and quantitative data
and enables to formulate structural hypotheses on unknown mass
signals linked to known metabolites.
For MS2 data, MetCirc uses molecular networking [28, 29]
based on spectral similarity scores between MS2 features based on a
modified normalized dot product (NDP, see [28] for further
details) taking into account either shared fragments or main neutral
losses. Small molecules, as produced by plants, often return few
fragments during MS2 fragmentation. Sometimes only a few frag-
ments are shared among MS2 spectra and diagnostic for a certain
compound family. Compared to the classically used MS/MS simi-
larity scoring based on shared fragments, we completed this scoring
procedure by taking into account a second similarity measure
incorporating shared neutral losses, which help to obtain
212 Thomas Naake et al.

biochemically meaningful MS/MS groupings that can be further

analyzed. While MetCirc was initially designed to facilitate the
exploration of MS2 metabolomics data obtained from cross-spe-
cies/cross-tissues comparative experiments, MetCirc also
improves the dereplication of known and unknown metabolites
within its interactive Shiny interface.
In this chapter, we provide stepwise protocols for the use of
MetNet and MetCirc in order to achieve compound class group-
ing of MS1 and MS2 data and for the annotation of structural
similarities among clustered MS2. To this end, we employed a
data set of MS2 spectra produced from the UPLC-qTOF+MS
analysis of polar to semipolar extracts of Nicotiana attenuata and
collection of MS2 spectra using a DIA method (Metabolights
MTBLS335). N. attenuata has been developed as ecological
model system to elucidate the role of specialized metabolites in
mediating plant-insect interactions. Main specialized metabolites
detected in positive ionization mode from leaf methanolic extracts
comprise alkaloids, flavonoid glycosides, hydroxycinnamic-
polyamine conjugates, O-acyl sugars, and
17-hydroxygeranyllinalool diterpene glycosides (17-HGL-DTG)
[30, 31]. A previous metabolomics study reported that approxi-
mately 360 nonredundant compound-level deconvoluted MS2
spectra are typically collected by DIA MS2 from N. attenuata leaf
extracts, of those around 60 compounds are annotated according
to levels 1 and 2 of the Metabolomics Standard Initiative [30–
32]. For this dataset, scores higher than 0.8 were retrieved for
less than 10% of the MS2 data set when querying (in 2015) MS2
spectra against entries from the Massbank public depository
[28]. The abovementioned 17-HGL-DTG compound class corre-
sponds to abundant 17-hydroxygeranyllinalool backbones deco-
rated with sugar and malonyl moieties and are employed as
efficient defenses against phytophagous insects [33]. The
17-HGL-DTG chemotype comprises more than 40 predicted com-
pounds in a given plant tissue and the rich fragmentation patterns
produced by these compounds provide sufficient resolution for a
priori structure annotation by mass spectral alignment
[32]. Hence, an MS2 cluster enriched in 17-HGL-DTGs is used
to illustrate MetCirc and MetNet annotation strategies from plant
metabolomics profiles.
Noteworthy, while both programs were written with the analy-
sis of plant metabolism in mind, they will be equally applicable to
annotate specialized metabolites of other organisms as well as
xenobiotics.
 Plant Metabolite Annotation 213

2 Materials

2.1 Software 1. R software (currently version
3.5 required) available via

Requirements https://www.r-project.org/bin/.
and Packages 2. optional: install RStudio (v1.0.153 or higher) available via
https://www.rstudio.com/products/rstudio/download/.
3. Internet connection to download BiocManager from CRAN
(https://cran.r-project.org/) and MetNet and/or MetCirc
package from Bioconductor (https://bioconductor.org/).
4. MetNet v1.0.0, MetCirc v1.12.0 (as used here, see Notes 1
and 2).
5. optional: amap v0.8-16 for clustering (as used here) or a similar
package.
6. Internet explorer, recommended: Firefox (version 63.0 or
higher) or Google Chrome (version 71.0.3578.80 or higher)
(MetCirc).

2.2 Input Data 1. m  n peak table of m MS1 features containing m/z values,
retention time and intensity values for n samples (MetNet, see
Notes 3–5). A peak table can be acquired by following the
protocol of Shimizu et al. [34] using an Orbitrap mass spec-
trometer or similar and running the below-mentioned xcms/
CAMERA script.
2. MS2 data from DDA or DIA mass spectrometry formatted as a
peak table or as a .msp file (MetCirc, see Note 4). A peak table
can be acquired according to Li et al. [28] using an Orbitrap or
qTOF mass spectrometer.

2.3 Hardware 1. Desktop computer.

Requirements 2. preferably: R Server environment with
64 GB RAM for large
peak tables with several thousand metabolic features (MetNet).

3 Methods

3.1 Prerequisites 1. if required: Install R (https://cran.r-project.org/bin/)

depending on the operating system.
2. optional: Install RStudio (https://www.rstudio.com/products/
rstudio/download/) depending on the operating system.
3. Open R Session, either by opening R, Rstudio or (under
Linux) by entering R to the Terminal and hitting Enter.
4. Install BiocManager by entering to the R Session and hitting
enter:

if (!requireNamespace("BiocManager", quietly = TRUE))

install.packages("BiocManager")
214 Thomas Naake et al.

3.2 MetNet 1. Install MetNet via Bioconductor framework by entering to the

R Session and hitting enter:

BiocManager::install("MetNet", version = "3.8")

2. Load the namespace of the package and attach it to the search

space to the R environment by entering to the R Session and
hitting Enter.

library(MetNet)

3. Load the peak table to the R session. This step will differ
depending on how the peak table is stored. If the peak table
is stored in tabular format in a file (here in the peaklist.txt file)
load the file by

peaklist <- read.table(“path_to_peaklist.txt_file/peaklist.

txt”)

MetNet is compatible with the output of xcms/CAMERA

[15, 35] pipeline in R. Alternatively to the step by loading an
existing peak table by read.table a peak list can also be
created from NetCDF/mzXML by using xcms/CAMERA, for
example, by entering to the R console.

xset <- xcmsSet(method = "centWave", ppm = 30, snthresh = 10, peakwidth = c(5,20))
xset2 <- group(xset, method = "density", minfrac = 0.5, minsamp = 2, bw = 2, mzwid =
0.025)
xset3 <- retcor(xset2, family = "s", plottype = "m", missing = 1, extra = 1, span = 1)
xset4 <- group(xset3, method = "density", bw = 2, mzwid = 0.025, minfrac = 0.5,
minsamp = 2)
xset5 <- fillPeaks(xset4, method = "chrom")
an <- xsAnnotate(xset5)
anF <- groupFWHM(an, perfwhm = 0.6)
anI <- findIsotopes(anF, mzabs = 0.01)
anIC <- groupCorr(anI, cor_eic_th = 0.75, graphMethod = "lpc")
anFA <- findAdducts(anIC, polarity = "positive")
peaklist <- getPeaklist(anFA)

Parameters will depend on the specifications of the

employed chromatography and mass spectrometry system
according to [36].
4. if required: After import of the peak table the peak table needs
to be normalized (e.g., each column by its 75% quantile and
log2 transformation of intensity values) (see Note 6),
5. Define the transformation that will be looked for (and if
required: associated changes in retention time, see Note 7). As
an example we want to search for transformations that are
 Plant Metabolite Annotation 215

based on hydroxylation (associated with earlier retention time

on a reverse-phase column, +15.99), malonylation (higher
retention time, +86.00), and addition of rhamnose (earlier
retention time, +146.06), glucose (earlier retention time,
+162.05). Create the table with R by entering the command
lines below and hitting Enter.

transformation <- rbind(

c("Hydroxylation (-H)", "O", 15.9949146221, "-"),
c("Malonyl group (-H2O)", "C3H2O3", 86.0003939305, "+"),
c("Rhamnose (-H2O)", "C6H10O4", 146.0579088094, "-"),
c("Glucose (-H2O)", "C6H10O5", 162.0528234315, "-"))
transformation <- data.frame(group = transformation[,1],
formula = transformation[,2],
mass = as.numeric(transformation[,3]), rt = transformation[,4])

6. Create the adjacency matrix that reports a connection between

two metabolic features if the respective difference in m/z values
matches to a transformation defined in the data.frame object
transformation. Molecular weight difference wX is defined by
wX ¼ |wA  wB|, where wA is the molecular weight of substra-
te A, and wB is the molecular weight of product B [24] (typi-
cally, m/z values will be used as a proxy for the molecular
weight since the molecular weight is not directly derivable
from MS1 data, see Note 7). The function createStructur-
alAdjacency (see Note 8) will then return this adjacency
matrix together with another matrix telling about the type of
connection (in the above case if there is addition/loss of
hydroxyl, malonyl, rhamnose or glucose moiety). Enter the
following to the R console and hit Enter.

struct_adj <- createStructuralAdjacency(peaklist, transformation,

ppm = 10)

7. optional: Check the proposed connection by their retention

time difference and remove false positive connection that had
by chance the same/similar m/z difference but elute before or
after the reference feature (see Notes 3 and 9–11). For example
on data acquired by reverse-phase LC-MS, if there is a connec-
tion between features based on the difference of glucose addi-
tion and the feature with the higher m/z value elutes later,
rtCorrection will remove the connection between these
features, since the addition of glucose should render the feature
more polar, thus eluting earlier. The expected retention time
shift has to be given in the column “rt” of the transforma-
tion object. To adjust for retention time enter to the R
console and hit Enter.
216 Thomas Naake et al.

struct_adj_rt <- rtCorrection(struct_adj, peaklist, transformation)

8. Create the adjacency matrix based on statistical associations of

intensity values between metabolic features. The createSta-
tisticalAdjacency function provided by MetNet allows
for an approach proposed by [37] in that it uses different
statistical frameworks and combines the results from the differ-
ent statistical models into a consensus adjacency matrix. Cur-
rently, the following methods are implemented: Least absolute
shrinkage and selection operator [38], Random Forest [39],
Pearson and Spearman correlation (including partial and semi-
partial correlation) [25], context likelihood of relatedness
(CLR) [40] the algorithm for the reconstruction of accurate
cellular networks (ARACNE) [41] and constraint-based struc-
ture learning [42] (Bayesian networks, see Notes 12–15). By
way of example, we will create the consensus adjacency matrix
by using Pearson and Spearman correlation and the models
CLR and ARACNE by entering to the R console and hitting
Enter.

stat_adj <- createStatisticalAdjacency(peaklist_cut,

model = c("pearson", "spearman", "clr", "aracne"), correlation_adjust = "BH")

9. Combine the adjacency matrices from structural and statistical

association. Only these connections will be reported in the final
“consensus” network that are shared in the two adjacency
matrix are present in both adjacency matrices (intersection is
taken, see Note 14). To create the consensus matrix enter to
the R console and hit Enter.

cons_adj <- combineStructuralStatistical(struct_adj[[1]], stat_adj)

cons_adj <- combineStructuralStatistical(struct_adj_rt[[1]], stat_adj)

10. if required: After this step the MetNet pipeline is finished.

Commonly used network analysis tools (e.g., from igraph
or sna package) can be employed to analyze the network.
Known metabolites should guide the putative annotation of
unknown metabolites to formulate structural hypotheses
(cf. Fig. 1).
Here we will retrieve the components from the network
(subgraphs in which any two vertices are connected to each
other) that will contain possible associated metabolites by
entering to the R console and hitting Enter.

library(igraph)
net <- graph_from_adjancency_matrix(cons_adj, mode = ”undirected”, diag = FALSE)
net_comp <- components(net)
 Plant Metabolite Annotation 217

Fig. 1 Typical result of MetNet pipeline. MetNet created based on a Nicotiana MS2 dataset, a network
considering the m/z shifts corresponding to hydroxylation, malonylation, rhamnosylation, and glucosylation
and statistical associations using Pearson and Spearman correlation as well as CLR and ARACNE as statistical
models (see also [22] for further details). For the whole dataset, 2,417,704 and 3924 links between metabolic
features were determined for statistical and structural associations (intersection of 2026 links). Displayed are
network components with 10 or more members. Network #13 is enriched in diterpene glycosides (DTG) but
also contains unknown metabolites collected in the dataset. Based on m/z differences to known DTG, these
unknowns can be putatively annotated. For network #13, 704 and 87 links between metabolic features were
determined for statistical and structural annotations (intersection of 71 links)

3.3 MetCirc 1. Install MetCirc via Bioconductor framework by entering to

the R Session and hitting enter:

BiocManager::install("MetCirc", version = "3.8")

2. Load the namespace of the package and attach it to the search

space to the R environment by entering to the R Session and
hitting Enter.

library(MetCirc)

3. Load MS2 data and convert to the MSP class by either two
ways:
(a) Load a data frame (here msms.txt) with column “id”
(unique identifiers for MS/MS features), and columns
“mz” and “intensity” comprising the fragment ions and
their intensities (see Notes 16–19). The data frame can be
loaded to the R session by entering and hitting Enter.

msms <- read.table(“path_to_msms.txt/msms.txt")

msp <- convert2MSP(msms)
218 Thomas Naake et al.

(b) Load a data frame in .msp file format, a typical data format
for storing MS/MS libraries. Required properties of such
a data frame are the name of the metabolite (row entry
“NAME:”), the m/z value of the precursor ion (“PRE-
CURSORMZ” or “EXACTMASS:”), the number of
peaks of the feature (“Num Peaks:”), and information
on fragments and peak areas (see Note 20). To load the
file in .MSP format and convert it to the MSP class object,
enter the following to the R session and hit Enter.

msms <- read.table(“path_to_msms.msp/msms.msp")

msp <- convertMSP2MSP(msms)

4. optional: Calculate neutral losses from precursor ions by enter-

ing to the R console and hitting Enter.

msp <- msp2FunctionalLossesMSP(msp)

5. optional: Combine several MSP objects (here msp1, msp2) to

one MSP object (here msp) by entering to the R console and
hitting Enter.

msp <- combine(msp1, msp2)

6. Bin the fragments. Due to slight differences in m/z values over

measurements, fragments might have m/z values which differ
from other fragments even though they are in theory identical.
Binning will bin together fragment ions which are similar to
allow for comparison between m/z values (see Notes 21–23) by
entering to the R console and hitting enter.

binnedMSP <- binning(msp = msp, tol = 0.01)

7. Similarity between two MS/MS features is calculated accord-

ing to the NDP. For a considered MS/MS pair, peak intensities
of shared m/z values for precursor/fragment ions and neutral
losses are employed as weights WS1,i and WS2,i within the
following formula:
P 2
W S1,i  W S2,i
NDP ¼ P 2 P ,
W S1,i  W 2S2,i
with S1 and S2 the spectra 1 and 2, respectively, of the ith
of j common peaks. Weights are calculated according to
W ¼ [peak intensity]m · [m/z]n, with m ¼ 0.5 and n ¼ 2 as
default values as suggested by MassBank [28]. Calculate simi-
larity (see Notes 24–25) according to the NDP by entering to
the R console and hitting enter.
 Plant Metabolite Annotation 219

A Clustering of MS/MS features

10000000 15000000 20000000 25000000 B MetCirc visualization

Species 1
Species 2
Height

all MS/MS
similarity
edges
Visualized similarity edges
for a given threshold
(here = 1, identical MS/MS)
5000000

A
Species 3
0

#1 #2 #3 #4 #5 Selected
Species 4 MSMS (here mz 498.26 / RT 596s)
C Cluster #2

N. attenuata
N. obtusifolia
N. clevelandii
N. quadrivalvis
N. x obtusiata 10/27
N. x obtusiata 57/126

MS/MS similarity = 1

Nicotianoside X

Fig. 2 Typical result of MetCirc pipeline. (a) Hierarchical clustering of MS/MS features based on distance
between similarity scores prior to MetCirc analysis. Clustering will extract clusters containing highly similar
MS/MS features (with high similarity scores). (b) Interface of the MetCirc Shiny application. Within the Shiny
application MS/MS features can be navigated, reordered, annotated, and links between MS2 features
thresholded and selected based on the type of links. (c) Exemplary MetCirc output for diterpene glycosides
(DTG) for six Nicotiana species. For all panels, MS2 spectra were collected from leaves of six Nicotiana
species and used for processing through the MetCirc pipeline. MetCirc visualization is built for Cluster #2
enriched in DTGs. From this compound class, nicotianoside X previously identified in Nicotiana attenuata is
selected in the interface. Similarity threshold is fixed to 1 (cross-species MS2 identity) which readily allows for
the dereplication (reidentification) of nicotianoside X in three of the five other Nicotiana species tested

similarityMat <- createSimilarityMatrix(binnedMSP)

8. optional: Clustering of MS/MS features based on similarity

score (cf. Fig. 2a). In case the library of MS/MS feature is
high in number, highly related features can be extracted using
clustering. By way of example, to define five clusters by Spear-
man correlation as distance measure and retrieve the similarity
matrix with members of cluster 2, enter the following to the R
console and hit Enter.
220 Thomas Naake et al.

hClustMSP <- amap::hcluster(similarityMat, method = "spearman")

cutTreeMSP <- cutree(hClustMSP, k = 5)
cluster2 <- names(cutTreeMSP)[as.vector(cutTreeMSP) == "2"]
similarityMat_cluster2 <- similarityMat[cluster2, cluster2]

9. The adjacency matrix can be visualized within an interactive

shiny session. To start the visualization enter to the R console
(see Note 26) and hit Enter.

selectedFeatures <- shinyCircos(similarityMat, msp = msp)

The shiny interface will be loaded, upon which the type of

link to be displayed can be selected in the left-side panel (inter-
class links for links between groups, intra-class links for links
within groups, all for all links within and between groups).
Ordering inside of groups can be changed by selecting the
respective option in the sidebar panel (based on similarity
(“clustering”), m/z value (“m/z,” in ascending order), reten-
tion time (“retention time,” in ascending order, see Fig. 2b). In
the “Appearance” tab, the size of the plot can be changed, the
precision of the displayed values for the m/z and retention time
values, as well as if a legend is displayed or not.
10. Change the slider input (“Threshold for similarity to display,”
see Fig. 2b) to threshold similarity scores within the plot (see
Notes 27 and 28).
11. Navigate through MS/MS features and display information by
single-clicking on the sectors in the circular plot. Update a
selected feature by changing information (name, class, infor-
mation, adduct) in the sidebar panel and click “Update anno-
tation” to save changes to the MSP-object msp. A highly
similar, known feature should be used as a reference
(cf. [32]). On exiting the application via “Stop and export
selected features” (see Note 29), the updated MSP-object will
be retrievable by.

selectedFeatures$msp

12. optional: Permanently select MS/MS features by double-

clicking on the sector. Reset all permanently selected features
by clicking on “Reset features.” Permanently selected features
will be returned to the selectedFeatures object when exiting
the application via “Stop and export selected features” (see
Note 29).
 Plant Metabolite Annotation 221

4 Notes

1. Consult the vignette of MetNet for updated functionality or

further information (https://bioconductor.org/packages/
release/bioc/vignettes/MetNet/inst/doc/MetNet.pdf).
2. Consult the vignette of MetCirc for updated functionality or
further information (https://bioconductor.org/packages/
release/bioc/vignettes/MetCirc/inst/doc/MetCirc.pdf).
3. Chromatographic resolution has to be sufficiently good and
reliable to assure that batch-wise retention time alignment
correction is technically feasible.
4. Required data-sets have to be generated using high-resolution
MS instruments (mass accuracy ideally <15 ppm).
5. Sampling of diverse conditions/tissues is preferred to have
strong differences between samples which will be captured by
statistical methods.
6. MetNet does not impose any requirements for data normaliza-
tion, filtering, etc. However, the user has to make sure that the
data is properly preprocessed. These include division by inter-
nal standard, log2 transformation of intensity values, noise
filtering, removal of features that do not represent mass fea-
tures/metabolites, removal of isotopes, etc.
7. Occurrence of stereoisomers and structural isomers impose
problems/redundancy for network inference that should be
taken into account during network interpretation.
8. The ppm parameter in the function createStructuralAd-
jacency should be set according to the mass spectrometer
setup.
9. For ambiguous metabolic transformations, the retention time
correction can also be omitted by setting the respective data
frame entry to “?”.
10. Expert knowledge is required to assess how a specific transfor-
mation will change the polarity of a molecule and running the
rtCorrection function should be handled with care.

11. Depending on the position of addition/loss of a specific group,

the retention time shift can be reverted (cf. [22]).
12. Make sure to use the correct abbreviations for the statistical
methods for createStatisticalAdjacency: “lasso” for
LASSO, “randomForest” for Random Forest, “clr” for CLR,
“aracne” for ARACNE, “pearson” for Pearson correlation,
“pearson_partial” or “pearson_semipartial” for partial or semi-
partial Pearson correlation, “spearman” for Spearman correla-
tion, “spearman_partial” or “spearman_semipartial” for partial
222 Thomas Naake et al.

or semipartial Spearman correlation, “bayes” for constraint-

based structure learning.
13. For the function createStatisticalAdjacency, addi-
tional parameters can be passed to the functions lasso, ran-
domForest, clr, aracne, correlation, bayes, and/or
consensusAdjacency (e.g., for alpha value adjustment or
adjustment of epsilon parameter in clr).
14. The parameter threshold in createStatisticalAdja-
cency and combineStructuralStatistical should be
set accordingly, if another method than “central.graph”
(default) is used.
15. Choice of statistical methods should depend on the specific
question in mind and the size of the peak table, typically
LASSO, Random Forest, and constraint-based structure
learning will take a long calculation time for large data sets.
16. When loading the data frame in tabular format, make sure that
in the “id” column, all entries belonging to one MS/MS
feature are identical for all fragments, that is, the column “id”
has to be a unique descriptor for MS/MS features.
17. When loading the data frame in tabular, make sure that the
intensity values are in percent (%, not decimal) with the inten-
sity value with the highest intensity has a value of 100.0 (%,
instead of 1.00) and all other fragments have the respective
relative intensity (e.g., 10% means 10% smaller intensity than
the feature with 100%).
18. When loading the data frame in tabular format, the “id” col-
umn has to contain the m/z value of the precursor ion, and
optionally the retention time (add “rt ¼ TRUE” to
convert2MSP).
19. When loading the data frame in tabular format, the data frame
may contain additional information that will be stored in the
MSP object and may help with dereplication: rt (retention time
of features), names (name of the MS/MS feature), information
(additional information about the MS/MS feature), classes
(class of the MS/MS feature), adduct (adduct ion name of
MS/MS feature). Make sure that for all instances (except rt),
all entries per MS/MS feature are identical.
20. When loading the data frame in .MSP file format, the data
frame may optionally contain the row entries “RETENTION-
TIME:” (containing the retention time of the precursor),
“CLASS” (class of MS/MS feature), “ADDUCTION-
NAME:”/“PRECURSORTYPE:” (adduct ion type of
MS/MS feature), and “INFORMATION:” (additional infor-
mation about MS/MS feature) for each MS/MS feature within
the data frame.
 Plant Metabolite Annotation 223

21. The tol parameter in the binning function has to be set

according to the mass spectrometry setting (e.g., 0.01 means
that fragments with a m/z difference of 0.01 to a bin will be put
in this bin).
22. The function binning bins fragments together based on min-
imal distance to bins which were calculated either by the mean
or the median of fragments according to the tol parameter.
The user can specify by the method argument if median or
mean will be used.
23. The function binning expects a vector (group argument) which
comprises membership of the entries in the msp object, to a
compartment, species, individual, and so on. If group is not
specified binning will create an internal dummy variable
group (“a” with the length of the msp object).
24. For calculation of the similarity matrix, a matrix that has per
row one MS/MS feature and in the corresponding columns
the fragments, entries of the matrix are intensity values (in %)
for a specific fragment ion (the function binning will return
such a matrix).
25. For calculation of the similarity matrix, care has to be taken
with setting the parameters m and n (default to m ¼ 0.5 and
n ¼ 2 as suggested by MassBank). When peak intensities
should not be taken into account, m should be set to m ¼ 0.
26. Before running shinyCircos, make sure that the features in
similarityMat and msp are identical and in same order.
27. Typically, MS/MS features are further analyzed that share a
NDP score
0.55 (cf. [43]); however, depending on the grade
of fragmentation also MS/MS pairs can be analyzed.
28. Threshold the similarity score <1 to exclude identical features
from visualization.
29. When stopping shinyCircos, always do via “Stop and export
selected features,” only via this way will the selected features
and modified msp file be exported to the R session.

Acknowledgments

T.N. acknowledges support by the IMPRS-PMPG program and

A.R.F. the support of Max Planck Society. E. G. acknowledges the
support by the Deutsche Forschungsgemeinschaft Excellence Ini-
tiative to the University of Heidelberg and by the Centre National
de la Recherche Scientifique.
224 Thomas Naake et al.

References
1. Dixon RA, Strack D (2003) Phytochemistry compound spectra extraction and annotation
meets genome analysis, and beyond. Phyto- of liquid chromatography/mass spectrometry
chemistry 62:815–816 data sets. Anal Chem 84:283–289
2. Rai A, Saito K, Yamazaki M (2017) Integrated 16. Alonso A, Julia A, Beltran A et al (2011)
omics analysis of specialized metabolism in AStream: an R package for annotating
medicinal plants. Plant J 90:764–787 LC/MS metabolomic data. Bioinformatics
3. Fernie AR, Trethewey RN, Krotzky AJ et al 27:1339–1340
(2004) Metabolite profiling: from diagnostics 17. Uppal K, Walker DI, Jones DP (2017) xMSan-
to systems biology. Nat Rev Mol Cell Biol notator: an R package for network-based anno-
5:763–769 tation of high-resolution metabolomics data.
4. Alseekh S, Fernie AR (2018) Metabolomics Anal Chem 89:1063–1067
20 years on: what have we learned and what 18. Qiu F, Fine DD, Wherritt DJ et al (2016)
hurdles remain? Plant J 94:933–942 PlantMAT: a metabolomics tool for predicting
5. Ziegler J, Facchini PJ (2008) Alkaloid biosyn- the specialized metabolic potential of a system
thesis: metabolism and trafficking. Annu Rev and for large-scale metabolite identifications.
Plant Biol 59:735–769 Anal Chem 88:11373–11383
6. Wink M (2004) Phytochemical diversity of sec- 19. Li SZ, Park Y, Duraisingham S et al (2013)
ondary metabolites. In: Goodman RM Predicting network activity from high through-
(ed) Encyclopedia of plant and crop science. put metabolomics. PLoS Comput Biol 9:
Marcel Dekker, New York, pp 915–919 e1003123
7. Wink M (2015) Modes of action of herbal 20. Van Der Hooft JJ, Wandy J, Barrett MP et al
medicines and plant secondary metabolites. (2016) Topic modeling for untargeted sub-
Medicines (Basel) 2:251–286 structure exploration in metabolomics. Proc
8. Tohge T, Alseekh S, Fernie AR (2014) On the Natl Acad Sci U S A 113:13738–13743
regulation and function of secondary metabo- 21. Treutler H, Tsugawa H, Porzel A et al (2016)
lism during fruit development and ripening. J Discovering regulated metabolite families in
Exp Bot 65:4599–4611 untargeted metabolomics studies. Anal Chem
9. Van Der Hooft JJJ, Wandy J, Young F et al 88:8082–8090
(2017) Unsupervised discovery and compari- 22. Naake T, Fernie AR (2019) MetNet: metabo-
son of structural families across multiple sam- lite network prediction from high-resolution
ples in untargeted metabolomics. Anal Chem mass spectrometry data in R aiding metabolite
89:7569–7577 annotation. Anal Chem 91:1768–1772
10. Perez De Souza L, Naake T, Tohge T et al 23. Naake T, Gaquerel E (2017) MetCirc: navigat-
(2017) From chromatogram to analyte to ing mass spectral similarity in high-resolution
metabolite. How to pick horses for courses MS/MS metabolomics data. Bioinformatics
from the massive web resources for mass spec- 33:2419–2420
tral plant metabolomics. Gigascience 6:1–20 24. Breitling R, Ritchie S, Goodenowe D et al
11. Kopka J, Schauer N, Krueger S et al (2005) (2006) Ab initio prediction of metabolic net-
[email protected]: the Golm metabolome data- works using Fourier transform mass spectrom-
base. Bioinformatics 21:1635–1638 etry data. Metabolomics 2:155–164
12. D’auria JC, Gershenzon J (2005) The second- 25. Steuer R (2006) Review: on the analysis and
ary metabolism of Arabidopsis thaliana: grow- interpretation of correlations in metabolomic
ing like a weed. Curr Opin Plant Biol data. Brief Bioinform 7:151–158
8:308–316 26. Morreel K, Saeys Y, Dima O et al (2014) Sys-
13. Li X, Svedin E, Mo HP et al (2014) Exploiting tematic structural characterization of metabo-
natural variation of secondary metabolism lites in Arabidopsis via candidate substrate-
identifies a gene controlling the glycosylation product pair networks. Plant Cell 26:929–945
diversity of dihydroxybenzoic acids in Arabi- 27. Gaquerel E, Kuhl C, Neumann S (2013)
dopsis thaliana. Genetics 198:1267 Computational annotation of plant metabolo-
14. Sweetlove LJ, Fernie AR (2013) The spatial mics profiles via a novel network-assisted
organization of metabolism within the plant approach. Metabolomics 9:904–918
cell. Annu Rev Plant Biol 64:723–746 28. Li D, Baldwin IT, Gaquerel E (2015) Navigat-
15. Kuhl C, Tautenhahn R, Bottcher C et al (2012) ing natural variation in herbivory-induced sec-
CAMERA: an integrated strategy for ondary metabolism in coyote tobacco
 Plant Metabolite Annotation 225

populations using MS/MS structural analysis. metabolite profiling using nonlinear peak
Proc Natl Acad Sci U S A 112:E4147–E4155 alignment, matching, and identification. Anal
29. Watrous J, Roach P, Alexandrov T et al (2012) Chem 78:779–787
Mass spectral molecular networking of living 36. Patti GJ, Tautenhahn R, Siuzdak G (2012)
microbial colonies. Proc Natl Acad Sci U S A Meta-analysis of untargeted metabolomic data
109:E1743–E1752 from multiple profiling experiments. Nat Pro-
30. Gaquerel E, Heiling S, Schoettner M et al toc 7:508–516
(2010) Development and validation of a liquid 37. Marbach D, Costello JC, Kuffner R et al
chromatography-electrospray ionization-time- (2012) Wisdom of crowds for robust gene net-
of-flight mass spectrometry method for work inference. Nat Methods 9:796
induced changes in Nicotiana attenuata leaves 38. Tibshirani R (1996) Regression shrinkage and
during simulated herbivory. J Agric Food selection via the lasso. J Roy Stat Soc B Met
Chem 58:9418–9427 58:267–288
31. Li DP, Heiling S, Baldwin IT et al (2016) 39. Breiman L (2001) Random forests. Mach
Illuminating a plant’s tissue-specific metabolic Learn 45:5–32
diversity using computational metabolomics 40. Faith JJ, Hayete B, Thaden JT et al (2007)
and information theory. Proc Natl Acad Sci Large-scale mapping and validation of Escher-
USA 113:E7610–E7618 ichia coli transcriptional regulation from a
32. Heiling S, Khanal S, Barsch A et al (2016) compendium of expression profiles. PLoS Biol
Using the knowns to discover the unknowns: 5:e8
MS-based dereplication uncovers structural 41. Margolin AA, Nemenman I, Basso K et al
diversity in 17-hydroxygeranyllinalool diter- (2006) ARACNE: an algorithm for the recon-
pene glycoside production in the Solanaceae. struction of gene regulatory networks in a
Plant J 85:561–577 mammalian cellular context. BMC Bioinfor-
33. Heiling S, Schuman MC, Schoettner M et al matics 7:S7
(2010) Jasmonate and ppHsystemin regulate 42. Scutari M (2010) Learning Bayesian networks
key malonylation steps in the biosynthesis of with the bnlearn R package. J Stat Softw
17-hydroxygeranyllinalool diterpene glyco- 35:1–22
sides, an abundant and effective direct defense
against herbivores in Nicotiana attenuata. Plant 43. Wolfender JL, Nuzillard JM, Van Der Hooft
Cell 22:273–292 JJJ et al (2019) Accelerating metabolite identi-
fication in natural product research: toward an
34. Shimizu T, Watanabe M, Fernie AR et al ideal combination of LC-HRMS/MS and
(2018) Targeted LC-MS analysis for plant sec- NMR profiling, in silico databases and chemo-
ondary metabolites. Methods Mol Biol metrics. Anal Chem 91(1):704–742
1778:171–181
35. Smith CA, Want EJ, O’maille G et al (2006)
XCMS: processing mass spectrometry data for
 Chapter 13

Feature-Based Molecular Networking for Metabolite

Annotation
Vanessa V. Phelan

Abstract
The Global Natural Product Social Molecular Networking (GNPS) platform leverages tandem mass
spectrometry (MS/MS) data for annotation of compounds. Molecular networks aid in the visualization
of the chemical space within a metabolomics experiment. Recently, molecular networking has been
combined with feature detection methods to yield Feature-Based Molecular Networking (FBMN).
FBMN allows for the discrimination of isomers within the molecular network, incorporation of quantitative
information generated by the feature detection tools into visualization of the molecular network, and
compatibility with forthcoming in silico annotation tools. This chapter provides step-by-step methods for
generating a molecular network to annotate microbial natural products using the Global Natural Product
Social Molecular Networking (GNPS) Feature-Based Molecular Networking (FBMN) workflow.

Key words Molecular networking, Secondary metabolism, Feature annotation, Specialized metabo-
lites, Natural Products, GNPS

1 Introduction

In liquid chromatography–mass spectrometry (LC-MS) based

metabolomics workflows, annotation of a metabolite as an “identi-
fied metabolite” (confidence level 1) [1] typically requires compar-
ing the accurate mass and retention time of a molecular feature to
an authentic standard. While this approach has proven to be very
robust in well-studied systems, it requires a priori knowledge of
metabolites of interest in the biological sample as well as availability
of authentic standards. In microbial metabolomics, specialized
metabolites (also known as secondary metabolites, natural pro-
ducts, quorum sensing molecules, and small molecular virulence
factors) are often of particular interest to the research community
due to their roles in treating human disease, in contributing to
infection, and as integral components of molecular communication
within microbiome communities [2–4]. However, microbial
specialized metabolite discovery is often hampered by the

Shuzhao Li (ed.), Computational Methods and Data Analysis for Metabolomics, Methods in Molecular Biology, vol. 2104,
https://doi.org/10.1007/978-1-0716-0239-3_13, © Springer Science+Business Media, LLC, part of Springer Nature 2020

227
228 Vanessa V. Phelan

rediscovery of characterized structures; most of these metabolites

are not commercially available.
To address the limitation of molecular annotation in microbial
metabolomics, Global Natural Products Social Molecular Network-
ing (GNPS, https://gnps.ucsd.edu) was developed [5]. GNPS in
an online bioinformatics platform for the analysis, sharing, and
curation of tandem mass spectrometry (MS/MS) data. While
GNPS was originally developed with microbial metabolites in
mind, the tools within GNPS can be applied to any MS/MS
metabolomics dataset.
The core principle of molecular networking at GNPS is the
assumption that molecules of similar structure will fragment simi-
larly during MS/MS data acquisition. Therefore, two structurally
related molecules will likely have similar fragment ion spectra. To
create a molecular network, all MS/MS spectra are aligned together
to determine fragmentation similarity, and assembled into a molec-
ular network. To aid in dereplication of molecules in the molecular
network, each MS/MS spectrum of a metabolite of interest is
scored based upon its spectral similarity to libraries of MS/MS
spectra of known compounds. In practice, annotation (confidence
level 2) [1] of MS/MS spectra of an experiment occurs in GNPS via
automated dereplication (identification of knowns), whereby all
MS/MS spectra collected during an experiment are scored against
all MS/MS libraries (https://gnps.ucsd.edu/ProteoSAFe/
libraries.jsp), including compound collections from the FDA and
NIH, third party spectral libraries, and user-submitted annotations.
If a user makes their data public via GNPS, that data will undergo
continuous identification—the periodic and automatic reanalysis of
the dataset against updated GNPS MS/MS libraries.
Further, the annotation of unknowns (not present in spectral
libraries) can be made based on upon the molecular networks.
Microorganisms often produce a variety of specialized metabolites
from a single biosynthetic pathway. By leveraging molecular net-
working’s grouping of similar MS/MS spectra of an entire dataset
into structurally related molecular families of metabolites [6], rapid
annotation (confidence level 3) [1] of analogues is possible. Molec-
ular networks can be visualized using either the in-browser visuali-
zation or in Cytoscape [7].
While molecular networking has been broadly utilized in meta-
bolomics, a general drawback of using the classic workflow is the
MSCluster algorithm [8], which combines nearly identical MS/MS
spectra to simplify visualization of a molecular network, does not
take into account retention time. Therefore, the MS/MS spectra of
isobaric species are combined into a single representative consensus
MS/MS spectrum in classic molecular networking, despite the
isobaric compounds separating chromatographically. Recently, fea-
ture finding programs have been integrated with molecular
 Feature Based Molecular Networking 229

networking [9] via the Feature-Based Molecular Networking

(FBMN) workflow to correct this limitation.
Feature finding is a data reduction step in LC-MS analysis,
whereby a feature is defined as a mass to charge (m/z) ratio,
retention time pair and the corresponding abundance. Ultimately,
processing LC-MS data using feature detection programs results in
a quantitative representation of all metabolites within a metabolo-
mics dataset. In FBMN, LC-MS/MS data is processed via feature
finding, yielding abundances from MS data and a single represen-
tative MS/MS spectrum is selected per feature. This workflow
allows users to import quantitative information derived from the
feature detection tools into the molecular networks, discriminate
isomers by retention time within the molecular network, and is
compatible with in silico annotation of MS/MS spectra. Addition-
ally, metadata such as bioactivity, media, and strain can be
incorporated into the network to allow for rapid visual interpreta-
tion of the LC-MS/MS data. This approach allows for annotation
of metabolites [1] based upon accurate mass and retention time of a
molecular feature to an authentic standard (confidence level 1), a
MS/MS spectral library (confidence level 2), and annotation of
analogues of known compounds (confidence level 3).
Herein, we use raw LC-MS/MS data from two Pseudomonas
aeruginosa PA14 strains, wild-type (WT) and a deletion mutant in
the rhlR gene (rhlR) to illustrate the utility of FBMN. The rhlR
gene is a key transcriptional regulator of the quorum sensing
machinery of P. aeruginosa and regulates the production of a
number of molecular families including the phenazines, quino-
lones, rhamnolipids, and siderophores [10–12]. FBMN provides a
platform to rapidly annotate known microbial metabolites and their
analogues as well as prioritizing features for further analysis by
visualizing differential abundance via the molecular network.

2 Materials

2.1 Data The raw data (Bruker .d file format and centroided mzXML file
format) and the metadata table used for the step-by-step instruc-
tions below, a batch file containing the settings for processing the
raw data in MZmine2 [13], and all resulting files from the
MZmine2 and GNPS processing workflows can be downloaded
free of charge at https://massive.ucsd.edu or accessed via the
GNPS menu option “MassIVE Datasets” via accession number
MSV000083500 (see Note 1).

2.2 mzMine 2 The current version of MZmine2 is available free of charge from the
project web page http://mzmine.github.io.
230 Vanessa V. Phelan

2.3 GNPS Academic, government, and not-for-profit researchers can create a

free account on the GNPS website https://gnps.ucsd.edu and use
the associated computational workflows free of charge.

2.4 Cytoscape The current version of Cytoscape is available free of charge from the
project web page https://cytoscape.org.

3 Methods

3.1 Create The metadata file provides the ability to incorporate sample prop-
a Metadata File erty data into visualization and analysis of FBMN in Cytoscape.
The metadata format is a tab separated text file, that must be
generated using a text editor (see Note 2).
In order for the metadata file to be processed properly in
GNPS, the format of the file must be correct (Fig. 1). The first
column must be titled “filename” in all lowercase. The file names in
the “filename” column must be the full filename. The filenames
should each be unique, identical to the names of the ones used for
the feature finding process described in Subheading 3.2, and not
contain any path information. The columns containing metadata
must be prefixed by “ATTRIBUTE_” (see Note 3). The “ATTRI-
BUTE_” columns should contain simple descriptors. For example,
the “ATTRIBUTE_Strain” column in Fig. 1 contains two descrip-
tors called “GNPSGROUPS”: WT and rhlR, which denote the two
strains of P. aeruginosa being compared.

3.2 Perform FBMN 1. Start MZmine2 (see Note 4). The MZmine2 settings used for
Compatible Feature the sample data are outlined in Table 1.
Finding 2. Import raw data files: Select the files to be analyzed under the
menu option “Raw data methods ! Raw data import” (see
Note 5).
3. In MZmine2, a sequence of steps must be performed to cor-
rectly process the raw data for both feature finding and export
to GNPS for molecular networking (see Note 6). The
MZmine2 settings used for the sample data are outline in
Table 1 (see Note 7).
(a) Mass detection:
l Select all raw data files. Perform mass detection on MS
level 1 under the menu option “Raw data methods !
Mass detection ! Set filter: MS level 1” (see Note 8).
l Select all raw data files. Perform mass detection on MS
level 2 under the menu option “Raw data methods !
Mass detection ! Set filter: MS level 2” (see Note 9).
(b) Build chromatograms: Select all raw data files. Perform
building the chromatograms by choosing and applying
 Feature Based Molecular Networking 231

A B C D
1 filename ATTRIBUTE_Strain
2 9177.mzXML rhIR
3 9178.mzXML WT
4 9208.mzXML rhIR
5 9209.mzXML WT
6 9238.mzXML rhIR
7 9239.mzXML WT
8

Fig. 1 Correct format for the metadata table file. The metadata file describes
sample properties which will allow greater flexibility for analysis and
visualization of the molecular network

appropriate settings (see Note 10) under the menu option

“Raw data methods ! Chromatogram builder.” New files
will be created within the “Peak lists” window with the
suffix chromatograms.
(c) Deconvolute chromatograms: Select all peak lists with the
suffix chromatograms. Perform chromatogram deconvolu-
tion by choosing and applying appropriate settings (see
Note 11) under the menu option “Peak list methods !
Peak detection ! Chromatogram deconvolution.” New
files will be created within the “Peak lists” window the
suffix chromatograms deconvoluted.
(d) Isotopic peaks grouper: Select all peak lists with the suffix
chromatograms deconvoluted. Perform isotopic peaks
grouper by choosing and applying appropriate settings
(see Note 12) under the menu options “Peak list methods
! Isotopes ! Isotopic peaks grouper.” New files will be
created within the “Peak lists” window the suffix chroma-
tograms deconvoluted deisotoped.
(e) Align features: Select all peak lists with the suffix chroma-
tograms deconvoluted deisotoped. Perform feature align-
ment by choosing and applying appropriate settings
under the menu option “Peak list methods ! Alignment
! Join Aligner.” A new file called Aligned peak list will be
created within the “Peak lists” window.
(f) Detect missing peaks: Select the peak list Aligned peak list.
Perform gap-filling by choosing and applying appropriate
settings (see Note 13) under the menu option “Peak list
methods ! Gap-filling ! Peak finder (multithreaded).”
A new file called Aligned peak list gap-filled will be created
within the “Peak lists” window.
(g) Filter for MS/MS peaks: Select the peak list Aligned peak
list gap-filled. Perform peak list filtering for MS/MS by
selecting Keep all rows that match all criteria under the
232 Vanessa V. Phelan

Table 1
MZmine2 parameter settings used for sample dataset

Parameter Setting
Mass detection (MS1)
Scans MS level: 1; any polarity; any spectrum type
Mass detector Centroid; noise level 1.0E3
Mass list name Masses
Mass detection (MS2)
Scans MS level: 2; any polarity; any spectrum type
Mass detector Centroid; noise level 1.0E2
Mass list name Masses
Chromatogram builder
Scans MS level: 1; any polarity; any spectrum type
Mass list Masses
Min time span (min) 0.05
Min height 3.0E3
m/z tolerance 0.01 m/z or 20.0 ppm
Suffix Chromatograms
Chromatogram deconvolution
Suffix Deconvoluted
Algorithm Baseline cut-off; min peak height 1.0E3; peak duration range (min)
0.00–2.00; baseline level 5.0E3
m/z center calculation MEDIAN
Isotopic peaks grouper
Name suffix Deisotoped
m/z tolerance 0.01 m/z or 20.0 ppm
Retention time tolerance 0.25 absolute (min)
Monotonic shape Checked
Maximum charge 4
Representative isotope Most intense
Join aligner
Peak list name Aligned peak list
m/z tolerance 0.01 m/z or 20.0 ppm
Weight for m/z 0.8
Retention time tolerance 0.5 absolute (min)

(continued)
 Feature Based Molecular Networking 233

Table 1
(continued)

Parameter Setting
Weight for RT 0.2
Require same charge state Checked
Peak finder (multithreaded)
Name suffix Gap-filled
Intensity tolerance 10.0%
m/z tolerance 0.01 m/z or 20.0 ppm
Retention time tolerance 0.5 absolute (min)
Peak list rows filter
Name suffix Filtered
Minimum peaks in a row 3
Peak duration range Checked 0.00–2.0
Keep or remove rows Keeps rows that match all criteria
Keep only peaks with MS2 Checked
scan (GNPS)
Reset the peak number ID Checked
Duplicate peak filter
Name suffix Filtered
Filter mode OLD AVERAGE
m/z tolerance 0.005 m/z or 10.0 ppm
RT tolerance 0.5 absolute (min)
Export for/submit to GNPS
Filename FBMN example
Mass list Masses
Filter rows ONLY WITH MS2

“Keep or remove rows” option and checking the box for

“Keep only peaks with MS2 scans (GNPS)” (see Note 14)
under “Peak list methods ! Filtering ! Peaklist row
filter.” A new file called “Aligned peak list gap-filled fil-
tered” will be created within the “Peak lists” window.
(h) Filter for duplicate peaks: Select the peak list called
Aligned peak list gap-filled filtered. Perform duplicate
peak rows filtering by choosing and applying appropriate
settings (see Note 15) under the menu option “Peak list
234 Vanessa V. Phelan

methods ! Filtering ! Duplicate peak filters.” A new file

called Aligned peak list gap-filled filtered filtered will be
created within the “Peak lists” window.
l Export files for GNPS: Select the peak list called
Aligned peak list gap-filled filtered filtered. Perform
export of files under the menu option “Peak list meth-
ods ! Export for/Submit to GNPS.” Under the “file-
name” option in the export window, provide a name
for the files that will be created in the chosen folder.
Under “mass list” choose masses and select ONLY
WITH MS2 under “Filter rows.” After pressing OK,
navigate to the export folder. Within this folder, two
files will be created: an mgf file which will be used for
molecular networking and the associated Peak Area
Quanitification Table (a csv file labeled with “quant”
which contains the MS1 peak area abundance informa-
tion for each sample).

3.3 Perform FBMN 1. Navigate to the GNPS website (https://gnps.ucsd.edu) and

sign in using your user name and password.
2. Navigate to the homepage (https://gnps.ucsd.edu/Pro
teoSAFe/static/gnps-splash.jsp) and scroll down to
“Advanced Analysis Tools” option. Under “Feature Network-
ing,” choose “Analyze.” The analysis web page specific to the
FBMN workflow will open.
3. Set up and run the GNPS FBMN workflow. The GNPS FBMN
settings used for the sample data are outlined in Table 2 (see
Note 16).
(a) Workflow Selection: Provide a descriptive title (see Note
17).
(b) File Selection: Upload files to GNPS by clicking “Select
Input Files.” The “Select Input Files” interface will open.
Navigate to the “Upload Files” pane. Upload the mgf file,
the Peak Area Quantification Table csv file, and the Sam-
ple Metadata Table created in the previous steps by drag-
ging the files to the “File Drag and Drop” area (see Note
18). After upload of the files to the appropriate folder,
navigate to the “Select Input Files” pane. Locate the files
you want to analyze. Select the mgf file and click the “MS2
mgf” button. The mgf file will now be listed under
“Selected MS2 MGF file” in the “Selected Files” pane.
Follow the same procedure for the Peak Area Quantifica-
tion Table and the Sample Metadata Table. Speclibs will be
automatically listed under the “Selected Library Files”
section of the “Selected Files” pane (see Note 19). Close
the window by clicking “Finish Selection.”
 Feature Based Molecular Networking 235

Table 2
GNPS FBMN workflow parameter settings used for sample dataset

Parameter Setting
Workflow selection
Title FBMN example PA14 rhlR cosine 0.70
File selection
MS2 MGF file FBMN example.mgf
Peak area quantification table FBMN example_quant.csv
Sample metadata table FBMN example metadata.txt
Basic options
Quantification table source MZmine2
Precursor ion mass tolerance 0.05 Da
Fragment ion mass tolerance 0.1 Da
Advanced network options
Min pairs Cos 0.70
Network TopK 10
Minimum matched fragment ions 4
Maximum connected component size 100
Maximum shift between precursors 500 Da
Advanced library search options
Library search min matched peaks 4
Search analogs Don’t search
Top results to report per query 1
Score threshold 0.7
Maximum analog search mass difference 100.0 Da (default value)
Advanced filtering options
Minimum peak intensity 0.0
Filter precursor window Filter
Filter peaks in 50 Da window Filter
Filter library Filter
Advanced quantification options
Normalization per file No norm
Aggregation method for peak abundances per group Mean
Advanced external tools

(continued)
236 Vanessa V. Phelan

Table 2
(continued)

Parameter Setting
Run dereplicator Don’t run
Advanced extras
Supplementary pairs

(c) Basic Options: Choose MZmine2 as the “Quantification

Table Source.” Change “Precursor Ion Mass Tolerance”
and “Fragment Ion Mass Tolerance” settings to a value
that is appropriate for the data (see Note 20).
(d) Workflow Submission: Verify that your e-mail address is
correct. Click “Submit.”
4. When the FBMN analysis job is completed, navigate to the
menu item “Jobs” on GNPS and click the DONE status.
Under “Advanced Views—Experimental Views,” Click “Direct
Cytoscape Preview/Download.” After the GNPS Cytoscape
Downloader finishes loading, click “Download Cytoscape
File” (see Note 21).

3.4 Visualize 1. Locate and open the Cytoscape file downloaded from GNPS.
the FBMN in Cytoscape In the resulting molecular network, each node represents the
MS/MS spectrum from an MS1 feature (see Note 22). All
metadata associated with the network is shown in the
“Table Panel,” including the “GNPSGROUPS” descriptors
of the “ATTRIBUTE_Strain” column of the metadata table.
Additionally, the “Table Panel” lists information about
matches to the GNPS libraries.
2. Different parameters can be set to enhance the visualization of
data features within the molecular network by modifying the
settings in the “Node,” “Edge,” or “Network” tabs within the
“Style” tab of the “Control Panel” window. The default set-
tings for the automatically generated Cytoscape file are node
color: blue; edge color: white; background color: gray; node label:
Compound name (see Note 23). To recapitulate the molecular
network in Fig. 2a, apply the following steps:
(a) To highlight the connectivity similarity between nodes,
load cosine settings. To do this, select the “Edge” tab
within the “Style” tab of the “Control Panel.” For the
“Width” setting, select cosine for the “Column” and Con-
tinuous Mapping for the “Mapping Type.” Double-click
the “Current Mapping” box and set the desired minimum
(0) and maximum (30) values. Within the molecular net-
work, the thicker edges indicate a stronger relationship
between the MS2 spectra represented by the nodes.
 Feature Based Molecular Networking 237

Fig. 2 Feature-based molecular network. (a) The total feature-based molecular

network. Each feature is represented by a node. The pie chart indicates the
abundance of each feature in the different GNPSGROUPS, rhlR (pink) and WT
(blue). Nodes that have MS/MS matches to the GNPS libraries are outlined in
black. (b) The rhamnolipid molecular family. Three features had MS/MS matches
to the GNPS libraries: m/z 527.3189 (Rha-C10-C10), m/z 673.3768 (Rha-Rha-
C10-C10), and m/z 699.3927 (Rha-Rha-C10-db:C12). As molecules with similar
MS/MS spectra will cluster into a molecular family, annotation of other nodes
can be propagated through the molecular family based upon their difference in
m/z values to library matches
238 Vanessa V. Phelan

(b) To visualize the precursor mass on each node, select the

“Node” tab. For the “Label” setting, select precursor mass
for the “Column” and Passthrough Mapping for the
“Mapping Type.”
(c) To visualize which nodes have corresponding MS2 spec-
tral matches to the GNPS libraries, select the “Node” tab.
For the “Border Paint” setting, select Compound_Name
for the “Column” and Discrete Mapping for the
“Mapping Type.” Select the boxes with compound names,
right click, select edit, then edit Select Discrete Mapping
Values and choose black.
(d) To visualize the group average abundances, pie charts are
used. Select the “Node” tab. For the “Image/Chart 1”
setting, click on the Def. box. In the “Data” tab, remove
the default columns from the “Selected Columns” panel
and add the appropriate columns (GNPSGROUP:rhlR
and GNPSGROUP:WT for the sample data) from the
“Available Columns” panel. These two columns will now
be listed within the “Selected Column” panel. Under the
“Options” tab, select the desired color for the two col-
umns—pink for 1 (GNPSGROUP:rhlR) and blue for
2 (GNPSGROUP:WT). Click “Apply” (see Note 24).

3.5 Use the FBMN 1. The visualization of experimental data can be used to guide
to Perform Initial secondary metabolite annotation and analysis.
Analysis (a) Differential abundance (see Note 25): The pie chart visu-
alization can be used to quickly identify features that have
differential abundance between sample groups (Fig. 2).
The “cluster index” column of the “Node Table” panel of
the “Table Panel” in Cytoscape corresponds to the feature
number in the Peak Area Quanitification Table generated
by the MZmine2 workflow. Therefore, the abundance
differences displayed by the pie chart in the molecular
network can be validated simply by finding the feature
number in the Peak Area Quanitification Table for each
sample and averaging the abundance values for each group
(see Note 26).
(b) Annotation: In the molecular network, all nodes with
matches to the GNPS libraries are indicated by a black
border. Click on a node with a black boarder. The anno-
tation of that node will be listed within the “Compound_-
Name” column of the “Node Table” panel of the
“Table Panel.”
(c) Propagating an annotation: Locate the node annotated
“Rhamnolipid Rha-C10-C10” (m/z 527.3197). As
molecular networking compares the similarity of
 Feature Based Molecular Networking 239

MS/MS spectra, the nodes connected to “Rhamnolipid

Rha-C10-C10” form a molecular family of structurally
related molecules based on the MS/MS spectra. There-
fore, the MS1 mass difference between nodes and the raw
MS/MS spectra (see Note 26) can be leveraged to anno-
tate (confidence level 3) the rest of the nodes within the
molecular family (Fig. 2b).

4 Notes

1. The step-by-step instructions for the analysis described was

performed using MZmine2 on the Windows platform. There-
fore, adjustments may be required if a different computing
platform or feature finding program is used. For information
regarding how to convert proprietary raw MS data to open
source formats, consult the instrument vendor.
2. Tab separated text files can be generated using text editors
including Microsoft Excel, Notepad++ (Windows), gedit
(Linux), TextWrangler (Mac OS).
3. Metadata in metabolomics studies is important for defining the
experimental methodology as well as facilitating reproducibil-
ity. In the context of FBMN, metadata can be used to differen-
tially label or color components of the molecular network.
Therefore, it is important to consider what scientific questions
you would like to investigate using FBMN. Relevant metadata
for FBMN includes, but is not limited to, sample type,
biological source, genotype, treatment, geolocation, sample
preparation, bioactivity measurements, and so on.
4. Information about the compatibility of FBMN with other
feature finding programs can be found on the GNPS documen-
tation website (https://ccms-ucsd.github.io/
GNPSDocumentation/featurebasedmolecularnetworking/).
5. Raw data import file compatibility requirements for MZmine2
can be found in the HELP menus. The sample data was col-
lected on a Bruker Maxis HD, lock mass was applied, and the
data was converted from .d to centroided .mzXML using Bru-
ker CompassXport.
6. MZmine2 has extensive documentation under HELP menus
describing the options available for data processing. There are
many more options available in MZmine2 than those illu-
strated in the step by step instructions. While each setting can
be set individually, if you anticipate analyzing many similar
datasets, a batch .xml file can be saved, loaded, and configured
for multiple uses.
240 Vanessa V. Phelan

7. The MZmine2 settings for FBMN need to be optimized for

each dataset. The settings appropriate for the example data may
not be appropriate for your data.
8. Set an appropriate intensity threshold for MS1 feature finding
by clicking the “. . .” box next to the mass detector option. You
can use the “Show Preview” window to assess the effect of your
intensity threshold on your data. Minimally, the value should
correspond the minimum value set for triggering an MS2 scan
event used when setting up your MS2 data collection.
9. Set an appropriate intensity threshold for MS2 feature finding
by clicking the “. . .” box next to the mass detector option. You
can use the “Show Preview” window to assess the effect of your
intensity threshold on your data. This threshold should be set
to the representative noise level in the MS2 data. If this thresh-
old is set too high, export for GNPS processing will be nega-
tively impacted.
10. Choose settings appropriate for the LC separation and mass
accuracy of your experiment.
11. While not used in this example, m/z range for MS2 scan pairing
(Da) and RT range for MS2 scan pairing (min) can be enabled
to more precisely match an MS2 scan with an MS1 feature.
12. The settings chosen for Isotope peaks grouper are dependent
upon the chromatographic peak shape, duty cycle time of the
mass spectrometer, and the MS1 mass accuracy.
13. Gap-filling is a critical step in MZmine2 feature finding. Fol-
lowing alignment, the resulting peak list may contain missing
peaks due to deficient peak detection or a mistake in alignment
of different peak lists. A missing peak does not imply that the
peak does not exist. In most cases, a missing peak indicates that
it went undetected in previous steps. To verify that the feature
finding is correct, the peak list table can be visualized by right-
clicking on the “aligned peak list gap-filled” and choosing
“Show peak list table.” It is important to always verify the
peak list by validating randomly selected features in the raw
data. Peak finding (multithreaded) allows the processing to be
performed in parallel instead of sequentially.
14. There are many different filtering options available in
MZmine2. For further details, access the HELP information.
15. Removal of duplicate peaks ensures that duplicate features
generated during the previous processing steps are removed.
This filtering step can be performed before applying the Peaks
rows filter, if desired.
16. The settings for the FBMN GNPS workflow must be opti-
mized for each individual dataset. For more information
about the various settings, consult the GNPS documentation
web page.
 Feature Based Molecular Networking 241

17. Using a descriptive title that encompasses the setting used for
FBMN is extremely useful. One of the advantages of the GNPS
platform is that a user can perform iterative analyses of the same
data set concurrently in order to identify the most optimal
settings for the analysis of their dataset.
18. If the files are too large for the Drag and Drop feature of the
Select Input Files interface, the files can be uploaded to GNPS
via an ftp server. Directions to perform file upload via ftp is
available within the GNPS documentation.
19. GNPS annotates MS/MS spectra by comparing each spectrum
to MS/MS libraries. The list of available MS/MS libraries can
be accessed via the GNPS website.
20. The appropriate “Precursor Ion Mass Tolerance” and “Frag-
ment Ion Mass Tolerance” settings are dictated by the mass
accuracy and mass resolution of the mass spectrometer used for
MS/MS data collection. In addition to Basic Options, the
FBMN workflow can be adjusted by modifying the various
Advanced Options. Learn more about these settings on the
GNPS documentation pages.
21. In addition to visualizing the molecular network in Cytoscape,
the network can be viewed and analyzed via the in-browser
visualization. For further information about the different types
of analyses that can be performed in GNPS, access the GNPS
documentation.
22. It is important to note that during the feature finding proces-
sing in MZmine2, only the features that have a corresponding
MS/MS spectrum are retained for molecular networking anal-
ysis. If a feature does not have an associated MS/MS spectrum,
it will not be in the molecular network. Whether a feature has a
corresponding MS/MS spectrum is dependent upon the set-
tings used during data acquisition as well as during data pro-
cessing in MZmine2.
23. Tutorials for how to use Cytoscape, including a manual and
videos, are available on the Cytoscape website.
24. If the visualization adjustments of the molecular network in
Cytoscape do not automatically change, they can be shown by
utilizing the “Show Graphic Details” option within the
“View” menu.
25. Cytoscape has many different settings that can be used to
differentiate groups. Node size, shape, color, etc. can be
adjusted to reflected different attributes of the samples such
as origin, and bioactivity.
26. FBMN is a tool to aid in the analysis of MS data. The settings
used in the FBMN workflow, including those for feature
finding and generating the molecular network, will influence
the final molecular network output. It is very important to
validate the observations from FBMN with the raw data.
242 Vanessa V. Phelan

Acknowledgments

We thank L. Dietrich (Columbia University) for kindly providing

the P. aeruginosa strains and M. Wang (University of California,
San Diego) for providing feedback on the manuscript. Our work on
microbial metabolomics has been supported by the National Insti-
tutes of Health (National Institute of General Medical Sciences
grants K01 GM103809 and R35 GM128690), the ALSAM Foun-
dation (L.S. Skaggs Professorship and Therapeutical Innovation
Award), and the University of Colorado.

References

1. Sumner LW, Amberg A, Barrett D, Beale MH, Gonzalez DJ, Silva DB, Marques LM, Demar-
Beger R, Daykin CA, Fan TW, Fiehn O, que DP, Pociute E, O’Neill EC, Briand E, Hel-
Goodacre R, Griffin JL, Hankemeier T, frich EJN, Granatosky EA, Glukhov E,
Hardy N, Harnly J, Higashi R, Kopka J, Lane Ryffel F, Houson H, Mohimani H, Kharbush
AN, Lindon JC, Marriott P, Nicholls AW, Reily JJ, Zeng Y, Vorholt JA, Kurita KL,
MD, Thaden JJ, Viant MR (2007) Proposed Charusanti P, McPhail KL, Nielsen KF,
minimum reporting standards for chemical Vuong L, Elfeki M, Traxler MF, Engene N,
analysis Chemical Analysis Working Group Koyama N, Vining OB, Baric R, Silva RR,
(CAWG) Metabolomics Standards Initiative Mascuch SJ, Tomasi S, Jenkins S, Macherla V,
(MSI). Metabolomics 3(3):211–221. https:// Hoffman T, Agarwal V, Williams PG, Dai J,
doi.org/10.1007/s11306-007-0082-2 Neupane R, Gurr J, Rodriguez AMC,
2. Garg N, Luzzatto-Knaan T, Melnik AV, Lamsa A, Zhang C, Dorrestein K, Duggan
Caraballo-Rodriguez AM, Floros DJ, BM, Almaliti J, Allard PM, Phapale P, Nothias
Petras D, Gregor R, Dorrestein PC, Phelan LF, Alexandrov T, Litaudon M, Wolfender JL,
VV (2017) Natural products as mediators of Kyle JE, Metz TO, Peryea T, Nguyen DT,
disease. Nat Prod Rep 34(2):194–219. VanLeer D, Shinn P, Jadhav A, Muller R,
https://doi.org/10.1039/c6np00063k Waters KM, Shi W, Liu X, Zhang L,
3. Newman DJ, Cragg GM (2016) Natural pro- Knight R, Jensen PR, Palsson BO,
ducts as sources of new drugs from 1981 to Pogliano K, Linington RG, Gutierrez M,
2014. J Nat Prod 79(3):629–661. https://doi. Lopes NP, Gerwick WH, Moore BS, Dorres-
org/10.1021/acs.jnatprod.5b01055 tein PC, Bandeira N (2016) Sharing and com-
munity curation of mass spectrometry data
4. Phelan VV, Liu WT, Pogliano K, Dorrestein with global natural products social molecular
PC (2011) Microbial metabolic exchange— networking. Nat Biotechnol 34(8):828–837.
the chemotype-to-phenotype link. Nat Chem https://doi.org/10.1038/nbt.3597
Biol 8(1):26–35. https://doi.org/10.1038/
nchembio.739 6. Nguyen DD, Wu CH, Moree WJ, Lamsa A,
Medema MH, Zhao X, Gavilan RG,
5. Wang M, Carver JJ, Phelan VV, Sanchez LM, Aparicio M, Atencio L, Jackson C,
Garg N, Peng Y, Nguyen DD, Watrous J, Ballesteros J, Sanchez J, Watrous JD, Phelan
Kapono CA, Luzzatto-Knaan T, Porto C, VV, van de Wiel C, Kersten RD, Mehnaz S, De
Bouslimani A, Melnik AV, Meehan MJ, Liu Mot R, Shank EA, Charusanti P, Nagarajan H,
WT, Crusemann M, Boudreau PD, Duggan BM, Moore BS, Bandeira N, Palsson
Esquenazi E, Sandoval-Calderon M, Kersten BO, Pogliano K, Gutierrez M, Dorrestein PC
RD, Pace LA, Quinn RA, Duncan KR, Hsu (2013) MS/MS networking guided analysis of
CC, Floros DJ, Gavilan RG, Kleigrewe K, molecule and gene cluster families. Proc Natl
Northen T, Dutton RJ, Parrot D, Carlson Acad Sci U S A 110(28):E2611–E2620.
EE, Aigle B, Michelsen CF, Jelsbak L, https://doi.org/10.1073/pnas.1303471110
Sohlenkamp C, Pevzner P, Edlund A,
McLean J, Piel J, Murphy BT, Gerwick L, 7. Shannon P, Markiel A, Ozier O, Baliga NS,
Liaw CC, Yang YL, Humpf HU, Wang JT, Ramage D, Amin N,
Maansson M, Keyzers RA, Sims AC, Johnson Schwikowski B, Ideker T (2003) Cytoscape: a
AR, Sidebottom AM, Sedio BE, Klitgaard A, software environment for integrated models of
Larson CB, CAB P, Torres-Mendoza D, biomolecular interaction networks. Genome
 Feature Based Molecular Networking 243

Res 13(11):2498–2504. https://doi.org/10. sensing in Pseudomonas aeruginosa. Proc Natl

1101/gr.1239303 Acad Sci U S A 96(24):13904–13909
8. Frank AM, Bandeira N, Shen Z, Tanner S, 12. Winson MK, Camara M, Latifi A, Foglino M,
Briggs SP, Smith RD, Pevzner PA (2008) Clus- Chhabra SR, Daykin M, Bally M, Chapon V,
tering millions of tandem mass spectra. J Pro- Salmond GPC, Bycroft BW, Lazdunski A,
teome Res 7(1):113–122. https://doi.org/10. Stewart GSAB, Williams P (1995) Multiple
1021/pr070361e N-acyl-L-homoserine lactone signal molecules
9. Olivon F, Grelier G, Roussi F, Litaudon M, regulate production of virulence determinants
Touboul D (2017) MZmine 2 data- and secondary metabolites in Pseudomonas-
preprocessing to enhance molecular network- aeruginosa. Proc Natl Acad Sci USA 92
ing reliability. Anal Chem 89(15):7836–7840. (20):9427–9431. https://doi.org/10.1073/
https://doi.org/10.1021/acs.analchem. pnas.92.20.9427
7b01563 13. Pluskal T, Castillo S, Villar-Briones A, Oresic
10. Schuster M, Greenberg EP (2007) Early acti- M (2010) MZmine 2: modular framework for
vation of quorum sensing in Pseudomonas aer- processing, visualizing, and analyzing mass
uginosa reveals the architecture of a complex spectrometry-based molecular profile data.
regulon. BMC Genomics 8:287. https://doi. BMC Bioinformatics 11:395. https://doi.
org/10.1186/1471-2164-8-287 org/10.1186/1471-2105-11-395
11. Whiteley M, Lee KM, Greenberg EP (1999)
Identification of genes controlled by quorum
 Chapter 14

A Bioinformatics Primer to Data Science, with Examples

for Metabolomics
W. Stephen Pittard, Cecilia “Keeko” Villaveces, and Shuzhao Li

Abstract
With the increasing importance of big data in biomedicine, skills in data science are a foundation for the
individual career development and for the progress of science. This chapter is a practical guide to working
with high-throughput biomedical data. It covers how to understand and set up the computing environ-
ment, to start a research project with proper and effective data management, and to perform common
bioinformatics tasks such as data wrangling, quality control, statistical analysis, and visualization, with
examples on metabolomics data. Concepts and tools related to coding and scripting are discussed. Version
control, knitr and Jupyter notebooks are important to project management, collaboration, and research
reproducibility. Overall, this chapter describes a core set of skills to work in bioinformatics, and can serve as
a reference text at the level of a graduate course and interfacing with data science.

Key words Bioinformatics, Metabolomics, Data science, Quality control, Data management, Cloud
computing, Scripting, Data visualization

1 Introduction

The field of bioinformatics received the first major boost by

genome sequencing [1]. Over the past three decades, bioinformat-
ics became infused with many disciplines. At times, it is difficult to
recognize bioinformatics as a single research field, because of its
coverage of broad topics, from molecular evolution, structure of
macromolecules, drug screening, and machine learning to network
models. When commemorating the 50th anniversary of the DNA
helix, Walter Gilbert said that “[it] seems to me that molecular
biology is dead. DNA-based thinking has penetrated the whole of
biology, and the separate field no longer exists” [2]. The same thing
can be said of bioinformatics today—biology has become an infor-
mation science, where larger and larger volumes of data lead the
new waves of innovation. Bioinformatics has penetrated biology.
The most exciting bioinformatics research has been combined with

Shuzhao Li (ed.), Computational Methods and Data Analysis for Metabolomics, Methods in Molecular Biology, vol. 2104,
https://doi.org/10.1007/978-1-0716-0239-3_14, © Springer Science+Business Media, LLC, part of Springer Nature 2020

245
246 W. Stephen Pittard et al.

particular scientific domains, which evolved along genomics, epi-

genomics, proteomics, microbiome, and now metabolomics.
The goal of this chapter is to describe a core set of skills to do
bioinformatics, at the level of a graduate course and interfacing
with data science. This is written as an introductory reference of
working with high throughput biomedical data, covering data
management, scripting, version control, data wrangling, quality
control, statistical analysis and visualization. We will use metabolo-
mics data as examples, but the principles are rather generic. An
in-depth guide of the essential tools, including Python and R
programming environments, Git for version control, and Docker
for virtualization, is provided in Chapter 15.

2 Computing Environment and Setup

Selecting the most appropriate computational environment is an

important decision for researchers involved in even modest forms
of data analysis. The choice will be a function of several things
including the type and size of data, whether collaboration is
involved, as well as the desired outcome (e.g., a software product
or package in support of a publication). In response to the prolifer-
ation of omics data, there has been a corresponding development of
proprietary and open source tools and frameworks designed to ease
the management and comprehension of this information. This is
advantageous since researchers and analysts do not have to develop
major software tools to accomplish in-depth data transformation
and analysis. However, they do have to invest time in understand-
ing how to deploy such tools to leverage their full potential. In
particular, combining disparate tools into pipelines, while selecting
appropriate options for each of the component programs, can
require significant experimentation when seeking to optimize for
the data to generate or test project hypotheses.
It is also possible that a project might involve the use of a
number of tools, languages, and operating systems which in turn
requires the familiarity with different platform’s exchange formats,
and user interfaces. This can be mitigated somewhat by the avail-
ability of web-based portals that insulate the user from the under-
lying details of data management and analysis while providing the
benefit of near immediate access to results that can be examined
using, for example, a genome browser or alignment tools common
to bioinformatics workflows. Tools such as BLAST, Galaxy, and the
UCSC Genome browser fall into this category. These frameworks
are also available for installation on local computers upon which the
user can install as much or as little of the toolset as desired while
retaining the option of using the public versions of these tools.
In this case the researcher, or someone acting on their behalf,
must manage the technical aspects of installation, deployment, and
 Bioinformatics Primer for Data Science 247

training for the tool. For purposes of classroom education, use of

Web services and GUIs might be preferable as it allows the students
to focus on the science behind a given tool without the added
distraction of having to understand the underlying computational
environment. However, for researchers and analysts, being able to
“look under the hood” is important since it permits a nuanced use
of the tool that is necessary or more productive. Moreover, being
able to author “scripts” and programs to obtain, manipulate, and
visualize data is useful to accelerate the pace of one’s research.
When considering the above, it becomes apparent that some
degree of self-sufficiency and familiarity with informatics tools is
useful if not necessary to accelerate the pace of research. Publication
agreements can also require the data and code made available. The
creation of digital assets that can be easily tracked greatly enhances
the reproducibility of published research, and it usually necessitates
the involvement of informatics aware personnel. Even before the
publication step, it is helpful for a researcher to have facility with
data management approaches and visualization packages essential
to data exploration and the formation of hypotheses. In this case,
being productive with the “command line” interface becomes
important as it provides greater flexibility than is customarily avail-
able in form-based web tools. It also enables the creation of “glue”
code to wrap, supplement, or replace existing software tools.

2.1 Desktops, Analysis can be accomplished in a number of environments depend-

Laptops, Clusters, ing on the scope and complexity of the project. Modern computing
and the Cloud hardware, even at the laptop level, provides sufficient memory,
multicore CPUs, and disk space to accomplish many data transfor-
mation and analysis activities, including the use of parallel libraries
(depending on the language). Desktop computers can provide even
greater flexibility and expandable components, which can also
accommodate virtual servers designed to host one or more
operating systems should there be a need for such. Cheap hardware,
in combination with comparably priced networking equipment,
motivated the formation of Beowulf clusters in the late 1990s and
early 2000s which allowed researchers to leverage the collective
power of various computers by presenting them as an integrated
cluster. Users submit “jobs” into a work “queue” and wait for
results according to an established priority. While the Beowulf
model is still in use, it is more common to experience formalized
clusters managed by knowledgeable personnel responsible for the
continued operation of the resource and, optionally, assist users
with questions. The advantage for the researcher is that the infra-
structure is managed on their behalf, there is a single point of entry,
and the resource can generally be expanded without impacting
existing work.
A significant evolution of this approach was introduced in 2006
when Amazon launched Amazon Web Services, which provided
scalable Infrastructure as a Service (IaaS) allowing knowledgeable
248 W. Stephen Pittard et al.

users the ability to provision and organize compute “instances,”

storage, and networking on-demand. This approach provides flexi-
bility in the architecture of environments customized to the needs
of a specific project. The service is available as a “pay what you use”
model, while significantly discounted reserved resources are avail-
able. Google and Microsoft followed suit and offer competing
resources at a similar level and cost. All of these vendors provide
additional services such as machine learning and deep learning,
database management, data streaming which can be arranged and
connected in novel ways which implies that fully leveraging these
services requires dedicated skills and training. These vendors, Ama-
zon, Microsoft, and Google are colloquially known as “Cloud
Computing” providers in reference to the idea that all hardware,
networking, and storage are managed off-premise as well in refer-
ence to the idea that the at-scale capabilities on offer are seemingly
infinite.

2.2 Operating A key aspect of a computational environment is the availability of

Systems software resources and libraries. Complex dependencies between
for Bioinformatics these resources and libraries may determine the cost and feasibility
of a project, even if new software development is not involved. Part
of the problem is that they can be dependent on a particular
operating system. Furthermore, how one interacts with the
operating system directly impact one’s productivity. Three major
operating systems are widely used today: Microsoft Windows,
Apple Mac OS, and Linux, with many people being more familiar
with the first two. Microsoft Windows is the most prevalent
operating system for computing at the consumer and corporate
levels wherein office productivity is a priority. Thus, the environ-
ment is developed primarily to support spreadsheet use, email, web
browsing, and business applications. Nonetheless, it is still possible
to use Microsoft Windows as a host operating system for computa-
tional research especially with the addition of Windows 10 “power
shell” which offers a “shell/interpreter” tool for those accustomed
to command line development within a UNIX environment.
Microsoft, like Apple Mac OS, is a proprietary system the cost of
which is typically embedded within the purchase price of a com-
puter although some organizations have established licensing
agreements with Microsoft to offer discounted or subsidized
licenses. Many open source tools are available for both Windows
and Mac OS.
Apple Mac OS is also a popular proprietary operating system
(including versions of Mac OS, OS X, macOS) which is tightly
integrated with Apple hardware to provide a unified, stable envi-
ronment in accordance with an established set of user interface
design considerations. The Mac OS operating system is based
upon a variant of UNIX called BSD (Berkeley Software Distribu-
tion). For the end user this means that a form of UNIX is accessible
 Bioinformatics Primer for Data Science 249

on any Apple hardware running OS. It is interesting to note that

the GUI (the “Finder”) is a process that runs on top of UNIX. For
those interested in command line access, it can be obtained via the
Terminal application. Apple Mac OS also offers a development
interface called Xcode which is primarily for those interested in
creating applications for Mac OS and iOS, while it is used to
manage projects in various languages. For those interested in treat-
ing their Apple Mac hardware primarily as a UNIX box, there is the
Homebrew project (https://brew.sh/), which enables the manage-
ment and installation of common UNIX packages such as the GNU
suite of tools (gcc, make, etc.). The advantage of this approach is
that one can continue to use the conveniences of the Apple Mac
GUI while having full access to a UNIX-based development and
shell scripting environment.
Virtualization technologies have been commonly used to serve
multiple operating systems on the same hardware, and to support
cloud computing. These include the two main categories of virtual
machines and containers. These technologies help modularize
computing and make the deployment of software applications
more compatible and efficient. The Docker containers are
described in details in Chapter 15.

2.3 UNIX/Linux The terms UNIX and Linux are often used interchangeably, but
technically that is not accurate. The term UNIX originally applied
to a proprietary operating system developed out of Bell Labs in the
early 1970s, which was originally intended for internal use. As the
system grew in utility and popularity, licenses were offered to
Berkeley University for further development, which resulted in
BSD, the first major branch of the ATT code base. Organizations
such as IBM and HP developed their own “flavors” of UNIX, some
based on ATT or BSD, and offered these distributions under vari-
ous business models. The GNU project was created by Richard
Stallman and became the leader of the free software movement. The
year of 1991 saw the release of the first Linux kernel by Linus
Torvalds. Subsequently, Linux became the dominant system to
run web servers, and led to the boom of an open source software
ecosystem. The name Linux is commonly employed generically
though there are a number of specific distributions (abbreviated
as “distros”) such as Ubuntu, Debian, GenToo, and Fedora as well
as commercially supported distros such as Redhat and SUSE.
While Linux is freely available, vendors can offer “for cost”
support models to help individuals and organizations become pro-
ductive. The organizations behind each of the respective distros can
offer enthusiastic support freely simply to encourage adoption of
those particular distributions. As an example, the Ubuntu distribu-
tion has a large number of “baked in” tools for the processing of
genomic data which is appealing to researchers within that domain.
250 W. Stephen Pittard et al.

Linux has become the standard for high performance computation

and is the foundation of major cloud computing vendors.

2.4 GUI vs. the The convenience and intuitive nature of GUI (graphical user inter-
Command Line face) is attractive as it simplifies the management of files, folders,
and computer settings by offering a visual interface to perform
these common actions without requiring in depth training or ori-
entation as a prerequisite to productivity. Moreover, there are a
number of desktop applications written specifically to leverage the
capabilities of these operating systems which can further simplify
the manipulation of data. For example, there are commercial
packages for the analysis of genomic data that run natively on
both Windows and Mac OS. However, in order for a GUI to be
available, the exact task has to be defined and coded into the GUI
design. This places a lot of limit on what one can do within a GUI.
A typical task in bioinformatics is converting data formats, which
relies mostly on the use of text manipulation package such as sed or
awk or the creation of a script in Python or R to first preprocess the
data. Without such knowledge the researcher will be at a disadvan-
tage and reliant upon others to accomplish what is a basic yet
essential task.
The phrase “command line” refers to the use of typed com-
mands at a shell (e.g., Bash, ksh) prompt that in turn invokes
specific programs available within a given computational environ-
ment. Microsoft Windows has a command prompt. Shell is native
to Linux and Mac OS, which runs on top of a variant of the UNIX
operating system. Bioinformatics often involves command lines to
execute and connect different programs. This creates the capability
to perform and automate complex tasks.

2.5 The Shell Learning how to interact with the UNIX operating system does
require a commitment on the user’s part, though a facility with it
can be acquired in stages over time and in response to the demands
of a given project. The basic components involve the use of a
SHELL which is the interpreter for any command a user might
enter at the command line prompt. Think of a shell as an interpreter
of interactive input although scripts containing an arbitrary number
of shell commands, including programming constructs, can be
developed for execution within the shell context. As an example,
it is possible to create a number of shell scripts written in bash for
the maintenance of data over time. These are typically designed and
implemented for system administration tasks, though Linux offers
powerful text manipulation tools which can be helpful in the man-
agement of scientific data. However, many researchers choose to
develop scripts in Python or R, with the benefit of packages written
specifically to work with high throughput data. Figure 1 is an
example of using the shell on Apple Mac OS. It is outside the
 Bioinformatics Primer for Data Science 251

Fig. 1 Screenshot of BASH example on Mac OSX

scope of this text to provide a detailed primer on the Bash shell,

though it represents an important development tool.
Shells are native to Linux and Mac OS. To communicate with
remote servers, including cloud infrastructure, one relies mostly on
Shells. Very recently, Microsoft has started including “Windows
subsystem for Linux,” to provide full Linux functions within the
Windows 10 operating system.

3 Data Management

3.1 Organizing Your Independent of the operating system you choose, one of the pri-
Projects mary uses of GUI is to manage files and folders on your local system
or attached to it in cases wherein you are using a shared filesystem
or a form of network attached storage. GUIs provide an easy way to
create, rename, move, and remove folders to reflect the needs of
your work. All of this can be accomplished using the command
line/shell, although that requires facility with the specific command
names to perform efficiently. One area in which the SHELL is quite
helpful is when needing to rename or manipulate a large number of
files according to a general pattern in which cases a Bash script can
be written. However, generally speaking it is fine to use a GUI to
manage your project organization. The key is to choose a setup and
naming convention that will make sense particularly after the pas-
sage of time such that when you return the structure will be under-
standable. Also consider that you might later involve a collaborator
in which case you will need to explain the structure if it is not
already evident. While organizing your project is a highly subjective
activity here are some tips that will help you:
1. Do not stockpile code, data, and spreadsheets into a single
folder. Create meaningfully sub folders up front to contain
highly related information even if you do not yet have a signifi-
cant amount of information to manage.
For example, all sequencing data should be in a folder and
perhaps sub folders depending on the experimental design of
your project. While this seems obvious, it is very easy to initiate
a project by creating one folder with some basic data, perhaps
252 W. Stephen Pittard et al.

Fig. 2 File directories as examples for data management. (a) A project folder on a remote Linux server,
mounted to MacOS desktop via sshfs. This is an automated directory structure on this server, so that scripts
are run periodically to gather project statistics and generate reports. (b) A project folder with a focus on DNA
sequencing. (c) A manuscript folder, where data are organized by figures

some invoices from a Core facility, PDFs of relevant publica-

tions, and example code. This might be manageable in the
short term, but as more data arrives, it begins to dominate
the folder which makes it difficult to locate the documents. It
is always possible to create new subfolders within the originally
created folder as more information accumulates though it
becomes difficult to cleanly move information into the newer
folders if naming conventions are not uniform. Figure 2 gives a
few examples of organizing data for a project.
2. Choose names that avoid spaces, especially if you have colla-
borators using Linux.
It is possible to create folders and filenames that contain
spaces within the name (e. g. “Sequencing Folder”). Note that
 Bioinformatics Primer for Data Science 253

when accessing these files from a program or command line

tool, they must be quoted so the shell (or code) will not
interpret the name as being two distinct files. While this is a
minor point, it is easier to use a fully specified name without
spaces up front rather than retro fitting such a change. While
doing this might not make a difference for your work it might
very well make a difference to your collaborator who is
attempting to programmatically manipulate your folder struc-
ture, particularly the sequencing files. Names should be specific
and informative, and adding the date is a good idea (e.g.,
“SmithJ_Sequencing_Data_May_20_2019”), because the
time stamp attached by the operating system may get lost
after the files are moved around a few places.
3. Place data on a volume that is backed up or replicated to an
off-premise site.
This may sound obvious, but projects frequently begin by
using a local hard drive that might be personal property or not
part of a research network and thus cannot access the drives or
sharing services available within an organization. Many univer-
sities and research organizations have shared drives that can be
mounted which is useful for backing up data although one
must ensure that such shares can be accessed off campus.
Services such as Dropbox or Box allow use of shared drives
anywhere, and in cases where network activity is interrupted,
the changes are synchronized once connectivity has returned.
There is wide variation in how organizations offer backup and
archival services so it is incumbent upon the researcher to make
regular copies of project information. Obviously, in cases where
the data is of considerable size (terabytes or petabytes) accom-
modations must be made at the outset to avoid moving the
data except where necessary given that network traversal can be
slow. At the individual level, it is possible to use Amazon Web
Services or Google Cloud to stockpile project data at a reason-
able cost, assuming your organization’s data management poli-
cies permit such an approach.
4. Use Git repositories to maintain source code. Use Github-like
services for distribution and collaboration.
Git is arguably the most popular version control tool. Git
and Github will be discussed in greater detail in a Chapter 15.
Git allows users to manage changes to source code over time in
a way that tracks all modifications with the added capability of
being able to arbitrarily revert to a previous version. While this
seems like more of a convenience for programmers it is useful
for anyone engaged in the development of scripts, involved
code, or pipelines that will eventually become part of a pub-
lished or shared work. Services such as Github allow users to
create repositories for hosting and sharing Git repositories
254 W. Stephen Pittard et al.

(abbreviated “repos”) at no cost. The repository is safely

backed up and archived long term. Keep in mind that Github
is only for source code as opposed to the data that the code will
apply to. While it is possible, and advisable, to include a set of
demo data within the repository, it should be small in size.

3.2 Notebook Tools Data reproducibility is fundamental to science because a scientific

conclusion is only valid if it can be validated by peers. With regard
to publications that include analysis and code to arrive at a set of
results, a general reader should be able to realize the same result
given reasonable access to the code and data referenced in the
publication.
Scripting notebooks are a useful tool for sharing the full process
of data analysis. With respect to R, there is the knitr (https://yihui.
name/knitr) package that provides dynamic report generation
which permits the mixture of text, equations, images, and R code
within a given document notebook that can be used to create
results based on user supplied data. The format of the notebook
can be Markdown, LaTeX, or Rnw all of which can be used to
develop large scale publications in PDF or HTML formats. Thus, it
is easy to create publication ready documents that can easily be
regenerated in response to newly supplied data. Another popular
tool is Jupyter Notebook (https://jupyter.org/), which also offers
a notebook style framework for report code sharing and report
generation. In addition to Python and R, Jupyter supports over
40 programming languages. By keeping code and results in a web
browser, it keeps research record, facilitates collaboration, and
makes good tutorials.

4 Coding and Scripting

4.1 Before You Start Software development usually requires specialized tools and work-
Coding flow. Most bioinformatics data analysis does not require software
development but scripting. Bioinformatics is often a lengthy pro-
cess of many steps, each step to accomplish a specific task. It is
critical to identify what tasks one need to complete in order to
accomplish the overall goal. One has to know what the question is
before choosing the right tool. Scripting is usually used to wrangle
data and “glue” different steps. Interactive terminals and notebook
tools (e.g., knitr and Jupyter notebooks) are often preferred over
complex IDEs (integrated development environments). But it is
still a good idea to version control the scripts. Should a different
version produce a different result, one need to track down the
source of the change.
Even without involving serious software development, a few
practical tips are good to have. Do not repeat yourself—if a piece of
code accomplishes a common task, make it reusable as a function,
 Bioinformatics Primer for Data Science 255

class or module. Keep it simple and stupid—readability will save

you and, more importantly, others, a lot of time later. Adopt con-
ventions whereas possible, especially common data formats. These
include csv, tsv, XML, and JSON. Since most programming lan-
guages have mature tools to handle common data formats, this
reduces development cost and make the work more interoperable.
Learning to use a code editor is mandatory, because one has to
know what is the command given to the computer and what is the
input data. Microsoft Word is the most popular WYSIWYG editor,
but it cannot be used for code editing. Because the user has to have
absolute transparency over the code, and WYSIWYG editors mask
the content by visual elements. A code editor works on plain text
and gives you total transparency. The default text editor on MS
Windows is Notepad, on Mac OS TextEdit. Among the popular
free code editors are vim, Atom and Visual Studio Code. Beginners
should avoid the complexity of IDEs (Integrated Development
Environment, e.g., Eclipse), in order to focus on the mastering of
the basic programming concepts. Rstudio, Matlab, and IPython are
not exactly programming IDEs, but instead are interactive envir-
onments for data analysis.

4.2 What Languages It is easy to incite hot debates when choosing a programing lan-
to Use guage. In the field of bioinformatics, Python has already become
the most popular language (replacing Perl). Java has wide applica-
tions in enterprise software development, and C/C++ is often
preferred for performance. Javascript is now gaining prominence
due to its central role in web development. With many mature
libraries and a clean, elegant design, Python enables rapid scripting
and development without compromising a lot of performance,
since many libraries are implemented in C. Also highly popular, R
is irreplaceable in the field of bioinformatics. R is designed as a
statistical programming environment, centered on data analysis.
Preferred by statisticians, many statistical tools were first published
in R. The successful Bioconductor project further boosted its pop-
ularity. Both Python and R are explained in detail in Chapter 15.
The most practical choice, however, is dependent on several
factors of a project. It takes time to gain proficiency in a language—
one has to weigh over the time investment when learning a new
tool. Domain specific applications are an important consideration.
To work with relational databases, knowledge of Structured Query
Language (SQL) is required to select and manipulate data from the
database. One may also consider embedding SQL capability within
software code and pipelines. Sometimes, the availability of one
function determines the design of a project. If a commercial license
is required, the cost and maintenance needs to be planned carefully.
Moving a piece of software from a laptop to cloud could require a
new license.
256 W. Stephen Pittard et al.

The real-world work environment often consists of mixed,

heterogeneous computing components, which demand multiple
programming languages. The maintenance cost of software pro-
grams could be grossly underestimated. The libraries the software is
built upon evolve over time, which can lead to version conflicts and
complex dependency issues. Docker container is a tool to produce a
snapshot of all related resources of a software project, thus addres-
sing this issue. If the components of software or different software
programs communicate via standard interfaces, especially via a
network protocol, the dependency issues will be isolated within
each component. This also means, programs do not have to be
written in the same language to communicate effectively. As a lot of
computing moves into the web and cloud, this new paradigm is well
accommodated in the latest trend in microservices.

5 Common Bioinformatics Tasks

There are bioinformatics tools for specific domains (e.g. BLAST

and BLAT for searching DNA sequences, and XCMS, mzMine, and
OpenMS for processing metabolomics data) (Chapters 2–4 in this
book). Different from those, this section covers a set of common
tasks that are often carried out by scripting using Shell, Python or
R. These include data handling, analysis, and visualization. One
needs the flexibility of scripting to accomplish the tasks, and some
general guiding principles are explained here.
To copy, move, rename and back up batch files, interactive
commands in the Shell are the right tool. Even simple commands
can be written in a script and automated. For example, one can use
one command to compress files and another to copy to a remote
computer; a script consisted of these two lines of code can be run
automatically at 12:15 am every night (using a systems service like
Cron on UNIX). Readers can explore more on data management
and Shell scripting in their own practice. In this section, we will
cover quality control, data wrangling, common statistics, and visu-
alization using metabolomics data as examples (Note 1).

5.1 Performing Before spending 2 years to analyze your data, it is critical to know
Quality Control that the data are not garbage due to some malfunction of an
(QC) and Quality instrument. For high-throughput data, abnormality is easy to spot
Assurance (QA) by examining two things: signal intensity acquired on a sample
(Figs. 3 and 4), and how the measurements in a sample correlate
with other samples (Fig. 5). If an instrument fails during the
analysis of a sample, the signal intensity will not be consistent for
this sample, resulting often low signal intensity. For samples of the
same type of biological matrix (e.g., human serum), the basic
pattern of intensity should be the same, therefore manifesting
high correlation coefficients between these samples.
 Bioinformatics Primer for Data Science 257

Fig. 3 Bar plot of average intensity of all replicates in the moDC dataset,
inspected for quality control

Quality control should already be incorporated into the experi-

mental protocols. Commonly, metabolomics experiments include
spike-in exogenous chemicals as controls, which are chosen as
stable isotope labeled molecules that are absent in biological sam-
ples. These control molecules should give consistent peaks in the
data of each sample, and visualized for quality control. Similarly,
once can plot the signals from common endogenous metabolites,
while their signals may not be consistent for biological reasons. An
example of phenylalanine (protonated molecular weight 166.0863,
consistently seen in human plasma using positive ESI on LC-MS) is
shown in Fig. 4. Sometimes, injection replicates are performed on
mass spectrometry metabolomics. These are technical replicates
and expected to generate the same result (Fig. 5a).
The above tasks are usually associated with the laboratory that
performs the experiments, and they should be completed before
the data are handed over to the next stage of projects. For people
responsible for the data analysis, these QC results should be exam-
ined before time and effort is spent on downstream analysis. In
addition, PCA and MA plots are good ways to check for anomalies
in the data (Figs. 6 and 7).
258 W. Stephen Pittard et al.

Fig. 4 Bar plot of phenylalanine intensity in all samples

Fig. 5 Using correlation between samples to examine data quality. (a) Scatter plot of all features between two
technical replicates. Each dot is a feature. The high number of values close to 10 (forming two lines) reflects
imputed values in the data. This scatter plot is typical for metabolomics and transcriptomics data, where
reproducibility is better for features of higher intensities. (b) Pair-wise Pearson correlation coefficients for all
samples plotted on a heatmap, where the color is scaled by the coefficient value
 Bioinformatics Primer for Data Science 259

Fig. 6 Using PCA plot to examine grouping patterns of biological samples. Each
dot represents a sample. One of the 0 h samples groups with the mock 6 h
samples. This raises a flag for further investigation

Fig. 7 The MA plot of two replicates or samples should have a flat trendline,
otherwise it indicates a bias in data distribution. The straight lines of features are
from imputed values, similar to those in Fig. 5a

5.2 Data For high-throughput data, it is often useful to filtering the data for
Transformation, missing values and low intensities, so that data of higher quality is
Scaling used for subsequent analysis. Many statistical methods do not cope
and Normalization with missing data, and imputation needs to be performed. It is
common in metabolomics data to replace missing values by limit
260 W. Stephen Pittard et al.

of detection. Similar to transcriptomics, the measurement errors in

metabolomics data tend to be multiplicative. After log transforma-
tion (usually by log2), the data are more likely to follow a normal
distribution.
Although variation is expected in data, it often arises due to
uncontrolled factors such as inconsistent specimen collection, or
instrument drift. This sort of variation can obscure insight one may
gain from what was supposed to be measured. Data normalization,
scaling and standardization attempt to suppress unwanted variation
while minimally affecting the quality of the data. These preproces-
sing methods are often required before further analysis can be done
as a way to prepare the data so that only what is meant to be
analyzed can be analyzed. Some algorithms, such as PCA and
Gradient descent in machine learning, require scaling, whereas for
others, such as t-test there is no effect but scaling may help clearer
presentation of the result.
Scaling is transforming the data be applying a function to
it. Normalization and standardization are both types of scaling.
Standardization is changing the data to have a mean of zero and a
standard deviation of one. Normalization for -omics data usually
refers to scaling each sample to a similar distribution of feature
intensities. Common methods of normalization include mean cen-
tering, LOESS (local weighted regression) and quartile normaliza-
tion. Methods more specific to metabolomics data are available
[3]. Of note, LC-MS metabolomics tends to have many missing
values for low-abundance features (similar to single-cell RNA-seq
data). Caution has to be taken using common normalization meth-
ods. In our example code, we used a strategy to guide normaliza-
tion using only the more abundant metabolites.

5.3 Common The field of metabolomics has a good share of complex statistics.
Statistical Analyses Multivariate methods are often used to project data into latent
spaces, and PLS-DA (partial least square discriminant analysis) is
among the popular methods to compare biological classes. Princi-
pal component analysis (PCA) is not directly used for statistical
analysis, but often used for dimension reduction and visualization.
Due to its unsupervised nature, PCA is also used for exploration
and quality control. In Fig. 6, one baseline sample is not grouped
with other baseline samples, but with a treatment group. This
figure raises a red flag on that particular sample and demands
further investigation of the cause.
Less complex but often preferred are the univariate statistics. If
the data are known to follow normal distribution, ANOVA (Analy-
sis of variance, three or more groups) or t-test (two groups) can be
used. They can be further divided into categories for independent
samples or related samples. If the distribution of data is not normal
or unknown, nonparametric methods such as Mann–Whitney U
test can be used. For continuous outcome, linear regression is
 Bioinformatics Primer for Data Science 261

Fig. 8 Visualization of results from a statistical test. (a) Volcano plot of all features. The Y-axis indicates the
significance of a feature in the statistical test, and the X-axis shows the direction and magnitude of the
difference between two groups. (b) A heatmap of the significant features (FDR < 0.05 here) the statistical test.
Each column represents a sample and cell of the heatmap represents a feature of one sample. The six
samples are separated into two biological groups, matched to the experimental design. (c) Boxplot to visualize
the mean level and standard deviation of one feature in the two biological groups

commonly used. A more extensive review on this topic can found in

Gardinassi et al., 2017 [4]. In our example code, we show the
comparison between two experimental groups using the Student’s
t-test, visualized via a volcano plot, a heatmap, and a box plot
(Fig. 8).
It is important to note that multiple testing is a major concern
in -omics data (also see Chapter 20). The chance of getting a small
p-value increases greatly when one performs the same test hundreds
or thousands of times. False discovery rate (FDR) is a common
approach to adjust the result by multiple test correction, imple-
mented with various algorithms. The reporting of FDR is consid-
ered necessary in scientific publications using metabolomics data.
These above statistical methods are available as both R and Python
programming tools. An example of more user-friendly tools is
MetaboAnalyst, which enables users to perform data analysis by
interacting with graphical web interface (see Chapter 17). After the
statistical analysis of metabolite features, pathway and network
analysis can be performed to bring the data to the relevant
biological context (see Chapter 19).

5.4 Visualization A large portion of scientific findings are communicated via figures.
For data intensive publications, the visual presentation has even
more special challenges. A good example is the rising of Circos
plots. Circos was initially developed for genome projects, but has
since been applied to many domains to illustrate the relationships
between complex data [5]. One can condense a great amount of
data of many different types of plots, such as line plots, histograms,
tiles and heatmap, into one figure. This type of figures has to be
262 W. Stephen Pittard et al.

generated by computer scripts, as manual artistry can no longer

cope with the complexity and precision.
A few principles apply to scientific figures, regardless how the
figures are generated. Firstly, the type of figures or plots should be
determined by the underlying message. For instance, a Venn dia-
gram is to visualize the overlap between two or more categories,
but will be ineffective if there are more than four categories, in
which case UpSet is a better solution (http://caleydo.org/). Sec-
ondly, illustrations are made to deliver a clear message. Optional
visual elements, if not related to the message, should be avoided.
This includes the parsimony of using colors. Readers will wonder
what the color codes for if a color is used in a figure. If the purpose
is not clear, the result will be more confusing than helpful. The real
estate on a figure should be prioritized for the elements that deliver
the message. If the pattern is shown by a heatmap is clear in an area
of 1 square inch, there is no need to plot the heatmap in 2 square
inches. But if the sample labels are important to the message, they
need to have an appropriate portion of the area and be clearly
readable. Thirdly, figures should adhere to the instructions by
print journals from the start. Journals generally require figures to
be submitted in separate files, and of enough resolution (e.g.,
300 dpi). Figures and their captions should not be embedded for
submission, because the publishers need to typeset the manuscripts
according to their standards. Many journals specify limited sizes to
choose for figures, for example, to accommodate 2-column or
3-column format of the print. Keeping fonts consistent for all
figures is required by many journals. All journals post their author
instructions on the web, and readers should at least familiarize with
at least a few of them. Finally, scientific results have to reproducible.
Most publications today undergo revisions. The manuscripts and
corresponding data analyses should be properly versioned. In fact,
for a manuscript involving complex data analysis, it is also a good
way to organize the underlying data clearly by each figure, so that
the figure can always be regenerated even after modifications
(Fig. 2c). We illustrate here a few common types of figures in our
example code, including barplots, boxplots, scatter plots, heat-
maps, and volcano plots (Note 2).

6 Notes

1. Code example. We use the moDC dataset from the Mummi-

chog paper [6] to illustrate major steps in processing, data
analysis and presentation. This was an experiment of infecting
immune cells by yellow fever virus (YFV). Three experimental
groups are baseline (0 h), mock infection (6 h of culture but no
virus), and YFV infection (6 h). Each group included three
biological samples, and each biological sample was run in
 Bioinformatics Primer for Data Science 263

triplicate on a mass spectrometer coupled with liquid chroma-

tography. The example data and code are available at the sup-
plemental website: https://metabolomics-data.github.io.
2. It is easier to learn the bioinformatics skills by working on a
project. The computer programming world changes quickly
and most people learn without taking a class. People learn
differently, and one may have to try various things to find a
suitable approach. Online help can be found via websites such
as Stack Overflow (https://stackoverflow.com/) and BioStars
(https://www.biostars.org).

Acknowledgments

This work has been funded, in part, by the US national Institutes of

Health via grants UH2 AI132345 (Li), U2C ES030163 (Jones, Li,
Morgan, Miller), U01 CA235493 (Li, Xia, Siuzdak), U2C
ES026560 (Miller), P30 ES019776 (Marsit), P50 ES026071
(McCauley), and the US EPA grant 83615301 (McCauley).

References

1. Zauhar RJ (2001) University bioinformatics metabolomics data. Curr Pharmacol Rep 3

programs on the rise. Nat Biotechnol 19(3):285 (6):374–383
2. Gilbert W (2003) Life after the helix. Nature 5. Krzywinski M et al (2009) Circos: an informa-
421:315–316 tion aesthetic for comparative genomics.
3. De Livera AM, Olshansky G, Simpson JA, Creek Genome Res 19:1639–1645
DJ (2018) NormalizeMets: assessing, selecting 6. Li S, Park Y, Duraisingham S, Strobel FH,
and implementing statistical methods for nor- Khan N, Soltow QA, Jones DP, Pulendran B
malizing metabolomics data. Metabolomics 14 (2013) Predicting network activity from high
(5):54 throughput metabolomics. PLoS Comput Biol
4. Gardinassi LG, Xia J, Safo SE, Li S (2017) Bio- 9(7):e1003123
informatics tools for the interpretation of
 Chapter 15

The Essential Toolbox of Data Science: Python, R, Git, and

Docker
W. Stephen Pittard and Shuzhao Li

Abstract
The daily work in data science involves a set of essential tools: the programming languages Python and R,
the version control tool Git and the virtualization tool Docker. Proficiency in at least one programming
language is required for data science. R is tied to a computing environment that focuses on statistics, in
which many new algorithms in genomics and biomedicine are first published. Python has a root in system
administration, and is a superb language for general programming. Version control is critical to managing
complex projects, even if software development is not involved. Docker container is becoming a key tool for
deployment, portability, and reproducibility. This chapter provides a self-contained practical guide of these
topics so that readers can use it as a reference and to plan their training.

Key words Bioinformatics, Data science, Python, R, Git, Docker, Version control, Virtualization

1 Introduction

The critical role of data science permeates many aspects of our

society. As for introductory bioinformatics, the basic concepts of
data management, analysis and visualization are covered in
Chapter 14, with real examples of metabolomics. Chapter 16
describes predictive modeling and machine learning. Other
specialized domains, such as high-performance computing, are
not covered in this book. This chapter covers the toolset of data
science in general: the programming languages Python and R, the
version control tool Git and the virtualization tool Docker. From
our experience, these are the essential tools required for data-
intense work on a daily basis.
Proficiency in at least one programming language is required
for data science, depending on application domains. Python has a
root in system administration, and is a superb language for general
programming. R is tied to a computing environment that focuses
on statistics, in which many new algorithms in genomics and bio-
medicine are first published. Both Python and R are interpreted,

Shuzhao Li (ed.), Computational Methods and Data Analysis for Metabolomics, Methods in Molecular Biology, vol. 2104,
https://doi.org/10.1007/978-1-0716-0239-3_15, © Springer Science+Business Media, LLC, part of Springer Nature 2020

265
266 W. Stephen Pittard and Shuzhao Li

high-level languages. They allow the direct interactions via com-

mand line with the language interpreter, thus easing the learning
curve and facilitating code prototyping. Between saving time
writing code and computing efficiency, both languages lean toward
the former. However, if the libraries used in an application are
implemented efficiently (e.g., written in C language), computing
performance needs not suffer. It is the excellent ecosystems of
available libraries and packages that establish the popularity of
these two languages in data science.
Version control is critical to managing complex projects, even if
software development is not involved. Docker container is becom-
ing a key tool for deployment, portability, and reproducibility.
Excellent tutorials and documentations are available for each of
these tools. This chapter is intended to be a self-contained practical
guide for readers to learn the topics in a structured pace. All these
tools are free to use and available on Microsoft Windows, Apple
Mac OS, and Linux operating systems.

2 Python for Data Science

The Python programming language was created by Guido van

Rossum in early 1990s. The name came from the British television
series “Monty Python” (not from the snake); thus, a flavor of
humor is seen occasionally in its code and documentation. It started
on a modern and elegant design, and code readability is enforced by
its syntax. Together with other features, Python is suited for both
small and large projects, and is used in many complex, enterprise-
level software applications.

2.1 Installation and Python is included by default in Apple Mac OS and Linux distribu-
Relevant Packages tions, as it also serves system administrative functions. Thus,
besides the Python package manager pip, Python libraries can be
installed via the software manager of the operating systems. For
scientific computing, several libraries are indispensable and called
the “SciPy stack” (as presented at https://www.scipy.org), includ-
ing NumPy, SciPy, Matplotlib, and pandas. The basic numerical
array data structure is provided by NumPy, and pandas enables
major modern data structures similar to those in R. Matplotlib
provides comprehensive plotting functions. The name of Matplotlib
reflects its initial similarity to Matlab, a popular commercial tool for
data analysis and visualization. But the library has been rewritten for
object oriented programming. Matplotlib is the basis of other
Python plotting tools, including seaborn. In addition, comprehen-
sive methods of machine learning are implemented in the scikit-
learn library.
Custom installation of Python on Mac OS and Linux is often
necessary to upgrade to newer versions, or accommodate multiple
versions. Python can be downloaded from its official site, https://
 Essential Toolbox of Data Science 267

Fig. 1 Common Python libraries for scientific computing that are shipped with the Anaconda distribution. The
logos may be trademarked by individual parties. The tool stack is compatible with all three major operating
systems

www.python.org, for Linux/UNIX, Mac OS, Windows, and other

operating systems. For beginners, the simplest way is to install
Anaconda, a software platform for data analytics and scientific
computing, which includes all the tools discussed in this section
(Fig. 1. Instruction at http://docs.continuum.io/anaconda/
install/). At the time of writing, users should start on Python 3.7
because Python 2 will not be supported beyond 2020.

2.2 Interactive Once Anaconda (or another Python distribution with the above
Computing with libraries) is installed, we can start working on data—the possibilities
Python SciPy Stack are almost endless! Evoke the interactive environment by typing
“python” in the command line shell and pressing the ENTERkey.
This interactive environment is marked by “>>>” starting every
line, and the code in the current line is executed every time the
ENTER key is pressed:

>>> 165.079 + 1.0073

166.0863

Import NumPy and pandas libraries:

>>> import numpy as np

>>> import pandas as pd

Try a few simple things:

>>> n = 5
>>> a = [2, 3, 4, 5, 6]
>>> a2 = np.array(a)
268 W. Stephen Pittard and Shuzhao Li

>>>
>>> type(n)
<class ’int’>
>>> type(a)
<class ’list’>
>>> type(a2)
<class ’numpy.ndarray’>
>>>
>>> # help will show the supporting documentation on the
object, “q” to quit
>>> help(a2)

This above snippet shows: n is defined as an integer, a as a list

and a2 as a n-dimensional array in NumPy. Same as in R, content
behind “#” is a comment and ignored from execution. Next, we
download the sample data “modc_ae_2012.txt” from our GitHub
site, and place it under this work directory (where we started this
Python interactive environment). We use the input function
read_csv in pandas to read the data:

>>> mydata = pd.read_csv("modc_ae_2012.txt", sep="\t")

>>> mydata.shape
(7995, 31)
>>> mydata.head()
mz time mz.min mz.max p_0hr_01_1 ...
0 85.02783 59.68820 85.02783 85.02783 15.5810 ...
1 85.04717 124.75120 85.04709 85.04739 14.4754 ...
2 85.06532 68.66651 85.06517 85.06547 14.4223 ...
3 85.10073 16.48022 85.10050 85.10078 14.5249 ...
4 86.05951 67.78485 86.05949 86.05980 10.6674 ...
[5 rows x 31 columns]
>>>

>>> mydata.mean(0)[:8]
mz 479.377606
time 165.949377
mz.min 479.327571
mz.max 479.427072
p_0hr_01_1 16.091406
p_0hr_01_3 16.144007
p_0hr_01_5 16.290922
p_0hr_02_1 16.295431
dtype: float64

In the above snippet, we read our example data into an object

mydata, which is a table of 7995 rows and 31 columns. A peek of
 Essential Toolbox of Data Science 269

the data (first 5 rows) is printed out by mydata.head(). We

calculate the mean value of each column by mydata.mean(0).
The dot “.” after an object calls for its attribute or method.

2.3 Concepts in Data For simplicity, we take a subset of the data (first 4 rows and
Structure 6 columns) for this section.

>>> sliced_data = mydata.iloc[0:4, 0:6]

>>> sliced_data
mz time mz.min mz.max p_0hr_01_1 p_0hr_01_3
0 85.02783 59.68820 85.02783 85.02783 15.5810
16.0425
1 85.04717 124.75120 85.04709 85.04739 14.4754
14.2709
2 85.06532 68.66651 85.06517 85.06547 14.4223
15.0515
3 85.10073 16.48022 85.10050 85.10078 14.5249
13.2573
>>> type(sliced_data)
<class ’pandas.core.frame.DataFrame’>
>>>
>>> sliced_data.iloc[0]
mz 85.02783
time 59.68820
mz.min 85.02783
mz.max 85.02783
p_0hr_01_1 15.58100
p_0hr_01_3 16.04250
Name: 0, dtype: float64
>>> type(sliced_data.iloc[0])
<class ’pandas.core.series.Series’>
>>>
>>> sliced_data.iloc[0].to_numpy()
array([85.02783, 59.6882 , 85.02783, 85.02783, 15.581 ,
16.0425 ])
>>> type( sliced_data.iloc[0].to_numpy() )
<class ’numpy.ndarray’>
>>>
>>> list(sliced_data.iloc[0].to_numpy())
[85.02783000000001, 59.6882, 85.02783000000001,
85.02783000000001, 15.581, 16.0425]

In the above example, we checked the type of each object. We

have sliced_data as a “DataFrame.” The method sliced_-
data.iloc[0] gets the first row of sliced_data, and is a
“Series.” These are data structures pandas uses to facilitate applica-
tions data science. Series is a one-dimensional labeled array of the
same data type, and DataFrame is a 2-dimensional labeled data
270 W. Stephen Pittard and Shuzhao Li

structure with columns of potentially different types. DataFrame is

similar to the data frame in R, and Series to the vector in R. We can
convert the Series to NumPy array by .to_numpy(), and further
to a regular list in Python by list().

2.4 Plotting We import Matplotlib here for data plotting.

>>> import matplotlib.pyplot as plt

>>> X = mydata[’p_0hr_01_1’]
>>> Y = mydata[’p_0hr_01_3’]
>>> plt.figure(figsize=(5,5))
<Figure size 500x500 with 0 Axes>
>>> plt.plot(X, Y, ’r.’, markersize=2)
[<matplotlib.lines.Line2D object at 0x7f359301b780>]
>>> plt.title("Scatter plot of two samples")
Text(0.5, 1.0, ’Scatter plot of two samples’)
>>> plt.savefig("Figure2.png")

The above snippet places the metabolite intensities of two

samples in our example data into variables X and Y, and creates a
scatter plot to visualize their relationship. The saved figure is shown
here as Fig. 2. The function plt.savefig() can further specify
figure resolution and formats such as PDF, PNG, or SVG.
This completes our introductory section on Python for data
science. To continue onto the next level, readers are encouraged to
complete two training materials: the 10-min tutorial linked on
pandas website (https://pandas.pydata.org/), and the 6-h course
of programming in Python on the website of Software Carpentry
(https://software-carpentry.org/lessons/). With additional experi-
ence, readers are also encouraged to use iPython and Jupyter note-
book in place of the default interactive environment.

3 R for Data Science

The R language is an open source implementation created by Ross

Ihaka and Robert Gentleman based on the S programming lan-
guage which was originally created at AT&T Bell Laboratories. R
provides an interactive environment for data manipulation, visuali-
zation, and statistical computing that includes strong support for
programming constructs and robust input and output capabilities.
R is extensible via the “package” mechanism which facilitates the
addition of user-contributed code. The Comprehensive Archive
Network (CRAN) offers 14,348 contributed packages at the time
of writing, across many application domains. R benefits from
enthusiastic user support and a highly motivated community of
developers, statisticians, researchers, and data scientists. An
extended version of this text in notebook format is posted at our
GitHub site https://metabolomics-data.github.io/.
 Essential Toolbox of Data Science 271

Fig. 2 Example of scatter plot in Python

3.1 Installation and The R project website https://cran.r-project.org/offers single click

Packages installers for Microsoft Windows, Apple Mac OS, and Linux (in the
form of packages) designed to provide the base R interactive envi-
ronment, language, graphics capability, and a foundational set of
statistical functions. A default installation of R is comprised of core
packages which offer basic functionality that can be extended with
user contributed packages via the install.packages() command. As
an example, the following will install the actuar package onto the
local system. The library() function is then used to load the pack-
age into the workspace after which functions within the package
may be used.

> install.packages("actuar")
Installing package into ‘/Users/esteban/r_packages’
(as ‘lib’ is unspecified)
--- Please select a CRAN mirror for use in this session ---
trying URL ’https://mirrors.nics.utk.edu/cran/bin/macosx/el-
capitan/contrib/3.5/actuar_2.3-1.tgz’
Content type ’application/x-gzip’ length 2044668 bytes (1.9
MB)
==================================================
downloaded 1.9 MB

The downloaded binary packages are in

/var/folders/wh/z0v5hqgx3dzdfgbr_3w0000gn/T//Rtmpy1m6A7/down-
loaded_packages

> library(actuar).
272 W. Stephen Pittard and Shuzhao Li

The associated Bioconductor project https://www.bio

conductor.org/ provides 1436 packages specifically for the explo-
ration of high-throughput genomic data. Management of
Bioconductor tools involves a different process that first requires
the installation of the BiocManager package.

> if (!requireNamespace("BiocManager", quietly = TRUE))

install.packages("BiocManager")
> BiocManager::install()

To then install a Bioconductor package, simply supply the

desired package name(s) as a vector:

> BiocManager::install("xcms") # Install the xcms package

> BiocManager::install(c("xcms","GenomicFeatures"))

3.2 Getting Help A good starting place for R assistance is the “Getting Help” page at
https://www.r-project.org/help.html which presents FAQs and
steps on how to walk through demonstrations of various functions
and packages. R also contains the browseVignettes() command
which provides a web listing of all available vignettes / guided
tours associated with a given package. Assistance is also available
via Stack Overflow http://stackoverflow.com with bioinformatics
specific support being available at the BioStars forum https://www.
biostars.org/. Please read the respective terms of use for both sites
prior to the submission of questions.

3.3 Variables and Variables are containers for data involving four primary types:
Data Structures numeric, character, logical, and factor. Variables are assigned
using the symbol “<
” or “¼“; however, the former is strongly
preferred. R is largely an interactive environment that retains vari-
ables and data structures during and, optionally, across sessions.
The interpreter accepts input after the “>” character.

> A <- 2.5 # The "<-" is the preferred method of assignment

> A = 2.5 # This is equivalent to the above although using

the "=" is
# strongly discouraged except when setting function
arguments.
> A
[1] 2.5
 Essential Toolbox of Data Science 273

R variables are case sensitive (“MyVar” vs. “myvar”), cannot

contain spaces (“my var”), or begin with symbols (.,$,%,^,&,∗,#,
(,),!). Character variables typically represent qualitative or labeled
information that might later be converted to factors, which are
variables which can only take a collection of predefined values.
Numeric variables typically represent measured information
although integers could also represent a coded or binary “Yes/
No” response. Logical variables take the value “TRUE” or
“FALSE,” and are usually the result of some comparison within a
programming construct (if else) or are used when indexing into
vectors, matrices, lists, or data frames.

3.3.1 Vectors Vectors are containers for homogeneous data that in turn can be
assembled into matrices. Vectors can be explicitly created using the
“c” function or be returned from other R functions.

> 1:5 # Create a vector from 1, 2, 3, 4, 5

[1] 1 2 3 4 5

> rnorm(8) # Generate 8 random values from a Normal

Distribution
[1] -0.75633261 0.70375541 -0.05123946 1.07488741
0.82288562 -0.45921131
[7] -0.48127334 0.58647522

> y <- 1:10 # A vector with 10 elements (1 .. 10)

> y <- c(1,2,3,4,5,6,7,8,9,10) # Same as above yet using the

"c" function

Consider the following example which simulates the collection

of height measurements of eight people. This example also intro-
duces the bracket notation which uses logical expressions to iden-
tify subsets of information.

> height <- c(59,70,66,72,62,66,60,60) # create a vector of

8 heights

> height[1:5] # Get first 5 elements

[1] 59 70 66 72 62

> height[c(1,5)] # Get just first and fifth elements

[1] 59 62
274 W. Stephen Pittard and Shuzhao Li

This example shows the use of more involved logical conditions

within the brackets.

# Find values between 60 and 70

> height[height > 60 & height < 70]
[1] 66 62 66

# Find heights between 60 and 70 inclusive

> height[height > 60 & height <= 70]
[1] 70 66 62 66

Vectors can be computed upon and used as input to functions.

> weight <- c(117,165,139,142,126,151,120,166) # weight

(in lbs)

> weight/100
[1] 1.17 1.65 1.39 1.42 1.26 1.51 1.20 1.66

> mean(weight)
[1] 140.75

3.3.2 Matrices A matrix can be considered as a vector with a dimension attribute.

To this end, the dim() function is used to impose a matrix structure
onto an existing vector. The matrix() function can also be used.

> set.seed(123) # Makes this example reproducible

> mymatrix <- rnorm(9)
> mymatrix
[1] -0.56047565 -0.23017749 1.55870831 0.07050839
0.12928774 1.71506499
[7] 0.46091621 -1.26506123 -0.68685285

> dim(mymatrix) <- c(3,3)

> mymatrix
[,1] [,2] [,3]
[1,] -0.5604756 0.07050839 0.4609162
[2,] -0.2301775 0.12928774 -1.2650612
[3,] 1.5587083 1.71506499 -0.6868529

> mymatrix <- matrix(rnorm(9),nrow=3,ncol=3)

> mymatrix
[,1] [,2] [,3]
 Essential Toolbox of Data Science 275

[1,] 0.7013559 -0.2179749 -0.6250393

[2,] -0.4727914 -1.0260044 -1.6866933
[3,] -1.0678237 -0.7288912 0.8377870

As with vectors, matrices can be indexed using bracket notation

which, in this case, will require two indices, one for rows and
another for columns. Omitting an index implies that all rows or
columns are desired.

> newmat[1,1] # First row and first column

[1] -0.5604756

> newmat[1,] # First row and all columns

[1] -0.56047565 0.07050839 0.46091621

> newmat[1:2,] # First two rows

[,1] [,2] [,3]
[1,] -0.5604756 0.07050839 0.4609162
[2,] -0.2301775 0.12928774 -1.2650612

3.3.3 Lists Lists provide a way to store heterogeneous data within a single
structure. Newcomers to R usually do not create lists except when
(1) writing a function that needs to return heterogeneous informa-
tion or (2) as a precursor to creating a data frame. Lists are more
commonly encountered in results returned by common statistical
modeling activities such as regression, decision trees, and random
forests.

> data(mtcars) # Load mtcars into the environment

> mylm <- lm(mpg ~ wt, data = mtcars)

Use the names() function to see the list elements.

> names(mylm)
[1] "coefficients" "residuals" "effects" "rank"
[5] "fitted.values" "assign" "qr" "df.
residual"
[9] "xlevels" "call" "terms" "model"

The “$” symbol is used to address specific list elements.

> mylm$rank
[1] 2
276 W. Stephen Pittard and Shuzhao Li

> mylm$call
lm(formula = mpg ~ wt, data = mtcars)

> mylm$coefficients
(Intercept) wt
37.285126 -5.344472

3.3.4 Data Frames Data frames are tightly coupled collections of variables organized
into a rectangular format. Data frames can be constructed from
vectors, lists, matrices or by reading in CSV files, or from informa-
tion being returned by APIs (Application Programming Inter-
faces). Best of all, data frames can hold different data types across
multiple columns. There are a number of data frames built into R
that can be used to illustrate how to work with them. The data()
function can be used to load a built-in data frame.

data(mtcars). # Load the mtcars data frame into the environ-

ment

# Determine structure. It looks like a list

> str(mtcars)
’data.frame’: 32 obs. of 11 variables:
$ mpg : num 21 21 22.8 21.4 18.7 18.1 14.3 24.4 22.8 19.2 ...
$ cyl : num 6 6 4 6 8 6 8 4 4 6 ...
$ disp: num 160 160 108 258 360 ...
$ hp : num 110 110 93 110 175 105 245 62 95 123 ...
$ drat: num 3.9 3.9 3.85 3.08 3.15 2.76 3.21 3.69 3.92 3.92
...
$ wt : num 2.62 2.88 2.32 3.21 3.44 ...
$ qsec: num 16.5 17 18.6 19.4 17 ...
$ vs : num 0 0 1 1 0 1 0 1 1 1 ...
$ am : num 1 1 1 0 0 0 0 0 0 0 ...
$ gear: num 4 4 4 3 3 3 3 4 4 4 ...
$ carb: num 4 4 1 1 2 1 4 2 2 4 ...

> nrow(mtcars) # How many rows does it have ?

[1] 32

> ncol(mtcars) # How many columns are there ?

[1] 11

The previously described bracket notation can be used with

data frames just as easily.

> mtcars[1:3,] # Get the first three rows

 Essential Toolbox of Data Science 277

mpg cyl disp hp drat wt qsec vs am gear carb

Mazda RX4 21.0 6 160 110 3.90 2.620 16.46 0 1 4 4
Mazda RX4 Wag 21.0 6 160 110 3.90 2.875 17.02 0 1 4 4
Datsun 710 22.8 4 108 93 3.85 2.320 18.61 1 1 4 1

# Get the first three rows and only columns one and four
> mtcars[1:3,c(1,4)]
mpg hp
Mazda RX4 21.0 110
Mazda RX4 Wag 21.0 110
Datsun 710 22.8 93

# Find all rows wherein the mpg is >= 30 and then select
columns 2 through 6

> mtcars[mtcars$mpg >= 30.0,2:6]

cyl disp hp drat wt
Fiat 128 4 78.7 66 4.08 2.200
Honda Civic 4 75.7 52 4.93 1.615
Toyota Corolla 4 71.1 65 4.22 1.835
Lotus Europa 4 95.1 113 3.77 1.513

3.4 Functions Functions are a very important part of the R language especially
given that all packages are comprised of predefined functions to
assist with a large variety of tasks. Users communicate within R
almost entirely through functions, thus consider writing one to
avoid repeating the interactive entry of commands. However,
before writing new code it is wise to first determine if there is not
already an existing function that performs the desired actions. A
significant idea associated with software development is “Don’t
Repeat Yourself” commonly abbreviated as “DRY.” Functions are
created using the function() directive and are stored as R objects
just like any other variable. Functions allow easy reuse of code that
can optionally be used in the creation of a package for distribution
on CRAN. Information on a preexisting function can be obtained
using the “?” character before the function name of interest.

> ?mean
mean package:base R
Documentation

Arithmetic Mean

Description:

Generic function for the (trimmed) arithmetic mean.

278 W. Stephen Pittard and Shuzhao Li

Usage:

mean(x, ...)

3.5 S3 and S4 Organizing high-throughput data types such as geospatial, time

Objects series, and experimental information can be challenging using a
single data frame or matrix. To provide a meaningful representation
might involve a combination of structures organized into an
“object” that makes intuitive sense for users. As an example, the
output from the linear modeling lm() function involves a list of
eleven elements (some of which are lists themselves) known to be
useful in model evaluation.

> mylm <- lm(mpg~.,data=mtcars)

> names(mylm)
[1] "coefficients" "residuals" "effects" "rank"
[5] "fitted.values" "assign" "qr" "df.
residual"
[9] "xlevels" "call" "terms" "model"

R provides “generic” functions such as plot(), summary(), and

print() which, based on the object type, will “dispatch” work to an
underlying “method” that knows what to do with the specific
object type in question. In the case of the print() function, there
are three methods relating to regression although the user need not
worry about which to call as the S3 Object system handles this:

> grep("lm",methods(print),value=TRUE)
[1] "print.glm" "print.lm" "print.summary.glm"
[4] "print.summary.lm"

> print(mylm)

Call:
lm(formula = mpg ~ ., data = mtcars)

Coefficients:
(Intercept) cyl disp hp drat
wt
12.30337 -0.11144 0.01334 -0.02148 0.78711
-3.71530
qsec vs am gear carb
0.82104 0.31776 2.52023 0.65541 -0.19942
 Essential Toolbox of Data Science 279

It is important to understand that various objects exist within

the typical R environment although novice and intermediate users
will primarily be consumers of objects as opposed to developers.
In R, there are three object systems, S3, S4, and Reference Classes.
Creating objects involve considerations customarily associated with
Object Oriented Programming (OOP) which are beyond the scope
of the text. Of note, to call for an attribute of an object, we use “$”
in R, which is the equivalent of the dot “.” in Python. The dot in a
string has no lexical meaning in R.

3.6 Graphics The R language provides a powerful environment for the visualiza-
tion of scientific data. It provides publication quality graphics,
which are fully programmable and reproducible. There are a num-
ber of useful output types (PDF, JPEG, PNG, SVG) in addition to
the default high resolution on-screen graphics capability. R Gra-
phics can be confusing given that there are three primary user-
focused graphics systems: Base, Lattice, and ggplot2.

3.6.1 Base Graphics Base graphics is the default display package included in every R
installation. It has both high and low level routines which provides
flexibility in developing customized plots. Base graphics is well
documented and there is significant support available via Google.

3.6.2 Lattice Graphics Lattice graphics http://lattice.r-forge.r-project.org/ is an imple-

mentation of Trellis graphics outlined in the “Visualization Data”
text by William Cleveland. Lattice simplifies the creation of
conditioned plots with automatic creation of axes, legends, and
other annotations. The user can specify a formula interface similar
to that used in predictive models and aggregation commands, thus
allowing the user to leverage previous R knowledge.

3.6.3 ggplot2 ggplot2 https://ggplot2.tidyverse.org/ is a relative new comer that

is written as an implementation of Leland Wilkinon’s “Grammar of
Graphics” which attempts to decompose a visualization into
semantic components designed to facilitate a better understanding
of the data. By discussing the visualization in terms of accepted
vocabulary, the ggplot2 package enables flexible exploration of data
without the need to commit to a fixed set of specific chart types
unless desired. ggplot2 is part of the “tidyverse” (http://tidyverse.
org) which is a collection of R packages that share a common set of
design philosophies, grammar, and data structures that seek to
simplify the management, organization, and display of data. How-
ever, the ggplot2 package can be used independently of the larger
tidyverse package.
280 W. Stephen Pittard and Shuzhao Li

4 Git for Version Control

Data science often involves writing scripts and software programs,

which are designed to perform discrete tasks during the exploratory
phases of the effort. As the needs of the project become more
apparent, these scripts are frequently updated and grow in com-
plexity which increases the likelihood of software errors (“bugs”)
being introduced at which point the developer must then locate the
last known functional version and also address the bug impacting
the latest version. This can be difficult given that folders can contain
files with arbitrary names which then necessitates the examination
of multiple files to locate the error free version. The difficulty in this
scenario can be compounded by the introduction of collaborators
and additional analysts seeking to participate in the effort. Also
consider that software developed in support of an eventual publica-
tion should have a transparent trackable version history to facilitate
resolution of reported bugs which in turn can enhance reproduc-
ibility of published results.

4.1 Git Git is an open source tool for managing revisions to a set of files that
experience change over time, that is, version control. Many tools
for version control have been developed in the past few decades,
but Git has become arguably the most popular. It was originally
developed in 2005 by Linus Torvalds to manage the development
of the Linux kernel. While Git can work with a general set of text
files, it is more commonly employed as part of collaborative soft-
ware development projects to track modifications to a “repository”
(a collection of files within a folder structure). Even If there is no
intent for public distribution, the developer benefits from the
ability to revert to previous versions or create new “branches” for
novel development without impacting the reference branch. Git
follows a distributed model that allows developers to “clone” repo-
sitories, along with any associated change history, and submit mod-
ifications for subsequent (re)integration into the reference copy,
commonly known as the “master branch.” Git can assist in the
detection and resolution of conflicts wherein changes to a file are
made by multiple people.
According to the 2018 Stack Overflow Developer Survey,
87.2% of respondents who use version control in their projects
use Git. It is ideal for managing large, distributed projects involving
hundreds of participants though it is also appropriate for the
laboratory-based informaticist developing code in support of a
publication. Git is language agnostic in that it views source code
as simple text files to be monitored for changes. The functionality
of Git is not a function of the selected programming language.
 Essential Toolbox of Data Science 281

4.2 Availability and Operating system specific installers are available at https://git-scm.
Installation com/downloads The clients provide a built-in GUI tool although
the method of interaction described in this text will relate to the
command line client available which is available via the Terminal
application in Apple Mac OS and Linux. The Windows install also
comes with a git shell that can be launched as a standard
application.

4.3 A Common Git There are formal methodologies for software development projects
Workflow although use of Git does not involve or require the adoption of any
specific approach. A common generic development workflow using
git involves the following steps:
1. Create a repository using “git init” or “git clone.”
2. Create/modify files in the working repository using a text
editor.
3. Add file(s) to be tracked using “git add.”
4. Use “git status” frequently to see the current state of files and
commits.
5. When the tracked files represent a logical point of progress then
commit changes using “git commit” or “git commit -a” for
subsequent commits.
Repeat steps 2–5 until the project is complete or at a logical
stopping point. There are many possible variations to this workflow
though it represents a useful starting point.

4.4 Key Concepts The following represent essential ideas for becoming productive
with Git. Each concept will be considered in detail.
Repository: A folder that has been placed under git control.
Origin: Every repository has an “origin,” which could be local
(on a hard drive) or a remote hosting service such as Gihub or
GitLab.
Branch: Each repository has a reference or “master” branch
from which copies can be made for purposes of experimentation,
testing, and bug fixing without impacting the availability of the
master branch.
Tracking: The activity of monitoring files for changes.
Commit: The activity of saving or taking a snapshot of existing
modifications. Each project will usually experience multiple com-
mits over time.

4.5 Getting Started Git requires some basic identity information such as e-mail address
although this does not have to be an actual working address.
Consider though that this information is what Git will use in the
meta data of the repository so should it be distributed in the future,
end users will have appropriate contact information. Launch a
282 W. Stephen Pittard and Shuzhao Li

terminal on Apple Mac OS or Linux or the Git Shell on Microsoft

windows. This will display a window containing a shell prompt that
typical includes a “$” prompt or some variation thereof.

/home/ubuntu/MyProject $ git config --global user.email "john-

[email protected]"
/home/ubuntu/MyProject $ git config --global user.name "John
Doe"

4.6 Create a Git repositories can be created from existing folders or new ones.
Repository The following commands demonstrate the creation of a new folder
and repository. Note that there are no files to manage just yet.

/home/ubuntu $ mkdir MyProject

/home/ubuntu $ cd MyProject/
/home/ubuntu/MyProject $ git init
Initialized empty Git repository in /home/ubuntu/MyProject/.
git/
/home/ubuntu/MyProject $ ls
/home/ubuntu/MyProject $
/home/ubuntu/MyProject $ ls -a # List hidden files
. .. .git

It is important to understand that working inside of a git

repository is no different than working in any other folder. Git
does not attempt to validate any code you develop nor does it
alter any user created files. However, git does manage the content
of the .git folder that is created by the “git init” command. This
folder exists to capture all the metadata for the repository. On
Linux and Apple systems any file beginning with a period character
is generally hidden to prevent accidental deletion. The presence of
the .git folder is what makes the MyProject folder an actual git
repository. Removing the .git folder would not remove any data
files, but any revision control information would be lost.

4.7 Adding Content Next, create some files using a text editor (e.g., vi, emacs).

/home/ubuntu/MyProject $ vi Readme.md
/home/ubuntu/MyProject $ vi regression.R
/home/ubuntu/MyProject $ cat Readme.md

# MyProject

This is the readme file for this project.

 Essential Toolbox of Data Science 283

More details will follow

/home/ubuntu/MyProject $ cat regression.R

data(mtcars)
mylm <- lm(mpg~.,data=mtcars)
summary(mylm)

Determine the status of these files according to git:

/home/ubuntu/MyProject $ git status

On branch master

No commits yet

Untracked files:
(use "git add <file>..." to include in what will be
committed)

Readme.md
regression.R

nothing added to commit but untracked files present (use "git

add" to track)

The results of the “status” command indicate that the two new
files are not currently being tracked. It also provides information on
how to initiate tracking which means that git will monitor the files
for changes. After adding the file, enter the “git status” command
which will show that the files are now “staged” for a “commit.”
Additional changes can be made and git will track them.

/home/ubuntu/MyProject $ git add ∗

/home/ubuntu/MyProject $ git status
On branch master

No commits yet

Changes to be committed:
(use "git rm --cached <file>..." to unstage)

new file: Readme.md

new file: regression.R
284 W. Stephen Pittard and Shuzhao Li

4.8 Making a Making a commit in git will take a “snapshot” of all changes and
Commit register them. This is accomplished using the “git commit” com-
mand. For an active development project, commits should be made
frequently. Remember that a benefit of the git system is that any
commit can be reverted should it result in a software error.

$ /home/ubuntu/MyProject $ git commit

This will invoke the default editor associated with the current
user account. For Ubuntu this will be “nano.” The following screen
illustrates a typical commit message which should be informative.

This is my first commit

# Please enter the commit message for your changes. Lines
starting
# with ’#’ will be ignored, and an empty message aborts the
commit.
#
# On branch master
#
# Initial commit
#
# Changes to be committed:
# new file: Readme.md
# new file: regression.R
#

After existing from the commit editor the following conforma-

tion of the commit activity can be observed:

[master (root-commit) f025b99] This is my first commit

2 files changed, 7 insertions(+)
create mode 100644 Readme.md
create mode 100644 regsession.R

Next, follow this up with a “git status.” Git indicates that all
changes have been noted and committed to the master branch. At
this point, more files can be created or new edits can be made to the
existing files followed by more commits.

/home/ubuntu/MyProject $ git status

On branch master
nothing to commit, working tree clean
 Essential Toolbox of Data Science 285

The “git log” provides a summary of the commit activity.

/home/ubuntu/MyProject $ git log

commit f025b99c53def7b6216d92d8113982d387610615 (HEAD -> mas-
ter)
Author: John Doe <[email protected]>
Date: Tue Jun 18 19:35:24 2019 +0000

This is my first commit

An alternative form of the log command provides a more

compact summary. Note that the number to the left is a “hash
number” that uniquely identifies each commit which simplifies the
process of retrieving or reverting to a previous commit.

/home/ubuntu/MyProject $ git log --pretty=oneline

f025b99c53def7b6216d92d8113982d387610615 (HEAD -> master)
This is my first commit

4.9 Inspecting This example demonstrates what happens when modifications are
Changes to Files made to a file that is currently being tracked for changes but has yet
to be committed. This is accomplished using the “git diff” com-
mand. At this point, neither of the two files in the repository have
been modified since the commit.

/home/ubuntu/MyProject $ git diff

/home/ubuntu/MyProject $

Modify the regression.R file which currently looks like:

data(mtcars)
mylm <- lm(mpg~.,data=mtcars)
summary(mylm)

Here are the changes.

# Predict the mpg of a car

data(mtcars)
mylm <- lm(mpg~hp+wt,data=mtcars)
summary(mylm)
286 W. Stephen Pittard and Shuzhao Li

The “git difference” command will note the recent changes

which in this case are (1) the addition of a new line at the top and
(2) a change in the line containing the call to the R lm() function. A
full discussion of the change format is beyond the scope of this text
though it is apparent that a new line 1 was added and the new line
3 (line 2 in the previous version) was modified.

/home/ubuntu/MyProject $ git diff

diff --git a/regression.R b/regression.R
index ccffd6c..b7db2aa 100644
--- a/regression.R
+++ b/regression.R
@@ -1,3 +1,4 @@
+# Predict the mpg of a car
data(mtcars)
-mylm <- lm(mpg~.,data=mtcars)
+mylm <- lm(mpg~hp+wt,data=mtcars)
summary(mylm)

Following this up with a call to “git status” confirms that a

change has been made and the file is staged for a subsequent
commit although it is not required. More changes can be made to
file including file deletions or additions.

/home/ubuntu/MyProject $ git status

On branch master
Changes not staged for commit:
(use "git add <file>..." to update what will be committed)
(use "git checkout -- <file>..." to discard changes in
working directory)

modified: regression.R

no changes added to commit (use "git add" and/or "git commit

-a")

At this point, apply a commit to take a snapshot of the current

changes. Note that output of the git status command indicates that
the “git commit -a” option may be used to update what will be
committed while the actual commit is taking place.

/home/ubuntu/MyProject $ git commit -a

Added a descriptive header line and changed lm predictor

variables
 Essential Toolbox of Data Science 287

# Please enter the commit message for your changes. Lines

starting
# with ’#’ will be ignored, and an empty message aborts the
commit.
#
# On branch master
# Changes to be committed:
# modified: regression.R
#

[master 457c2ee] Added a descriptive header line and changed lm

predictor variables
1 file changed, 2 insertions(+), 1 deletion(-)

Now check the status and commit event log to confirm that the
master branch is up to date.

/home/ubuntu/MyProject $ git status

On branch master
nothing to commit, working tree clean

/home/ubuntu/MyProject $ git log

commit 457c2ee2a189af1f067bc2d82eb01b01f4d9f206 (HEAD -> mas-
ter)
Author: John Doe <[email protected]>
Date: Tue Jun 18 20:42:27 2019 +0000

Added a descriptive header line and changed lm predictor

variables

commit f025b99c53def7b6216d92d8113982d387610615
Author: John Doe <[email protected]>
Date: Tue Jun 18 19:35:24 2019 +0000

This is my first commit

4.10 Using Branches An attractive feature of Git is the ability to create copies of the
master branch into arbitrarily named experimental branches with-
out impacting the master branch code. Think of any new branches
as being derivative safe copies of previously committed code which
can be managed using the usual git command set. Changes to a
derivative branch can, if desired, be integrated back into the master
branch or remain in a separate branch for more testing. The “git
branch” command indicates any existing branches and places an
288 W. Stephen Pittard and Shuzhao Li

asterisk next to the currently active branch which in this case is the
“master” branch.

/home/ubuntu/MyProject $ git branch

∗ master
/home/ubuntu/MyProject $

This example will create a branch called “experiment” although

the current branch is still the master branch. The “git branch
<branchname>” creates a new branch name “branchname.”

/home/ubuntu/tester $ git branch experiment

/home/ubuntu/tester $ git branch
experiment
∗ master

The command “git checkout experiment” will checkout the

contents of the master branch into the branch named “experiment”
and make “experiment” the current branch:

/home/ubuntu/tester $ git checkout experiment

Switched to branch ’experiment’
/home/ubuntu/tester $ git branch
∗ experiment
master

The only thing that has changed is the current branch. The
underlying folder contents are the same. Create a new file called
“logistic_regression.R” with the following contents and commit
the change with a commit message of “Added a file for logistic
regression.”

# Logistic Regression Example

mtcars$am <- factor(mtcars$am)
myglm <- glm(am~.,data=mtcars)
summary(myglm)

/home/ubuntu/tester $ git add logistic_regression.R

/home/ubuntu/tester $ git commit
[experiment 3c8a02f] Added a file for logistic regression
1 file changed, 4 insertions(+)
create mode 100644 logistic_regression.R

/home/ubuntu/tester $ ls
 Essential Toolbox of Data Science 289

Readme.md logistic_regression.R regression.R

Remember that the new file has been added within the experi-
mental branch and will NOT impact the master branch. To verify
this, switch back to the master branch by using the “git checkout
master” command.

/home/ubuntu/tester $ git branch

∗ experiment
master
/home/ubuntu/tester $ git checkout master
Switched to branch ’master’

/home/ubuntu/tester $ git branch

experiment
∗ master

/home/ubuntu/tester $ ls
Readme.md regression.R

Observe that the master branch does not include the logisti-
c_regression.R file since it exists only within the experiment
branch. It is possible to merge the changes made in the experiment
branch into the master branch by using the “git merge” command.

/home/ubuntu/tester $ git branch

experiment
∗ master

/home/ubuntu/tester $ git merge experiment

Updating 457c2ee..3c8a02f
Fast-forward
logistic_regression.R | 4 ++++
1 file changed, 4 insertions(+)
create mode 100644 logistic_regression.R

/home/ubuntu/tester $ ls
Readme.md logistic_regression.R regression.R

Once the merge is complete the experiment branch can be

removed.

/home/ubuntu/tester $ git branch

experiment
∗ master
290 W. Stephen Pittard and Shuzhao Li

/home/ubuntu/tester $ git branch -d experiment

Deleted branch experiment (was 3c8a02f).

/home/ubuntu/tester $ git branch

∗ master

4.11 Sharing Using git is an effective way to manage any number of development
Repositories projects in a way that preserves changes made during the project
lifecycle which enhances reproducibility and serves as documenta-
tion of the project’s development trajectory. The examples pre-
sented thus far use a locally created repository which could be
shared with other users although that is optional. However, it is
common for developers to share projects using a hosting service
such as Github that supports over 36 million developers and
100 million repositories. Ref https://github.com/ Github offers
a free hosting plan for unlimited public and private repositories,
issue and bug tracking, and project management tools. Proceed to
https://github.com/join?source¼header-home to create an
account and select the free plan.
Many newcomers to Git are already familiar with GitHub
having been directed to download a repository by a publication or
a colleague. Knowledge of git commands is usually not essential to
simply using software in a repository although to participate in
development efforts does require such knowledge. Login to the
previously created Github account and create a New repository by
clicking the “+” button in the upper right corner of the Github
screen (Fig. 3):
After selecting “New Repository” a screen will be displayed
that prompts for information including the desired repository
name, an optional description, whether the repository should be
public (leave this selected), and whether to initialize it with a
README file (leave selected). After providing this information,
click the green “Create repository” at the bottom of the screen
(Fig. 4):
The next screen will display a summary of the repository from
which the green “Clone or Download” button can be clicked. After
clicking this button, in the pop-up box the URL for the repository
will be shown. Click the clipboard icon to copy the URL.

4.12 Cloning the As Git is decentralized, a repository can be cloned by any number of
Repository Locally parties interested in contributing to a project. To clone the reposi-
tory locally, open a Terminal and enter the following git command
which will download the repository into the current folder.

/home/ubuntu $ git clone https://github.com/wspem/mycoolrepo.

git
 Essential Toolbox of Data Science 291

Fig. 3 Creating a new repository in Github

Fig. 4 Repository information screen on Github

Cloning into ’mycoolrepo’...

remote: Enumerating objects: 3, done.
remote: Counting objects: 100% (3/3), done.
remote: Total 3 (delta 0), reused 0 (delta 0), pack-reused 0
292 W. Stephen Pittard and Shuzhao Li

Fig. 5 Github Repository main page

Unpacking objects: 100% (3/3), done.

/home/ubuntu $ ls -a
. .. .git README.md

The contents of the repository will match the contents reflected

on the Github website. There is only a single README.md file
which was created at the same time that the repository itself was
created.

/home/ubuntu $ cd mycoolrepo/
/home/ubuntu/mycoolrepo $ ls
README.md

An important git concept mentioned earlier was that of “ori-

gin” which up until now has not been a consideration since the
repository created in the previous section was local. However, since
the “"mycoolrepo” repository was created on Github and then
cloned locally, the origin of the repository should reflect that
sequence of events. The command “git remote –v” will confirm
the correct point of origination of “mycoolrepo.”

/home/ubuntu $ cd mycoolrepo/
/home/ubuntu/mycoolrepo $ git remote -v
origin https://github.com/wspem/mycoolrepo.git (fetch)
origin https://github.com/wspem/mycoolrepo.git (push)

Files can still be added, modified, or deleted just as with the

earlier local repository so create a new file called plotter.R with the
following contents.
 Essential Toolbox of Data Science 293

# Program to plot MPG vs Wt

data(mtcars)
plot(mpg~wt,data=mtcars)

Now use “git status”.

/home/ubuntu/mycoolrepo $ git status

On branch master
Your branch is up to date with ’origin/master’.

Untracked files:
(use "git add <file>..." to include in what will be
committed)

plotter.R

nothing added to commit but untracked files present (use "git

add" to track)

This is a familiar result based on the earlier work with the local
MyProject repository. At this point add and commit the file and
then check the status.

/home/ubuntu/mycoolrepo $ git add plotter.R

/home/ubuntu/mycoolrepo $ git commit
git commit
[master 07874ac] Added a program to plot MPG vs Wt
1 file changed, 4 insertions(+)
create mode 100644 plotter.R

/home/ubuntu/mycoolrepo $ git status

On branch master
Your branch is ahead of ’origin/master’ by 1 commit.
(use "git push" to publish your local commits)

nothing to commit, working tree clean

The message indicates that the commit worked though given

that a new file has been added the local version of the repository, it
is now “ahead” of the origin repository currently residing on
Github. Work could continue and commits could be made but
this local copy of the mycoolrepo repository will be out of sync
with the version on Github. There is a command that will allow one
to “push” these changes to the master branch located at the origin
which is
294 W. Stephen Pittard and Shuzhao Li

/home/ubuntu/mycoolrepo $ git remote -v

origin https://github.com/wspem/mycoolrepo.git (fetch)
origin https://github.com/wspem/mycoolrepo.git (push)

Push up the changes. This will require knowing the user id and
password for the Github account created earlier.

/home/ubuntu/mycoolrepo $ git push origin

Username for ’https://github.com’: wspem
Password for ’https://[email protected]’:
Counting objects: 3, done.
Compressing objects: 100% (3/3), done.
Writing objects: 100% (3/3), 347 bytes | 347.00 KiB/s, done.
Total 3 (delta 0), reused 0 (delta 0)
To https://github.com/wspem/mycoolrepo.git
e61a151..07874ac master -> master

/home/ubuntu/mycoolrepo $ git status

On branch master
Your branch is up to date with ’origin/master’.

nothing to commit, working tree clean

Refer back to the web page associated with the Github reposi-
tory which in this example is https://github.com/wspem/
mycoolrepo Note that the file plotter.R that was added locally is
now present in the Github repository (Fig. 5).

4.13 Next Steps This walkthrough has provided an overview of Git and Github
basics although there remains a significant degree of capability
that cannot be covered due to space constraints. Understanding
how to manage software conflicts and work with repositories
owned by other developers is also important information. In any
case, the material presented in this section represents a solid basis
upon which laboratory-based developers can efficiently build repro-
ducible software projects.

5 The Docker Container

A software application of any real-world complexity will depend on

other software and libraries to function. These software and
libraries will further depend on computing environments including
services from the operating systems. Some of those may be available
only on certain operating systems (e.g., on Linux but not Win-
dows). The access to multiple operating systems, without
 Essential Toolbox of Data Science 295

Fig. 6 Diagram of virtualized systems

purchasing a new computer, can be provided by the use of virtual

machines. For example, VirtualBox allows Linux to run on a local
Windows system. In this example the following terms would apply
(Fig. 6):
Server: The actual physical hardware.
Host OS: The operating system installed on the Server (Micro-
soft Windows).
Hypervisor: The software which enables virtualized services
(e.g., Virtual Box).
Guest OS: The operating system installed within a virtualized
environment (Linux).
Use of a hypervisor allows for one or more guest operating
systems, assuming the server hardware provides adequate storage
and memory resources. The guest operating system must be
installed including any supporting system and applications libraries.
The hypervisor provides access to the functioning virtual compu-
ters, and corresponding operating systems, each of which can access
the local system’s hardware (CD/DVD, mouse, USB ports, etc.).
Figure 2 shows an Apple laptop running VirtualBox that is used to
host a copy of Mint Linux. Virtualized systems are particularly
useful in situations that require a significant number of applications
which justifies the effort involved in creating and configuring a
virtual machine along with the supporting OS and application
libraries (Fig. 7).
A disadvantage of virtual machines is the resource requirement
(CPU cycles, memory and disk space) on the host computer. In
effect, the host system can become over-subscribed. A second
disadvantage is it assumes the user has sufficient knowledge to
296 W. Stephen Pittard and Shuzhao Li

Fig. 7 Linux Mint running in VirtuLBox on Apple Mac OS

install and manage the necessary operating systems and application

packages. Many times, a project requires a single application in
which case, using a virtual machine represents more cost that it is
worth. A lighter alternative technology is containers, and the best
example of containers is Docker.
Another reason of using containers, perhaps more important to
data science, is the dependency/compatibility issue of computer
software. The libraries that a software application depends on will
change over time and have many versions, and some be out of
maintenance after a while. Therefore, a software application often
fails to run because its dependency on those libraries breaks. The
solution is to have a snapshot of the exact versions and copies of
these libraries that the software application is created to run on
(which is also important to the reproducibility of science). Even
though one could create a virtual machine image for this, but to do
it regularly as part of daily routine, containers are the answer. In
practice, containers often reduce the complexity of installing or
deploying a software application.

5.1 Installing Docker Free installers for Apple Mac OS and Microsoft Windows are
available from Docker Hub which requires the creation of an
account. The install provides both the Docker daemon and client
CLI command line interface.
Microsoft Windows—https://docs.docker.com/docker-for-
windows/install/.
 Essential Toolbox of Data Science 297

Fig. 8 Comparing a virtual machine and docker

Apple Mac OS—https://docs.docker.com/docker-for-mac/

install/.
Linux—Refer to package repository for the specific
distribution.
After the installation of Docker the “hello-world” image can be
used to validate the installation.

$ docker run hello-world

Unable to find image ’hello-world:latest’ locally
latest: Pulling from library/hello-world
1b930d010525: Pull complete
Digest: sha256:41a65640635299bab090f783209-
c1e3a3f11934cf7756b09cb2f1e02147c6ed8
Status: Downloaded newer image for hello-world:latest

Hello from Docker!

This message shows that your installation appears to be working
correctly.

To generate this message, Docker took the following steps:

1. The Docker client contacted the Docker daemon.
2. The Docker daemon pulled the "hello-world" image from the
Docker Hub.
(amd64)
3. The Docker daemon created a new container from that image
which runs the
executable that produces the output you are currently
reading.
298 W. Stephen Pittard and Shuzhao Li

4. The Docker daemon streamed that output to the Docker

client, which sent it
to your terminal.

To try something more ambitious, you can run an Ubuntu

container with:
$ docker run -it ubuntu bash

Share images, automate workflows, and more with a free Docker

ID:
https://hub.docker.com/

For more examples and ideas, visit:

https://docs.docker.com/get-started/

In the Docker domain, the concept of a virtual machine is

replaced by a “container” which employs an “image” of a specific
tool or application, as well as any dependencies, that can be exe-
cuted under the control of a Docker process running on the indi-
vidual’s computer. Docker containers can be executed in isolation
although it is easy to enable communication between containers to
create services from independent components. The differences
between a Virtual Machine and a Docker setup can be observed
in the following figure. Note the absence of a guest operating
system in the Docker architecture. Also, the Hypervisor is replaced
by a Docker process that manages containers and image (Fig. 8).

5.2 Key Docker From an end user point of view, the terminology for Docker
Concepts involves a basic knowledge of the following ideas:
Client/CLI—This is the CLI (command line interface) that
allows the user to issue various docker commands. The client
usually runs on the same host as the Docker daemon process. The
CLI offers a suite of commands for managing containers, images,
volumes, and networking between containers.
Docker Process/Daemon—This is the persistent process that
handles client requests, checks with the registry, and manages con-
tainers. This usually runs on the same host as the Docker client.
Container—A packaging construct designed to insulate docker
images from the environment in which they will execute. A con-
tainer can be thought of as an actively running image with which
the end user can interact.
Image—A collection of bundled dependencies necessary to
execute an application. Base images are those not reliant upon
other images. Examples are operating systems (e.g., Ubuntu,
Debian). Child images are those created on top of Base images.
Docker Registry—A registry is a collection of images that can
be searched and deployed. Consider it as a reference point for
locating and initiating containers. An example is Docker Hub
https://www.docker.com/products/docker-hub.
 Essential Toolbox of Data Science 299

Fig. 9 The search command in docker

5.3 Finding Useful There are a number of prebuilt images relating to bioinformatics
Images and supporting languages including Python and R. For example,
the Rocker project https://www.rocker-project.org/ offers Docker
Containers for the R Environment as well as instructions on how to
run the images. Image discovery is straightforward by using the
search capability which is part of the default installation of Docker.
The “search” command is used to locate images of interest. Note
also that there is a web-based search tool Located at https://hub.
docker.com/search?q¼rocker&type¼image that offers an intuitive
user interface (Fig. 9)
The “docker search” command has a number of options that
can help filter output though the basic format given above is typical.
The NAME and DESCRIPTION files document the image label
and any supplied information. The STARS field is a community
rating of the image which is how the output is sorted. Once the
desired image is found it is simple to execute it. Below is an example
to run “rocker/r-base.”

$ docker run -it rocker/r-base

Unable to find image ’rocker/r-base:latest’ locally
latest: Pulling from rocker/r-base
2666d10a4f80: Pull complete
2c0f31f3b229: Pull complete
8978e71a606b: Pull complete
3a18d5b41e17: Pull complete
3b9876199949: Pull complete
1ecd21a8af49: Pull complete
Digest: sha256:35099e4073c8aff2c616add01d7ffd452bf79c78a893-
b08a9ac30e8114cab247
Status: Downloaded newer image for rocker/r-base:latest

R version 3.6.0 (2019-04-26) -- "Planting of a Tree"

Copyright (C) 2019 The R Foundation for Statistical Computing
Platform: x86_64-pc-linux-gnu (64-bit)

R is free software and comes with ABSOLUTELY NO WARRANTY.

You are welcome to redistribute it under certain conditions.
300 W. Stephen Pittard and Shuzhao Li

Type ’license()’ or ’licence()’ for distribution details.

Natural language support but running in an English locale

R is a collaborative project with many contributors.

Type ’contributors()’ for more information and
’citation()’ on how to cite R or R packages in publications.

Type ’demo()’ for some demos, ’help()’ for on-line help, or

’help.start()’ for an HTML browser interface to help.
Type ’q()’ to quit R.

>

The command “docker run -it rocker/r-base” will

(1) contact the local Docker process/daemon to see if there is a
local image named “rocker/r-base:latest”. If not found, the
Docker process can be configured to (2) consult a registry to
determine the existence of the image. Once located, the Docker
process will (3) “pull” (download) the image along with any sup-
porting images. The Docker process will (4) then start a container
on the host which the user can then engage to experiment with the
new metabolomics package without impacting the version of R
residing on the laptop. Once the experiment has concluded the
user can stop the container, by existing R or using “docker stop,”
and remove the image if desired or develop a new analysis pipeline
that could then be converted to an image (Fig. 10).

Fig. 10 Workflow of a command run on docker

 Essential Toolbox of Data Science 301

5.4 Managing A notable side effect or running containers is that any associated
Containers and Images image might require additional image layers that contribute to the
overall download size. Images remain on the local hard drive unless
explicitly removed. The Docker client includes commands to man-
age containers and images as well as tools to create them. Start the
following container using the ubuntu:latest image command.

$ docker run -it --name=local_ubuntu ubuntu:latest

Unable to find image ’ubuntu:latest’ locally
latest: Pulling from library/ubuntu
5b7339215d1d: Pull complete
14ca88e9f672: Pull complete
a31c3b1caad4: Pull complete
b054a26005b7: Pull complete
Digest: sha256:9b1702dcfe32c873a770a32cfd306dd7fc1c4f-
d134adfb783db68defc8894b3c
Status: Downloaded newer image for ubuntu:latest
root@06949ec991be:/#

The “-it” options indicate a requested interactive session with

the container via a terminal / command line prompt. The “---
name ¼ local_ubuntu” is an optionally assigned label for the
container which simplifies identification should there be other
active containers on the system.
The docker “ps” command will show all resident containers.
The “-a” option shows inactive containers also. From another
terminal / command line, enter$ docker ps -a (Fig. 11).
Notice that the container has a “CONTAINER ID” assigned
by the Docker process although the more intuitive name “loca-
l_ubuntu” may also be used to identify and manage the container.
The container can be stopped using the Docker CLI via “stop”
(Fig. 12).
Stopping the container does not remove it from the system. If a
later attempt is made to restart it via the docker run command then
the following would occur:

Fig. 11 Using the ps command

Fig. 12 Using the stop command

302 W. Stephen Pittard and Shuzhao Li

$ docker run -it --name=local_ubuntu ubuntu:latest

docker: Error response from daemon: Conflict. The container
name "/local_ubuntu" is already in use by container
"c4ec515779368cb14d4b9d79fac86d53e4bb018125277eaf6a-
b65e9438d89814". You have to remove (or rename) that container
to be able to reuse that name.
See ’docker run --help’.

To remove the local_ubuntu container from the system

(Fig. 13).
The container has been removed; however, the supporting
image still remains locally. The “rmi” command can be used to
delete the local image (Fig. 14):
In general, it is not necessary to remove containers and images
from the system unless disk space becomes an issue in which case
these approaches can be used to perform cleanup. One way to
prevent a container from persisting locally is to specify the “-rm”
option when running the docker. After the container is stopped,
the container will be deleted. Run the ubuntu:latest image as
before:

$ docker run -it --rm --name=local_ubuntu ubuntu:latest

Unable to find image ’ubuntu:latest’ locally
latest: Pulling from library/ubuntu
5b7339215d1d: Pull complete
14ca88e9f672: Pull complete
a31c3b1caad4: Pull complete
b054a26005b7: Pull complete
Digest: sha256:9b1702dcfe32c873a770a32cfd306dd7fc1c4f-

Fig. 13 Using the rm. command to remove Ubuntu container

Fig. 14 Using rmi command to delete local image

 Essential Toolbox of Data Science 303

Fig. 15 Using the ps command to verify the existence of the container and image

Fig. 16 Using the ps command to verify the existence of image but not the container

Fig. 17 Rocker images which support the R language

d134adfb783db68defc8894b3c
Status: Downloaded newer image for ubuntu:latest
root@ee1ef120852d:/#

From another Terminal window use images and ps commands

to verify the existence of the container and image (Fig. 15):
Back in the ubuntu:latest shell, type the word “exit” to end
the container session:

root@ee1ef120852d:/# exit
exit

Next, verify that the container has not been maintained

although the image has (Fig. 16).

5.5 A More Practical As seen earlier, the rocker/r-base image offers command line
Use Case access to the R environment although the rocker project offers
images in support of the R language (Fig. 17) .
The rocker/rstudio image in particular provides R lan-
guage support along with the RStudio IDE which in this case is a
web-based service accessible locally via a browser. Because of this,
some options will need to be supplied to appropriately map the
RStudio port (typically 8787) to a local system port which can also
be 8787 assuming no other local service is using it. The RStudio
IDE also requires the specification of a password at run time. The
304 W. Stephen Pittard and Shuzhao Li

default user is “rstudio.” This is how the associated docker run

command would look:

$ docker run -e PASSWORD="testpass" --name=local_rstudio -p

8787:8787 rocker/rstudio

[s6-init] making user provided files available at /var/run/s6/

etc...exited 0.
[s6-init] ensuring user provided files have correct perms...
exited 0.
[fix-attrs.d] applying ownership & permissions fixes...
[fix-attrs.d] done.
[cont-init.d] executing container initialization scripts...
[cont-init.d] add: executing...
Nothing additional to add
[cont-init.d] add: exited 0.
[cont-init.d] userconf: executing...
[cont-init.d] userconf: exited 0.
[cont-init.d] done.
[services.d] starting services
[services.d] done.

The “-e” option will set the PASSWORD within the container.
Note that this command will be running in attached mode which
means that the terminal window in which it was entered will be
blocked/engaged until the container is terminated. This is not
necessarily a problem since the point of user interaction will be
the RStudio login panel. In order to run rocker/rstudio as a
detached container, include the “-d” option which will leave the
terminal free for input.
Next, launch a web browser and provide a URL of http://
localhost:8787. This will bring up the RStudio login page. Provide
a user name of “rstudio” and a password of “testpass”
(Fig. 18).
After clicking the “Sign In” button, the RStudio IDE will be
presented for engagement. At this point the user can install any
additional packages and create code as with any instance of R
(Fig. 19).

5.6 Sharing Data Remember that the container is running under the supervision of
the Docker process and the containerized R environment will in no
way impact any locally existing versions of R. However, a problem
with this specific example is that there is nowhere to save code and
have it persist except within the container itself. For example, a user
might have existing R code in a local folder (e.g., metabolo-
mics_r_code) intended for use with the container. Or, in working
with the container, the user might want to save code to the local
 Essential Toolbox of Data Science 305

Fig. 18 Login panel for Rocker/R-studio

Fig. 19 Rocker/R-studio container

system outside of the container or with future containers. Docker

has a solution for this in the form of volume mapping which can be
specified on the run command line. Stop the running rocker/
verse instance either by typing Control-C in the terminal window
used to launch the container or by using “docker stop.”
306 W. Stephen Pittard and Shuzhao Li

The user wants to expose the metabolomics_r_code folder

to the rocker/rstudio container. The folder resides within /
Users/esteban.

$ cd ~
$ cd metabolomics_r_code/
$ pwd
/Users/esteban/metabolomics_r_code
$ ls
linear_regression.R. logistic_regression.R support_vector_-
machine.R

$ docker run -e PASSWORD="testpass" --name=local_rstudio_vol

-v /Users/esteban/metabolomics_r_code:/home/rstudio/metabolo-
mics_r_code -p 8787:8787 rocker/rstudio

[s6-init] making user provided files available at /var/run/s6/

etc...exited 0.
[s6-init] ensuring user provided files have correct perms...
exited 0.
[fix-attrs.d] applying ownership & permissions fixes...
[fix-attrs.d] done.
[cont-init.d] executing container initialization scripts...
[cont-init.d] add: executing...
Nothing additional to add
[cont-init.d] add: exited 0.
[cont-init.d] userconf: executing...
[cont-init.d] userconf: exited 0.
[cont-init.d] done.
[services.d] starting services
[services.d] done.

As before, use a browser to login to the interface at http://

localhost:8787.
Once the RStudio interface is displayed, notice the existence of
the metabolomics_r_code folder which is now associated with
the same folder within /Users/esteban (Fig. 20).
Clicking the folder will verify the existence of the user code
(Fig. 21).
These files may be edited or removed with any changes within
this container being reflected in /Users/esteban/metabolo-
mics_r_code. Even after the local_rstudio_vol container is
stopped or removed the changes will persist in the local folder. The
“-v” option is what makes this possible. In this case the run
command included:

-v /Users/esteban/metabolomics_r_code:/home/rstudio/metabolo-
mics_r_code
 Essential Toolbox of Data Science 307

Fig. 20 Shared folder in R-studio container

Fig. 21 User files

Notice the semicolon character that separates the two path

specifications. The path on the left, /Users/esteban/metabolo-
mics_r_code refers to the folder currently residing on the user’s
computer. The second path, /home/rstudio/metabolomics_r_-
code, indicates where in the container the local folder should be
mapped to. The /home/rstudio folder exists in the folder, and
depending on the image being used, this might need to change,
though within the rocker domain the /home/rstudio folder will
exist.
There are many R packages available in a Docker image format,
so it is important to thoroughly search Docker Hub to take advan-
tage of their existence. The rocker project is useful as it is the
Bioconductor project which provides tools for the analysis of
high-throughput genomic data.
More specifically there is a Docker image for the xcms R
package for processing and visualization of chromatographically
separated and single-spectra mass spectral data.

$ docker search bioconductor --limit=5 --format ’{{.Name}} \t

308 W. Stephen Pittard and Shuzhao Li

{{.Description}}’
bioconductor/release_base release base container
bioconductor/release_core release core container
bioconductor/devel_core2 Automated Build Bioconductor
Develement Core. . .
bioconductor/devel_base2 Automated Build Bioconductor
Develement Base. . .
bioconductor/devel_mscore2 Automated Build Bioconductor
Develement msco. . .
$
$ docker search xcms --limit=5 --format ’{{.Name}} \t {{.
Description}}’
yufree/xcmsrocker Rocker image for metabolomics study
payamemami/xcms-container
yufree/xcms Non-Target Data Analysis Environment with CR. . .
pcm32/xcms-camera
wilsontom/xcms-dockerdev

5.7 Dockers for The Jupyter project (https://jupyter.org/) provides a notebook

Python system that supports over 40 languages, although it is strongly
associated with the Python language. There are docker images
available for Jupyter (Fig. 22):
Note that the “docker search” command has an option called
“format” which permits selection of specific output fields. The
default is to provide information on five fields that, depending on
the length of the Description field, might result in too much
information. The format option accepts a template specification
to extract only those field names of interest.

$ docker search jupyter --limit=10 --format ’{{.Name}}’

jupyter/datascience-notebook
jupyter/all-spark-notebook
jupyterhub/jupyterhub
jupyter/scipy-notebook
jupyter/tensorflow-notebook

Fig. 22 Jupyter images on docker

 Essential Toolbox of Data Science 309

jupyter/pyspark-notebook
jupyter/minimal-notebook
jupyter/base-notebook
jupyterhub/singleuser
jupyter/nbviewer

$ docker search jupyter --limit=10 --format ’{{.Name}} \t {{.

Description}}’
jupyter/datascience-notebook Jupyter Notebook Data Science
Stack from htt. . .
jupyter/all-spark-notebook Jupyter Notebook Python,
Scala, R, Spark, Me. . .
jupyterhub/jupyterhub JupyterHub: multi-user Jupyter
notebook serv. . .
jupyter/scipy-notebook Jupyter Notebook Scientific Python
Stack fro. . .
jupyter/tensorflow-notebook Jupyter Notebook Scientific Python
Stack w/ . . .
jupyter/pyspark-notebook Jupyter Notebook Python, Spark,
Mesos Stack . . .
jupyter/minimal-notebook Minimal Jupyter Notebook Stack
from https://. . .
jupyter/base-notebook Small base image for Jupyter Notebook
stacks. . .
jupyterhub/singleuser single-user docker images for use
with Jupyt. . .
jupyter/nbviewer Jupyter Notebook Viewer

Launching Jupyter.
Based on the previous search for Jupyter notebooks, there is a
jupyter/minimal-notebook image. The command will start
the image in a local container and associate the local folder /
Users/esteban/notebooks to the /home/joyvan folder
inside the container. Also note that the container provides specific
URL information to access the notebook (Fig. 23).

$ docker run -p 8888:8888 -v /Users/esteban/notebooks:/home/

jovyan jupyter/minimal-notebook
Executing the command: jupyter notebook
[I 21:10:08.657 NotebookApp] JupyterLab extension loaded from
/opt/conda/lib/python3.7/site-packages/jupyterlab
[I 21:10:08.658 NotebookApp] JupyterLab application directory
is /opt/conda/share/jupyter/lab
[I 21:10:08.661 NotebookApp] Serving notebooks from local
directory: /home/jovyan
[I 21:10:08.662 NotebookApp] The Jupyter Notebook is running
at:
310 W. Stephen Pittard and Shuzhao Li

Fig. 23 Jupyter main page container

[I 21:10:08.662 NotebookApp] http://(aa292bbbba5f or

127.0.0.1):8888/?token=664e4773930180eacef53877390f30a39-
d758595ab0b61dc
[I 21:10:08.663 NotebookApp] Use Control-C to stop this server
and shut down all kernels (twice to skip confirmation).
[C 21:10:08.683 NotebookApp]

To access the notebook, open this file in a browser:

file:///home/jovyan/.local/share/jupyter/runtime/
nbserver-6-open.html
Or copy and paste one of these URLs:
http://(aa292bbbba5f or 127.0.0.1):8888/?toke-
n=664e4773930180eacef53877390f30a39d758595ab0b61dc

http://127.0.0.1:8888/?token=664e4773930180eacef53877390-
f30a39d758595ab0b61dc

5.8 Summary This introduction has provided basic information on how to iden-
tify and run Docker containers which represent a lightweight alter-
native to virtualization techniques involving the use of hypervisor
software. Dockers are particularly useful for containing applications
involving multiple languages or run times as well as those delivered
via the web such as RStudio, Jupyter Notebooks, and Galaxy. The
Docker mechanism neatly packages these tools, and supporting
components, into images on behalf of the end user who does not
need to install or compile code. All that is required is installation of
the Docker package for the local host. Docker also offers the ability
 Essential Toolbox of Data Science 311

to create images using the Dockerfile mechanism as well as the

docker-compose command which enables the deployment of mul-
tiple containers simultaneously.
Because these four tools, Python, R, Git, and Docker, are
widely used in the information technologies, these are very trans-
ferable skills. For the same reason, one can easily find good online
resources and get help. Stack Overflow (http://stackoverflow.com)
is a popular site to search answers for or ask these technical ques-
tions. We also encourage readers to use and contribute to our
supplemental site on GitHub (https://metabolomics-data.github.
io/).

Acknowledgments

This work has been funded in part by the US National Institutes of

Health via grants UH2 AI132345 (Li), U2C ES030163 (Jones, Li,
Morgan, Miller), U01 CA235493 (Li, Xia, Siuzdak), U2C
ES026560 (Miller), P30 ES019776 (Marsit), P50 ES026071
(McCauley), and the US EPA grant 83615301 (McCauley).
 Chapter 16

Predictive Modeling for Metabolomics Data

Tusharkanti Ghosh, Weiming Zhang, Debashis Ghosh,
and Katerina Kechris

Abstract
In recent years, mass spectrometry (MS)-based metabolomics has been extensively applied to characterize
biochemical mechanisms, and study physiological processes and phenotypic changes associated with dis-
ease. Metabolomics has also been important for identifying biomarkers of interest suitable for clinical
diagnosis. For the purpose of predictive modeling, in this chapter, we will review various supervised learning
algorithms such as random forest (RF), support vector machine (SVM), and partial least squares-
discriminant analysis (PLS-DA). In addition, we will also review feature selection methods for identifying
the best combination of metabolites for an accurate predictive model. We conclude with best practices for
reproducibility by including internal and external replication, reporting metrics to assess performance, and
providing guidelines to avoid overfitting and to deal with imbalanced classes. An analysis of an example data
will illustrate the use of different machine learning methods and performance metrics.

Key words Metabolomics, Mass spectrometry, Supervised learning, Performance Metrics, Predictive
Modeling

1 Introduction

In the past 20 years, there has been a dramatic increase in the

development and use of high-throughput technologies for measur-
ing various types of biological activity. Common examples include
transcriptomics (the measurement of gene expression) and proteo-
mics (the measurement of protein levels). The focus of this chapter
is on metabolomics, which involves the measurement of small
compounds, referred to here as metabolites, on a high-throughput
basis. As products of activity at the protein level, metabolites repre-
sent an intermediate level between regulatory processes such as
methylation and transcription, and the full spectrum of physiologi-
cal and disease states. One appealing feature of metabolites is their
ability to be used as clinical biomarkers, and for this reason, meta-
bolomics has been extensively applied for finding biomarkers and

Shuzhao Li (ed.), Computational Methods and Data Analysis for Metabolomics, Methods in Molecular Biology, vol. 2104,
https://doi.org/10.1007/978-1-0716-0239-3_16, © Springer Science+Business Media, LLC, part of Springer Nature 2020

313
314 Tusharkanti Ghosh et al.

studying physiological processes and phenotypic changes associated

with disease [1–4].
Metabolomics experiments fall into two categories: targeted
and untargeted. Targeted metabolomics experiments measure
ions from known biochemically annotated metabolites. By contrast,
untargeted metabolomics experiments measure all possible ions
within a predefined mass range and as a result may also include
ions that do not map to known metabolites [5–8]. The main
objective of metabolomics is to quantify and characterize the
whole spectrum of metabolites. There are a variety of platforms
by which metabolites can be measured. Examples include gas chro-
matography mass spectrometry (GC-MS) and liquid chromatogra-
phy mass spectrometry (LC-MS) [9–11]. At a high level, these
platforms input a sample, fragment it into ions, and separate them
using physical properties in order to generate spectra for the sam-
ple. The fragmented ion spectra are then selected based on their
physical properties (e.g., the retention time and the mass–charge
ratio). In many instances, these properties can be used to map the
ions to known metabolites.
Metabolomics data pose a variety of analytical challenges
[12, 13]; thus, carefully constructed analytical pipelines need to
be developed in order to preprocess and normalize the data. Once
the data are normalized, one can proceed with various downstream
analytical tasks, such as differential expression analysis, clustering,
classification, network discovery, and visualization. In this chapter,
we focus on the particular task of classification, which also goes by
the name of prediction, supervised learning and biomarker discov-
ery. We give an overview on some of the most commonly used
methods for classification, along with an illustrative example using a
dataset from our group. The structure of this chapter is as follows.
In Subheading 2, we provide a short review of missing values and
techniques for missing value imputation in metabolomics data. We
then briefly describe the most commonly used supervised learning
methods (Random Forest, Support Vector Machine, and Partial
Least Square-Discriminant Analysis). In Subheading 3, we lay out a
framework on fitting prediction models and their practical issues.
This is followed by an illustrative example of data analysis and
performance evaluation in Subheading 4, and the chapter ends
with a short discussion in Subheading 5. In this chapter, we inter-
changeably use the term supervised learning, predictive modeling
and machine learning.

2 Methods

2.1 Missing Values Supervised learning methods require complete data; however,
untargeted metabolomic data is prone to missing values, where
the data matrix contains zeros in one or more entries. Some studies
 Predictive Modeling of Metabolomics Data 315

have reported 20–30% missing values in datasets generated using

untargeted MS [14, 15]. It is difficult to deduce whether a missing
value is a genuine absence of a feature, a feature below the lower
limit of detection of the machine, or the failure of the algorithms
employed to identify real signals from the background. In practice,
statisticians have defined three types of mechanisms that lead to
missing values: missing at random (MAR), missing not at random
(MNAR) and missing completely at random (MCAR) [16–
18]. MCAR means that the missingness mechanism is completely
random and depends neither on the observed data nor on the
missing data. Scientifically plausible reasons that are compatible
with missing completely at random include random errors or sto-
chastic fluctuations of peak detection during the acquisition process
of the raw data (incomplete derivations of signals). MAR means
that the probability of a variable being missing is fully accounted for
by other observed variables. Missing not at random (MNAR)
means that the missingness mechanism depends on the unobserved
values. If analysts believe MNAR to hold, there are unfortunately
no ways to assess this assumption using observed data. A practical
strategy is to collect as much covariate information as possible in
order to make the MAR assumption plausible with the
observed data.
There have been several attempts in the literature to deal with
missing values for metabolomics data. For example, fillPeaks [19]
in the XCMS software package has many missing value imputation
tools available. A practical rule of thumb is to impute missing values
by a small value or zero. This is problematic in that this leads to
distortions of the distribution of missing variables and can cause the
standard deviations to be underestimated [20]. Finally, Zhan et al.
developed kernel-based approaches which explicitly modeled the
missingness into a differential expression analysis [21]. Other
imputation strategies include imputing missing values by zero,
half of the minimum value or by the mean or median of observed
values. More advanced methods use, random forest (RF) [20, 22],
singular value decomposition (SVD) [23, 24], and k-nearest neigh-
bors (kNN) [25]. The choice of these methods can influence the
data analyses and inferences [14, 22, 26]. It is therefore extremely
crucial to select the most suitable method for tackling missing
values before moving forward with prediction. Recent work has
compared performance of various missing value imputation meth-
ods [14, 25, 27, 28] on MS metabolomics data [20].

2.2 Classification It is important to note that what we now call supervised learning
Methods: An dates back to over 80 years ago, when Sir R. A. Fisher introduced
Early Look the use of linear discriminant analysis (LDA) [29]. This was a
generative model in which the features conditional on class label
were modeled as a multivariate normal distribution with a mean
vector that depended on group, and a common covariance matrix.
316 Tusharkanti Ghosh et al.

This was generalized to quadratic discriminant analysis, in which

the covariance matrix also depends on the group. Linear discrimi-
nant analysis possessed two desirable properties:
1. Since the multivariate normal distribution is fully specified by
the mean vector and covariance matrix, it is relatively simple to
compute.
2. The classification rule from LDA is linear in the predictors and
thus simple to interpret.
While this methodology is well established, there are two chal-
lenges with modern metabolomics data that make the utility of
LDA less effective. First, in most situations, the number of meta-
bolites being measured is greater than the sample size, which means
that the covariance matrix will not be directly estimable from the
observed data. Second, there is an increasing recognition that the
linear classification rule might be too restrictive and that analysts
should consider other nonlinear classifiers. This will motivate the
classification tools we describe in Subheadings 2.3–2.5.
A second technique that dates back to the 1950s and has been
used extensively in machine learning is Naı̈ve Bayes [30]. In this
framework, we assume the features are conditionally independent
given the group label and model the likelihood ratio of the feature
given the group label. Based on the product of these likelihood
ratios, we are able to assign a new observation to a predicted group.
The term “Naı̈ve” comes from the fact that we assume that the
features are statistically independent when, in fact, we know that
they are not. That said, Naı̈ve Bayes has been shown to be an
effective tool in classification problems [31], and it can handle the
situation when the number of metabolites measured is greater than
the sample size.

2.3 Decision Tree A Decision Tree (DT) is a supervised machine learning model, that
outputs a hierarchical structure to classify subjects [32]. It is a
nonlinear classifier which is mainly used for classifying nonlinearly
separable data. The objective of a decision tree is to develop a model
that predicts the value of a response variable based on several
predictor variables. Figure 1 shows an example of a hypothetical
DT, which divides the data into two categories based on two input
variables. DT used in data mining can be classified into two groups:
l Classification tree: The predicted outcome is a categorical vari-
able, representing two or more classes to which the observation
belongs.
l Regression tree: The predicted value is a continuous variable.
DT is also known as Classification and Regression Trees
(CART), which was first introduced in the machine learning litera-
ture [33]. The main difference between classification and regres-
sion trees is the criteria on which the split-point decision is made.
 Predictive Modeling of Metabolomics Data 317

Fig. 1 A simple decision tree that splits the data into two gender groups based on two metabolites

2.4 Random Forest A Random Forest (RF) is an extremely reliable classifier and robust
to overfitting. It constructs an ensemble of DTs, which means an
aggregation of tree-structured predictors [34]. In RF, each tree is
independently constructed using a bootstrap sample of the original
data (the “bagged sample”). This training data is used to build the
classification model. The data that was not sampled using the
bootstrap is referred to as the out-of-bag sample. Since these data
were not used in model building, they can be used as a test data set,
which can be used to evaluate classification accuracy in an unbiased
manner, by calculating the “out-of-bag error” [35]. A measure of
the variable importance of classification is also computed by con-
sidering the difference between the results from the original and
randomly permuted versions of the data set. Cross-validation is not
needed since RF is estimated from the bootstrap samples.
RF has become popular as a biomarker detection tool in various
metabolomics studies [36, 37]. RF has the strength to deal with
missing and data [34, 38] and overfitting issues [39, 40]. In addi-
tion, it can also tackle high-dimensional data sets without feature
elimination as a requirement [41].

2.5 Support Vector Support Vector Machines (SVM) have been previously used in the
Machines analysis of several omics studies, particularly gene expression data
[42–44]. A simple figure of an SVM is shown in Fig. 2. The main
characteristics that define the concept of SVMs are (a) the criteria
they use to categorize nonlinear relationships (b) the set of training
318 Tusharkanti Ghosh et al.

Fig. 2 A simple graphical representation of SVM

sets that are necessary to optimize the linear classifier; (c) the use of
kernel machines to transform the variable into a higher order
nonlinear space where linear separability holds; (d) utility in terms
of performance and efficiency for high dimensional data sets.
A major drawback of SVM is its restrictions to binary classifica-
tion problems. For example, it can only discriminate between two
classes where the data points are categorized by two classes in
n-dimensional space, where n corresponds to the number of meta-
bolites in our context. A hyperplane is constructed that separates
the data points from the two classes. The hyperplane coefficients are
determined based on the variable (metabolite) importance for dis-
criminating between two classes.
SVM can yield a hyperplane of p-1 dimension in p dimensional
space. The main purpose of SVM is to optimize the largest margin.
In practice, a separation often does not exist as the data points
cannot always be linearly separated. In such nonlinear cases, a
kernel substitution is adopted to map the data to a higher order
dimension. The maximum-margin hyperplane was the original
algorithm developed as a linear classifier [45]. An extension to
create nonlinear classifiers was proposed by applying the kernel
trick to maximum-margin hyperplanes [46]. The advantage of
using the kernel trick is that it can substitute the linear kernel
with other robust kernels, such as the Gaussian kernel [47]. Also
in the family of nonlinear supervised learners are deep neural net-
works (DNN), which construct a nonlinear function from input
variables to outcome variables using a combination of convolution
filters and hidden layers [48].
 Predictive Modeling of Metabolomics Data 319

2.6 Partial Least Partial least squares-discriminant analysis (PLS-DA) is a supervised

Squares-Discriminant technique widely used in metabolomics studies [49–52]. It is
Analysis mainly constructed on the rotation of metabolite abundances in
order to maximize the covariance between the independent vari-
ables (metabolite abundances) and the corresponding response
variable (classes) in high-dimension by finding a linear subspace of
the predictors [53]. PLS-DA is an extension of classical PLS regres-
sion which was implemented for solving linear equations and esti-
mating parameters of interest. PLS-DA method has been
extensively used in various metabolomics studies for disease classi-
fication and biomarker detection [50, 54–56]. Furthermore,
PLS-DA can also be used for dimension reduction, and feature
selection by ranking the loading vectors in decreasing order [52,
57, 58].
Orthogonal PLS (OPLS)-DA was developed as an improve-
ment to PLS-DA in order to discriminate two or more classes of
metabolites using multivariate data [59, 60]. The main advantage
of OPLS-DA over PLS-DA is that a single component is used as a
predictor where the other components constitute the orthogonal
contrasts for analysis of variance, which are independent linear
comparisons between the classes of a component.
Multilevel PLS-DA is another classification technique that can
be used to classify multivariate data from crossover designed studies
[61]. For example, each subject in a controlled experimentation
setup undergo treatment in a random order [62]. Multilevel
PLS-DA can be thought as a multivariate extension of the paired
t-test [61].

3 Practical Issues in Fitting Prediction Models

3.1 Feature Selection Feature Selection (FS) is an important step in successful data
mining procedures [63], such as SVMs [64, 65] and Naı̈ve Bayes
[66], to enhance performance and reduce computational efficiency.
However, FS is not a necessary criterion for some supervised algo-
rithms, such as SVM due to its reliance on regularization, which is
the process of adding information to prevent overfitting in order to
enhance the predictive accuracy and interpretability of the super-
vised learning model. The purpose of feature selection is similar to
model selection [67], which tries to find a compromise between
high predictive accuracy and a model with few predictors. The
insignificant input features in a supervised model may lead to over-
fitting. Hence, it is reasonable to ignore those input features with
negligible or no effect on the output. For example, in the example
later in this chapter, the objective is to infer the relationship
between gender and their corresponding metabolite features.
However, if the sample identifier or any other redundant column
is included as one of the input features, it may cause overfitting. FS
320 Tusharkanti Ghosh et al.

is generally used as a preprocessing tool, in order to reduce the

dimension of a data set by only selecting subsets of features (meta-
bolites), on which a supervised learning is employed. Some well-
known extensions of these FS techniques are Recursive Feature
Elimination, L1 norm SVM [68], and Sequential Minimal Optimi-
zation (SMO) [69].
One of the most commonly used measure in FS is the Variable
Importance Score (VIS), which evaluates features using a model-
based approach [70] by ranking the features according to their
relevance in a classification problem [71]. The main advantage of
using VIS is that incorporates the correlation structure between the
predictors (metabolite features) into the importance calculation.

3.2 Cross-Validation The classification performance of supervised learners is crucial to

determine their predictive power and accuracy. Generally, the vali-
dation procedures are implemented by assuming the model on a
training set and then testing it on an independent set (validation
data set). However, in practical situations, due to the relatively
small number of samples and unavailability of an unbiased indepen-
dent validation data set, cross-validation (CV) can be applied by
splitting a data set into training and test sets. Using k-fold cross
validation [36], the training data set is split into k subsets (folds) of
almost equal size, that is, where k-1 training sets consist of x% of the
data and the remaining (100-x)% data is contained in the kth test
data set. Ideally, x% far exceeds (100-x)%, and x is usually chosen as
90, 80 or 70. Leave-One-Out-CV is a special case of CV, where k is
equal to the total number of data points.

3.3 Metrics There are several potential metrics by which one can evaluate a
for Evaluation prediction model. The most common metric that is used in practice
is the classification accuracy, meaning the proportion of predictions
from the model that are correct based on the gold standard label.
An alternative classification metric is given by the receiver operating
characteristic (ROC) curve. Assume that we have two groups,
disease and control and that higher values of the model correspond
to a greater probability of having disease. We will let the model
output be Y and group label be D, where D ¼ 0 means control and
D ¼ 1 means diseased. One can define the false positive rate based
on a cutoff c by FP(c) ¼ P(Y > c| D ¼ 0). Similarly, the true positive
rate is TP(c) ¼ P(Y > c| D ¼ 1). The true and false positive rates can
then be summarized by the receiver operating characteristic (ROC)
curve, which is a graphical presentation of TP(c),FP(c) for all possi-
ble cutoff values of c. The ROC curve shows the tradeoff between
increasing true positive and false positive rates. Then, the area
under the ROC curve (AUC) can be measured for the curve and
is a summary based on how well the model can distinguish between
two diagnostic groups (diseased/control). Other commonly used
metrics are defined in terms of TP(c) and FP(c) as below:
 Predictive Modeling of Metabolomics Data 321

Sensitivity (SENS): SENS ¼ TP(c).

Specificity (SPEC): SPEC ¼ 1  FP(c).
Precision (PREC): TP(c)/(TP(c) + FP(c).
Recall (REC): REC¼SENS ¼ TP(c).
False Discovery Rate (FDR): FP(c)/(TP(c) + FP(c).
The predicted classes are conventionally computed based on
the cut-off c (¼50%) for the probabilities. However, the cutoff
(threshold) value can be tuned to control the FDR depending on
the problem setting in order to attain maximum predictive
accuracy.
Calibration is another property that has been espoused for risk
prediction models. Well-calibrated models are those in which the
predicted risk matches the observed risk for individuals. The man-
ner in which this is typically assessed is by comparing the risk
predictions from the model to some nonparametric (i.e., non–
model-based) estimate; the closer the predictions are, the better
calibrated the model is. Calibration has been advocated in the risk
prediction [72]. As a matter of course, nonparametric estimates of
risk models require binning of covariates or categorization of pre-
dicted values in order to deal with the inherent sparsity that exists
with using continuous covariates. One method of performing cali-
bration, in the binary outcome setting, is to use the Hosmer–
Lemeshow goodness of fit statistic [73]; smaller values of the
statistic correspond with better calibrated models.
In the calibration setting, what is important is understanding
the distribution of the predicted probabilities, or equivalently, the
risk scores, from the fitted model. Calibration of the model then is
equivalent to modeling the distribution of risk scores; a useful
quantity for accomplishing this is the predictiveness curve [74].

3.4 Imbalanced In numerous data sets, there are unequal numbers of cases in each
Classes class. In this instance, the classifier is biased toward better perfor-
mance of the larger (or majority) class, compared to the smaller
(or minority) class. Often, the research question is much more
focused on performance of discriminating the minority class from
the majority class. But the size of the minority class may be limited
by the difficulty, expense, or time of obtaining the rarer type of
sample. This unequal distribution between classes of a data set is
referred to as the imbalanced class problem [75].
In such cases, the main interest lies in the correct classification
of the “minority class” [76]. Classes with fewer samples or no
sample have a low prior probability and low error cost [77]. The
relation between the distribution of samples in the training set and
costs of misclassification can be controlled by setting a prior proba-
bility at each class.
322 Tusharkanti Ghosh et al.

Several methods have been discussed for tackling imbalanced

data [78, 79], and two techniques which have been extensively
applied in the last decade are resampling and cost-sensitive learning.
In resampling, the approach is to either oversample the minority
class or undersample the majority class. For example, the minority
class can be oversampled by producing duplicates [80] or under
sampling (removing samples) of the majority class [81, 82]. One
major drawback of under sampling is that the majority class may
lose some information, if a large part of majority class in a small
training set is not considered. In cost-sensitive learning, the
approach is to assign a cost misclassification of the minority class
and minimize the overall cost function [83, 84]. Both the resam-
pling and cost-sensitive learning approaches are considered to be
more effective in terms of predictive accuracy than by using equal
class prior constraints [85].

3.5 TRIPOD A recent scientific initiative has focused on developing more repro-
Guidelines ducible approaches to the building, evaluation and validation of
prediction models. A document resulting from this effort that helps
in this goal is the TRIPOD (Transparent reporting of a multivari-
able prediction model for individual prognosis or diagnosis) state-
ment, which provides recommendations for fair reporting of
studies developing, validating, or updating a prediction model. It
consists of a 22-item checklist detailing vital information that must
be incorporated in a prediction model study report [86]. For our
example analysis below, we provide supporting information to
illustrate how our analysis satisfies the 22-item TRIPOD checklist
(Appendix 1).

4 Illustrative Example

4.1 Data For the predictive comparison of classifiers, we obtained a LC-MS

metabolomics dataset from https://www.
metabolomicsworkbench.org (Project ID: PR00038). The data
were generated from subjects enrolled in the Genetic Epidemiology
of Chronic Obstructive Pulmonary Disease Gene study (COPD-
Gene) [87, 88]. Plasma from 131 subjects was collected from the
COPDGene study cohort and analyzed using untargeted LC-MS
(C18+ and HILIC+) metabolomics. The lipid fraction of the
human plasma collected from current and former smokers was
analyzed using Time of Flight (ToF) liquid chromatograph
(LC) (Agilent 6210 Series) and a Quadrupole ToF mass spectrom-
eter (Agilent 6520) which yielded combined data on 2999 metab-
olite features. Data were annotated, normalized and preprocessed
using the methods described in [87, 88].
COPD is an extremely heterogeneous disease comprising mul-
tiple phenotypes. The 131 subjects were either current or former
 Predictive Modeling of Metabolomics Data 323

smokers with various chronic obstructive pulmonary disease

(COPD) phenotypes including airflow obstruction, radiologic
emphysema, and exacerbations. Within this set there were
56 males and 75 females. For additional information about the
cohort, sample collection and data storage data generation,
see [87].

4.2 Training We split the data (131 samples) into 70% (93 samples) training and
and Test Sets 30% (38 samples) test (evaluation) data. For the training data, we
use fivefold CV, where we split the training data (93 training sam-
ples) into 5 different subsets (or fivefolds). We used the first four-
folds to train the data and left the last (fifth) fold as holdout-test
dataset. We then trained the algorithms against each of the folds
and computed (average over fivefolds) the metrics for the training
dataset. The test dataset (n ¼ 38 samples) is used to provide an
unbiased evaluation of the best model fit on the training dataset.
The test dataset can be regarded as an external dataset which
provides the gold standard used to evaluate the models, using
ROC curves and other metrics for evaluation. For model validation,
we predicted the performance of the test data using the trained
models for all the three classifiers.

4.3 Feature Ranking In this section, we implemented different predictive models using
and Variable metabolite abundances as the predictor variables and Gender
Importance (Male/Female) as the response based on the training dataset. We
then computed the Variable Importance Score, which is a measure
of feature relevance to gender for each metabolite (see Subheading
3.1). These scores are nonparametric in nature, and range between
0 and 100. They are subsequently used to rank all the features to
the classification of our response variable, that is, Gender. Metabo-
lites with high values are considered to more relevant features in
classification problem.
In the dataset, the top five metabolites are detected as feature
metabolites out of 2999 metabolites in the training set with fivefold
Cross-Validation for three different classifiers (Fig. 3a–c). Among
them, C39 H79 N7 O + 7.3314843, N-palmitoyl-D-sphingosyl-1-
(2-aminoethyl)phosphonate, and C43 H86 N2 O2 are considered
to be significant metabolite features based on RF and SVM classi-
fiers. However, zeta-Carotene, unannotated metabolite (mass:
2520.6355 and retention time: 1.5409486), 5-Hydroxyisourate
+4.668069, C13 H28 N2 O4, and Tyrosine∗ +2.3151746 are
identified as good predictors based on PLS-DA.

4.4 Model Validation In this section, we evaluated the performance of all the three
classifiers based on the 30% test data of 38 samples using the trained
models. Here, we present ROC curves for all the predictive models
of the testing data used to compute the diagnostic potential of a
classifier in this clinical metabolomics application. From the ROC
324 Tusharkanti Ghosh et al.

Fig. 3 Metabolite relevant feature ranking bar plots (top five metabolites) using Variable Important Scores
ranging from 0 to 100. (a) Random Forest, (b) Support Vector Machine (SVM), and (c) Partial Least Square-
Discriminant Analysis (PLS-DA) for the training dataset

curves, the three methods perform similarly (Fig. 4). Table 1 shows
the performance metrics of the testing data evaluated for all the
classifiers. In this testing dataset, we use AUC as our metric to
choose the best performing classifier. Based on this metric RF has
a small advantage over the other methods (0.87 versus 0.86), but
with other metrics the other methods have a small advantage. In
addition, we also computed the Variable Importance Score on the
test dataset. The top five metabolites for all the three classifiers
using the test dataset were exactly the same selected using the
training dataset with fivefold CV in the previous section.
 Predictive Modeling of Metabolomics Data 325

1.0
0.8
0.6
Sensitivity
0.4
RF
SVM

0.2
PLS-DA

0.0
1.0 0.8 0.6 0.4 0.2 0.0
Specificity

Fig. 4 ROC curves of the testing dataset obtained from three classification
algorithms (RF, SVM, and PLS-DA)

Table 1
Metrics (area under curve (AUC), sensitivity (SENS), specificity (SPEC), precision (PREC), recall (REC))
to evaluate the performance of classification on testing dataset

Metrics/methods AUC SENS SPEC PREC REC

RF 0.87 0.71 0.64 0.71 0.71
SVM 0.86 0.76 0.71 0.76 0.76
PLS-DA 0.86 0.81 0.65 0.74 0.81

5 Summary

Biomarker detection in the field of metabolomics is popular both in

the context of prognostic and diagnostic studies. In this chapter, we
discussed the most commonly used supervised learning algorithms,
feature selection methods, and performance metrics, used in the
downstream analyses of metabolomics studies. In addition, we also
reported predictive accuracy of three classifiers on an example
human plasma LC/MS test dataset to predict gender. Even though
there were advantages of one method compared to the other
depending on the metric, our results cannot be held as a compre-
hensive comparison of these methods, since different classifiers
perform differently depending on the datasets. We encourage
investigators to explore a variety of methods. For more detailed
discussions of biomarker detection and predictive accuracy, see [89–
92]. The R code for this chapter is posted at the supplemental
website, https://metabolomics-data.github.io/. Appendix 2 lists
selected open source tools that implement supervised learning
algorithms.
326 Tusharkanti Ghosh et al.

TRIPOD Checklist for Predictive Modeling for Metabolomics Data

Section/topic Item Checklist item Section

Title and abstract
Title 1 Identify the study as See title
developing
and/or
validating a
multivariable
prediction
model, the target
population, and
the outcome to
be predicted
Abstract 2 Provide a summary See abstract
of objectives,
study design,
setting,
participants,
sample size,
predictors,
outcome,
statistical
analysis, results,
and conclusions
Introduction
Background and objectives 3a Explain the medical Subheading 4.1
context
(including
whether
diagnostic or
prognostic) and
rationale for
developing or
validating the
multivariable
prediction
model, including
references to
existing models
3b Specify the Internal
objectives, validation,
including Subheading 4.4
whether the
study describes
the development
or validation of
the model or
both

(continued)
 Predictive Modeling of Metabolomics Data 327

Section/topic Item Checklist item Section

Methods
Source of data 4a Describe the study Subheading 4.1
design or source
of data (e.g.,
randomized trial,
cohort, or
registry data),
separately for the
development and
validation data
sets, if applicable
4b Specify the key Subheading 4.1,
study dates, see [87]
including start of
accrual; end of
accrual; and, if
applicable, end of
follow-up
Participants 5a Specify key N/A
elements of the
study setting
(e.g., primary
care, secondary
care, general
population)
including
number and
location of
centers
5b Describe eligibility Subheading 4.1,
criteria for see [87]
participants
5c Give details of N/A
treatments
received, if
relevant
Outcome 6a Clearly define the Subheading 4.1
outcome that is
predicted by the
prediction
model, including
how and when
assessed
6b Report any actions N/A
to blind
assessment of the
outcome to be
predicted

(continued)
328 Tusharkanti Ghosh et al.

Section/topic Item Checklist item Section

Predictors 7a Clearly define all 2999 predictors,
predictors used in for more details
developing or see [87]
validating the
multivariable
prediction
model, including
how and when
they were
measured
7b Report any actions N/A
to blind
assessment of
predictors for the
outcome and
other predictors
Sample size 8 Explain how the Subheading 4.1,
study size was see [87]
arrived at
Missing data 9 Describe how The data was
missing data were already
handled (e.g., preprocessed
complete-case and imputed,
analysis, single see Subheading
imputation, 4.1
multiple
imputation) with
details of any
imputation
method
Statistical analysis methods 10cFor validation, Subheading 3.3
describe how the
predictions were
calculated
10d Specify all measures Subheading 3.3
used to assess
model
performance and,
if relevant, to
compare multiple
models
10e Describe any model N/A
updating (e.g.,
recalibration)
arising from the
validation, if
done

(continued)
 Predictive Modeling of Metabolomics Data 329

Section/topic Item Checklist item Section

Risk groups 11 Provide details on N/A
how risk groups
were created, if
done
Development vs. validation 12 For validation, N/A
identify any
differences from
the development
data in setting,
eligibility criteria,
outcome, and
predictors
Results
Participants 13a Describe the flow of Subheading 4.1,
participants see [87]
through the
study, including
the number of
participants with
and without the
outcome and, if
applicable, a
summary of the
follow-up time. A
diagram may be
helpful
13b Describe the Subheading 4.1,
characteristics of see [87]
the participants
(basic
demographics,
clinical features,
available
predictors),
including the
number of
participants with
missing data for
predictors and
outcome
13c For validation, Subheadings 4.3
show a and 4.4
comparison with
the development
data of the
distribution of
important
variables
(demographics,
predictors, and
outcome)

(continued)
330 Tusharkanti Ghosh et al.

Section/topic Item Checklist item Section

Model performance 16 Report N/A
performance
measures (with
CIs) for the
prediction model
Model-updating 17 If done, report the Subheading 4.4
results from any
model updating
(i.e., model
specification,
model
performance)
Discussion
Limitations 18 Discuss any Subheading 4.1,
limitations of the see [87]
study (such as
nonreprese
ntative sample,
few events per
predictor,
missing data)
Interpretation 19a For validation, N/A
discuss the results
with reference to
performance in
the development
data, and any
other validation
data
19b Give an overall Subheadings 4.4
interpretation of and 5
the results,
considering
objectives,
limitations,
results from
similar studies,
and other
relevant evidence
Implications 20 Discuss the Subheadings 4.4
potential clinical and 5.
use of the model However,
and implications performance of
for future the model is
research data-driven
Other information

(continued)
 Predictive Modeling of Metabolomics Data 331

Section/topic Item Checklist item Section

Supplementary 21 Provide Subheading 4.1,
information information see [87]
about the
availability of
supplementary
resources, such as
study protocol,
web calculator,
and data sets
Funding 22 Give the source of NIH
funding and the
role of the
funders for the
present study

Selected Open Source (R/Bioconductor/Web-Based) Tools for Supervised Learning

Algorithms

Method Source Reference

PLS-DA Bioconductor (ropls) [93]
PLS-DA, RF, and SVM Bioconductor (biosigner) [94]
SVM, RF Bioconductor (MLSeq) [95]
RF, SVM, PLS-DA Metaboanalyst [96]
http://www.
metaboanalyst.ca/
PCA, PLS-DA, RF Bioconductor (statTarget) [97]
Feature selection, metric evaluation Bioconductor [98]
(OmicsMarker)
Sparse PLS-DA Bioconductor [99]
(mixOmics)
Feature selection, metric evaluation CRAN (lilikoi) [100]
Probabilistic principal component CRAN (MetabolAnalyze) [101]
analysis
Kernel-based metabolite CRAN (KMDA) [21]
differential analysis
PLS-DA, OPLS-DA CRAN (muma) [102]
RF CRAN [103]
(RFmarkerDetector)
RF, SVM, PLS-DA CRAN (caret) [104]
332 Tusharkanti Ghosh et al.

References
1. Maniscalco M, Fuschillo S, Paris D, nitrogen supplementation in oats. Metabolo-
Cutignano A, Sanduzzi A, Motta A (2019) mics 15(3):42. https://doi.org/10.1007/
Clinical metabolomics of exhaled breath con- s11306-019-1501-x
densate in chronic respiratory diseases. Adv 10. Fang J, Zhao H, Zhang Y, Wong M, He Y,
Clin Chem 88:121–149. https://doi.org/ Sun Q, Xu S, Cai Z (2019) Evaluation of gas
10.1016/bs.acc.2018.10.002 chromatography-atmospheric pressure chem-
2. Pujos-Guillot E, Petera M, Jacquemin J, ical ionization tandem mass spectrometry as
Centeno D, Lyan B, Montoliu I, Madej D, an alternative to gas chromatography tandem
Pietruszka B, Fabbri C, Santoro A, mass spectrometry for the determination of
Brzozowska A, Franceschi C, Comte B polychlorinated biphenyls and polybromi-
(2018) Identification of pre-frailty sub-phe- nated diphenyl ethers. Chemosphere
notypes in elderly using metabolomics. Front 225:288–294. https://doi.org/10.1016/j.
Physiol 9:1903. https://doi.org/10.3389/ chemosphere.2019.03.011
fphys.2018.01903 11. Lohr KE, Camp EF, Kuzhiumparambil U,
3. Sarode GV, Kim K, Kieffer DA, Shibata NM, Lutz A, Leggat W, Patterson JT, Suggett DJ
Litwin T, Czlonkowska A, Medici V (2019) (2019) Resolving coral photoacclimation
Metabolomics profiles of patients with Wilson dynamics through coupled photophysiologi-
disease reveal a distinct metabolic signature. cal and metabolomic profiling. J Exp Biol
Metabolomics 15(3):43. https://doi.org/10. 222:jeb195982. https://doi.org/10.1242/
1007/s11306-019-1505-6 jeb.195982
4. Wang X, Zhang A, Sun H (2013) Power of 12. Baumeister TUH, Ueberschaar N, Schmidt-
metabolomics in diagnosis and biomarker dis- Heck W, Mohr JF, Deicke M, Wichard T,
covery of hepatocellular carcinoma. Hepatol- Guthke R, Pohnert G (2018) DeltaMS: a
ogy 57(5):2072–2077 tool to track isotopologues in GC- and
5. Caesar LK, Kellogg JJ, Kvalheim OM, Cech LC-MS data. Metabolomics 14(4):41.
NB (2019) Opportunities and limitations for https://doi.org/10.1007/s11306-018-
untargeted mass spectrometry metabolomics 1336-x
to identify biologically active constituents in 13. Gilmore IS, Heiles S, Pieterse CL (2019)
complex natural product mixtures. J Nat Prod Metabolic imaging at the single-cell scale:
82:469. https://doi.org/10.1021/acs. recent advances in mass spectrometry imag-
jnatprod.9b00176 ing. Annu Rev Anal Chem (Palo Alto Calif)
6. Liu LL, Lin Y, Chen W, Tong ML, Luo X, Lin 12:201. https://doi.org/10.1146/annurev-
LR, Zhang HL, Yan JH, Niu JJ, Yang TC anchem-061318-115516
(2019) Metabolite profiles of the cerebrospi- 14. Do KT, Wahl S, Raffler J, Molnos S,
nal fluid in neurosyphilis patients determined Laimighofer M, Adamski J, Suhre K,
by untargeted metabolomics analysis. Front Strauch K, Peters A, Gieger C,
Neurosci 13:150. https://doi.org/10.3389/ Langenberg C, Stewart ID, Theis FJ,
fnins.2019.00150 Grallert H, Kastenmuller G, Krumsiek J
7. Sanchez-Arcos C, Kai M, Svatos A, (2018) Characterization of missing values in
Gershenzon J, Kunert G (2019) Untargeted untargeted MS-based metabolomics data and
metabolomics approach reveals differences in evaluation of missing data handling strategies.
host plant chemistry before and after infesta- Metabolomics 14(10):128. https://doi.org/
tion with different pea aphid host races. Front 10.1007/s11306-018-1420-2
Plant Sci 10:188. https://doi.org/10.3389/ 15. Liggi S, Hinz C, Hall Z, Santoru ML,
fpls.2019.00188 Poddighe S, Fjeldsted J, Atzori L, Griffin JL
8. Wang R, Yin Y, Zhu ZJ (2019) Advancing (2018) KniMet: a pipeline for the processing
untargeted metabolomics using data- of chromatography-mass spectrometry meta-
independent acquisition mass spectrometry bolomics data. Metabolomics 14(4):52.
technology. Anal Bioanal Chem 411:4349. https://doi.org/10.1007/s11306-018-
https://doi.org/10.1007/s00216-019- 1349-5
01709-1 16. Fielding S, Fayers PM, McDonald A,
9. Allwood JW, Xu Y, Martinez-Martin P, McPherson G, Campbell MK (2008) Simple
Palau R, Cowan A, Goodacre R, Marshall A, imputation methods were inadequate for
Stewart D, Howarth C (2019) Rapid missing not at random (MNAR) quality of
UHPLC-MS metabolite profiling and pheno- life data. Health Qual Life Outcomes 6(1):57
typic assays reveal genotypic impacts of
 Predictive Modeling of Metabolomics Data 333

17. Schafer JL, Graham JW (2002) Missing data: 27. Chen MX, Wang SY, Kuo CH, Tsai IL (2019)
our view of the state of the art. Psychol Meth- Metabolome analysis for investigating host-
ods 7(2):147 gut microbiota interactions. J Formos Med
18. Steyerberg EW, van Veen M (2007) Imputa- Assoc 118(Suppl 1):S10–S22. https://doi.
tion is beneficial for handling missing data in org/10.1016/j.jfma.2018.09.007
predictive models. J Clin Epidemiol 60 28. Shen X, Zhu ZJ (2019) MetFlow: an interac-
(9):979 tive and integrated workflow for metabolo-
19. Smith CA, Want EJ, O’Maille G, Abagyan R, mics data cleaning and differential metabolite
Siuzdak G (2006) XCMS: processing mass discovery. Bioinformatics 35:2870. https://
spectrometry data for metabolite profiling doi.org/10.1093/bioinformatics/bty1066
using nonlinear peak alignment, matching, 29. McLachlan, Geoffrey J (2004) Discriminant
and identification. Anal Chem 78 analysis and statistical pattern recognition.
(3):779–787. https://doi.org/10.1021/ Wiley-Interscience, Hoboken, N.J. John
ac051437y Wiley & Sons. & Wiley InterScience (Online
20. Wei R, Wang J, Su M, Jia E, Chen S, Chen T, Service)
Ni Y (2018) Missing value imputation 30. McCallum A, Nigam K (1998) A comparison
approach for mass spectrometry-based meta- of event models for naive Bayes text classifica-
bolomics data. Sci Rep 8(1):663. https://doi. tion. In: AAAI-98 workshop on learning for
org/10.1038/s41598-017-19120-0 text categorization, vol 1. Citeseer, pp 41–48
21. Zhan X, Patterson AD, Ghosh D (2015) Ker- 31. Wang Q, Garrity GM, Tiedje JM, Cole JR
nel approaches for differential expression anal- (2007) Naive Bayesian classifier for rapid
ysis of mass spectrometry-based assignment of rRNA sequences into the new
metabolomics data. BMC Bioinformatics bacterial taxonomy. Appl Environ Microbiol
16:77. https://doi.org/10.1186/s12859- 73(16):5261–5267
015-0506-3 32. Quinlan JR (1986) Induction of decision
22. Gromski PS, Xu Y, Kotze HL, Correa E, Ellis trees. Mach Learn 1(1):81–106
DI, Armitage EG, Turner ML, Goodacre R 33. Breiman L (2017) Classification and regres-
(2014) Influence of missing values substitutes sion trees. Routledge, Boca Raton
on multivariate analysis of metabolomics data. 34. Liaw A, Wiener M (2002) Classification and
Metabolites 4(2):433–452. https://doi.org/ regression by randomForest. R News 2
10.3390/metabo4020433 (3):18–22
23. Kumar N, Hoque MA, Shahjaman M, Islam 35. Gislason PO, Benediktsson JA, Sveinsson JR
SM, Mollah MN (2017) Metabolomic bio- (2006) Random forests for land cover classifi-
marker identification in presence of outliers cation. Pattern Recogn Lett 27(4):294–300
and missing values. Biomed Res Int
2017:2437608. https://doi.org/10.1155/ 36. Chen T, Cao Y, Zhang Y, Liu J, Bao Y,
2017/2437608 Wang C, Jia W, Zhao A (2013) Random forest
in clinical metabolomics for phenotypic dis-
24. Sun X, Langer B, Weckwerth W (2015) Chal- crimination and biomarker selection. Evid
lenges of inversely estimating Jacobian from Based Complement Alternat Med
metabolomics data. Front Bioeng Biotechnol 2013:298183
3:188. https://doi.org/10.3389/fbioe.
2015.00188 37. Scott I, Lin W, Liakata M, Wood J, Vermeer
CP, Allaway D, Ward J, Draper J, Beale M,
25. Lee JY, Styczynski MP (2018) NS-kNN: a Corol D (2013) Merits of random forests
modified k-nearest neighbors approach for emerge in evaluation of chemometric classi-
imputing metabolomics data. Metabolomics fiers by external validation. Anal Chim Acta
14(12):153. https://doi.org/10.1007/ 801:22–33
s11306-018-1451-8
38. Ho TK (1998) Nearest neighbors in random
26. Di Guida R, Engel J, Allwood JW, Weber subspaces. In: Joint IAPR International
RJM, Jones MR, Sommer U, Viant MR, Workshops on Statistical Techniques in Pat-
Dunn WB (2016) Non-targeted UHPLC- tern Recognition (SPR) and Structural and
MS metabolomic data processing methods: a Syntactic Pattern Recognition (SSPR).
comparative investigation of normalisation, Springer, pp 640–648
missing value imputation, transformation
and scaling. Metabolomics 12(5):93. 39. Biau G (2012) Analysis of a random forests
https://doi.org/10.1007/s11306-016- model. J Mach Learn Res 13
1030-9 (Apr):1063–1095
40. Hapfelmeier A, Hothorn T, Ulm K, Strobl C
(2014) A new variable importance measure
334 Tusharkanti Ghosh et al.

for random forests with missing data. Stat 51. Quiros-Guerrero L, Albertazzi F, Araya-
Comput 24(1):21–34 Valverde E, Romero RM, Villalobos H,
41. Menze BH, Kelm BM, Masuch R, Poveda L, Chavarria M, Tamayo-Castillo G
Himmelreich U, Bachert P, Petrich W, Ham- (2019) Phenolic variation among Chamae-
precht FA (2009) A comparison of random crista nictitans subspecies and varieties
forest and its Gini importance with standard revealed through UPLC-ESI()-MS/MS
chemometric methods for the feature selec- chemical fingerprinting. Metabolomics 15
tion and classification of spectral data. BMC (2):14. https://doi.org/10.1007/s11306-
Bioinformatics 10(1):213 019-1475-8
42. Maker AV, Hu V, Kadkol SS, Hong L, 52. Wang J, Yan D, Zhao A, Hou X, Zheng X,
Brugge W, Winter J, Yeo CJ, Hackert T, Chen P, Bao Y, Jia W, Hu C, Zhang ZL, Jia W
Buchler M, Lawlor RT, Salvia R, Scarpa A, (2019) Discovery of potential biomarkers for
Bassi C, Green S (2019) Cyst fluid biosigna- osteoporosis using LC-MS/MS metabolomic
ture to predict Intraductal papillary mucinous methods. Osteoporos Int 30:1491. https://
neoplasms of the pancreas with high malig- doi.org/10.1007/s00198-019-04892-0
nant potential. J Am Coll Surg 228:721. 53. Grissa D, Petera M, Brandolini M, Napoli A,
https://doi.org/10.1016/j.jamcollsurg. Comte B, Pujos-Guillot E (2016) Feature
2019.02.040 selection methods for early predictive bio-
43. Tkachev V, Sorokin M, Mescheryakov A, marker discovery using untargeted metabolo-
Simonov A, Garazha A, Buzdin A, mic data. Front Mol Biosci 3:30. https://doi.
Muchnik I, Borisov N (2018) FLOating- org/10.3389/fmolb.2016.00030
window projective separator (FloWPS): a 54. Bayci AWL, Baker DA, Somerset AE,
data trimming tool for support vector Turkoglu O, Hothem Z, Callahan RE,
machines (SVM) to improve robustness of Mandal R, Han B, Bjorndahl T, Wishart D,
the classifier. Front Genet 9:717. https:// Bahado-Singh R, Graham SF, Keidan R
doi.org/10.3389/fgene.2018.00717 (2018) Metabolomic identification of diag-
44. Yerukala Sathipati S, Ho SY (2018) Identify- nostic serum-based biomarkers for advanced
ing a miRNA signature for predicting the stage melanoma. Metabolomics 14(8):105.
stage of breast cancer. Sci Rep 8(1):16138. https://doi.org/10.1007/s11306-018-
https://doi.org/10.1038/s41598-018- 1398-9
34604-3 55. Catav SS, Elgin ES, Dag C, Stark JL, Kucuka-
45. Cortes C, Vapnik V (1995) Support-vector kyuz K (2018) NMR-based metabolomics
networks. Mach Learn 20(3):273–297 reveals that plant-derived smoke stimulates
46. Boser BE, Guyon IM, Vapnik VN (1992) A root growth via affecting carbohydrate and
training algorithm for optimal margin classi- energy metabolism in maize. Metabolomics
fiers. In: Proceedings of the fifth annual work- 14(11):143. https://doi.org/10.1007/
shop on computational learning theory. s11306-018-1440-y
ACM, pp 144–152 56. Guo JG, Guo XM, Wang XR, Tian JZ, Bi HS
47. Bishop CM (2006) Pattern recognition and (2019) Metabolic profile analysis of free
machine learning. Springer, Berlin amino acids in experimental autoimmune
uveoretinitis rat plasma. Int J Ophthalmol 12
48. Ripley BD (1994) Flexible non-linear (1):16–24. https://doi.org/10.18240/ijo.
approaches to classification. In: From statistics 2019.01.03
to neural networks. Springer, Berlin, pp
105–126 57. Rodrigues-Neto JC, Correia MV, Souto AL,
Ribeiro JAA, Vieira LR, Souza MT Jr, Rodri-
49. Contreras-Jodar A, Nayan NH, Hamzaoui S, gues CM, Abdelnur PV (2018) Metabolic fin-
Caja G, Salama AAK (2019) Heat stress modi- gerprinting analysis of oil palm reveals a set of
fies the lactational performances and the uri- differentially expressed metabolites in fatal
nary metabolomic profile related to yellowing symptomatic and
gastrointestinal microbiota of dairy goats. non-symptomatic plants. Metabolomics 14
PLoS One 14(2):e0202457. https://doi. (10):142. https://doi.org/10.1007/
org/10.1371/journal.pone.0202457 s11306-018-1436-7
50. Park HG, Jang KS, Park HM, Song WS, 58. Wong M, Lodge JK (2012) A metabolomic
Jeong YY, Ahn DH, Kim SM, Yang YH, Kim investigation of the effects of vitamin E sup-
YG (2019) MALDI-TOF MS-based total plementation in humans. Nutr Metab (Lond)
serum protein fingerprinting for liver cancer 9(1):110. https://doi.org/10.1186/1743-
diagnosis. Analyst 144:2231. https://doi. 7075-9-110
org/10.1039/c8an02241k
 Predictive Modeling of Metabolomics Data 335

59. Li Y, Chen M, Liu C, Xia Y, Xu B, Hu Y, values. IEEE Geosci Remote Sens Lett 14
Chen T, Shen M, Tang W (2018) Metabolic (11):1988–1992
changes associated with papillary thyroid car- 72. Van Calster B, Vickers AJ (2015) Calibration
cinoma: a nuclear magnetic resonance-based of risk prediction models: impact on decision-
metabolomics study. Int J Mol Med 41 analytic performance. Med Decis Making 35
(5):3006–3014. https://doi.org/10.3892/ (2):162–169
ijmm.2018.3494 73. Agresti A (2002) Categorical data analysis.
60. Rezig L, Servadio A, Torregrossa L, Miccoli P, Wiley, New York
Basolo F, Shintu L, Caldarelli S (2018) Diag- 74. Huang Y, Sullivan Pepe M, Feng Z (2007)
nosis of post-surgical fine-needle aspiration Evaluating the predictiveness of a continuous
biopsies of thyroid lesions with indeterminate marker. Biometrics 63(4):1181–1188
cytology using HRMAS NMR-based metabo-
lomics. Metabolomics 14(10):141. https:// 75. Holder LB, Haque MM, Skinner MK (2017)
doi.org/10.1007/s11306-018-1437-6 Machine learning for epigenetics and future
medical applications. Epigenetics 12
61. Westerhuis JA, van Velzen EJ, Hoefsloot HC, (7):505–514. https://doi.org/10.1080/
Smilde AK (2010) Multivariate paired data 15592294.2017.1329068
analysis: multilevel PLSDA versus OPLSDA.
Metabolomics 6(1):119–128 76. Chen C, Liaw A, Breiman L (2004) Using
random forest to learn imbalanced data, vol
62. Liquet B, Le Cao KA, Hocini H, Thiebaut R 110. University of California, Berkeley, pp
(2012) A novel approach for biomarker selec- 1–12
tion and the integration of repeated measures
experiments from two assays. BMC Bioinfor- 77. Breiman L, Friedman J, Olshen RA, Stone CJ
matics 13:325. https://doi.org/10.1186/ (1984) Classification and regression trees.
1471-2105-13-325 Chapman & Hall, New York
63. Liu H, Motoda H (1998) Feature extraction, 78. Japkowicz N (2000) Learning from imbal-
construction and selection: a data mining per- anced data sets: a comparison of various stra-
spective, vol 453. Springer Science & Business tegies. In: AAAI workshop on learning from
Media, Norwell imbalanced data sets. Menlo Park, CA, pp
10–15
64. Guyon I, Elisseeff A (2003) An introduction
to variable and feature selection. J Mach Learn 79. Maloof MA (2003) Learning when data sets
Res 3:1157–1182 are imbalanced and when costs are unequal
and unknown. In: ICML-2003 workshop on
65. Weston J, Elisseeff A, Schölkopf B, Tipping M learning from imbalanced data sets II, pp 2–1
(2003) Use of the zero-norm with linear
models and kernel methods. J Mach Learn 80. Ling CX, Li C (1998) Data mining for direct
Res 3(Mar):1439–1461 marketing: problems and solutions. In: KDD
1998, pp 73–79
66. Mladenic D, Grobelnik M (1999) Feature
selection for unbalanced class distribution 81. Chawla NV, Bowyer KW, Hall LO, Kegel-
and naive bayes. In: ICML 1999, pp 258–267 meyer WP (2002) SMOTE: synthetic minor-
ity over-sampling technique. J Artif Intell Res
67. Bozdogan H (1987) Model selection and 16:321–357
Akaike’s information criterion (AIC): the
general theory and its analytical extensions. 82. Kubat M, Matwin S (1997) Addressing the
Psychometrika 52(3):345–370 curse of imbalanced training sets: one-sided
selection. In: ICML 1997. Citeseer, pp
68. Guan W, Zhou M, Hampton CY, Benigno 179–186
BB, Walker LD, Gray A, McDonald JF, Fer-
nández FM (2009) Ovarian cancer detection 83. Domingos P (1999) Metacost: a general
from metabolomic liquid chromatography/ method for making classifiers cost-sensitive.
mass spectrometry data by support vector In: KDD 1999, pp 155–164
machines. BMC Bioinformatics 10(1):259 84. Cateni S, Colla V, Vannucci M (2014) A
69. Platt J (1998) Sequential minimal optimiza- method for resampling imbalanced datasets
tion: a fast algorithm for training support vec- in binary classification tasks for real-world
tor machines problems. Neurocomputing 135:32–41
70. Kuhn M, Johnson K (2013) Applied predic- 85. Drummond C, Holte RC (2003) C4. 5, class
tive modeling, vol 26. Springer, New York imbalance, and cost sensitivity: why under-
sampling beats over-sampling. In: Workshop
71. Behnamian A, Millard K, Banks SN, White L, on learning from imbalanced datasets
Richardson M, Pasher J (2017) A systematic II. Citeseer, pp 1–8
approach for variable selection with random
forests: achieving stable variable importance
336 Tusharkanti Ghosh et al.

86. Collins GS, Reitsma JB, Altman DG, Moons 93. Thévenot EA (2016) ropls: PCA, PLS (-DA)
KG (2015) Transparent reporting of a multi- and OPLS (-DA) for multivariate analysis and
variable prediction model for individual prog- feature selection of omics data
nosis or diagnosis (TRIPOD): the TRIPOD 94. Rinaudo P, Boudah S, Junot C, Thévenot EA
statement. BMC Med 13(1):1 (2016) Biosigner: a new method for the dis-
87. Cruickshank-Quinn CI, Jacobson S, covery of significant molecular signatures
Hughes G, Powell RL, Petrache I, from omics data. Front Mol Biosci 3:26
Kechris K, Bowler R, Reisdorph N (2018) 95. Zararsiz G, Goksuluk D, Korkmaz S,
Metabolomics and transcriptomics pathway Eldem V, Duru IP, Unver T, Ozturk A, Zar-
approach reveals outcome-specific perturba- arsiz MG, klaR M, biocViews Sequencing, R
tions in COPD. Sci Rep 8(1):17132 (2014) Package ‘MLSeq’
88. Regan EA, Hokanson JE, Murphy JR, 96. Xia J, Psychogios N, Young N, Wishart DS
Make B, Lynch DA, Beaty TH, Curran- (2009) MetaboAnalyst: a web server for meta-
Everett D, Silverman EK, Crapo JD (2010) bolomic data analysis and interpretation.
Genetic epidemiology of COPD (COPD- Nucleic Acids Res 37(suppl_2):W652–W660
Gene) study design. COPD 7(1):32–43. 97. Luan H, Ji F, Chen Y, Cai Z (2018) statTar-
https://doi.org/10.3109/ get: a streamlined tool for signal drift correc-
15412550903499522 tion and interpretations of quantitative mass
89. Andersen SL, Briggs FBS, Winnike JH, spectrometry-based omics data. Anal Chim
Natanzon Y, Maichle S, Knagge KJ, Newby Acta 1036:66–72
LK, Gregory SG (2019) Metabolome-based 98. Determan Jr CE, Determan Jr MCE (2015)
signature of disease pathology in MS. Mult Package ‘OmicsMarkeR’
Scler Relat Disord 31:12–21. https://doi.
org/10.1016/j.msard.2019.03.006 99. Rohart F, Gautier B, Singh A, Le Cao K-A
(2017) mixOmics: an R package for ‘omics
90. Lee HS, Seo C, Hwang YH, Shin TH, Park feature selection and multiple data integra-
HJ, Kim Y, Ji M, Min J, Choi S, Kim H, Park tion. PLoS Comput Biol 13(11):e1005752
AK, Yee ST, Lee G, Paik MJ (2019) Metabo-
lomic approaches to polyamines including 100. Al-Akwaa FM, Yunits B, Huang S, Alhajaji H,
acetylated derivatives in lung tissue of mice Garmire LX (2018) Lilikoi: an R package for
with asthma. Metabolomics 15(1):8. personalized pathway-based classification
https://doi.org/10.1007/s11306-018- modeling using metabolomics data. Giga-
1470-5 Science 7(12):giy136
91. Long NP, Yoon SJ, Anh NH, Nghi TD, Lim 101. Gift N, Gormley IC, Brennan L, Gormley
DK, Hong YJ, Hong SS, Kwon SW (2018) A MC (2010) Package ‘MetabolAnalyze’
systematic review on metabolomics-based 102. Gaude E, Chignola F, Spiliotopoulos D,
diagnostic biomarker discovery and validation Spitaleri A, Ghitti M, Garcı̀a-Manteiga JM,
in pancreatic cancer. Metabolomics 14 Mari S, Musco G (2013) Muma, an R package
(8):109. https://doi.org/10.1007/s11306- for metabolomics univariate and multivariate
018-1404-2 statistical analysis. Curr Metabol 1
92. Regan EA, Hersh CP, Castaldi PJ, DeMeo (2):180–189
DL, Silverman EK, Crapo JD, Bowler RP 103. Palla P (2015) Information management and
(2019) Omics and the search for blood bio- multivariate analysis techniques for metabolo-
markers in COPD: insights from COPD- mics data. Universita’degli Studi di Cagliari
Gene. Am J Respir Cell Mol Biol 61:143. 104. Kuhn M (2008) Building predictive models in
https://doi.org/10.1165/rcmb.2018- R using the caret package. J Stat Softw 28
0245PS (5):1–26
 Chapter 17

Using MetaboAnalyst 4.0 for Metabolomics Data Analysis,

Interpretation, and Integration with Other Omics Data
Jasmine Chong and Jianguo Xia

Abstract
MetaboAnalyst (www.metaboanalyst.ca) is an easy-to-use, comprehensive web-based tool, freely available
for metabolomics data processing, statistical analysis, functional interpretation, as well as integration with
other omics data. This chapter first provides an introductory overview to the current MetaboAnalyst
(version 4.0) with regards to its underlying design concepts and user interface structure. Subsequent
sections describe three common metabolomics data analysis workflows covering targeted metabolomics,
untargeted metabolomics, and multi-omics data integration.

Key words Web server, Multivariate statistics, Enrichment analysis, Metabolic pathway analysis,
Multi-omics integration

1 Introduction and Exploring MetaboAnalyst 4.0

Rapid advances in analytical chemistry, mass spectrometry (MS),

and nuclear magnetic resonance (NMR) spectroscopy have dramat-
ically increased the size and speed at which metabolomics data can
be obtained. However, the inability of researchers to extract mean-
ingful biological insights from these increasingly large and complex
datasets has now become a major barrier in current metabolomics
research and applications. There is an urgent demand for user-
friendly and easily accessible bioinformatic tools to bridge the gap
between data generation and biological insights.
MetaboAnalyst is an easy-to-use web-based tool suite that
permits users to perform a wide variety of metabolomics data
analysis tasks. It was first released in 2009 with a single module
for metabolomics data processing and statistical analysis [1] and has
since been continuously updated to meet the ever-evolving needs
for metabolomics data analysis [2–4]. The current release (v4.0)
contains 12 modules which can be placed into four general cate-
gories: (1) exploratory statistical analysis, (2) functional analysis,
(3) data integration and systems biology, and (4) data processing

Shuzhao Li (ed.), Computational Methods and Data Analysis for Metabolomics, Methods in Molecular Biology, vol. 2104,
https://doi.org/10.1007/978-1-0716-0239-3_17, © Springer Science+Business Media, LLC, part of Springer Nature 2020

337
338 Jasmine Chong and Jianguo Xia

and utility functions. The MetaboAnalyst web server has been

carefully designed so that each module workflow follows an intui-
tive yet consistent process beginning with data uploading, followed
by data processing and analysis. MetaboAnalyst supports the analy-
sis of various common data types obtained from both targeted and
untargeted metabolomics. Its companion R package, MetaboAna-
lystR, was made available in 2018 to complement the web server
with improved flexibility, transparency, and scalability for advanced
and high-throughput metabolomics data analysis [5]. The package
was recently upgraded to version 2.0 to support raw spectral pre-
processing and functional interpretation for LC-MS based global
metabolomics [6].
The objective of this chapter is to introduce the reader to three
typical metabolomics data analysis workflows using MetaboAnalyst
4.0. The chapter is divided into four sections. Section 1 focuses on
introducing the overall layout and design concepts of MetaboAna-
lyst. Section 2 provides an example of a targeted metabolomics data
workflow, including commonly used statistical analyses as well as
pathway and enrichment analysis. Section 3 goes through an untar-
geted metabolomics workflow from raw MS peaks to functional
interpretation. Finally, Section 4 describes an example of using
MetaboAnalyst for integrative analysis of metabolomics and
transcriptomics data.
This section will guide users through MetaboAnalyst, specifi-
cally how to understand its modular workflow, navigation tree,
analysis report generation, and R Command History.

1.1 Understanding 1. To begin, open a preferred web browser and enter the Meta-
the MetaboAnalyst boAnalyst web address: https://www.metaboanalyst.ca/.
Layout 2. At the top of the home page, click the “>>click here to
start<<” hyperlink to enter the “Module Overview” page.
3. In the center of the “Module Overview” page is a circular
“clock” showing 12 different modules currently available in
MetaboAnalyst 4.0 (Fig. 1). All modules follow the same gen-
eral flow: data upload, data processing and data analysis. To
briefly demonstrate this, select the “Statistical Analysis” button
on the top left of the circular panel.
4. The Statistical Analysis data upload page should now be visible
(Fig. 2). On the left-hand side is the navigation tree, which
guides users through all analysis steps of the selected module.
This navigation tree is specific for each module. Note the
“Upload” hyperlink is highlighted in blue, representing the
current step. On the right-hand side is the R Command His-
tory panel, which displays all underlying R commands as they
occur in real-time. The aim of this panel is to improve the
reproducibility and transparency of the MetaboAnalyst web
server. These R commands can be used directly by the compan-
ion MetaboAnalystR package to reproduce one’s results [5, 6].
 Metabolomics Data Analysis Using MetaboAnalyst 4.0 339

Fig. 1 A screenshot of the MetaboAnalyst module overview

Fig. 2 A screenshot of the Statistical Analysis data upload page within MetaboAnalyst
340 Jasmine Chong and Jianguo Xia

Fig. 3 A screenshot of the heatmap generated in the Statistical Analysis module within MetaboAnalyst. To the
left of the image is the navigation tree and to the right is the R Command History panel

5. Select any example dataset listed at the bottom of the page and
click “Submit.” For the purposes of this section, we will speed
through the next few pages. A detailed guide to data processing
and statistical analysis is provided below in Subheading 2.
6. The next page is the Data Integrity Check, click “Skip” on the
bottom of the page. Note that the R commands are updated on
the right upon each user action. After is the Normalization
page, keep all options set to “None”, press “Normalize” (bot-
tom left) and then “Proceed” (bottom right). Following this is
the “Analysis View” page, which shows all the analysis options
available in MetaboAnalyst. Click “Heatmaps” under the
“Cluster Analysis” subheading. A heatmap should now be
visible in the center of the screen (Fig. 3). Take some time to
adjust all available parameters at the top of the page to update
the heatmap, such as switching the “View Mode” from “Over-
view” to “Detail View.”
7. To download this plot (and any other plot in MetaboAnalyst)
as a high-resolution image, click the paint palette icon at the
top right of the page. The “Graphics Center” dialog will now
appear, where users can adjust the format, resolution, and size
of their image to download. Click “Submit” and the download
link will now appear at the bottom of the dialog. Right click the
hyperlink and select “Save link as. . .” to download the image.
 Metabolomics Data Analysis Using MetaboAnalyst 4.0 341

8. Click the “Download” hyperlink from the left-side navigation

tree to enter the “Results Download” page. Here, users can
download all results and the R Command History from the
“Download.zip” hyperlink in the downloads table. Press the
blue “Generate Report” button to create a PDF report of the
entire analysis. A blue “Analysis Report” hyperlink will appear
on the right, containing a link to the created report. This report
contains all selected parameters, generated figures and tables.
Click on the link and the report will appear in a new tab of the
web-browser. Scroll through the report, which first begins with
an introduction to the selected module, then data processing
steps, followed by the analyses and R Command History. Note
that the Analysis Report contains the updated heatmap figure.
To save the report, right-click the page and select “Save as. . .”
to download it.
9. Press the “Logout” button to exit the session and return to the
MetaboAnalyst home page.

2 A Targeted Metabolomics Workflow

A targeted metabolomics study involves the identification and

quantification of a predefined set of metabolites. It is often applied
for biomarker discovery (i.e., identification of key metabolites asso-
ciated with a phenotype) or to study the dynamics of certain
metabolic pathways of interest. The aim of analyzing targeted
metabolomics data is therefore to uncover important metabolites
that are (1) significantly different between phenotypes, and or
(2) functionally coherent with known molecular and metabolic
mechanisms. The next section demonstrates how to use MetaboA-
nalyst to perform various statistical and functional analyses of data
generated from targeted metabolomics studies.

2.1 Data Upload Metabolomics data contains both biologically relevant variation—
and Processing that is, changes in the metabolome induced by different experi-
mental conditions, and unwanted variation introduced during sam-
ple preparation, instrumentation, and even data preprocessing
(e.g., peak picking, grouping, and alignment). The aim of data
processing is to enhance biologically relevant signals while reducing
unrelated variations in the data [7]. Caution must be applied at this
step to select the most appropriate methods as it can greatly influ-
ence the downstream interpretation of results [8, 9]. Researchers
should keep in mind the aim of their study, the structure of their
data, as well as the statistics they wish to use [7–9]. Methods for
data processing within MetaboAnalyst include missing value impu-
tation, filtering, scaling, centering, and transformations. In this
section we demonstrate how users can process their data using
MetaboAnalyst 4.0.
342 Jasmine Chong and Jianguo Xia

1. Go to the MetaboAnalyst home page and click “>>click here

to start<<” to enter the “Module Overview” page. Select the
“Statistical Analysis” button to enter the corresponding data
upload page. There are two sections—the top section allows
users to upload their own data, and the bottom section allows
users to explore the module using an example data. Select the
first dataset (77 urine metabolomics data from a study on
cancer cachexia [10]) and click the “Submit” button at the
bottom of the page.
2. The “Data Integrity Check” page should appear next. This
page informs users how their data was read in by MetaboAna-
lyst (e.g., summary of samples, features, missing values, and
metadata) and if their data passed all quality checks. As this
dataset passed the integrity check and no missing values were
detected, click the “Skip” button on the bottom right of
the page.
3. The “Normalization” page should now appear. The normali-
zation methods are categorized into three groups: (1) sample
normalization, (2) data transformation, and (3) data scaling.
Sample normalization is used to adjust potential overall differ-
ences among samples (i.e., those caused by volume or mass
differences during sample preparation). Metabolite concentra-
tions in urine samples should be adjusted for the dilution
effect. MetaboAnalyst provides many different methods for
this purpose including the traditional normalization by a refer-
ence compound (i.e., Creatinine in this case), normalization by
sum/median, as well as the probabilistic quotient normaliza-
tion (PQN) [11] using a particular reference sample or a
pooled average sample from a reference group. For demonstra-
tion purposes, we will use the traditional approach. Select
“Normalize by a reference feature” under the subheading
“Sample normalization.” A dialog will appear. Select “Creati-
nine” and click “Submit” to close the dialog. Next, set Data
transformation to “Log transformation” and Data scaling to
“None.” Click “Normalize” to process the data.
4. To view a graphical summary of the resulting data normaliza-
tion, click “View Result.” The distribution of the normalized
data now follows a “bell-shaped” curve compared to the origi-
nal data. Click the “X” on the top right of the dialog to close
the dialog and press “Proceed” on the bottom right of
the page.
5. The “Data Analysis Exploration” page should now be dis-
played. MetaboAnalyst 4.0 currently offers 18 statistical meth-
ods for metabolomics data analysis. These methods are
organized into five general categories: (1) univariate analysis,
(2) chemometrics analysis, (3) significant feature identification,
 Metabolomics Data Analysis Using MetaboAnalyst 4.0 343

Fig. 4 A screenshot of a volcano plot created using an example dataset within MetaboAnalyst highlighting
glucose in a boxplot

(4) cluster analysis, and (5) classification & feature selection.

For the sake of brevity, only a handful of commonly used
statistical methods will be demonstrated.

2.2 Univariate 6. Click “Volcano plot” to view the volcano plot of the data using
Analysis to Identify default parameters. Points highlighted in red are significant
Significant Features based on the default p-value cutoff of 0.1 and fold-change
threshold of 2.0. Click on the single red point to view a boxplot
showing the concentrations of the selected feature within each
group (Fig. 4). Glucose shows a greater concentration in
cachexic patients as compared to controls.
7. There are two icons on the top-right of the plot. Click on the
mini table icon to enter the “Feature Details View” of the
detailed result table from volcano analysis. This table shows
compound names, fold-changes and p-values of statistically
significant features. Click on any feature name to view boxplot
summaries of both the original and normalized concentrations.

2.3 Multivariate Multivariate methods can be classified as unsupervised or super-

Exploratory Data vised. Unsupervised methods do not consider class labels during
Analysis model building and are therefore suitable for data exploration to
identify inherent patterns in the data. In comparison, supervised
methods consider class labels and are often used for biomarker
discovery and classification. The next part will provide instructions
344 Jasmine Chong and Jianguo Xia

on how to perform one unsupervised method—principal compo-

nent analysis (PCA), and two supervised methods—partial least
squares discriminant analysis (PLS-DA) and random forest (RF).
8. To obtain a visual overview of the data, click “Principal Com-
ponent Analysis (PCA)” from the “Data Analysis Exploration”
page or the “PCA” link from the navigation tree to enter the
PCA analysis page. The results from PCA are shown in six tabs.
The first tab shows pairwise score plots between the top five
principal components (PCs). From this default view, it appears
that the two groups overlap significantly with each other.
9. Click the “3D Plot” (the fourth tab) to explore the PCA results
in an interactive 3D score (on the left) and loading plot (on the
right) of the first three PCs. The score plot is used to explore
the underlying structure of the data, and the loading plot is
used to asses which features have the greatest influence on each
component. Use your mouse to rotate and zoom in and out of
each plot (Fig. 5). Note that rotations are synchronized
between the two 3D plots. From the direction of separation
in the score plot, users can easily identify features with strong
contributions to the separation in the loading plot. Click on
any feature in the loading plot to view a boxplot comparing the
concentrations of the selected feature between the two groups.
10. Move on to supervised multivariate statistical methods. Click
the “PLS-DA” hyperlink from the navigation tree to enter the
PLS-DA analysis page. This page contains seven tabs. The
second tab shows the 2D Score Plot of the PLS-DA classifica-
tion model. Compared to PCA, the PLS-DA score plot shows
better separation between the two groups.

Fig. 5 A screenshot of a 3D PCA score and loading plot. Hypoxanthine is highlighted with a boxplot
 Metabolomics Data Analysis Using MetaboAnalyst 4.0 345

Fig. 6 A screenshot of a PLS-DA VIP scores plot highlighting the top 15 most important features

11. Next, click “3D Plot” to view the three-dimensional PLS-DA

score and loading plots (further details in step 9 above). The
3D score plot shows good separation between cachexic and
control patients. The synchronized loading plot can be used
to explore which features drive the separation between the
two groups. For instance, click upon the upper right most
feature (Acetone). The box plot of Acetone shows higher
concentration in controls compared to cachexic patients.
12. To identify the important features of the PLS-DA model,
select the “Imp. Features” tab. A plot summarizing the top
most important features will appear (Fig. 6). MetaboAnalyst
supports two importance measures that summarize the con-
tribution a variable makes to the PLS-DA model—the vari-
able importance in projection (VIP) and the weighted sum of
346 Jasmine Chong and Jianguo Xia

absolute regression coefficients (coef.). By default, it shows

the top 15 features according to their VIP scores. The colored
boxes to the right of the plot indicate whether the average
concentration of a feature is high or low relative to each other.
From the plot, acetone, uracil, and methylguanidin are higher
in controls, while creatine and succinate are higher in cachexic
patients.
13. Lastly, we will explore the Random Forest (RF) method. RF is
an ensemble machine learning method. It builds and com-
bines multiple decision trees to improve accuracy and predic-
tion. Click the “RandomForest” link on the navigation tree.
By default, MetaboAnalyst sets the number of trees built to
500. Users can adjust the number of decision trees to see if
the results can be improved. For instance, change this to 2000
and then click “Update.” Notice that the OOB error
improves from 0.26 to 0.247. However, these values may
change due to the random nature of this algorithm.
14. To view the important features selected by the RF classifier,
select the “Var. Importance” tab. The plot is similar to the
VIP score plot from step 12 except here, features are selected
based on how much they increase OOB error when left out.
The top three features are uracil, isoleucine, and creatine.

2.4 Functional MetaboAnalyst provides two modules for functional interpretation

Interpretation of targeted metabolomics data: (1) metabolite set enrichment anal-
of Targeted ysis, and (2) metabolic pathway analysis. Both methods are based
Metabolomics Data on existing biological knowledge to determine if specific groups of
metabolites (i.e., those involved in pathways [12] or metabolite sets
[13]) are significantly enriched in users’ data. Users can directly
upload a metabolite concentration table, or a list of important
metabolites obtained from the previous statistical analysis. The
next section will first go through metabolic enrichment analysis
followed by metabolic pathway analysis.

2.4.1 Metabolite Set 1. Return to the MetaboAnalyst home page by clicking the home
Enrichment Analysis icon on the top of the navigation tree. Click “>>click here to
start<<” to enter the “Module Overview” page and select the
“Enrichment Analysis” button to enter the data upload page.
2. On the Enrichment Analysis upload page, click the “A concen-
tration table” panel and select the example data named “Data
1”. This is the same data used previously for statistical analysis.
Click “Submit” to upload the data.
3. The “Compound Name/ID Standardization” page should
appear. The aim of compound name standardization is to
match compounds from the user’s data to those curated in
the MetaboAnalyst knowledgebase. Compounds without
 Metabolomics Data Analysis Using MetaboAnalyst 4.0 347

exact matches will be highlighted in yellow with a “View”

button shown in the last column to allow users to perform
approximate search. As there are no compounds without
matches, press the “Submit” button at the bottom of the page.
4. Next is the Data Integrity Check page. Click “Skip” to
continue.
5. The “Normalization overview” page should now be displayed.
Keep the Sample Normalization and Data Transformation
options to none, and select “Auto scaling” under the Data
Scaling options. Click “Normalize” and then “Proceed”.
6. The following page shows the parameters for enrichment anal-
ysis. For demonstration purposes, select the “Pathway-asso-
ciated metabolite sets [SMPDB],” which consists of
99 pathways from the Small Molecule Pathway Database
[14], and keep all other parameters to their default. Click
“Submit” to continue.
7. The Enrichment Analysis results page consists of two tabs, each
with two parts. The first part of each tab consists of a graphical
representation of the results, either as a network or a bar chart.
The bottom part of each tab shows a detailed table of the
enrichment analysis results.
8. Beginning with the Network View (Fig. 7), each circle within
the network (node) represents a metabolite set and is colored
and sized based on its p-value and fold enrichment (hits/
expected), respectively. Two metabolite sets are linked together
with a line (edge) if they share >20% of metabolites. The
network is also interactive, permitting users to drag and drop
nodes, move left and right, as well as zoom in and out. From
the network view, Betaine Metabolism and Glycine and Serine
Metabolism are two of the most enriched metabolite sets and
are also connected with several other metabolite sets.
9. Scroll down to the enrichment analysis results table. Listed are
the pathway names, pathway size, number of compound hits,
and enrichment statistics. Click “View” under the Details col-
umns for further information about a pathway. A popup will
appear. At the top of the popup is a bar-chart of all matched
compounds, colored based on its association to a phenotype.
On the bottom are the common names of all metabolites in the
selected pathway, with matches highlighted in red.
10. Return to the parameters page and this time select “Pathway-
associated metabolite sets (KEGG),” which consists of 80 path-
ways from the KEGG database [12]. Keep all other parameters
as default and click “Submit” at the bottom of the page.
11. From the network, based on their larger size and red color-
ing, Glycine, serine and threonine metabolism and Valine,
348 Jasmine Chong and Jianguo Xia

Fig. 7 A screenshot of the interaction network of the SMPDB pathway metabolite set created using the
Enrichment Analysis module of MetaboAnalyst

leucine and isoleucine biosynthesis are two of the top most

enriched pathway sets. Enrichment analysis using the two dif-
ferent metabolite sets consistently identified alterations in gly-
cine metabolism, which is consistent with cachexia
literature [13, 15–17].

2.4.2 Metabolic Pathway 1. Return to the MetaboAnalyst home page by selecting the home
Analysis icon on the top of the navigation tree. Click “>>click here to
start<<” to enter the “Module Overview” page and select the
“Pathway Analysis” button to enter the data upload page.
2. On the Pathway Analysis upload page, click the “A concentra-
tion table” panel and select the empty box next to “Use the
example data.” This is the same data used previously for enrich-
ment analysis. Click “Submit” to upload the data.
3. The compound standardization, data integrity check, and nor-
malization are identical to steps 3–5 in Section 2.4.1 above.
 Metabolomics Data Analysis Using MetaboAnalyst 4.0 349

4. The next page shows all the options for metabolic pathway
analysis. Keep the default parameters for the pathway analysis
algorithm and KEGG Version. As the example samples were
obtained from humans, select “Homo sapiens (KEGG)” and
click “Submit” at the bottom of the page to perform pathway
analysis.
5. The Pathway Analysis results are displayed on the following
page. The “Overview of Pathway Analysis” plot on the left of
the screen provides a graphical summary of the pathway analy-
sis. In the plot, all matched pathways are represented as circles.
The color and size of each circle corresponds to its p-value and
pathway impact value, respectively. From this plot, Glycine,
serine and threonine metabolism was the most enriched pathway
(largest p-value) and had a pathway impact of 0.612. Click
upon any circle and its corresponding pathway plot will appear
to the right (Fig. 8). Matched compounds in the pathway plot
are colored in red, orange, or yellow based on its
corresponding p-value. Press on any of these matched com-
pounds and a popup will appear showing a boxplot of the
concentrations of the selected metabolite between the two

Fig. 8 A screenshot of the graphical overview of Pathway Analysis results in MetaboAnalyst, highlighting
Glycine, serine and threonine metabolism
350 Jasmine Chong and Jianguo Xia

groups, as well as its p-value and pathway importance score.

The pathway viewer is interactive, permitting zooming in and
out and moving around the viewer.
6. Scroll down the page to view the numeric details of the Path-
way Analysis results. The table contains the names of all
matched pathways, the number of compound matches from
the uploaded data, statistical values, and links to KEGG and
SMPDB. Clicking on any pathway name will update the path-
way plot above.
7. Click the “Download” hyperlink from the navigation tree. This
page permits users to download all generated plots, tables, and
analysis report. Click “Exit” on the navigation tree to end this
session.

3 An Untargeted Metabolomics Workflow

Untargeted metabolomics, also known as global metabolomics,

aims to measure all possible metabolites within samples without a
priori knowledge of the metabolome. Due to its high sensitivity and
wide coverage, high resolution LC-MS is currently the dominant
platform in global metabolomics [18, 19]. A typical LC-MS based
metabolomics experiment can generate 10,000 or morepeaks (fea-
tures) characterized by their mass and retention times. However, as
a single peak can potentially match multiple compounds within the
given mass range, peak annotation requires significant efforts to
search through compound databases and perform tandem MS
experiments, Due to this challenge, functional interpretation of
global metabolomics data is not straightforward, as classical meta-
bolic pathway enrichment analysis requires metabolites as input,
not MS peaks. To address this bottleneck, Li et al. proposed a novel
approach, named mummichog, to directly infer pathway activities
from peak lists by leveraging the collective power of metabolic
pathways, without requiring a priori metabolite identification
[20]. This algorithm assumes that a certain degree of random errors
during individual peak assignment will not change the collective
behavior jointly determined by all metabolites involved in the path-
ways. This concept has been recently adapted to the popular Gene
Set Enrichment Analysis (GSEA) algorithm [21] in MetaboAna-
lystR 2.0 and MetaboAnalyst 4.0. Mummichog is based on over
representation analysis (ORA) to test if certain pathways are
enriched in the significant peaks as compared to null models
based on peak lists of the same size randomly drawn from the
inputted peak list. In comparison, GSEA is a cutoff free method
that evaluates the overall differences of two distributions based on
Kolmogorov–Smirnov tests. The following section will go through
 Metabolomics Data Analysis Using MetaboAnalyst 4.0 351

this module to showcase the functional interpretation of preranked

MS peaks.
3.1 Functional 1. From the MetaboAnalyst home page, click “>>click here to
Interpretation of MS start<<” to enter the “Module Overview” page. Next, select
Peaks the “MS Peaks to Paths” button to enter the data upload page.
Notice that the page is divided into two parts, the top where
users can upload their own peak list or peak intensity table and
the bottom where users can select an example dataset to
explore the module.
2. To use the example data, select the radio button next to IBD.
The example data contains peak list or peak intensity table data
obtained from a subset of fecal samples from 12 pediatric IBD
patients and 12 controls collected from the Integrative Human
Microbiome Project [22]. Click “Submit” to continue.
3. The “Data Integrity Check” page should appear, which
informs users whether the quality of the uploaded data was
suitable for further analysis. Click “Proceed” to move forward.
4. Next is the “Library View,” where users can select which algo-
rithm to use to calculate pathway enrichment (mummichog,
GSEA, or their integration) and their desired species-specific
pathway library. For the mummichog algorithm, users can
either use the “Default cutoff,” which uses the top 10% or
top 500 peaks (whichever is smaller based on the user’s
uploaded MS peaks) as significant or specify a p-value cutoff
to indicate which features will be considered significant. For
demonstration purposes, we will first use the mummichog
algorithm and then the third approach—integrating mummi-
chog and GSEA. Select the mummichog algorithm and click
the button next to “Default p-value cutoff.” Keep the pathway
library selection to “Homo sapiens [MFN],” which was con-
structed from the human BiGG and Edinburgh genome-scale
metabolic networks [23].
5. The following page shows the predicted pathway activity using
the mummichog algorithm. The pathway summary plot at the
top of the page displays all matched pathways as circles (Fig. 9).
The color and size of each circle corresponds to its p-value and
enrichment factor, respectively. The enrichment factor of a
pathway is calculated as the ratio between the number of sig-
nificant pathway hits and the expected number of hits within
the pathway. The plot is also interactive. Hover over any circle
to view the name of the pathway, the enrichment factor, and
-log10 p-value. Further, double-click any circle to view the list
of potential compound hits from the uploaded peak list to
those in the selected pathway. Significant hits are shown in
red and nonsignificant hits are blue. Meanwhile, the pathway
enrichment analysis results table below contains the total size of
the pathway, the number of compound matches, and
352 Jasmine Chong and Jianguo Xia

Fig. 9 A screenshot of the pathway activity profile plot in MetaboAnalyst summarizing the results of the
mummichog algorithm in the MS Peaks to Paths module

raw/adjusted p-values. The top ranked pathways are Vitamin E

metabolism and Bile acid biosynthesis.
6. Users can download the pathway results table from the “Path-
way Hits” button at the top left of the table, as well as all
matched metabolites from the uploaded list of MS peaks from
the “Compound Hits” button at the top center of the table.
Users can also visualize the pathway results in the context of a
KEGG Global Metabolic Network by clicking the “Explore
Results in Network” button.
7. Return to the “Library View” by selecting “Set parameter”
from the navigation tree. For the pathway algorithms, maintain
the mummichog algorithm as checked and also select GSEA.
Keep the pathway library selection to “Homo Sapiens [MFN]”
and press “Submit” to perform the integrated mummichog
and GSEA analysis.
8. The results of the combined mummichog and GSEA p-values
are shown on the following page (Fig. 10). The page is split
into two parts, with the top showing a plot summarizing the
combined pathway analysis, and the bottom containing a
detailed results table. The integrated pathway summary plot
shows the results of the mummichog (y-axis) and GSEA (x-
axis) p-value combination, with the circles representing
 Metabolomics Data Analysis Using MetaboAnalyst 4.0 353

Fig. 10 A screenshot of the pathway summary plot integrating mummichog (y-axis) and GSEA (x-axis) p-
values. The size and color of the circles correspond to the transformed combined p-values

matched pathways. The size and color of the circles correspond

to their transformed combined p-values. Bile acid biosynthesis
was consistently identified as one of the most perturbed path-
ways in pediatric IBD patients as compared to healthy controls
using each algorithm individually as well as integrated. Impor-
tantly, bile acids are well-known to play an important role in the
onset or progression of inflammatory bowel disease [24–26].
9. Click on the “Download” hyperlink from the navigation tree to
enter the “Download” page. Here, the analysis report, R com-
mand history, images, and results tables can be downloaded.
Alternatively, click the home icon to exit the session.

4 Knowledge-Based Multi-Omics Data Integration

Generation of multiple omics data from the same set of samples is

becoming increasingly common nowadays. However, efficient
multi-omics data integration and interpretation has yet to be fully
354 Jasmine Chong and Jianguo Xia

realized. One common approach is to analyze each omics individu-

ally and then combine the resulting lists of significant molecules
onto existing biological knowledgebases for joint analysis and visu-
alization. For instance, to integrate metabolomics and proteomics/
transcriptomics data, a common approach is to co-project metabo-
lites and genes/proteins onto the same metabolic pathways or
networks for visual exploration as well as for enrichment analysis.
The following section will illustrate multi-omics integration first
based on pathways using the Joint Pathway Analysis module, fol-
lowed by network-based integration using the Network Explorer
module.

4.1 Joint Pathway 1. From the MetaboAnalyst home page, click “>>click here to
Analysis start<<” to enter the “Module Overview” page. Select the
“Joint Pathway Analysis” button on the middle-right of the
circular panel to enter the Data Upload page.
2. On the “Joint Upload” page, click the checkbox next to “Use
our example data” to use the example data. A list of genes and
metabolites will show in the corresponding text areas and the
organism will be set to Homo sapiens. This example data con-
tains a subset of transcriptome and metabolome data from a
study of intrahepatic cholangiocarcinoma [27]. Click “Submit”
to continue.
3. The “Compound/Gene Mapping” page is similar to the
“Compound Name Standardization” page from the enrich-
ment/pathway analysis modules. The page contains two tabs,
showing the results of the compound and gene name mapping,
respectively. Compounds or genes with no matches will be
highlighted in red. Scroll down and press “Submit.”
4. The “Analysis Parameters” page allows users to specify para-
meters for enrichment analysis, topology analysis, underlying
pathway databases, and integration method. There are three
types of pathways—the gene–metabolite pathways, gene-
centric pathways, and metabolite-centric pathways. Please
note, many KEGG pathways contain only genes/proteins and
are only available when gene-centric pathways are selected. By
default genes and metabolites are merged into a single list
which is used to perform enrichment analysis against pathways
containing both genes and metabolites. Keep all default para-
meters and press “Submit.”
5. The result from Joint Pathway Analysis follows the same design
as the Pathway Analysis result (Fig. 11). Briefly, the page is
separated into two, with the top half containing the pathway
visualization and the bottom half showing a detailed results
table (refer to steps 5–6 in Section 2.4.2 for further details).
The “Overview of Pathway Analysis” plot displays all matched
pathways as circles, with the color and size of each circle
corresponding to its p-value and pathway impact value,
 Metabolomics Data Analysis Using MetaboAnalyst 4.0 355

Fig. 11 A screenshot of the graphical overview of the Joint Pathway Analysis results in MetaboAnalyst,
highlighting Arginine and proline metabolism

respectively. Click upon any circle for its corresponding path-

way plot to appear on the right.
6. Scroll down to the results table. One of the top ranked path-
ways is Arginine and proline metabolism, with four matched
metabolites/genes and a p-value of 0.0050855. Click upon the
pathway name to visualize the pathway above. The matched
features are highlighted red if upregulated and green if down-
regulated. Further, circles represent matched metabolites and
squares represent genes. Double-click any highlighted node to
get more details about the selected gene or metabolite. At this
moment, it is advisable to also check the results from gene-
centric pathways, as transcriptome measurement is much more
comprehensive as compared to targeted metabolomics.
7. After further exploration, click the “Download” link on the
navigation tree (left-side menu) to download all results. After
downloading, click the “Logout” button to exit the session.

4.2 Network In the previous section, multi-omics integration was limited to

Exploration of Multi- predefined pathways. Pathways represent high-quality, well-charac-
Omics Data terized knowledge that can provide explicit mechanistic under-
standing of the data. However, there are many known interactions
and associations that are not captured in current pathway databases,
which has significantly limited the utility of this approach. A natural
extension is to use a network-based approach. Compared to path-
ways, networks are very flexible and inclusive—they can integrate
356 Jasmine Chong and Jianguo Xia

multiple types of information (including pathways) and may repre-

sent systems-level behavior. In the following section, we demon-
strate how to use the Network Explorer module to map
metabolites/genes onto known molecular interactions and
molecule-phenotype association networks.
8. From the MetaboAnalyst home page, press “>>click here to
start<<” to enter the “Module Overview” page. Next, select
the “Network Explorer” button in the lower right corner of the
circle to enter the Data Upload page.
9. For demonstration purposes, click the “try our example data”
hyperlink on the top right of the page. A dialog will appear with
two options of example datasets, providing details such as ID
type, a brief description of the data, and instrumentation used.
Select the Metabolite-genes data, which is the same intrahepa-
tic cholangiocarcinoma (ICC) dataset used above in the Joint
Pathway Analysis module. Press “Yes” to use the data and then
“Submit” to continue.
10. The “Compound/Gene Mapping” page is identical to the
one from the Joint Pathway Analysis. Refer to step 3 in
Section 4.1 above for further details. Scroll down and press
the “Submit” button.
11. The next page shows the five knowledge-based networks:
(1) the KEGG Global Metabolic Network, (2) metabolite–
disease interaction networks, (3) gene–metabolite interaction
networks, (4) metabolite–metabolite interaction networks,
and (5) metabolite–gene–disease interaction networks.
Details of how these networks were obtained can be found
in the MetaboAnalyst 4.0 and FAQs from the website
[2]. Click on the hyperlink for the “Metabolite-Gene-Disease
Interaction Network” at the bottom of the page to use this
network.
12. The “Mapping Overview” page should appear next. This page
provides a detailed summary of the mapping of the uploaded
data to the selected “Metabolite-Gene-Disease Interaction
Network.” Click “Proceed” to continue to the network
visualization.
13. The “Network Viewer” page displays the default layout of the
uploaded data onto the “Metabolite-Gene-Disease Interac-
tion Network.” This network viewer consists of four parts, the
top toolbar for network customization, the left showing the
Node Explorer table, the center with the interaction network,
and the right with the Function Explorer menu.
14. In the interaction network, metabolites are represented as
diamonds, genes as circles, and diseases as squares. The
matched features are sized based on their position in the
 Metabolomics Data Analysis Using MetaboAnalyst 4.0 357

interaction network, with important nodes (i.e., bigger) hold-

ing key positions in the network. The Node Explorer table
provides two topological measures: node degree and
betweenness centrality. Click the checkbox next to L-Argi-
nine to zoom into this node. L-Arginine is linked to several
blood-related disorders. Details for all selected nodes will
appear in the “Current Selections” box in the bottom left
corner of the page.
15. For further functional insights into the uploaded multi-omics
data, pathway and enrichment analysis can be performed. To
demonstrate this, in the Function Explorer panel (right-side),
keep “Query” set to all nodes and select “Reactome” from the
“Database” dropdown menu. Click “Submit” to perform
pathway analysis. The top pathways are Peptide ligand-
binding receptors, Alternative complement activation, and
Vasopressin-like receptors. To view the features involved in
these pathways, click the checkboxes next to the pathway
names (Fig. 12).
16. MetaboAnalyst provides several features for users to custom-
ize the network and perform other analyses such as finding the
shortest path between any two nodes in the network. Take
time to explore these features.
17. Following network visualization, click on the “Download”
hyperlink from the navigation tree (now at the top of the

Fig. 12 A screenshot of the metabolite-gene-disease interaction network created using example data within
the Network Explorer module of MetaboAnalyst
358 Jasmine Chong and Jianguo Xia

screen) to enter the “Download” page. Alternatively, click the

home icon to exit the session.

5 Conclusion

MetaboAnalyst is one of the most widely used bioinformatics tools

for metabolomics data analysis and interpretation. This chapter first
introduced the general design concepts of MetaboAnalyst 4.0, and
worked through several common workflows for targeted metabo-
lomics, untargeted metabolomics, and multi-omics data integra-
tion. The underlying concepts and approaches of individual
modules have been previously discussed in-depth [7, 28–33]. It
should be noted that study design and raw data preprocessing are
also essential components of a metabolomics data workflow and
heavily influence the interpretation of results. We therefore encour-
age readers to familiarize themselves with proper experimental and
preprocessing approaches to ensure high data quality and reduce
unwanted variation [34–40]. For advanced users with basic knowl-
edge in R, we recommend coupling the MetaboAnalystR package
together with the web server to leverage the strengths of both
approaches to maximize the potential of their data from raw spectra
to biological insights.

Acknowledgement

This work has been supported in part by the US National Institutes

of Health grant U01 CA235493, Natural Sciences and Engineer-
ing Research Council of Canada and Canada Research Chairs
program.

References
1. Xia J, Psychogios N, Young N, Wishart DS more meaningful. Nucleic Acids Res 43:
(2009) MetaboAnalyst: a web server for meta- W251–W257
bolomic data analysis and interpretation. 5. Chong J, Xia J (2018) MetaboAnalystR: an R
Nucleic Acids Res 37:W652–W660 package for flexible and reproducible analysis of
2. Chong J, Soufan O, Li C, Caraus I, Li S, metabolomics data. Bioinformatics
Bourque G, Wishart DS, Xia J (2018) Meta- 34:4313–4314. https://doi.org/10.1093/bio
boAnalyst 4.0: towards more transparent and informatics/bty528
integrative metabolomics analysis. Nucleic 6. Chong J, Yamamoto M, Xia J (2019) Meta-
Acids Res 46:W486–W494. https://doi.org/ boAnalystR 2.0: from raw spectra to biological
10.1093/nar/gky310 insights. Meta 9:57
3. Xia J, Mandal R, Sinelnikov IV, Broadhurst D, 7. van den Berg RA, Hoefsloot HC, Westerhuis
Wishart DS (2012) MetaboAnalyst 2.0—a JA, Smilde AK, van der Werf MJC (2006) Scal-
comprehensive server for metabolomic data ing, and transformations: improving the
analysis. Nucleic Acids Res 40:W127–W133 biological information content of metabolo-
4. Xia J, Sinelnikov IV, Han B, Wishart DS (2015) mics data. BMC Genomics 7:142. https://
MetaboAnalyst 3.0—making metabolomics doi.org/10.1186/1471-2164-7-142.
 Metabolomics Data Analysis Using MetaboAnalyst 4.0 359

8. Temmerman L, Livera A, Bowne JB, Sheedy spectrometry in metabolomics. Springer,

JR, Callahan DL, Nahid A, Souza D, New York, pp 3–12
Schoofs L, Tull DL, McConville M (2012) 19. Theodoridis G, Gika HG, Wilson ID (2008)
Cross-platform urine metabolomics of experi- LC-MS-based methodology for global metab-
mental hyperglycemia in type 2 diabetes. Dia- olite profiling in metabonomics/metabolo-
betes Metab S 6:002 mics. Trac Trend Anal Chem 27:251–260.
9. De Livera AM, Dias DA, De Souza D, https://doi.org/10.1016/j.trac.2008.01.008
Rupasinghe T, Pyke J, Tull D, Roessner U, 20. Li S, Park Y, Duraisingham S, Strobel FH,
McConville M, Speed TP (2012) Normalizing Khan N, Soltow QA, Jones DP, Pulendran B
and integrating metabolomics data. Anal Chem (2013) Predicting network activity from high
84:10768–10776. https://doi.org/10.1021/ throughput metabolomics. PLoS Comput Biol
ac302748b 9:e1003123
10. Eisner R, Stretch C, Eastman T, Xia J, Hau D, 21. Subramanian A, Tamayo P, Mootha VK,
Damaraju S, Greiner R, Wishart DS, Baracos Mukherjee S, Ebert BL, Gillette MA,
VE (2011) Learning to predict cancer- Paulovich A, Pomeroy SL, Golub TR, Lander
associated skeletal muscle wasting from ES et al (2005) Gene set enrichment analysis: a
1 H-NMR profiles of urinary metabolites. knowledge-based approach for interpreting
Metabolomics 7:25–34 genome-wide expression profiles. Proc Natl
11. Dieterle F, Ross A, Schlotterbeck G, Senn H Acad Sci U S A 102:15545–15550. https://
(2006) Probabilistic quotient normalization as doi.org/10.1073/pnas.0506580102
robust method to account for dilution of com- 22. Integrative HMPRNC (2014) The integrative
plex biological mixtures. Application in 1H human microbiome project: dynamic analysis
NMR metabonomics. Anal Chem of microbiome-host omics profiles during per-
78:4281–4290 iods of human health and disease. Cell Host
12. Kanehisa M, Goto S, Sato Y, Furumichi M, Microbe 16:276–289. https://doi.org/10.
Tanabe M (2012) KEGG for integration and 1016/j.chom.2014.08.014
interpretation of large-scale molecular data 23. Li S, Pozhitkov A, Ryan RA, Manning CS,
sets. Nucleic Acids Res 40:D109–D114. Brown-Peterson N, Brouwer M (2010) Con-
https://doi.org/10.1093/nar/gkr988 structing a fish metabolic network model.
13. Puchalska P, Crawford PA (2017) Multi- Genome Biol 11:R115. https://doi.org/10.
dimensional roles of ketone bodies in fuel 1186/gb-2010-11-11-r115
metabolism, signaling, and therapeutics. Cell 24. Tiratterra E, Franco P, Porru E, Katsanos KH,
Metab 25:262–284. https://doi.org/10. Christodoulou DK, Roda G (2018) Role of
1016/j.cmet.2016.12.022 bile acids in inflammatory bowel disease. Ann
14. Jewison T, Su Y, Disfany FM, Liang Y, Knox C, Gastroenterol 31:266–272. https://doi.org/
Maciejewski A, Poelzer J, Huynh J, Zhou Y, 10.20524/aog.2018.0239
Arndt D et al (2014) SMPDB 2.0: big 25. Ogilvie LA, Jones BV (2012) Dysbiosis mod-
improvements to the small molecule pathway ulates capacity for bile acid modification in the
database. Nucleic Acids Res 42:D478–D484. gut microbiomes of patients with inflammatory
https://doi.org/10.1093/nar/gkt1067 bowel disease: a mechanism and marker of dis-
15. Flint TR, Janowitz T, Connell CM, Roberts ease? Gut 61:1642–1643. https://doi.org/10.
EW, Denton AE, Coll AP, Jodrell DI, Fearon 1136/gutjnl-2012-302137.
DT (2016) Tumor-induced IL-6 reprograms 26. Duboc H, Rajca S, Rainteau D, Benarous D,
host metabolism to suppress anti-tumor immu- Maubert MA, Quervain E, Thomas G,
nity. Cell Metab 24:672–684. https://doi. Barbu V, Humbert L, Despras G et al (2013)
org/10.1016/j.cmet.2016.10.010 Connecting dysbiosis, bile-acid dysmetabolism
16. Ham DJ et al (2014) Glycine administration and gut inflammation in inflammatory bowel
attenuates skeletal muscle wasting in a mouse diseases. Gut 62:531–539. https://doi.org/
model of cancer cachexia. Clin Nutr 33 10.1136/gutjnl-2012-302578
(3):448–458 27. Murakami Y, Kubo S, Tamori A, Itami S,
17. Cui P et al (2019) Metabolic derangements of Kawamura E, Iwaisako K, Ikeda K, Kawada N,
skeletal muscle from a murine model of glioma Ochiya T, Taguchi Y (2015) Comprehensive
cachexia. Skelet Muscle 9(1):3 analysis of transcriptome and metabolome
18. Gowda GN, Djukovic D (2014) Overview of analysis in intrahepatic Cholangiocarcinoma
mass spectrometry-based metabolomics: and hepatocellular carcinoma. Sci Rep 5:16294
opportunities and challenges. In: Mass 28. Worley B, Powers R (2013) Multivariate analy-
sis in metabolomics. Curr Metabolomics
360 Jasmine Chong and Jianguo Xia

1:92–107. https://doi.org/10.2174/ Anal Chem 30:1685–1698. https://doi.org/

2213235X11301010092. 10.1016/j.trac.2011.04.019
29. Vinaixa M, Samino S, Saez I, Duran J, Guino- 35. Tautenhahn R, Patti GJ, Rinehart D, Siuzdak
vart JJ, Yanes O (2012) A guideline to univari- G (2012) XCMS online: a web-based platform
ate statistical analysis for LC/MS-based to process untargeted metabolomic data. Anal
untargeted metabolomics-derived data. Meta Chem 84:5035–5039. https://doi.org/10.
2:775–795. https://doi.org/10.3390/ 1021/ac300698c
metabo2040775 36. Myers OD, Sumner SJ, Li S, Barnes S, Du X
30. Rubingh CM, Bijlsma S, Derks EP, Bobeldijk I, (2017) Detailed investigation and comparison
Verheij ER, Kochhar S, Smilde AK (2006) of the XCMS and MZmine 2 chromatogram
Assessing the performance of statistical valida- construction and chromatographic peak detec-
tion tools for megavariate metabolomics data. tion methods for preprocessing mass spectrom-
Metabolomics 2:53–61. https://doi.org/10. etry metabolomics data. Anal Chem
1007/s11306-006-0022-6 89:8689–8695. https://doi.org/10.1021/
31. Bartel J, Krumsiek J, Theis FJ (2013) Statistical acs.analchem.7b01069
methods for the analysis of high-throughput 37. Dudzik D, Barbas-Bernardos C, Garcia A, Bar-
metabolomics data. Comput Struct Biotechnol bas C (2018) Quality assurance procedures for
J 4:e201301009. https://doi.org/10.5936/ mass spectrometry untargeted metabolomics.
csbj.201301009 A review. J Pharm Biomed Anal
32. Marco-Ramell A, Palau-Rodriguez M, Alay A, 147:149–173. https://doi.org/10.1016/j.
Tulipani S, Urpi-Sarda M, Sanchez-Pla A, jpba.2017.07.044
Andres-Lacueva C (2018) Evaluation and 38. Lange E, Tautenhahn R, Neumann S, Gröpl C
comparison of bioinformatic tools for the (2008) Critical assessment of alignment proce-
enrichment analysis of metabolomics data. dures for LC-MS proteomics and metabolo-
BMC Bioinformatics 19(1). https://doi.org/ mics measurements. BMC Bioinformatics
10.1186/s12859-017-2006-0 9:375
33. Hackstadt AJ, Hess AM (2009) Filtering for 39. Tautenhahn R, Bottcher C, Neumann S
increased power for microarray data analysis. (2008) Highly sensitive feature detection for
BMC Bioinformatics 10:11. https://doi.org/ high resolution LC/MS. BMC Bioinformatics
10.1186/1471-2105-10-11 9:504. https://doi.org/10.1186/1471-2105-
34. Hendriks MMWB, van Eeuwijk FA, Jellema 9-504
RH, Westerhuis JA, Reijmers TH, Hoefsloot 40. Zhou B, Xiao JF, Tuli L, Ressom HW (2012)
HCJ, Smilde AK (2011) Data-processing stra- LC-MS-based metabolomics. Mol BioSyst
tegies for metabolomics studies. Trac Trend 8:470–481
 Chapter 18

Using Genome-Scale Metabolic Networks for Analysis,

Visualization, and Integration of Targeted
Metabolomics Data
Jake P. N. Hattwell, Janna Hastings, Olivia Casanueva,
Horst Joachim Schirra, and Michael Witting

Abstract
Interpretation of metabolomics data in the context of biological pathways is important to gain knowledge
about underlying metabolic processes. In this chapter we present methods to analyze genome-scale models
(GSMs) and metabolomics data together. This includes reading and mining of GSMs using the SBTab
format to retrieve information on genes, reactions, and metabolites. Furthermore, the chapter showcases
the generation of metabolic pathway maps using the Escher tool, which can be used for data visualization.
Lastly, approaches to constrain flux balance analysis (FBA) by metabolomics data are presented.

Key words Metabolic networks, Metabolomics, Analysis, Visualization, Integration

1 Introduction

1.1 Genome-Scale Similar to how an insect struggling in a spider’s web sends vibra-
Models tions across the threads, any originally localized disturbance can
result in effects “reverberating” in a complex pattern across the
entire metabolic network. To improve understanding of the com-
plex metabolic networks, present in nature, genome-scale modeling
has emerged in recent years as a means of representing a biological
system in a format that is informative and usable for experiments. A
genome-scale model (GSM, or GSMN for genome-scale metabolic
network) is a digital reconstruction that represents the global met-
abolic processes that occur within an organism [1, 2]. Such a model
should be able to be constrained and optimized to facilitate both
qualitative and quantitative investigation of the biological system of
interest, usually taking the form of in silico simulations. GSMs help
to expedite research by providing a comprehensive knowledge base
of the organism’s metabolism as well as reducing time and financial
costs.

Shuzhao Li (ed.), Computational Methods and Data Analysis for Metabolomics, Methods in Molecular Biology, vol. 2104,
https://doi.org/10.1007/978-1-0716-0239-3_18, © Springer Science+Business Media, LLC, part of Springer Nature 2020

361
362 Jake P. N. Hattwell et al.

Genome-scale models aim to amass as much information about

a biochemical system as possible. The core of a GSM is built using
information about genes, the reactions that the products of those
genes catalyze, and the metabolites and chemical compounds inside
the system [1, 3, 4]. Currently, genes map directly to reactions
using Boolean operations, creating gene associations that control
whether reactions can be active. Reactions consume and produce
metabolites, forming a network of metabolites linked by reactions.
The information in GSMs must be carefully cultivated through
both automated and manual curation [5]. Refinement through
correction of errors and addition of new information ensures that
the model can be used as a reliable source of information, both as a
reference tool and for contextualizing data. This results in a
resource that allows for rapid identification of active pathways and
trends in the biological network, which can provide context to
-omics data.
As advancements in -omics technologies continue to reveal the
complexity of biological systems, different ways to analyze metabo-
lism have evolved. Genetic manipulation of metabolic genes allows
for the study of their influence on different phenotypic traits, while
metabolomics allows for the characterization of the metabolic sta-
tus of a given organism or biological system. Both techniques can
be performed with decent throughput, but each only captures a
small part of what is a genome-wide multi-omic system. Genome-
scale modeling and an accompanying technique, flux balance anal-
ysis (FBA), allows for the computational prediction of intracellular
turnover rates (or fluxes) for all metabolic reactions in a cell or in an
organism. This integrated method overcomes several of the limita-
tions of traditional techniques.
These in silico tools have become very powerful, and a large
number of GSMs for different organisms have been reconstructed
from genomic information and published [6–9]. While the first
GSM of the gram-negative bacteria Haemophilus influenzae was
rather small [10], a trend toward more complex, multicellular
organisms and full-body metabolic reconstructions is underway as
our technology and methods advance [11, 12].

1.2 Caenorhabditis The small nematode Caenorhabditis elegans was the first multicel-
elegans GSM lular organism with a completely sequenced genome [13]. Several
GSMs have been published for C. elegans [14–17]. Recently, these
different models have been merged into a consensus GSMN chris-
tened WormJam (from “worm jamboree,” i.e., curation meetings
of the WormJam community) [18, 19]. This model currently
represents one of the best curated models for C. elegans, but work
is still constantly being undertaken to improve it further. The
WormJam model will be used as the example model throughout
this chapter.
 Genome Scale Metabolic Networks 363

1.3 Metabolomics While GSMs represent large knowledge bases of metabolism in a

and GSMs given organism, the discrepancies between these models and meta-
bolomic measurements can be significant [20]. One of the key
challenges that faces the use of GSMs with metabolomic data is
accurately mapping the identities of metabolites with the model
with those experimentally measured. For example, metabolites
found in a GSM may fall below the limit of detection of current
metabolomics techniques or it may be present in very specific stages
or conditions. Conversely, due to a large gap in the knowledge of
metabolites, many of them might not yet be annotated and
incorporated into the GSM. Despite these limitations, metabolo-
mics data can be successfully combined with FBA for several
applications.
In addition to visualizing metabolomic data on pathway dia-
grams, it is also possible to improve the in silico modeling potential
of genome-scale models through the integration of metabolo-
mics and transcriptomics data. Flux balance analysis (FBA) is a
constraint-based technique that can be used to predict fluxes of
various conditions, which can be of a genetic or environmental
nature. Attractively, FBA does not require kinetic information
about reactions, but instead uses a combination of elementary
pathway analysis and linear programming techniques to optimize
a model. This optimization is commonly for the production of
biomass, but can be changed to other biological processes of
interest.
When unconstrained, the number of possible flux distributions
is essentially infinite. For this reason, experimental -omics data are
used to reduce the system to an allowable solution space, which is
then optimized, or to clarify the objective for optimization.
A variety of -omics data can be integrated with GSMs in a
number of different ways to obtain biologically relevant FBA solu-
tions, as illustrated in Fig. 1. But as mentioned in the introduction,
only those experimentally measured genes and metabolites can be
used to constrain an FBA that are present in the GSM. In this
chapter we present different protocols demonstrating how GSM
and metabolomics data can work together. This includes the use of
GSMs as a metabolome database for annotation of metabolomics
data and the use of targeted metabolomics data to combine with
flux balance analysis. In particular, we explore two methods for the
integration of metabolomic data into in silico experiments. The first
(MetabFBA) uses endometabolomic data to enhance the objective
function of the model [20], and the second (MetaboTools) is the use
of exometabolomic data to constrain an FBA via exchange fluxes
[21, 22].
364 Jake P. N. Hattwell et al.

Fig. 1 Integration of -omics data with flux balance analysis. This illustration shows different ways that -omics
data can be incorporated together with a GSM for FBA. The methods are sorted with respect to the different
omics data sources and whether the data are used to apply constraints to the model (left) or to inform the
objective function that will be optimized (right)

2 Materials

The protocols in this chapter were tested on a computer with

Windows 10 installed. Since the R and Python programs are com-
patible with all major operating systems, the examples here are
expected to work on other computer systems. The specific software
components used here are the following:
1. R 3.5.2
l The tidyverse package (1.2.1).
l Rstudio 1.1.463.
2. Python 3.6+
l Python is freely downloadable from https://www.python.
org/downloads/. The methods within this chapter have
been tested to work in both Python 3.6 and Python 3.7.
l COBRApy and Escher packages for Python (Installation
instructions detailed in Subheading 3.2.1.
3. Demo Scripts: Demonstration scripts and test data are avail-
able for download at the following git repository: https://
github.com/michaelwitting/MiMB_CompMetabo.
 Genome Scale Metabolic Networks 365

3 Methods

These protocols assume basic knowledge and experience of both R

and Python. All R scripts use the tidyverse package and espe-
cially the “pipe” %>% operator to avoid nested function calls. If you
are not familiar with the tidyverse we suggest reading the chapters
in this online web tutorial https://r4ds.had.co.nz/.
In the first example, the statistical programing language R (see
Note 1) is used for inspection of the SBTab files and preparation of
metabolomics data. The tidyverse package is used to allow for
easy access to the data as well as to facilitate data filtering and
manipulation. However, similar scripts, data inspection, and manip-
ulation can be performed in other languages such as Python. The
second part utilizes Python and the packages cobra and escher to
generate Escher maps which can be used for metabolomics data
visualization. Python is then used to demonstrate a basic flux
balance analysis.

3.1 Reading Metabolic models are typically stored in the SBML format, an XML
and Searching derivative customized for the use in systems biology. However, this
for Information format is hardly human readable and contains a lot of information
in SBTab Files to aid the computer in processing the file. In contrast to SBML files,
the Systems Biology Table (SBTab) format allows humans directly
to work with the model. Both formats are interconvertible and free
web tools for conversion are available (https://www.sbtab.net/
sbtab/default/converter.html). The SBTab format comprises a
series of separate files that individually contain the reactions, meta-
bolites, genes, and other information present in the GSM. Because
the data in the individual data files are related to each other, the
same identifiers are used across all data. This makes it possible to
cross-reference and search for information (e.g., isolate metabolites
from a single reaction and retrieve information on them from the
compound table).
We use an example model, which is a truncated version of the
WormJam model containing only information related to the tricar-
boxylic acid (TCA) cycle, pentose phosphate pathway (PPP), and
glycolysis pathway from the WormJam model. This minimal exam-
ple contains the following SBTab files:
– compartments-SBTab.tsv contains the information on all the
model involved compartments,
– compounds-SBTab.tsv contains the chemical information of all
metabolites,
– genes-SBTab.tsv contains all related metabolic genes,
– pathways-SBTab.tsv contains the information on pathways and
mapping to external pathway databases (e.g., KEGG),
366 Jake P. N. Hattwell et al.

Fig. 2 (a) The three most common classes of entities in genome-scale models, and some of the information
they contain. In this representation, Genes 1 and 2 are metabolic genes required for the catalysis of the
reaction, which converts A and B into C. (b) Example tables from SBTab format

– reactions-SBTab.tsv contains all reactions related to the three

pathways mentioned above.
The model contains in total 29 reactions, 49 genes, and
51 compounds. These reactions, genes and compounds were
derived from other previously published models and have been
additionally manually checked and curated. A specialty of the
WormJam model is the detailed curation of metabolite structures.
GSMs use metabolite structures with charges appropriate for the
cytosolic pH of 7.3 and mass balanced reactions, that is, reactions in
which atoms are neither created nor destroyed. However, in meta-
bolomics experiments typically neutral structures are reported. In
order to overcome problems in mapping neutral metabolite names
from metabolomics studies to charged structures, the compound
table contains both the charged and neutral form of the metabolites
(Fig. 2).
 Genome Scale Metabolic Networks 367

3.1.1 Reading of SBTab 1. All SBTab files have to be stored in the same folder. All files are
Files read at the same time by the function below, which reads all files
having a file extension of “-SBTab.tsv” and creates a data frame
(called a tibble) for each file.

# load required libraries -------------------------------------------

---–-----–
library(tidyverse)

# helper function -------------------------------------------

-----------
read_sbtab <- function(folderPath) {

# get all files

sbtab_files <- list.files(folderPath,
pattern = "-SBTab.tsv$",
full.names = TRUE)

# make new list

sbtab_list <- list()

# iterate over all files

for(i in 1:length(sbtab_files)) {

#get current file and add to list

sbtab_file <- sbtab_files[i]
sbtab_list[[i]] <- read_tsv(sbtab_file,
comment = "!!")

# correct names
sbtab_names <- str_replace_all(basename(sbtab_files),
"-SBTab.tsv", "")
names(sbtab_list) <- paste0(sbtab_names,

# make tibble for each table

for(i in 1:length(sbtab_list)) {

assign(names(sbtab_list)[i],
sbtab_list[[i]],
envir = parent.frame())

}
368 Jake P. N. Hattwell et al.

As this function is highly used, we store it in a separate .R

file which is sourced at the beginning of each script. This
custom function is stored in an .R file named “loadSBTab.R”,
which can be afterward sourced from any other .R file to have
access to the reading function.

# load required libraries ------------------------------------

-----------
library(tidyverse)

# source
source("R/loadSBTab.R")

# load tables -----------------------------------------------

------------
read_sbtab("SBTab")

2. Reactions can be printed by typing reactions_table at the

command prompt, which prints the first ten reactions in the
output console. In order to more closely inspect individual
reactions the $ operator can be used to access the reaction
formula ($‘!ReactionFormula‘). This command prints all
29 reactions. Let us now examine closer the reaction catalyzed
by pyruvate synthase.

> reactions_table$‘!ReactionFormula‘[20]
[1] "M_coa_m + M_nad_m + M_pyr_m <=> M_co2_m + M_nadh_m +
M_accoa_m"

The WormJam model uses BiGG identifiers for the meta-

bolites. The leading M denotes the entity as a metabolite. The
ID ends with a c, m, n, or e coding for cytosolic, mitochondrial,
nuclear, and extracellular location of the metabolite, respec-
tively. The files in this chapter only contain entities from the
cytosol and mitochondria. The numbers and letters in between
denote the metabolite itself. M_pyr_c and M_pyr_m are both
identifiers for pyruvate, but with different subcellular locations.
Based on the metabolite identifiers it can be seen that the
reaction occurs in the mitochondrion.
3. Since the BiGG metabolite identifier follows certain rules, it
can be identified by regular expressions. Metabolite identifiers
follow the following regular expression “M_\\w+_(c|m|n|
e|)”. Using the str_extract_all() function it is possible
to isolate all metabolites from the reactions stored in the
 Genome Scale Metabolic Networks 369

reaction table. The following section of code isolates all unique

metabolites IDs and counts their occurrence in the reactions.

# isolate all compounds and count occurrence -----------------

--------–-
metabolite_counts <- reactions_table$‘!ReactionFormula‘ %>%
map(.f=~str_extract_all(.x, "M_\\w+_(c|m|e|n)")) %>%
unlist() %>%
tibble(‘!ID‘ = .) %>%
add_count(‘!ID‘) %>%
distinct(‘!ID‘, .keep_all = TRUE) %>%
arrange(desc(n))

The result is a data frame with the BiGG identifiers fol-

lowed by the count. The highest count was found for the
metabolite M_h_c which represents cytosolic protons. These
protons are often required for mass balancing of reactions and
not informative. The data frame can be filtered to remove
uninformative metabolites or hub metabolites (e.g., H2O).

# filter to remove hub metabolites --------------------------

--------–-
hub_metabolites <- c("M_h_c", "M_h_m", "M_h2o_c", "M_h2o_m")

metabolite_counts %>%
filter(!‘!ID‘ %in% hub_metabolites)

After this filtering, ADP remains as the metabolite with the

highest occurrence.
4. In the next step the genes related to the pyruvate synthase are
extracted. Gene assignments are stored in the column $‘Gen-
eAssociation‘ and WormBase identifiers “WBGene\d+”.

> reactions_table$‘!GeneAssociation‘[20]
[1] "(WBGene00010794 and WBGene00007824 and WBGene00009082 and
WBGene00015413 and WBGene00011510)"

Five different genes are associated with this reaction. They

are connected with an AND operator which indicates that all of
them are required. If a list of genes is required, they can be
isolated from this field with the str_extract_all() func-
tion similar to the metabolites. The regular expression for gene
identifiers is “WBGene\\d+”
370 Jake P. N. Hattwell et al.

reactions_table$‘!GeneAssociation‘[20] %>%
map(.f=~str_extract_all(.x, "WBGene\\d+")) %>%
unlist()

This returns all the gene identifiers as vector. The enzymes

required for this reaction form a large complex and genes
returned encode different subparts of this complex.
5. Lastly, we can Isolate all reactions that contain a specific metab-
olite either as reactant or product using the str_detect()
function. The code below isolates all reactions that contain
pyruvate (independent of subcellular location).

# get all reactions with a pyruvate involved -----------------

--------––––-
filtered_reactions_table <- reactions_table %>%
filter(str_detect(.$‘!ReactionFormula‘, "M_pyr"))

filtered_reactions_table

3.1.2 Interaction and ID 1. So far only entries from the reaction table have been isolated,
Mapping Between Tables without additionally retrieving data from the compound or
gene table. In order to obtain this linked information (e.g.,
gene symbols for each gene related to pyruvate synthase), we
can isolate all WormBase identifiers and query the gene table.
First the gene identifiers are isolated and stored in a vector.

# get genes involved in pyruvate synthase reaction ----------

-----------
gene_list <- reactions_table$‘!GeneAssociation‘[20] %>%
map(.f=~str_extract_all(.x, "WBGene\\d+")) %>%
unlist()

This vector can be then used to filter the gene table using
the filter function.

# filter gene table according to entries in gene_list -------

----------–
genes_table %>%
filter(.$‘!ID‘ %in% gene_list)

The two commands can be combined into a single com-

mand, where first a new data frame containing the WormBase
IDs is generated and is then joined with the information from
the gene table.
 Genome Scale Metabolic Networks 371

# single command to obtain a filter gene table --------------

-----------
reactions_table$‘!GeneAssociation‘[20] %>%
map(.f=~str_extract_all(.x, "WBGene\\d+")) %>%
unlist() %>%
tibble(‘!ID‘ = .) %>%
left_join(., genes_table, by = c("!ID"))

Since Wormbase identifiers act as machine-readable unique

resource identifiers, and are very rarely used in normal work,
this command returns the human-readable gene symbols, in
our case dld-1, dlat-1, dlat-2, pdhb-1, and pdhb-2.
2. Likewise chemical information for a metabolite can be
retrieved. In contrast to genes, the function str_detect()
is used, which utilizes regular expression. The search string
“M_pyr” is used, which returns the entries for the cytosolic
and mitochondrial pyruvate.

# get chemical information for pyruvate ----------------------

----------
compounds_table %>% filter(str_detect(.$‘!ID‘, "M_pyr"))

The data table for the metabolites contains all information,

such as sum formula, chemical structure, and identifiers for the
charged and neutral structures. A specific metabolite with its
subcellular localization can be retrieved using the following.

# get chemical information for pyruvate in the cytosol -------

----------
compounds_table %>% filter(.$‘!ID‘ == "M_pyr_c")

3. Multiple entries can be retrieved using a vector with the IDs of

the metabolites. The resulting code lines are analogous to the
ones used for genes in step 1.

# get all metabolites related to pyruvate synthase reaction

--------
metabolite_list <- reactions_table$‘!ReactionFormula‘[20] %>%
map(.f=~str_extract_all(.x, "M_\\w+_(c|m|e|n)")) %>%
unlist()

# filter gene table according to entries in metabolite list

----------
compounds_table %>% filter(.$‘!ID‘ %in% metabolite_list)
372 Jake P. N. Hattwell et al.

# single command to obtain a filtered compound table ----------

------------
reactions_table$‘!ReactionFormula‘[20] %>%
map(.f=~str_extract_all(.x, "M_\\w+_(c|m|e|n)")) %>%
unlist() %>%
tibble(‘!ID‘ = .) %>%
left_join(., compounds_table, by = c("!ID"))

4. Data in the compound table contain both the charged and

neutral versions of metabolites. The neutral version can be
used to calculate m/z values for metabolite annotation. All
columns that contain information on the neutral metabolite
have a “neutral” in the column name and can be selected in the
code using this. Based on the last script, we will now only select
columns containing the neutral part for all metabolites related
to pyruvate synthase reaction.

# single command to obtain a filtered compound table and select

only neutral
reactions_table$‘!ReactionFormula‘[20] %>%
map(.f=~str_extract_all(.x, "M_\\w+_(c|m|e|n)")) %>%
unlist() %>%
tibble(‘!ID‘ = .) %>%
left_join(., compounds_table, by = c("!ID")) %>%
select(contains("neutral"))

3.1.3 Generation of a 1. The complete information for all metabolites in the reactions
Suspect List for Pathway can be retrieved from the previous script (see Note 2). Based on
Screening these data, a suspect list for inspection or annotation of meta-
bolomics data can be generated. First all metabolites are
isolated from the reactions and their chemical data is retrieved.
The list of metabolites might contain certain molecules that
cannot be measured by the relevant analytical technique (e.g.,
protons or water in the case of mass spectrometry). These
compounds and others that represent only generic compounds
with no explicit chemical structure are removed from the list.
Afterward, only unique entries are selected using the dis-
tinct() function.

# isolate all metabolites from all reactions and leave

unique ---------
neutral_metabolites <- reactions_table$‘!ReactionFormula‘ %>%
map(.f=~str_extract_all(.x, "M_\\w+_(c|m|e|n)")) %>%
unlist() %>%
 Genome Scale Metabolic Networks 373

tibble(‘!ID‘ = .) %>%
left_join(., compounds_table, by = c("!ID")) %>%
filter(!.$‘!ID‘ %in% metabolite_exclusion) %>%
select(contains("neutral")) %>%
distinct()

2. One package that allows for the calculation of m/z values from
neutral masses or formulas is masstrixR (Witting and Schmitt-
Kopplin, under review). Data has to be supplied in a specific
format. Therefore a new data frame called suspect_List is
created.

# create new tibble that can be used with masstrixR ----------

-----------
suspect_List <- tibble(
id = neutral_metabolites$‘!Notes:ChEBI_neutral‘,
smiles = NA,
inchi = neutral_metabolites$‘!Notes:InChI_neutral‘,
inchikey = neutral_metabolites$‘!Notes:InChIKey_neutral‘,
formula = neutral_metabolites$‘!Notes:FORMULA_Neutral‘,
name = neutral_metabolites$‘!Notes:ChEBI_Name_neutral‘,
exactmass = NA
)

3. The masstrixR package can be obtained from GitHub (see

Note 3). After loading the package, the adducts that shall be
covered have to be defined. Metabolites from glycolysis, PPP,
and TCA cycle are typically detected in negative ionization
mode as [M
H]- adduct.

# load masstrixR library -------------------------------------

----------
library(masstrixR)

# create compound list with adduct masses

----------------------------
neg_adducts <- c("[M-H]-")
neg_suspect_List <- as_tibble(prepareCompoundList(suspect_-
List,
adductList = neg_adducts))

This compound list can be used to create EIC traces (e.g.,

in XCMS, or to annotate nontargeted metabolomics data).
374 Jake P. N. Hattwell et al.

Fig. 3 (a) The Escher workflow. Starting with a blank window, reactions are added to the map in a chainlike
fashion, creating a pathway. (b) Overlaying an example table of metabolites onto the generated pathway map.
The Escher Builder window allows for customization of the overlay, through the settings pane

3.2 Generation Escher is a versatile tool for the generation of biological pathway
of Escher Pathway maps that can be used to visualize metabolic systems and give
Map for Visualization context to omics data [23]. Several template Escher maps are avail-
of Metabolic Fluxes able for use on the Escher website; however, for organism-specific
pathways or figure generation, a custom map may be desired. This
section details the generation of an Escher pathway map from
scratch (Fig. 3).
 Genome Scale Metabolic Networks 375

To create a map, a Constraint Based Reconstruction and Anal-

ysis (COBRA) formatted model is needed. Instructions for the
generation of a COBRA model can be found in a Nature Protocols
paper [2].

3.2.1 Installation 1. To install the COBRApy and Escher modules, after installing
of COBRApy and Escher Python 3.6+, open a command prompt window, and enter the
following commands:

python -m pip install --user cobra

python -m pip install --user escher

This will install the COBRApy and Escher packages onto

your account on the computer. To install the packages for all
users, simply omit the “--user” parameter; however, adminis-
trator privileges are required.

3.2.2 Generation 1. A map can be generated using the SBML file of the model,
of Escher Pathway Map Python 3, and the COBRApy and Escher packages. In Python,
enter the following commands.
When a file name is presented (nonitalicized), replace this
with the path to the corresponding file from the Demo Scripts
in Subheading 2.
Lines beginning with a # represent comments to explain
the code and do not need to be entered.

#import dependancies
import cobra
from escher import Builder
#read the SBML model, create an Escher Builder object, and open
the Escher server
model = cobra.io.read_sbml_model(“WormJam_Toy_Network.xml”)
escher_model = Builder(model=model)
escher_model.display_in_browser()

This will start up a local server on your computer and open

up the Escher Builder in a browser window.
2. To begin adding reactions, select “Edit” ! “Add reaction
mode” from the top menu, and click anywhere to add your
first reaction. Type in the Reaction identifier from your model,
and Escher will automatically place it on the canvas. You can
then click on any of the metabolites that participate in the
reaction to add another reaction. The selection tool and rota-
tion tool can be used to arrange items on the canvas. More
information about high level use of Escher can be found in the
376 Jake P. N. Hattwell et al.

publication [23] and the documentation site (https://escher.

readthedocs.io). A sample Escher diagram of the WormJam
model’s glycolysis, PPP, and TCA cycle can be found with the
accompanying files.
3. It is important to save your work using “Map” ! “Save map
JSON”. Whilst Escher is hosted locally, it does not save the
work you have done, so this must be done manually. To load
a Escher map you have worked on previously, use
“Map” ! “Load map JSON”. SVG and PNG formatted
images of your pathway maps can also be exported using the
“Map” menu. The sample map can be loaded in this way.

3.3 Overlay Data obtained from targeted metabolomics can be visualized on

of Metabolomics Data Escher maps (see Subheading 3.3.2). Here we present how such
on Escher dataset can be prepared for the visualization using functions from
Pathway Maps the R tidyverse package. As an example, data from Hastings et al. is
used [20]. Different strains of C. elegans have been followed over
3.3.1 Preparation some time and examined using HILIC-MS/MS.
of Metabolomics Data
1. First the metabolomics data is loaded from .tsv file. Addition-
ally, a file containing mapping information between the metab-
olite names from the result file and the model IDs is read.

# read files ------------------------------------------------

-----------
metabolomics_data <- read_tsv("MetabolomicsData/metabolomic-
s_table.tsv")
model_mapping <- read_tsv("MetabolomicsData/metabo-model-map-
ping.tsv")

2. Next, metabolites that belong to the glycolysis, TCA cycle, or

pentose phosphate pathway are filtered and retained based
using the filter() function. This reduces the number of
metabolites to 18.

# filter metabolites from glycolysis, TCA and PPP ------------

--------------
metabolomics_data_filtered <- metabolomics_data %>%
filter(str_detect(‘METABOLIC PATHWAY‘,
"Glycolysis|TCA|Pentose phosphate pathway"))

3. In order to calculate the group means and standard deviation

we convert the data to a long format using the gather()
function.
 Genome Scale Metabolic Networks 377

# generate long format from data table -----------------------

-----------
metabolomics_long <- gather(metabolomics_data_filtered, key,
value,
-‘Compound (Metabolite)‘,
-‘METABOLIC PATHWAY‘,
convert = TRUE, na.rm = TRUE)

4. The grouping of the different samples is hard-coded in the

sample name. The column with the sample name can be split
into separate columns using the separate() function. This
will generate additional columns, which are named time, strain,
Day, and sampleID.

# generate long format and split sample column to sub groups

---------
metabolomics_long <- gather(metabolomics_data_filtered, key,
value,
-‘Compound (Metabolite)‘,
-‘METABOLIC PATHWAY‘,
convert = TRUE, na.rm = TRUE) %>%
separate(into = c("time", "strain", "Day", "sampleID"),
col = key,
sep = "-")

5. The newly generated data can be then grouped accordingly and

the summarize function is used to calculate the mean, standard
deviation, and the number of samples (n) for each metabolite
and sample group.

# group data and summarize -----------------------------------

----------
metabolomics_summary <- metabolomics_long %>%
group_by(‘Compound (Metabolite)‘, time, strain) %>%
summarize(mean = mean(value),
sd = sd(value),
n = n())

6. Next the data is merged with the table containing the mapping
of the measured and model metabolites. Since not all metabo-
lites that have been measured so far are on the metabolic model
only metabolites that have a model ID are kept.
378 Jake P. N. Hattwell et al.

metabolomics_summary <- metabolomics_summary %>% left_join(.,

model_mapping,
by = c("Compound (Metabolite)" =
"METABOLITE")) %>%
filter(!is.na(MODEL))

7. On Escher maps only two sample conditions can be compared

at the moment. Therefore, data has to be selected accordingly.
The code below uses only metabolite and data points fitting to
the glp condition and measured at 65 or 73 h. Afterward, a data
frame suitable for mapping to Escher maps is generated and
written to a .csv file.

# use filter function to get strain and time points of

interest ------
metabolomics_summary %>% filter(strain == "glp",
time %in% c("65", "73")) %>%
ungroup() %>%
unite(sampleID, strain, time) %>%
select(sampleID, MODEL, mean) %>%
spread(key = sampleID, value =mean) %>%
rename(ID = MODEL) %>%
write_csv("Escher/escher_metabolites.csv")

3.3.2 Mapping The resulting CSV can then be loaded into Escher through the map
and Interpretation of Data JSON file created in Subheading 3.2.2. An example map of the
TCA cycle, PPP, and glycolysis pathway is available in the Demo
Scripts in Subheading 2.
1. To overlay data onto the pathway maps, we begin by importing
dependencies.

import cobra
from escher import Builder

2. We next load the previously constructed pathway map and

open it in a browser window. The file name here can be
replaced with the path to the map you generated in Subheading
3.2.2.

escher_model = Builder(map_json=” TCA_PPP_Glycolysis.json”)

escher_model.display_in_browser()
 Genome Scale Metabolic Networks 379

3. In the Escher Builder window, use the menu to select

“Data” ! “Load Metabolite Data”. Navigate to the “escher_-
metabolites.csv” file and select it to overlay the data onto the
pathway map. Transcriptomic data can also be used in this step
to further augment the visualization, by selecting
“Data” ! “Load Gene Data” or “Data” ! “Load Reaction
Data”, depending on how your data is formatted. The heatmap
is customizable from the “View” “Settings” option of the
main menu.

3.4 Integrating Endo- To incorporate data from the endometabolome into the model,
and Exometabolomic time series metabolomic data of the cells of interest are required.
Data with FBA MetabFBA is a method which integrates intracellular metabolomic
changes with the objective function of the FBA problem, requiring
3.4.1 Using the model to allow for the production or consumption of metabo-
Endometabolome Data lites within the system between time points [20]. The central logic
to Constrain a Flux Balance of MetabFBA is shown in Fig. 4.
Analysis If we compare an example metabolite (Fig. 4b), called M, the
concentration drops during Timeframe f1, indicating that the
demand of M by reactions is greater than the generation of M by
reactions. Conversely in Timeframe f2, the concentration of
M increases; thus, the supply must have exceeded the demand.
The changes in concentration of metabolites across a timeframe
provide information on the changes that are occurring within the
system of interest. The implication that these changes are a result of
net production or consumption of metabolites over time is used by
MetabFBA to augment the objective function, adding the demand
and supply of changing metabolites as variables to be maximized. In
this section we present a basic Python implementation of
MetabFBA.
1. The experimental metabolomics data must first be mapped to
the model before any further analysis can occur. A series of t-
tests are used to assess if the concentrations of metabolites have
changed significantly; thus, replicates are needed. To begin, we
generate a table of metabolites and their concentrations at
various time points. An example of this can be seen in the file
“MetabolomicsData/metabolomics_table.tsv”. Additionally, a
table that allows for the conversion of experimental metabolite
identifiers in the dataset with the identifiers used within the
model must also exist (e.g., “MetabolomicsData/metabo-
model-mapping.tsv”).
2. The metabolite data table is then converted to a “Differences”
table (example: “Metabolomicsdata/metabo-diffs.tsv”), in
which each column represents the difference in concentrations
for each timeframe (the time between adjacent time points),
and each row is a different metabolite. To do this, we perform
t-tests for each combination of metabolite and timeframe in to
380 Jake P. N. Hattwell et al.

Fig. 4 Overview of MetabFBA approach: (a) Schematic overview of the MetabFBA approach, where t1 and t2
are two consecutive time points (t2 > t1). (b) Schematic representation of the underlying assumption that
metabolite level changes between time points correspond to sustained differences in fluxes within the
timeframe

determine significance. If the change is significantly increased

or decreased, then the value of that cell in the table is set to
“UP” or “DOWN,” respectively. If there is no significant
change, we use “NODIFF” as the value of the cell.
3. To integrate the metabolomics data into the model, Python is
used, alongside the COBRApy module, and Python’s in-built
CSV module. In this example, the data tables are saved as TSV
files. First, we import the necessary modules, create variables
for the locations of the relevant files and load the data into
Python. We load the mapping table as a dictionary of key–value
pairs and the differences table are loaded into nested lists. A list
 Genome Scale Metabolic Networks 381

of metabolites present in the data is also generated. The code

presented here is also available from the Demo Scripts.

import csv
import cobra
import cobra.flux_analysis
from cobra.core import Reaction

mapping_file = "PATH_TO_MAPPING_TABLE"
metabolite_file = “PATH_TO_DIFFERENCES_TABLE”
model_file = “PATH_TO_MODEL_FILE

with open(mapping_file,"r") as file:

mapping_data_input = list(csv.reader(file,delimiter="\t"))
mapping_data = {entry[0]:entry[1] for entry in mapping_da-
ta_input}
with open(metabolite_file,’r’) as file:
metabolite_data = list(csv.reader(file,delimiter="\t"))
metabolites = [row[0] for row in metabolite_data][1:]

4. Next, we define a function that will allow us to add reactions for

the increase and decrease of metabolites that change signifi-
cantly in the data.

def addMetabolomicsReaction(model, metabolites, react_name,

coefficient_str=1):
print(react_name,":",metabolites)
# Build a reaction
coefficient_str = str(coefficient_str)
reaction_str = coefficient_str + (" + "+coefficient_str).join
(metabolites) + " <==> " + react_name+"_c"
reaction = Reaction(react_name)
reaction.name = react_name
reaction.subsystem = ’Metabolomics integration’
# Add the reaction to the model
model.add_reactions([reaction])
reaction.build_reaction_from_string(reaction_str)
#print(reaction,":::::",reaction_str)
return reaction

This function takes a COBRA model object, a list of meta-

bolites, the name of the reaction, and a coefficient that is set to
1 by default. It then constructs the COBRA Reaction object
that converts the metabolites list to a single “lumped” metab-
olite and returns it.
382 Jake P. N. Hattwell et al.

5. We next define a function to integrate the metabolomics data.

This function accepts the model and a timeframe as arguments,
then finds the column in the data relating to the timeframe.
The “UP” and “DOWN” tagged metabolites are then sepa-
rated into their own lists and converted to the metabolite
identifiers used in the model using the mapping table. The
previously defined addMetabolomicsReaction function is used
to generate a demand reaction for both the increasing and
decreasing metabolites, before the new demand reactions are
incorporated alongside the Biomass reaction of the model into
the objective function of the FBA problem.

def integrate_metabolomics(model,group):

## generate a list of the UP, DOWN and NODIFF assignments to

metabolites for the column "group"

metabo_diff_values = [row[metabolite_data[0].index(group)]
for row in metabolite_data][1:]

# filter the increased and decreased metabolites and

translate the identifiers to model identifiers using the
mapping data

up_mets_codes = [metabolites[i] for i in range(len

(metabolites)) if metabo_diff_values[i] == "UP"]

up_mets_codes = [mapping_data[i] for i in up_mets_codes if

i in mapping_data and mapping_data[i] != ""]

down_mets_codes = [metabolites[i] for i in range(len

(metabolites)) if metabo_diff_values[i] == "DOWN"]

down_mets_codes = [mapping_data[i] for i in down_mets_codes

if i in mapping_data and mapping_data[i] != ""]

# arbitrary threshold

threshold = 2.5

# If there are changed metabolites, add a sink reaction for

them.

if (len(up_mets_codes)>0):

up_mets_react = addMetabolomicsReaction(model,
up_mets_codes, "up_metabolites")
 Genome Scale Metabolic Networks 383

model.metabolites.up_metabolites_c.compartment = ’c’

model.add_boundary(model.metabolites.up_metabolites_c,
type=’sink’,lb=0,ub=threshold)

if (len(down_mets_codes)>0):

down_mets_react = addMetabolomicsReaction(model,
down_mets_codes, "down_metabolites")

model.metabolites.down_metabolites_c.compartment = ’c’

model.add_boundary(model.metabolites.down_metaboli-
tes_c, type=’sink’,lb=-1∗threshold,ub=0)

# Add the demand reactions to the model objective

if (len(up_mets_codes)>0 and len(down_mets_codes)>0):

model.objective = model.reactions.BIO0100.flux_expres-
sion + model.reactions.up_metabolites.flux_expression - model.
reactions.down_metabolites.flux_expression

elif (len(up_mets_codes)>0):

model.objective = model.reactions.BIO0100.flux_expres-
sion + model.reactions.up_metabolites.flux_expression

elif (len(down_mets_codes)>0):

model.objective = model.reactions.BIO0100.flux_expres-
sion - model.reactions.down_metabolites.flux_expression

return model

6. Finally, we utilize the function to constrain the model with

metabolomic data and perform a parsimonious FBA.

model = cobra.io.read_sbml_model(model_file)
model = integrate_metabolomics(model,“INSERT_COLUMN_NAME”)
solution = cobra.flux_analysis.pfba(model)
print(solution.fluxes[“BIO0100”])

The solution of the flux balance analysis will be displayed.

The methods presented here may improve the quality of pre-
dictions alone; however, the model has not yet been
384 Jake P. N. Hattwell et al.

constrained in any other fashion, and thus configuring the

growth medium and incorporating transcriptomic data will
both greatly improve the predictive accuracy of the model
through further constraining the solution space (Note 4).

3.4.2 Using The exometabolome can also be used to constrain an FBA. The
Exometabolome Data logic of the approach is related to the use of endometabolomic data,
to Constrain a Flux Balance but rather than modifying the objective function, the data are used
Analysis to constrain the FBA directly via using exchange fluxes between the
cells and the external medium:
Extracellular metabolomic data can be obtained through the
analysis of spent medium of cell culture. By analyzing the changes
in metabolite concentration over time and the quantity of cells in
the culture, it is then possible to constrain an FBA experiment.
Differences in the exometabolome over time should be a result of
transport of metabolites in and out of cells, which can then be used
to place qualitative and/or quantitative constraints on the
exchange reactions: The limits of detection of a specific metabolite
and the concentration of that metabolite in the media can be used
to place qualitative constraints on the respective exchange reaction.
Furthering the use of extracellular metabolomic data, the absolute
differences in metabolite concentrations can be used to constrain
exchange reactions in a quantitative fashion.
This method to constrain an FBA with exometabolomic data
has been incorporated into the MetaboTools toolbox [22], and the
MetaboTools protocols article provides an excellent step-by-step
description of the workflow and also contains tutorials of two use
cases. Thus, the present chapter will not provide detailed instruc-
tions on the use of MetaboTools, as any description by us would
merely duplicate already existing resources. Instead, the reader is
referred to the detailed description of the protocol in the Metabo-
Tools paper [22]. MetaboTools is part of the COBRA toolbox
(https://github.com/opencobra/cobratoolbox), and detailed
descriptions on the installation and use of the COBRA toolbox
are provided in a previous protocol paper [24]. An applied example
of the MetaboTools technique is the use of extracellular metabo-
lomic data to constrain the human genome-scale model in order to
study the metabolic profiles of cancer cells [21, 25].

4 Notes

1. Similar data reading and manipulation can be performed in

other programming languages (e.g., Python). R was chosen
since it is used frequently in metabolomics data analysis.
2. The generation of a suspect list for LC-MS is possible with the
WormJam model, since it stores also neutral metabolites. In
 Genome Scale Metabolic Networks 385

case of other models only containing the charged version the

charged formula and the charge state can be used to generate a
neutral formula, which, in turn, can be used to calculate neutral
mass and adduct m/z values.
3. The R package masstrixR can be installed from GitHub like a
normal package using the devtools::install_github()
function.

# install masstrixR -----------------------------------------

------------
devtools::install_github("michaelwitting/masstrixR")

4. Colijn et al. provide a convenient method for the integration of

transcriptomics data in [26].

References
1. Palsson BØ (2015) Systems biology: genome-scale models. Nucleic Acids Res 44
constraint-based reconstruction and analysis. (D1):D515–D522
Cambridge University Press, Cambridge 10. Edwards JS, Palsson BO (1999) Systems prop-
2. Thiele I, Palsson BØ (2010) A protocol for erties of the Haemophilus influenzae Rd meta-
generating a high-quality genome-scale meta- bolic genotype. J Biol Chem 274
bolic reconstruction. Nat Protoc 5:93 (25):17410–17416
3. Lee D-S (2010) Interconnectivity of human 11. Krauss M et al (2012) Integrating cellular
cellular metabolism and disease prevalence. J metabolism into a multiscale whole-body
Stat Mech 2010(12):14 model. PLoS Comput Biol 8(10):e1002750
4. Bergdahl B, Sonnenschein N, Machado D, 12. Yilmaz LS, Walhout AJM (2017) Metabolic
Herrgård M, Förster J (2015) Genome‐scale network modeling with model organisms.
models. In: Villadsen J (ed) Fundamental bio- Curr Opin Chem Biol 36:32–39
engineering, 1st edn. John Wiley & Sons, Inc., 13. C. elegans Sequencing Consortium (1998)
Hoboken, New Jersey. https://doi.org/10. Genome sequence of the nematode
1002/9783527697441.ch06 C. elegans: a platform for investigating biology.
5. Tian M, Kumar P, STP G, Reed JL (2017) Science 282(5396):2012–2018
Metabolic modeling for design of cell factories. 14. Büchel F et al (2013) Path2Models: large-scale
In: Nielsen J, Hohmann S (eds) Systems biol- generation of computational models from bio-
ogy. John Wiley & Sons, Inc., Hoboken, New chemical pathway maps. BMC Syst Biol 7
Jersey. https://doi.org/10.1002/ (1):116
9783527696130.ch3 15. Gebauer J et al (2016) A genome-scale data-
6. Brunk E et al (2018) Recon3D enables a three- base and reconstruction of Caenorhabditis ele-
dimensional view of gene variation in human gans metabolism. Cell Syst 2(5):312–322
metabolism. Nat Biotechnol 36:272 16. Yilmaz LS, Walhout AJ (2016) A Caenorhabdi-
7. de Oliveira Dal’Molin CG et al (2010) Ara- tis elegans genome-scale metabolic network
GEM, a genome-scale reconstruction of the model. Cell Syst 2(5):297–311
primary metabolic network in Arabidopsis. 17. Ma L et al (2017) Systems biology analysis
Plant Physiol 152(2):579–589 using a genome-scale metabolic model shows
8. Sigurdsson MI et al (2010) A detailed genome- that phosphine triggers global metabolic sup-
wide reconstruction of mouse metabolism pression in a resistant strain of C. elegans.
based on human recon 1. BMC Syst Biol 4 bioRxiv
(1):140 18. Witting M et al (2018) Modeling meets meta-
9. Ebrahim A et al (2015) BiGG models: a plat- bolomics—the WormJam consensus model as
form for integrating, standardizing and sharing basis for metabolic studies in the model
386 Jake P. N. Hattwell et al.

organism Caenorhabditis elegans. Front Mol visualizations of biological pathways. PLoS

Biosci 5:96 Comput Biol 11(8):e1004321
19. Hastings J et al (2017) WormJam: a consensus 24. Schellenberger J et al (2011) Quantitative pre-
C. elegans metabolic reconstruction and meta- diction of cellular metabolism with constraint-
bolomics community and workshop series. based models: the COBRA Toolbox v2.0. Nat
Worm 6(2):e1373939 Protoc 6:1290
20. Hastings J et al (2019) Multi-omics and 25. Aurich MK, Fleming RMT, Thiele I (2017) A
genome-scale modeling reveal a metabolic systems approach reveals distinct metabolic
shift during C. elegans aging. Front Mol Biosci strategies among the NCI-60 cancer cell lines.
6:2 PLoS Comput Biol 13(8):e1005698
21. Aurich MK et al (2015) Prediction of intracel- 26. Colijn C, Brandes A, Zucker J, Lun DS,
lular metabolic states from extracellular meta- Weiner B, Farhat MR, Cheng T-Y, Moody
bolomic data. Metabolomics 11(3):603–619 DB, Murray M, Galagan JE (2009) Interpret-
22. Aurich MK, Fleming RMT, Thiele I (2016) ing expression data with metabolic flux mod-
MetaboTools: a comprehensive toolbox for els: predicting Mycobacterium tuberculosis
analysis of genome-scale metabolic models. mycolic acid production. PLoS Comput Biol
Front Physiol 7:327 5(8):e1000489. https://doi.org/10.1371/
23. King ZA et al (2015) Escher: a web application journal.pcbi.1000489
for building, sharing, and embedding data-rich
 Chapter 19

Pathway Analysis for Targeted and Untargeted

Metabolomics
Alla Karnovsky and Shuzhao Li

Abstract
Recent advances in analytical techniques, particularly LC-MS, generate increasingly large and complex
metabolomics datasets. Pathway analysis tools help place the experimental observations into relevant
biological or disease context. This chapter provides an overview of the general concepts and common
tools for pathway analysis, including Mummichog for untargeted metabolomics. Examples of pathway
mapping, MetScape, and Mummichog are explained. This serves as both a practical tutorial and a timely
survey of pathway analysis for label-free metabolomics data.

Key words Metabolomics, Untargeted metabolomics, MetScape, Mummichog, Pathway analysis,

Metabolic network

1 Introduction

The wealth of information of biochemical reactions, accumulated

over more than a century of scientific investigations, is often found
in biochemistry textbooks organized into one pathway per chapter.
Many of the metabolic pathways, such as urea cycle, tricarboxylic
acid cycle, glycolysis, and oxidative phosphorylation, were eluci-
dated in the first half of the past century. Pathways consist of
reactions, and reactions are defined by reactants and products and
often catalyzed by enzymes. Enzymes are proteins that are encoded
in the genome sequence. The emergence of high-throughput gene
expression studies in the 1990s prompted the development of
pathway analysis techniques to help contextualize and interpret
the results. The purpose of these techniques is to help reduce the
data involving hundreds of altered genes to smaller and more
interpretable sets of altered biological “concepts,” helping generate
testable hypotheses. Application of these techniques allowed
researchers to move from assessing changes in individual genes, to
evaluating significance of pathways and other meaningful biological
categories. As pathways represent prior knowledge, the collective

Shuzhao Li (ed.), Computational Methods and Data Analysis for Metabolomics, Methods in Molecular Biology, vol. 2104,
https://doi.org/10.1007/978-1-0716-0239-3_19, © Springer Science+Business Media, LLC, part of Springer Nature 2020

387
388 Alla Karnovsky and Shuzhao Li

statistical power can also identify patterns that are missed at indi-
vidual gene level. Later, as metabolomics data became available,
similar methods were applied to perform pathway analysis on
metabolites.

1.1 Pathway Various mathematical and computational formalisms can be used to

Analysis Approaches model metabolic pathways. For example, an ordinary differential
for Metabolomics Data equation can incorporate the concentrations of metabolites, reac-
tion rates, and mass balance, to model a metabolic reaction nicely.
However, most metabolomics experiments measure the averaged
concentration of each metabolite at a particular moment of a
biological specimen, and do not capture the precise information
that is required for such models. A simplified approach is flux
balance analysis, which assumes steady state of all reactions in a
closed system. More often, statistical models not kinetic models are
used in biomedical investigations of the metabolomes. Stable iso-
tope labeling combined with metabolomics provides a powerful
tool to trace reactions and pathways. We refer the readers to
Chapter 6 for isotope tracing and Chapter 18 for flux balance
analysis. This chapter will focus on the pathway-based bioinformat-
ics analysis of label-free metabolomics data.
Analytical approaches in metabolomics are often referred in two
categories: targeted and untargeted. Targeted metabolomics is the
measurement of defined groups of chemically characterized and
biochemically annotated metabolites [1]. In contrast, untargeted
metabolomics is the comprehensive analysis of all measurable meta-
bolites in a sample. In addition to a set of identified metabolites,
untargeted experiments can yield thousands unidentified features,
of which, a fraction may represent unique metabolites [2]. Com-
mon bioinformatics approaches, such as enrichment tests, can be
used for the analysis of targeted and the known portion of untar-
geted metabolomics data (Fig. 1a). Incorporating the unknown
portion of untargeted experiments requires distinct bioinformatics
techniques (Fig. 1b). Therefore, in this chapter, we will discuss
these approaches separately.

1.2 Pathway A number of databases catalog information on metabolic pathways,

Databases metabolites, metabolic reactions, enzymes and the genes that
and Genome Scale encode them (Note 1). Kyoto Encyclopedia of Genes and Genome
Metabolic Models (KEGG) pioneered this area and is still widely used [3]. Extensive
literature search and expert curation set the initial stage. Later,
genome sequences enabled a new approach of metabolic recon-
struction. That is, proteins encoding metabolic enzymes can be
inferred from genome sequences, and further linked to their respec-
tive biochemical reactions. The reactions are connected into meta-
bolic pathways and networks, under the term “genome scale
metabolic models” (see Chapter 18). BioCyc [4], RECON [5],
and EHMN [6] were some early notable examples. The latter two
were merged into a community model now named RECON3D
 Pathway Analysis 389

Fig. 1 Pathway analysis for metabolomics. (a) Measured metabolites are

mapped to metabolic pathways. Each circle represents a metabolite, whereas
filled circles are significant metabolites selected based on in study question. A
metabolomics experiment only measures a subset of metabolites. Most common
analysis is the enrichment of significant metabolites on a pathway, as indicated
by the red box with dashed lines. The background is based on Chakazul’s
Metabolic Metro Map (https://en.wikipedia.org/wiki/User:Chakazul). (b)
Flowchart to choose an approach of pathway analysis. For untargeted
metabolomics, one can analyze it directly using the Mummichog algorithm, or
perform annotation and treat the annotated subset as targeted metabolomics

[7]. The rapidly expanding number of available genomic sequences

and amount of gene expression data popularized this approach.
Subsequently, a number of more detailed organism-specific meta-
bolic reconstructions have been developed [5–9]. In addition to
detailed information about metabolic pathway topology and indi-
vidual components of pathways, some of these include information
about subcellular compartments where the metabolic reactions
occur [9], and describe metabolic enzyme complexes and
390 Alla Karnovsky and Shuzhao Li

transporters [5]. Different from these genome-centric approaches,

a recent more metabolite-centric effort is Small Molecule Pathway
Database (SMPDB) [10].

2 Materials

This chapter gives an overview of the concepts in metabolomics

pathway and network analysis. One needs a computer with a mod-
ern web browser and Internet access to replicate the exercises,
which are given on two software programs, CytoScape with the
MetScape app and Mummichog. Two datasets are provided on our
GitHub site (https://metabolomics-data.github.io/) to be used
for these examples.

3 Metabolite Identifiers and Pathway Mapping

To query a database and a pathway, a compatible identifier is

required for each metabolite. Due to the head start, KEGG identi-
fiers are commonly used by other software. When one uses other
databases such as PubChem and HMDB, options are given to link
to KEGG identifiers. Metabolite identifier conversion can be via
some web tools, which are listed as notes at the end of this Chapter.
Conventions like InChI and SMILES provide structural informa-
tion of the molecule and are encouraged to be used along with
identifiers and common names.
Structural code and common name and structural code (e.g.,
InChI or SMILES).
If the goal is to map a list of metabolites on pathways, statistical
tests are not necessarily involved. KEGG and Biocyc provide tools
for mapping and “painting” metabolites on pathways. To map
human metabolites to KEGG pathways:
Step 1. Go to https://www.genome.jp/kegg/tool/map_path
way2.html.
Step 2. Enter “hsa” in the “Search against” box.
Step 3. Copy the list of metabolites in KEGG identifiers, one
metabolite per line, paste to use here, then click “Exec.” If
users do not have a list of metabolites to use, example data
can be found on our companion GitHub page.
Step 4. On the result page, click individual pathways to explore.
The metabolites from your input should be colored by pink or
your designated color.
 Pathway Analysis 391

4 Statistical Tests and Application to Targeted Metabolomics

From the field of transcriptomics, pathway analysis can be classified

into three groups: overrepresentation analysis (e.g., DAVID [11]),
functional class scoring (e.g., GSEA [12]), and approaches that
utilize pathway topology (reviewed in [13]) (Note 2). Tools in
the first two categories do not take into account the interactions
between members of a pathway, while the third group of methods
considers such interactions. Extensive experience of pathway and
network analysis of genomic data, particularly in cancer, has shown
that consensus pathway topology can be altered in a disease condi-
tion [14]. In addition to gene expression data, the development of
functional enrichment techniques had a significant impact on the
analysis of epigenetics and GWAS data [12, 15–18]. These pathway
analysis tools provide intuitive results that can be easily interpreted
by biologists. Their popularity has stimulated the development of
similar tools for metabolomics [19–22].
The overrepresentation test is found in almost all software
packages for pathway analysis, using either Fisher’s exact test or
hypergeometric test (the two are similar for large sample sizes).
Using the example in Fig. 1a, there are 40 metabolites measured in
the illustrated experiment. In the pathway defined by the red
dashed circle, 4 of the 7 metabolites are significant. Out of the
total 40 metabolites, 8 are significant. Thus, the significance of the
pathway is computed using Fisher’s exact test as:
  
7 33
4 4
P¼   ¼ 0:02
40
8
A few limitations apply to this approach. The coverage of many
targeted analysis is small, thus insufficient to investigate all the
pathways. The curation of pathways themselves could be limited,
and only a handful of metabolites from a given pathway are typically
measured. Overlap between pathways and correlations between
metabolites from different pathways can induces positive depen-
dence between pathways, requiring special adjustments in order not
inflate the significance of such pathways. Overrepresentation analy-
sis and functional class scoring methods used by most existing
metabolite enrichment tools do not address these problems effi-
ciently. The application of GSEA method should be cautioned on a
separate issue: the level of metabolites in a pathway can change in
different directions. Should an enzyme be dysfunctional in a path-
way, it would result in the accumulation of the upstream metabolite
but depletion of the downstream metabolite. Therefore, the rank-
ing method in GSEA may not apply to such data.
392 Alla Karnovsky and Shuzhao Li

Commercial software programs are available for metabolomics

pathway analysis. MetaCore and Ingenuity Pathway Analysis are
among the popular choices. So far, these software programs only
work with targeted metabolomics data (Note 3).

5 Pathway Visualization

Automated layout of networks has progressed well in the recent

years. However, pathway visualization is more challenging because
logics of chemical structures and conversions are often desired, and
human conventions are often preferred. As a result, manual layout
of metabolic pathways is prevailing. Early in the development,
KEGG has coded their pathways in a markup language, which are
reused by many other bioinformatic tools. Some of these tools
utilize the static pathways charts [23, 24], while others make
them interactive [25, 26]. One such tools called Paintomics can
load metabolite and gene expression measurements and visualize
them over KEGG pathway maps [24]. A more interactive tool,
Vanted, has been developed for exploration of experimental meta-
bolomics data in the context of metabolic pathways, originally from
plants [25, 26]. However, it can be used for any data: users can load
KEGG maps or build their own pathways. Another metabolomics
pathways analysis tool MetPA [27], that is now part of the compre-
hensive data analysis package MetaboAnalyst, in addition to path-
way mapping, calculates pathway impact based on normalized
centrality measure of a given compound relative to the other
compounds.
One of the limitations of visualizing data over pathways charts
stems from the fact that metabolites are often involved in multiple
pathways. In order to understand the overall effect of altered level
of a given metabolite, the user has to go through multiple pathways
and to understand the connections between them. An alternative to
this approach is building a network of genes/metabolites where
each node is unique and nodes from multiple pathways can be
linked together. Such networks provide an easy way to connect
multiple pathways and build gene/compound centric maps
enabling quick data exploration and logical, well-informed hypoth-
esis generation.
As the software ecosystem moves toward the web, native java-
script based pathway visualization starts to emerge. Esher is an
excellent example of this [28]. Esher allows for building, viewing,
and sharing pathway visualization online. A practical example of
Esher is given in Chapter 18.
 Pathway Analysis 393

6 Example Using Metscape

MetScape [29] is a plugin for a widely used network visualization

program CytoScape [30]. It allows users to upload a list of meta-
bolites with experimentally determined concentrations and map
them to reactions, genes and pathways. It also supports identifica-
tion of enriched biological pathways from expression profiling data,
building the networks of genes and metabolites involved in these
pathways, and allow users to visualize the changes in the gene/
metabolite data over time/experimental conditions. MetScape is
using human metabolic pathways, although, it can also map mouse
and rat genes to their human homologs. A screenshot of CytoScape
3.7.1/MetScape 3.1.3 is shown in Fig. 2. To replicate this example:
Step 1. Obtain CytoScape from https://cytoscape.org/. Install
CytoScape by following instructions.
Step 2. Launch CytoScape. From the “Apps” menu, open “App
Manger.” Search “MetScape” and install it.
Step 3. Now MetScape can be found under the “Apps” menu. Click
the “MetScape” tab, open “Build Network” then “Pathway-
based.”
Step 4. Download a copy of the example data from our companion
GitHub page if not done so. This is the same file as the KEGG
mapping exercise.
Step 5. Input the example data file using the “Select” button in the
input box in the left panel. After the data file is read (ignore
warnings), click “Build Network” at the bottom of the left
panel. The network of metabolites similar to Fig. 2 should
appear in the main panel.

7 Untargeted Metabolomics Analysis

The advantage of untargeted metabolomics is to measure as many

metabolites as possible without requiring a priori hypothesis. For
data generated on high-resolution mass spectrometry, most peaks
are likely to be unique individual features. When some of these
peaks match to a chemical library that was generated on the same
platform in the same lab, the annotation of the library can be
transferred to these peaks, resulting in a data table of partial anno-
tation. Sometimes, the annotated peaks are presented separately as
targeted data. In addition to reference-based annotation, computa-
tional methods are available to perform annotation by combining
database searches with heuristic rules (e.g., isotopic patterns, corre-
lation of peak intensities). Furthermore, it becomes possible to
retrospectively annotate peaks in archived high-resolution data.
394 Alla Karnovsky and Shuzhao Li

Fig. 2 Screenshot of MetScape within the CytoScape framework. Using our example data, a metabolite
network is formed by connecting known metabolic reactions that contain the input metabolites. The full
network is shown in the thumbnail and the metabolites of tryptophan pathway are shown in the center of the
window

There is no consensus on the best approach of processing and

analyzing untargeted metabolomics. It is reasonable to expect a
computational annotation tool to group related features into meta-
bolites and to produce a list of metabolites of high-quality annota-
tion. This “in silico” list can be treated as targeted data and
analyzed using targeted tools as discussed in previous sections,
while the peaks without high-quality annotation will be ignored.
A common mistake in untargeted LC-MS metabolomics is to
search each feature in a public database, and use the top or an
arbitrary match for pathway analysis. Since the chromatographic
information is usually irrelevant in such database search, the top
match is usually only one of many possible candidates for the true
metabolites. The arbitrary selection of a candidate bears a small
probability of finding the true metabolite, and this selection is not
 Pathway Analysis 395

necessarily better based on the smallest mass difference between the

query and the candidate. When this process is repeated for many
metabolites, one ends up using a list of mostly false metabolite
identifiers for downstream analysis.
The proper approach should accommodate the ambiguity of
metabolite identification in the statistical framework of pathway
analysis. This was the design of the Mummichog algorithm
[31]. From each feature measured in LC-MS, a list of possible
candidate metabolites is computed (no more than one metabolite
can be true per feature). The algorithm searches pathways using the
combination of all candidates, and calculates an enrichment p-value
for each combination per pathway. Given that most of the candi-
dates and combinations are false predictions, only few results are
expected to land on the true metabolites and correct pathways. The
problem becomes how to distinguish those real signals from ran-
dom data. The distribution of random data can be simulated by
permutating the input data. By contrasting the pathway enrichment
result started from statistically significant features against random
data, the likelihood of finding the correct pathways and metabolites
can be quantified.
Not all metabolites are assigned to a pathway, and many meta-
bolites are shared by multiple pathways. A network model of meta-
bolic reaction is free of the arbitrary boundaries between pathways,
and the unbiased network topology allows us to test network
modules that are enriched by a metabolomics experiment. Both
pathway analysis and network module analysis are provided by
Mummichog. This enables rapid generation of high-quality of
hypotheses directly from untargeted metabolomics without requir-
ing prior annotation, thus accelerating the rate of scientific discov-
ery. Mummichog has been incorporated into the popular web
platforms of on XCMS Online [32] and MetaboAnalyst [33]. It is
also available as an open-source Python package, and a standalone
server (linked via mummichog.org), which is a simple place to use
the tool. Mummichog does not deal with feature level statistics.
Users will produce a p-value per mass feature using their method of
choice (e.g., t-test or linear regression to test significance according
to the study design). A table of mass features (mass to charge ratio,
retention time and p-values) is used as input to Mummichog.
Mummichog only supports LC-MS data at the moment, but ongo-
ing development aims to support more data types and more
upstream processing pipelines. The example output from Mummi-
chog is shown in Fig. 3. Users can use the example data from our
GitHub site to replicate these results.
This strategy of matching patterns in projected spaces between
mass spectrometry and biological models can be broadly applied,
including the integration with other omics data. PIUMet is a
software tool that extends the Mummichog approach by including
protein interaction data, thus integrating protein and gene
396 Alla Karnovsky and Shuzhao Li

Fig. 3 Example output from Mummichog (version 2.1). (a) The summary report contains a replot of user input
data as two Manhattan plots. Significant pathways are shown as a table and a bar plot. Not included but
scrollable in the screenshot is the module analysis and list of metabolites of interest. (b) Example of a significant
network module. (c) Mummichog computes an empirical p-value by comparing the pathway FET p-values (red)
against those from permutation data (blue). Details of the algorithm are given in Li et al. (2013) [31]

expression data with untargeted metabolomics [34]. Because meta-

bolomics data are often orthogonal to gene expression and other
data types, multi-omics integration is often a data-driven process
(see Chapter 23). Prior annotation is not required to perform such
data driven integration until the step of interpreting the data, where
Mummichog can be applied. This works well to include the global
measurement, since discovery of novel biology is a main motivation
of many studies. The online resources related to this chapter are
listed in Note 4.

8 Notes

1. There is a relatively low coverage of experimentally measured

metabolites by biological pathway databases. Most existing
tools rely on the same set of pathway databases that provide a
genome-centric view of the metabolome. In contrast, experi-
mental metabolomics provides a chemistry-centric view of
 Pathway Analysis 397

metabolome determined by a chosen analytical method. The

areas of overlap of these two views typically include primary
metabolism, while the coverage of secondary and lipid metab-
olism is scarce [35]. Further, exogenous compounds that are
routinely detected by metabolomics assays are not included in
pathway maps. As a result, often in current studies, only half of
detected named polar metabolites and even fewer lipids can be
mapped onto metabolic pathways. The improvement of path-
way curation based on experimental measurement of metabo-
lites will be an important area of scientific investigation.
2. Both pathways and networks can be mathematically repre-
sented as graphs. Since pathways are predefined with scientific
conventions, this infusion of human knowledge could increase
the sensitivity of analysis. Analysis via network structure can be
less biased, because networks do not have the boundaries
between pathways.
3. Statistical significance is not the same thing as biological signif-
icance. If one or few metabolites lead to critical biological
insight, pathway analysis is not even necessary. After pathway
analysis, one should validate the findings in some form, espe-
cially with untargeted metabolomics data. Mummichog can
shift the step of metabolite annotation and confirmation after
pathway analysis. Chemical standards and MSn experiments are
a good way to validate metabolites (see Chapter 1 for reporting
standards).
4. Resources.
Metabolic pathways and models.
KEGG: http://www.genome.jp/kegg/kegg2.html
BioCyc: https://biocyc.org
Recon: https://www.vmh.life/#human/all
Metabolite ID conversion.
https://cts.fiehnlab.ucdavis.edu
https://www.metaboanalyst.ca/faces/upload/ConvertView.
xhtml
Metabolomics data analysis tools.
MetaboAnalyst: https://www.metaboanalyst.ca
MetScape: http://metscape.ncibi.org/
MetExplore: https://metexplore.toulouse.inra.fr/
Metabox/Met-DA: http://metda.fiehnlab.ucdavis.edu
XCMS Online: https://xcmsonline.scripps.edu
Mummichog: http://mummichog.org
Network visualization tools.
Cytoscape: http://cytoscape.org
Gehpi: https://gephi.org
398 Alla Karnovsky and Shuzhao Li

Acknowledgments

This work has been funded in part by the US national Institutes of

Health via grants U01CA235487 (Karnovsky, Michailidis), U19
AI090023 (Pulendran), U2C ES026560 (Marsit), and
U01 CA235493 (Li, Xia, Siuzdak).

References
1. Roberts LD, Souza AL, Gerszten RE, Clish CB ND, Selkov E Sr, Sigurdsson MI, Simeonidis E,
(2012) Targeted metabolomics. Curr Protoc Sonnenschein N, Smallbone K, Sorokin A, van
Mol Biol. Chapter 30:Unit 30.32.31-24. Beek JH, Weichart D, Goryanin I, Nielsen J,
https://doi.org/10.1002/0471142727. Westerhoff HV, Kell DB, Mendes P, Palsson
mb3002s98 BO (2013) A community-driven global recon-
2. Mahieu NG, Patti GJ (2017) Systems-level struction of human metabolism. Nat Biotech-
annotation of a metabolomics data set reduces nol 31(5):419–425. https://doi.org/10.
25000 features to fewer than 1000 unique 1038/nbt.2488
metabolites. Anal Chem 89 8. Sigurdsson MI, Jamshidi N, Steingrimsson E,
(19):10397–10406. https://doi.org/10. Thiele I, Palsson BO (2010) A detailed
1021/acs.analchem.7b02380 genome-wide reconstruction of mouse metab-
3. Kanehisa M, Goto S, Hattori M, Aoki- olism based on human Recon 1. BMC Syst Biol
Kinoshita KF, Itoh M, Kawashima S, 4:140. https://doi.org/10.1186/1752-0509-
Katayama T, Araki M, Hirakawa M (2006) 4-140
From genomics to chemical genomics: new 9. Hao T, Ma HW, Zhao XM, Goryanin I (2010)
developments in KEGG. Nucleic Acids Res 34 Compartmentalization of the Edinburgh
(database issue):D354–D357 human metabolic network. BMC Bioinformat-
4. Caspi R, Altman T, Dreher K, Fulcher CA, ics 11:393. https://doi.org/10.1186/1471-
Subhraveti P, Keseler IM, Kothari A, 2105-11-393
Krummenacker M, Latendresse M, Mueller 10. Jewison T, Su Y, Disfany FM, Liang Y, Knox C,
LA, Ong Q, Paley S, Pujar A, Shearer AG, Maciejewski A, Poelzer J, Huynh J, Zhou Y,
Travers M, Weerasinghe D, Zhang P, Karp Arndt D, Djoumbou Y, Liu Y, Deng L, Guo
PD (2012) The MetaCyc database of metabolic AC, Han B, Pon A, Wilson M, Rafatnia S,
pathways and enzymes and the BioCyc collec- Liu P, Wishart DS (2014) SMPDB 2.0: big
tion of pathway/genome databases. Nucleic improvements to the small molecule pathway
Acids Res 40(Database issue):D742–D753. database. Nucleic Acids Res 42(Database
https://doi.org/10.1093/nar/gkr1014 issue):D478–D484. https://doi.org/10.
5. Duarte NC, Becker SA, Jamshidi N, Thiele I, 1093/nar/gkt1067
Mo ML, Vo TD, Srivas R, Palsson BO (2007) 11. Huang DW, Sherman BT, Tan Q, Collins JR,
Global reconstruction of the human metabolic Alvord WG, Roayaei J, Stephens R, Baseler
network based on genomic and bibliomic data. MW, Lane HC, Lempicki RA (2007) The
Proc Natl Acad Sci U S A 104(6):1777–1782. DAVID gene functional classification tool: a
https://doi.org/10.1073/pnas.0610772104 novel biological module-centric algorithm to
6. Ma H, Sorokin A, Mazein A, Selkov A, functionally analyze large gene lists. Genome
Selkov E, Demin O, Goryanin I (2007) The Biol 8(9):R183. https://doi.org/10.1186/
Edinburgh human metabolic network recon- gb-2007-8-9-r183
struction and its functional analysis. Mol Syst 12. Subramanian A, Tamayo P, Mootha VK,
Biol 3:135. https://doi.org/10.1038/ Mukherjee S, Ebert BL, Gillette MA,
msb4100177 Paulovich A, Pomeroy SL, Golub TR, Lander
7. Thiele I, Swainston N, Fleming RM, Hoppe A, ES, Mesirov JP (2005) Gene set enrichment
Sahoo S, Aurich MK, Haraldsdottir H, Mo analysis: a knowledge-based approach for inter-
ML, Rolfsson O, Stobbe MD, Thorleifsson preting genome-wide expression profiles. Proc
SG, Agren R, Bolling C, Bordel S, Chavali Natl Acad Sci U S A 102(43):15545–15550.
AK, Dobson P, Dunn WB, Endler L, Hala D, https://doi.org/10.1073/pnas.0506580102
Hucka M, Hull D, Jameson D, Jamshidi N, 13. Khatri P, Sirota M, Butte AJ (2012) Ten years
Jonsson JJ, Juty N, Keating S, Nookaew I, Le of pathway analysis: current approaches and
Novere N, Malys N, Mazein A, Papin JA, Price outstanding challenges. PLoS Comput Biol 8
 Pathway Analysis 399

(2):e1002375. https://doi.org/10.1371/jour PaintOmics 3: a web resource for the pathway

nal.pcbi.1002375 analysis and visualization of multi-omics data.
14. Creixell P, Reimand J, Haider S, Wu G, Nucleic Acids Res 46(W1):W503–w509.
Shibata T, Vazquez M, Mustonen V, https://doi.org/10.1093/nar/gky466
Gonzalez-Perez A, Pearson J, Sander C, 23. Paley SM, Karp PD (2006) The pathway tools
Raphael BJ, Marks DS, Ouellette BFF, cellular overview diagram and Omics viewer.
Valencia A, Bader GD, Boutros PC, Stuart Nucleic Acids Res 34(13):3771–3778.
JM, Linding R, Lopez-Bigas N, Stein LD https://doi.org/10.1093/nar/gkl334
(2015) Pathway and network analysis of cancer 24. Garcia-Alcalde F, Garcia-Lopez F, Dopazo J,
genomes. Nat Methods 12(7):615–621. Conesa A (2011) Paintomics: a web based tool
https://doi.org/10.1038/nmeth.3440 for the joint visualization of transcriptomics
15. Huang DW, Sherman BT, Tan Q, Kir J, Liu D, and metabolomics data. Bioinformatics 27
Bryant D, Guo Y, Stephens R, Baseler MW, (1):137–139. https://doi.org/10.1093/bioin
Lane HC, Lempicki RA (2007) DAVID bioin- formatics/btq594
formatics resources: expanded annotation data- 25. Junker BH, Klukas C, Schreiber F (2006)
base and novel algorithms to better extract VANTED: a system for advanced data analysis
biology from large gene lists. Nucleic Acids and visualization in the context of biological
Res 35(Web Server issue):W169–W175. networks. BMC Bioinformatics 7:109.
https://doi.org/10.1093/nar/gkm415 https://doi.org/10.1186/1471-2105-7-109
16. Lee PH, O’Dushlaine C, Thomas B, Purcell 26. Klukas C, Schreiber F (2010) Integration of
SM (2012) INRICH: interval-based enrich- -omics data and networks for biomedical
ment analysis for genome-wide association research with VANTED. J Integr Bioinform 7
studies. Bioinformatics 28(13):1797–1799. (2):112. https://doi.org/10.2390/biecoll-
https://doi.org/10.1093/bioinformatics/ jib-2010-112
bts191 27. Xia J, Wishart DS (2010) MetPA: a web-based
17. Segre AV, Groop L, Mootha VK, Daly MJ, metabolomics tool for pathway analysis and
Altshuler D (2010) Common inherited varia- visualization. Bioinformatics 26
tion in mitochondrial genes is not enriched for (18):2342–2344. https://doi.org/10.1093/
associations with type 2 diabetes or related gly- bioinformatics/btq418
cemic traits. PLoS Genet 6(8):e1001058. 28. King ZA, Drager A, Ebrahim A,
https://doi.org/10.1371/journal.pgen. Sonnenschein N, Lewis NE, Palsson BO
1001058 (2015) Escher: a web application for building,
18. Cavalcante RG, Lee C, Welch RP, Patil S, sharing, and embedding data-rich visualiza-
Weymouth T, Scott LJ, Sartor MA (2014) tions of biological pathways. PLoS Comput
Broad-enrich: functional interpretation of Biol 11(8):e1004321. https://doi.org/10.
large sets of broad genomic regions. Bioinfor- 1371/journal.pcbi.1004321
matics 30(17):i393–i400. https://doi.org/10. 29. Karnovsky A, Weymouth T, Hull T, Tarcea VG,
1093/bioinformatics/btu444 Scardoni G, Laudanna C, Sartor MA, Stringer
19. Cavalcante RG, Patil S, Weymouth TE, Bend- KA, Jagadish HV, Burant C, Athey B, Omenn
inskas KG, Karnovsky A, Sartor MA (2016) GS (2012) Metscape 2 bioinformatics tool for
ConceptMetab: exploring relationships the analysis and visualization of metabolomics
among metabolite sets to identify links among and gene expression data. Bioinformatics 28
biomedical concepts. Bioinformatics 32 (3):373–380. https://doi.org/10.1093/bioin
(10):1536–1543. https://doi.org/10.1093/ formatics/btr661
bioinformatics/btw016 30. Shannon P, Markiel A, Ozier O, Baliga NS,
20. Lopez-Ibanez J, Pazos F, Chagoyen M (2016) Wang JT, Ramage D, Amin N,
MBROLE 2.0-functional enrichment of chem- Schwikowski B, Ideker T (2003) Cytoscape: a
ical compounds. Nucleic Acids Res 44(W1): software environment for integrated models of
W201–W204. https://doi.org/10.1093/ biomolecular interaction networks. Genome
nar/gkw253 Res 13(11):2498–2504. https://doi.org/10.
21. Xia J, Wishart DS (2016) Using MetaboAna- 1101/gr.1239303
lyst 3.0 for comprehensive metabolomics data 31. Li S, Park Y, Duraisingham S, Strobel FH,
analysis. Curr Protoc Bioinformatics Khan N, Soltow QA, Jones DP, Pulendran B
55:14.10.11–14.10.91. https://doi.org/10. (2013) Predicting network activity from high
1002/cpbi.11 throughput metabolomics. PLoS Comput Biol
22. Hernandez-de-Diego R, Tarazona S, 9(7):e1003123. https://doi.org/10.1371/
Martinez-Mira C, Balzano-Nogueira L, Furio- journal.pcbi.1003123
Tari P, Pappas GJ Jr, Conesa A (2018)
400 Alla Karnovsky and Shuzhao Li

32. Huan T, Forsberg EM, Rinehart D, Johnson 34. Pirhaji L, Milani P, Leidl M, Curran T, Avila-
CH, Ivanisevic J, Benton HP, Fang M, Pacheco J, Clish CB, White FM, Saghatelian A,
Aisporna A, Hilmers B, Poole FL, Thorgersen Fraenkel E (2016) Revealing disease-associated
MP, Adams MWW, Krantz G, Fields MW, Rob- pathways by network integration of untargeted
bins PD, Niedernhofer LJ, Ideker T, Majum- metabolomics. Nat Methods 13(9):770–776.
der EL, Wall JD, Rattray NJW, Goodacre R, https://doi.org/10.1038/nmeth.3940
Lairson LL, Siuzdak G (2017) Systems biology 35. Barupal DK, Haldiya PK, Wohlgemuth G,
guided by XCMS online metabolomics. Nat Kind T, Kothari SL, Pinkerton KE, Fiehn O
Methods 14(5):461–462. https://doi.org/ (2012) MetaMapp: mapping and visualizing
10.1038/nmeth.4260 metabolomic data by integrating information
33. Chong J, Soufan O, Li C, Caraus I, Li S, from biochemical pathways and chemical and
Bourque G, Wishart DS, Xia J (2018) Meta- mass spectral similarity. BMC Bioinformatics
boAnalyst 4.0: towards more transparent and 13:99. https://doi.org/10.1186/1471-
integrative metabolomics analysis. Nucleic 2105-13-99
Acids Res 46(W1):W486–W494. https://doi.
org/10.1093/nar/gky310
 Chapter 20

Application of Metabolomics to Renal and Cardiometabolic

Diseases
Casey M. Rebholz and Eugene P. Rhee

Abstract
Metabolomics has been increasingly applied to study renal and related cardiometabolic diseases, including
diabetes and cardiovascular diseases. These studies span cross-sectional studies correlating metabolites with
specific phenotypes, longitudinal studies to identify metabolite predictors of future disease, and physio-
logic/interventional studies to probe underlying causal relationships. This chapter provides a description of
how metabolomic profiling is being used in these contexts, with an emphasis on study design considerations
as a practical guide for investigators who are new to this area. Research in kidney diseases is underlined to
illustrate key principles. The chapter concludes by discussing the future potential of metabolomics in the
study of renal and cardiometabolic diseases.

Key words Metabolomics, Study design, Cardiometabolic disease, Renal disease, Cardiovascular
disease, Biomarker, Replication

1 Introduction

1.1 Current Use Biomarkers of cardiometabolic disease, including kidney disease,

of Cardiometabolic are already being widely used in both research and clinical settings.
and Renal Disease There are many biomarkers that can help with diagnosing and
Biomarkers monitoring disease progression, stratifying the population based
in Research on risk of developing a clinical outcome, and serving as surrogate
and Practice markers of hard clinical outcomes in clinical trials. Some key bio-
markers are peptides or proteins, such as brain natriuretic peptide
and cardiac troponin, respectively. Elevated levels of cardiac tropo-
nin indicate myocardial injury, which, in the setting of myocardial
ischemia, is sufficient to diagnose myocardial infarction [1] and in
the appropriate clinical context trigger acute treatment. Other
diagnostic biomarkers are metabolites, such as glucose in diabetes
or uric acid with gout. Blood levels of cholesterol and triglycerides
are used to diagnose dyslipidemias, but can also be used along with
other clinical and lifestyle factors to estimate risk of cardiovascular
disease, originally with the Framingham risk equation and more

Shuzhao Li (ed.), Computational Methods and Data Analysis for Metabolomics, Methods in Molecular Biology, vol. 2104,
https://doi.org/10.1007/978-1-0716-0239-3_20, © Springer Science+Business Media, LLC, part of Springer Nature 2020

401
402 Casey M. Rebholz and Eugene P. Rhee

recently with the American Heart Association/American College of

Cardiology Pooled Cohort Risk Equation, and guide decisions
regarding whether or not to treat with lipid-lowering therapy
[2, 3]. For kidney disease, blood levels of cystatin C (a peptide)
and creatinine (a metabolite), as well as urine levels of protein
(albuminuria) are the basis for diagnosis and staging [4, 5]. Further-
more, change in estimated glomerular filtration rate (eGFR), which
is calculated with blood levels of creatinine, has been adopted as
surrogate endpoint for research studies [6].
Despite their utility in various settings, current biomarkers also
have important limitations. This is well illustrated with existing
renal biomarkers. Creatinine, the main endogenous marker of kid-
ney function—or more specifically, kidney filtration (“glomerular
filtration rate”, GFR)—is confounded by muscle mass, dietary
intake of protein, and tubular secretion [7]. Cystatin C, a more
recently characterized kidney filtration marker, is influenced by
other non-GFR determinants including inflammation [8]. Blood
elevations of creatinine and cystatin C, although indicative of kid-
ney function, are not pathologic; that is, they do not make a causal
contribution to disease pathogenesis, in contrast to elevated glu-
cose levels in diabetes or elevated uric acid levels in gout. Although
there is considerable evidence from both clinical trials and observa-
tional studies for change in albuminuria as a surrogate outcome for
chronic kidney disease progression, and albuminuria may be both a
marker and mediator of disease progression, its use is limited in part
due to high within person variability thereby hampering the ability
to differentiate between clinically significant change and random
error [9–12]. Finally, creatinine, cystatin C, and urine albumin all
share the limitation of nonspecificity—they do not provide strong
insight into the underlying cause of kidney disease. Common con-
ditions such as diabetes and hypertension, as well as less common
conditions such as autoimmune disease, allergic drug reaction, or
inherited structural anomalies can all cause kidney disease, and
alterations in these biomarkers are unable to differentiate specific
etiologies.

1.2 Opportunity There is a substantial need for the discovery of new biomarkers of
for Metabolomics renal and other cardiometabolic diseases to overcome existing lim-
to Improve Biomarker itations. As outlined in more detail elsewhere in this book, current
Research metabolomic platforms can detect >1000 known metabolites and
often many more unknown ion features, many of which will be
characterized and added to the growing list of known metabolites
over time [13–15]. Alterations in metabolites could yield new
insights into disrupted metabolism characteristic of prevalent dis-
ease states as well as future disease prognosis, and ideally, provide
mechanistic insight and outline new treatment targets as well.
Further, by integrating both endogenous metabolism and exoge-
nous inputs, such as diet, medications, and the microbiome,
 Metabolomics for Renal and Cardiometabolic Diseases 403

metabolite alterations have the potential to capture functional

mediators of disease that are not assessed by other molecular
domains.

2 How to Apply Metabolomics to Clinical Questions

In the process of designing and implementing a metabolomics

study to address a clinical question, it is essential for the investigator
to take certain factors into consideration ranging from study design
and analytic considerations to clinical characteristics of study parti-
cipants and details about biospecimen collection and storage. New
studies on the metabolomics of cardiometabolic diseases should at
least meet the standards set by previously conducted research of this
kind and ideally would expand upon the existing research in terms
of rigor and validity.

2.1 Designing Different study designs have their own strengths and weaknesses,
a Population or Clinical and their suitability depends on the study question of interest as
Study well as practical constraints such as sample size and available
resources. Cross-sectional, cohort, and case-control studies are types
of epidemiologic, observational study designs that allow for the
description of free-living populations with respect to their expo-
sures (including metabolites) and disease outcomes.
Cross-sectional metabolomics studies assess the metabolome
and the outcome of interest at the same point in time, allowing
for the metabolomic characterization of a prevalent disease state.
Cross-sectional studies can compare metabolite levels in individuals
with or without a disease—for example, coronary disease, type
2 diabetes, CKD—or can test the association (or correlation) of
metabolite levels with related continuous traits. For example, one
study measured ~500 blood metabolites in 1735 participants of the
KORA F4 study, followed by replication in 1164 individuals in the
TwinsUK registry. This study demonstrated the substantial impact
of kidney function—in KORA F4, 112 metabolites were signifi-
cantly associated with eGFR, 54 of which replicated in TwinsUK
[16]. For select metabolites most strongly correlated with eGFR,
the authors then examined their association with measured GFR
among 200 participants of the AASK study, highlighting
c-mannosyltryptophan and pseudouridine as alternative or comple-
mentary markers of kidney function. Cross-sectional studies have
also identified numerous blood metabolites that correlate with
other cardiometabolic traits such as body mass index and insulin
sensitivity, including branched chain amino acids and lipids
enriched for saturated fatty acids [17, 18]. Thus, cross-sectional
studies have been useful in highlighting metabolite perturbations
that coincide with abnormal pathophysiologic conditions.
Although these studies have the potential to identify novel disease
404 Casey M. Rebholz and Eugene P. Rhee

biomarkers and risk factors, the fact that metabolites and outcomes
are being assessed at the same time precludes conclusions about the
direction of causality.
Unlike cross-sectional analyses, observational studies allow for
the establishment of temporality, whereby metabolomic profiling
can be conducted at a baseline time point and the outcome events
occur subsequently such that the exposure (metabolite alteration)
precedes disease onset. These types of studies are useful for identi-
fying early (preclinical) markers of disease risk that could be used to
risk stratify the population, and potentially to identify causal disease
pathways and therapeutic targets. Observational studies can have
different designs, including cohort and case-control studies. A cohort
study considers all subjects available in a given study population,
some of whom have the outcome of interest (cases). Cohort studies
have the advantage of having a well-characterized underlying study
population and allow for multiple study questions to be addressed.
For example, metabolomics data generated at a baseline time point
in a large cohort can be tested for association with any longitudinal
outcome that has been rigorously assessed over time, such as new
onset kidney disease, diabetes, or cardiovascular disease. For meta-
bolites that emerge as potential risk markers, normal reference
ranges can be determined across all individuals in the cohort and
the correlation with various demographic factors and comorbidities
can be explored. Extensive profiling of a given cohort also permits
other avenues of discovery, for example conducting a genome wide
association study of metabolite levels if genotyping data has also
been generated [19], and ultimately even pursuing Mendelian
randomization analyses [20].
Assaying all samples in a given cohort study is not always
feasible, considering the cost of metabolomic profiling and the
value of nonrenewable biological specimens, and alternatives are
required to answer specific questions of interest more efficiently. In
a case-control study, one compares individuals who have the out-
come of interest (cases) with individuals who do not but are other-
wise similar (controls). Often, cases and controls are selected from a
larger cohort, or in other words are “nested” within the cohort
study. For example, one study utilized a nested case-control study
design within the Chronic Renal Insufficiency Cohort, consisting
of 200 cases with rapid progression of kidney disease (defined as
GFR loss >3 ml/min per year) and 200 controls with stable kidney
function over time who were matched to cases on baseline eGFR
and proteinuria [21]—this subset of 400 individuals was selected
from the nearly ~4000 recruited into the parent cohort. This study
identified a panel of three metabolites (arginine, methionine, and
threonine) for which higher levels were strongly and independently
associated with rapid kidney disease progression, were not asso-
ciated with baseline eGFR, and might serve as markers of renal
metabolic capacity.
 Metabolomics for Renal and Cardiometabolic Diseases 405

Selection bias is an important limitation that can arise when

designing a case-control study, as it can be challenging to assemble
an appropriate control group. Case-control studies often are more
appropriate for highlighting associations for wider validation and
characterization. Indeed, the nested case-control study outlined
above serves as a springboard for consortial studies that seek to
identify metabolomic predictors of CKD progression across many
thousands of individuals within the context of the Chronic Kidney
Disease Biomarkers Consortium (https://www.
ckdbiomarkersconsortium.org/).
As already noted, the main advantage of a case-control design is
the ability to enrich for cases, which may represent a small fraction
of the overall cohort, and the ability to constrain overall sample
number through the identification of otherwise similar controls
(typically ~1–3 fold the number of selected cases). A third kind of
observational study design, a case-cohort study, allows for preserva-
tion of the underlying cohort through random selection of the
subcohort, rather than trying to find “matching” controls that
closely resemble cases, while making sure to enrich for cases with
the outcome of interest. In other words, cases are selected as in a
case-control study, but are ultimately compared to a random sam-
pling of the overall cohort rather than matched controls. In either
case, conducting metabolomics research using a case-control or
case-cohort study limits the use of the data to analyses of the disease
status of the case group or to subsetting the data for analyses within
the subcohort.
Physiologic or interventional studies involve the purposeful
manipulation of a parameter, such as medication or diet, or
performing some other manipulation of the study subject. These
types of studies provide a controlled setting in which to investigate
the metabolomic effects of the intervention, that is, to probe cause
and effect in a way that is not possible with observational studies. In
part, this advantage derives from the temporal relationship between
intervention and change in the metabolome. In addition, these
studies sometimes permit use of each participant as his or her own
biologic control, comparing the metabolome pre- and postinter-
vention, thus reducing the concerns about confounding intro-
duced by comparisons across different individuals.
One subtype of a physiologic or interventional study is a meta-
bolic challenge study. For example, the human metabolome
(HuMet) study was designed to assess the dynamics of the human
metabolome in 15 healthy, young men in response to different
metabolic challenges: 36-h fasting period, standard liquid diet,
oral glucose tolerance test, physical activity test, oral lipid tolerance
test, and cold stress test [22]. The investigators collected a variety
of biospecimens including plasma, urine, breath gas, and exhaled
breath condensate at multiple time points throughout each inter-
vention and conducted metabolomic profiling on the
406 Casey M. Rebholz and Eugene P. Rhee

biospecimens. They were able to demonstrate expected changes in

known metabolites (e.g., low levels of insulin and glucose during
fasting and an increase in their plasma levels 30 min after the oral
glucose tolerance test). The repeated measures of the metabolome
is a notable strength of this study, which, in the context of a
metabolic challenge allows one to attribute change in the metabo-
lome to the intervention as well as to determine the kinetics and
stability of these changes.
Feeding studies and other types of diet intervention studies are
a valuable design for understanding the effect of diet on cardiome-
tabolic outcomes. The food metabolome reflects metabolites that
are derived from food itself as well as those altered through the
metabolism of food by tissues and the microbiota [23]. Character-
izing the metabolomics of dietary intake, especially within feeding
studies, allows for the identification of biomarkers of food as well as
insights about diet-modifiable pathways leading to disease out-
comes [24]. For example, one group conducted a controlled cross-
over feeding trial in which 19 healthy volunteers received four
dietary interventions that varied in level of adherence to the
World Health Organization healthy eating guidelines for 3 days
each in a randomized order [25]. Metabolomic profiling was con-
ducted on biospecimens that were pooled from urine collected at
three time points each day. The investigators were able to develop
diet-specific urinary metabolite models which were then validated
in two independent study populations, that is, the INTERMAP UK
cohort (n ¼ 225) and a healthy eating Danish cohort (n ¼ 66).
These results provide proof-of-concept for metabolomic phenotyp-
ing of dietary intake, an objective approach that would vastly
improve upon currently used tools for assessing diet such as patient
recall and daily food diaries.
Another important type of intervention study is one that eval-
uates the effect of a medication on metabolites pertinent to cardio-
metabolic and renal outcomes. Of course, these kinds of studies
have already been conducted for individual metabolites such as
glucose, uric acid, and cholesterol, which are known to be causal
disease mediators. In theory, unbiased metabolomic approaches
could be used to provide a more nuanced understanding of meta-
bolic pathways that are affected by pharmacologic agents, particu-
larly if the mechanism of action is unknown or incompletely
elucidated. Once the metabolomic response to treatment has been
characterized, metabolomics could then be used to distinguish
between individuals who would be expected to respond to treat-
ment and individuals who would not be expected to respond to
treatment. Ultimately, conducting metabolomics in treatment trials
could inform a personalized medicine strategy of providing benefi-
cial treatment to responders and avoiding cost and harmful effects
by not prescribing such therapy to nonresponders [26]. These are
important avenues of future investigation in the field.
 Metabolomics for Renal and Cardiometabolic Diseases 407

Finally, in addition to a focused examination of how dietary or

pharmacologic inputs impact the metabolome, physiologic study
designs can seek to understand organ specific metabolite produc-
tion and handling. For example, one study profiled individuals who
underwent “planned” myocardial infarctions, that is, patients with
hypertrophic obstructive cardiomyopathy undergoing alcohol
induced ablation of their septal coronary arteries—this is an elective
surgical procedure that intentionally infarcts a portion of septal
muscle in patients whose thickened heart chambers obstruct opti-
mal flow of blood out of the heart [27]. In addition to enabling
precisely timed blood sampling, this study utilized invasive cathe-
terization of the coronary sinus, a venous channel that receives
blood from all of the coronary veins and empties into the right
atrium, permitting direct assessment of metabolite release from the
ischemic heart. In another example leveraging invasive catheteriza-
tion, metabolites were profiled in plasma obtained from the aorta
and renal vein, permitting direct assessment of the heterogeneous
impact of kidney function across the metabolome [28]. As
expected, this analysis showed that many metabolite levels are
substantially lower in the renal vein than the aorta because of
renal clearance; however, there was substantial heterogeneity in
the magnitude of decrease, attributable to the different chemical
characteristics of circulating metabolites (size, polarity, protein
binding, etc.) as well as the different mechanisms by which the
kidneys handle metabolites, including filtration, reabsorption,
secretion, and metabolism.

2.2 Statistics Multiple testing is a fundamental challenge in metabolomics

and the Challenge research. The large number of compounds detectable using cur-
of Multiple Testing rently available metabolomic platforms is an important strength,
permitting an expansive and unbiased discovery approach, with the
opportunity not just to identify single metabolites of interest but to
also highlight select metabolic pathways. However, the large num-
ber of metabolites assayed and tested statistically also raises impor-
tant analytical challenges, specifically an increase in the type one
error rate, the detection of false positive findings. There are a
variety of analytic approaches that can be used to account for
multiple testing, including adjusting the threshold for determining
statistical significance. One widely used approach is a Bonferroni
correction, where statistical significance is adjusted for the number
of metabolites measured (0.05/n), and if relevant the number of
outcomes examined as well. This can be overly conservative for
study designs that include both discovery analysis and independent
replication, particularly given the significant intercorrelation
between metabolite measurements and the ability to evaluate
metabolite alterations in the context of biochemical pathways.
A widely used alternative to Bonferroni correction is the Benja-
mini–Hochberg procedure (false discovery rate) [29, 30]. Some
408 Casey M. Rebholz and Eugene P. Rhee

investigators use data reduction techniques to reduce the number

of statistical tests, for example, by limiting the analysis to only
known metabolites (and not unknown metabolites or ion features).
An additional approach is to restrict analysis to metabolites with
relatively low intra-person, day to day variability, on the premise
that highly variable metabolites are unlikely to serve as reliable
disease biomarkers in the clinic. In one study of diabetic kidney
disease and progression to end stage kidney disease, the investiga-
tors only considered metabolites that had a correlation coeffi-
cient  0.4 on paired, fasting samples drawn several years apart,
narrowing the number of metabolites considered from 262 to
119 [31].
Traditional approaches to examining the association between
metabolites and outcomes examine each metabolite individually.
Given the high-dimensional nature of metabolomics data as well
as the intercorrelation among metabolites, more systems-based
approaches (covered in more detail elsewhere in the textbook)
that consider metabolites in aggregate and in the context of bio-
chemical pathways may be more appropriate depending on the
study of interest. Analytical approaches need to be aligned with
the overarching goal of the study. If the goal is to identify and
develop a clinically useful biomarker, then traditional regression
methods that model each metabolite individually and apply the
most stringent significance thresholds may be most appropriate. If
the goal is to use metabolomics data primarily for biologic insight
and as a springboard for laboratory-based interrogation, then
systems-based approaches with more permissive significance
thresholds may be most appropriate (and the full breadth of data,
including metabolites that have high intra-person variability, should
be analyzed). In either case, the concern for false positive discovery
can be ameliorated by increasing statistical power through
increased sample size, an option that will become more tractable
as the throughput and costs of metabolomics platforms continue to
improve. In clinical research of cardiometabolic and renal disease,
fortunately, there are large samples of individuals available through
medical registries and cohort studies who are available for metabo-
lomic profiling (although primarily for European and Caucasian
study populations). Enhancing sample size may remain a challenge
for less common diseases or for non-Caucasian study populations.

2.3 Controlling The measurement and adjustment for potential confounding fac-
for Confounding tors is a critical consideration in metabolomics studies of renal and
Factors cardiometabolic disease. Body mass index and diabetes status have
an enormous influence on the metabolome, in large part because
insulin action and sensitivity modulates blood levels of sugars,
amino acids, and lipids [32, 33]. Kidney function also has an
enormous influence on the metabolome, as highlighted by both
observational and physiologic studies described above. Age,
 Metabolomics for Renal and Cardiometabolic Diseases 409

gender, dietary habits, and other demographic factors also impact

the metabolome [34, 35]. Careful consideration of these factors is
essential in both study design, to ensure that they are catalogued
and measured, as well as in analysis, where models may need to be
adjusted to account for these potential confounders.
A discussion of confounding in kidney disease illustrates some
key challenges. As noted, the kidneys significantly impact the meta-
bolome, particularly small, polar molecules which are normally
excreted in urine. However, inverse correlation with GFR does
not necessarily indicate that a metabolite undergoes glomerular
filtration, as the kidneys also contribute to metabolite clearance
through reabsorption, secretion, and metabolism. In addition,
loss of kidney function can cause alterations in diet, gut microbiota,
insulin sensitivity, and other parameters that may impact the meta-
bolome [36]. Finally, a metabolite in theory could have an inverse
relationship with GFR if it contributes directly to the pathogenesis
of kidney disease, that is, it causes kidney damage. Hypothetically,
statistical approaches should vary depending on the questions of
interest. For studies that seek to identify novel causal mediators of
kidney disease, the association between metabolites and outcomes
(e.g., progression of kidney disease) should not necessarily be
adjusted for clinical measures of kidney function, as this could
attenuate or abrogate potentially meaningful signals. In practice,
however, almost all metabolomics studies of kidney disease to date
have adjusted for clinical measures of kidney function. In part, this
is because so many metabolites end up reaching statistical signifi-
cance without adjusting for kidney function, making it difficult to
prioritize select findings. In addition, this reflects the fact that most
investigators in the field have a background in clinical science and
epidemiology, with a focus on the discovery of clinically useful
biomarkers, which mandates rigorous adjustment to identify signals
with predictive power beyond GFR.
When adjusting for kidney function, the standard in the field is
to adjust for GFR and proteinuria. Multiple studies have been
conducted to examine which metabolites are cross-sectionally asso-
ciated with both measured GFR and estimated GFR [16, 37–
40]. There is comparatively less evidence examining the cross-
sectional association between blood metabolites and proteinuria.
One study has examined the cross-sectional association between
the metabolome and proteinuria as assessed using 24-h urine col-
lections in the African American Study of Kidney Disease and
Hypertension and the Modification of Diet in Renal Disease
study [41] and found that 58 metabolites (~9% of measured com-
pounds) were significantly associated with proteinuria after
accounting for demographic characteristics, serum albumin,
mGFR, and other clinical risk factors.
410 Casey M. Rebholz and Eugene P. Rhee

Medications are another major determinant of the metabo-

lome. Kidney disease patients as well as participants in other clinical
studies of preexisting cardiometabolic diseases are often treated
with medications for comorbid conditions such as hypertension,
diabetes, and dyslipidemia. However, self-report of medication use
is suboptimal, and there is variability in the type of medications
(even within specific classes of drugs) and degree to which patients
adhere to their prescribed medications. Metabolomics has the
potential to identify markers of drug use, including the drug itself
or one of its catabolites, but this has been a relatively underutilized
application in studies to date [42].
For all metabolomics studies, diet needs to be considered as a
potential confounder. As already noted, the metabolome is
impacted by dietary intake, reflecting breakdown products of
foods that are consumed, compounds that are metabolized in the
body by tissue and microbiota, and secondary effects of the hor-
monal responses triggered by caloric input [23, 24]. When the
nutritional metabolome is not of interest to the study question
but rather considered to be a confounding factor between the
metabolome and kidney disease status or risk, the investigator
needs to adjust for the fasting versus fed state. Study participants
will vary in terms of time since their last meal and with respect to the
composition of their diet. It may suffice to measure and adjust for
time since last meal. As an additional measure to account for
variability in food composition, one could adjust for an overall
diet quality score (e.g., the Healthy Eating Index) [43].

2.4 Specimen Type, The nature of samples used for metabolomics analysis requires
Storage, and Handling careful consideration. For large cohort studies that have completed
enrollment, samples will have already been collected with
corresponding demographic and clinical data. Access to these sam-
ples often requires direct collaboration with investigators oversee-
ing the cohort studies, or in some cases, can be obtained through
public channels (e.g., the National Institute of Diabetes and Diges-
tive and Kidney Diseases Central Repository or the National Heart,
Lung, and Blood Institute Biologic Specimen and Data Repository
Information Coordinating Center). In addition to large observa-
tional cohorts, biospecimens are also often collected for rando-
mized clinical trials of different pharmacologic or lifestyle
interventions, although access to these specimens may be restricted
in use. Finally, prospective collection by the investigator or a group
of collaborators is an attractive approach to sample accrual when
appropriate, for example for physiologic or interventional studies or
when rare or unusual study populations are required.
Whether samples have already been collected or will be col-
lected prospectively, it is important to recognize that procedures for
sample collection, processing, and storage can all impact the levels
of metabolites in a given biological specimen. First, all sampling
 Metabolomics for Renal and Cardiometabolic Diseases 411

approaches, along with the subsequent generation of metabolomics

data and integration with clinical data, need to be approved by the
relevant regulatory board for human research. Fasting status is
critical to determine for all samples. Timing of sample collection,
independent of feeding state, is also important as the role of circa-
dian oscillation on the metabolome is increasingly recognized
[44]. A few small studies have directly compared results between
serum and plasma, suggesting overall concordance in results,
although serum appears to have higher levels of some metabolites
including some reflective of the clotting process [45–48]. For
plasma, different anticoagulants may have different effects. For
example, citrate will clearly impact the measurement of citrate levels
and may interfere with the measurement of isomers, whereas hepa-
rin may impact lipids based on its ability to inhibit lipoprotein lipase
[49]. Samples that are shipped, stored for an extended period of
time, or subaliquoted may undergo freeze–thaw cycles which could
alter the composition of the metabolome. Previous studies have
reported on the stability of the metabolome at different storage
temperatures and after undergoing multiple freeze–thaw cycles
[50, 51]. For example, one study demonstrated stability of metab-
olite concentrations over four freeze–thaw cycles, but degradation
of samples was apparent when they were left at room temperature
for 12 h or longer before processing and storing [52]. Thus, inves-
tigators should be aware of exactly how samples for their study were
collected and processed, and the duration of time intervals between
phlebotomy and storage at 80  C. Stability of course also varies
depending on the metabolite, with amino acids and neutral lipids
demonstrating particular robustness across various sample handling
and storage conditions, whereas molecules with high energy phos-
phate bonds would be expected to dissipate more readily.
In theory, a range of biospecimens are amenable to metabolo-
mic profiling, including blood (plasma or serum), urine, saliva,
cerebral spinal fluid, stool, and even tissue homogenates. All have
been examined to date, but in practice blood and urine have been
studied the most extensively in renal and cardiometabolic disorders.
As such, a brief discussion of urine metabolomics is warranted. One
challenge with urinary analyses relates to the wide range of dilution,
resulting in metabolite levels that may fall above or below the linear
dynamic range of detection. To account for between-person varia-
bility in dilution, metabolite levels can be indexed to urine creati-
nine. Alternatively, urine collected over a 24-h period may be more
informative than spot urine which is collected at a random point in
time (during a clinical or study visit) since 24-h urine specimens
provide a cumulative measure of the metabolome which varies
diurnally. The choice of blood versus urine may depend on the
pathophysiologic aspect of kidney disease of interest. For example,
urine may be a more suitable medium for studying tubular secretion
or reabsorption [53]. The simultaneous measure of urine and
412 Casey M. Rebholz and Eugene P. Rhee

blood could provide unique insights about the transport of com-

pounds across kidney cells, that is, fractional excretion which is the
fraction of the metabolite filtered by the kidney that is ultimately
excreted, providing insight on metabolite absorption and secretion
[53]. However, the current availability of studies with metabolo-
mics results in both urine and blood is limited.

2.5 Replication Currently available metabolomic platforms, particularly untargeted

of Discovery Findings approaches, allow for the identification of many known metabolites
and many more unknown compounds or peaks/features. Follow-
ing such discovery work, it is necessary to replicate findings to
provide more confidence in their validity. One technique for vali-
dation is to randomly subset the data (e.g., two-thirds and
one-third samples) using one subset for discovery and the second
subset for validation. The difficulty with this technique is that one is
reducing statistical power by analyzing small datasets. If feasible, a
preferred alternative is to use an independent, external study popu-
lation to replicate the initial discovery findings. However, with an
external replication, lack of consistent results could be due to
differences between the discovery and replication cohorts. An addi-
tional challenge for replication across studies is nonoverlap of
metabolomic platforms with respect to hardware, procedures, qual-
ity control methods, and metabolite identification, thereby pre-
cluding the ability to systematically meta-analyze data generated
across disparate research groups. The Chronic Kidney Disease Bio-
markers Consortium recently conducted a methodological study in
which the same specimens were run on two nontargeted platforms
[14], yielding several observations. First, there was strong overlap
of known metabolites across platforms, particularly among abun-
dant metabolites such as amino acids and lipids, with generally
better coefficients of variation (CVs) across technical replicates for
these than the other known metabolites. Second, the exercise of
using different platforms on the same sample set can lead to the
expansion of known metabolite coverage for both platforms. Third,
agreement for the majority of overlapping measurements across
platforms was strong, suggesting that meta-analysis across plat-
forms is feasible, although there is a relative paucity of these kinds
of studies to date, in stark contrast to genomics and genome wide
association studies. Despite the areas of overlap, however, this study
also reinforced a general challenge in metabolomics, that no single
analytic approach is truly comprehensive and that careful under-
standing of the methods utilized is crucial. For kidney and cardio-
metabolic disease biomarker research, the simultaneous
deployment of different methodologies across different cohorts
may be prudent, to cast the widest net for discovery, and to mini-
mize the risk inherent in selecting one analytical approach.
 Metabolomics for Renal and Cardiometabolic Diseases 413

2.6 Data Analysis As metabolomics methods develop further, the analytic cost of
and Computational measuring the metabolome decreases, but the computational cost
Cost of Metabolomics of big data will remain, especially as the ability to run larger batches
Research and identify more metabolites increases. Key advances that have
allowed for more efficient measurement of the metabolome include
the use of robotics, automated data processing software for metab-
olite identification and quantification, and improvement in quality
control procedures [54]. As such, it is now possible to conduct
metabolomics in large-scale epidemiologic studies, yielding an
increasingly large amount of data in terms of sample size and
identified metabolites.
The most productive metabolomics projects providing mean-
ingful and translatable findings will be led by a team of scientists,
each offering their own expertise in basic science, physiology, clini-
cal medicine, bioinformatics, statistics, and epidemiology. This type
of team science requires the support by home institutions, to
recognize the need of infrastructure and career development of
young scientists. There is a need to train early career researchers
to be able to meaningfully analyze metabolomics data and to
appreciate both the physiological significance of detectable meta-
bolites as well as how to effectively use statistical techniques. Chap-
ters 14–16 in this book provide a foundation of data science and
statistics. Other chapters give more specific treatment of topics in
metabolomics research. Readers are also encouraged to visit the
supplemental website (https://metabolomics-data.github.io) for
more training materials.

3 Summary

In summary, the potential of metabolomics research of renal and

cardiometabolic diseases is great, with substantial progress made to
date. The majority of past studies have been observational in
nature, either cross-sectional or observational cohort studies. How-
ever, the identification of novel clinical biomarkers, elucidation of
new biology, and identification of actionable therapeutic targets
will require larger and more diverse studies permitting rigorous
validation and meta-analysis, as well as more physiologic and inter-
ventional study designs (e.g., to understand how diet and medica-
tions alter the relevant metabolic pathways and how these responses
impact disease risk and to assign organ specificity to metabolomic
alterations). When applying metabolomics to study questions in
kidney disease and related cardiometabolic diseases, it is important
to consider the impact of conducting many statistical tests on
statistical power, to control for confounding factors, to understand
details of specimen collection and handling, and to select a speci-
men type best suited for the study question. Given the broad range
414 Casey M. Rebholz and Eugene P. Rhee

of considerations, the application of metabolomics to clinical dis-

ease is best conducted by a multidisciplinary team with sufficient
technical, computational, and content-related expertise.

References
1. Thygesen K, Alpert JS, Jaffe AS, Chaitman BR, and the US Food and Drug Administration.
Bax JJ, Morrow DA, White HD, Executive Am J Kidney Dis 64(6):821–835
Group on behalf of the Joint European Society 7. Stevens LA, Coresh J, Greene T, Levey AS
of Cardiology/American College of Cardiol- (2006) Assessing kidney function--measured
ogy/American Heart Association/World and estimated glomerular filtration rate. N
Heart Federation Task Force for the Universal Engl J Med 354(23):2473–2483. https://doi.
Definition of Myocardial I (2018) Fourth uni- org/10.1056/NEJMra054415
versal definition of myocardial infarction 8. Stevens LA, Schmid CH, Greene T, Li L, Beck
(2018). Glob Heart 13(4):305–338. https:// GJ, Joffe MM, Froissart M, Kusek JW, Zhang
doi.org/10.1016/j.gheart.2018.08.004 YL, Coresh J, Levey AS (2009) Factors other
2. Anderson KM, Odell PM, Wilson PW, Kannel than glomerular filtration rate affect serum
WB (1991) Cardiovascular disease risk profiles. cystatin C levels. Kidney Int 75(6):652–660.
Am Heart J 121(1 Pt 2):293–298 https://doi.org/10.1038/ki.2008.638
3. Goff DC Jr, Lloyd-Jones DM, Bennett G, 9. HJL H, Greene T, Tighiouart H, Gansevoort
Coady S, D’Agostino RB, Gibbons R, RT, Coresh J, Simon AL, Chan TM, Hou FF,
Greenland P, Lackland DT, Levy D, O’Donnell Lewis JB, Locatelli F, Praga M, Schena FP,
CJ, Robinson JG, Schwartz JS, Shero ST, Levey AS, Inker LA, Chronic Kidney Disease
Smith SC Jr, Sorlie P, Stone NJ, Wilson PW, Epidemiology C (2019) Change in albumin-
Jordan HS, Nevo L, Wnek J, Anderson JL, uria as a surrogate endpoint for progression of
Halperin JL, Albert NM, Bozkurt B, Brindis kidney disease: a meta-analysis of treatment
RG, Curtis LH, De Mets D, Hochman JS, effects in randomised clinical trials. Lancet Dia-
Kovacs RJ, Ohman EM, Pressler SJ, Sellke betes Endocrinol 7(2):128–139. https://doi.
FW, Shen WK, Smith SC Jr, Tomaselli GF, org/10.1016/S2213-8587(18)30314-0
American College of Cardiology/American 10. Coresh J, Heerspink HJL, Sang Y,
Heart Association Task Force on Practice G Matsushita K, Arnlov J, Astor BC, Black C,
(2014) 2013 ACC/AHA guideline on the Brunskill NJ, Carrero JJ, Feldman HI, Fox
assessment of cardiovascular risk: a report of CS, Inker LA, Ishani A, Ito S, Jassal S,
the American College of Cardiology/American Konta T, Polkinghorne K, Romundstad S,
Heart Association task force on practice guide- Solbu MD, Stempniewicz N, Stengel B,
lines. Circulation 129(25 Suppl 2):S49–S73. Tonelli M, Umesawa M, Waikar SS, Wen CP,
https://doi.org/10.1161/01.cir. Wetzels JFM, Woodward M, Grams ME,
0000437741.48606.98 Kovesdy CP, Levey AS, Gansevoort RT,
4. Kidney Disease: Improving Global Outcomes Chronic Kidney Disease Prognosis C, Chronic
(KDIGO) (2013) KDIGO clinical practice Kidney Disease Epidemiology C (2019)
guideline for the evaluation and management Change in albuminuria and subsequent risk of
of chronic kidney disease. Kidney Int 3 end-stage kidney disease: an individual
(1):1–150 participant-level consortium meta-analysis of
5. Inker LA, Schmid CH, Tighiouart H, Eckfeldt observational studies. Lancet Diabetes Endo-
JH, Feldman HI, Greene T, Kusek JW, crinol 7(2):115–127. https://doi.org/10.
Manzi J, Van Lente F, Zhang YL, Coresh J, 1016/S2213-8587(18)30313-9
Levey AS, Chronic Kidney Disease Epidemiol- 11. Inker LA, Levey AS, Pandya K, Stoycheff N,
ogy Collaboration Investigators (2012) Esti- Okparavero A, Greene T, Chronic Kidney Dis-
mating glomerular filtration rate from serum ease Epidemiology C (2014) Early change in
creatinine and cystatin C. N Engl J Med 367 proteinuria as a surrogate end point for kidney
(1):20–29. https://doi.org/10.1056/ disease progression: an individual patient meta-
NEJMoa1114248 analysis. Am J Kidney Dis 64(1):74–85.
6. Levey AS, Inker LA, Matsushita K, Greene T, https://doi.org/10.1053/j.ajkd.2014.02.020
Willis K, Lewis E, De Zeeuw D, Cheung AK, 12. Waikar SS, Rebholz CM, Zheng Z, Hurwitz S,
Coresh J (2014) GFR decline as an endpoint Hsu CY, Feldman HI, Xie D, Liu KD, Mifflin
for clinical trials in CKD: a scientific workshop TE, Eckfeldt JH, Kimmel PL, Vasan RS, Bon-
sponsored by the National Kidney Foundation ventre JV, Inker LA, Coresh J, Chronic Kidney
 Metabolomics for Renal and Cardiometabolic Diseases 415

Disease Biomarkers Consortium I (2018) update. Hum Mol Genet 24(R1):R93–R101.

Biological variability of estimated GFR and https://doi.org/10.1093/hmg/ddv263
albuminuria in CKD. Am J Kidney Dis 72 20. Sekula P, Del Greco MF, Pattaro C, Kottgen A
(4):538–546. https://doi.org/10.1053/j. (2016) Mendelian randomization as an
ajkd.2018.04.023 approach to assess causality using observational
13. Schrimpe-Rutledge AC, Codreanu SG, Sher- data. J Am Soc Nephrol 27(11):3253–3265.
rod SD, McLean JA (2016) Untargeted meta- https://doi.org/10.1681/ASN.2016010098
bolomics strategies-challenges and emerging 21. Rhee EP, Clish CB, Wenger J, Roy J,
directions. J Am Soc Mass Spectrom 27 Elmariah S, Pierce KA, Bullock K, Anderson
(12):1897–1905. https://doi.org/10.1007/ AH, Gerszten RE, Feldman HI (2016) Meta-
s13361-016-1469-y bolomics of chronic kidney disease progres-
14. Rhee EP, Waikar SS, Rebholz CM, Zheng Z, sion: a case-control analysis in the chronic
Perichon R, Clish CB, Evans AM, Avila J, Den- renal insufficiency cohort study. Am J Nephrol
burg MR, Anderson AH, Vasan RS, Feldman 43(5):366–374. https://doi.org/10.1159/
HI, Kimmel PL, Coresh J, Consortium CKDB 000446484
(2019) Variability of two metabolomic plat- 22. Krug S, Kastenmuller G, Stuckler F, Rist MJ,
forms in CKD. Clin J Am Soc Nephrol 14 Skurk T, Sailer M, Raffler J, Romisch-Margl W,
(1):40–48. https://doi.org/10.2215/CJN. Adamski J, Prehn C, Frank T, Engel KH,
07070618 Hofmann T, Luy B, Zimmermann R,
15. Mahieu NG, Patti GJ (2017) Systems-level Moritz F, Schmitt-Kopplin P, Krumsiek J,
annotation of a metabolomics data set reduces Kremer W, Huber F, Oeh U, Theis FJ,
25000 features to fewer than 1000 unique Szymczak W, Hauner H, Suhre K, Daniel H
metabolites. Anal Chem 89 (2012) The dynamic range of the human meta-
(19):10397–10406. https://doi.org/10. bolome revealed by challenges. FASEB J 26
1021/acs.analchem.7b02380 (6):2607–2619. https://doi.org/10.1096/fj.
16. Sekula P, Goek ON, Quaye L, Barrios C, Levey 11-198093
AS, Romisch-Margl W, Menni C, Yet I, 23. Scalbert A, Brennan L, Manach C, Andres-
Gieger C, Inker LA, Adamski J, Gronwald W, Lacueva C, Dragsted LO, Draper J, Rappaport
Illig T, Dettmer K, Krumsiek J, Oefner PJ, SM, van der Hooft JJ, Wishart DS (2014) The
Valdes AM, Meisinger C, Coresh J, Spector food metabolome: a window over dietary expo-
TD, Mohney RP, Suhre K, Kastenmuller G, sure. Am J Clin Nutr 99(6):1286–1308.
Kottgen A (2016) A metabolome-wide associ- https://doi.org/10.3945/ajcn.113.076133
ation study of kidney function and disease in 24. Guasch-Ferre M, Bhupathiraju SN, Hu FB
the general population. J Am Soc Nephrol 27 (2018) Use of metabolomics in improving
(4):1175–1188. https://doi.org/10.1681/ assessment of dietary intake. Clin Chem 64
ASN.2014111099 (1):82–98. https://doi.org/10.1373/
17. Ho JE, Larson MG, Ghorbani A, Cheng S, clinchem.2017.272344
Chen MH, Keyes M, Rhee EP, Clish CB, 25. Garcia-Perez I, Posma JM, Gibson R, Cham-
Vasan RS, Gerszten RE, Wang TJ (2016) bers ES, Hansen TH, Vestergaard H,
Metabolomic profiles of body mass index in Hansen T, Beckmann M, Pedersen O,
the Framingham heart study reveal distinct Elliott P, Stamler J, Nicholson JK, Draper J,
Cardiometabolic phenotypes. PLoS One 11 Mathers JC, Holmes E, Frost G (2017) Objec-
(2):e0148361. https://doi.org/10.1371/jour tive assessment of dietary patterns by use of
nal.pone.0148361 metabolic phenotyping: a randomised, con-
18. Newgard CB, An J, Bain JR, Muehlbauer MJ, trolled, crossover trial. Lancet Diabetes Endo-
Stevens RD, Lien LF, Haqq AM, Shah SH, crinol 5(3):184–195. https://doi.org/10.
Arlotto M, Slentz CA, Rochon J, Gallup D, 1016/S2213-8587(16)30419-3
Ilkayeva O, Wenner BR, Yancy WS Jr, 26. Tolstikov V (2016) Metabolomics: bridging
Eisenson H, Musante G, Surwit RS, Millington the gap between pharmaceutical development
DS, Butler MD, Svetkey LP (2009) A and population health. Metabolites 6(3):E20.
branched-chain amino acid-related metabolic https://doi.org/10.3390/metabo6030020
signature that differentiates obese and lean 27. Lewis GD, Wei R, Liu E, Yang E, Shi X,
humans and contributes to insulin resistance. Martinovic M, Farrell L, Asnani A, Cyrille M,
Cell Metab 9(4):311–326. https://doi.org/ Ramanathan A, Shaham O, Berriz G, Lowry
10.1016/j.cmet.2009.02.002 PA, Palacios IF, Tasan M, Roth FP, Min J,
19. Kastenmuller G, Raffler J, Gieger C, Suhre K Baumgartner C, Keshishian H, Addona T,
(2015) Genetics of human metabolism: an Mootha VK, Rosenzweig A, Carr SA, Fifer
MA, Sabatine MS, Gerszten RE (2008)
416 Casey M. Rebholz and Eugene P. Rhee

Metabolite profiling of blood from individuals among Asian Indians living in the United
undergoing planned myocardial infarction States have distinct Metabolomic profiles that
reveals early markers of myocardial injury. J are associated with Cardiometabolic risk. J
Clin Investig 118(10):3503–3512. https:// Nutr 148(7):1150–1159. https://doi.org/
doi.org/10.1172/JCI35111 10.1093/jn/nxy074
28. Rhee EP, Clish CB, Ghorbani A, Larson MG, 36. Ramezani A, Raj DS (2014) The gut micro-
Elmariah S, McCabe E, Yang Q, Cheng S, biome, kidney disease, and targeted interven-
Pierce K, Deik A, Souza AL, Farrell L, tions. J Am Soc Nephrol 25(4):657–670.
Domos C, Yeh RW, Palacios I, Rosenfield K, https://doi.org/10.1681/ASN.2013080905
Vasan RS, Florez JC, Wang TJ, Fox CS, Gersz- 37. Coresh J, Inker LA, Sang Y, Chen J, Shafi T,
ten RE (2013) A combined epidemiologic and Post WS, Shlipak MG, Ford L, Goodman K,
metabolomic approach improves CKD predic- Perichon R, Greene T, Levey AS (2018) Meta-
tion. J Am Soc Nephrol 24(8):1330–1338. bolomic profiling to improve glomerular filtra-
https://doi.org/10.1681/ASN.2012101006 tion rate estimation: a proof-of-concept study.
29. Curtin F, Schulz P (1998) Multiple correla- Nephrol Dial Transplant 34(5):825–833.
tions and Bonferroni’s correction. Biol Psychi- https://doi.org/10.1093/ndt/gfy094
atry 44(8):775–777 38. Titan SM, Venturini G, Padilha K, Tavares G,
30. Benjamini Y, Hochberg Y (1995) Controlling Zatz R, Bensenor I, Lotufo PA, Rhee EP,
the false discovery rate - a practical and power- Thadhani RI, Pereira AC (2019) Metabolites
ful approach to multiple testing. J R Stat Soc B related to eGFR: evaluation of candidate mole-
57(1):289–300 cules for GFR estimation using untargeted
31. Niewczas MA, Sirich TL, Mathew AV, metabolomics. Clin Chim Acta 489:242–248.
Skupien J, Mohney RP, Warram JH, Smiles A, https://doi.org/10.1016/j.cca.2018.08.037
Huang X, Walker W, Byun J, Karoly ED, Ken- 39. Goek ON, Doring A, Gieger C, Heier M,
sicki EM, Berry GT, Bonventre JV, Koenig W, Prehn C, Romisch-Margl W,
Pennathur S, Meyer TW, Krolewski AS Wang-Sattler R, Illig T, Suhre K, Sekula P,
(2014) Uremic solutes and risk of end-stage Zhai G, Adamski J, Kottgen A, Meisinger C
renal disease in type 2 diabetes: metabolomic (2012) Serum metabolite concentrations and
study. Kidney Int 85(5):1214–1224. https:// decreased GFR in the general population. Am J
doi.org/10.1038/ki.2013.497 Kidney Dis 60(2):197–206. https://doi.org/
32. Warren B, Rebholz CM, Sang Y, Lee AK, 10.1053/j.ajkd.2012.01.014
Coresh J, Selvin E, Grams ME (2018) Diabetes 40. Ng DP, Salim A, Liu Y, Zou L, Xu FG,
and trajectories of estimated glomerular filtra- Huang S, Leong H, Ong CN (2012) A meta-
tion rate: a prospective cohort analysis of the bolomic study of low estimated GFR in
atherosclerosis risk in communities study. Dia- non-proteinuric type 2 diabetes mellitus. Dia-
betes Care 41(8):1646–1653. https://doi. betologia 55(2):499–508. https://doi.org/
org/10.2337/dc18-0277 10.1007/s00125-011-2339-6
33. Bell EK, Gao L, Judd S, Glasser SP, 41. Luo S, Coresh J, Tin A, Rebholz CM, Appel
McClellan W, Gutierrez OM, Safford M, Lack- LJ, Chen J, Vasan RS, Anderson AH, Feldman
land DT, Warnock DG, Muntner P (2012) HI, Kimmel PL, Waikar SS, Kottgen A, Evans
Blood pressure indexes and end-stage renal dis- AM, Levey AS, Inker LA, Sarnak MJ, Grams
ease risk in adults with chronic kidney disease. ME, Chronic Kidney Disease Biomarkers Con-
Am J Hypertens 25(7):789–796. https://doi. sortium I (2019) Serum Metabolomic altera-
org/10.1038/ajh.2012.48 tions associated with proteinuria in CKD. Clin
34. Yin X, Subramanian S, Willinger CM, Chen G, J Am Soc Nephrol 14(3):342–353. https://
Juhasz P, Courchesne P, Chen BH, Li X, doi.org/10.2215/CJN.10010818
Hwang SJ, Fox CS, O’Donnell CJ, 42. Guo L, Milburn MV, Ryals JA, Lonergan SC,
Muntendam P, Fuster V, Bobeldijk-Pastorova I, Mitchell MW, Wulff JE, Alexander DC, Evans
Sookoian SC, Pirola CJ, Gordon N, AM, Bridgewater B, Miller L, Gonzalez-Garay
Adourian A, Larson MG, Levy D (2016) ML, Caskey CT (2015) Plasma metabolomic
Metabolite signatures of metabolic risk factors profiles enhance precision medicine for volun-
and their longitudinal changes. J Clin Endocri- teers of normal health. Proc Natl Acad Sci U S
nol Metab 101(4):1779–1789. https://doi. A 112(35):E4901–E4910. https://doi.org/
org/10.1210/jc.2015-2555 10.1073/pnas.1508425112
35. Bhupathiraju SN, Guasch-Ferre M, Gadgil 43. Krebs-Smith SM, Pannucci TE, Subar AF,
MD, Newgard CB, Bain JR, Muehlbauer MJ, Kirkpatrick SI, Lerman JL, Tooze JA, Wilson
Ilkayeva OR, Scholtens DM, Hu FB, Kanaya MM, Reedy J (2018) Update of the healthy
AM, Kandula NR (2018) Dietary patterns eating index: HEI-2015. J Acad Nutr Diet
 Metabolomics for Renal and Cardiometabolic Diseases 417

118(9):1591–1602. https://doi.org/10. metabolomic profiles in samples obtained dur-

1016/j.jand.2018.05.021 ing cardiac catheterization. Circ Cardiovasc
44. Grant LK, Ftouni S, Nijagal B, De Souza DP, Genet 4(6):695–700. https://doi.org/10.
Tull D, McConville MJ, Rajaratnam SMW, 1161/CIRCGENETICS.111.960575
Lockley SW, Anderson C (2019) Circadian 50. Gika HG, Theodoridis GA, Wilson ID (2008)
and wake-dependent changes in human plasma Liquid chromatography and ultra-performance
polar metabolites during prolonged wakeful- liquid chromatography-mass spectrometry fin-
ness: a preliminary analysis. Sci Rep 9 gerprinting of human urine: sample stability
(1):4428. https://doi.org/10.1038/s41598- under different handling and storage condi-
019-40353-8 tions for metabonomics studies. J Chromatogr
45. Breier M, Wahl S, Prehn C, Fugmann M, A 1189(1–2):314–322. https://doi.org/10.
Ferrari U, Weise M, Banning F, Seissler J, 1016/j.chroma.2007.10.066
Grallert H, Adamski J, Lechner A (2014) Tar- 51. Teahan O, Gamble S, Holmes E, Waxman J,
geted metabolomics identifies reliable and sta- Nicholson JK, Bevan C, Keun HC (2006)
ble metabolites in human serum and plasma Impact of analytical bias in metabonomic stud-
samples. PLoS One 9(2):e89728. https://doi. ies of human blood serum and plasma. Anal
org/10.1371/journal.pone.0089728 Chem 78(13):4307–4318. https://doi.org/
46. Wedge DC, Allwood JW, Dunn W, Vaughan 10.1021/ac051972y
AA, Simpson K, Brown M, Priest L, Blackhall 52. Anton G, Wilson R, Yu ZH, Prehn C,
FH, Whetton AD, Dive C, Goodacre R (2011) Zukunft S, Adamski J, Heier M, Meisinger C,
Is serum or plasma more appropriate for inter- Romisch-Margl W, Wang-Sattler R, Hveem K,
subject comparisons in metabolomic studies? Wolfenbuttel B, Peters A, Kastenmuller G,
An assessment in patients with small-cell lung Waldenberger M (2015) Pre-analytical sample
cancer. Anal Chem 83(17):6689–6697. quality: metabolite ratios as an intrinsic marker
https://doi.org/10.1021/ac2012224 for prolonged room temperature exposure of
47. Lin Z, Zhang Z, Lu H, Jin Y, Yi L, Liang Y serum samples. PLoS One 10(3):e0121495.
(2014) Joint MS-based platforms for compre- https://doi.org/10.1371/journal.pone.
hensive comparison of rat plasma and serum 0121495
metabolic profiling. Biomed Chromatogr 28 53. Kottgen A, Raffler J, Sekula P, Kastenmuller G
(9):1235–1245. https://doi.org/10.1002/ (2018) Genome-wide association studies of
bmc.3152 metabolite concentrations (mGWAS): rele-
48. Ishikawa M, Tajima Y, Murayama M, Senoo Y, vance for nephrology. Semin Nephrol 38
Maekawa K, Saito Y (2013) Plasma and serum (2):151–174. https://doi.org/10.1016/j.
from nonfasting men and women differ in their semnephrol.2018.01.009
lipidomic profiles. Biol Pharm Bull 36 54. Tzoulaki I, Ebbels TM, Valdes A, Elliott P,
(4):682–685 Ioannidis JP (2014) Design and analysis of
49. Brunner MP, Shah SH, Craig DM, Stevens metabolomics studies in epidemiologic
RD, Muehlbauer MJ, Bain JR, Newgard CB, research: a primer on -omic technologies. Am
Kraus WE, Granger CB, Sketch MH Jr, Newby J Epidemiol 180(2):129–139. https://doi.
LK (2011) Effect of heparin administration on org/10.1093/aje/kwu143
 Chapter 21

Using the IDEOM Workflow for LCMS-Based Metabolomics

Studies of Drug Mechanisms
Anubhav Srivastava and Darren J. Creek

Abstract
Rapid advancements in metabolomics technologies have allowed for application of liquid chromatography
mass spectrometry (LCMS)-based metabolomics to investigate a wide range of biological questions. In
addition to an important role in studies of cellular biochemistry and biomarker discovery, an exciting
application of metabolomics is the elucidation of mechanisms of drug action (Creek et al., Antimicrob
Agents Chemother 60:6650–6663, 2016; Allman et al., Antimicrob Agents Chemother 60:6635–6649,
2016). Although it is a very useful technique, challenges in raw data processing, extracting useful informa-
tion out of large noisy datasets, and identifying metabolites with confidence, have meant that metabolomics
is still perceived as a highly specialized technology. As a result, metabolomics has not yet achieved the
anticipated extent of uptake in laboratories around the world as genomics or transcriptomics. With a view to
bring metabolomics within reach of a nonspecialist scientist, here we describe a routine workflow with
IDEOM, which is a graphical user interface within Microsoft Excel, which almost all researchers are familiar
with. IDEOM consists of custom built algorithms that allow LCMS data processing, automatic noise
filtering and identification of metabolite features (Creek et al., Bioinformatics 28:1048–1049, 2012). Its
automated interface incorporates advanced LCMS data processing tools, mzMatch and XCMS, and
requires R for complete functionality. IDEOM is freely available for all researchers and this chapter will
focus on describing the IDEOM workflow for the nonspecialist researcher in the context of studies
designed to elucidate mechanisms of drug action.

Key words Metabolomics, IDEOM, Drug mechanism, Mode of action, Microsoft Excel, LCMS,
Data processing

1 Introduction

The main challenges in untargeted metabolomics studies include

the low signal to noise ratio, metabolite identification and compar-
ative data visualization. IDEOM attempts to address all these issues
through its user-friendly graphical user interface built in Microsoft
Excel. The MS Excel spreadsheet operates through a number of
VBA macros that allow automated data processing of high-
resolution LCMS data. These macros allow for automatic genera-
tion and execution of scripts in the VBA and R environment [1]

Shuzhao Li (ed.), Computational Methods and Data Analysis for Metabolomics, Methods in Molecular Biology, vol. 2104,
https://doi.org/10.1007/978-1-0716-0239-3_21, © Springer Science+Business Media, LLC, part of Springer Nature 2020

419
420 Anubhav Srivastava and Darren J. Creek

(www.r-project.org) using adjustable parameters. Using powerful

inbuilt tools for LCMS data processing—mzMatch [2] and XCMS
[3]—it generates a list of putative metabolites with associated
confidence levels and peak intensities, with intuitive data visualiza-
tion and standard statistical tests. This chapter will provide a step-
by-step guide for the use of IDEOM to investigate the mode of
action (MoA) of a drug.
The main approaches used during the early stages of drug
discovery are target-based screens and phenotypic screens. While
target-based screens are a popular approach to identify small mole-
cules that will act by a predefined mode of action, phenotypic
screening has proved to be a successful avenue for the identification
of many drug candidates [4]. For example, in the field of infectious
diseases it can be very efficient to screen large chemical libraries for
their ability to inhibit pathogen growth, giving rise to a large
number of hit compounds that can be triaged and optimized to
provide lead drug candidates [5, 6]. The obvious issue with drug
discovery based on phenotypic screens has been limited informa-
tion about the drug target responsible for the observed phenotype.
The knowledge of the mode of action or target is vital as it can
provide a structural basis for chemical optimization of the hit
compound (e.g., for improving binding), it can underpin the ratio-
nal design of combination therapies and can also help in monitoring
for resistance to the therapeutic compounds in the clinical setting.
Metabolomics studies have been shown to help in finding the
mode of action of novel and existing pharmaceutical compounds
[7–11]. Typically, this involves in vitro incubation of the relevant
cell population with the investigational drug, followed by untar-
geted metabolomics analysis. However, care must be taken to
ensure that the metabolic profile observed following drug incuba-
tion is representative of the specific action of the drug. While higher
drug concentrations and longer incubations may lead to a more
definitive assay outcome (e.g., resulting in cell death/growth
arrest), these conditions will likely lead to off-target affects, thus
confounding identification of the metabolic perturbation primarily
responsible for the action of the drug. On the other hand, low
concentrations and/or short incubations may not allow for suffi-
cient pharmacological activity to an extent whereby the metabolic
profile can be differentiated from the untreated control. Therefore,
a balance must be achieved both in duration and concentration of
the drug in MoA studies using metabolomics, whereby the optimal
conditions induce specific perturbations to metabolism without
causing cell death. Ideally, some simple activity assays should be
conducted with escalating drug exposure to determine the highest
achievable sublethal concentration and duration conditions that
will be suitable for metabolomics analysis. The addition of time-
course, concentration-ranging, and negative control (e.g., inactive
chemical analogues and compounds that kill by distinct
 IDEOM for Drug Mechanism 421

mechanisms) samples also improves interpretation of the link

between specific metabolic responses and mechanisms of drug
action. Consideration should also be given to the sampling condi-
tions, as the biomass requirements for metabolomics analysis (e.g.,
>1 million cells/sample for most human cell lines) may differ from
standard activity assay conditions, and the metabolism must be
quenched for analysis in a short but practical time window which
will allow experimental reproducibility.
In this chapter, we present an excerpt from a study [7] where
LCMS-based metabolomics data was analyzed using the IDEOM
workflow to find the mode of action of antimalarial compounds.
This study used compounds from the Malaria Box, an open-access
collection of “hit” antimalarial compounds that were identified by
phenotypic screening and assembled by the Medicines for Malaria
Venture [12]. The lack of information on the mode of action of
these compounds is a significant hurdle to further development of
these hits. This study used simple in vitro incubation of cultures of
the malaria parasite, Plasmodium falciparum, with the test com-
pounds followed by LCMS-based metabolomics analysis using
IDEOM, and generated MoA information which will be crucial
for future drug discovery efforts based on these hits.

2 Materials and Methods

2.1 Cell Culture Trophozoite stage asexual P. falciparum (3D7) parasites (causative
and Drug Incubations agents of malaria) were cultured using a previously defined method
for Metabolomics [13], using human RBCs (obtained from the Australian Red Cross
Analysis Blood Service) at 7–8% parasitemia (iRBC) and 3% hematocrit in
RPMI 1640 with hypoxanthine and 0.5% Albumax (Gibco) at
37
C under a specialized gas mix (95% N2, 4% CO2, 1% O2). For
drug-induced metabolic perturbation experiments, 200 μL iRBCs
were incubated with 1 μM of test compounds for 5 h in 96-well
plates in quadruplets. This short incubation with test compounds
was intended to induce metabolic perturbations in the malaria
parasite without causing significant mortality in the cell population,
as P. falciparum in vitro activity assays are usually performed using a
48 – 72-h incubation.

2.2 Metabolite Following sublethal drug incubation, culture medium was removed
Extraction and the metabolism of cells was quenched by resuspending the cells
for Metabolomics in ice-cold PBS. All the subsequent steps were performed on ice.
Analysis Cells were washed by centrifugation at 1000  g for 5 min and PBS
was removed. Metabolite extraction was performed by adding
135 μL methanol (containing internal standards) followed by
rapid mixing. Samples were then gently mixed for 1 h at 4
C and
then centrifuged at 3000  g for 5 min to precipitate proteins and
other insoluble molecules. The supernatant, which contained the
422 Anubhav Srivastava and Darren J. Creek

metabolite extract, was transferred to glass LC-MS vials and stored

at 80
C before LC-MS analysis. A 10 μL aliquot of each sample
was pooled together to generate a quality control (QC) sample
which was used to monitor sample stability and analytical
reproducibility.

2.3 LC-MS Based The method of choice for this study was LC-MS, using hydrophilic
Metabolomics interaction liquid chromatography (HILIC) and high-resolution
Analysis (Orbitrap) mass spectrometry (MS). Briefly, Samples (10 μL) were
injected onto a Dionex RSLC U3000 LC system (Thermo) fitted
with a ZIC-pHILIC column (5 μm, 4.6 by 150 mm; Merck).
20 mM ammonium carbonate (A) and acetonitrile (B) were used
as the mobile phases with a 30 min gradient starting from 80% B to
40% B over 20 min, followed by washing at 5% B for 3 min and
reequilibration at 80% B. Mass spectrometry utilized a Q-Exactive
MS (Thermo) with a heated electrospray source operating in posi-
tive and negative modes (rapid switching) and a mass resolution of
35,000 from m/z 85 to 1050. Before each batch of samples, blanks
and mixtures of authentic standards (234 metabolites) were ana-
lyzed, and 2 pooled extracts were analyzed in data-dependent
tandem mass spectrometry (MS/MS) mode to facilitate down-
stream metabolite identification where necessary. Pooled QC sam-
ples were analyzed periodically throughout each batch.

2.4 LC-MS Data Data was analyzed using IDEOM pipeline as described in this
Analysis Using IDEOM chapter (see Subheading 3). Briefly, raw files were converted to
mzXML with msconvert [14], LC-MS peak signals were extracted
with the Centwave algorithm in XCMS [15], samples were aligned
and artifacts were filtered with mzMatch [2] and additional data
filtering and feature identification was performed based on accurate
mass and retention time [16]. The parameters used in automated
IDEOM filtering for this study are demonstrated in Fig. 1, and the
annotated peak metadata and chromatogram images allowed fur-
ther manual data filtering to remove features that were not reliably
detected across replicates and those with very low-quality chro-
matographic peaks. This resulted in reduction of the 15,000
detected LC-MS peaks, which includes many noise and electrospray
artifacts, to a manageable 460 reproducible peaks with reasonable
putative identifications. Metabolite identification (level 1 confi-
dence according to the Metabolomics Standards Initiative (MSI)
[17] corresponded to confidence score of 10 in IDEOM based on
accurate mass and retention time for metabolites that were present
in the standard mixture. Other features were putatively annotated
(MSI level 2) based on accurate mass and predicted retention time
using the IDEOM database. Metabolite abundance was deter-
mined by peak height and normalized to the average for untreated
samples from the same batch. Univariate statistical analyses utilized
Welch’s t test (α ¼ 0.05) and Pearson’s correlation (Microsoft
 IDEOM for Drug Mechanism 423

Fig. 1 Parameters used for automated IDEOM filtering and feature identification from raw metabolomics data
in the atovaquone mode of action study [7]

Excel) and multivariate analysis was performed by using the Meta-

bolomics package in R [18] through IDEOM.

2.5 Mode of Action Although 100 compounds were tested in this study, for the pur-
of Atovaquone poses of this chapter we focused on the mode of action of a known
Confirmed by IDEOM- antimalarial drug, atovaquone. The comparison sheet (see IDEOM
Based Analysis processing) showed the metabolite intensities of drug-treated sam-
ples expressed relative to the “control” group, which was DMSO
(vehicle control) in this study (IDEOM file with entire data set can
be downloaded from the supplemental material for this study [7]).
As expected [19], specific changes in the levels of metabolites
involved in pyrimidine biosynthesis were observed in metabolite
extracts from parasites treated with atovaquone. The observed
increase in levels of N-carbamoyl-L-aspartate and dihydroorotate
in atovaquone-treated cultures, and the decrease in downstream
424 Anubhav Srivastava and Darren J. Creek

Fig. 2 Top panel: Snapshot of the comparison sheet highlighting the metabolites involved in pyrimidine
biosynthesis in the atovaquone mode of action study [7]. Bottom panel: Schematic representation of the
pyrimidine biosynthesis pathway in malaria parasite and the step inhibited by atovaquone (ATV: Atovaquone,
DHODH: dihydroorotate dehydrogenase)

metabolites orotate, UMP, and UDP, compared to DMSO control

(Fig. 2) was indicative of inhibition of dihydroorotate dehydroge-
nase (DHODH) activity. This is consistent with the known mecha-
nism of action of atovaquone involving the primary inhibition of
ubiquinol oxidation by the cytochrome bc1 complex [19], which
provides the essential cofactor for DHODH. This indirect inhibi-
tion of P. falciparum DHODH by atovaquone has been indepen-
dently genetically validated as the primary mechanism of action of
atovaquone [20]. Furthermore, this specific impact of atovaquone
on pyrimidine pathway metabolites has been demonstrated in two
independent studies that analyzed these metabolites using targeted
LCMS analysis [8, 21].

3 IDEOM Workflow

3.1 Installation The MS Excel-based IDEOM template can be downloaded from

http://mzmatch.sourceforge.net/ideom.html. It does not require
any installation, and can be loaded directly in a recent version of
Microsoft Excel (2010 or later). For full functionality, users require
 IDEOM for Drug Mechanism 425

Fig. 3 Screenshot of the IDEOM template as it appears when the file is first opened. Areas circled in red
indicate the location of I: “Save As” button, II: “Enable Content” button to allow macros to run, III: Button for
installation of the necessary R packages, IV: Adjustable parameters for IDEOM, mzMatch and XCMS functions,
V: Cells to enter names of the internal standards, VI: worksheets containing metabolite database, retention
time and fragment information. See information in the “Getting Started” section for more details

a current installation of R (www.r-project.org) and proteowizard

tools, which can be downloaded from http://proteowizard.
sourceforge.net/downloads.shtml. Hyperlinks to Xcalibur
(Thermo) and web browsers are included in the IDEOM template;
however, these programs are not required for successful implemen-
tation of IDEOM.

3.2 Getting Started Figure 3 gives a screenshot of the IDEOM template as it appears
when the file is first opened.
1. Open the IDEOM.xlsb file in MS Excel and “Save As” with
your study name (e.g., mystudy1.xlsb).
2. Enable Macros by clicking the “Enable content” button in the
“Security Warning” ribbon.
426 Anubhav Srivastava and Darren J. Creek

3. (First time only) Install all required R packages by clicking the

appropriate blue button in the help section at the bottom of the
page. (You may need to scroll down to row 57 to find this). The
blue buttons execute functions in R (not within Excel); you
may be prompted to select the “Rgui.exe” file on your com-
puter. Please agree to any Windows security warnings.
4. All settings for XCMS, mzMatch, and IDEOM processing are
located in column E.
5. (optional) Up to 10 internal/external standards may be
entered into cells U2:AD2.
6. (optional) Red sheets contain advanced settings and databases.

3.3 General Ensure to save the IDEOM template as a macro-enabled workbook

Information in “.xlsb” Excel binary format. It is recommended to save the file
before running each step. Macros are activated by clicking the
colored buttons. In-cell hyperlinks in Excel are activated by
double-click and web links are activated by single-click. The
IDEOM template includes a number of Excel “worksheets”
(Fig. 3) whose features are described in Table 1.

3.4 Processing Steps The raw data processing steps can be executed by clicking the
corresponding blue buttons at the top-left of the settings sheet
3.4.1 Raw Data
(Fig. 4)
Processing
1. Manually sort files into folders according to study group.
(a) Create a working directory which is not within “My Docu-
ments” or “Program Files,” create folders for each group
of replicates in the study (control, treatment, blank, QC,
etc.) and move relevant raw files into these folders.
(b) This step may be skipped if the user wishes to process all
files without grouping.
(c) Avoid spaces in folders and filenames. It is recommended
to replace space with underscore in the names. For exam-
ple; use “IDEOM_trial” rather than “IDEOM trial.”
2. Convert RAW to mzXML files and split polarity.
(a) Run step (b) (click blue button) if dealing with raw files.
(Skip this step if you already have mzxml files).
(b) This step uses msconvert (or ReAdW), through R, to
convert raw LCMS files to the .mzXML format. Mscon-
vert is recommended.
(c) The location of msconvert.exe (or ReAdW.exe) on your
computer needs to be assigned, and can be stored in cell
E43.
(d) msconvert.exe needs to be in the same folder as the Pwiz
dll files (keep all Pwiz files together). (ReAdW requires
zlib1.dll to be in your windows directory to work
correctly.)
 IDEOM for Drug Mechanism 427

Table 1
Description of worksheets in the IDEOM file

Sheet Description
Settings Home page and the starting point for all analysis. It contains many basic settings,
help documentation and the main macro buttons for processing data. Default
settings are suitable for 4.6 mm ZIC-pHILIC chromatography (with
ammonium carbonate/ACN, 0.3 mL/min) coupled to the Q-Exactive
Orbitrap
MZmatch Blank sheet to allow import of the peak list from mzMatch (or other tools)
Alldata Information about every peak set in the peak list is written to this sheet
Rejected Peak sets with putative identification, but confidence below 5, are copied to this
sheet. Most of these peaks are noise or artifacts, but users may wish to scan
them manually for false rejections
Identification All identified metabolites with confidence of 5 or above are copied to this sheet
allBasepeaks All base peaks (from mzMatch related peaks function) are copied to this sheet
unless the peaks are not significant in any group (less than blank)
Comparison Comparative peak intensity data is placed here after running the “compare all
sets” macro. Many functions for evaluation and visualization are available on
this sheet
DB This sheet contains the full metabolite database and associated metadata. If
required, additional metabolites may be added to the bottom of the list, or
additional property columns to the right of existing columns. Additional
information may also be added to existing columns, but do not insert new
columns between existing data
RTcalculator and These contain important tables of information required for the IDEOM macros.
fragments Advanced users may wish to update these tables with instrument-specific values
Targeted This sheet allows a targeted analysis of specific metabolites
Tools This sheet demonstrates the utilization of IDEOM’s user-defined Excel
functions, suitable for manual calculations such as mass determinations
Method and samples Extra sheets to allow uploading of experiment metadata

(e) TIC CHECKER: This function is available in the R script

menu at the bottom of the page and allows users to run
this additional script in R. It is not essential, but provides a
preliminary opportunity to check signal reproducibility of
the mzXML files.
3. Run XCMS (and mzmatch) to pick peaks and convert to
peakML files.
(a) This step uses the Centwave function in xcms (through R)
to pick peaks, and then mzmatch converts each individual
file to peakml format.
428 Anubhav Srivastava and Darren J. Creek

Fig. 4 Screenshot from the settings sheet with the raw data processing steps (blue buttons) and IDEOM data
processing steps (green buttons)
 IDEOM for Drug Mechanism 429

Table 2
List of automated mzMatch functions

mzMatch function Description

Group Groups peaks across replicate samples based on specified mass and RT windows
(optional, requires files to be assigned to group folders before starting)
RSD filter (Relative Standard Deviation) removes features with high variability in peak
intensity for replicates within a group (optional, requires previous group
function)
Combine Combines all groups/samples into one peakml file (peak list)
Noise filter Peak shape is assessed by CoDA-DW score (0–1)
Intensity filter Features are removed if no sample has a peak above the intensity threshold
Detections filter Peaks must be present in a minimum number of samples
Gap-filler Fills gaps (missing values) in the peak intensity table where peaks were detected in
some samples but not others. This requires access to the raw data (mzxml files) to
extract the relevant ion signals from the given m/z and retention time windows
Related peaks Annotates peaks that are likely to be ESI artifacts (e.g., isotopes, adducts (Na+, K+,
annotation Cl, ACN, etc.), fragments, multiply charged species, dimers, multimers,
complex adducts orFT artifact signals) based on retention time, peak shape and
correlation of peak intensities across samples
Output Generates output files in peakml format and a peak list table in .txt format

(b) Settings for the Centwave function can be changed in cells

E4–E11 of the Settings sheet — for full documentation
see xcms help (metlin.scripps.edu/xcms).
(c) Briefly, the parameters are:
l PPM: mass deviation from scan to scan.
l Peakwidth: range for baseline peak width.
l S/N threshold: Minimum signal to noise ratio.
l Prefilter: number of scans greater than a given intensity
threshold.
4. Run MZmatch to combine data and annotate related peaks.
This step uses mzMatch to group individual peakml files,
filter peaks and annotate related peaks. Automated mzMatch
functions are listed in Table 2.
Settings can be changed in cells E14–E20, and E25 on the
Settings sheet, additional parameters can be altered by saving
the R script and running it externally — for documentation see
mzmatch help (mzmatch.sourceforge.net). The default set-
tings should be a good starting point, though you might
need to change RSD filter (E16) if the data is variable, or
Minimum Detections # (E19) if there are less than 3 replicates
in any group. Gapfiller requires mzXML files to be in the same
430 Anubhav Srivastava and Darren J. Creek

location as when they were converted to peakml files (step c),

therefore it is a good idea to run steps (c) and (d) together
(on the same computer) by using the “COMBINED” button.
Gapfiller is very memory intensive so if the process crashes at
this step, try closing the R session and continuing the script
manually from the gapfiller stage. If this does not work then
tighten the filters (RSD, min detections) to make smaller
peakml files. Another option would be to run it on a computer
with a higher RAM capacity.

3.5 IDEOM The main IDEOM data processing steps and associated parameters
Processing are found on the settings sheet (Fig. 4). While most IDEOM
processing is automated, some study-specific input is required
from the user. Optimal results are achieved by clicking steps 1–9
(green buttons) and following the on-screen prompts. Steps 1 and
2 can be run in any order, but must both be completed before
running the main processing macro (step 3)
1. Import MZmatch data and enter grouping info.
(a) Use this function to import data from the mzMatchout-
put.txt file produced by mzMatch. (If you have already
entered data manually, or by the “import example data” or
“import Mzmine data” buttons, you may press cancel at
the “import file” screen to skip this step.)
(b) The second part of this function asks the user to enter
grouping information. “Autofill” can be used if the prefix
of the sample names refers to the grouping, otherwise
manually select groups using the “add” buttons.
(c) Set-Type needs to be selected for each group using the
drop-down lists (Table 3). Always set one group as “Treat-
ment” and another as “Control” to allow statistical
comparisons.
(d) If there are more than 15 sample groups, there is a second
tab with space for 15 more groups. IDEOM currently
supports a maximum of 30 groups (Fig. 5).
(e) The third part of this function plots average sample inten-
sities to allow a quick check of whether the data is consis-
tent. Internal (external) standards will also be plotted if
details have been entered in cells U2-AD2 of the settings
sheet.
(f) The fourth part gives the option of normalizing the data
either by TIC, median, or user-defined values (column R
on settings sheet). Normalization is not routinely recom-
mended for LCMS data due to nonlinear responses and
the unpredictability of ion-suppression.
 IDEOM for Drug Mechanism 431

Table 3
Description of set types (groups) in the settings sheet

Set types
(groups) Description
Blank Solvent blank(s), used as the background reference for the groups significance filter
Control Base group for intensity comparisons. Only one control can be set at any time. If you
wish to compare results to multiple controls please use “‘Save As” to get different
Excel files corresponding to each control when you run the “‘Compare All Sets”
macro
Treatment Initial 2-sample intensity comparison, and used for the QC: RSD if no QC is present.
Only one treatment can be set at any time
Sample Any real sample groups that are not selected as the initial “control” or “treatment”
group should be set as “‘Sample.” Undefined groups are assumed to be “sample”
QC This group is used for the RSD filter and is included in graphs of individual samples,
but not in the comparison of results
Standards and These groups are excluded from all functions, but the peak intensity data is retained
Exclude in the data matrix on all sheets

Fig. 5 Screenshot of the window where group information can be entered while
importing the mzMatch data. Refer to Table 3 for description of Group Types
432 Anubhav Srivastava and Darren J. Creek

Fig. 6 Retention time prediction model on the RTcalculator sheet, showing

values associated with the current multiple linear regression model, and a plot
of the model fit for the current metabolite standards

(g) The grouping functionality in IDEOM assumes that repli-

cates in the mzMatch output data are in adjacent columns.
This sample ordering should automatically occur if using
the mzMatch preprocessing workflow described above,
otherwise the column order should be adjusted in the
txt file prior to import.
2. Update DB with Retention Times.
(a) This function enters standard retention times into the
database (DB sheet), and (optional) enters predicted
retention times for other metabolites [22].
(b) A list of retention times for the particular LCMS method,
obtained from analysis of authentic standards (ideally in
 IDEOM for Drug Mechanism 433

Table 4
List of RSD parameters for each group of replicates

RSD
parameters Description
Strict If any group has RSD larger than this setting, the metabolite is considered unreliable and
given a low confidence score (0.4)
Technical If the QC group (or Treatment group in the absence of a QC group) has RSD larger than
this setting, the metabolite is given a low confidence score (0.4)
Generous No specific RSD filter is applied, although if no group has an RSD below the filter
threshold then the peak is not significant and has confidence level of 0
Exclusions Groups with mean intensity below the LOQ threshold are excluded from this test
(as they often have higher RSD)

the same batch), is required (create this list either using

the Targeted Sheet in IDEOM, or externally using tar-
geted analysis software such as Toxid, Tracefinder, or
Xcalibur).
(c) The list of standard RTs may be either imported from any
Excel-readable file, or entered directly into columns A and
B on the “RTcalculator” sheet.
(d) If importing .csv files of retention times: “_” in metabolite
names will be replaced with “,”.
(e) All authentic standards (column A) must have names that
exactly match those in the DB sheet.
(f) RT calculator uses physicochemical properties in the DB
sheet to predict retention times based on a multiple linear
regression model with the authentic standards (QSPR
approach [22]).
(g) You have the opportunity to check the model fit before
annotating all metabolites in the database (Fig. 6).
(h) If there is no good prediction model you can still use this
function to upload standard retention times for those
metabolites where you have authentic standards.
(i) RTcalculator sheet parameters:
l The column (cell O8) and dead volume (mins) should
be entered before running this macro (cell O9). Other
data in columns N-U show the accuracy of the current
RT prediction model.
l The mass range for application of the prediction model
is defined in cells O10 and Q10. The default model,
“pHILIC_Scotmet”, is suitable for the ZIC-pHILIC
method with ammonium carbonate and acetonitrile
mobile phase (see LCMS method in the example
above) for metabolites within the MW range 70–400.
434 Anubhav Srivastava and Darren J. Creek

Table 5
Description of columns in Identification, Rejected, All data, and all Base peaks sheets

Column Description
A Neutral exact mass (from mzMatch)
B Retention time (from mzMatch) in minutes
C Formula from DB with closest match to mass (if within ppm window)
D Number of isomers in DB with this exact formula
E Metabolite name: Best match from DB for this mass and RT (bold type if there is a
standard RT for this metabolite in the DB)
F Confidence level (arbitrary out of 10) according to parameters on “settings” sheet
G Records whether the metabolite is in a “preferred database” (from DB)
H Map: The general area of metabolism for this metabolite (usually from KEGG)
Note: Columns G and H can be changed by choosing a different header in cell G1 or H1
I Mass error (in ppm) from nearest match in DB
J RT error relative to authentic standard (name is bold) or predicted RT as % of RT (NB:
Where no predicted RT exists in database the cell is colored red)
K Alt ppm: Mass error for the next closest mass in the DB (if within ppm window) — If font
is red it is a double-charged match, if background is colored it is an adduct match — See
settings D36:F38 for adduct colors
L Groups: Records which sample groups have significant peaks detected for this metabolite
(peaks > blanks, and RSD <RSDfilter)
M BP: Basepeak (if identified) for that peak (from mzMatch: Largest coeluting peak with
correlated peak shape and intensity trend across samples) NB: This does not update
when other identifications are changed to a different isomer
N Mzdiff: Mass difference between this peak and the basepeak
O Relationship: Relationship to the basepeak (according to mzMatch)
P Addfrag: Common adduct, isotope, fragment or neutral-loss: Based on filters and formulas
the “fragments” sheet. Possible complex adducts of larger coeluting peaks
(in duplicatepeaks window) are also annotated here. Note: a “complex adduct” may
sometimes be the parent of two fragment ions
Q % error of detected 13C-isotope intensity from the theoretical 13C-isotope intensity (note:
This relies on filtered peaks, and does not go back to the raw data)
R Related peaks: A list of common related peaks (isotopes, multicharged, adducts,
fragments) that the macro has detected for this metabolite
S RSD (relative standard deviation) for QC samples (or for treatment group if no QC is
assigned)
T Maximum RSD for all included sample groups
U Maximum intensity from all included samples
V Relation id (from mzMatch)

(continued)
 IDEOM for Drug Mechanism 435

Table 5
(continued)

Column Description
W Peak intensity ratio for mean of “treatment” group vs. mean of “control” group
X P-value for unpaired t-test between “treatments” and “controls”
Y Adduct of formula match to mass (H, Na, double-charge, etc.)
Z Polarity (in combined files the first-named polarity is that with the biggest peak. All sample
intensities in combined file are taken from the polarity with the biggest peak for each
metabolite)
AA Number of detected peaks in included groups
NEXT # Peak intensities from all samples present on the mzmatch sheet
Extra Other functions such as “compare with medium,” “compare 2 other groups,” and
columns “isotope search” will add extra columns. Users may add additional columns to the right
without affecting macro performance

l Columns W:X allow standard retention times to be

uploaded to the database without being included in
the prediction model (e.g., for large metabolites out-
side the validated mass range).
l Columns Z:AI allow predicted retention times to be
uploaded to the database based on class properties,
according to specific annotations in the DB sheet
(Default values are for the pHILIC method above).
l Headers E1:J1 can be adjusted to other phys-chem
properties (from drop-down menus) if a different RT
prediction model is required (select “manual”
column).
(j) If this function is not implemented then metabolite iden-
tification will only be based on exact mass (not retention
time), hence there will be many more false-identifications.
3. Run Identification Macro.
Before running this macro, check all settings in column E
of the Settings sheet. It is worth noting that the settings for
RSD filter, intensity filter, minimum detections and related
peaks window (under mzMatch settings, cells E16:E20) are
also used in IDEOM functions.
(a) RT and mass windows (E23:E25) may need to be
changed depending on the quality of the data.
(b) Preferred DB (E33) should be chosen from the drop-
down list. “Central” refers to central pathways in KEGG
(Cofactors and vitamins, amino acid, nucleotide, carbo-
hydrate, lipid, and energy metabolism). Additional
436 Anubhav Srivastava and Darren J. Creek

Table 6
Description of columns in comparisons sheet

Column Description
A Neutral exact mass (from mzMatch)
B Retention time (from mzMatch) in minutes
C Formula from DB with closest match to mass (if within ppm window)
D Number of isomers in DB with this exact formula
E Metabolite name: Best match from DB for this mass and RT (bold type if there is a standard
RT for this metabolite in the DB)
F Confidence level (arbitrary out of 10) according to parameters on “settings” sheet
G Map: The general area of metabolism for this metabolite (usually from KEGG)
H Pathway: List of biochemical pathways for this metabolite (usually from KEGG)
I Max intensity: Maximum LCMS intensity in any sample for that metabolite. Note:
Columns G, H, and I can be changed by choosing a different header in cell G1, H1, or I1
J Mean intensity of each included group relative to the “control” group (as set when the
“comparison” macro was run). Significant (t-test) values are in bold
Next # P-values for unpaired t-test between each included group and the control
Next # Mean intensity for each included group
Next # Standard deviation for each included group
Next # Relative standard deviation for each included group
Next # Fisher ratio for each included group, relative to the control group
Additional Additional columns are added every time you sort by correlation of intensity trends relative
to a specific metabolite. You may add your own additional columns to the right of existing
data without affecting performance. Please do not insert columns between existing data

organisms/annotations may be added to the database

using the purple “Annotate DB” macro, or by manually
appending a column to the end of the DB sheet (column
BN).
(c) Additional adducts (other than H+/H) can be searched
by selecting them in cells D36:F38 on the settings sheet.
(mzMatch has already corrected the data for 1 proton). It
is recommended to select “Na” and “NH3” in positive
mode data, “Cl” and “Formic acid” in negative mode
data, and “double-charge” for both ionization polarities.
(d) This macro runs most of the IDEOM functions to filter
data and identify peaks, data is recorded in the “alldata”
sheet, and then copied to the other results sheets.
The following list of functions is applied to each peak:
 IDEOM for Drug Mechanism 437

Table 7
List of shortcuts obtained by double-clicking in columns of Identification, Rejected and all Base
peaks sheets

Column Description of shortcut

A (Mass) Plots the mass spectrum for all coeluting peaks. Red peaks are RelatedPeaks (from
mzMatch). Hover mouse over lines to get annotation information
B (Retention time) plots intensity and RT of all isomeric peaks for this mass from all samples.
There is also the option for a quick link to look at EIC for a specific RAW file in Xcalibur, or
to produce EIC(s) of this mass from a peakml file
C (Formula) checks all possible adducts for formulas, then gives option to copy a mass for you to
paste into either Xcalibur or ChemSpider search engine, or run a process in R to find possible
chemical formulae
D (Isomers) shows a list of isomers in DB with this exact formula, also shows % retention time
error, databases and pathways (from DB)
E (Name) plots a graph of intensities for each sample. Also, single-click gives a dropdown list that
enables you to select any isomer from the DB with the selected formula (data in columns F:J
will change depending on this selection)
F (Confidence) - In rejected sheet only - looks to see if this metabolite has been detected in the
“identification” sheet. If present, the retention time of each peak is shown
G (PreferredDB) — If hyperlinks activated — Searches ChemSpider for the metabolite name
(search engine can be changed in the hyperlinks table, e.g., PubChem or Google)
H (DB) — If hyperlinks activated — Searches online databases for the metabolite (e.g., KEGG,
HMDB, Metacyc, or Lipidmaps)
J (RT% err) shows a list of isomers in DB with this exact formula, also shows % retention time
error, databases, and pathways (same as column D)
K (altPPM) shows information about metabolites with the alternate formula that gives this ppm
error
W (TvsCTRL) plots a graph of intensities for each sample on a log scale

Table 8
List of shortcuts obtained by double-clicking in columns of Comparison sheet

Column Description of shortcut

A–E Same as identification sheet (see Table 7 above)
F (Confidence) gives all the information about this metabolite from the “identification” sheet
G (Map) — If hyperlinks activated — Searches for metabolite in ChemSpider (search engine can
be changed in the hyperlinks table, e.g., PubChem or Google)
H (Pathway) — If hyperlinks activated — Searches for metabolite in online databases
I (Groups) gives a plot of individual sample intensities
438 Anubhav Srivastava and Darren J. Creek

(a) Group detection vs. noise: If a peak is present for every

sample in a group and each peak is greater than every peak
in the blanks, and the RSD is below the RSD filter
threshold, then the metabolite detection is considered
significant and that group name is appended to the
“groups” column (L). Peaks that are not significant in
any group are assigned the lowest confidence (0).
(b) Charge: If 13C isotope is present, the detected charge-
state is recorded and taken into consideration for all mass
(or m/z)-dependent functions.
(c) The RSD (relative standard deviation) is calculated for
each group of replicates. The maximum RSD is entered in
column T, and the RSD for the QCs is entered in column
S. If no QCs are assigned then the RSD for the “Treat-
ment” group is entered in column S. The RSD para-
meters are listed in Table 4.
(d) Intensity filter: The maximum intensity for all included
samples is entered in column U. If this is less than the
Intensity filter (LOQ) then a low confidence is given
(0.5).
(e) Formula assignment: based on the closest match to the
database (within mass window) — if 2 matches within
specified mass window the mass error for the furthest
formula is noted in “altppm” column (K).
(f) Other adducts: All Basepeaks (see mzmatch) are checked
for other “search adducts” according to the user selec-
tions in cells D36:F38 on the settings sheet — If a for-
mula has already been assigned, the mass error for the
adduct is noted in the “altppm” column (K).
(g) Metabolite identification: for all formulae is based on
(a) standard retention time (if present), or (b) predicted
retention time (if present).
(h) Isomers: If metabolite assignment remains ambiguous,
the first matching metabolite in the DB is assigned.
NOTE: DB has been sorted to give preference to the
selected “preferred DB,” then to “central” metabolites,
then to metabolites present in more source databases
(KEGG, MetaCyc, etc.), then to lowest metabolite ID
number in source databases.
(i) Fragments: Known fragments of specific metabolites are
removed (according to the table in “Fragments” sheet
columns Q:U, from standards analyzed on the HILIC-
Orbitrap system).
(j) Other related peaks: Related peaks determined by
mzMatch are annotated in the “addfrag” column
 IDEOM for Drug Mechanism 439

(P) according to possible mass differences in the “Frag-

ments” sheet (columns A:C). Annotations in Column C
of the “Fragments” sheet designate whether specific
adducts/fragments are automatically rejected with confi-
dence level (0.4) or merely annotated with this informa-
tion about a potential relationship.
(k) Isotopes: 13C Isotope abundances are recorded in column
Q as the % error relative to the theoretical isotope abun-
dance for the putative metabolite formula, or for uniden-
tified peaks, as the theoretical number of carbon atoms in
the formula.
(l) Confidence: Arbitrary confidence levels are assigned based
on retention time error, or annotation as “Xenobiotic” in
“Map” column of DB sheet. Confidence is modified by
whether the metabolite is in the “preferred” database,
and mzMatch annotation as either “base” or “related”
peak. See columns N:O of settings sheet for more details.
Confidence levels for each filter may be adjusted if
desired.
(m) Filtering: All metabolites with a confidence score of 5 or
greater are copied to the “Identification” sheet, those
below 5 are copied to the “rejected” sheet.
(n) Duplicate annotations: If duplicates remain with the same
identification, the identity of the smaller peak is changed
to the next most likely isomer that satisfies the retention-
time window.
(o) Additional calculations: detailed information is entered
into columns A:AA for each metabolite.
4. Manually move any false rejections from “Rejected” list to
“Identification” list (optional).
This is an optional step for an expert user to thoroughly
interrogate the list of putatively identified metabolites that
were rejected by the IDEOM algorithms. If you believe a
rejected metabolite should be retained in the data-set, dou-
ble-click the “Confidence” value to see if that metabolite is
already in the identification sheet. Wrongly rejected metabo-
lites can be moved to the “Identification” sheet by using the
“Retrieve Row” function (Table 5).
5. Recalibrate mass (ppm).
This macro can be started from either this settings sheet, or
from the button at the top of the “identification” sheet. It will
plot the relationship between mass and mass accuracy (ppm
error) for all “identified” metabolites, with authentic standards
in red, and fit a fifth order polynomial function (this should
allow for the calibration errors observed on Thermo Orbitrap).
If the polynomial function appears to fit the data, agree to
440 Anubhav Srivastava and Darren J. Creek

recalibrate masses. If the curve is not a good fit, but you see a
trend, consider manual recalibration efforts. After calibration,
check the new plot of mass errors, and set a new ppm window
to remove outliers (false-identifications). In some cases, it is
worth checking the rejected peaks (bottom of “rejected” list)
for alternative identifications by clicking the “altppm” column.
If significant adjustment was made (e.g., >2 ppm) consider
rerunning analysis from step 3 (with the recalibrated data) to
improve identification.
6. Manually check related peaks and isomers (optional).
For thorough analysis, check all the information supplied
for each peak in the “identification” sheet. Another common
approach is to skip this step initially, and return later to double-
check specific metabolites of interest. While it is always a good
idea to return to raw data for confirmation of specific metabo-
lites, the identification sheet allows rapid access to a large
amount of meta-information to simplify the process of manual
data curation and metabolite identification.
(a) Isomers: Sort by the Formula column and remove dupli-
cates that appear to be due to poor chromatography (click
a cell in the RT column (B:B) for a plot of relative inten-
sities and retention times of all isomeric peaks).
(b) Click the light-blue “add chromatograms” button to add
cropped peak chromatogram images for each putative
metabolite (click or hover in column A to view).
(c) Related Peaks: Click the orange “Sort By Relation ID”
button and look in columns M:P for ESI artifacts that
were not filtered by the common adducts search (e.g.,
Metabolite-specific fragments). Double-click in the mass
(column A) to see all coeluting peaks—red masses are
“related” according to mzMatch.
(d) Also look at “C isotope error” (column Q) to see if the
main isotope intensity does not match the formula,
Related Peaks (column R) to see if all the fragments and
isotopes are possible from the identified structure, and
“adduct” (column Y) to check whether the identified
adduct (e.g., 2+, Na) is likely. (The “charge” column at
the end may help this if a 13C isotope was present, but it is
sometimes not correct in noisy data).
(e) Alternative identification: Easily change identification by
clicking the metabolite name (column E), then select from
the isomers in the dropdown list. Further information
about these isomers can be obtained by double-clicking
the isomer number (column D). Alternative formulas can
be investigated for individual masses by double-clicking
the formula (column C).
 IDEOM for Drug Mechanism 441

(f) Use the “Remove row” button to easily transfer metabo-

lites from the “identification” sheet to the “rejected”
sheet. There is an option to merge peaks, which would
be used for compounds with poor peak shape (e.g., lipids
on a HILIC column) that appear with the same mass but
slightly different retention times. If merging peaks it keeps
the largest intensity from each sample.
(g) To speed up viewing data you may set Excel’s calculation
to “manual” (Formulas  Calculation Options), but be
aware that you need to manually calculate if you make any
changes. Some macros will automatically return calcula-
tion to “Automatic” (Formulas  Calculation Options).
(h) There is an option to normalize data at this stage (in the
Tools button menu; purple button) to allow normaliza-
tion based on (putatively) identified metabolite peaks,
rather than normalizing on total LC-MS signal or peaks
(which includes much more noise).
(i) Columns W and X allow a rapid comparison of relative
intensity and p-value (from t-test) between the 2 groups
specified as “control” and “treatment.” Other 2-group
comparisons may be generated from the “Tools” menu.
7. Combine Pos and Neg modes (optional).
If data is acquired in switching polarity mode (or in both
polarity modes separately with identical chromatography), data
from each polarity must be processed separately up to this
point. This step combines all data from the positive and nega-
tive mode IDEOM files. If a metabolite gives peaks in both
polarities (same formula and RT within “duplicatepeaks” win-
dow setting), data from the polarity with the highest maximum
intensity is selected.
8. Compare all sets.
This macro takes the data from the “identification” sheet
and summarizes it in the “Comparison” sheet, providing the
most convenient presentation of data for biological interpreta-
tion (examples in Tables 6, 7 and 8). An option exists to also
include all significant BasePeaks (regardless of their identifica-
tion status) to allow statistical analysis of unidentified metabo-
lites. Note that the metabolite names in the “identification”
sheet are dependent on the name in the “comparison” sheet, so
if a peak in the “comparison” sheet is renamed to an alternative
isomer, it will be reflected in the “identification” sheet. The
following functions are available in the “Comparison” sheet:
(a) Relative Intensities for all metabolites in all groups are
expressed relative to the “control” group (from the Set-
tings sheet).
442 Anubhav Srivastava and Darren J. Creek

(b) P values are for unpaired t-test against “control” group;

mean peak intensities, standard deviations, Relative Stan-
dards Deviations (RSD), and Fisher ratios are also
provided.
(c) Correlation Sort allows the metabolite list to be sorted by
intensity correlation (across samples) relative to a speci-
fied metabolite.
(d) The Sort function provides a quick link to sort by any
column (e.g., intensity, pathway, p-value).
(e) Excel’s native Sort and Autofilter functionality should
also be used to improve data visualization.
(f) Columns G and H can be changed to any variable in the
DB sheet.
(g) Column I can be changed to any variable in the Identifi-
cation sheet.
(h) Double-clicking in column E (name) gives an intensity
plot (with standard deviations) for that metabolite.
Double-clicking in column I gives individual sample
intensities.
(i) Double-clicking in column D (Isomers) gives informa-
tion about the isomers in the database.
(j) Double-clicking in column F (confidence) gives informa-
tion from the “identification” sheet that helps with deter-
mination of the identity confidence.
(k) Various summary plots are available from the orange
“Graphs” button.
(l) Various functions are available from the purple “Tools”
button.
(m) Various export formats are available from the light blue
“Export” button.
9. Assign BasePeaks (optional).
This macro attempts to improve the identification of
unknown BasePeaks by looking at the current “Identification”
and “Rejected” sheets to update the information in the “allBa-
sePeaks” sheet to reflect any changes since the initial identifica-
tion step. It also annotates any “related peaks” if the
“basepeak” itself could not be identified.

4 Other Tools Available Within IDEOM

A range of additional metabolomics data analysis tools are available

within IDEOM and can be accessed through the “Tools,”
“Graphs,” and “Export” menus (Table 9). Two of the more
 IDEOM for Drug Mechanism 443

Table 9
Additional functions are provided which may be used in any cell just like the native Excel functions

Excel functions Description

Exact mass fx ¼ ExactMass (formula, Clabels (optional), Nlabels (optional), Olabels
(optional), Dlabels (optional))
Returns the exact mass of a given formula. Only works for the following
atoms: C, H, N, O, S, P, Cl, F, I, Br, and Se. Optional arguments allow
exact mass calculation for isotopically labeled compounds with the
specified number of 13C, 15N, 18O, or deuterium atoms
Mass error (ppm) calculation fx ¼ ppmcalc (mass, Theoretical mass (optional), formula (optional))
Calculates the mass difference (in ppm) between a given mass and a
theoretical formula or mass
Formula reactor fx ¼ FormulaReactor (Formula1, Formula2 (optional), formula loss
(optional))
Returns the formula that results from the addition of the 2 input formulae.
If the reaction involves the loss of a second product (e.g., H2O), this
must be entered as the “formulaloss.” Alternatively, this function can
determine the fragment resulting from the loss of “formulaloss” from
“Formula1.” only works for the following atoms: C, H, N, O, S, P,
Cl, F, I, and Br. (e.g., Formula Reactor (C6H12O6, H3PO4,
H2O) ¼ C6H13O9P)
Formula match from exact fx ¼ Formula MATCH (mass, ppm, Mass list, Formula list)
mass Finds a matching formula in a database of ascending masses (e.g., the DB
sheet). Mass list and Formula list need to be selected as columns in a
database. If two masses either side of the search mass are within the
allowable ppm error the answer is italicized
Formula validity check fx ¼ Formulavalid (formula)
Checks the validity of a proposed chemical formula against 5 of Kind &
Fiehn’s 7 golden rules (excluding isotope and TMS rules) [23]
Theoretical isotope fx ¼ IsotopeAbundance (formula, atom)
abundance calculator Calculates the theoretical natural isotope abundance of a specified atom in
a given formula. Only works for 13C, 2H, 15N, 18O, 34S, 37Cl, 81Br
Positive charge (average) fx ¼ Pos (pH, cation, pKa1, pKa2 (optional), pKa3 (optional), pKa4
(optional), pKa5 (optional))
Calculates the average number of positive charges on a molecule at a given
pH, based on the formal positive charge (cation) and a list of basic pKa
values
Negative charge (average) fx ¼ Neg (pH, anion, pKa1, pKa2 (optional), pKa3 (optional), pKa4
(optional), ..., pKa8 (optional))
Calculates the average number of negative charges on a molecule at a
given pH, based on the formal negative charge (anion) and a list of
acidic pKa values

commonly used additional tools are for targeted analysis of a set of

metabolites, and for the analysis of data from stable-isotope label-
ing experiments.
444 Anubhav Srivastava and Darren J. Creek

4.1 Targeted Targeted analysis for specific metabolites can be undertaken from
Analysis the “Targeted” sheet by following the process indicated by the
buttons at the top of the page.
1. Enter the Metabolite names in Column A.
2. Click step 2 to upload information for each metabolite from
the metabolite database (DB sheet). For unique metabolites
and/or masses enter the RT, formula and/or mass directly.
3. This step uses msconvert/mzMatch (through R) to process
either raw, mzXML or peakML files and filters the results
according to the specified masses. Parameters are taken from
the Settings sheet.
4. The results from step III are searched to return results for the
specific metabolites to the targeted page. Alternatively, any text
file from mzMatch, or the data on the mzMatch sheet of
IDEOM, can be used for this step. If multiple peaks are present
it first takes the most intense peak within the RT window
setting for standard RT (settings sheet; cell E23), then the
most intense peak within the RT window setting for calculated
RT (settings sheet; cell E24). Metabolites with no expected RT
are assigned the peak with the largest intensity.
5. The targeted analysis is commonly used to obtain retention
times for authentic standards, and hence there is an option to
export these results to the RT calculator which is used in the
regular untargeted IDEOM processing method.

4.2 Isotope Search This tool provides an untargeted search for labeled metabolites in
data that was obtained using stable-isotope labeled precursors. Run
this from the “tools” menu in “comparison,” “allbasepeaks” or
“Identification” sheets to search for isotopes of putatively identified
compounds. The isotope search macro looks within the RT win-
dow for related peaks. The isotope search result is the relative
abundance, that is, the ratio of the isotope peak to the unlabeled
peak in each specified sample.
The result is shaded yellow if the relative abundance is more
than 10% greater than the expected abundance of the natural
isotope of the unlabeled metabolite. This macro will only find
isotopes if they were imported to IDEOM from mzMatch. In
some cases, isotopes are missed because the peaks were not detected
by XCMS. To conduct a more comprehensive analysis of isotopes in
raw data use the “Targeted Isotopes” macro in the “export” menu.

References
1. Team RC (2014) R: a language and environ- 2. Scheltema RA, Jankevics A, Jansen RC, Swertz
ment for statistical computing. R Foundation MA, Breitling R (2011) PeakML/mzMatch: a
for Statistical Computing, Vienna, Austria file format, Java library, R library, and tool-
 IDEOM for Drug Mechanism 445

chain for mass spectrometry data analysis. Anal 14. Chambers MC, Maclean B, Burke R,
Chem 83:2786–2793 Amodei D, Ruderman DL, Neumann S,
3. Smith CA, Want EJ, O’Maille G, Abagyan R, Gatto L, Fischer B, Pratt B, Egertson J et al
Siuzdak G (2006) XCMS: processing mass (2012) A cross-platform toolkit for mass spec-
spectrometry data for metabolite profiling trometry and proteomics. Nat Biotechnol
using nonlinear peak alignment, matching, 30:918–920
and identification. Anal Chem 78:779–787 15. Tautenhahn R, Bottcher C, Neumann S
4. Moffat JG, Vincent F, Lee JA, Eder J, Prunotto (2008) Highly sensitive feature detection for
M (2017) Opportunities and challenges in high resolution LC/MS. BMC Bioinformatics
phenotypic drug discovery: an industry per- 9:504
spective. Nat Rev Drug Discov 16:531–543 16. Creek DJ, Jankevics A, Burgess KE,
5. Gamo FJ, Sanz LM, Vidal J, de Cozar C, Breitling R, Barrett MP (2012) IDEOM: an
Alvarez E, Lavandera JL, Vanderwall DE, excel interface for analysis of LC-MS-based
Green DV, Kumar V, Hasan S et al (2010) metabolomics data. Bioinformatics
Thousands of chemical starting points for anti- 28:1048–1049
malarial lead identification. Nature 17. Sansone SA, Fan T, Goodacre R, Griffin JL,
465:305–310 Hardy NW, Kaddurah-Daouk R, Kristal BS,
6. Hovlid ML, Winzeler EA (2016) Phenotypic Lindon J, Mendes P, Morrison N et al (2007)
screens in antimalarial drug discovery. Trends The metabolomics standards initiative. Nat
Parasitol 32:697–707 Biotechnol 25:846–848
7. Creek DJ, Chua HH, Cobbold SA, Nijagal B, 18. De Livera AM, Dias DA, De Souza D,
Macrae JI, Dickerman BK, Gilson PR, Ralph Rupasinghe T, Pyke J, Tull D, Roessner U,
SA, McConville MJ (2016) Metabolomics- McConville M, Speed TP (2012) Normalizing
based screening of the malaria box reveals and integrating metabolomics data. Anal Chem
both novel and established mechanisms of 84:10768–10776
action. Antimicrob Agents Chemother 60 19. Biagini GA, Fisher N, Shone AE, Mubaraki
(11):6650–6663 MA, Srivastava A, Hill A, Antoine T, Warman
8. Allman EL, Painter HJ, Samra J, AJ, Davies J, Pidathala C et al (2012) Genera-
Carrasquilla M, Llinas M (2016) Metabolomic tion of quinolone antimalarials targeting the
profiling of the malaria box reveals antimalarial Plasmodium falciparum mitochondrial respira-
target pathways. Antimicrob Agents Che- tory chain for the treatment and prophylaxis of
mother 60:6635–6649 malaria. Proc Natl Acad Sci U S A
9. Kwon YK, Lu W, Melamud E, Khanam N, 109:8298–8303
Bognar A, Rabinowitz JD (2008) A domino 20. Ganesan SM, Morrisey JM, Ke H, Painter HJ,
effect in antifolate drug action in Escherichia Laroiya K, Phillips MA, Rathod PK, Mather
coli. Nat Chem Biol 4:602–608 MW, Vaidya AB (2011) Yeast dihydroorotate
10. Vincent IM, Creek DJ, Burgess K, Woods DJ, dehydrogenase as a new selectable marker for
Burchmore RJ, Barrett MP (2012) Untargeted Plasmodium falciparum transfection. Mol Bio-
metabolomics reveals a lack of synergy between chem Parasitol 177:29–34
nifurtimox and eflornithine against Trypano- 21. Cobbold SA, Chua HH, Nijagal B, Creek DJ,
soma brucei. PLoS Negl Trop Dis 6:e1618 Ralph SA, McConville MJ (2016) Metabolic
11. Zampieri M, Szappanos B, Buchieri MV, dysregulation induced in Plasmodium falci-
Trauner A, Piazza I, Picotti P, Gagneux S, parum by dihydroartemisinin and other front-
Borrell S, Gicquel B, Lelievre J et al (2018) line antimalarial drugs. J Infect Dis
High-throughput metabolomic analysis pre- 213:276–286
dicts mode of action of uncharacterized antimi- 22. Creek DJ, Jankevics A, Breitling R, Watson
crobial compounds. Sci Transl Med 10: DG, Barrett MP, Burgess KE (2011) Toward
eaal3973 global metabolomics analysis with hydrophilic
12. Spangenberg T, Burrows JN, Kowalczyk P, interaction liquid chromatography-mass spec-
McDonald S, Wells TN, Willis P (2013) The trometry: improved metabolite identification
open access malaria box: a drug discovery cata- by retention time prediction. Anal Chem
lyst for neglected diseases. PLoS One 8:e62906 83:8703–8710
13. Trager W, Jensen J (1976) Human malaria 23. Kind T, Fiehn O (2007) Seven Golden rules for
parasites in continuous culture. Science heuristic filtering of molecular formulas
193:673–675 obtained by accurate mass spectrometry. BMC
Bioinformatics 8:105
 Chapter 22

Analyzing Metabolomics Data for Environmental Health

and Exposome Research
Yuping Cai, Ana K Rosen Vollmar, and Caroline Helen Johnson

Abstract
The exposome is the cumulative measure of environmental influences and associated biological responses
across the life span, with critical relevance for understanding how exposures can impact human health.
Metabolomics analysis of biological samples offers unique advantages for examining the exposome. Simul-
taneous analysis of external exposures, biological responses, and host susceptibility at a systems level can
help establish links between external exposures and health outcomes. As metabolomics technologies
continue to evolve for the study of the exposome, metabolomics ultimately will help provide valuable
insights for exposure risk assessment, and disease prevention and management. Here, we discuss recent
advances in metabolomics, and describe data processing protocols that can enable analysis of the exposome.
This chapter focuses on using liquid chromatography–mass spectrometry (LC-MS)-based untargeted
metabolomics for analysis of the exposome, including (1) preprocessing of untargeted metabolomics
data, (2) identification of exposure chemicals and their metabolites, and (3) methods to establish associa-
tions between exposures and diseases.

Key words Exposome, Human health, Untargeted metabolomics, Data processing

1 Introduction

In 2002, the first genome-wide association study (GWAS) was

published, analyzing genetic susceptibility to myocardial infarction
[1]. One year later, the Human Genome Project was completed
[2]. Since then, we have witnessed the increasing success of GWAS
to show relationships between base-pair/gene patterns in genomic
loci and many diseases [3], including autoimmune diseases [4],
breast cancer [5], colorectal cancer [6], schizophrenia [7], and
type 1 and 2 diabetes [8, 9]. However, genetic linkage association
studies have only been able to explain a small fraction of the
estimated heritability of complex diseases. Instead, gene–environ-
ment interactions appear to underlie the etiology of major non-
communicable chronic diseases such as cancer, diabetes, and
vascular and neurodegenerative diseases. A meta-analysis of twin

Shuzhao Li (ed.), Computational Methods and Data Analysis for Metabolomics, Methods in Molecular Biology, vol. 2104,
https://doi.org/10.1007/978-1-0716-0239-3_22, © Springer Science+Business Media, LLC, part of Springer Nature 2020

447
448 Yuping Cai et al.

studies over the past 50 years found that heritability across all
human traits is 49% [10]. These results indicate that nongenetic
influences such as environmental and lifestyle risk factors could be
related to different phenotypes and underlie the etiology of many
diseases.
The importance of the environment as a key determinant of
health, both separate from and in interaction with genetics, has
motivated large-scale collaborative efforts to develop methods for
comprehensively measuring environmental exposures [11–
13]. While particulate matter (ambient particles and household
smoke), smoking, nutrition (diets high in cholesterol, salt, and
sugar), and occupational exposures are linked to about 50% of
deaths globally, there are many diseases for which the exposure
risk factors remain unknown [14]. In addition, the nature of these
risks is different for each person due to individual susceptibility,
which includes host factors such as differences in metabolism,
microbiome and endogenous hormones, and social determinants
of health like socioeconomic status or neighborhood. All of these
factors fluctuate over the life course, with susceptibility increasing
during critical windows, such as in pregnancy or times of ill health
[15], making it difficult to attribute a specific exposure as causal of a
particular outcome. The exposome, or “the cumulative measure of
environmental influences and associated responses throughout the
lifespan,” encompasses these diverse and changing external and
internal exposures [16]. The exposome concept was developed in
part to address the need for more comprehensive exposure assess-
ment methods, and to provide a framework for the measurement
and analysis of an individual’s exposure portfolio over an entire
lifetime [16–19].
Developing methods for measuring the exposome remains a
challenge given the difficulty of cataloguing an individual’s expo-
some at even a single point in time. High-throughput methods
enable a systems-level evaluation of genes (genomics), gene expres-
sion (transcriptomics), proteins (proteomics), and metabolites
(metabolomics), offering a comprehensive view of the biological
changes that can occur in an individual after a specific exposure.
Multi-omics approaches which combine these data can further help
identify downstream chemical changes that contribute to an expo-
sure phenotype or exposotype, or the accrued biological changes
within a system that has undergone an exposure event (Fig. 1).
Exposures can thus be linked to disease end points through
identification of associated biomolecular changes, or biomarker
measurements [20]. Compared to other -omics technologies,
metabolomics is advantageous for studying the exposome because
the metabolome profiles both the exposure and the biological
response. Peaks from environmental exposures are routinely
observed in untargeted mass spectrometry data. Indeed, the
increased incorporation of metabolomics into exposure research
 Metabolomics for Environmental Health and Exposome Research 449

Fig. 1 Exposotype, the exposure phenotype. (a) The environmental health paradigm: exposures modify
biological molecules and lead to disease. (b) Possible actions of environmental exposures on DNA, proteins,
and metabolites. Please note that metabolites are not in the original Central Dogma of Molecular Biology, but
informative of biological functions at both cellular and organismal levels. (Figure reproduced from [19])

can be visualized through the steady annual increase in the number

of publications in PubMed from 2005, when the term “exposome”
was coined, to 2019 (Fig. 2).
The metabolome, or the pool of low molecular-weight mole-
cules in a sample, encompasses both endogenous and externally
derived exposure metabolites. Endogenous metabolites comprise
approximately 3000 substrates and products, the concentrations of
which are determined by dietary intake, de novo synthesis, and
actions of genome-encoded protein enzymatic reactions that
450 Yuping Cai et al.

Fig. 2 Number of publications searched by keywords “exposure, metabolomics”

in PubMed

convert these metabolite substrates into products. These metabo-

lites are required to serve as intermediates of metabolic pathways
that produce energy and molecules for cell growth, essential
biological functions that can be monitored as surrogates of aberrant
biochemistry if these processes go awry. Metabolites derived from
specific external exposures include environmental chemicals (e.g.,
pesticides, plasticizers, heavy metals, food contaminants, deodor-
ants, cosmetics) and pharmaceutical drugs; these exposures and
their biotransformed phase I and phase II metabolites can be
identified and quantified simultaneously through metabolome
profiling. Untargeted metabolomics, the measurement of the meta-
bolome at a systems-level, allows for a comprehensive, agnostic,
and high-throughput capture of low molecular-weight metabolites
within a biological sample [21, 22]. The nature of the approach,
along with the continued advancement of high-resolution plat-
forms such as ultraperformance liquid chromatography–mass spec-
trometry (UPLC-MS) and 1H nuclear magnetic resonance (NMR)
spectroscopy, offers advantages for exposomics. Simultaneous
profiling of exposures and perturbed metabolic pathways or net-
works, along with a deep breadth of coverage, can establish the
connection between exposures, biological responses, and pheno-
types [23–25]. Implementing metabolomics for large-scale and
systematic exposome research presents a number of challenges,
especially with respect to data analysis [19, 26]. Typically, an LC-
MS-based untargeted metabolomics experiment can measure more
than 30,000 features/variables [27]. This results in a large, high-
dimensional dataset that requires nontrivial bioinformatics analysis
to identify environmental chemicals, endogenous metabolites, and
the enriched metabolic pathways linking exposures to human
health outcomes.
 Metabolomics for Environmental Health and Exposome Research 451

Challenges from both analytical and statistical perspectives

must be addressed when analyzing metabolomics data for environ-
mental health and exposome research, and have been detailed in
multiple reviews [28–30]. From an analytical perspective, chal-
lenges include selection of appropriate data preprocessing proce-
dures to quantify the relevant mass spectral features of exposures
and metabolites; structural elucidation, or identification of external
exposures and endogenous metabolites; and discovery of metabolic
networks affected by specific exposures [17]. From a statistical
perspective, notable challenges include not only assessing associa-
tions between a single exposure and an outcome, but accounting
for mixtures of external and internal environmental exposures that
change over time [31]; addressing complex correlation structures
of exposome data as strong correlations between exposures are
common when many exposures are measured [30, 32]; and diffi-
culty differentiating between normal variation in outcomes and
environmentally induced changes, since exposures are dynamic
and there is no “reference exposome” against which to compare
findings [29]. In this chapter, we describe data processing protocols
that can be used to analyze metabolomics data in exposomics
research, focusing on the LC-MS-based untargeted discovery of
metabolites and metabolic pathways which link environmental
chemical exposures to adverse human health outcomes.

2 Methods

2.1 Metabolomic Choosing the most suitable analytical platform is a key foundation
Technologies for exposome research. Developments in high-resolution metabo-
for Exposome lomics (HRM) have accelerated exposome research due to their
Research improved ability to profile a broad range of endogenous metabo-
lites and environmental chemicals simultaneously. Modern time-of-
2.1.1 High-Resolution flight (ToF) and Fourier-transform (FT) mass spectrometers confer
Metabolomics (HRM) high resolution, high accuracy, and provide the most sensitive
analysis for untargeted metabolomics methodologies [22]. Benefit-
ing from a high scan speed, ToF mass spectrometers are the most
commonly used instruments for high-throughput metabolite
profiling. FT mass spectrometers such as the Orbitrap™, however,
have high mass resolution and mass accuracy which are not affected
by low ion abundance. Therefore, they are superior for metabolite
quantification, especially for environmental chemicals with a low
abundance and a wide dynamic range in biological samples
[33]. The data output from these HRM platforms contains a wealth
of information regarding enriched metabolic pathways which can
facilitate the discovery of relationships between environmental
exposures and adverse health outcomes.
452 Yuping Cai et al.

2.1.2 Measurement The wide dynamic range of chemical compounds that are present in
of Low Level of Exposures human samples presents a major challenge for accurate quantifica-
tion, which is essential for distinguishing biological differences
between groups. Typically, blood concentrations of environmental
chemicals (fmol to μmol) are 1000 times lower than endogenous
metabolites (nmol to mmol) and metabolites from nutritional or
pharmaceutical drug sources [34]. Biotransformed environmental
chemicals produced through endogenous enzymatic reactions may
also be present at very low levels. Electrospray ionization (ESI)
coupled with LC is the most commonly applied technique in
MS-based metabolomics research [35]. In order to quantify
low-abundance environmental chemicals such as polychlorinated
biphenyls [36], polycyclic aromatic hydrocarbons [37], and poly-
brominated diphenyl ethers [38, 39], other ionization sources are
required due to their hydrophobic chemical structures.
Gas-chromatography (GC) with electron impact (EI) ionization is
routinely used for the analyses of these common environmental
pollutants [40–42]. LC hybrids with soft ionization approaches
such as atmospheric pressure chemical ionization (APCI) and
atmospheric pressure photoionization (APPI) can also readily mea-
sure these low-level hydrophobic chemicals [43]. In addition, APPI
uses a charge carrier (acetone or toluene) which increases sensitiv-
ity. Ion mobility LC-MS analysis has shown promise for the char-
acterization of structurally similar environmental chemicals, due to
increased resolving power and improved signal-to-noise ratio
[43, 44].

2.2 Metabolomics An MS-based metabolomics experiment can result in a high-

Data Processing for dimensional dataset containing information on tens of thousands
Exposome Research of ions. Thus, the primary task for data preprocessing is to extract
mass spectral ions that are related to signals from environmental
2.2.1 Data Preprocessing chemicals and endogenous metabolites. In recent years, much
effort has been devoted to developing algorithms and bioinformat-
ics tools for peak detection and peak annotation. These tools have
been developed in different programming languages, such as
XCMS [45] using R, MZmine [46] using Java, and MetAlign
[47] using C++. In addition, instrument vendors have their own
proprietary software that can be used. A recent review provides a
comprehensive list of current software available for metabolomics
analysis [48].
Some metabolomics data preprocessing decisions such as the
type of normalization approach can be influenced by exposome
study design and specimen collection considerations. For example,
because of the impossibility of measuring the entire life course
exposome, many studies focus on exposure measurements and
specimen collection during critical windows for the health out-
comes being investigated [29, 49]. These critical windows may
have unique physiologic signatures that must be considered when
 Metabolomics for Environmental Health and Exposome Research 453

selecting normalization approaches. A prime example is pregnancy,

which is a common window targeted by exposome studies explor-
ing early-life exposures [30]. However, the normalization of urine
metabolites is unreliable when using the gold standard creatinine,
due to the normal physiologic changes of pregnancy [50]. Instead,
approaches like specific gravity or probabilistic quotient normaliza-
tion may be more suitable for metabolomics analyses of urine
during this critical window [51, 52].

2.2.2 Characterization A prerequisite for identifying causal pathways linking exposures to

and Identification health outcomes is to first identify the statistically significant mass
of Exposures spectral signals within the metabolomics data. Since both exposures
and Metabolites and endogenous metabolite signals will be captured at the same
time in a data set, such chemical complexity can make it difficult to
accurately characterize the relationships between exposures and
metabolic consequences. One of the most significant analytical
bottlenecks of metabolomics remains the structural elucidation of
chemicals. Fortunately, the availability of bioinformatic tools and
chemical databases that incorporate environmental chemicals and
other exposures can help to overcome this challenge. For example,
the R package Collection of Algorithms for Metabolite pRofile
Annotation (CAMERA), which is embedded in the XCMS soft-
ware, organizes isotopic peaks, adducts, and fragment ions by
grouping ions based on retention time, integration of peak shape,
and correlation of peak intensity. This reduces feature redundancy
by ~50% and helps with metabolite identification [53]. xMSanno-
tator is an R package and automated computational workflow to
annotate ions in untargeted metabolomics. This annotation tool
uses an integrative multicriteria scoring algorithm to assign identi-
fication to an ion, accounting for correlation strength, difference in
retention time, number of matching adducts, and isotopes to cate-
gorize database matches containing information about both envi-
ronmental chemicals and endogenous metabolites (ChemSpider,
KEGG, HMDB, T3DB, and LipidMaps) into different confidence
levels [54].
However, to obtain absolute confidence regarding metabolite
identification, MS/MS data need to be acquired and the spectra
compared either against metabolite databases, or preferably to
standards (to also obtain matching retention time information).
In the following sections, various databases developed for identifi-
cation and characterization of environmental chemicals and endog-
enous metabolites are introduced and discussed, respectively.
To better characterize environmental chemicals from metabo-
lomics data, several useful databases with MS/MS information have
been developed and are readily available for identification of envi-
ronmental metabolites in exposome research (Table 1). The
METLIN database was recently modified to incorporate
>700,000 chemicals from the United States Environmental
454 Yuping Cai et al.

Table 1
Databases of environmental metabolites for exposome research

Number of
exposures or
Database metabolites Content URL
METLIN Exposome >700,000 Chemical information, MS/MS https://metlin.scripps.
database spectra edu/
Comparative 130,796 Chemical–gene/protein http://ctdbase.org/
Toxicogenomics interactions, chemical/
Database (CTD) gene–disease relationships
Drugbank 11,926 Drugs, drug targets, MS/MS https://www.
spectra drugbank.ca/
Hazardous 6016 Chemical information, https://toxnet.nlm.
Substances Data environmental fate, toxicity nih.gov/
Bank (HSDB) newtoxnet/hsdb.
htm
Toxin Exposome 3678 Both toxin and toxin target, gene http://www.t3db.ca/
Database (T3DB) expression data, MS/MS spectra
Exposome-Explorer 876 Dietary and pollutant biomarkers, http://exposome-
with concentration and explorer.iarc.fr/
correlation values

Protection Agency’s (EPA) Distributed Structure-Searchable Tox-

icity (DSSTox) database, which includes toxicants and xenobiotics
[55]. Furthermore, the database consists of a number of phase I
and II mammalian biotransformation products that can help to
investigate xenobiotic metabolism for exposome research, and can
be better observed by LC-ESI-MS compared to the nontrans-
formed parent compounds, which cannot be ionized through this
approach. In addition, the METLIN database contains MS/MS
spectra for many metabolites listed on the database, enabling
greater confidence in metabolite identification. Moreover, the data-
base is integrated into the XCMS Online data processing platform,
enabling instant putative metabolite identification and toxicologi-
cal pathway analysis, facilitating both metabolite identification and
knowledge of altered metabolic pathways [56].
The Toxin Exposome Database (T3DB) houses 3678 toxins,
with approximately one-third having MS/MS spectrum informa-
tion [57]. T3DB is a comprehensive exposome resource that con-
tains information on toxins, toxin targets, and gene expression
datasets, with detailed information on chemical properties, toxic
concentration in human biofluids, toxic effects, and mechanisms.
The informative descriptions of each exposure can be useful to
epidemiologists and toxicologists in the exposome research field.
Another source of externally derived chemicals is pharmaceutical
Metabolomics for Environmental Health and Exposome Research 455

drugs. The database Drugbank was developed to help identify these

compounds, and contains information on 12,064 drugs, also with
mass spectral data [58].
Other environmental chemical-specific databases are important
for exposure characterization, but do not contain the MS/MS
information required for metabolite identification. For example,
the Comparative Toxicogenomics Database (CTD) is a large online
database containing concentration values of environmental chemi-
cals in blood, urine, and other biospecimens extracted from the
scientific literature [59]. CTD includes information for chemicals,
relevant genes, and proteins on toxicology, sequence, reference,
species, and microarray, along with manually curated information
about chemical–gene/protein interactions, and chemical–disease
and gene–disease relationships. The Hazardous Substances Data
Bank (HSDB) is managed by the US National Institutes of Health
and focuses on the toxicology of potentially hazardous chemicals
[60]. It contains information on environmental fate, human expo-
sure, detection methods, and regulatory requirements, and is
updated several times a year by a Scientific Review Panel.
Exposome-Explorer, which currently houses 876 exposures, is the
first database dedicated to exposure biomarkers. This database was
constructed by consolidating exposure biomarker data scattered
throughout the literature [61]. To identify metabolic pathways
modulated by environmental chemicals, the structural elucidation
of metabolites is of critical importance. For MS-based metabolo-
mics data, the robust identification of metabolites is obtained by
comparing the MS1 mass-to-charge (m/z) ratio (the initial/precur-
sor m/z spectrum), retention time, and MS/MS spectra to stan-
dards. Spectral similarity matching to standard metabolite libraries
housed on metabolite databases has helped expedite this process.
Over the last 20 years, several freely available databases have
been developed to facilitate identification of metabolites in meta-
bolomic experiments, including METLIN [62], Human Metabo-
lome Database (HMDB) [63], MassBank [64], and LipidMaps
[65]. METLIN has a dual function in metabolomics, as it is useful
for identifying both environmental exposures and endogenous
metabolites. These databases were generated using pure standard
compounds to provide accurate identification of metabolites from
biological samples. Most of these repositories allow for a compari-
son of MS/MS data from a research sample to MS/MS data on the
database for metabolite identification. Other platforms such as
MyCompoundID [66] enable identification of unknown metabo-
lites by comparison to a library of predicted MS/MS data from
known metabolites, and from those formed by biotransformation
reactions. The Global Natural Products Social Molecular Network-
ing (GNPS) [67] resource is another platform for the storage,
analysis, and sharing of MS/MS spectral data; it enables researchers
456 Yuping Cai et al.

to annotate MS/MS data, and compare it to publicly available data.

In addition, novel tools such as Mummichog [68] and PIUMet
[69] provide putative dysregulated metabolic pathway information
through bypassing metabolite identification and assessing the inter-
connectivity of metabolites in known metabolic pathways. To
increase accuracy in the discovery of dysregulated metabolic path-
ways, Metabolite Identification and Dysregulated Network Analysis
(MetDNA) software was recently developed by incorporating
MS/MS spectral matching followed by dysregulated network anal-
ysis (http://metdna.zhulab.cn). Of note, metabolite identification
is optimized through the implementation of a metabolic reaction
network (MRN)-based recursive algorithm. These approaches
could advance the discovery of metabolic pathways affected by
environmental exposures. Most of these resources and methods
are described in depth in other chapters of this book.

2.2.3 Analyzing Across the life course, an individual will encounter a myriad of
Associations Between environmental exposures which can influence human health, either
the Exposome and Health individually or synergistically. A notable challenge of implementing
Outcomes the exposome is understanding the direct relationship between one
or more exposures and a disease outcome, and identifying the
factors that link environmental chemicals to adverse health out-
comes. From a data analysis perspective, the core data processing
step essential for the exposome lies in establishing the relationship
between the exposure and health outcome.
Conventional biomonitoring of environmental exposures tar-
gets the measurement of a limited number of chemicals in biospeci-
mens such as blood and urine. Establishing the effects of an
exposure on human health is therefore restricted to a relatively
small exposure panel. Traditional environmental epidemiology
reflects this biomonitoring approach, as it assesses the association
between a single exposure and a single response. However, these
methods can result in biased and fragmentary pictures of environ-
ment–health relationships, since environmental exposures often
occur as mixtures and change frequently [29]. The exposome
more realistically represents the many exposures an individual will
encounter throughout life, but the enormous quantity of data
collected in exposome research presents challenges for analysis
and interpretation.
To analyze the thousands of features typical of LC-MS experi-
ments, metabolomics-based exposome research often employs
agnostic or hypothesis-free statistical methods, examining data for
patterns and correlations between exposures/metabolites and out-
comes [30]. These findings can later be validated through targeted
methods to describe potential mechanisms. Proponents of agnostic
methods see such techniques as allowing previously unknown rela-
tionships to emerge given the complexity of exposures and out-
comes, while critics warn such approaches have high chances of
Metabolomics for Environmental Health and Exposome Research 457

false discovery, reverse causation, and spurious findings

[28, 70]. Even hypothesis-free analytic methods should continue
to employ core environmental epidemiology practices to assess and
adjust for confounders, selection bias, measurement error, and
exposure misclassification, and testing for multiple comparisons
must be controlled for in analysis [29, 30, 70].
Another challenge in analyzing associations between the expo-
some and health outcomes is that when many exposures are
measured, it is assumed that there will be strong correlations and
collinearity between exposures. For example, cumulative exposure
to radon, γ-ray, and long-lived radionuclides (LLR) was shown to
be associated with a significant risk of lung cancer [71]. However,
the Pearson correlation coefficient between radon and γ-ray expo-
sure was 0.60 ( p < 0.001), and the correlation between radon and
LLR exposure 0.52 ( p < 0.001) in the study. This collinearity
between exposures limits our ability to elicit the effect on human
health specific to each exposure. Describing the data’s correlation
structure and key exposures or subject attributes responsible for
variation allows researchers to more precisely identify indicator
exposures, which may represent groups of exposures [30]. Such
indicator exposures could guide future research and measurement,
or have implications for regulatory policy. For example, a study of
the exposome during pregnancy integrated biomonitoring, ques-
tionnaire, environmental monitoring, and geospatial modeling data
from 728 pregnant women to determine the relationships between
81 environmental exposures [30]. Strong correlations were
observed within certain “families” of exposure, such as noise,
water pollutants, perfluoroalkyl substances (PFAS), and air pollu-
tants, while weak correlations were observed for groups including
home and built environment exposures, metals, and phthalates,
reflecting more diverse sources of exposure.
Many univariate and multivariate statistical methodologies exist
to address these and other statistical challenges of exposome
research, allowing for relationships between environmental expo-
sures and health outcomes to be more effectively established. Regres-
sion methods are most commonly used in this field. To assess the
statistical significance of the association between a single environ-
mental exposure and the target biological response or health out-
come, linear regression can be applied when the outcome is
continuous, and logistic regression can be applied when the outcome
is binary. This can be repeated for each environmental exposure
measured within a cohort, and correction for multiple comparisons
is performed to reduce false positive results. To obtain robust find-
ings, the results should be validated using other datasets or cohorts.
This was how the initial “environment-wide association studies” or
“exposome-wide association studies” (EWAS) were conducted
[72]. Because EWAS is a hypothesis-free, agnostic method of evalu-
ating associations between exposures and health outcomes, it can
help uncover novel associations [73–75].
458 Yuping Cai et al.

A common problem of high-dimensional data sets is having a

larger number of predictors than the sample size [76]. LASSO
(least absolute shrinkage and selection operator) is a penalized
regression model and variable selection method that identifies the
most informative predictors for inclusion in the final model from a
larger set of covariates. LASSO improves prediction precision by
forcing the sum of the regression coefficients to be shrunk to zero,
with the least informative coefficients given a value of zero
[77]. For LASSO, the number of nonpenalized variables is limited
by the number of observations. Elastic net (ENET) is another
penalized regression model, which combines the advantages of
ridge and LASSO regression [78]. Ridge regression effectively
accommodates correlated variables, while LASSO selects variables
through shrinkage [79]. The elastic net approach enables the selec-
tion of a set of correlated exposures, with the interpretability of a
simple regression model. Bayesian methods are another venue for
variable selection. Graphical Unit Evolutionary Stochastic Search
(GUESS) and its wrapper tool R2GUESS allows for model param-
eterization and handling of up to several thousand of observations,
hundreds of thousands of predictors, and a few outcomes simulta-
neously, taking advantage of graphic hardware capabilities [80, 81].
For exposome research, where the number of covariates is
greater than the number of observations, dimension reduction
methods such as principal components analysis (PCA) [82], canon-
ical correlation analysis (CCA) [83], and partial least-squares (PLS)
[84] are approaches for reducing the redundancy and complexity of
results. By using dimension reduction methods, several compo-
nents consisting of multiple covariates/predictors are defined iter-
atively to explain as much of the covariance between the exposures
and the health outcome as possible. Table 2 lists some commonly
used statistical methods and tools to assess exposome-health
associations.

Table 2
Common informatics tools freely available as R packages for analyzing associations between the
exposome and health outcomes

Statistical method R package

Linear regression and logistic regression Stats and rexposome
LASSO and elastic net Glmnet
Bayesian variable selection regression of multivariate responses R2GUESS
Principal components analysis (PCA) Prcomp
Canonical correlation analysis (CCA) CCA
Partial least squares (PLS) Pls
 Metabolomics for Environmental Health and Exposome Research 459

2.3 Multi-omics The primary goal of exposome research is to unravel biological

Integration responses to environmental exposures, ultimately linking exposures
for Exposome to health outcomes. Multiple layers of physiologic responses can be
Research stimulated by environmental exposures, including DNA expression
(genome), RNA expression (transcriptome), proteins (proteome),
metabolites (metabolome), and microbiota (microbiome). There-
fore, the diverse and rich information obtained from genomics,
epigenomics, proteomics, transcriptomics, and metabolomics can
be exploited to capture comprehensive biological responses to
exposures. A recent example demonstrated the strength of -omics
integration in environmental health through the use of transcrip-
tomics and metabolomics to reveal the neurotoxicologic mechan-
isms of a herbicide (paraquat) and a fungicide (maneb) [85].
-Omics data are high-dimensional, often with thousands or
tens of thousands of data points, presenting major challenges for
multi-omics data integration that aims to identify exposure-
mediated biomarkers. Both unsupervised and supervised algo-
rithms form the basis of multi-omics analyses. Based on the statisti-
cal methodologies utilized, integration strategies can be
categorized into matrix factorization methods, Bayesian methods,
network-based methods, multiple kernel learning, and multistep-
based methods. As an example, Guida et al. applied a two-step
strategy to integrate epigenomics and transcriptomics data to char-
acterize the biologically relevant long-term markers of tobacco
smoke exposure [86]. Initially, the epigenomics data were regressed
against the outcome, resulting in a set of outcome-associated vari-
ables (smoking-related CpG sites). Then, another regression was
performed on the smoking-related epigenomics against transcrip-
tomics data. In this study, the gene LRRN3 was identified as a
promising biomarker in which both methylation and gene expres-
sion were associated with tobacco smoke exposure. The topic of
multi-omics integration is covered in more detail in Chapter 23 of
this book.
Multi-omics technologies allow for agnostic analyses of the
biological effects of exposure in a high-throughput manner, and
integrating these data can help identify and reinforce molecular
biomarkers of chemical exposures, as discussed above. However,
several challenges need to be addressed in the near future for better
characterization of the exposome. First, methods for efficiently
extracting useful information need to be developed, as -omics
approaches generate a large amount of data accompanied by redun-
dancy. For example, it is typical for an untargeted metabolomics
dataset with over 25,000 measured metabolic features to harbor
fewer than 1000 unique metabolites [87]. Additionally, there are
both data storage and data processing challenges associated with
the large data sets generated by multi-omics experiments. This
challenge can be partially addressed by improving the computa-
tional power and performance efficiency of new algorithms. Data
460 Yuping Cai et al.

sharing for multiple data types is another critical aspect for consid-
eration. At present, domain-specific data repositories are available
for data sharing, including repositories for EWAS data sets (https://
nhanes.hms.harvard.edu/transmart/datasetExplorer/index),
metabolomics data (Metabolomics Workbench, https://www.
metabolomicsworkbench.org; MetaboLights, https://www.ebi.ac.
uk/metabolights/), and genomics data (Gene Expression Omni-
bus, https://www.ncbi.nlm.nih.gov/geo/). Future directions for
multi-omics data sharing could include platforms integrating data-
sets on specific health outcomes.

3 Conclusions and Perspectives

Metabolomics is a vital tool for exposome research, capable of

profiling both environmental chemicals and biological effects
through the measurement of metabolites. In this chapter, we intro-
duced HRM technologies and presented data processing proce-
dures for using metabolomics data in exposome research,
including methods for processing raw MS data, identifying envi-
ronmental chemicals and endogenous metabolites, and analyzing
associations between exposures and health outcomes. We anticipate
that exposome research will flourish by integrating metabolomics
and other -omics approaches, especially since several tools for data
processing with application to the exposome paradigm are now
available.
A series of different challenges still remain for exposome
research, including analytical challenges and difficulties in data
processing. From an analytical perspective, the broad dynamic
range of environmental chemicals in biological specimens (e.g.,
blood and urine) prompts the need for new detection technologies
using reconfigured mass spectrometers coupled with chro-
matographic separation (for example, GC/APCI/MS/MS techni-
ques) [88]. Even though several publicly available databases now
exist that house environmental chemicals, the accurate identifica-
tion of these metabolites captured by metabolomics data is an
urgent challenge that needs to be addressed. One promising solu-
tion to facilitate structural elucidation is to first identify the chemi-
cal class to which exposures and metabolites belong. For example,
differential chemical isotope labeling LC-MS methods have been
developed to profile high coverage of the submetabolome, as with
the danslyation labeling method for the hydroxyl submetabolome
[89]. Other informatics tools such as MetFamily [90] and CluM-
SID [91] apply unsupervised clustering methods to the analysis of
MS/MS spectra in untargeted metabolomics. Both approaches
found structural-related families to aid in compound identification.
MS2LDA, utilizing the Latent Dirichlet Allocation (LDA) algo-
rithm, was recently shown to decompose molecules into
Metabolomics for Environmental Health and Exposome Research 461

Mass2Motifs (substructures) for molecular grouping [92]. The

updated version MS2LDA+ can support the discovery and compar-
ison of structural families across multiple samples [93].
The greatest challenges often present the greatest opportu-
nities. With advances in personalized medicine, exposome research
may play a vital role in improving medical treatments for patients, as
human variability in response to exposures can impact the treat-
ment of diseases, especially for complex cancer [94]. In addition to
individual-specific biological signatures originating from inherited
genes, transcripts, proteins, and metabolites, externally derived
environmental exposures also contribute to human variation.
Exposures can modulate host biological processes, or use the
same phase I and II biotransformation enzymes, thus potentially
affecting pharmaceutical drug efficacy and toxicology. Therefore,
exposome studies evaluating the effects of environmental factors on
human health have the potential to contribute to the development
of individualized health care and personalized medicine. Single
-omics data provides limited information on the pathogenesis of
diseases, but by combining multiple -omics data sources (e.g.,
genomics, transcriptomics, proteomics, and metabolomics), expo-
some research can support a comprehensive understanding of the
molecular profiles of each individual. One example project,
Integrated Personal Omics Profiling (iPOP), aims to integrate a
wide range of data to better characterize the molecular phenotypes
underlying a participant’s individual health status, drawing on data
sources including the genome, transcriptome, proteome, and meta-
bolome, along with information on cytokines, immune cells, clini-
cal laboratory test results, diet, stress, and activity levels.
Future individualized treatment may take into account exposome
information such as exposures to environmental chemicals, dietary
influences, and lifestyle.
The exposome framework encompasses not only internal expo-
sures from metabolism and the microbiome but also external expo-
sures ranging from socioeconomic status and neighborhood to air
quality and chemical exposures. Therefore, the exposome intersects
with structural and population-level determinants of health. While
metabolomics research in the context of the exposome is oriented
around generating individual-level data about biologically relevant
exposures, innovative exposome study designs are starting to con-
nect longitudinal and hierarchical nested exposure measures, such
as measures at the community, family, and individual levels. Such
studies will begin to describe how exposures change and are inter-
related at multiple scales and across time in the life course [29]. This
research has the potential to inform not only individual-level clini-
cal decisions, but also risk assessment, exposure prevention, and
policy interventions [95]. Metabolomics is embedded in the expo-
some paradigm, with the capacity to quantify both environmental
exposures and biological responses, thus offering exciting tools for
environmental health and exposome research.
462 Yuping Cai et al.

References
1. Ozaki K, Ohnishi Y, Iida A, Sekine A, nucleotide polymorphisms (SNPs) previously
Yamada R, Tsunoda T, Sato H, Sato H, associated with type 2 diabetes replicates asso-
Hori M, Nakamura Y, Tanaka T (2002) Func- ciation with 12 SNPs in nine genes. Diabetes
tional SNPs in the lymphotoxin-alpha gene 56(1):256–264
that are associated with susceptibility to myo- 10. Polderman TJC, Benyamin B, de Leeuw CA,
cardial infarction. Nat Genet 32(4):650–654 Sullivan PF, van Bochoven A, Visscher PM,
2. Collins FS, Lander ES, Rogers J, Waterston Posthuma D (2015) Meta-analysis of the heri-
RH, Conso IHGS (2004) Finishing the tability of human traits based on fifty years of
euchromatic sequence of the human genome. twin studies. Nat Genet 47(7):702
Nature 431(7011):931–945 11. Vrijheid M, Slama R, Robinson O, Chatzi L,
3. Buniello A, MacArthur JAL, Cerezo M, Harris Coen M, van den Hazel P, Thomsen C,
LW, Hayhurst J, Malangone C, McMahon A, Wright J, Athersuch TJ, Avellana N,
Morales J, Mountjoy E, Sollis E, Suveges D, Basagana X, Brochot C, Bucchini L,
Vrousgou O, Whetzel PL, Amode R, Guillen Bustamante M, Carracedo A, Casas M,
JA, Riat HS, Trevanion SJ, Hall P, Junkins H, Estivill X, Fairley L, van Gent D, Gonzalez
Flicek P, Burdett T, Hindorff LA, JR, Granum B, Grazuleviciene R, Gutzkow
Cunningham F, Parkinson H (2019) The KB, Julvez J, Keun HC, Kogevinas M, McEa-
NHGRI-EBI GWAS Catalog of published chan RRC, Meltzer HM, Sabido E, Schwarze
genome-wide association studies, targeted PE, Siroux V, Sunyer J, Want EJ, Zeman F,
arrays and summary statistics 2019. Nucleic Nieuwenhuijsen MJ (2014) The human early-
Acids Res 47(D1):D1005–D1012 life exposome (HELIX): project rationale and
4. Eyre S, Worthington J (2014) Take your PICS: design. Environ Health Perspect 122
moving from GWAS to immune function. (6):535–544
Immunity 41(6):883–885 12. Kawamoto T, Nitta H, Murata K, Toda E,
5. Cuzick J, Brentnall A, Dowsett M (2017) Tsukamoto N, Hasegawa M, Yamagata Z,
SNPs for breast cancer risk assessment. Onco- Kayama F, Kishi R, Ohya Y, Saito H, Sago H,
target 8(59):99211–99212 Okuyama M, Ogata T, Yokoya S, Koresawa Y,
6. Yao L, Tak YG, Berman BP, Farnham PJ Shibata Y, Nakayama S, Michikawa T,
(2014) Functional annotation of colon cancer Takeuchi A, Satoh H, Ch WGER (2014)
risk SNPs. Nat Commun 5:5114 Rationale and study design of the Japan envi-
ronment and children’s study (JECS). BMC
7. Need AC, Ge D, Weale ME, Maia J, Feng S, Public Health 14:25
Heinzen EL, Shianna KV, Yoon W,
Kasperaviciute D, Gennarelli M, Strittmatter 13. Vineis P, Chadeau-Hyam M, Gmuender H,
WJ, Bonvicini C, Rossi G, Jayathilake K, Cola Gulliver J, Herceg Z, Kleinjans J,
PA, McEvoy JP, Keefe RS, Fisher EM, St Jean Kogevinas M, Kyrtopoulos S,
PL, Giegling I, Hartmann AM, Moller HJ, Nieuwenhuijsen M, Phillips DH, Probst-
Ruppert A, Fraser G, Crombie C, Middleton Hensch N, Scalbert A, Vermeulen R, Wild
LT, St Clair D, Roses AD, Muglia P, Francks C, CP, Consortium E (2017) The exposome in
Rujescu D, Meltzer HY, Goldstein DB (2009) practice: design of the EXPOsOMICS project.
A genome-wide investigation of SNPs and Int J Hyg Environ Health 220(2):142–151
CNVs in schizophrenia. PLoS Genet 5(2): 14. Rappaport SM, Barupal DK, Wishart D,
e1000373 Vineis P, Scalbert A (2014) The blood expo-
8. Reddy MVPL, Wang H, Liu S, Bode B, Reed some and its role in discovering causes of dis-
JC, Steed RD, Anderson SW, Steed L, ease. Environ Health Perspect 122
Hopkins D, She JX (2011) Association (8):769–774
between type 1 diabetes and GWAS SNPs in 15. Wild CP (2012) The exposome: from concept
the southeast US Caucasian population. Genes to utility. Int J Epidemiol 41(1):24–32
Immun 12(3):208–212 16. Miller GW, Jones DP (2014) The nature of
9. Willer CJ, Bonnycastle LL, Conneely KN, nurture: refining the definition of the expo-
Duren WL, Jackson AU, Scott LJ, Narisu N, some. Toxicol Sci 137(1):1
Chines PS, Skol A, Stringham HM, Petrie J, 17. Wild CP (2005) Complementing the genome
Erdos MR, Swift AJ, Enloe ST, Sprau AG, with an “exposome”: the outstanding chal-
Smith E, Tong M, Doheny KF, Pugh EW, lenge of environmental exposure measurement
Watanabe RM, Buchanan TA, Valle TT, Berg- in molecular epidemiology. Cancer Epidemiol
man RN, Tuomilehto J, Mohlke KL, Collins Biomark Prev 14(8):1847–1850
FS, Boehnke M (2007) Screening of 134 single
 Metabolomics for Environmental Health and Exposome Research 463

18. Louis GMB, Sundaram R (2012) Exposome: human health and disease. Curr Environ
time for transformative research. Stat Med 31 Health Rep 4(1):89–98
(22):2569–2575 29. Stingone JA, Louis GMB, Nakayama SF, Ver-
19. Rattray NJW, Deziel NC, Wallach JD, Khan meulen RCH, Kwok RK, Cui YX, Balshaw
SA, Vasiliou V, Ioannidis JPA, Johnson CH DM, Teitelbaum SL (2017) Toward greater
(2018) Beyond genomics: understanding implementation of the Exposome research par-
exposotypes through metabolomics. Hum adigm within environmental epidemiology.
Genomics 12(1):4 Annu Rev Public Health 38(38):315–327
20. Steinberg CEW, Sturzenbaum SR, Menzel R 30. Robinson O, Basagana X, Agier L, de
(2008) Genes and environment - striking the Castro M, Hernandez-Ferrer C, Gonzalez JR,
fine balance between sophisticated biomonitor- Grimalt JO, Nieuwenhuijsen M, Sunyer J,
ing and true functional environmental geno- Slama R, Vrijheid M (2015) The pregnancy
mics. Sci Total Environ 400(1–3):142–161 Exposome: multiple environmental exposures
21. Nicholson JK, Lindon JC, Holmes E (1999) in the INMA-Sabadell birth cohort. Environ
Metabonomics: understanding the metabolic Sci Technol 49(17):10632–10641
responses of living systems to pathophysiologi- 31. Chung MK, Kannan K, Louis GM, Patel CJ
cal stimuli via multivariate statistical analysis of (2018) Toward capturing the Exposome:
biological NMR spectroscopic data. Xenobio- exposure biomarker variability and Coexposure
tica 29(11):1181–1189 patterns in the shared environment. Environ
22. Patti GJ, Yanes O, Siuzdak G (2012) Metabo- Sci Technol 52(15):8801–8810
lomics: the apogee of the omics trilogy. Nat 32. Rappaport SM (2016) Genetic factors are not
Rev Mol Cell Biol 13(4):263–269 the major causes of chronic diseases. PLoS One
23. Ellis JK, Athersuch TJ, Thomas LDK, 11(4):e0154387
Teichert F, Perez-Trujillo M, Svendsen C, 33. Go YM, Walker DI, Liang YL, Uppal K, Soltow
Spurgeon DJ, Singh R, Jarup L, Bundy JG, QA, Tran V, Strobel F, Quyyumi AA, Ziegler
Keun HC (2012) Metabolic profiling detects TR, Pennell KD, Miller GW, Jones DP (2015)
early effects of environmental and lifestyle Reference standardization for mass spectrome-
exposure to cadmium in a human population. try and high-resolution metabolomics applica-
BMC Med 10:61 tions to Exposome research. Toxicol Sci 148
24. Maitre L, Villanueva CM, Lewis MR, (2):531–543
Ibarluzea J, Santa-Marina L, Vrijheid M, 34. Dennis KK, Marder E, Balshaw DM, Cui YX,
Sunyer J, Coen M, Toledano MB (2016) Lynes MA, Patti GJ, Rappaport SM, Shaugh-
Maternal urinary metabolic signatures of fetal nessy DT, Vrijheid M, Barr DB (2017) Biomo-
growth and associated clinical and environ- nitoring in the era of the Exposome. Environ
mental factors in the INMA study. BMC Med Health Perspect 125(4):502–510
14:177 35. Lei ZT, Huhman DV, Sumner LW (2011)
25. Baker MG, Simpson CD, Lin YS, Shireman Mass spectrometry strategies in metabolomics.
LM, Seixas N (2017) The use of metabolomics J Biol Chem 286(29):25435–25442
to identify biological signatures of manganese 36. Ulbrich B, Stahlmann R (2004) Developmen-
exposure. Ann Work Expo Health 61 tal toxicity of polychlorinated biphenyls
(4):406–415 (PCBs): a systematic review of experimental
26. Johnson CH, Athersuch TJ, Collman GW, data. Arch Toxicol 78(5):252–268
Dhungana S, Grant DF, Jones DP, Patel CJ, 37. Balcioglu EB (2016) Potential effects of poly-
Vasiliou V (2017) Yale school of public health cyclic aromatic hydrocarbons (PAHs) in marine
symposium on lifetime exposures and human foods on human health: a critical review. Toxin
health: the exposome; summary and future Rev 35(3–4):98–105
reflections. Hum Genomics 11:32 38. Frederiksen M, Vorkamp K, Thomsen M,
27. Ivanisevic J, Zhu ZJ, Plate L, Tautenhahn R, Knudsen LE (2009) Human internal and exter-
Chen S, O’Brien PJ, Johnson CH, Marletta nal exposure to PBDEs—a review of levels and
MA, Patti GJ, Siuzdak G (2013) Toward sources. Int J Hyg Environ Health 212
Omic scale metabolite profiling: a dual (2):109–134
separation-mass spectrometry approach for 39. Herbstman JB, Sjodin A, Kurzon M, Leder-
coverage of lipid and central carbon metabo- man SA, Jones RS, Rauh V, Needham LL,
lism. Anal Chem 85(14):6876–6884 Tang D, Niedzwiecki M, Wang RY, Perera F
28. Buck Louis GM, Smarr MM, Patel CJ (2017) (2010) Prenatal exposure to PBDEs and neu-
The Exposome research paradigm: an opportu- rodevelopment. Environ Health Perspect 118
nity to understand the environmental basis for (5):712–719
464 Yuping Cai et al.

40. Pleil JD, Stiegel MA, Sobus JR, Tabucchi S, 51. Dieterle F, Ross A, Schlotterbeck G, Senn H
Ghio AJ, Madden MC (2010) Cumulative (2006) Probabilistic quotient normalization as
exposure assessment for trace-level polycyclic robust method to account for dilution of com-
aromatic hydrocarbons (PAHs) using human plex biological mixtures. Application in H-1
blood and plasma analysis. J Chromatogr B NMR metabonomics. Anal Chem 78
Analyt Technol Biomed Life Sci 878 (13):4281–4290
(21):1753–1760 52. Gagnebin Y, Tonoli D, Lescuyer P, Ponte B, de
41. Marek RF, Thorne PS, Wang K, DeWall J, Seigneux S, Martin PY, Schappler J, Boccard J,
Hornbuckle KC (2013) PCBs and OH-PCBs Rudaz S (2017) Metabolomic analysis of urine
in serum from children and mothers in urban samples by UHPLC-QTOF-MS: impact of
and rural US communities. Environ Sci Tech- normalization strategies. Anal Chim Acta
nol 47:3353–3361 955:27–35
42. Awad AM, Martinez A, Marek RF, Hornbuckle 53. Johnson CH, Ivanisevic J, Benton HP, Siuzdak
KC (2016) Occurrence and distribution of two G (2015) Bioinformatics: the next frontier of
hydroxylated polychlorinated biphenyl conge- metabolomics. Anal Chem 87(1):147–156
ners in Chicago air. Environ Sci Technol Lett 3 54. Uppal K, Walker DI, Jones DP (2017) xMSan-
(2):47–51 notator: an R package for network-based anno-
43. Zheng XY, Dupuis KT, Aly NA, Zhou YX, tation of high-resolution metabolomics data.
Smith FB, Tang KQ, Smith RD, Baker ES Anal Chem 89(2):1063–1067
(2018) Utilizing ion mobility spectrometry 55. Warth B, Spangler S, Fang M, Johnson CH,
and mass spectrometry for the analysis of poly- Forsberg EM, Granados A, Martin RL,
cyclic aromatic hydrocarbons, polychlorinated Domingo-Almenara X, Huan T, Rinehart D,
biphenyls, polybrominated diphenyl ethers and Montenegro-Burke JR, Hilmers B, Aisporna A,
their metabolites. Anal Chim Acta Hoang LT, Uritboonthai W, Benton HP,
1037:265–273 Richardson SD, Williams AJ, Siuzdak G
44. Marquez-Sillero I, Aguilera-Herrador E, (2017) Exposome-scale investigations guided
Cardenas S, Valcarcel M (2011) Ion-mobility by global metabolomics, pathway analysis, and
spectrometry for environmental analysis. TrAC cognitive computing. Anal Chem 89
Trends Anal Chem 30(5):677–690 (21):11505–11513
45. Smith CA, Want EJ, O’Maille G, Abagyan R, 56. Tautenhahn R, Patti GJ, Rinehart D, Siuzdak
Siuzdak G (2006) XCMS: processing mass G (2012) XCMS online: a web-based platform
spectrometry data for metabolite profiling to process untargeted Metabolomic data. Anal
using nonlinear peak alignment, matching, Chem 84(11):5035–5039
and identification. Anal Chem 78(3):779–787 57. Wishart D, Arndt D, Pon A, Sajed T, Guo AC,
46. Katajamaa M, Miettinen J, Oresic M (2006) Djoumbou Y, Knox C, Wilson M, Liang YJ,
MZmine: toolbox for processing and visualiza- Grant J, Liu YF, Goldansaz SA, Rappaport SM
tion of mass spectrometry based molecular pro- (2015) T3DB: the toxic exposome database.
file data. Bioinformatics 22(5):634–636 Nucleic Acids Res 43(D1):D928–D934
47. Lommen A (2009) MetAlign: Interface- 58. Wishart DS, Feunang YD, Guo AC, Lo EJ,
driven, versatile metabolomics tool for Marcu A, Grant JR, Sajed T, Johnson D,
hyphenated full-scan mass spectrometry data Li C, Sayeeda Z, Assempour N, Iynkkaran I,
Preprocessing. Anal Chem 81(8):3079–3086 Liu YF, Maciejewski A, Gale N, Wilson A,
48. Misra BB, Mohapatra S (2019) Tools and Chin L, Cummings R, Le D, Pon A, Knox C,
resources for metabolomics research commu- Wilson M (2018) DrugBank 5.0: a major
nity: a 2017–2018 update. Electrophoresis 40 update to the DrugBank database for 2018.
(2):227–246 Nucleic Acids Res 46(D1):D1074–D1082
49. Chadeau-Hyam M, Athersuch TJ, Keun HC, 59. Davis AP, Grondin CJ, Johnson RJ, Sciaky D,
De Iorio M, Ebbels TMD, Jenab M, McMorran R, Wiegers J, Wiegers TC, Mat-
Sacerdote C, Bruce SJ, Holmes E, Vineis P tingly CJ (2019) The comparative toxicoge-
(2011) Meeting-in-the-middle using meta- nomics database: update 2019. Nucleic Acids
bolic profiling - a strategy for the identification Res 47(D1):D948–D954
of intermediate biomarkers in cohort studies. 60. Jordan S, Fonger G, Hazard G (2017) Hazard-
Biomarkers 16(1):83–88 ous substances data bank: recent features and
50. MacPherson S, Arbuckle TE, Fisher M (2018) enhancements. Abstr Am Chem Soc 254
Adjusting urinary chemical biomarkers for 61. Neveu V, Moussy A, Rouaix H, Wedekind R,
hydration status during pregnancy. J Expo Sci Pon A, Knox C, Wishart DS, Scalbert A (2017)
Environ Epidemiol 28(5):481–493 Exposome-explorer: a manually-curated
 Metabolomics for Environmental Health and Exposome Research 465

database on biomarkers of exposure to dietary Demarque DP, Pociute E, O’Neill EC,

and environmental factors. Nucleic Acids Res Briand E, Helfrich EJN, Granatosky EA,
45(D1):D979–D984 Glukhov E, Ryffel F, Houson H,
62. Smith CA, O’Maille G, Want EJ, Qin C, Trau- Mohimani H, Kharbush JJ, Zeng Y, Vorholt
ger SA, Brandon TR, Custodio DE, JA, Kurita KL, Charusanti P, McPhail KL, Niel-
Abagyan R, Siuzdak G (2005) METLIN—a sen KF, Vuong L, Elfeki M, Traxler MF,
metabolite mass spectral database. Ther Drug Engene N, Koyama N, Vining OB, Baric R,
Monit 27(6):747–751 Silva RR, Mascuch SJ, Tomasi S, Jenkins S,
63. Wishart DS, Knox C, Guo AC, Eisner R, Macherla V, Hoffman T, Agarwal V, Williams
Young N, Gautam B, Hau DD, Psychogios N, PG, Dai JQ, Neupane R, Gurr J, Rodriguez
Dong E, Bouatra S, Mandal R, Sinelnikov I, AMC, Lamsa A, Zhang C, Dorrestein K, Dug-
Xia JG, Jia L, Cruz JA, Lim E, Sobsey CA, gan BM, Almaliti J, Allard PM, Phapale P,
Shrivastava S, Huang P, Liu P, Fang L, Nothias LF, Alexandrovr T, Litaudon M, Wol-
Peng J, Fradette R, Cheng D, Tzur D, fender JL, Kyle JE, Metz TO, Peryea T,
Clements M, Lewis A, De Souza A, Zuniga A, Nguyen DT, VanLeer D, Shinn P, Jadhav A,
Dawe M, Xiong YP, Clive D, Greiner R, Muller R, Waters KM, Shi WY, Liu XT, Zhang
Nazyrova A, Shaykhutdinov R, Li L, Vogel LX, Knight R, Jensen PR, Palsson BO,
HJ, Forsythe I (2009) HMDB: a knowledge- Pogliano K, Linington RG, Gutierrez M,
base for the human metabolome. Nucleic Acids Lopes NP, Gerwick WH, Moore BS, Dorres-
Res 37:D603–D610 tein PC, Bandeira N (2016) Sharing and com-
munity curation of mass spectrometry data
64. Horai H, Arita M, Kanaya S, Nihei Y, Ikeda T, with global natural products social molecular
Suwa K, Ojima Y, Tanaka K, Tanaka S, networking. Nat Biotechnol 34(8):828–837
Aoshima K, Oda Y, Kakazu Y, Kusano M,
Tohge T, Matsuda F, Sawada Y, Hirai MY, 68. Li SZ, Park Y, Duraisingham S, Strobel FH,
Nakanishi H, Ikeda K, Akimoto N, Maoka T, Khan N, Soltow QA, Jones DP, Pulendran B
Takahashi H, Ara T, Sakurai N, Suzuki H, (2013) Predicting network activity from high
Shibata D, Neumann S, Iida T, Tanaka K, throughput metabolomics. PLoS Comput Biol
Funatsu K, Matsuura F, Soga T, Taguchi R, 9(7):e1003123
Saito K, Nishioka T (2010) MassBank: a public 69. Pirhaji L, Milani P, Leidl M, Curran T, Avila-
repository for sharing mass spectral data for life Pacheco J, Clish CB, White FM, Saghatelian A,
sciences. J Mass Spectrom 45(7):703–714 Fraenkel E (2016) Revealing disease-associated
65. Fahy E, Sud M, Cotter D, Subramaniam S pathways by network integration of untargeted
(2007) LIPID MAPS online tools for lipid metabolomics. Nat Methods 13(9):770–776
research. Nucleic Acids Res 35:W606–W612 70. Langston MA, Levine RS, Kilbourne BJ,
66. Huan T, Tang CQ, Li RH, Shi Y, Lin GH, Li L Rogers GL, Kershenbaum AD, Baktash SH,
(2015) MyCompoundID MS/MS search: Coughlin SS, Saxton AM, Agboto VK, Hood
metabolite identification using a library of pre- DB, Litchveld MY, Oyana TJ, Matthews-
dicted fragment-ion-spectra of 383,830 possi- Juarez P, Juarez PD (2014) Scalable combina-
ble human metabolites. Anal Chem 87 torial tools for health disparities research. Int J
(20):10619–10626 Environ Res Public Health 11
(10):10419–10443
67. Wang MX, Carver JJ, Phelan VV, Sanchez LM,
Garg N, Peng Y, Nguyen DD, Watrous J, 71. Vacquier B, Rage E, Leuraud K, Caer-Lorho S,
Kapono CA, Luzzatto-Knaan T, Porto C, Houot J, Acker A, Laurier D (2011) The influ-
Bouslimani A, Melnik AV, Meehan MJ, Liu ence of multiple types of occupational exposure
WT, Criisemann M, Boudreau PD, to radon, gamma rays and long-lived radionu-
Esquenazi E, Sandoval-Calderon M, Kersten clides on mortality risk in the French “post-55”
RD, Pace LA, Quinn RA, Duncan KR, Hsu sub-cohort of uranium miners: 1956–1999.
CC, Floros DJ, Gavilan RG, Kleigrewe K, Radiat Res 176(6):796–806
Northen T, Dutton RJ, Parrot D, Carlson 72. Patel CJ, Ioannidis JP (2014) Studying the
EE, Aigle B, Michelsen CF, Jelsbak L, elusive environment in large scale. JAMA 311
Sohlenkamp C, Pevzner P, Edlund A, (21):2173–2174
McLean J, Piel J, Murphy BT, Gerwick L, 73. Patel CJ, Bhattacharya J, Butte AJ (2010) An
Liaw CC, Yang YL, Humpf HU, environment-wide association study (EWAS)
Maansson M, Keyzers RA, Sims AC, Johnson on type 2 diabetes mellitus. PLoS One 5(5):
AR, Sidebottom AM, Sedio BE, Klitgaard A, e10746
Larson CB, Boya CA, Torres-Mendoza D, 74. Patel CJ, Cullen MR, Ioannidis JP, Butte AJ
Gonzalez DJ, Silva DB, Marques LM, (2012) Systematic evaluation of environmental
466 Yuping Cai et al.

factors: persistent pollutants and nutrients cor- combined biological effects of multiple expo-
related with serum lipid levels. Int J Epidemiol sures. J Epidemiol Community Health 72
41(3):828–843 (7):564–571
75. Manrai AK, Cui YX, Bushel PR, Hall M, 85. Roede JR, Uppal K, Park Y, Tran V, Jones DP
Karakitsios S, Mattingly CJ, Ritchie M, (2014) Transcriptome-metabolome wide asso-
Schmitt C, Sarigiannis DA, Thomas DC, ciation study (TMWAS) of maneb and paraquat
Wishart D, Balshaw DM, Patel CJ (2017) neurotoxicity reveals network level interactions
Informatics and data analytics to support in toxicologic mechanism. Toxicol Rep
exposome-based discovery for public health. 1:435–444
Annu Rev Public Health 38(38):279–294 86. Guida F, Sandanger TM, Castagne R,
76. Sun ZC, Tao YB, Li S, Ferguson KK, Meeker Campanella G, Polidoro S, Palli D, Krogh V,
JD, Park SK, Batterman SA, Mukherjee B Tumino R, Sacerdote C, Panico S, Severi G,
(2013) Statistical strategies for constructing Kyrtopoulos SA, Georgiadis P, Vermeulen
health risk models with multiple pollutants RCH, Lund E, Vineis P, Chadeau-Hyam M
and their interactions: possible choices and (2015) Dynamics of smoking-induced
comparisons. Environ Health 12:85 genome-wide methylation changes with time
77. Tibshirani R (1996) Regression shrinkage and since smoking cessation. Hum Mol Genet 24
selection via the Lasso. J R Stat Soc Series B (8):2349–2359
Stat Methodol 58(1):267–288 87. Mahieu NG, Patti GJ (2017) Systems-level
78. Zou H, Hastie T (2005) Regularization and annotation of a 25 000 features to fewer than
variable selection via the elastic net (vol B G metabolomics data set reduces 1000 unique
67, pg 301, 2005). J R Stat Soc Series B Stat metabolites. Anal Chem 89(19):10397–10406
Methodol 67:768–768 88. Geng DW, Jogsten IE, Dunstan J, Hagberg J,
79. Agier L, Portengen L, Chadeau-Hyam M, Wang T, Ruzzin J, Rabasa-Lhoret R, van Bavel
Basagana X, Giorgis-Allemand L, Siroux V, B (2016) Gas chromatography/atmospheric
Robinson O, Vlaanderen J, Gonzalez JR, Nieu- pressure chemical ionization/mass spectrome-
wenhuijsen MJ, Vineis P, Vrijheid M, Slama R, try for the analysis of organochlorine pesticides
Vermeulen R (2016) A systematic comparison and polychlorinated biphenyls in human
of linear regression-based statistical methods to serum. J Chromatogr A 1453:88–98
assess Exposome-health associations. Environ 89. Zhao S, Luo X, Li L (2016) Chemical isotope
Health Perspect 124(12):1848–1856 Labeling LC-MS for high coverage and quanti-
80. Bottolo L, Richardson S (2010) Evolutionary tative profiling of the hydroxyl submetabolome
stochastic search for Bayesian model explora- in metabolomics. Anal Chem 88
tion. Bayesian Anal 5(3):583–618 (21):10617–10623
81. Liquet B, Bottolo L, Campanella G, 90. Treutler H, Tsugawa H, Porzel A, Gorzolka K,
Richardson S, Chadeau-Hyam M (2016) Tissier A, Neumann S, Balcke GU (2016) Dis-
R2GUESS: a graphics processing unit-based covering regulated metabolite families in untar-
R package for Bayesian variable selection geted metabolomics studies. Anal Chem 88
regression of multivariate responses. J Stat (16):8082–8090
Softw 69(2):1–32 91. Depke T, Franke R, Bronstrup M (2017) Clus-
82. Jiang C, Wang X, Li XY, Inlora J, Wang T, tering of MS2 spectra using unsupervised
Liu Q, Snyder M (2018) Dynamic human envi- methods to aid the identification of secondary
ronmental Exposome revealed by longitudinal metabolites from Pseudomonas aeruginosa. J
personal monitoring. Cell 175(1):277 Chromatogr B: Anal Technol Biomed Life Sci
83. Wang XH, Eijkemans MJC, Wallinga J, 1071:19–28
Biesbroek G, Trzcinski K, Sanders EAM, 92. van der Hooft JJJ, Wandy J, Barrett MP, Bur-
Bogaert D (2012) Multivariate approach for gess KEV, Rogers S (2016) Topic modeling for
studying interactions between environmental untargeted substructure exploration in meta-
variables and microbial communities. PLoS bolomics. Proc Natl Acad Sci U S A 113
One 7(11):e50267 (48):13738–13743
84. Jain P, Vineis P, Liquet B, Vlaanderen J, 93. van der Hooft JJJ, Wandy J, Young F,
Bodinier B, van Veldhoven K, Kogevinas M, Padmanabhan S, Gerasimidis K, Burgess KEV,
Athersuch TJ, Font-Ribera L, Villanueva CM, Barrett MP, Rogers S (2017) Unsupervised
Vermeulen R, Chadeau-Hyam M (2018) A discovery and comparison of structural families
multivariate approach to investigate the across multiple samples in untargeted metabo-
lomics. Anal Chem 89(14):7569–7577
 Metabolomics for Environmental Health and Exposome Research 467

94. Lu YF, Goldstein DB, Angrist M, Cavalleri G MG, Estes SM, Agboto VK, Robinson P,
(2014) Personalized medicine and human Wilson S, Lichtveld MY (2014) The public
genetic diversity. Cold Spring Harb Perspect health exposome: a population-based, expo-
Med 4(9):a008581 sure science approach to health disparities
95. Juarez PD, Matthews-Juarez P, Hood DB, research. Int J Environ Res Public Health 11
Im W, Levine RS, Kilbourne BJ, Langston (12):12866–12895
MA, Al-Hamdan MZ, Crosson WL, Estes
 Chapter 23

Network-Based Approaches for Multi-omics Integration

Guangyan Zhou, Shuzhao Li, and Jianguo Xia

Abstract
Network-based approach is rapidly emerging as a promising strategy to integrate and interpret different
-omics datasets, including metabolomics. The first section of this chapter introduces the current progresses
and main concepts in multi-omics integration. The second section provides an overview of the public
resources available for creation of biological networks. The third section describes three common applica-
tion scenarios including subnetwork identification, network-based enrichment analysis, and systems meta-
bolomics. The section four introduces the concept of hierarchical community network analysis. The section
five discusses different tools for network visualization. The chapter ends with a future perspective on multi-
omics integration.

Key words Systems biology, Multi-omics integration, Data integration, Network enrichment analysis,
Hierarchical community networks, Network visualization

1 Introduction

The widespread applications of various omics technologies across

life sciences have shifted the main bottleneck from data generation
to data analysis and interpretation. Researchers now can easily
obtain comprehensive landscapes of DNA/RNA molecules, pro-
teins or metabolites using genomics, transcriptomics, proteomics,
or metabolomics technologies. Fueled by the decreasing costs of
omics technologies, the past few years have witnessed a growing
number of studies employing more than one type of omics tech-
nologies (i.e., multi-omics) to investigate complex diseases and
biological processes [1–4]. Despite being a relatively new member
of the omics family, metabolomics has gradually become a key
component in recent multi-omics studies. Small compounds are
not only the downstream products of the complex interactions
between host genetics, life styles and environmental exposures,
they also play active roles in physiology and disease through modu-
lation of other “omics” levels [5]. How to extract biologically
meaningful information from these multi-omics data have become

Shuzhao Li (ed.), Computational Methods and Data Analysis for Metabolomics, Methods in Molecular Biology, vol. 2104,
https://doi.org/10.1007/978-1-0716-0239-3_23, © Springer Science+Business Media, LLC, part of Springer Nature 2020

469
470 Guangyan Zhou et al.

a key area of research in current bioinformatics and systems biology

[6–9].
Different methods have been explored to help integrate and
interpret multi-omics data in order to gain systems insights. Two
general categories have emerged—the statistical integration and the
network-based integration. Statistical integration aims to identify
overall patterns or shared signatures across multiple omics datasets
by drawing inference from the data themselves, without necessarily
using prior knowledge. Some of the commonly used integration
methods include univariate correlation, matrix-based correlation
and matrix factorization [10–12]. Univariate correlation methods
attempt to identify relationships between individual molecules from
one omics layer to those of another omics layer. Matrix-based
correlation methods seek to identify linear relationships that best
explain correlation and variability both within and across omics
layers. Lastly, matrix factorization methods focus on decomposing
the datasets and projecting variations onto dimensionally reduced
space. Although very powerful, they tend to produce very complex
results which are very challenging to explain the underlying
biological processes. This review will not go into detail about
these statistical approaches. Please refer to the several excellent
reviews for more details [12–15].
The network-based integration views the biological system as
interconnected networks of molecular entities, and is primarily
concerned with creating, analyzing, and visually exploring such
networks [6]. To perform network-based integration, researchers
need to first obtain a network relevant to the biological systems
under investigation. A network can be constructed based on our
current knowledge (i.e., genetic, physical, or biochemical interac-
tions obtained from databases or previous studies). The most com-
monly used biological networks are protein–protein interaction
(PPI) networks and metabolic networks [16]. These networks
have proven to be invaluable frameworks to aid in global under-
standing of complex interactions among different molecules of
interest [17]. Researchers first project the molecules of interest to
these networks, and then try to identify important connections and
modules to gain insights into the underlying system. Many algo-
rithms have been developed to identify subnetworks enriched for
user input data, which correspond to context-specific functional
modules characteristic of the biological systems under study [18].
This chapter will introduce the three key components in
network-based multi-omics integration—the prior knowledgebase
to create the networks; the algorithms to mine the networks; and
the visualization strategies to explore the networks. Researchers
need to develop a good understanding of the three components
in order to efficiently apply the network-based approach to obtain
systems insights and to develop hypotheses. A further conceptual
development of hierarchical community networks will also be intro-
duced and illustrated.
 Network-Based Multi-omics Integration 471

2 Knowledge Bases and Multi-omics Data Resources

The general idea of network-based integration is to map the multi-

omics experimental data into the context of our knowledge frame-
work in the forms of different networks. For instance, proteins can
bind to DNA or RNA to regulate the expression of genes, and
interactions between proteins, metabolites, and lipids can control
the function of molecular apparatus and signal cascades. The
ensemble of these different interactions provides the knowledge
framework for the use of network-based approaches to multi-
omics integration.
Over the last decade, significant efforts have been invested into
creating comprehensive databases on annotations of molecules,
biological pathways and molecular interaction networks. Currently,
curated metabolic pathways can be found in several comprehensive
databases including KEGG [19], Reactome [20], MetaCyc [21],
BiGG [22], Recon3D [23], and WikiPathways [24]. KEGG path-
way database covers ~8500 reactions and ~16,500 compounds
combining different species. MetaCyc offers a comparable number
of pathways to KEGG, however, it provides a greater set of pathway
attributes and larger coverage in plants, fungi, metazoans, and
actinobacteria [25]. BiGG specializes in collecting genome-scale
metabolic models for a given organism. Recon3D is specialized in
the characterization of human metabolic model. WikiPathways is
unique with its collaborative pathway editing feature which allows
research communities to contribute to pathway annotation efforts.
Pathways are functionally coherent and are easy to use and to
understand. However, they may miss important reactions that
cross the boundaries of multiple pathways. In addition, many
metabolites detected in metabolomics are not covered in any meta-
bolic pathways. To address these limitations, researchers have com-
bined manual curation and text-mining to create more
confederated biochemical reaction networks or protein–chemical
interactions databases such as RHEA [26], CTD [27], and
STITCH [28].
Besides metabolic pathways and networks, PPI is widely used to
create molecular interaction networks. There are several compre-
hensive PPI databases such as IntAct [29], BioGRID [30], MINT
[31], and InnateDB [32]. To control the high false positives in PPI,
these databases focus on experimentally validated interactions,
which only cover a subset of the proteins in several model organ-
isms. To address these limitations, STRING database collects and
integrates experimentally validated, literature-curated, and compu-
tationally predicted PPI data covering more than 2000 different
organisms [33].
In addition to metabolic and PPI networks, gene regulatory
networks (GRN) such as transcription factor (TF)–gene and
472 Guangyan Zhou et al.

miRNA–gene interactions offer critical mechanistic insights into

gene regulations and downstream effects. These interactions used
to be based primarily on computational predictions which tend to
contain very high false positives [34, 35]. To improve the situation,
high-throughput experimental technologies have been developed
in recent years that enabled the development of several high-quality
databases on TF–gene and miRNA–gene interactions [36–38].
Multi-omics datasets have become increasingly available to the
public, providing exciting opportunities for developing and bench-
marking computational methods for data integration. There have
been concerted efforts by several large consortia to generate multi-
omics datasets to study biological responses and complex diseases
(e.g., the NCI60 cell line datasets [39], The Cancer Genome Atlas
(TCGA) [40], the integrative Human Microbiome Project (iHMP)
[41], Metabolic Syndrome in Men (METSIM) Study [42]). Other
projects such as Japanese Multi Omics Reference Panel (jMorp)
[43] and MuTHER [44] aim at characterizing the molecular phe-
notypes of healthy individuals and twins. Finally, the Omics Dis-
covery Index (OmicsDI) platform aims to facilitate the discovery
and coordination of different omics dataset [45]. Table 1 provides a
list of these examples.

3 Network-Based Integration

Biological systems can be abstracted as a series of networks com-

prised of interconnected molecular entities. When the networks are
relatively small, researchers can directly visualize the relationships to
gain global insights. For large networks, it is often necessary to first
apply some graph theory to extract the most relevant subnetworks
before they can be visualized. Figure 1 illustrates three different
purposes using network-based approaches. In the following sec-
tion, we introduce these three types of network analysis: subnet-
work analysis, network enrichment analysis, and systems
metabolomics.

3.1 Subnetwork A common network-based approach is to build context-specific

Identification multi-omics subnetwork from experimental data [46–48]. This is
and Analysis usually achieved by connecting significant molecules identified in
individual omics layers through known molecular interactions. For
instance, to integrate transcriptomics and metabolomics, the list of
differentially expressed genes can be mapped into PPI databases
through their corresponding proteins to form PPI subnetworks
while the list of differentially abundant metabolites can connect to
the PPI subnetwork through known biochemical reactions. The
resulting subnetworks contain both input genes/proteins and
metabolites as well as a subset of their direct interacting partners.
Thy reveal not only the relationships among those important
 Network-Based Multi-omics Integration 473

Table 1
Example large-scale multi-omics projects

Project Description Links

NCI60 Small compound screening on 60 human tumor https://dtp.cancer.gov/discovery_
cell lines development/nci-60/
TCGA Cancer genomics project: ~20,000 samples from https://www.cancer.gov/about-nci/
33 cancer types and matched normal samples organization/ccg/research/
structural-genomics/tcga
TopMed Large project to study heart, lung, blood, and https://www.nhlbi.nih.gov/science/
sleep disorders trans-omics-precision-medicine-
topmed-program
jMorp Large multi-omics database generated using https://jmorp.megabank.tohoku.ac.jp
samples from 5000 healthy Japanese
volunteers
iHMP Multi-omics data characterizing microbiome- https://www.hmpdacc.org/ihmp/
host profiles in health and disease (pregnancy
and preterm birth, inflammatory bowel
disease, type II diabetes)
MuTHER Multi-omics resources from a range of tissues http://www.muther.ac.uk/
collected from a set of ~850 UK twins
METSIM Population-based study to investigate http://www.nationalbiobanks.fi/index.
nongenetic and genetic factors associated with php/studies2/10-metsim
the risk of T2D and CVD
OmicsDI A meta-database for multi-omics data discovery https://www.omicsdi.org/
and indexing

molecules, but also additional molecules (i.e., interaction partners)

which may be more suitable as potential biomarkers or therapeutic
targets.
Different algorithms can be used to build the context-specific
subnetworks. The most basic approach is to build subnetworks by
identifying known interactions that exist between molecules that
are prioritized from multi-omics experiments. For instance, for
proteins and metabolites, we can query the metabolic interaction
database to identify known relationships between enzymes and
metabolites from our input (also known as zero-order interaction
network). The main limitation with this approach is the potential
lack of direct interactions between molecules across omics layers,
resulting in very sparse subnetworks connecting a very small num-
ber of input elements. To address this issue, researchers often create
subnetworks containing the input molecules, the relationships
among them, and their direct interacting partners (also known as
first-order interaction network). However, such networks can
sometimes become too big to be visualized for meaningful inter-
pretation. Graph mining algorithms, such as Prize-Collecting
474 Guangyan Zhou et al.

Fig. 1 The overall workflow of knowledge-driven network-based approach. Experimental data obtained from
single-omics analysis are mapped onto existing molecular interaction or pathway databases to build a multi-
omics network. The resulting network can be used to identify context-specific subnetworks, perform network
enrichment analysis and for compound identification

Steiner Forest (PCSF) [49, 50] or jActiveModule [51], are often

applied to further identify and extract the relevant context specific
modules while minimizing the number of irrelevant nodes and
connections.
PCSF is a well-known graph theory problem [49, 50]. Given an
undirected global network and a list of input nodes with prizes, the
objective is to identify one or multiple subnetworks while maximiz-
ing the prizes (input nodes), minimizing the costs (edges) and the
number of subnetworks. For instance, the node prizes can corre-
spond to p-values obtained from differential expression analysis and
edge costs correspond to the inverse of interaction confidence
scores. As a general method, it can be used to interpret multi-
omics dataset. This algorithm has been used to identify novel
Parkinson-related druggable targets [52] and enriched pathways
that are affected by Kaposi’s Sarcoma associated Herpesvirus
infection [53].
 Network-Based Multi-omics Integration 475

In contrast to PCSF, jActiveModule scans the molecular inter-

action network to identify “network hotspots” (i.e., subnetwork
with significant differential expression changes), without having a
list of predefined nodes to be included in the resulting subnet-
works. Specifically, it computes p-value at node level from expres-
sion data, which is then converted to z-values and aggregated at
subnetwork level. To identify highest scoring subnetworks, it uses
an algorithm based on simulated annealing, a method that models
the physical process of heating a material and then slowly lowering
the temperature. For each iteration of the algorithm, it randomly
selects a node and keeps it if it improves the solution (subnetwork
score). Multiple methods based on the similar underlaying concept
have been developed subsequently to identify active modules or
network hotspots [54, 55]. Researchers have used this type of
approach to identify novel biomarkers in gastric cancer [56] and
novel metabolic modules that regulate macrophage
polarization [57].
Random walk with restart (RWR) is another promising graph
mining algorithm for the detection of novel candidates relevant to
condition of interest. It aims at identifying nodes with similar
functions to seed nodes, with the assumption that nodes with
similar functions tend to group more closely in a network. The
algorithm estimates affinity score of nodes in relation to seed nodes.
The basis of the algorithm is a walker that either randomly moves to
an immediate neighboring node or goes back to the starting node.
After some iterations, it reaches a stable state which outputs a steady
probability vector containing the affinity scores associated with
nodes. The RWR algorithm has been used to predict drug–target
interactions [58], lncRNA–disease associations [59], and miRNA–
disease associations [60].

3.2 Network-Based Enrichment analysis aims to identify functions that are significantly
Enrichment Analysis changed by comparing experimental data with known pathways or
other functionally related gene sets. A wide range of methods have
been developed. Among them, overrepresentation analysis (ORA)
and gene set enrichment analysis (GSEA) are two widely methods
for this task [61]. Although very powerful approaches, these
approaches ignore topological properties of these molecules within
pathways or networks, which are essential in order to understand
systems behaviour. In addition, they rely on direct overlap between
input molecules and functions, and molecules with missing path-
way or gene set annotation are not considered.
The network-based enrichment methods have been developed
to overcome some of these issues. These approaches take advantage
of graph-based statistics to exploit the connectivity information in
biological pathways or molecular interaction networks. A typical
approach consists of two main steps: (1) mapping the experimental
data onto the pathway or network; and (2) use “topology-aware”
476 Guangyan Zhou et al.

statistics that integrates structural information to calculate enrich-

ment scores. Proximity measure is often used to relate user input to
known biological pathways, as a fundamental principle is “guilt-by-
association” which assumes that nodes with similar functions tend
to be closer to each other in the networks. This approach allows the
detection of enriched pathways/gene sets whose associated mem-
bers do not directly overlap with those in user input. The connec-
tivity information can also complement classical enrichment
methods as an extra weighing factor, to increase discriminative
power and interpretability within the network context. For
instance, topological properties of a node can be used to evaluate
its significance in a given pathway. Below, we will briefly introduce
three methods for network-based enrichment analysis—the signal-
ing pathway impact analysis (SPIA), the network enrichment anal-
ysis (NEA), and EnrichNet.
Signaling pathway impact analysis (SPIA) [62] is one of the
earliest “topology-aware” enrichment methods. SPIA combines
scores from classical enrichment analysis with network-based
scores. Firstly, the algorithm matches elements from experimental
data to curated biological pathways and calculate a score using ORA
or GSEA. Secondly, using seed nodes as the origin, it propagates
the measured expression/abundance changes to the rest of nodes
in the pathways, generating a gene-level perturbation score for each
node. This perturbation analysis will assign different significance to
genes according to its position in the pathway (upstream or down-
stream) and other topological properties (i.e., degrees and
betweenness). Finally, the two results are combined to generate a
final score for a given pathway. By incorporating pathway structure
information, SPIA shows better specificity and more sensitivity than
those based on expression changes alone [62].
The network enrichment analysis (NEA) method [63] aims at
integrating functional information and network connectivity. It
quantifies functional set enrichment by assessing the interactions
between reference gene sets and input genes. Specifically, it com-
pares the observed number of links between them and with the
expected number for the given size of input genes under null
models via a network randomization algorithm to rewire the exist-
ing edges while preserving the degree distribution.
EnrichNet is another network-based enrichment analysis
method, in which the enrichment score is computed by measuring
proximity between input genes and reference gene sets, and does
not require the nodes to be direct interacting partner of each other
[64]. Firstly, RWR algorithm is used to score its distance to all
reference gene sets for each seed node. Secondly, node-level dis-
tance scores are converted into distance score vectors for reference
gene set. The individual distance vectors are then aggregated to
form a distribution corresponding to the background model.
Finally, the enrichment score of a specific reference gene set is
 Network-Based Multi-omics Integration 477

calculated by measuring the deviation of its distance vector to the

average distribution of the background model.

3.3 Systems Unlike transcriptomics or proteomics technologies, current meta-

Metabolomics bolomics platforms can only cover a small portion of the metabo-
lome. Given the chemical diversities of metabolites, no single
analysis method offers a “one-size-fits-all” solution. Nonetheless,
global metabolomics aims to provide an unbiased global measure-
ment of the metabolome. Modern mass spectrometers provide
ultrahigh resolution in characterizing molecular mass, and LC-MS
(liquid chromatography coupled mass spectrometry) has become
the dominant methods for global metabolomics in recent years
[65]. These LC-MS systems routinely detect tens of thousands of
metabolite “features,” defined mainly by mass-to-charge ratio (m/
z) and retention time. However, metabolite identification remains a
bottleneck of the field, as ambiguity is common in compound
matching and robust fragmentation data are difficult to acquire
for many metabolites. Fast and accurate compound annotation
remains a key obstacle in interpretation of global metabolomic
data and its effective integration with other omics data for systems
insights [66]. Systems metabolomics has emerged in recent years to
help address these issues. Here we define systems metabolomics as
the approach of inferring functional insights from global metabolite
profiles by leveraging prior knowledge and network structures.
One solution is shift the focal plane of analysis from individual
metabolites to pathways or networks (see Chapter 19 by Karnovsky
and Li), which are known to be more tolerant against random
errors [67]. Mummichog is among one of the earliest bioinformat-
ics tools using this concept [68]. It takes as input two lists of
spectral features—a significant list and a background reference
peak list (i.e., all features detected in the experiment). Computa-
tionally predicted metabolites from the significant peaks are then
mapped into known metabolic pathways or networks from existing
databases. Enrichment analysis is performed on these putative
metabolites and then compared with the results based on peaks
drawn randomly from the reference peak list. Thus, mummichog
borrows knowledge from prior metabolic reactions to prioritize the
prediction of metabolites and to perform pathway/network analy-
sis in one step. This method has since been successful applied to
many multi-omics studies [69–71] and has been implemented in
popular metabolomics tools such as XCMSonline [72] and Meta-
boAnalyst [73]. Figure 2 shows an example mummichog output as
implemented in MetaboAnalyst 4.0.
PIUMet employs a similar concept [74] but extends the search
space from metabolic pathways to an integrated weighted network
composed of protein–protein interaction combined with metabolic
interactions. Additionally, it allows the integrative analysis of spec-
tral data with other omics data such as transcriptomics or
478 Guangyan Zhou et al.

Fig. 2 An example mummichog result implemented in MetaboAnalyst. Mummichog algorithm directly maps
peaks as putative metabolites onto metabolic pathways and performs enrichment analysis to identify
perturbed pathways (shown in red circle)

proteomics. The first step consists of mapping the significant peak

features into the global network as sets of putative metabolites. To
help elucidate putative identities of altered metabolite feature, it
uses PCSF algorithm to identify minimal subnetworks containing
metabolite matches, assuming that metabolic reactions tend to be
parsimonious.

4 Data Driven Hierarchical Community Networks

The molecular networks discussed above are the workhorses of

systems biology. It is also important to consider the spatial and
temporal parameters in real biological systems. If a metabolic reac-
tion is limited in the mitochondria, a large generic PPI network
may contain a lot of irrelevant information and mislead investiga-
tors to many false positives. Given that different data types are often
from different compartments in human studies (e.g., gut micro-
biome from digestive tract, metabolomics from serum, and tran-
scriptomics from blood cells), direct coupling based on gene
activities can be very problematic. Kinetic rates may not be captured
in experiments, but they are critical to data interpretation. For
example, an intracellular metabolic reaction can occur in minutes,
but the activation and differentiation of T cells in an immune
 Network-Based Multi-omics Integration 479

response can take weeks. Therefore, computational models should

match the sampling and resolution of the experimental data. One of
the motivations of data driven approaches is to be close to the
biological truth.
Since different technologies often produce different variance
structure, PLS (partial least squares, or projection to latent struc-
ture) based methods are popular for integrating multi-omics data
[75]. After identifying the statistical relationships between features
in different -omics spaces, the interpretation is still challenging,
because the result can be still very redundant and is not necessarily
fit into existing knowledge structure (e.g., a known pathway). The
problem is well exemplified in transcriptomics from whole blood or
white blood cells: a large portion of the transcriptomic change can
be caused by the change of frequency of cell populations. A univar-
iate parameter at the cellular scale corresponds to hundreds of
features at the gene transcript level. To recognize as well as to
model these hierarchical organizations in data, a useful conceptual
development is hierarchical community networks (HiCoNet,
https://github.com/shuzhao-li/hiconet). The prototype of HiC-
oNet was published as a multiscale, multifactorial response network
to herpes zoster vaccine [76]. The authors integrated plasma meta-
bolomics, PBMC (peripheral blood mononuclear cell) transcrip-
tomics, plasma cytokines, and blood cell populations in the network
model and used it to interpret adaptive immune responses to the
vaccination. Within each data type, local communities were first
detected. The association between these communities across data
types was then assessed by PLS regression and permutation. Their
results established significant associations between blood metabo-
lites and intracellular transcriptome, and the importance of meta-
bolic phenotypes during human immune responses [76]. The
intermediate layer of communities provides a powerful means to
incorporate knowledge mapping. For untargeted metabolomics,
the community structure is likely to combine signals from isotopes,
adducts and metabolites of the same classes, therefore achieving
meaningful dimension reduction and cleaner result, as illustrated in
Fig. 3.
This HiCoNet approach has already been applied to infectious
disease [77] and exposome research (Li et al., Reproductive Toxi-
cology, in press). The concept of leveraging community structure in
each data type is compatible and operational with real biological
problems. The microbiome data are easily organized into commu-
nities based on transcription profiles and taxonomy. The emerging
single cell sequencing data are naturally organized by communities
of subpopulations. For exposome research, because humans are
usually exposed to a mixture of many environmental factors includ-
ing chemicals, exposure communities are a useful means to investi-
gate mixed exposures. The community detection algorithms are a
well-studied and mature area, easily adapted to HiCoNet. The
480 Guangyan Zhou et al.

Fig. 3 A schematic illustration of hierarchical community network. The

communities are detected within each data type. The association between
communities of different data types is tested via PLS regression, and the
significnat associations form the hierarchical community network

continued computational development will aid the visualization

and exploration of the resulting networks.

5 Network Visualization

Statistics alone is often insufficient to communicate the complex

information within multi-omics data to researchers. A key compo-
nent of network-based integration is visualization. Human eyes are
very good at pattern detection and to perform visual data mining to
gain deep insight or for hypothesis generation [78]. When pre-
sented properly, a network is very intuitive to navigate and to
interpret. In multi-omics networks, nodes represent a biological
entity such as metabolite, gene, protein or miRNA, and edges
represent known interactions shared between pairs of nodes. To
allow researchers to visually process large quantity of information,
different visual encoding can be applied to facilitate pattern detec-
tion. For instance, node size can vary according to topological
degree, node color based on expression or abundance value, and
node shape based on molecule types.
Several applications are available for interactive network visual-
ization of multi-omics data. Cytoscape is arguably the most widely
used tool for analysis and visualization of biological networks
[79]. Aside of the powerful features in terms of interactive
 Network-Based Multi-omics Integration 481

exploration and manipulation of networks, it has a plug-in system

with well-defined application programming interface (API) to
allow the community to contribute a wide range of analytics fea-
tures. KeyPathwayMiner is an example of Cytoscape plug-in devel-
oped for integrated analysis of multi-omics data [80]. Given a
global network, it attempts to identify perturbed subnetworks
using expression/abundance data of omics elements. Another
example is PathVisio, a popular Cytoscape plug-in developed for
the visualization and analysis of biological pathways [81]. It sup-
ports multi-omics visualization of biological pathways by mapping
expression values of transcriptomics and metabolomics datasets,
allowing users to visually assess the general patterns. There are
also several web-based applications such as Pathview and Pain-
tOmics that support the similar types of pathway analysis [82, 83].
The past few years have witnessed a trend toward web-based
network visualization. For instance, MetaboAnalyst 4.0 contains
two modules—Joint Pathway Analysis and Network Explorer to
support the integrative analysis and visualization of data from
metabolomics, transcriptomics, and metagenomics [73]. The
3Omics is a web application developed for combined analysis of
human transcriptomics, metabolomics, and proteomics data
[84]. Another web-based application, OmicsNet, was developed
for the creation and visualization of multi-omics biological net-
works in 3D space [47]. Users can map different molecules
(genes/proteins, transcription factors (TF), miRNAs, and/or
metabolites) of interest into the integrated knowledgebase to con-
struct context-specific subnetworks. The interactive 3D network
viewer can arrange the molecules from different omics into differ-
ent “layers” to reduce the visual cluttering and to facilitate the
pattern detection. An example network visualization generated
using OmicsNet is given in Fig. 4, showcasing a multi-omics net-
work composed of TF-gene, protein-protein and metabolite–
enzyme interactions. Users can perform a variety of network analy-
sis such as enrichment analysis, module detection, topology analy-
sis, and shortest path computing. The visualization system was
development using the cutting-edge WebGL technology. Given
the fast development of virtual reality (VR) technology, it is envi-
sioned that VR will play a significant role in visual analytics for
multi-omics integration and systems biology [85].

6 Future Prospects

This chapter introduces several key concepts and current status in

network-based multi-omics integration. Network-based
approaches rely heavily on the quality and details of the underlying
molecular interactions or pathways. Therefore, this approach is
limited by our current knowledge and understanding of biological
482 Guangyan Zhou et al.

Fig. 4 An example multi-omics network generated using OmicsNet. Transcription factors are colored in purple.
Genes/proteins are colored by their expression value using a green-red gradient. Grey nodes refers to
predicted interacting partner. Metabolites are colored in yellow. A shortest path between a transcription
factor and metabolite is highlighted in blue

processes and is primarily used to refine hypothesis or to clarify the

mechanistic links that explain molecular underpinning of pheno-
types within well-annotated model organisms. Despite the fact that
multivariate statistics are inherently complex and difficult to inter-
pret, there have been growing interests in recent years to develop
robust mathematical approaches to reveal coherent biological
changes across different omics profiles [13, 86]. Leveraging both
approaches represents a promising route toward comprehensive
understanding the complex multi-omics dataset.

7 Notes

1. We have reviewed network-based data integration in this chap-

ter, while mechanism-based integration is usually based on
metabolic models (see Chapter 18 by Witting et al.).
2. The common network analysis and visualization algorithms are
now implemented in popular software libraries (e.g., igraph
(https://igraph.org) [87], NetworkX (https://networkx.
github.io/) [88], Sigma (http://sigmajs.org), 3D network
(https://github.com/vasturiano/3d-force-graph), and graph-
 Network-Based Multi-omics Integration 483

tools (https://graph-tool.skewed.de/)). Popular desktop

based software tools include Cytoscape (https://cytoscape.
org/) [79] and Gephi (https://gephi.org/) [89].
Network-based multi-omics integration tools that are rel-
evant to metabolomics: MetaboAnalyst (https://meta
boanalyst.ca) [73], XCMS Online (https://xcmsonline.
scripps.edu/) [72], Metabox (https://github.com/
kwanjeeraw/metabox) [90], PIUMet (http://fraenkel-nsf.
csbi.mit.edu/PIUMet) [74], HiCoNet (https://github.
com/shuzhao-li/hiconet), 3Omics (https://3omics.cmdm.
tw/) [84], Metscape (http://metscape.ncibi.org) [91], Pain-
tOmics (http://www.paintomics.org) [83], and OmicsNet
(https://omicsnet.ca) [47].

Acknowledgments

This work has been funded in part by the US National Institutes of

Health via grants UH2 AI132345 (Li), R01 GM124061 (Yu),
U2C ES030163 (Jones, Li, Morgan, Miller), U01 CA235493
(Li, Xia, Siuzdak), Genome Canada, Genome Quebec, and
Canada Research Chairs program.

References
1. Hasin Y, Seldin M, Lusis A (2017) Multi-omics 7. Casci T (2012) Bioinformatics: next-genera-
approaches to disease. Genome Biol 18(1):83 tion omics. Nat Rev Genet 13(6):378
2. Coleman WB (2017) Next-generation breast 8. Rattray NJ, Deziel NC, Wallach JD, Khan SA,
cancer omics. Am J Pathol 187 Vasiliou V, Ioannidis JP et al (2018) Beyond
(10):2130–2132 genomics: understanding exposotypes through
3. Mach N, Ramayo-Caldas Y, Clark A, Moroldo metabolomics. Hum Genomics 12(1):4
M, Robert C, Barrey E et al (2017) Under- 9. Ritchie MD, Holzinger ER, Li R, Pendergrass
standing the response to endurance exercise SA, Kim D (2015) Methods of integrating data
using a systems biology approach: combining to uncover genotype–phenotype interactions.
blood metabolomics, transcriptomics and miR- Nat Rev Genet 16(2):85
Nomics in horses. BMC Genomics 18(1):187 10. Chong J, Xia J (2017) Computational
4. Villar M, Ayllon N, Alberdi P, Moreno A, Mor- approaches for integrative analysis of the meta-
eno M, Tobes R et al (2015) Integrated meta- bolome and microbiome. Metabolites 7(4):
bolomics, transcriptomics and proteomics E62
identifies metabolic pathways affected by Ana- 11. Gligorijevic V, Przulj N (2015) Methods for
plasma phagocytophilum infection in tick cells. biological data integration: perspectives and
Mol Cell Proteomics 14(12):3154–3172 challenges. J R Soc Interface 12
5. Rinschen MM, Ivanisevic J, Giera M, Siuzdak (112):20150571
G (2019) Identification of bioactive metabo- 12. Meng C, Zeleznik OA, Thallinger GG, Kuster
lites using activity metabolomics. Nat Rev Mol B, Gholami AM, Culhane AC (2016) Dimen-
Cell Biol 20:353–367 sion reduction techniques for the integrative
6. Yan J, Risacher SL, Shen L, Saykin AJ (2018) analysis of multi-omics data. Brief Bioinform
Network approaches to systems biology analy- 17(4):628–641
sis of complex disease: integrative methods for 13. Bersanelli M, Mosca E, Remondini D, Giam-
multi-omics data. Brief Bioinform 19 pieri E, Sala C, Castellani G et al (2016) Meth-
(6):1370–1381 ods for the integration of multi-omics data:
484 Guangyan Zhou et al.

mathematical aspects. BMC Bioinformatics 17 comparative toxicogenomics database: update

(2):S15 2019. Nucleic Acids Res 47(D1):D948–D954
14. Huang S, Chaudhary K, Garmire LX (2017) 28. Szklarczyk D, Santos A, von Mering C, Jensen
More is better: recent progress in multi-omics LJ, Bork P, Kuhn M (2015) STITCH 5: aug-
data integration methods. Front Genet 8:84 menting protein–chemical interaction net-
15. Tini G, Marchetti L, Priami C, Scott-Boyer M- works with tissue and affinity data. Nucleic
P (2019) Multi-omics integration—a compari- Acids Res 44 (Database issue):D380–D384
son of unsupervised clustering methodologies. 29. Kerrien S, Aranda B, Breuza L, Bridge A,
Brief Bioinform 20(4):1269–1279 Broackes-Carter F, Chen C et al (2011) The
16. Barabasi AL, Gulbahce N, Loscalzo J (2011) IntAct molecular interaction database in 2012.
Network medicine: a network-based approach Nucleic Acids Res 40 (Database issue):
to human disease. Nat Rev Genet 12(1):56–68 D841–D846
17. Vidal M, Cusick ME, Barabasi AL (2011) 30. Chatr-Aryamontri A, Oughtred R, Boucher L,
Interactome networks and human disease. Rust J, Chang C, Kolas NK et al (2017) The
Cell 144(6):986–998 BioGRID interaction database: 2017 update.
18. Mitra K, Carvunis AR, Ramesh SK, Ideker T Nucleic Acids Res 45 (Database issue):
(2013) Integrative approaches for finding D369–D379
modular structure in biological networks. Nat 31. Licata L, Briganti L, Peluso D, Perfetto L,
Rev Genet 14(10):719–732 Iannuccelli M, Galeota E et al (2011) MINT,
19. Kanehisa M, Goto S (2000) KEGG: Kyoto the molecular interaction database: 2012
encyclopedia of genes and genomes. Nucleic update. Nucleic Acids Res 40 (Database
Acids Res 28(1):27–30 issue):D857–D861
20. Fabregat A, Jupe S, Matthews L, Sidiropoulos 32. Breuer K, Foroushani AK, Laird MR, Chen C,
K, Gillespie M, Garapati P et al (2017) The Sribnaia A, Lo R et al (2013) InnateDB: sys-
Reactome pathway knowledgebase. Nucleic tems biology of innate immunity and beyond—
Acids Res 46 (Database issue):D481–D487 recent updates and continuing curation.
Nucleic Acids Res 41 (Database issue):
21. Caspi R, Billington R, Fulcher CA, Keseler IM, D1228–D1233
Kothari A, Krummenacker M et al (2017) The
MetaCyc database of metabolic pathways and 33. Szklarczyk D, Franceschini A, Wyder S, For-
enzymes. Nucleic Acids Res 46 (Database slund K, Heller D, Huerta-Cepas J et al
issue):D633–D639 (2014) STRING v10: protein–protein interac-
tion networks, integrated over the tree of life.
22. King ZA, Lu J, Dr€ager A, Miller P, Federowicz Nucleic Acids Res 43 (Database issue):
S, Lerman JA et al (2015) BiGG models: a D447–D452
platform for integrating, standardizing and
sharing genome-scale models. Nucleic Acids 34. Khan A, Fornes O, Stigliani A, Gheorghe M,
Res 44 (Database issue):D515–D522 Castro-Mondragon JA, van der Lee R et al
(2018) JASPAR 2018: update of the open-
23. Brunk E, Sahoo S, Zielinski DC, Altunkaya A, access database of transcription factor binding
Dr€ager A, Mih N et al (2018) Recon3D profiles and its web framework. Nucleic Acids
enables a three-dimensional view of gene varia- Res 46 (Database issue):D260–D266
tion in human metabolism. Nat Biotechnol 36
(3):272–281 35. Agarwal V, Bell GW, Nam JW, Bartel DP
(2015) Predicting effective microRNA target
24. Slenter DN, Kutmon M, Hanspers K, Riutta A, sites in mammalian mRNAs. Elife 4:e05005
Windsor J, Nunes N et al (2017) WikiPath-
ways: a multifaceted pathway database bridging 36. Han H, Shim H, Shin D, Shim JE, Ko Y, Shin J
metabolomics to other omics research. Nucleic et al (2015) TRRUST: a reference database of
Acids Res 46 (Database issue):D661–D667 human transcriptional regulatory interactions.
Sci Rep 5:11432
25. Altman T, Travers M, Kothari A, Caspi R, Karp
PD (2013) A systematic comparison of the 37. Chou C-H, Shrestha S, Yang C-D, Chang N-
MetaCyc and KEGG pathway databases. BMC W, Lin Y-L, Liao K-W et al (2017) miRTarBase
Bioinformatics 14(1):112 update 2018: a resource for experimentally
validated microRNA-target interactions.
26. Alcántara R, Axelsen KB, Morgat A, Belda E, Nucleic Acids Res 46 (Database issue):
Coudert E, Bridge A et al (2011) Rhea—a D296–D302
manually curated resource of biochemical reac-
tions. Nucleic Acids Res 40(D1):D754–D760 38. Karagkouni D, Paraskevopoulou MD, Chatzo-
poulos S, Vlachos IS, Tastsoglou S, Kanellos I
27. Davis AP, Grondin CJ, Johnson RJ, Sciaky D, et al (2017) DIANA-TarBase v8: a decade-long
McMorran R, Wiegers J et al (2018) The collection of experimentally supported
 Network-Based Multi-omics Integration 485

miRNA–gene interactions. Nucleic Acids Res 51. Ideker T, Ozier O, Schwikowski B, Siegel AF
46 (Database issue):D239–D245 (2002) Discovering regulatory and signalling
39. Shoemaker RH (2006) The NCI60 human circuits in molecular interaction networks. Bio-
tumour cell line anticancer drug screen. Nat informatics 18(suppl_1):S233–S240
Rev Cancer 6(10):813–823 52. Khurana V, Peng J, Chung CY, Auluck PK,
40. Tomczak K, Czerwińska P, Wiznerowicz M Fanning S, Tardiff DF et al (2017) Genome-
(2015) The cancer genome atlas (TCGA): an scale networks link neurodegenerative disease
immeasurable source of knowledge. Contemp genes to α-synuclein through specific molecu-
Oncol 19(1A):A68 lar pathways. Cell Syst 4(2):157–170. e14
41. The Integrative HMP (iHMP) Research Net- 53. Sychev ZE, Hu A, DiMaio TA, Gitter A, Camp
work Consortium (2014) The Integrative ND, Noble WS et al (2017) Integrated systems
human microbiome project: dynamic analysis biology analysis of KSHV latent infection
of microbiome-host omics profiles during per- reveals viral induction and reliance on peroxi-
iods of human health and disease. Cell Host some mediated lipid metabolism. PLoS Pathog
Microbe 16(3):276–289 13(3):e1006256
42. Laakso M, Kuusisto J, Stančáková A, Kuulas- 54. Beisser D, Klau GW, Dandekar T, Müller T,
maa T, Pajukanta P, Lusis AJ et al (2017) The Dittrich MT (2010) BioNet: an R-package for
metabolic syndrome in men study: a resource the functional analysis of biological networks.
for studies of metabolic and cardiovascular dis- Bioinformatics 26(8):1129–1130
eases. J Lipid Res 58(3):481–493 55. Alcaraz N, List M, Dissing-Hansen M,
43. Tadaka S, Saigusa D, Motoike IN, Inoue J, Rehmsmeier M, Tan Q, Mollenhauer J et al
Aoki Y, Shirota M et al (2017) jMorp: Japanese (2016) Robust de novo pathway enrichment
multi Omics reference panel. Nucleic Acids Res with KeyPathwayMiner 5. F1000Res 5:1531
46(D1):D551–D557 56. Anvar MS, Minuchehr Z, Shahlaei M, Kheitan
44. Nica AC, Parts L, Glass D, Nisbet J, Barrett A, S (2018) Gastric cancer biomarkers; a systems
Sekowska M et al (2011) The architecture of biology approach. Biochem Biophys Rep
gene regulatory variation across multiple 13:141–146
human tissues: the MuTHER study. PLoS 57. Jha AK, Huang SC-C, Sergushichev A, Lam-
Genet 7(2):e1002003 propoulou V, Ivanova Y, Loginicheva E et al
45. Perez-Riverol Y, Bai M, da Veiga Leprevost F, (2015) Network integration of parallel meta-
Squizzato S, Park YM, Haug K et al (2017) bolic and transcriptional data reveals metabolic
Discovering and linking public omics data sets modules that regulate macrophage polariza-
using the omics discovery index. Nat Biotech- tion. Immunity 42(3):419–430
nol 35(5):406–409 58. Chen X, Liu M-X, Yan G-Y (2012) Drug–tar-
46. Yugi K, Kubota H, Hatano A, Kuroda S (2016) get interaction prediction by random walk on
Trans-omics: how to reconstruct biochemical the heterogeneous network. Mol BioSyst 8
networks across multiple ‘omic’layers. Trends (7):1970–1978
Biotechnol 34(4):276–290 59. Liu Y, Zeng X, He Z, Zou Q (2017) Inferring
47. Zhou G, Xia J (2018) OmicsNet: a web-based microRNA-disease associations by random
tool for creation and visual analysis of walk on a heterogeneous network with multi-
biological networks in 3D space. Nucleic ple data sources. IEEE/ACM transactions on
Acids Res 46(W1):W514–W522 computational biology. Bioinformatics 14
48. Creixell P, Reimand J, Haider S, Wu G, Shibata (4):905–915
T, Vazquez M et al (2015) Pathway and net- 60. Chen X, You Z-H, Yan G-Y, Gong D-W
work analysis of cancer genomes. Nat Methods (2016) IRWRLDA: improved random walk
12(7):615–621 with restart for lncRNA-disease association
49. Akhmedov M, Kedaigle A, Chong RE, Mon- prediction. Oncotarget 7(36):57919
temanni R, Bertoni F, Fraenkel E et al (2017) 61. Subramanian A, Tamayo P, Mootha VK,
PCSF: an R-package for network-based inter- Mukherjee S, Ebert BL, Gillette MA et al
pretation of high-throughput data. PLoS (2005) Gene set enrichment analysis: a knowl-
Comput Biol 13(7):e1005694 edge-based approach for interpreting genome-
50. Tuncbag N, Gosline SJ, Kedaigle A, Soltis AR, wide expression profiles. Proc Natl Acad Sci U
Gitter A, Fraenkel E (2016) Network-based S A 102(43):15545–15550
interpretation of diverse high-throughput 62. Tarca AL, Draghici S, Khatri P, Hassan SS,
datasets through the omics integrator software Mittal P, Kim JS et al (2009) A novel signaling
package. PLoS Comput Biol 12(4):e1004879 pathway impact analysis. Bioinformatics 25
(1):75–82
486 Guangyan Zhou et al.

63. Alexeyenko A, Lee W, Pernemalm M, Guegan of response to vaccination in humans. Cell 169
J, Dessen P, Lazar V et al (2012) Network (5):862–877. e17
enrichment analysis: extension of gene-set 77. Gardinassi LG, Arévalo-Herrera M, Herrera S,
enrichment analysis to gene networks. BMC Cordy RJ, Tran V, Smith MR et al (2018)
Bioinformatics 13(1):226 Integrative metabolomics and transcriptomics
64. Glaab E, Baudot A, Krasnogor N, Schneider R, signatures of clinical tolerance to Plasmodium
Valencia A (2012) EnrichNet: network-based vivax reveal activation of innate cell immunity
gene set enrichment analysis. Bioinformatics and T cell signaling. Redox Biol 17:158–170
28(18):i451–i457 78. Pavlopoulos GA, Malliarakis D, Papanikolaou
65. Dettmer K, Aronov PA, Hammock BD (2007) N, Theodosiou T, Enright AJ, Iliopoulos I
Mass spectrometry-based metabolomics. Mass (2015) Visualizing genome and systems biol-
Spectrom Rev 26(1):51–78 ogy: technologies, tools, implementation tech-
66. da Silva RR, Dorrestein PC, Quinn RA (2015) niques and trends, past, present and future.
Illuminating the dark matter in metabolomics. Gigascience 4(1):38
Proc Natl Acad Sci U S A 112 79. Shannon P, Markiel A, Ozier O, Baliga NS,
(41):12549–12550 Wang JT, Ramage D et al (2003) Cytoscape: a
67. Albert R, Jeong H, Barabasi AL (2000) Error software environment for integrated models of
and attack tolerance of complex networks. biomolecular interaction networks. Genome
Nature 406(6794):378–382 Res 13(11):2498–2504
68. Li S, Park Y, Duraisingham S, Strobel FH, 80. Alcaraz N, Pauling J, Batra R, Barbosa E, Junge
Khan N, Soltow QA et al (2013) Predicting A, Christensen AG et al (2014) KeyPathway-
network activity from high throughput meta- Miner 4.0: condition-specific pathway analysis
bolomics. PLoS Comput Biol 9(7):e1003123 by combining multiple omics studies and net-
69. Xu X, Araki K, Li S, Han JH, Ye L, Tan WG et works with Cytoscape. BMC Syst Biol 8(1):99
al (2014) Autophagy is essential for effector 81. Kutmon M, van Iersel MP, Bohler A, Kelder T,
CD8(+) T cell survival and memory formation. Nunes N, Pico AR et al (2015) PathVisio 3: an
Nat Immunol 15(12):1152–1161 extendable pathway analysis toolbox. PLoS
70. Li S, Todor A, Luo R (2016) Blood transcrip- Comput Biol 11(2):e1004085
tomics and metabolomics for personalized 82. Luo W, Pant G, Bhavnasi YK, Blanchard SG Jr,
medicine. Comput Struct Biotechnol J 14:1–7 Brouwer C (2017) Pathview web: user friendly
71. Stewart CJ, Embleton ND, Marrs ECL, Smith pathway visualization and data integration.
DP, Fofanova T, Nelson A et al (2017) Longi- Nucleic Acids Res 45(W1):W501–W508
tudinal development of the gut microbiome 83. Garcia-Alcalde F, Garcia-Lopez F, Dopazo J,
and metabolome in preterm neonates with Conesa A (2010) Paintomics: a web based tool
late onset sepsis and healthy controls. Micro- for the joint visualization of transcriptomics
biome 5(1):75 and metabolomics data. Bioinformatics 27
72. Huan T, Forsberg EM, Rinehart D, Johnson (1):137–139
CH, Ivanisevic J, Benton HP et al (2017) Sys- 84. Kuo T-C, Tian T-F, Tseng YJ (2013) 3Omics: a
tems biology guided by XCMS online metabo- web-based systems biology tool for analysis,
lomics. Nat Methods 14(5):461–462 integration and visualization of human tran-
73. Chong J, Soufan O, Li C, Caraus I, Li S, Bour- scriptomic, proteomic and metabolomic data.
que G et al (2018) MetaboAnalyst 4.0: towards BMC Syst Biol 7(1):64
more transparent and integrative metabolomics 85. Sommer B, Baaden M, Krone M, Woods A
analysis. Nucleic Acids Res 46(W1): (2018) From virtual reality to immersive ana-
W486–W494 lytics in Bioinformatics. J Integr Bioinform 15
74. Pirhaji L, Milani P, Leidl M, Curran T, Avila- (2):20180043
Pacheco J, Clish CB et al (2016) Revealing 86. Argelaguet R, Velten B, Arnol D, Dietrich S,
disease-associated pathways by network inte- Zenz T, Marioni JC et al (2018) Multi-Omics
gration of untargeted metabolomics. Nat factor analysis-a framework for unsupervised
Methods 13(9):770 integration of multi-omics data sets. Mol Syst
75. Rohart F, Gautier B, Singh A, Lê Cao K-A Biol 14(6):e8124
(2017) mixOmics: an R package for ‘omics 87. Csardi G, Nepusz T (2006) The igraph soft-
feature selection and multiple data integration. ware package for complex network research.
PLoS Comput Biol 13(11):e1005752 InterJournal, Complex Systems 1695(5):1–9
76. Li S, Sullivan NL, Rouphael N, Yu T, Banton S, 88. Hagberg A, Swart P, S Chult D (2008) Explor-
Maddur MS et al (2017) Metabolic phenotypes ing network structure, dynamics, and function
using NetworkX (No. LA-UR-08-05495; LA-
 Network-Based Multi-omics Integration 487

UR-08-5495). Los Alamos National Lab. interpretation and integrative exploration.

(LANL), Los Alamos, NM (United States) Plos One 12(1):e0171046
89. Bastian M, Heymann S, & Jacomy M (2009) 91. Gao J, Tarcea VG, Karnovsky A, Mirel BR,
Gephi: an open source software for exploring Weymouth TE, Beecher CW, Cavalcoli JD,
and manipulating networks. In: Third interna- Athey BD, Omenn GS, Burant CF, Jagadish
tional AAAI conference on weblogs and social HV (2010) Metscape: a Cytoscape plug-in for
media visualizing and interpreting metabolomic data
90. Wanichthanarak K, Fan S, Grapov D, Barupal in the context of human metabolic networks.
DK, Fiehn O, Orešič M (2017) Metabox: a Bioinformatics 26(7):971–973
toolbox for metabolomic data analysis,
 INDEX

A Data processing .................................. v, 5, 11–22, 49–60,

79, 149, 239, 337, 338, 340, 341, 413, 419, 420,
Alignment ................................. 6, 11–15, 17, 22, 26, 27, 426, 428, 430, 451, 452, 454, 456, 459, 460
30, 38, 40, 42, 45, 54, 55, 58, 77, 79, 87, 88, 212, Data science.................................................... 7, 245–263,
221, 231, 240, 246, 341
265–311, 413
Annotation ................................................. v, 6, 9, 12, 13, Data visualization .................................... 7, 51, 202, 246,
19–21, 26, 47, 76, 87, 122, 123, 131, 141, 247, 256, 257, 266, 270, 279, 361–385, 392,
150–152, 185–204, 209–223, 227–241, 350,
419, 420, 442
363, 372, 389, 393–397, 429, 435–437, 439, Disease ...................................... 2, 63, 99, 121, 139, 166,
452, 453, 471, 475, 477 227, 313, 353, 391, 401, 420, 447, 469
Automated data analysis pipeline
Docker ................................... 7, 246, 249, 256, 265–311
(ADAP) ......................................................... 25–47 Drug mechanisms ........................................... 8, 419–444
B E
Bioinformatics .................................................. 64, 66, 78, Enrichment analysis ....................................... 85, 89, 123,
89, 90, 141, 228, 245–263, 265, 272, 299, 337, 127, 133, 338, 346–348, 350, 351, 354, 357,
358, 388, 392, 413, 450, 453, 470
472, 474–476, 478
Biomarkers............................................. 8, 63, 69, 79–83, Environmental health ......................................... 447–461
88, 90, 149, 185, 313, 314, 317, 319, 325, 341, Exposome ............................... v, 5, 8, 150, 447–461, 479
343, 401, 402, 404–406, 408, 409, 412, 413,
Extraction ........................................ 67, 70, 80, 122, 421
448, 454, 455, 459, 473, 475
Blood ........................................ 9, 70, 71, 105, 113, 124, F
125, 178, 357, 401–403, 407–412, 421, 455,
456, 460, 473, 478, 479 Feature annotation..................................... 227–241, 422
Feature extraction ........................................................ 219
C Flux ...................................... 66, 99–102, 104, 108, 109,
111–115, 157, 362–365, 374, 379, 380, 383,
Cardiometabolic disease ..................................... 401–414 384, 388
Cardiovascular disease......................................... 401, 404
Cloud computing.......................................................... 87, G
90, 248–250
Gas chromatography mass spectrometry
D (GC-MS)..................................................... 3, 6, 7,
25–27, 38, 41–47, 141, 179, 181, 187, 210, 214
Data analysis ................................................ v, 5, 7, 11, 22, Genome scale metabolic networks ................................. 8,
25, 26, 49–51, 58, 63, 66, 67, 69, 72, 75–89, 246, 351, 361–385
254, 255, 257, 261, 266, 314, 337–358, 384,
Git ....................................................... 246, 253, 265–311
392, 397, 413, 422, 442, 470 Global Natural Product Social Molecular Networking
Databases .......................................... v, 2, 12, 56, 64, 123, (GNPS) .......................................... 143, 150, 200,
140, 149, 165, 185, 210, 248, 347, 363, 388,
228–230, 233–241, 455
422, 453, 470
Data exploration................................ 165–183, 212, 247, H
260, 279, 342–344, 354, 355, 392, 480, 481
Data integration ...................................... 9, 26, 337–358, Hierarchical community networks ............ 470, 478–480
361–385, 396, 411, 459, 472, 479, 482 Human..........................................v, 1, 69, 122, 165, 227,
Data management ....................................... 49, 246, 247, 256, 322, 349, 365, 390, 405, 421, 448, 471
251–254, 256, 265 Human health ................................. v, 122, 451, 456, 461

Shuzhao Li (ed.), Computational Methods and Data Analysis for Metabolomics, Methods in Molecular Biology, vol. 2104,
https://doi.org/10.1007/978-1-0716-0239-3, © Springer Science+Business Media, LLC, part of Springer Nature 2020

489
 COMPUTATIONAL METHODS AND DATA ANALYSIS FOR METABOLOMICS
490 Index
Human metabolome database Metabolomics ............................................v, 2, 11, 25, 49,
(HMDB) .......................................... 6, 56, 58, 84, 61, 139, 149, 165, 185, 209, 227, 246, 265, 313,
140, 141, 144, 145, 150, 165–170, 173, 175, 337, 362, 388, 402, 419, 448, 469
176, 178–183, 199, 390, 437, 453, 455 Metabotyping ............................................. 62–65, 68–70,
73, 80–82, 88, 90
I METLIN ............................................. 6, 12, 20, 21, 140,
IDEOM ............................................................... 419–444 141, 149–161, 453–455
Integration..............................................9, 12, 22, 25, 87, MetScape .............................................................. 89, 390,
393, 394, 397
137, 197, 202, 280, 351, 354, 361–385, 395,
396, 411, 453, 459, 470, 472–478, 482 Microsoft Excel ................................. 239, 419, 422, 424
Interpretation ......................................................... v, 8, 87, Mode of action ........................................... 420, 421, 423
Molecular formula ........................................... 7, 57, 139,
88, 102, 108, 112, 114, 121–124, 133, 146, 221,
330, 337–358, 378, 421, 456, 469, 477–479 151, 153, 155, 183, 185–204
Isotope ................................................... 8, 12, 13, 52, 53, Molecular networking............................ 6, 211, 227–241
MS/MS spectra ......................................... 2, 3, 6, 20, 21,
62, 66, 71, 84, 85, 101–103, 113, 115, 116,
156–158, 186–189, 191–193, 210, 221, 231, 57, 140–144, 150–159, 161, 183, 187, 188, 228,
232, 240, 257, 388, 429, 434, 435, 438–440, 229, 237, 239, 241, 454–456, 460
Multi-omics integration............................... 19, 353–358,
443, 444, 453, 460, 479
396, 459, 469–483
L Multivariate data analysis ................................. 14, 81–83,
86–88, 343–346
LipidMaps.................................................. 122, 123, 125, Multivariate statistics................. 8, 38, 87, 344, 357, 482
128–131, 133, 145, 437, 453, 455 Mummichog............................................ 8, 21, 140, 146,
Lipidome .................................................... 122, 124–134 211, 350–353, 389, 390, 395–397, 456, 477, 478
Lipidomics .......................................................... 5, 9, 116, MZmine............................................... 6, 11, 25–47, 187,
121–134 229, 230, 232, 235, 236, 238–241, 256, 430, 452
Liquid chromatography (LC).............................. 2, 4, 13,
25, 210, 322, 452 N
Liquid chromatography-mass spectrometry
(LC-MS) ........................ 210, 227, 419–444, 450 Natural products ............................... 143, 210, 228, 455
Liquid-chromatography tandem mass spectrometry Network enrichment analysis .................... 472, 474, 476
Network visualization ....................... 357, 393, 480, 481
(LC-MS/MS) .................................. 13, 140, 145,
179, 181, 185, 187, 227, 229 Nuclear magnetic resonance
(NMR) ....................... 3, 61, 144, 166, 337, 450
M
O
Mass spectrometry (MS)......................................... 3, 5–9,
OpenMS ............................................................. 6, 49–59,
13, 14, 21, 25, 30, 49, 63, 66, 87, 88, 115, 121,
122, 134, 143, 145, 149, 160, 161, 166, 204, 187, 256
209–223, 257, 337, 372, 393, 422, 448, 477
P
MetaboAnalyst .......................................... 8, 41, 87, 261,
337–358, 392, 395, 397, 477, 478, 481 Pathway................................................ 3, 21, 63, 99, 130,
Metabolic activity ....................................... 100, 112, 115 140, 150, 166, 210, 228, 261, 338, 362, 387,
Metabolic modification................................................ 161 424, 450, 471, 505
Metabolic networks.......................................... 21, 66, 89, Pathway analysis ................................................... 5, 8, 21,
351, 361–385, 470 87–89, 346, 348, 349, 352, 354–357, 363,
Metabolite databases............................. 6, 139–147, 149, 387–397, 481
425, 427, 433, 444, 453–455 Peak............................................... 4, 11, 26, 50, 65, 104,
Metabolite identification ............................... 7, 8, 21, 26, 139, 153, 179, 187, 210, 231, 251, 315, 338,
56–58, 71, 72, 76, 78, 83–86, 88, 139–147, 393, 412, 420, 448, 477
149–155, 157, 158, 161, 170, 180, 185–201, Performance metrics ........................................... 324, 325
204, 211, 350, 395, 412, 413, 419, 422, 427, Plant ...................................................................... 5, 6, 90,
428, 440, 451, 453–456, 477 209–223, 392, 471
 COMPUTATIONAL METHODS AND DATA ANALYSIS FOR METABOLOMICS
Index 491
Plasma ........................................................ 63, 65, 68–70, Systems biology................................ 12, 19, 21, 66, 337,
72, 73, 100, 104, 105, 108, 112, 116, 122, 365, 470, 478, 481
124–127, 257, 322, 325, 405–407, 411, 479 Systems medicine ......................................................... 8, 9
Precision medicine .......................................... 1, 2, 9, 473
Predictive modeling ............................................. 38, 265, T
279, 313–331 Tandem mass spectrum ............................................... 139
Preprocessing............................................................... 5, 6,
25–47, 57, 67, 76, 78–81, 87, 187–189, 191, U
320, 341, 358, 432, 451–453
Processing .................................................. v, 5, 11–22, 26, Unknown metabolite ................................... 84, 150–156,
28, 35, 49–59, 66, 79, 86–88, 149, 187, 210, 219, 158, 210, 212, 216, 217, 408
229, 239, 240, 249, 256, 262, 307, 337, 338, Untargeted ..................................... 8, 11, 13, 20, 25–27,
340–343, 365, 394, 395, 410, 411, 413, 419, 35, 47, 52–26, 58, 64, 67, 69, 85, 86, 121–133,
420, 423, 426–428, 430–442, 444, 451, 452, 141, 151, 157, 161, 186, 314, 315, 322, 338,
454, 456, 459, 460 350–353, 358, 387–398, 412, 419, 444, 448,
Python ................................................. 49, 51, 79, 81, 87, 450, 451, 453, 459, 460, 479
246, 250, 254–256, 261, 265–311, 364, 365, Untargeted metabolomics ..................................... 11, 13,
375, 379, 380, 384, 395 25, 27, 35, 69, 141, 151, 157, 314, 338,
350–353, 358, 387–398, 419, 450, 451, 453,
Q 459, 460, 479

Quality control ............................................ 7, 22, 50, 67, V

72–75, 82, 88, 256, 257, 260, 422
Version control ...................................... 7, 246, 253, 254,
R 265, 266, 280–294
Virtualization...................................................... 246, 249,
R.............................................. 18, 22, 86, 143, 213–218, 265, 310
220, 223, 246, 254, 265–311, 338, 358, 364,
Visualization ................................... 7, 22, 27, 28, 30–32,
373, 375, 385, 425, 426, 430, 453, 458 41, 47, 50, 51, 53, 81, 86–89, 146, 189, 202, 219,
Recommendation ............................................ 7, 101, 322 220, 223, 228, 230, 231, 236, 238, 241, 246,
Renal disease...................................... 401, 402, 408, 409 247, 256, 260, 261, 265, 266, 279, 281, 307,
Replication.................................................. 403, 407, 412
314, 345, 356, 357, 361–385, 392, 393, 397,
Reproducible science ............................. 3, 6, 61, 66, 262 419, 427, 442, 470, 480, 482
S W
Scripting................................. 49, 51, 246, 249, 254–256 Web server ................................................... 84, 141, 143,
Secondary metabolism ................................................. 185
144, 146, 249, 338, 358
Sirius ....................................................... 57, 58, 185–204 Workflows ......................................................... 11–15, 18,
Specialized metabolites ...................................... 209–223, 26, 27, 41–47, 49, 51, 52, 57, 58. 59, 64, 66–72,
227, 228 79, 86–90, 151, 187, 202–204, 227–230,
Spectra .......................................... 2, 12, 26, 50, 65, 139,
234–236, 238, 240, 241, 246, 281, 298, 300,
150, 166, 185, 211, 228, 307, 314, 338, 451, 477 338, 341–353, 358, 374, 384, 419–444,
453, 474
Spectral database .............. 140, 141, 143–145, 149–161
Structure elucidation .................. 64, 186, 188, 200, 202 X
Structure prediction ................................................ 6, 140
Study design ........................................ 1–9, 67, 101–112, XCMS ........................................................... 6, 11–22, 35,
327, 358, 395, 403, 405, 407, 409, 413, 452, 461 210, 213, 214, 249, 256, 272, 307, 308, 315,
Supervised learning ............................................ 314, 315, 373, 395, 397, 420, 422, 425–247, 444,
320, 325, 331 452–454, 477, 483