Basant K.

Tiwary

Bioinformatics
and Computational
Biology
A Primer for Biologists
Bioinformatics and Computational Biology
Basant K. Tiwary

Bioinformatics
and Computational
Biology
A Primer for Biologists
Basant K. Tiwary
Department of Bioinformatics, School of Life Sciences
Pondicherry University
Puducherry, India

ISBN 978-981-16-4240-1 ISBN 978-981-16-4241-8 (eBook)

https://doi.org/10.1007/978-981-16-4241-8

# Springer Nature Singapore Pte Ltd. 2022

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, expressed 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 Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore
Preface

The goal of writing this book is to explain the theoretical foundation and mathemati-
cal concepts of bioinformatics and computational biology to a biologist who is new
to this subject. This book is aimed at undergraduate and graduate students of any
biological discipline like life sciences, genetics, medicine, agriculture, animal hus-
bandry and bioengineering in addition to the students of bioinformatics and compu-
tational biology. Like any other technical subject, bioinformatics has its own
language and terms. All terms are explained in simple language avoiding mathemat-
ical jargon. This textbook starts with basic theoretical concepts in bioinformatics
building in a stepwise fashion leading to an optimum level where the application of
bioinformatics in medicine, agriculture, animal husbandry and bioengineering is
apparent. This work examines the underlying principles and methods in bioinfor-
matics applied to answer real biological problems. The last four chapters are
completely applied in nature although theoretical concepts discussed in earlier
eight chapters are a prerequisite for understanding the applications. The exercises
and multiple-choice questions at the end of each chapter will reinforce the under-
standing of the concepts discussed in the text. The worked-out answers to exercises
in each chapter will provide a thorough understanding of the problems in bioinfor-
matics and their stepwise solutions. Some useful references are added at the end of
each chapter for further reading.
I am indebted to all my teachers who have shaped my understanding of interdisci-
plinary science. I thank my Ph.D. advisor, Prof. Arun K. Ray, who has given me the
freedom to explore the interdisciplinary and experimental biological research at Bose
Institute. I express my heartfelt gratitude to Prof. Wen-Hsiung Li, James Watson
Professor at the University of Chicago to initiate me in the field of computational
genomics during my stay in his lab as a visiting faculty. I am also grateful to all
research workers across the world in the field of bioinformatics and computational
biology for building the knowledge base of the book. I thank all my colleagues and
students at the Department of Bioinformatics, Pondicherry University, for creating a
wonderful teaching-learning environment. Special thanks to Dr. R. Krishna for
reading and commenting on the chapter on structural bioinformatics. I also thank the
team at Springer Nature for making this academic endeavour successful.

Puducherry, India Basant K. Tiwary

v
Contents

1 Introduction to Bioinformatics and Computational Biology . . . . . . 1

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Definitions of Bioinformatics and Computational Biology . . . . . 2
1.3 Interdisciplinary Science . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 Computational Thinking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.5 Bioinformatics Skills . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.6 Historical Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.7 Goals of the Book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.8 Multiple Choice Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Suggested Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2 Biological Databases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Types of Biological Databases . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2.1 Sequence and Structure Databases . . . . . . . . . . . . . . . . 13
2.2.2 Expression Databases . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2.3 Pathway Databases . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.2.4 Disease Databases . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.5 Organism-Specific and Virus Databases . . . . . . . . . . . . 23
2.3 Database Searching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.5 Multiple Choice Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Suggested Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3 Statistical Computing Using R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.1 Introduction to R Language . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2 Data Structures in R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.3 R Packages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.4 Data Input and Output in R . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.5 Biological Data Analysis in R . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.5.1 Statistical Distributions . . . . . . . . . . . . . . . . . . . . . . . . 36
3.5.2 Testing of Hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.5.3 Parametric Tests Vs. Non-parametric Tests . . . . . . . . . 42

vii
viii Contents

3.5.4 Common Statistical Tests . . . . . . . . . . . . . . . . . . . . . . 42

3.5.5 Graphics in R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.5.6 R Packages for Graphics . . . . . . . . . . . . . . . . . . . . . . . 47
3.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.7 Multiple Choice Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
Suggested Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4 Sequence Alignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.2 Pairwise Sequence Alignment . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.3 Multiple Sequence Alignment . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.4 Alignment Visualization and Editing . . . . . . . . . . . . . . . . . . . . 58
4.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.6 Multiple Choice Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
Suggested Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5 Structural Bioinformatics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.2 Protein Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5.3 Modelling of Protein Structure . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.3.1 Homology Modelling . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.3.2 Ab Initio Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.3.3 Threading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.3.4 Integrative (Hybrid) Modelling . . . . . . . . . . . . . . . . . . 69
5.4 Structural Genomics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.5 Scoring Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.6 Model Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.7 Classification of Protein Structures . . . . . . . . . . . . . . . . . . . . . . 72
5.8 Molecular Dynamics Simulations . . . . . . . . . . . . . . . . . . . . . . . 73
5.9 Docking of Ligands and Proteins . . . . . . . . . . . . . . . . . . . . . . . 75
5.10 Structure-Based Drug Design . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.11 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.12 Multiple Choice Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
Suggested Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6 Molecular Evolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.2 Neutral and Nearly Neutral Theories of Evolution . . . . . . . . . . . 88
6.3 The Concept of Homology . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
6.4 Genetic Distances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
6.5 Nucleotide Substitution Models . . . . . . . . . . . . . . . . . . . . . . . . 91
6.6 Phylogenetic Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
6.7 Methods for Reconstruction of Phylogenetic Tree . . . . . . . . . . . 93
6.7.1 Distance-Based Methods . . . . . . . . . . . . . . . . . . . . . . . 94
6.7.2 Maximum Parsimony . . . . . . . . . . . . . . . . . . . . . . . . . 96
6.7.3 Maximum Likelihood Methods . . . . . . . . . . . . . . . . . . 97
Contents ix

6.7.4 Bayesian Phylogenetic Inference . . . . . . . . . . . . . . . . . 98

6.8 Molecular Clock . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.9 Ohno Hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
6.10 Molecular Signatures of Selection . . . . . . . . . . . . . . . . . . . . . . 101
6.11 Evolution of Protein Structure . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.12 Evolution of Viruses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6.13 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
6.14 Multiple Choice Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
Suggested Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
7 Next-Generation Sequencing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
7.2 Sequencing Platforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
7.2.1 454 (Pyrosequencing) . . . . . . . . . . . . . . . . . . . . . . . . . 118
7.2.2 Illumina (Sequencing-by-Synthesis) . . . . . . . . . . . . . . . 119
7.2.3 SOLiD (Sequencing by Ligation) . . . . . . . . . . . . . . . . 119
7.2.4 Ion Torrent (Semiconductor Sequencing) . . . . . . . . . . . 120
7.2.5 PacBio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
7.2.6 Oxford Nanopore . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
7.3 Data Formats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
7.4 Visualization of NGS Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
7.5 Alignment of NGS Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
7.6 Reference-Based Assembly . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
7.6.1 Seed-and-Extend Method . . . . . . . . . . . . . . . . . . . . . . 124
7.6.2 Suffix Array and Suffix Tree-Based Alignment . . . . . . 125
7.6.3 Burrows–Wheeler Transformation (BWT)-Based
Alignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
7.7 De Novo Assembly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
7.7.1 Overlap-Layout-Consensus (OLC) . . . . . . . . . . . . . . . . 127
7.7.2 de Bruijn-Graph (DBG) . . . . . . . . . . . . . . . . . . . . . . . 128
7.8 Scaffolding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
7.9 Applications of Next-Generation Sequencing . . . . . . . . . . . . . . 129
7.9.1 Whole-Genome Sequencing . . . . . . . . . . . . . . . . . . . . 129
7.9.2 Transcriptomics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
7.9.3 Epigenomics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
7.9.4 Exome Sequencing . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
7.10 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
7.11 Multiple Choice Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
Suggested Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
8 Systems Biology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
8.2 Complex Biological System . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
8.3 Computational Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
x Contents

8.4 Network Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138

8.5 Genome-Scale Metabolic Model (GSMM) . . . . . . . . . . . . . . . . 143
8.6 Kinetic Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
8.7 Motifs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
8.8 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
8.9 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
8.10 Multiple Choice Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
Suggested Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
9 Clinical Bioinformatics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
9.2 NGS Applications in Clinical Research . . . . . . . . . . . . . . . . . . 164
9.3 Network Medicine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
9.4 Biomarker Discovery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
9.5 Multi-Target Drugs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
9.6 Artificial Intelligence in Medicine . . . . . . . . . . . . . . . . . . . . . . 168
9.6.1 Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
9.6.2 Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
9.7 Genome-Wide Association Mapping . . . . . . . . . . . . . . . . . . . . 173
9.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
9.9 Multiple Choice Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
Suggested Readings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
10 Agricultural Bioinformatics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
10.2 Pan-Genome of Crops . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
10.3 Assembly of Crop Genomes . . . . . . . . . . . . . . . . . . . . . . . . . . 188
10.4 Identification of Homeologs . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
10.5 Genomic Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
10.6 Crop Phenomics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
10.7 Crop Systems Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
10.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
10.9 Multiple Choice Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
Suggested Readings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
11 Farm Animal Informatics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
11.2 Whole-Farm Systems Model . . . . . . . . . . . . . . . . . . . . . . . . . . 204
11.3 Nutritional Models of Farm Animals . . . . . . . . . . . . . . . . . . . . 205
11.3.1 Dairy Cattle Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
11.3.2 Pig Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
11.3.3 Sheep and Goat Model . . . . . . . . . . . . . . . . . . . . . . . . 207
11.3.4 Laying Hen Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
11.3.5 Lactation Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
11.4 Livestock Genomics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
11.5 Livestock Phenomics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
Contents xi

11.6 Genomic Selection in Farm Animals . . . . . . . . . . . . . . . . . . . . 212

11.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
11.8 Multiple Choice Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
Suggested Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
12 Computational Bioengineering . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
12.2 Control and Systems Theory in Biology . . . . . . . . . . . . . . . . . . 220
12.3 Strategies in Bioengineering . . . . . . . . . . . . . . . . . . . . . . . . . . 221
12.4 Metabolic Engineering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
12.5 Evolutionary Engineering . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
12.6 Synthetic Biology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
12.7 Computational Design of Synthetic Systems . . . . . . . . . . . . . . . 226
12.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
12.9 Multiple Choice Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
Suggested Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
Introduction to Bioinformatics
and Computational Biology 1

Learning Objectives
You will be able to understand the following after reading this chapter:

• Definition of bioinformatics and computational biology.

• Connection between bioinformatics and other biological and physical
sciences.
• Difference between computational thinking and computational
programming.
• Evolution of bioinformatics as an interdisciplinary subject.

1.1 Introduction

Biology has undergone a sea change in the last two decades and transformed itself
into a true quantitative science akin to mathematics, physics and statistics. Although
biology was fortunate enough to be enriched by great mathematicians like
G.H. Hardy, R.A. Fisher, J.B.S. Haldane and Sewall Wright during the last century,
it continued to remain largely a descriptive natural subject with less emphasis on
quantitative methods until the end of last century. The advent of fast microcomputers
and advanced molecular biology techniques such as PCR, microarray and next-
generation sequencing has led to the emergence of high-throughput data in biology.
Surprisingly, the generation of genomic data has surpassed astronomical data in
terms of volume and became the most abundant available big data today. These
drastic developments in generation of biological data has necessitated a parallel
development of computational methods to analyse the data in a meaningful way and
ultimately led to the emergence of a new discipline known as bioinformatics.
Although biology and bioinformatics are currently being treated as separate

# Springer Nature Singapore Pte Ltd. 2022 1

B. K. Tiwary, Bioinformatics and Computational Biology,
https://doi.org/10.1007/978-981-16-4241-8_1
2 1 Introduction to Bioinformatics and Computational Biology

disciplines, the time is not far away when a thin line of division between these two
disciplines will gradually disappear in favour of a more holistic biology.

1.2 Definitions of Bioinformatics and Computational Biology

Bioinformatics and computational biology has been defined as distinct subjects

initially albeit a significant overlap at the interface of two disciplines. The NIH
Biomedical Information Science and Technology Initiative Consortium in the year
2000 proposed the working definitions of Bioinformatics and Computational Biol-
ogy. Bioinformatics was defined as the research, development or application of
computational tools and approaches for expanding the use of biological, medical,
behavioural or health data, including those to acquire, store, organize, archive,
analyse or visualize such data. On the other hand, computational biology was
defined as the development and application of data-analytical and theoretical
methods, mathematical modelling and computational simulation techniques to the
study of biological, behavioural and social systems. In simple words, bioinformatics
is the biological application of information technology with focus on data storage,
whereas computational biology is an application of information technology in
understanding biology with more emphasis on analytical algorithms. In other
words, computational biology is a computational approach to analyse and under-
stand biological processes. However, under current scenario, many areas in these
two disciplines are overlapping, henceforth, we will use the term interchangeably in
the rest of the book.

1.3 Interdisciplinary Science

The development of bioinformatics occurred at the interface of various scientific

disciplines such as biology, computer science, mathematics and statistics, physics
and chemistry (Fig. 1.1). The key areas in biology required for understanding
bioinformatics are biophysics, biochemistry, cell and molecular biology, genomics
and evolutionary biology. A good understanding of some fundamental concepts in
computer science such as programming, database, data structures, machine learning
and artificial intelligence are prerequisite for learning bioinformatics. In addition,
biostatistics, probability theory, linear algebra, discrete mathematics, differential
equations, Bayesian statistics and calculus have contributed significantly in the
development of bioinformatics. Bioinformatics also draws ideas from fundamental
concepts in physics and chemistry for finding solutions to a computational problem.
1.5 Bioinformatics Skills 3

Fig. 1.1 Bioinformatics as an interdisciplinary subject at the interface of biological and physical
sciences

1.4 Computational Thinking

Can we develop expertise in bioinformatics with minimal knowledge in computer

programming? Surprisingly the answer is yes! The simple logic behind this para-
doxical answer is that computational thinking is different from computer program-
ming. Computational thinking is a logical thought process encompassing
formulation of a complex problem and its subsequent possible computational
solutions in four steps (Fig. 1.2). First, a complex problem is broken down into
smaller parts (decomposition), followed by finding some similarities among smaller
parts (pattern recognition), and focussing on key information (signals) only avoiding
concomitant unnecessary noise (abstraction) within and among the parts. The final
step of computational thinking is the development of solution to a problem in
stepwise manner (algorithms). On the other hand, computer programming is merely
one aspect of software development having four successive steps: analysis, design,
implementation and evaluation. The first and last step of software development is
only based on computation thinking and does not need any knowledge in program-
ming. Thus, a computational biologist should be able to ask a biological question
and find its solution through algorithmic thinking. He should also be capable of
writing a new computational program or customize an existing program to his
biological analysis.

1.5 Bioinformatics Skills

A bioinformatics scientist is expected to have an array of computational skills. These

include ability to manage, store and analyse large biological datasets using available
algorithms and software. In addition, he should be an expert in mathematical and
4 1 Introduction to Bioinformatics and Computational Biology

Fig. 1.2 Four essential steps

in computational thinking

statistical modelling with strong foundation in the probability theory, graph theory,
descriptive and inferential statistics, differential equation and statistical program-
ming in the R language and environment. He is also expected to regularly maintain
and upgrade systems and servers in his lab. The knowledge of one scripting language
like R or Python will have an additional advantage. A good understanding in the core
biological subjects such as genetics, genomics, biochemistry, molecular biology and
evolution are prerequisite for a bioinformatics scientist. A continuous learning of
advanced technologies like next-generation sequencing and mass spectrometry is
also required. Above all, a bioinformatics scientist should be highly motivated and
dedicated person and have a teamwork spirit with excellent analytical ability.

1.6 Historical Perspective

Bioinformatics has undergone a magnificent journey with an extraordinary pace in

the last seven decades. The first biological problem solved by bioinformatics
approach was to decipher the protein primary structure using Edman peptide
sequencing data. Margaret Dayhoff, the mother and father of bioinformatics, devel-
oped a FORTRAN program called COMPROTEIN between 1958 and 1962 for
finding this solution. Emile Zuckerkandl and Linus Pauling started evolutionary
1.6 Historical Perspective 5

analysis of several protein biomolecules including haemoglobin in 1963. This

evolutionary analysis conclusively demonstrated two facts: orthologous proteins in
vertebrates have a high degree of similarity in sequence over a long evolutionary
period and the degree of differences among orthologous sequences from different
species were proportional to evolutionary distances between species. Zuckerkandl
and Pauling concluded that all these orthologous sequences were evolved from a
single common ancestor based on the above-mentioned facts. This novel idea in
evolutionary biology further paved the way for prediction of an ancestral sequence
from the available sequences of extant species.
Both molecular biology and computer science developed at fast pace albeit in
parallel from 1980 to 1990. The emergence of new molecular techniques like gene
cloning and PCR revolutionized the biological and clinical research. Concomitantly,
microcomputers were invented in 1977. Needleman and Wunsch developed an
algorithm for the first time for aligning protein sequences in 1970. This algorithm
further led to emergence of multiple sequence alignment (MSA) in early 1980s
although this algorithm was useful but impractical to apply due to their high
computational running time. Consequently, Feng and Doolittle in 1987 developed
another approach to MSA known as progressive sequence alignment based on a
guide tree. Most popular program for multiple sequence alignment CLUSTAL was
derived from this algorithm. First probabilistic model of amino acid substitution was
developed by Dayhoff, Schwartz and Orcutt in 1978 in form of point accepted
mutations (PAMs). With the introduction of this PAM matrix, amino acid
substitutions became a popular metrics for measuring evolutionary changes in a
sequence. The development of DNA sequencing technologies has paved the way for
whole-genome sequencing of many organisms.
After completion of human genome project, massive worldwide efforts started to
sequence various animal and plant genomes. DNA sequencing of genomes has
become much cheaper and faster with time creating huge amount of genome
sequences as a challenging big data for computational biologists. A new software,
Staden was developed to analyse the Sanger sequencing reads. The first phylogenetic
tree was reconstructed using protein sequences based on the least number of amino
acid changes (maximum parsimony method). But DNA sequences carry more
information than the protein sequences in terms of synonymous mutations which
does not manifest in a protein sequence. Statistically more robust method known as
maximum likelihood method was developed for inferring phylogenetic tree by
Felsenstein using DNA sequences. This method finds a tree which has maximum
probability of evolving the observed data. Bayesian approach to molecular phylog-
eny gained momentum in the 1990s and is a very popular and statistically most
robust method of phylogenetics.
The first software developed for sequence analysis was the GCG software suite in
1984 to manipulate sequences on small-scale mainframe computer. DNASTAR,
alternative software suite for personal computer was developed in the same year.
Three international databases, namely European Molecular Biology Laboratory
(EMBL) database, GenBank and DNA Data Bank of Japan (DDBJ) were developed
independently in Europe, the USA and Japan, respectively. Finally, these three
6 1 Introduction to Bioinformatics and Computational Biology

databases were integrated in 1986–1987 and are still existing under the umbrella of
the International Nucleotide Sequence Database Collaboration. Many scripting
languages were developed in the 1980s for application in bioinformatics. These
scripting languages are interpreted whenever launched and do not require compila-
tion from C code. Perl (Practical Extraction and Reporting Language) was developed
in 1987 and soon became a popular language to manipulate biological sequences.
Due its flexibility, it became very popular among bioinformaticians and improved
further in form of BioPerl in 1996. Today, R and Python are two major programming
languages for bioinformaticians.
The first genome of H. influenzae was sequenced by Craig Venter and his team in
1995. But human genome project was completed with combined private and public
efforts using Sanger sequencing technology in 2003. At present, genome sequencing
is much cheaper after the advent of second-generation or next-generation sequencing
technologies. The 454 pyrosequencing technology was the first next-generation
sequencing platform. But Illumina platforms like HiSeq soon became a popular
sequencing platform for high-throughput sequencing. This has led to development of
many alignment software for assembly of short reads generated by various sequenc-
ing platforms. Today, high-performance computing facility is being provided by
various public and private agencies.
Now, modern biology is moving from reductionist phase which focuses on a
single gene or a protein to a more holistic approach of systems biology. A mathe-
matical model representing a biological process can be generated from genome or
transcriptome or proteome or metabolome data. This endless journey of bioinfor-
matics from sequence analysis to systems biology will continue with a much faster
pace in near future. Computational biology is likely to be an integral part of
agriculture and medicine worldwide soon for a better food production and
healthcare.

1.7 Goals of the Book

This book is intended for biologists who have a basic understanding of mathematical
concepts. Bioinformatics is a diverse discipline with a potential to solve real
problems in biological, medical, agricultural, veterinary science and bioengineering.
This book is a humble attempt to communicate the concepts in bioinformatics to
experimental biologists with minimum emphasis on the underlying mathematical
principles. The mathematical descriptions are simplified in favour of a better under-
standing of biological processes. The first part of this book deals with fundamental
concepts in biological databases, statistical computing in R, sequence alignment,
structural bioinformatics and molecular evolution. A good concept in biological
database and sequence alignment are essential for understanding any advanced field
in bioinformatics. R language and environment has become the lingua franca of
bioinformatics and computational biology with largest collection of statistical
packages available for biological data analysis and visualization. I will discuss the
molecular structure and dynamics of macromolecules in connection to their
1.8 Multiple Choice Questions 7

functions in a chapter on structural bioinformatics. This will be followed by under-

standing the molecular basis of genome evolution in the next chapter on molecular
evolution with various methods of inferring evolutionary relationship from DNA
and protein sequences. The second part of this book will focus on two advanced
areas of bioinformatics, namely the next-generation sequencing and systems biol-
ogy. The third part of this book deals with application of bioinformatics in clinical,
agricultural, farm animal informatics and bioengineering research.
All chapters have exercises and multiple choice questions for self-assessment test
by the reader. Two exercises are included in each chapter to illustrate the application
of bioinformatics in various branches of basic and applied biological sciences. I wish
that this book will be a useful companion to not only experimental biologists and
other colleagues from different branches of physical and mathematical sciences but
to a computational biologist as well.

1.8 Multiple Choice Questions

1. Name a biological discipline which is NOT an integral part of bioinformatics:

(a) Molecular Biology
(b) Evolutionary Biology
(c) Environmental Biology
(d) Biochemistry
2. Who is known as the mother and father of bioinformatics?
(a) Margaret Dayhoff
(b) Francis Crick
(c) Sydney Brenner
(d) Motoo Kimura
3. Who proposed for the first time that orthologous sequences were derived from a
common ancestor?
(a) Zuckerkandl and Pauling
(b) Kimura and Nei
(c) Nei and Mayer
(d) Mayr and Goldman
4. Who developed progressive multiple sequence alignment based on a guide tree?
(a) Feng and Doolittle
(b) Smith and Waterman
(c) Wunsch and Needleman
(d) Smith and Needleman
5. The first method used for reconstruction of a phylogenetic tree from protein
sequences is:
(a) Maximum parsimony
(b) Maximum likelihood
(c) Neighbour joining
(d) UPGMA
8 1 Introduction to Bioinformatics and Computational Biology

6. Maximum likelihood method for reconstruction of a phylogenetic tree was

developed by:
(a) Felsenstein
(b) Kimura
(c) Nei
(d) Goldman
7. First software for sequence analysis is:
(a) GCG
(b) MEGA
(c) BIOEDIT
(d) PHYLIP
8. Which of the following is NOT a scripting language?
(a) R language
(b) Perl
(c) Python
(d) HTML
9. Which of the following is NOT as part of International nucleotide sequence
collaboration?
(a) GenBank
(b) EMBL
(c) DDBJ
(d) UNIPROT
10. Name the microorganism whose genome was sequenced for the first time:
(a) H. influenzae
(b) E. coli
(c) M. tuberculosis
(d) Mycobacterium pullorum

Answer: 1. c 2. a 3. a 4. a 5.a 6. a 7. a 8. d 9. d 10. a

Summary
• Bioinformatics and computational biology are integral parts of holistic biology.
• Bioinformatics deals with development of computational tools for storage and
analysis of biomedical data.
• Computational biology deals with mathematical modelling and simulations for
understanding biological systems.
• Computational thinking is more important than computer programming in devel-
oping bioinformatics and computational biology skills.
Suggested Reading 9

Suggested Reading
Moody G (2004) Digital code of life: how bioinformatics is revolutionizing science, medicine, and
business. Wiley, London
Mount D (2004) Bioinformatics: sequence and genome analysis, 2nd edn. Cold Spring Harbor
Laboratory Press, Cold Spring Harbor
Gauthier J, Vincent AT, Charette SJ, Derome N (November 2019) A brief history of bioinformatics.
Briefings in Bioinformatics 20(6):1981–1996
Biological Databases
2

Learning Objectives
You will be able to understand the following after reading this chapter:

• Different types of biological databases based on the nature of stored data.

• Disease databases focusing on an array of human genetic diseases.
• Organism-specific databases with special reference to economically impor-
tant plants and animals.
• Database searching using different variants of BLAST.

2.1 Introduction

Biological data have accumulated at a faster pace in the recent past due to advent of
high-throughput and cheaper next-generation sequencing technologies. Conse-
quently, new databases are being developed in order to manage this ever increasing
biological data. Databases are digital repositories based on a computerized software
for storage of information in a system and their retrieval from the system using
search tools. Biological databases are an important component of bioinformatics
research and are usually well-annotated and cross-referenced to other databases. The
primary objective of a database is to organize the data in a structured and searchable
form allowing easy retrieval of useful data. The simplest form of a database is a
single file containing multiple entries of same set of information. Currently, there are
more than thousand biological databases providing access to multifarious omics data
to biologists. Biological databases are classified based on different criteria such as
levels of data coverage and data curation. Based on the extent of data coverage,
biological databases consist of two main categories: comprehensive and specialized
databases. Comprehensive database such as GenBank includes a variety of data

# Springer Nature Singapore Pte Ltd. 2022 11

B. K. Tiwary, Bioinformatics and Computational Biology,
https://doi.org/10.1007/978-981-16-4241-8_2
12 2 Biological Databases

Table 2.1 Entrez databases

Database Area Description
PubMed Literature Biomedical abstracts and citations
PubMed central Literature Full text articles from journals
Nucleotide Genomes DNA and RNA sequences
Genome Genomes Genome sequencing projects of different species
Gene Genes Detailed information on gene loci
GEO profiles Genes Gene expression profiles
HomoloGene Genes Homologous genes from different species
Protein Proteins Protein sequences
Structure Proteins Biomolecular structures
PubChem compound Chemicals Chemical information of compounds with structures
Online Mendelian Genes A catalogue of human genes and genetic disorders
inheritance in man including phenotypes and linkage data
(OMIM)
BioProject Diverse Comprehensive collection of research studies
data including diverse data types
BioSample Diverse A resource of annotated biological samples from
data diverse studies
LitCovid Literature A COVID-19-specific curated literature database

collected from numerous species. On the other hand, specialized databases contain
data from one particular species, for example, WormBase contains data on nematode
worm. Similarly, biological databases are classified into two groups, namely primary
databases and secondary databases, based on the levels of data curation. Primary
databases such as GenBank are created from experimentally derived raw data
generated and submitted by experimental biologists. On the other hand, secondary
databases such as Ensembl (maintained at the EMBL-EBI, UK), UCSC Genome
Browser (maintained at the University of California, Santa Cruz, USA) and TIGR
(maintained at the Institute of Genomic Research, Maryland, USA) are highly
curated and are usually created from analysis of various sources of primary data.
Some databases such as UniProt have characteristics of both primary and secondary
databases. For example, Uniprot database stores peptide sequences generated from
sequencing experiments and computationally inferred from genomic data as well.
Since majority of biological databases do not have a complete information, an
integrated database retrieval system like Entrez maintained by NCBI provides
integrated access to 35 distinct databases. These databases may be grouped into
six areas: Literature, Genomes, Genes, Proteins, Health and Chemicals (Table 2.1).
It can be searched using Boolean phrases (AND, OR and NOT) and useful data is
downloaded in various formats. For example, most common Entrez databases
include PubMed, SRA, Nucleotide, Protein and Structure. Similarly, Sequence
Retrieval System (SRS) also provide integrated database retrieval system regarding
a search term. BioStudies (www.ebi.ac.uk/biostudies) is a recent public multimodal
database launched by EMBL-EBI in 2017 and it currently holds metadata
2.2 Types of Biological Databases 13

descriptions from various biological studies with a plan to archive all functional
genomics data in near future.

2.2 Types of Biological Databases

2.2.1 Sequence and Structure Databases

(a) Nucleic acid databases: All published DNA and RNA sequences are usually
deposited in three parallel public databases, namely GenBank (Fig. 2.1) (avail-
able at the National Centre for Biotechnology Information), EMBL (European
Molecular Biology Laboratory) and DDBJ (DNA Data Bank of Japan available
at the National Institute of Genetics), maintained in the USA, the UK and Japan,
respectively. These three public databases exchange their data under the Inter-
national Nucleotide Sequence Database Collaboration (INSDC) framework.
The important nucleotide databases are listed in Table 2.2. GenBank has
grown exponentially since its inception in 1982 and currently contains more
than 2.1 billion nucleotide sequences (Fig. 2.1). The current exponential
increase with a doubling time of 20.8 months is very close to the doubling
time of 18 months as proposed by Moore’s law. There was an exponential
increase in the number of taxa as well in the publically available databases
(Fig. 2.2). The European Nucleotide Archive (ENA) is maintained by EMBL
and provides open access to a wide range of nucleotide sequences from raw
reads to finished genome sequences. The DDBJ Sequence Read Archive (DRA),
maintained by the DDBJ, provides free access to raw read data and assembled
genomic data from next-generation sequencing platforms. The DNA and RNA
sequences are directly submitted by the researchers in these databases and are

13
Log 10 number of nucleotides

12

11

10

1990 2000 2010 2020

Year

Fig. 2.1 The growth of nucleotide sequences on a logarithmic scale in public databases (GenBank/
ENA/DRA) from June 1982 to March 2020. The central line is the best least square fit line
representing doubling time of 20.8 months, whereas external lines indicate a doubling time of
18 months
14 2 Biological Databases

Table 2.2 Nucleotide databases

Database Description URL
NCBI Genetic sequence database www.ncbi.nlm.nih.
GenBank gov/genbank
Ensembl A genome browser of vertebrates www.ensembl.org
NCBI A collection of non-redundant and well-annotated www.ncbi.nlm.nih.
Refseq genomic sequences, transcripts and proteins gov/refseq
UCSC Interactive genome visualization browser https://genome.ucsc.
genome edu/
browser
1000 A catalogue of human genomic variation www.
genomes internationalgenome.
org
GeneCards An integrative database of predicted and annotated www.genecards.org
human genes
lncRNAdb Database containing annotated long non-coding RNAs rnacentral.org/expert-
in eukaryotes database/lncrnadb
miRBase Database of published miRNA sequences along with www.mirbase.org
annotation
DIANA- A database of experimentally supported miRNA www.microrna.gr/
TarBase targets. tarbase

Fig. 2.2 The exponential growth in the number of taxa in the available public databases

cross-submitted in other two databases as well if submitted in one database.

Each sequence in the database has a unique accession number and a version
number which is common for all three databases. The GenBank (Fig. 2.3) is
searched online using Entrez, whereas both EMBL and DDBJ databases are
searched through SRS servers. Expressed Sequence Tags (ESTs), small
fragments of mRNA with high error rate are also available on nucleic acid
databases. Similarly, Genome Survey Sequences (GSS), a single-pass fragments
of genomic sequences with high error rate are also integral part of these
2.2 Types of Biological Databases 15

Fig. 2.3 Partial view of GenBank home page showing information on human oxytocin mRNA

databases. Moreover, whole-genome sequences of various species are also

deposited in the nucleic acid databases and are regularly updated with the
release of new genomic sequences. Nucleic acid databases are primarily
customized for human DNA and RNA sequences useful in biomedical research.
For example, a reference human genome in form of NCBI Refseq database is
created and human genetic variation is also profiled in a database known as
dbSNP. Ensembl is an integrated platform for genome annotation and distribu-
tion of genomic data with comprehensive annotation of genomic variants,
transcript structures and regulatory regions (Fig. 2.4). It provides a valuable
resource for evolutionary studies using large-scale comparative genomics data
of 227 vertebrate and model species. In addition, it also provides a genome
browser showing genomic variations emerging frequently in SARS-CoV2 virus
(https://covid-19.ensembl.org). The University of California Santa Cruz
(UCSC) Genome Browser currently provides a web-based view of 211 genome
assemblies of more than hundred species (Fig. 2.5). The SARS-CoV-2 Genome
Browser is recently added to UCSC Genome Browser including datasets from
major annotation databases. There are many databases dedicated for various
non-coding RNAs such as micro RNA (miRNA) and long non-coding RNA (lnc
RNA) and a popular database for unified access to non-coding RNA sequences
is RNAcentral. DIANA-TarBase is a reference database of experimentally tested
numerous miRNA targets providing cell-specific miRNA-gene interaction.
(b) Protein databases: The protein sequences available in the protein databases
(Table 2.3) are obtained from protein sequencing methods such as Edman
degradation and peptide mass spectrometry. In addition, they are also inferred
from three-dimensional structures obtained through X-ray crystallography and
16 2 Biological Databases

Fig. 2.4 Partial view of Ensembl genome database

Fig. 2.5 Partial view of UCSC Genome browser

NMR. Moreover, a significant amount of protein sequence data is also obtained

from translating DNA and RNA sequences. Universal protein resource
(UniProt) provides a comprehensive access to high-quality protein sequences
(Fig. 2.6). UniProt knowledge base (UniProtKB) is the primary source of
universal protein sequence information containing more than 189 million
sequences obtained from experimental sequencing as well as translated ORF
sequences from EMBL. This database is endowed with well-annotated protein
2.2 Types of Biological Databases 17

Table 2.3 Protein databases

Database Description URL
UniProt A public database of protein sequences with their www.uniprot.org
functional information
CATH-Gene3D Database of protein classification and prediction of www.cathdb.info
domain structure
GenPept Translated coding sequences from GenBank www.ncbi.nlm.
nih.gov/protein
DIP A database of experimentally determined protein– http://dip.doe-
protein interactions mbi.ucla.edu
HPRD Integrated platform depicting various information www.hprd.org
regarding human proteome
InterPro Integrated database providing functional analysis of www.ebi.ac.uk/
proteins interpro
ModBase Database containing theoretically calculated protein modbase.
structure models compbio.ucsf.edu
RCSB protein data Public database containing experimentally www.rcsb.org
Bank (PDB) determined protein and other macromolecule
structures
Pfam Database of large collection of protein families pfam.xfam.org
PROSITE Database of protein domains, families and functional prosite.expasy.
sites org
ProteomicsDB A multi-omics database including proteomics, www.
transcriptomics and cell line viability data ProteomicsDB.
org
CoV3D A database of coronavirus protein structures cov3d.ibbr.umd.
edu
STRING A database of protein–protein interactions string-db.org

Fig. 2.6 Partial view of UniProt home page

18 2 Biological Databases

sequences and a preliminary assignment of motifs present in the sequences. It is

also cross-referenced with other useful databases. Uniprot database consists of
two divisions: SwissProt and TrEMBL. TrEMBL is an automated database
requiring minimum human intervention, whereas Swissprot is a well-curated
database having manual entry of useful information from available literature.
Other important protein databases are CATH-Gene3D, GenPept, PRF and
DAD. The CATH-Gene3D database has the classification of about 151 million
protein domains into 54,881 superfamilies and prediction of structural domains
on the publicly available protein sequences. GeneBank Gene Products Databank
(GenPept) and DAD contain translated coding sequences from GenBank and
DDBJ, respectively. The Database of Interacting proteins (DIP) has a detailed
documentation of experimentally determined protein–protein interactions
involved in a biological process. The Human Protein Reference Database
(HPRD) is an integrated platform regarding the domain architecture and post-
translational modifications of proteins available in the human proteome along
with their interactions and association with diseases. Pfam is a protein database
dedicated to protein families and domains. The Protein Data Bank (PDB) is an
international consortium consisting of four partners, namely Research
Collaboratory for Structural Bioinformatics Protein Data Bank (RCSB-PDB),
Protein Data Bank in Europe (PDBe), Protein Data Bank Japan (PDBj) and
Biological Magnetic Resonance Bank (BMRB) (Fig. 2.7). It is a global archive
for protein structures and other macromolecules determined using X-ray crys-
tallography and NMR. This global database was launched in 1971 and contains
more than 150,000 macromolecular structures adding about 12,000 new
structures each year. ProteomicsDB is a protein-centric database having large
quantities of human proteomics data generated using mass spectrometry. It

Fig. 2.7 A screenshot of RCSB-PDB home page

2.2 Types of Biological Databases 19

provides a real-time exploration of protein abundance in different tissues, body

fluids and cell lines. In addition, a visual representation of diverse drug-target
interaction data is also allowed in this database. STRING database provides an
integrated access to all known and predicted protein–protein interactions, both
physical interactions and functional associations, in more than 14,000
organisms.

2.2.2 Expression Databases

There are some useful expression databases hosting gene expression data for func-
tional genomics studies (Table 2.4). High-throughput gene expression data derived
from both microarray and RNA-Seq platforms are usually archived in two public
databases: NCBI Gene Expression Omnibus (GEO) and EBI ArrayExpress (AE).
Gene expression Omnibus database was started in 2000 and is maintained by NCBI.
The overall structure of this database consists of platform record, sample record and
series record. Each GEO DataSet is an upper-level object having unique GEO
accession number starting with prefix GDS. DataSets are represented both in terms
of experiment-centric and gene-centric information. The gene-centric information
provides quantitative expression of a single gene across DataSets and is known as
GEO Profile. In addition to gene expression data, GEO contains other functional
genomic data such as copy number variations and transcriptional factor binding.
Another public database ArrayExpress is maintained by EBI and was launched in

Table 2.4 Expression databases

Database Description URL
NCBI GEO Gene expression data from microarray and sequencing. www.ncbi.nlm.
nih.gov/geo
EBI Gene expression data from microarray and sequencing. www.ebi.ac.uk/
ArrayExpress arrayexpress
Expression Gene and protein expression data from various species and www.ebi.ac.uk/
atlas conditions. gxa
Gene Gene expression data from microarray and sequencing. www.ddbj.nig.
expression ac.jp/gea
archive
Human A map of all human proteins in cells, tissues and organs www.
protein atlas proteinatlas.org
ONCOMINE Gene expression and sample data from different cancer www.oncomine.
types. org
TiGER Tissue-specific gene expression profile and gene bioinfo.wilmer.
regulation data jhu.edu/tiger
DualSeqDB A host-bacterial pathogen RNA-sequencing database www.
having combined gene expression data during infection tartaglialab.com/
process. dualseq
LncExpDB An expression database of human long non-coding RNAs https://bigd.big.
ac.cn/lncexpdb
20 2 Biological Databases

2002. It consists of three components: ArrayExpress Repository, ArrayExpress

Warehouse and ArrayExpress Atlas. The DDBJ recently started a similar database
known as Genomic Expression Archive (GEA) for hosting the functional genomics
data. All three of the public databases follow the guidelines of reporting standards
the Minimum Information About a Microarray Experiment (MIAME) for microarray
data and the Minimum Information about a high-throughput Nucleotide SEQuenc-
ing Experiment (MINSEQE) for sequencing data. Expression atlas of EBI provides
information regarding the gene and protein expression in different species under
various biological conditions. Moreover, Human Protein Atlas (HPA) has a detailed
map of all proteins expressed and located in various human cells, tissues and organs.
It consists of tissue atlas, subcellular atlas, pathology atlas, blood atlas, metabolic
atlas and brain atlas. Tissue-specific Gene Expression and Regulation (TiGER)
provides large-scale data sets not only for tissue-specific genes but also for tissue-
specific transcriptional regulatory elements. A user can retrieve useful information
regarding a gene, transcriptional factors or a specific tissue from TiGER using gene
view, TF view and tissue view, respectively. Oncomine is a database having
heterogeneous cancer profiles and provides access to gene expression signatures
and sample data of more than 500 cancer types and cancer cell lines. DualSeqDB is a
resource for differential gene expression in pathogenic bacteria and their natural
hosts upon infection at different time points. LncExpDB is an expression database
and a comprehensive resource for expression landscape of long non-coding RNAs
under various biological conditions.

2.2.3 Pathway Databases

Pathway databases provide a roadmap connecting various omics data such as

genomics, transcriptomics, proteomics, metagenomics and metabolic data
(Table 2.5). Metabolic pathways database hosts curated data regarding metabolic

Table 2.5 Metabolic pathway databases

Database Description URL
KEGG Database of biological systems including www.genome.jp/
drugs and diseases kegg
MetaCyc Reference database of metabolic pathways metacyc.org
PathBank Metabolic and signalling pathway database of www.pathbank.org
model organisms
Reactome Curated pathway database reactome.org
HMDB Human metabolome database hmdb.ca
BiGG models A database of genome-scale metabolic models bigg.ucsd.edu
knowledge base
Plant Reactome Comparative plant pathway database plantreactome.
gramene.org
Ingenuity pathway Commercial pathway database developed by digitalinsights.
analysis (IPA) Qiagen qiagen.com
2.2 Types of Biological Databases 21

pathways from all domains of life along with chemical compounds, reactions and
enzymes. They are useful in complex bioinformatics analysis such as flux balance
analysis and cellular modelling. KEGG and MetaCyc are the oldest and most
popular metabolic databases. MetaCyc is the largest reference database of experi-
mentally determined metabolic pathways and enzymes from all domains of life. It
contains 2749 pathways currently extracted from thousands of publications. Patho-
Logic, a component of Pathway Tools program, performs computational prediction
of metabolic pathways in an organism using this reference database. The major
limitations of this database are lack of any information regarding protein signalling,
disease and developmental process pathways. KEGG (Kyoto Encyclopaedia of
Genes and Genomes) is a comprehensive functional knowledge base of genes and
genomes at the molecular level and higher biological levels as well. KEGG
Orthology (KO) database hosts the molecular functions of various genes and
proteins. The higher-level biological functions are represented as KEGG pathway
maps describing the biological networks of molecular interactions and metabolic
reactions. In addition, the KEGG database also provides chemical information
regarding the metabolites and other small molecules in form of KEGG LIGAND
and health-related information of drugs and diseases in form of KEGG MEDICUS.
However, KEGG database does not provide information regarding lipid synthesis,
cellular location and metabolite signalling. PathBank is a comprehensive pathway
database of ten model organisms including human, rat, cow, yeast, etc. It provides a
detailed account of map of metabolic pathways and signalling, protein signalling,
disease, drug and physiological pathways in these model organisms. Reactome is a
relational database of signalling, transcriptional regulation, disease and metabolic
pathways. It provides a bioinformatics tool for analysis of pathways for systems
biology research. The human metabolome database (HMDB) is a comprehensive
electronic knowledge base of small-molecule metabolites present in human body
along with hyperlinks to other useful databases such as KEGG, MetaCyc and PDB.
BiGG models knowledge base is a high-quality genome-scale metabolic model
repository containing more than 100 BiGG models of different organisms including
a human-specific model known as Recon3D. In addition, there are 515 strain-
specific draft genome-scale models across three organisms. Plant Reactome is a
comparative and systems-level plant pathway database using rice as a reference
species and the pathway knowledge derived from rice is extended to other 82 plant
species using gene-orthology prediction. A total of 298 reference pathways are
hosted by this database related to metabolic pathways, transcriptional networks,
hormone signalling pathways and developmental processes. Ingenuity Pathway
Analysis (IPA) is a commercial database developed by QIAGEN for visualization
and analysis of omics data using a network or pathway.

2.2.4 Disease Databases

Disease database is one of the most important resource for biomedical research
(Table 2.6). The Online Mendelian Inheritance in Man (OMIM) of NCBI contains
22 2 Biological Databases

Table 2.6 Disease databases

Database Description URL
Online Mendelian Comprehensive compendium of human genes www.omim.org
inheritance in man and genetic phenotypes
(OMIM)
The Cancer genome A cancer database having genomic, www.cbioportal.org
atlas (TCGA) transcriptomic, proteomic and epigenomic
data
Human gene mutation Database of all published gene mutations www.hgmd.cf.ac.uk
database (HGMD) involved in human inherited diseases
GWAS central Comprehensive repository of genome-wide www.gwascentral.
association study data org
HbVar Database of human haemoglobin variants and globin.cse.psu.edu/
thalassemia mutations globin/hbvar/
MalaCards Integrated database of human diseases www.malacards.org
miR2Disease A comprehensive database on miRNA-disease www.mir2disease.
relationship org
DisGeNET Database of genes and variants associated with www.disgenet.org
human diseases
STAB A cell atlas providing cellular landscape of http://stab.comp-
human brain and neuropsychiatric diseases. sysbio.org
canSAR A knowledge base for cancer translational http://cansar.icr.
research and drug discovery ac.uk
CNCDatabase Cornell non-coding Cancer driver database is a https://cncdatabase.
manually curated database regarding med.cornell.edu/
non-coding cancer drivers
PAGER-CoV Pathways, annotated gene-lists and gene discovery.
signatures electronic repository for Corona informatics.uab.edu/
virus PAGER-CoV/

useful information regarding human genes and genetic diseases along with gene
sequences and associated phenotypes. It provides a complete description regarding a
disease gene, the phenotypes associated with the disease gene and other genes
associated with the disease. The Cancer Genome Atlas (TCGA) is joint programme
of National Cancer Institute and National Human Genome Research Institute
launched in 2006. It hosts more than 2.5 pentabytes of genomic, epigenomic,
transcriptomic and proteomic data of 33 cancer types. canSAR is a largest public
resource of cancer multidisciplinary data for integrative translational research and
drug discovery. Similarly, CNCDatabase provides a detailed information about
predicted non-coding cancer drivers located on promoters, untranslated regions,
enhancers and non-coding RNAs. The Human Gene Mutation Database (HGMD)
is a collection of all published gene mutations involved in human inherited diseases.
The GWAS Central database is most comprehensive repository of genome-wide
association study (GWAS) data related to more than 1400 phenotypes. It provides
integrative access and graphic visualization of GWAS data collected from more than
3800 studies. The HbVar is a database of human genomic sequence changes not only
2.2 Types of Biological Databases 23

involved in the generation of different variants of haemoglobin but also causing

different forms of haemoglobinopathies and thalassemia. A detailed information is
provided regarding genomic sequence alterations along with biochemical,
haematological and pathological changes. The MalaCards human disease database
is an integrated database of human diseases having a collection of more than 21,000
diseases. The miR2Disease is a manually curated database hosting detailed informa-
tion on a microRNA-disease relationship. The current version of the database
accounts for 3273 relationship between 349 miRNA and 163 diseases. The
DisGeNET is a knowledge platform providing data regarding disease-associated
genes and their variants collected from various sources including the scientific
publication. The current version of this comprehensive database hosts more than
17,000 genes and 117,000 genomic variants. STAB is a spatio-temporal cell atlas of
regional cellular heterogeneity of human brain providing a landscape of
transcriptome dynamics during neuropsychiatric disorders. PAGER-CoV is a
newly developed web-based disease database to interpret the functional genomics
studies of coronavirus disease such as inflammatory response and organ damage in
the host.

2.2.5 Organism-Specific and Virus Databases

There are specialized organism-specific databases available for some model animals,
economically important plants and animals (Table 2.7). In addition, specific genomic
databases of pathogenic viruses are also available. The Alliance of Genome
Resources Portal (www.alliancegenome.org) provides an integrated access to
genome of diverse model species used to study human biology. WormBase which
is one of the founding member of this alliance is a repository of detailed biological

Table 2.7 Organism-specific and Virus databases

Database Description URL
WormBase Database of experimental data of wormbase.org
nematode C. elegans
SilkDB Database of silkworm genome silkdb.bioinfotoolkits.
net
MBKbase-rice Integrated omics database of rice www.mbkbase.org/rice
Bovine genome Bovine genomics database bovinegenome.org
database (BGD)
Pig genome database Pig genomics database www.animalgenome.
(PGD) org/pig/genome/db
Zebrafish information Zebrafish genomics database zfin.org
network (ZFIN)
GISAID A global database of influenza viruses and www.gisaid.org
SARS-CoV-2 virus
GESS A database of global evaluation of SARS- wan-bioinfo.shinyapps.
CoV-2/hCoV-19 sequences io/GESS
24 2 Biological Databases

information from experimental observations from more than 1400 laboratories

across the world on multiple nematode species. Sericulture is an important compo-
nent of agriculture and silkworm is an insect species which laid the foundation of
sericulture. SilkDB is a comprehensive database of silkworm genome containing
high-quality chromosome level assembly along with various visualization tools.
MBKbase is an integrated molecular breeding knowledge base of important crop
genome sub-databases for rice, soybean, wheat and maize. The rice sub-database of
this knowledge base provides rice germplasm information including pan-genomic
view of multiple reference genomes. There are genomic databases for economically
important livestock species useful for research and molecular breeding of these
species. The Bovine Genome Database (BGD) provides genome browsing, genome
annotation and data mining of the bovine genome. In addition, this database also
offers bovine sequence searching using BLAST. The Pig Genome Database (PGD)
is a rich collection of quantitative trait loci (QTL), whole-genome association study
(WGAS), candidate genes and gene expression in the pig genome. Zebrafish is a
popular model organism for developmental biology and toxicological research. The
Zebrafish Information Network (ZFIN) has a wide variety of experimental data
related to the zebrafish research including genetic and genomic data. The genome
analysis is essential to understand the evolutionary history of pathogenic bacterial
strains and different variants of pathogenic viruses. The genomic database of these
pathogens is a useful resource for understanding the outbreak of any epidemic or
pandemic disease. For example, the recent pandemic caused by human CoV-19 can
be better understood and prevented using its genomic diversity across the world. In
this connection, the GISAID (Global Initiative on Sharing All Influenza data)
database provides open access to the genomic sequences of all influenza viruses
and human SARS-CoV-2. Moreover, Global Evaluation of SARS-CoV-2/hCoV-19
Sequences (GESS) is a web tool providing analysis of single-nucleotide variants
(SNV) collected from more than 300,000 high-quality and high-coverage human
CoV-19 viral genomes from GISAID database. It allows a user to search and
download SNVs at any single or multiple viral genomic position(s) from any
particular country or region and is a useful resource to monitor the migration and
evolution of CoV-19 virus. Moreover, CoV3D is a comprehensive resource for
coronavirus protein structures such as spike glycoprotein and their complexes with
antibodies, receptors and small molecules and provides useful information for
structure-based design of drugs and vaccines.

2.3 Database Searching

The Basic Local Alignment Search Tool (BLAST) is most widely used search tool
for sequence databases. It finds a region of local similarity (conserved sequence
pattern) between a query DNA or protein sequence against a target database. The
ultimate aim of a BLAST program is to infer an evolutionary and functional
relationship between two DNA or protein sequences. The original version of
BLAST was launched in 1990 followed by development of various variants of
2.4 Exercises 25

Table 2.8 Different variants of NCBI-BLAST program

BLAST
program Description
BLASTN A nucleotide sequence is searched against a nucleotide sequence database
BLASTP A protein sequence is searched against a protein sequence database
BLASTX All reading frames of a nucleotide sequence is against a protein sequence
database
TBLASTN A protein sequence is searched against all reading frames of a nucleotide
sequence database
TBLASTX Six-frame translations of nucleotide sequence is searched against six-frame
translations of a nucleotide sequence database
MEGABLAST Suitable for finding alignment between closely related sequences
PSI-BLAST Suitable for searching remote homologs or members of a protein family
PHI-BLAST Suitable for finding protein sequences with specific pattern in the database

BLAST specialized for different types of databases. Table 2.8 describes various
variants of BLAST along with their applications. There are three important aspects
of a search process: the input query sequence, target database and choosing a
customized BLAST program. For example, PSI-BLAST is a useful program to
identify remote homologs in different species. BLAST finds a local alignment in
each match in the database above a certain threshold S. These matches are known as
high scoring segment pairs (HSPs) along with their E-values. An E-value of 10 5 is
usually taken as cut-off during a BLAST search underlying the fact that the score or a
greater value than the score is likely to occur by chance in one out of 100,000
matches. The Position-Specific Iterated (PSI)-BLAST searches remote protein
homologs in a protein database in multiple iterative steps using position-specific
scoring matrix. FASTA is a similarity search program like BLAST and its input
sequence format is widely known as FASTA format. Blast-like alignment tool
(BLAT) is another efficient program to find sequences of greater similarity (more
than 95%) but has less sensitivity to divergent sequences. The BLAT algorithm
keeps the index of complete genome in the computer memory requiring 2GB RAM.
However, sensitivity of this program is significantly inferior to NCBI-BLAST.

2.4 Exercises

1. Aromatase, a member of Cytochrome P450 family of enzymes, performs the

conversion of androgens to oestrogens in brain, testes and ovary across different
species of vertebrates. Teleost fishes have two structurally and functionally
different forms of aromatase, aromatase A and aromatase B. Find the remote
homologs of human aromatase in different species of teleost fishes using an
appropriate variant of BLAST program.
26 2 Biological Databases

Solution
1. Retrieve the protein sequence of the human aromatase from GenBank or
Ensembl.
2. Use this protein sequence as a query sequence to perform PSI-BLAST (Figs. 2.8,
2.9, and 2.10).
2. The TP53 gene encodes for a tumour suppressor protein responding to various
kind of cellular stresses. It plays an important role in the prevention of cancer
and is also known to be mutated in majority of human cancer. Find the location
and number of alternative transcripts of TP53 gene in human and infer the

Fig. 2.8 PSI-BLAST submission page using default parameters

Fig. 2.9 Part of the Iteration 1 table for the human aromatase search showing closely related
sequences
2.4 Exercises 27

Fig. 2.10 New remote homologs are identified with each iteration of PSI-BLAST search. Part of
the interaction table shows newly identified fish homologous sequences (highlighted in yellow)

Fig. 2.11 Partial view of Ensembl genome browser showing details regarding human TP53 gene

evolutionary history of TP53 gene gain and loss in different genomes available in
the Ensembl database.

Solution
1. Search the term “TP53 gene” under human species in the Ensembl database.
2. It is located on the reverse strand of Chromosome 17: 7,661,779-7,687,538
(Fig. 2.11).
3. The TP53 gene consists of 27 alternative transcripts in human (Fig. 2.12).
28 2 Biological Databases

Fig. 2.12 The Ensembl genome browser showing 27 alternative transcripts of TP53 gene

4. View the gene gain/loss tree to see the evolutionary history of human TP53 gene
(Fig. 2.13). The species tree shows significant gene gain events or expansions
(indicated in red branch colour), gene loss events or contractions (indicated in
green branch colour) and no changes (indicated in blue branch colour). Each node
indicates the number of gene family present in the respective extant or extinct
species. In TP53 gene family tree, two species, namely Siamese fighting fish
(B. splendens) and African elephant (L. africana), have undergone remarkable
expansion in form of 27 members and 14 members, respectively. In addition,
16 species show significant expansions and three species show significant
contractions.
2.4 Exercises 29

Fig. 2.13 The gene gain/loss

tree of TP53 gene
30 2 Biological Databases

2.5 Multiple Choice Questions

1. Which of the following is a specialized database?

(a) GenBank
(b) EMBL
(c) WormBase
(d) DDBJ
2. Which of the following database is a secondary database?
(a) Ensembl
(b) GenBank
(c) DDBJ
(d) EMBL
3. The database retrieval system of NCBI is known as:
(a) Entrez
(b) SRS
(c) GenBank
(d) BLAST
4. A unified access to non-coding RNA data is provided by:
(a) RNAcentral
(b) Genome
(c) Gene
(d) GEO profiles
5. The primary source of universal protein sequences is:
(a) UniProtKB
(b) ProteomicsDB
(c) HPRD
(d) PROSITE
6. Which of the following is NOT a partner in the Protein Data Bank consortium?
(a) RCSB
(b) PDBe
(c) BMRM
(d) HPRD
7. A repository for high-quality genome-scale metabolic model is known as:
(a) BiGG models knowledge base
(b) GSMM models knowledge base
(c) RGMM models knowledge base
(d) HGSM models knowledge base
8. Which of the following database is a rich resource for cancer data?
(a) TCGA
(b) ATGA
(c) TGAA
(d) ATGC
Suggested Reading 31

9. Which of the following database is a useful resource for human SARS-CoV-

2 viruses?
(a) GISAID
(b) SARSV
(c) HSCOV
(d) DSCOV
10. Name the variant of BLAST useful for searching remote homologs of a protein:
(a) PHI-BLAST
(b) PSI-BLAST
(c) MEGABLAST
(d) TBLASTX

Answer: 1. c 2. a 3. a 4. a 5. a 6. d 7. a 8. a 9. a 10. b

Summary
• Biological database provides a structured platform for retrieval of useful data.
• Entrez is an integrated database retrieval system for 35 diverse databases.
• GenBank, ENA, DDBJ and Ensembl are major nucleotide databases.
• UniProt Knowledgebase (UniProtKB) provides a comprehensive resource for
protein sequences.
• Protein Data Bank (PDB) is a global archive for protein structures and other
macromolecules.
• GISAID database is a comprehensive genomic resource for different variants of
human SARS-CoV-2 virus.

Suggested Reading
Letovsky SI (1999) Bioinformatics: databases and systems. Springer, Boston
Lesk AM (2004) Database annotation in molecular biology: principles and practice. Wiley,
Chichester
Revesz P (2010) Introduction to databases: from biological to Spatio-temporal. Springer, London
Annual nucleic acid research database issues, Oxford University Press: Oxford, 2015-2021
Statistical Computing Using R
3

Learning Objectives
You will be able to understand the following after reading this chapter:

• R environment, data structures and functions.

• Data input and output in the R environment.
• Statistical analysis of biological data.
• Common statistical tests using R.
• Creating graphs using R.

3.1 Introduction to R Language

Statistical computing is an interface of computer science and statistics for analysis of

large data sets. R language is an object-oriented functional programming language
used for statistical computing. Robert Gentleman and Ross Ihaka developed this
program in early 1990s from the S language. It is more popular than any other
commercial statistical software in biological research because it is not only open
source but continuously being enriched and updated by the contributions from the
experts across the globe. R can be operated in both interactive and batch modes and it
stores its workspace in RAM. In interactive modes, commands are typed in succes-
sion one after other and consequently the result is displayed. However, batch modes
do not require regular interaction with the user and usually run in an automated
fashion for a longer duration. All R commands are executed in a window called the R
console (Fig. 3.1). Even, the R codes are stored in a file with a suffix .R or .r and later
executed with source command. The usual assignment operator in R is <- or ->,
although ¼ can also be used. The R programming deals with writing functions. A
function consists of a number of instructions for computing a result from inputs.

# Springer Nature Singapore Pte Ltd. 2022 33

B. K. Tiwary, Bioinformatics and Computational Biology,
https://doi.org/10.1007/978-981-16-4241-8_3
34 3 Statistical Computing Using R

Fig. 3.1 The R console window showing implementation of some codes

3.2 Data Structures in R

The R language has a variety of data structures such as vectors, character strings,
matrices, lists, data frames and classes. Vector is the R workhorse and consists of
elements of similar data type such as integer or character. However, scalars or
individual numbers in the R are essentially one-element vector. On the other hand,
character strings are truly single-element vectors having the character mode. An R
matrix consists of rectangular array of numbers. This mathematical structure has two
attributes: rows and columns. An R list consists of multiple values of different data
types. So, different components are packaged a single list that can be returned by an
R function. The internal structure of a list can be printed by command str(), where str
stands for structure. Data frame is a two-dimensional structure with rows and
columns akin to a matrix. However, each column has a different mode unlike matrix.
For example, one column may be numeric, whereas another column may consist of
character strings in a data frame. A class defines the type of an object in connection
with object-oriented programming in R. It is usually defined by the user during R
programming. Generic functions operate on an object based on its class.

3.3 R Packages

R packages are organized libraries developed for different functionalities. These

packages are usually downloaded from two specific repositories known as Compre-
hensive R Archive Network (CRAN) and Bioconductor. The standard installation
3.4 Data Input and Output in R 35

Fig. 3.2 CRAN, a network of ftp and webservers providing access to R codes, documentation and
packages

process installs two basic packages, namely base and stats by default. CRAN
provides thousands of packages for a wide range of statistical analysis and graphics
(Fig. 3.2).

3.4 Data Input and Output in R

Any session in R starts with choosing its working directory. R imports a data from a
datafile and builds an R object in the workspace. Here, the workspace is an abstract
concept that indicates the storage of data in RAM during a specific R session. The
workspace can be permanently stored in a binary file known as .RData file. In
addition to .RData file, all the commands executed in a particular R session can be
saved in a simple text file known as .RHistory file and reloaded again. First, the
current working directory needs to be checked with the function getwd(). The current
working directory may be changed through another function setwd(). R always
searches any file in the working directory and writes a file in the same directory.
The command line starts with a character > and continuation of a line is indicated by
+. The generic function read.table reads a data from a file and writes the insides into
an R object called data frame. We may add the separator in terms of tabulation
characters (sep ¼ “\t”), white spaces (sep ¼ “ ”) or commas (sep ¼ “,”).
36 3 Statistical Computing Using R

3.5 Biological Data Analysis in R

The measurement data obtained through biological experiments can be stored in

form of a data vector. For example, gene expression values in five samples of a gene
are stored using concatenate command c() which is as follows:

>gene <- c(1.25, 1.35, 1.00, 1.5, 1.75)

Here, an object gene is created including five gene expression values in different
samples. The simple computations such as sum, mean, median, standard deviation
and variance of these values can be computed using respective built-in functions.

> mean (gene)

[1] 1.37
> median(gene)
[1] 1.35
> sd(gene)
[1] 0.279732
> var(gene)
[1] 0.07825

However, data values are often stored in form of matrix consisting of rows and
columns.

3.5.1 Statistical Distributions

In order to test a biological hypothesis, the nature of statistical distribution should be

known for proper application. The important statistical distributions are binomial
distribution, normal distribution, T-distribution, F-distribution and chi-squared dis-
tribution. Only the binomial distribution is a discrete distribution, whereas rest of the
distributions are continuous. The binomial distribution is common in biological
experiments with repeated trials with two outcomes like success/failure, healthy/
disease, etc. (Figs. 3.3 and 3.4). The number of success is denoted by k and number
of trials is denoted by n. The mean is np, the variance is np(1-p) and the standard
deviation is square root of np(1-p) in case of a binomially distributed variable. The
density function of a binomial distribution in R can be computed by

>dbinom(k,n,p)

The normal distribution of data is assumed for many statistical analyses such as
gene expression values obtained using microarray data. The data values in normal
distribution are represented by a bell-shaped curve or normal density with mean mu
and variance sigma squared. This curve is symmetric around the mean value. We can
3.5 Biological Data Analysis in R 37

Fig. 3.3 Binomial

probabilities with n ¼ 20 and
p ¼ 0.6

Fig. 3.4 Binomial

cumulative probabilities with
n ¼ 20 and p ¼ 0.6

draw a random sample of size 100 with the population mean of 1.5 and the
population standard deviation of 0.6 using the R command

>rnorm(100,1.5,0.6)

A variable with normal distribution can be standardized into a standard normally

distributed Z by subtracting mu and dividing the result with sigma. The standard
normally distributed Z has a fixed mean zero and standard deviation one (Figs. 3.5
and 3.6). The chi-squared distribution Figs. 3.7 and 3.8), T-distribution (Figs. 3.9
and 3.10) and F-distribution (Figs. 3.11 and 3.12) are modifications of normal
distribution and are useful in testing hypothesis.
38 3 Statistical Computing Using R

Fig. 3.5 Normal probability

density function with mean
zero and standard
deviation one

Fig. 3.6 Cumulative normal

distribution function with
mean zero and standard
deviation one

3.5.2 Testing of Hypothesis

In a biological research, we build a theory and would like to support our hypothesis
with experimental data. Thus, the testing procedure consists of proposing a hypoth-
esis, assuming the statistical distribution and inferring the conclusion. First, we have
an idea that a certain drug is effective against a disease. This is also known as
alternative hypothesis. Secondly, we propose a null hypothesis that the drug is not
effective against a disease in order to remove the bias. A suitable statistic such as a
t-value is computed which is not only a function of random variables but a function
3.5 Biological Data Analysis in R 39

Fig. 3.7 Chi-squared density

function with degrees of
freedom ¼ 10

Fig. 3.8 Chi-squared

distribution with degrees of
freedom ¼ 10

of data values as well. A p-value is computed by comparing the distribution of

statistic with the value of statistic. A large p-value indicates that the model is fitting
well with the data and the null hypothesis is true with high probability. On the other
hand, a low p-value indicates that the experimental outcome is very unlikely under
the given assumption of distribution and the null hypothesis is rejected. If the data
are found reasonable given the null hypothesis, we accept the null hypothesis and
reject our idea. On the other hand, if the data are unreasonable given null hypothesis,
the null hypothesis is rejected and idea is provisionally accepted. An acceptance and
rejection range of the data are also defined for acceptance and rejection of our idea,
40 3 Statistical Computing Using R

Fig. 3.9 Probability density

function of t-distribution with
degrees of freedom ¼ 20

Fig. 3.10 t-distribution with

degrees of freedom ¼ 20

respectively. When a null hypothesis is rejected even if it is true, we call it a Type I

error. In contrast, a false null hypothesis is accepted in case of Type II error. It is
expected that rejection range should be small to minimize Type I errors and
acceptance range should also be small to minimize Type II errors. The probability
of making a Type I error is known as the significance level of the test. It is generally
kept at the 5% level indicating that we are creating Type I error five times in
100 experiments. The power of the test is measured by one minus probability of
making Type II error. We always want power of the test to be exceptionally high in
case probability of making Type II error is small. If we reduce the significance level
3.5 Biological Data Analysis in R 41

Fig. 3.11 Probability density

function of F-distribution with
numerator degrees of
freedom ¼ 20 and
denominator degrees of
freedom ¼ 10

Fig. 3.12 F-distribution with

numerator degrees of
freedom ¼ 20 and
denominator degrees of
freedom ¼ 10

by broadening the acceptance range, the power of the test is also reduced making us
prone to Type II errors. The power of the test or the probability of drawing correct
conclusion can be improved by increasing the sample size prior to conducting
experiments.
42 3 Statistical Computing Using R

3.5.3 Parametric Tests Vs. Non-parametric Tests

Parametric tests have some assumptions about the underlying distribution of data.
For example, the data collected from a population with normal distribution. The
normality of the data is assessed by Shapiro–Wilk test which is based on a Q-Q plot.
The Shapiro–Wilk test can be computed in R using built-in function as follows.

>shapiro.test (data)

Sometimes, parametric tests also assume homogeneity of variance between

groups. Most common examples of parametric tests are linear regression, t-tests
and analysis of variance. These tests are more powerful than the non-parametric
tests.
In case of non-normal distribution due to skewness or heavy tails in the data,
rank-based tests are preferred over parametric tests because they do not assume any
specific distribution. These tests are also applicable for small sample size.

3.5.4 Common Statistical Tests

3.5.4.1 The Z-Test

This test is applicable in cases where we would like to test null hypothesis against
alternative hypothesis and the standard deviation is known. First, the standardized
value z is computed from a variable such as gene expression values following a
normal distribution. Next, the p-value is defined as the standard normal probability
of Z attaining values being more extreme than |z| either in left or right direction. If the
p-value is larger than the significance level alpha, then the null hypothesis is not
rejected. On the other hand, when the p-value is lesser than the significance level,
null hypothesis is eventually rejected.
The R script for computing z-value can be computed as follows.

z.value<-sqrt(n)*(mean(x)-mu0)/sigma
p.value<-2*pnorm(-abs(z.value),0,1)

Thus, the null hypothesis is rejected when z-value is either highly positive or
highly negative. When z-value falls between 95% confidence interval, then null
hypothesis is not rejected and consequently this region is known as the acceptance
region.

3.5.4.2 One-Sample T-Test

The population standard deviation is unknown in most of the biological research.
Therefore, the Z-test cannot be applied under this scenario. The one-sample t-test is
an appropriate method under this condition. If the p-value is larger than the signifi-
cance level, the null hypothesis is not rejected. On the other hand, the null hypothesis
is rejected in case p-value is smaller than the significance level. The t-value is
computed in R using the following script.
3.5 Biological Data Analysis in R 43

>t.value<-sqrt(n)*(mean(x)-mu0)/sd(x)

However, there is a built-in function in R which can be easily applied as follows.

>t.test(x,mu=0)

This one command will generate t-value, p-value and 95% confidence interval for
population mean at value 0. The large value of t indicates that the mean is different
from 0, whereas very small p-value tending towards zero supports the rejection of
null hypothesis.

3.5.4.3 Two-Sample T-Test

One of the most common tests in experimental biology is two-sample t-test. In case,
we want to compare the population means of two groups such as healthy individuals
and patients. When the variance is unequal between two groups, the test is known as
the Welch two-sample t-test. The equal variances between two groups are tested
using the built-in R function var.test. A t-statistic is computed based on the mean and
the variance of two groups. The computed t-value is large when the sample sizes are
large and the standard deviations are small. The t-value is computed in R using the
command which is as follows.

>t.test (x, y, var.equal= FALSE)

If the variances for two groups are equal, the t-value is computed in R using the
following script.

>t.test(x, y, var.equal = TRUE)

Thus, the null hypothesis of equal population means between two groups is
rejected if the p-value is less than 0.05.

3.5.4.4 Wilcoxon Rank Test

This two-sample test is widely used in bioinformatics as an alternative to the t-test.
We compare the overall distribution of two groups rather than focussing on means.
The null hypothesis assumes that both distributions are equal, whereas alternative
hypothesis does not support an equal distribution, for example, one distribution is
larger or smaller than another one. Here, the data in both groups are ranked and then
summed over to compute a statistics W following a correction. The distribution of
sum of ranks provides an estimate of p-value which can be used further to reject the
null hypothesis in case smaller than the significance level. This test can be applied to
a dataset, for example gene expression values between two groups, using the built-in
function as follows.

>wilcox.test (x,y)
44 3 Statistical Computing Using R

3.5.4.5 Chi-Squared Test

This distribution free test is suitable for testing relationship between categorical
variables such as gender. The chi-square statistic measures the independence
between two variables. The association between two variables is assessed by com-
paring the observed frequency with the expected frequency in case both variables are
independents. The chi-square statistic computed is compared against a critical value
under a chi-square distribution. It is also used as a goodness of fit test to compare a
sample data with population. A small test statistic indicates that observed value fits
well with the expected value and, thus, there is a relationship between two values.
On the other hand, a large test statistic indicates that both observed and expected data
do not fit well and, therefore, there is no relationship between two types of data. This
test can be performed only on numbers not with proportions or percentages, and the
number should be large usually more than five. This test can be performed in R using
the following script.

>chisq.test (data)

3.5.4.6 Analysis of Variance (ANOVA)

This test is an alternative to t-test when there are three or more groups under testing. It
is common test in bioinformatics and is based on a linear model. We assume that the
residuals or errors are normally distributed and have equal variance (homogeneity)
across three or more groups. Here, we test a hypothesis that three or more population
means are equal. Two sums of squares are important in computation during analysis of
variance. Firstly, sum of squares within (SSW) is the sum of the squared deviation of
the individual measurements to their group means. Secondly, the sum of squares
between (SSB) is the sum of squares of the deviances of the group means in reference
to the total mean. Finally, the F-value is computed based on the definition as follows.

SSB ðG  1Þ
F¼
SSW ðN  GÞ

where N is the total number of measurements and G is the number of groups

under comparison. G1 and NG are the degrees of freedom. If the data are
normally distributed, then the F-value follows F-distribution and the p-value is
computed based on this distribution. The value of SSB and F tends to be small
under the null hypothesis of equality of group means. If the p-value is equal to or
higher than the significance level, then the null hypothesis of equal group means is
accepted. In case the p-value is smaller than the significance level, the null hypothe-
sis is rejected. The R function aov() is used to compute one-way ANOVA test and
function summary.aov() is used to summarize the ANOVA model. The one-way
ANOVA test indicates that some of the group means are different. In order to
identify the pairs of groups showing significant difference, Tukey multiple pairwise
comparison can be performed using the R function TukeyHSD() where the fitted
ANOVA is included as an argument.
3.5 Biological Data Analysis in R 45

3.5.4.7 Kruskal–Wallis Rank Sum Test

This test is a non-parametric alternative to the ANOVA when normality is violated.
It is based on ranking of the data and is highly robust against the violation of
normality assumption. However, the size of the experimental effects is not estimated
in this test. The R function kruskal.test () is used to perform this non-parametric test.

3.5.4.8 Correlation Coefficient

The correlation coefficient is a measure to find linear relationship between two
variables. The value of this measure is always between +1 and -1. If the value of
the coefficient lies close to either of these two extreme values, then two variables are
linearly related. The correlation coefficient in R is computed by the function cor (x,
y). The null hypothesis that the correlation coefficient is zero is tested against the
alternative hypothesis that correlation coefficient is not zero. Since this test assumes
the normal distribution of data, and therefore it provides t-value and corresponding
p-value. The correlation between two variables is established by rejecting the null
hypothesis of zero correlation if the p-value is very small.

3.5.4.9 Principal Components Analysis

Principal component analysis is a descriptive method to find dependencies between
variables using correlations. It finds new directions in the dataset based on the
maximal variation. The amount of variance for a particular component is measured
by the eigenvalue. Each eigenvalue represents more variance than any other
observed variable. The direction of maximal variation is defined as the linear
combination with maximal variance. First few variables are important because
they have large eigenvalues in comparison to other variables. This property is
known as elbow rule. The first principal component is defined as the linear combi-
nation of data with first eigenvector. In R, correlations matrix is computed using the
built-in function cor followed by computation of eigenvalues and eigenvectors using
the built-in function eigen. Alternatively, another built-in function princomp is used
to obtain the scores (component scores) and the loadings (eigenvectors).

3.5.5 Graphics in R

R language is useful in creating customized publication quality figures with few

commands only. We can draw common plots such as scatter plot, box plot and
histogram to compare the distribution of data under different conditions. The R
graphics consists of three types of plotting functions: high level, low level and the
layout. The high-level plotting involves creation of a complete plot, whereas
low-level plotting augments to an existing plot created by a high-level function.
Some optional arguments can be included in a high-level function to enhance the
quality of the plot. The common arguments are pch for point characters, lty for line
type, lwd for line thickness and col for colour. The complete list of colours available
in the R is obtained by the command colors (). The x-axis limit and y-axis limit can
be fixed using the arguments xlim and ylim, respectively. The arguments xlab and
46 3 Statistical Computing Using R

32

30
30

25
28

20
data
26

15
24

10
22

5
20

0
2 4 6 8 10 12 14
Index
Scatterplot Bar plot

32
3.0

30
1.0 1.5 2.0 2.5

Sample Quantiles
28
Frequency

26
24
22
0.5

20
0.0

20 22 24 26 28 30 32 −1 0 1
Theoretical Quantiles

Histogram Normal Q-Q plot

32

A
30

C
28
26
24

T
22

G
20

Box plot Pie chart

Fig. 3.13 Different plots created using R built-in functions

ylab are included to label the x-axis and y-axis, respectively. Moreover, a command
par is used to change the default layout of the graphics device. Finally, the plot is
exported by saving in common file formats like jpg, postscript, tiff, png and pdf.
Some of the important plots (Fig. 3.13) that can be created using high-level
functions are described as follows.

Scatterplot This plot can be used to draw any number of points in a figure window
using the command plot. It displays number of dots to show the relationship between
3.5 Biological Data Analysis in R 47

two variables representing x-axis and y-axis. For example, the expression of genes in
disease and healthy conditions can be visualized using scatter plot.

Histogram It is a graphical display of distribution of numerical or categorical data

using the command hist. It shows the distribution of a dataset into ranges and the
height of the bar indicates the number of data included in a particular range. It is
usually applied to observe the underlying distribution of data. For example, we can
create a histogram to see whether the dataset follows a normal distribution prior to
performing a parametric test.

Box Plot A box plot displays the distribution of numerical data based on different
levels of quartiles using the command boxplot. These levels are minimum, first
quartiles (Q1), median, third quartiles (Q3) and maximum. It also provides graphical
information not only regarding symmetry and skewness of distribution but indicates
the presence of an outlier.

Bar Plot Bar plots are the graphical display of categorical variables in form of
rectangles created in R using the barplot() function. The height or the length of the
bar is proportional to the numerical values.

Normal quantile-quantile (Q-Q) plot The normal quantile plots are the graphical
display of the theoretical normal distribution on the x-axis against the dataset on the
y-axis to assess whether your dataset deviates from the normal distribution. If the
dataset follows normal distribution, we can proceed with the standard statistical
analysis using t-test or ANOVA.

Pie Chart The pie chart describes a data graphically in a circular graph. Each slice
of a pie represents relative proportion of data. Pie chart is created in R using pie()
function.

The low-level functions are only executed on a plot created by high-level

functions. For example, abline is a low-level function to add a straight line on a
scatter plot. Moreover, two functions lines and points are implemented for adding a
line plot and a point to an existing plot, respectively. Similarly, a text on the plot is
added by function text and on the margin of the plot using a function mtext. Finally, a
title can be added to an existing plot using the function title.

3.5.6 R Packages for Graphics

The base graphics of R has a facility to explore and plot various types of data. Most
common package to perform general plotting functions is known as lattice devel-
oped by Deepayan Sarkar and it is an integral part of the R base installation. It
provides a high-level data visualization of multivariate data. The graphical
parameters can be customized in lattice for a publication quality display. Moreover,
48 3 Statistical Computing Using R

Table 3.1 Commonly used R built-in functions

Category Function Description
Input/output functions getwd() Find the working directory
setwd(x) Set a new working directory
library (x) Load a package
load (x) Load a saved file
read.table Read a space delimited file
read.csv Read comma delimited (csv) files
write.table Write a tabular data in a file
x<-c(1,2,3. . .) Create a data vector x with specified elements
Mathematical sum(x) Sum of the elements of x
functions mean (x) Mean of the elements of x
var(x) Variance of the elements of x
sd(x) Standard deviation of elements of x
cor (x,y) Linear correlation between x and y
t(x) Transpose of x
%*% Matrix multiplication
eigen(x) Computes eigenvalues and eigenvectors of a
matrix
Statistical functions aov (formula) Analysis of variance model
t.test(x,y) t-test to find significant difference between x and y
chisq.test Chi-square test
var.test F-test for equal variance
shapiro.test Shapiro-Wilk test for normality
rbinom (n,size,p) Generate random samples from binomial
distribution
rnorm (n,mean, Generate random samples from normal
sd) distribution
Graphical functions qqplot Q-Q plot to check normality of data
plot (x) Scatterplot of x
hist(x) Histogram of frequencies of x
barplot(x) Histogram of values of x
boxplot(x) Box and Whiskers plot of x
pie(x) Pie chart of x
abline(a,b) A straight line of slope b and intercept a

ggplot2 is another powerful package developed by Hadely Wickham to generate

publication ready plots. The learning of this package is little challenging but we can
expect best return considering the time invested in learning this package. The qplot()
is a general function in this package that can be implemented to create a wide range
of graphs. The options in the graph can be further modified using the theme()
function. The R built-in functions commonly used are described in Table 3.1.
3.6 Exercises 49

3.6 Exercises

1. SRGAP2 and FOXP2 are the key genes involved in extraordinary cognitive
development and speech in human lineage. The expression levels of both genes
in the prefrontal cortex of human brain are given below.

SRGAP2: 5.54341, 5.38406, 5.68797, 5.48020, 5.71367, 5.16755, 5.54771,

5.50315, 5.29877, 5.26532, 5.58837, 5.93008, 5.85553, 5.55625.
FOXP2: 2.53967, 3.22162, 2.35884, 2.04029, 2.67516, 2.27110, 1.62437, 2.15451,
1.99223, 2.74845, 2.78851, 2.98586, 2.44189, 2.55143.

1. Compute the mean and standard deviation of gene expression levels of SRGAP2
and FOXP2 genes.
2. Perform appropriate tests to check the normality and equality of variances in
expression levels of both genes.
3. Perform two-sample t-test to find significant differences in the gene expression
levels of both genes.

Solution:
>SRGAP2<-c
(5.54341,5.38406,5.68797,5.48020,5.71367,5.16755,5.54771,
5.50315,5.29877,5.26532,5.58837,5.93008,5.85553,5.55625)
>FOXP2<-c
(2.53967,3.22162,2.35884,2.04029,2.67516,2.27110,1.62437,
2.15451,1.99223, 2.74845,2.78851,2.98586,2.44189,2.55143)
> mean(SRGAP2)
[1] 5.537289
> mean(FOXP2)
[1] 2.456709
> sd(SRGAP2)
[1] 0.216255
> sd(FOXP2)
[1] 0.4243888
> shapiro.test(SRGAP2)
Shapiro-Wilk normality test
data: SRGAP2
W = 0.97621, p-value = 0.9469

> shapiro.test(FOXP2)
Shapiro-Wilk normality test
data: FOXP2
W = 0.99442, p-value = 1
> var.test(SRGAP2,FOXP2)
F test to compare two variances
data: SRGAP2 and FOXP2
50 3 Statistical Computing Using R

F = 0.25966, num df = 13, denom df = 13, p-value = 0.02122

alternative hypothesis: true ratio of variances is not equal to 1
95 percent confidence interval:
0.0833569 0.8088493
sample estimates:
ratio of variances
0.2596597

The p-value of F-test is 0.02122 which is lower than the significance level 0.05.
Therefore, there is a significant difference in between two variances.

> t.test(SRGAP2, FOXP2,var.equal=F)

Welch Two Sample t-test
data: SRGAP2 and FOXP2
t = 24.199, df = 19.325, p-value = 6.436e-16
alternative hypothesis: true difference in means is not equal to 0
95 percent confidence interval:
2.814441 3.346717
sample estimates:
mean of x mean of y
5.537289 2.456709

2. We want to test a hypothesis that the nucleotides of a gene have equal probability.
The null hypothesis to be tested is that the frequency of each nucleotide is 0.25.
Compute the total number of nucleotide bases and frequency of each nucleotide
in the gene. Perform chi-square test whether nucleotides in the gene occur with
equal probability. The coding sequence of the gene is as follows.

ATGAACAGCCAGGGGTCAGCCCAGAAAGCAGGGACACTCCTCCTGT-
TGCTGATATCAAAC

Solution
> library(Biostrings)
>gene<-DNAString("ATGAACAGCCAGGGGTCAGCCCAGAAAGCAGGGACACT
CCTCCTGTTGCTGATATCAAAC")
> length(gene)
[1] 60
> alphabetFrequency(gene, baseOnly=TRUE, as.prob=TRUE)
A C G T other
0.3000000 0.2833333 0.2500000 0.1666667 0.0000000
> gene.freq<-alphabetFrequency(gene, baseOnly=TRUE, as.prob=TRUE)
> chisq.test(gene.freq)
Chi-squared test for given probabilities
3.7 Multiple Choice Questions 51

data: gene.freq
X-squared = 0.30278, df = 4, p-value = 0.9896

The number of nucleotide bases in the gene is 60 and frequency of A, C, G, and T

is 0.3, 0.28, 0.25, and 0.17, respectively. Since, p-value of chi-square test is close to
1, the null hypothesis is not rejected. Thus, the nucleotides of given gene have equal
probability.

3.7 Multiple Choice Questions

1. A stored R file can be executed using the command:

(a) execute
(b) source
(c) open
(d) None of the above
2. The assignment operator in R code is written as:
(a) <-
(b) ->
(c) ¼
(d) All the above
3. Which of the following statistical distribution is a discrete distribution:
(a) Binomial distribution
(b) Normal distribution
(c) t-distribution
(d) F-distribution
4. The command to perform two-sample t-test with unequal variance in R is:
(a) t.test (x, y, var.equal ¼ TRUE)
(b) t.test (x, y, var.equal ¼ FALSE)
(c) t.test(x, y, var.equal ¼ EQUAL)
(d) t.test (x, y, var.equal ¼ UNEQUAL)
5. Tukey multiple pairwise comparison test is performed in R using command:
(a) TukeyHSD()
(b) Tukey.test
(c) T.test
(d) TuckeyMPC()
6. The statistical test which is non-parametric alternative to ANOVA test is
known as:
(a) Wilcoxon rank test
(b) Kruskal–Wallis rank sum test
(c) Kolmogorov–Smirnov test
(d) Shapiro–Wilk test
52 3 Statistical Computing Using R

7. High-level plotting in R graphics refers to:

(a) Creation of a complete plot
(b) Addition to an existing plot
(c) Transforming an existing plot
(d) Modification of an existing plot
8. Which of the following is/are a low-level function in R:
(a) point
(b) abline
(c) text
(d) All the above
9. The popular R package(s) for graphics is/are:
(a) lattice
(b) ggplot2
(c) lattice and ggplot2
(d) None of the above
10. Which of the following is a characteristic feature of a data frame?
(a) Data is numeric, factor or character
(b) Two-dimensional array-like structure
(c) Row name is unique and same number of data in each column
(d) All the above

Answer: 1. b 2. d 3. a 4. b 5. a 6. b 7. a 8. d 9. c 10. d

Summary
• R language is an object-oriented functional programming language used for
statistical computing.
• Common data structures in the R environment are vectors, character strings,
matrices, lists, data frames and classes.
• R packages are organized libraries for different functionalities available at CRAN
and bioconductor.
• Useful statistical parametric and non-parametric tests can be performed in the R
environment.
• Both high-level and low-level graphics functions are available in the R
environment.

Suggested Reading
Gerrard P, Johnson RM (2015) Mastering scientific computing with R. Packt Publishing,
Birmingham
Sinha PP (2014) Bioinformatics with R cookbook. Packt Publishing, Birmingham
Hahne F, Huber W, Gentleman R, Falcon S (2008) Bioconductor case studies. Springer, New York
Lewis PD (2010) R for medicine and biology. Jones and Bartlett Series, Burlington
Wickham H (2016) ggplot2: elegant graphics for data analysis. Springer, Cham
Sequence Alignment
4

Learning Objectives
You will be able to understand the following after reading this chapter:

• Pair-wise and multiple sequence alignment of nucleotide and protein

sequences.
• Various scoring matrices for optimal alignment.
• Global and local sequence alignments.

4.1 Introduction

The homologous relationship between different residues in the nucleotide and

protein sequences diverged from a common ancestor is inferred using optimal
sequence alignment. This idea is based on that fact that highly similar sequences
are likely to have same biological functions. During optimal alignment, the homolo-
gous residues are arranged in columns either manually or using automated programs
in a best possible way even allowing some gaps between the residues. The presence
of gaps in the alignment indicates evolutionary insertions or deletions (collectively
known as indels) events in the homologous sequences. The sequence alignment is an
essential prerequisite computational step before conducting phylogenetic analysis
and prediction of protein structure. It is comparatively easy to align closely related
sequences and, in turn, the optimal alignment of sequences is an indicator of degree
of closeness. However, it is very difficult to perform sequence alignment in case the
sequence identity (i.e. number of identical residues/total number of aligned residues)
is below 25%. Amino acid alignment is preferred over nucleotide alignment for
protein coding sequences because it is easy to perform and less ambiguous. More-
over, the reading frame of the coding nucleotides are prone to break down during

# Springer Nature Singapore Pte Ltd. 2022 53

B. K. Tiwary, Bioinformatics and Computational Biology,
https://doi.org/10.1007/978-981-16-4241-8_4
54 4 Sequence Alignment

nucleotide alignment. The best approach is to first align the sequences at the amino
acid level and then reverse translate the residues to corresponding nucleotide
sequence alignment. Some programs like DAMBE and TRANSALIGN can perform
this kind of analysis. However, it is still very difficult to find optimal alignment of
non-coding nucleotide sequences such as introns and repeats.
During alignment, the alignment score for each pair of nucleotides or residues is
computed in form of a scoring matrix. This scoring matrix is useful in finding a score
of each pair of nucleotides or residues. For example, NCBI-BLAST program assigns
a score of +1 for identical nucleotides and 3 for non-identical nucleotides for
highly conserved sequences. The scoring matrix for amino acid sequences are
developed considering their observed physicochemical properties and actual substi-
tution frequencies in nature. Point accepted mutation (PAM) matrix is derived from
the observed substitution rates. PAM matrices signify the substitution probabilities
of amino acids over a constant unit of evolutionary change. For example, PAM-1 is
one PAM unit which represents one substitution or accepted point mutation per
100 amino acid residues. This matrix is usually used to compare closely related
protein sequences. On the other hand, PAM-1000 matrix may be useful to compare
distantly related sequences. Usually, PAM-250 matrix is most commonly used in
alignment of protein sequences. BLOSUM matrix is an alternative to the PAM
matrix derived from observed substitution rates among related proteins. In
BLOSUM matrix, similar protein sequences are first clustered and then substitution
rates between clusters are computed. BLOSUM matrices are derived for protein
sequences having variable degree of similarity like PAM matrix. However, number-
ing pattern of BLOSUM matrix is inverse to the numbering pattern of PAM matrix.
As an example, a higher number BLOSUM matrix such as BLOSUM-62 is suitable
to compare closely related protein sequences, whereas a lower number BLOSUM
matrix such as BLOSUM-30 is appropriate for divergent protein sequences having
distant relationship. Most commonly used matrices like BLOSUM-62 and
BLOSUM-80 are appropriate for comparing protein sequences having approxi-
mately 62% and 80% similarities, respectively.

4.2 Pairwise Sequence Alignment

A dot plot is the simplest graphical method for visual comparison of two sequences
using a two-dimensional matrix and identify the regions of high similarity. In dot
plot, first sequence is plotted on the X-axis, whereas second sequence is plotted on
the Y-axis. Presence of identical residues in each position of both sequences is
indicated by a dot in the plot space. The adjacent regions with high similarity are
indicated by a diagonal in a dot plot. The abundance of dots along a diagonal line is
an indicator of multiple identical amino acids in homologous sequences. Thus, this
method is useful in finding similarity in large genomic sequences and identifying
internal repetitive sequences in protein sequences. However, this method is currently
not so popular in bioinformatics analysis and is widely used for educational
purpose only.
4.2 Pairwise Sequence Alignment 55

Fig. 4.1 Optimum global pairwise alignment using Needleman–Wunsch Algorithm. The optimal
traceback path is shown in the scoring matrix table

Dynamic programming is an efficient and fast approach to align homologous

sequences with gaps. This algorithm was developed by Needleman and Wunsch
(1970) for optimal sequence alignment. The Needleman–Wunsch algorithm breaks
down the original problem into smaller sub-problems and then finds an optimal
solution. The optimal path starts from beginning of the sequence and terminates at
the end of the sequence. Therefore, this alignment is known as global alignment. For
example, two sequences ACTGCA and ACTCA are aligned using Needleman–
Wunsch algorithm (Fig. 4.1). First, a scoring matrix having dimension of the grids
as (m + 1) x (n + 1) is created where m and n are the lengths of the sequences being
aligned. The horizontal and vertical axes are labelled with two sequences being
aligned. The alignment of two sequences is indicated by a diagonal path from upper
left-hand corner to lower right-hand corner. On the other hand, insertion of a gap in
the alignment is indicated by a horizontal move and a vertical move in the sequence
in the left axis and top axis, respectively. The first row and column of the matrix are
filled with multiples of gap penalty. Each grid is assigned a value based on a user-
defined match score, mismatch score and gap score. The first grid (2,2) representing
first column of the alignment is filled with a score based on one of three possible
values: insertion of gap in the first sequence, insertion of gap in the second sequence
and alignment of residues. Similarly, the first grid (1,1) is also filled with a score
using any of three possibilities. The first possibility is to take a value from the left
grid (2,1) and add the gap penalty to obtain the final score indicating insertion of a
gap in the sequence along the left axis. The second possibility is to take a value from
upper grid (1,2) and add the gap penalty in order to obtain a final score indicating
insertion of a gap in the sequence on the top axis. The third alternative is to take a
56 4 Sequence Alignment

Fig. 4.2 Local optimal three pairwise alignments using Smith–Waterman algorithm. The lowest
score of this matrix is zero and alignment starts from highest score and tracked back to a score of
zero. The traceback paths of three alignments are shown in specific colours

value from the grid located diagonally above in the left side (1,1) and add either
match or mismatch score to get a final score indicating alignment of residues. The
maximum score obtained out of all three possibilities is taken as a final score.
Finally, the optimal path of alignment is traced back starting from lower right-
hand corner to upper left-hand corner following the alignment score of the matrix.
This algorithm finds the best alignment based on the overall alignment score. In
order to avoid excessive gaps in the alignment, gaps are usually penalized by
subtracting gap penalty from overall score. The gap penalty includes both gap
opening penalty and gap extension penalty collectively known as affine gap
penalties. Gaps are usually found in blocks in an alignment and a simple linear
penalty for gaps is not a method of choice. Instead, affine gap penalties are
implemented in many algorithms where gap opening penalty is always higher than
the gap extension penalty. Thus, introduction of sequential gaps with affine gap
penalty in an alignment has a substantial reduction in the overall cost of penalty.
Smith and Waterman (1981) proposed a modification in the Needleman–Wunsch
algorithm in order to allow local alignment (Fig. 4.2). A local alignment finds best
matching regions within two sequences being compared. Here, the alignment path
does not start from the right-hand corner and end at the left-hand corner. Instead, it
can start and end internally anywhere and any direction in the matrix. A zero term is
added additionally in the Smith–Waterman algorithm and therefore no value exists
in the scoring matrix which is below zero. If a score below zero is obtained by any
method, a zero is placed in that position in the matrix. The traceback process begins
from the highest value in the matrix representing alignment score and ends after
reaching a value of zero.
4.3 Multiple Sequence Alignment 57

4.3 Multiple Sequence Alignment

The concept of dynamic programming can be easily extended to alignment of more

than two sequences. It finds an alignment which generates best score for the WSP
(weighted sum of pairs) objective function. First, a score for each pair of sequences is
computed followed by summing over all the pair scores to obtain a WSP function.
Here, an extra weight term for each pair is added for more reliable pairs but less
weights are assigned to closely related sequences. Overall, this method lacks an
evolutionary model and is not suitable for alignment of more than four sequences of
moderate length due to high memory and time requirements.
There are some heuristic methods available for progressive multiple sequence
alignment based on phylogeny. This method is simple and more attractive and
therefore implemented by most of the popular programs. This fast approach can
generate a phylogenetic tree from hundreds of sequences followed by their multiple
sequence alignment based on the initial tree. First, a neighbour joining phylogenetic
tree (also known as guide tree) is generated using pairwise alignment between all
sequences and then determining the pairwise distances between them. Subsequently,
the sequences are aligned based on the branching order in the tree. First, most closely
related sequences are aligned using dynamic programming followed by adding next
closely related sequences to the cluster of most closely related sequences in iterative
fashion. This iterative process follows the guide tree in placement of the indels
among all clustered sequences. This algorithm for progressive alignment has been
implemented in most widely used programs such as Clustal W and Clustal X. Clustal
W is suitable for alignment of small number of extraordinarily long sequences.
However, any misalignment of nucleotides or residues introduced during early
alignment phase cannot be corrected during latter phases in the progressive align-
ment. A better approach to overcome this problem of alignment error is to include
consistency-based scoring. This consistency-based scoring is based on WSP func-
tion. This approach improves the accuracy of the pairwise alignment by including
information from intermediate sequences. Moreover, a new version of Clustal,
Clustal Omega generates high-quality alignment of large number of protein
sequences based on hidden Markov model. T-COFFEE is a popular program for
multiple sequence alignment based on the consistency-based alignment method.
However, this program is slower and more computationally demanding than the
Clustal W. There are some programs like MUSCLE do iterative refinement towards
optimal alignment. MUSCLE can align large number of sequences in a short span of
time and, therefore, is implemented in most of the high-throughput pipeline.
PROBCONS is another popular program which implements posterior probability-
based scoring. Although this method is highly accurate but computationally expen-
sive for large datasets. DIALIGN performs multiple sequence alignment using short
conserved regions of similarity. Therefore, it is useful in finding similarity in the
conserved isolated domains of small number of sequences. The accuracy of protein
sequence alignment can be significantly improved by including structural informa-
tion. The 3D COFFEE, a variant of T-COFFEE algorithm is available on the
webserver EXPRESSO and uses protein structures from PDB database to guide
58 4 Sequence Alignment

Table 4.1 Common programs for multiple sequence alignment

Program Algorithm URL
CLUSTAL W/ Progressive alignment www.clustal.org/
CLUSTAL X
CLUSTAL Progressive alignment based on hidden Markov www.clustal.org/
OMEGA model.
DIALIGN Greedy and progressive segment-to-segment dialign.gobics.de/
alignment
T-COFFEE Tree and consistency-based alignment www.tcoffee.org/
3DCOFFEE Structure-based alignment www.tcoffee.org/Projects/
expresso/index.html
MAFFT Iterative refinement and consistency-based mafft.cbrc.jp/alignment/
alignment software/
MUSCLE Log-expectation score-based progressive www.drive5.com/muscle/
alignment and tree-dependent refinement
PROBALIGN Maximum expected accuracy alignment based probalign.njit.edu/
on posterior probability estimation standalone.html
PROBCONS Probabilistic modelling and consistency-based probcons.stanford.edu/
alignment
msa R package An R package for multiple sequence alignment bioconductor.org/
using Clustal W, Clustal omega, and MUSCLE packages/release/bioc/
algorithms html/msa.html

the alignment process. All popular multiple sequence alignment programs are listed
in Table 4.1.

4.4 Alignment Visualization and Editing

The multiple sequence alignment generated through automated methods is likely to

have some errors. These alignment errors can be removed using manual editing. It is
preferable to perform manual editing at the amino acid level for coding sequences.
The unambiguously aligned regions of the alignment are expected to be deleted. The
decision to remove the gaps from alignment should be taken judiciously. Some gaps
contain useful phylogenetic information. It is advisable to retain gaps if they are not
inserted in ambiguous manner and alignment columns with gaps are less than 50% of
total sites. The presence of gap at a particular site in a sequence may not provide any
phylogenetic information about that sequence per se but it does provide useful
information regarding its phylogenetic relationship with other aligned sequences.
The common software programs available for visualization and manual editing of
alignment are listed in Table 4.2. BioEdit is one of the most popular program for
visualization and editing of multiple sequence alignment (Fig. 4.3).
4.5 Exercises 59

Table 4.2 Common programs for visualization and editing of multiple sequence alignment
Editor Operating system URL
BIOEDIT Windows https://bioedit.software.informer.com
GENEDOC Windows https://genedoc.software.informer.com
JALVIEW Windows, Linux and Mac https://www.jalview.org/
MACCLADE Mac http://macclade.org/macclade.html
SEAVIEW Windows, Linux and Mac http://doua.prabi.fr/software/seaview

Fig. 4.3 BioEdit sequence alignment editor showing multiple sequence alignment of FOXP2
protein in mammals

4.5 Exercises

1. Find the optimal global alignment and local alignments of two sequences “AT
GAAATATACAAGTTATATCTTGGCTTTTCAGCTCTGCATCGTTTTGGG
TTCTCTTGGC” and “ATGAACGCTACACACTGCATCTTGGCTTTGCAGC
TCTTCCTCA TGGCTGTTTCTGGCTGT” using Biostrings package in the R
environment.

Solution
We will use the pairwiseAlignment function in the Biostrings to perform both
optimal global and local alignments. The scores for match and mismatch are
+1and 1, respectively. The gap opening penalty is 5 and gap extension penalty is 2.

>library(Biostrings)
>s1<-DNAString("ATGAAATATACAAGTTATAT")
>s2<-DNAString("ATGAACGCTACACACTGCAT")
>mat <- nucleotideSubstitutionMatrix(match = 1, mismatch = -1,
baseOnly = TRUE)
60 4 Sequence Alignment

>globalAlign <-pairwiseAlignment(s1, s2, substitutionMatrix = mat,

gapOpening = 5, gapExtension = 2)
>localAlign <-pairwiseAlignment(s1, s2, type = "local",
substitutionMatrix = mat,gapOpening = 5, gapExtension = 2)
> globalAlign
Global PairwiseAlignments
pattern: ATGAAATATACAAGT;TATAT
subject: ATGAACGCTACACACTGCAT
score: 4
> localAlign
Local PairwiseAlignments
pattern: [1] ATGAAATATACA
subject: [1] ATGAACGCTACA
score: 6

2. Kisspeptin is a recently discovered neuropeptide regulating reproductive function

in vertebrates. The protein sequences of human and chimpanzee are retrieved
from the Ensembl database and are described as follows:

Human:
MNSLVSWQLLLFLCATHFGEPLEKVASVGNSRPTGQQLESLGLLAPGE-
QSLPCTERKPAATARLSRRGTSLSPPPESSGSPQQPGLSAPHSRQIPAPQGA-
VLVQREKDLPNYNW NSFGLRFGKREAAPGNHGRSAGRG.
Chimpanzee:
MNSLVSWQLLLFLCATHFGEPLEEVASVGNSRPTGQQLESLGLLAPGE-
QSLPCTERKPAATARLSRRGTSLSPPPESSGSPQPGLSAPNSRQIPAPQGAV-
LVQREKDLPNYNWNSFGLRFGKREAAPGNHGRSAGRG.
Compare the similarity between human and chimpanzee sequences based on a dot
plot using seqinr package in the R environment.

Solution Two orthologous sequences are included as a text file “kisspeptin.fasta” in

working directory of R environment.

>library(seqinr)
> kisspeptin<-read.fasta("kisspeptin.fasta")
> attach(kisspeptin)
> dotPlot(Human, Chimpanzee, wsize = 10, wstep = 6, nmatch = 6)

The Fig. 4.4 indicates very high similarities between two neuropeptide
sequences from human and chimpanzee as dots are located on the main diagonal
in the dot plot.
4.6 Multiple Choice Questions 61

Fig. 4.4 A dot plot showing

similarity between human and
chimpanzee sequences

4.6 Multiple Choice Questions

1. The presence of a gap in an alignment is due to:

(a) Insertions only
(b) Deletions only
(c) Both insertions and deletions
(d) None of the above
2. Most appropriate PAM matrix for alignment of closely related sequences is:
(a) PAM-1
(b) PAM-250
(c) PAM-500
(d) PAM-1000
3. Most commonly used PAM matrix for alignment of protein sequences is:
(a) PAM-1
(b) PAM-250
(c) PAM-500
(d) PAM-1000
4. Most appropriate BLOSUM matrix for comparing closely related protein
sequences having 80% similarity is:
(a) BLOSUM-20
(b) BLOSUM-40
(c) BLOSUM-60
(d) BLOSUM-80
62 4 Sequence Alignment

5. Which of the following statement is correct for dot plot?

(a) It is a two-dimensional plot
(b) Identify regions of similarity between two sequences
(c) Identify internal repetitive regions in protein sequences
(d) All the above
6. The affine gap penalties consist of:
(a) Gap opening penalty
(b) Gap extension penalty
(c) Gap opening and gap extension penalties
(d) None of the above
7. In Smith and Waterman algorithm for local alignment, minimum value allowed
in a scoring matrix is:
(a) 1
(b) 0
(c) +1
(d) +2.
8. Which of the following multiple sequence alignment program is based on the
protein three-dimensional structures?
(a) T-COFFEE
(b) 3D-COFFEE
(c) PROBCONS
(d) DIALIGN
9. Most appropriate multiple sequence alignment program for large number of
sequences:
(a) CLUSTAL X
(b) MUSCLE
(c) T-COFFEE
(d) PROBCONS
10. Which of the following program is not used for manual editing and visualization
of multiple sequence alignment?
(a) BIOEDIT
(b) SEAVIEW
(c) PROBALIGN
(d) GENEDOC

Answer: 1. c 2. a 3. b 4. d 5. d 6. c 7. b 8. b 9. b 10. c

Summary
• Amino acid alignment is preferred over nucleotide alignment for protein coding
sequences.
• The scoring matrix for amino acid sequences is based on their observed physico-
chemical properties and actual substitution frequencies in nature.
• A pair-wise global alignment is performed using Needleman–Wunsch algorithm.
• A pairwise local alignment is performed using Smith and Waterman algorithm.
Suggested Reading 63

• The progressive alignment algorithm based on phylogeny is implemented for

multiple sequence alignment using Clustal W.

Suggested Reading
Nguyen K, Guo X, Pan Y (2016) Multiple biological sequence alignment: scoring functions,
algorithms and evaluation. Wiley, New York
Rosenberg MS (2009) Sequence alignment: methods, models, concepts, and strategies. University
of California Press, Berkeley
DeBlasio D, Kececioglu JD (2017) Parameter advising for multiple sequence alignment. Springer,
Cham
Chao K-M, Zhang L (2009) Sequence comparison: theory and methods. Springer, London
Structural Bioinformatics
5

Learning Objectives
You will be able to understand the following after reading this chapter:

• Different approaches in computational protein structure modelling such as

homology modelling, ab initio modelling and threading.
• Common scoring functions for prediction of protein structure.
• Different approaches in assessment of predicted model.
• Structural classification of protein structures.
• Principles and methods of molecular dynamic simulations.
• Principles and methods of molecular docking.
• Principles and methods of structure-based drug designing.

5.1 Introduction

The structural information regarding the three-dimensional structure of a biological

molecule is necessary to understand the mechanism of its function. For example, a
detailed understanding of three-dimensional structure of an enzyme is helpful in
understanding its mechanism of action in a metabolic reaction. Structural bioinfor-
matics applies various in silico approaches to establish the sequence-structure-
function relationship of a biological molecule. In this discipline, diverse computa-
tional methods are being developed to predict the protein structure, analyse the
protein structure and function and simulate the dynamical behaviour of a protein
molecule. Since experimental determination of a protein structure using a high-
throughput X-ray crystallography or an NMR spectroscopy method is a tedious
and expensive process, computational methods are indispensable in bridging the

# Springer Nature Singapore Pte Ltd. 2022 65

B. K. Tiwary, Bioinformatics and Computational Biology,
https://doi.org/10.1007/978-981-16-4241-8_5
66 5 Structural Bioinformatics

wide gap between the abundant protein sequence information and protein structure
information of millions of protein sequences.

5.2 Protein Structure

A protein structure can be studied at four levels: primary, secondary, tertiary and
quaternary structures. The primary structure of a protein consists of a polypeptide
sequence of amino acids which starts with the N-terminal amino acid and ends with
the C-terminal amino acid. There are 20 standard amino acids present universally
among all organisms. The amino acids are linked by a peptide bond which forms an
amide linkage between NH2 group of one amino acid with COOH group of another
adjacent amino acid. The peptide bond is unique in terms of having a rigid double-
bond character with restricted rotation. The polypeptide backbone forms various
conformations known as secondary structures. The alpha helices, beta-pleated sheets
and turns or bends are the examples of secondary structures (Fig. 5.1). The second-
ary structure of a protein further folds in a three-dimensional space to form a tertiary
structure. When a protein consists of multiple polypeptides (subunits), each subunit
combines together to form a quaternary structure of a functional protein. For

Fig. 5.1 A protein structure

of malate dehydrogenase
(PDB: 2dfd) showing
secondary structures: alpha
helices (red) and beta-pleated
sheets (blue)
5.3 Modelling of Protein Structure 67

Table 5.1 Molecular graphics applications

Application Description
RasMol A popular molecular graphics program written in C
PyMol A high-performance graphics program written in Java
jmol An open-source java-based program for molecular graphics
ccp4mg A comprehensive fully featured program written in Python
VMD A visual molecular dynamics program for large biomolecules

example, a haemoglobin molecule consists of four functional subunits, two alpha

subunits and two beta subunits, forming a functional molecule. Although protein
molecules have a definitive three-dimensional structure, many proteins are fully or
partially intrinsically disordered under native functional conditions and therefore, are
known as intrinsically disordered proteins. Popular molecular graphics applications
for visualization of protein structure are listed in Table 5.1.

5.3 Modelling of Protein Structure

The computational modelling of three-dimensional structure of a protein has been

one of the grand challenges in bioinformatics. Here, the three-dimensional structure
of a protein is predicted from its amino acid sequence. There are four different
computational approaches to model the structure of a protein: homology modelling,
threading, ab initio modelling (Fig. 5.2) and integrative modelling. Common protein
modelling software and severs are listed in Table 5.2.

5.3.1 Homology Modelling

The homology or comparative modelling of a protein is based on the evolutionary

principle that protein sequences having common evolutionary descent are likely to
have similar three-dimensional structures. Here, available templates with known
experimental structures from a related family of proteins are used to model the
protein of interest (target protein). First, few appropriate template structures related
to the target protein are identified from the Protein Data Bank (PDB) and then correct
alignment of template sequence with the target sequence is performed to evaluate the
percentage identity between template-target sequence. It is always advisable to
prefer sequence identity of 40% or more to attain a good model accuracy where
the model deviates less than 2A0 RMSD from the template structure. However, we
are prone to errors in comparative modelling and the accuracy of the model falls
drastically if the sequence identity is below 30%. After alignment, all-atom model of
the target protein is generated based on the alignment with one or more available
template structures. This all-atom model further undergoes refinement process using
an energy minimization step. Finally, the model is evaluated for overall geometrical
accuracy of the model based on some specialized scoring methods. We can obtain a
68 5 Structural Bioinformatics

Fig. 5.2 Schematic flowcharts of homology modelling, ab initio prediction and threading

Table 5.2 Protein modelling software and servers

Software/server Modelling approach URL
Modeller Homology modelling salilab.org/modeller
ModWeb Homology modelling salilab.org/modweb
HHPred Homology modelling toolkit.tuebingen.mpg.de/tools/hhpred
Rosie Ab initio prediction rosie.rosettacommons.org
Robetta Ab initio prediction robetta.bakerlab.org
SwissModel Homology modelling swissmodel.expasy.org
I-Tasser Threading zhanglab.ccmb.med.umich.edu/I-TASSER

very good model with less than 1A0 RMSD for the main-chain atoms using
homology modelling in case the identity level is above 50%. On the other hand,
low accuracy models with multiple types of structural errors are expected with an
identity level less than 30%.

5.3.2 Ab Initio Prediction

The structure prediction methods which do not rely on the availability of a template
structure are known as de novo methods. Ab initio prediction methods are a subset of
de novo methods which depends on energy functions or information inferred from
other protein structures. This method assumes a global free energy minimum for the
5.3 Modelling of Protein Structure 69

native state of the protein based on the conformational space for available protein
tertiary structures with low free energy for a sequence of amino acids. Rosetta
method is a popular ab initio method which relies on an ensemble of short structural
fragments obtained from the Protein Data Bank (PDB). A Monte Carlo search
method is used to assemble the structural fragments based on a scoring function.
This method has been applied widely for modelling protein structures and structur-
ally variable regions such as insertions and loops. Another popular method is
TASSER which relies on a threading approach to find the local fragments that are
subsequently connected into a full-length model by means of a random walk
followed by a refinement process. Although de novo methods have predicted
accurate models of some small proteins, they suffer from some inherent limitations
like huge computational demand and poor quality of models for large proteins.

5.3.3 Threading

The protein fold recognition and threading methods are useful in generating three-
dimensional models based on evolutionary distant fold templates. Threading is most
popular approach for fold recognition of proteins by fitting a sequence onto the
backbone coordinates of available protein structures. Threading has become a
generic term for fold recognition although it is a sub-class of fold recognition
methods. We can identify the proteins with known structures within the Protein
Data Bank (PDB) sharing common folds with the target sequence using fold
recognition. These proteins are subsequently used as templates to model the folds
of the target sequence. Main focus of this method is to assign folds to the target
sequence even if there is low sequence identity to known protein structures. The
basic concepts involved in threading is just reverse of the comparative modelling. In
comparative modelling, we search for a protein structure best fitting several target
sequences. In contrast, each available potential protein structure is fitted with a target
sequence in threading process. Here, each target sequence is compared with a library
of potential fold templates based on energy potentials and other similarity scores;
templates having either least energy score or best similarity score are assumed to best
fit of the fold of the target structures.

5.3.4 Integrative (Hybrid) Modelling

Integrative modelling approach computationally combines the structural information

from various sources such as X-ray crystallography, NMR spectroscopy, cryo-
electron microscopy, cross-linking coupled to mass spectroscopy, sequence analysis,
etc. to obtain a structural model of large protein complexes with better precision,
accuracy and resolution. Here, the strength of each method complements each other
in order to get a better structural model. Integrative modelling iterates through the
process of collecting data, proposing an initial model based on the data and again
collecting more data for further refinement and validation of the proposed initial
70 5 Structural Bioinformatics

model. There are four steps commonly involved in an integrative modelling. First,
the input data is collected from various experimental methods and other statistical
measures such as atomic statistical potentials and molecular mechanics force fields.
A scoring function is developed to evaluate the consistency of the model with the
input data. Multiple models are generated and their scores are improved using
various optimization algorithms such as Monte Carlo, genetic algorithm and gradient
decent. The good scoring models after optimization represent the conformations of
the system. Finally, the ensemble of all good scoring models is clustered and
analysed to obtain the overall precision and accuracy of the ensemble. It also gives
a clue to most probable experiment to be conducted in the next iteration.

5.4 Structural Genomics

The explosive growth of genomic sequences using high-throughput sequencing

technology has led to the emergence of a new discipline called functional genomics
to determine the function of genes at the genome-wide scale. Structural genomics is
the sub-discipline of the functional genomics dealing with structure-based inferences
of molecular functions of gene products of all genomic sequence families. This
robotics aided large-scale determination of protein structures is known as structural
genomics or structural proteomics. This technology-driven field has developed with
a fast pace in the last one decade due to recent developments in modern genomic
technologies such as cloning and expression of proteins as well as advanced
structure determination methods such as macromolecular crystallography and
NMR. It provides a systems-level view of protein structural world and further aids
in identification of remote homologs which are unlikely to be detected by only
sequence comparison.

5.5 Scoring Functions

The three-dimensional structure of a protein is predicted based on the potential

energy function. Scoring functions are multidimensional matrices commonly used
for structure prediction of proteins. These functions are useful in assessment of both
experimentally determined and computationally predicted protein structures. Since
scoring functions account for both atom–atom interaction and solvation effect, as a
result these functions are also known as effective energy functions. There are four
main components of a scoring function: a body definition, a geometrical descriptor, a
reference system and a set of restraints. A body definition may consist of either
single atoms or centroids of a group of atoms or different atoms or centroids sharing
a common physicochemical property. A geometrical descriptor describes
interactions between different bodies using some measures such as angles, pairwise
distances, dihedral or torsion angles and radial or angular densities. A reference
system is the weighted average state of all possible states of a system and used as a
reference setting to calculate a specific score for a particular state. A set of restraints
5.6 Model Assessment 71

either defines the upper and lower limit of the score or divides into different classes
requiring a specific treatment.
Some of the most common scoring functions are contact scoring functions,
distance-dependent scoring functions, accessible surface scoring functions and com-
bined scoring functions. Contact scoring function is the simplest form of scoring
function due to minimum size of the matrix filled with few experimental data. For
example, contact potential is a squared two-dimensional matrix of 20 rows and
20 columns for a body definition of alpha carbons of 20 standard amino acids. These
functions are generally derived symmetrically for beta carbons or side chain
centroids of standard amino acids. Distance-dependent scoring functions are most
widely used in the prediction of protein structure. It is represented by a matrix of four
dimensions: the first two dimensions include total number of bodies defined, the
third dimension distinguishes between local and non-local interactions and the
fourth dimension describes distance ranges or bins in order to represent interactions.
Accessible surface scoring functions tend to capture the propensity of interaction of
defined bodies with the solvent. It describes the solvent accessibility at the residue or
atomic level. The combined scoring function is a combination of information from
both accessible surface and contact plus distance-dependent approaches. The contact
and distance-dependent approach deals with intramolecular protein interactions,
whereas accessible surface approach is concerned with the interactions between
the solvent and the protein. A normalization step is required to combine these two
independent scoring functions.
There are various applications of scoring functions in protein structure prediction.
It helps in selecting the best model in terms of accuracy out of different alternative
models. Moreover, the fold assessment of a predicted model is performed using a
scoring function to evaluate the correct fold. The geometrical accuracy of the overall
model as well as individual regions can also be assessed using scoring functions. The
stability of a protein structure can also be predicted using scoring functions in case of
mutant screening.

5.6 Model Assessment

The primary goal of model assessment is to evaluate the overall accuracy of the
predicted model. In addition, it also selects the most accurate model from a set of
alternative models and evaluates the accuracy of the different regions of the model.
There are four common approaches of model assessment, namely physics-based
energies, knowledge-based potentials, combined scoring functions and clustering
approaches. Chemical force fields are the physics-based energy function of particles
in a system and are derived from both quantum mechanical calculations and experi-
mental findings. The force field energy function is represented by the energy from
chemical bonds between interacting atoms (i.e. distances, angles and dihedral
angles) and interactions between non-bonded atoms (i.e. electrostatic and van der
Waals interactions). These physics-based approaches are useful in selecting near-
native structure models from a set of predicted models. The second approach is the
72 5 Structural Bioinformatics

knowledge-based potential based on empirical observations on the known protein

structures. It is implemented in form of statistical potentials or potentials of mean
force. Several statistical potentials such as contact, distance, solvent accessibility and
pairwise interaction along with solvent accessibility are derived to assess the struc-
tural features of a model. The third approach is combined scoring functions which
are an optimized weighted combination of various scores from physics and
knowledge-based approaches. However, all three scoring approaches often fail to
detect the near-native structure. Finally, multiple models are generated from the
same sequence and are compared structurally to select the best model in the
clustering approach. The quality of the scoring functions used to generate an
ensemble of conformations determines the accuracy of the model. Thus, useful
information from an ensemble of possible conformations are converted to statistical
probabilities in order to identify the near-native structure of the protein. However,
clustering approach cannot assess the quality of the model alone and aided by other
scoring functions.

5.7 Classification of Protein Structures

The structural classification of proteins is largely based on the identification of

individual domains and their evolutionary relationships. The structural similarity
between two proteins is usually assessed by superimposing the C alpha atoms of a
protein structure on the top of another one. The distances between two equivalent
atoms in both superimposed proteins are measured in terms of root mean square
deviation (RMSD). A low RMSD value less than 3.5 A0 indicates a close similarity
in protein structures and consequently a possible structural homology between two
proteins. There are only two manually curated protein structure classification
databases, namely CATH database and SCOP database covering the entire protein
structures deposited in the PDB. The CATH stands for four levels of classification
hierarchy, namely Class, Architecture, Homologous superfamily and Topological
motif, and is regularly updated using automated algorithms followed by manual
curation. Domains are classified in the CATH database using both automated and
manual approaches based on similarity at the sequence, structure and functional
levels. Firstly, domains are classified into three main classes based on the secondary
structures: mainly alpha-helical structures, mainly beta sheet structures and mixed
alpha-beta structures. Further, they are classified into 40 groups based on the
architecture or gross arrangement of secondary structures independent of their
connectivity in a three-dimensional space. Thirdly, they are classified into more
than 1000 topological motifs or fold groups based on the arrangement of secondary
structures and their connectivity as well. Finally, domains are clustered into more
than 2000 homologous superfamilies based on their evolutionary relationships.
These domains also share substantial structural similarity along with similarity at
the sequence and/or functional level as well.
The structural classification of proteins (SCOP) database also divides a protein
structure into one or more domains like the CATH database but the entire
5.8 Molecular Dynamics Simulations 73

classification process is manual. It follows a classification hierarchy at various major

levels. The highest level is the class which is arranged according to the secondary
structure content of the protein. The SCOP contains five major classes such as all
alpha, all beta, alpha/beta (interspersed alpha helices and beta strands), alpha+beta
(segregated alpha helices and beta strands) and multidomain (having two or more
domains from different classes). The architecture level is missing in the SCOP
database unlike CATH database. However, fold groups and superfamily groups
are included in the hierarchy of SCOP database like CATH database.

5.8 Molecular Dynamics Simulations

Molecular dynamics simulations are linked to the concept of intermolecular

interactions in a complex system. These models are based on the force field and
provide useful insights into various intermolecular processes such as protein folding,
protein-ligand interactions and protein dynamics. Molecular dynamics uses simple
approximations to the laws of Newtonian physics in order to simulate atomic
motions (atomistic representation). The process of molecular dynamic simulation
can be performed in a few steps (Fig. 5.3). The initial model of the molecular system
is obtained from the experimental structure derived from nuclear magnetic resonance
(NMR) and X-ray crystallography or computational homology modelling. The
forces operating on each atom of the system are estimated from a potential energy
function equation describing the forces derived from interactions between bonded
and non-bonded atoms which are as follows:

Total energy function E total ¼ ½E b þ E θ þ Eϕ þ E ω  þ ½E vdW þ E elc 

Fig. 5.3 The workflow of a molecular dynamic (MD) simulation of a PDB structure (PDB:2dfd)
generating a large dataset of trajectories for further graphical analysis
74 5 Structural Bioinformatics

Table 5.3 MD simulation software

Program Description
GROMOS A package for dynamic modelling of biomolecules
GROMACS A package for high-performance molecular dynamics of biomolecules
NAMD A parallel molecular dynamics program for high-performance simulation.
AMBER A package of biomolecular simulation programs
CHARMM A molecular simulation program for biological systems
Bio3d An R package to analyse protein structure and trajectory data
VMD A software to analyse and animate trajectories of MD simulation
MDplot An R package providing plotting functions to MD simulation output

where Eb is a bond length potential term, Eθ is a bond angle potential term, Eφ is a

dihedral angle (torsion) potential term, EvdW is a non-bonded van der Waals interac-
tion term and Eelc is a non-bonded electrostatic term. In addition to these energy
functions, some constraints are often applied to force a molecule in a desired
conformation as an optional term of the total potential energy function.
Chemical bonds and atomic angles are modelled based on simple virtual springs
(bond stretching and angle bending), whereas dihedral angles (rotations about a
bond) are modelled using a sinusoidal function. Non-bonded forces such as van der
Waals interactions and electrostatic interactions are modelled using Lennard-Jones
potential and Coulomb’s law, respectively. The actual behaviour of real molecules in
motion can only be modelled after parameterizing the energy terms to fit quantum
mechanical calculations. The parameterizing process consists of identification of
ideal stiffness and length of the spring representing chemical bonds and atomic
angles, estimation of best partial atomic charges for calculating electrostatic interac-
tion energies, identification of proper van der Waals atomic radii, and so forth. All
these parameters are known as force field due to their individual contributions of
atomic forces controlling molecular dynamics. There are some useful engines for
molecular dynamic simulation (Table 5.3). The GROMOS (GROningen MOlecular
Simulation), AMBER (Assisted Model Building with Energy Refinement),
CHARMM (Chemistry at HARvard Molecular Mechanics) and GROMACS
(GROningen MAchine for Chemical Simulations) are commonly used force fields
that only differ in their mode of parameterization. After estimation of the forces
acting on the system atoms, the position of atoms is moved following the Newton’s
laws of motion. Further, simulation time is repeatedly advanced by a small fraction
of a second for millions of times. These simulations are usually performed on
computer clusters or supercomputers due to large number of calculations required.
The MD simulation output are analysed using standard plots such as root mean
square deviation (RMSD) and root mean square fluctuation (RMSF). RMSD is the
measure of positional divergence of atoms during simulation over a time range
(Fig. 5.4), whereas RMSF indicates the degree of positional variation of a certain
atom over time (Fig. 5.5). The major limitations of molecular dynamics simulations
are requirement of further refinement of force fields and very high computational
demand.
5.9 Docking of Ligands and Proteins 75

0.25 0.30
0.15 0.20
RMSD [nm]
0.10
0.05

Legend
Wild-type peptide
Mutant peptide
0.00

0 5 10 15 20 25 30 35
time [ns]

Fig. 5.4 The RMSD curves of two independent MD simulation trajectories in 35 nanoseconds

Legend
0.5

Wild-type peptide
Mutant peptide
0.4
RMSF [nm]
0.3
0.2
0.1
0.0

0 50 100 150 200 250 300

atom number

Fig. 5.5 The RMSF curves of two independent MD simulations showing positional variation of
first 300 atoms

5.9 Docking of Ligands and Proteins

The interaction between small-molecule ligands or proteins with their protein

receptors in an aqueous solution is crucial in understanding the mechanistic basis
of pharmaceutically active compounds. The ligand-receptor complex is not only
stabilized by the intermolecular van der Waals interactions, hydrogen bonds and
electrostatic interactions but also by the desolvating nonpolar parts of the molecular
surface. Both ligand and receptor interact with solvent before binding. The
desolvation of the nonpolar parts of the surface releases water molecules resulting
in the increase of entropy which further strengthens complex formation. This
interaction between the ligand and its receptor is known as docking (Fig. 5.6) and
has laid a foundation for the computational structure-based drug design. Some
76 5 Structural Bioinformatics

Fig. 5.6 Docking of a small ligand into the crystal structure of its receptor to form a stable complex
(PDB:5toa)

popular docking programs such as Autodock, Glide, GOLD and LigandFit have
been developed on the fundamental principles of physical chemistry. Molecular
docking is useful in finding and optimizing lead compounds and thereby proved as
a powerful tool of drug discovery. A docking program has two important
components: a docking algorithm to search the configurational and conformational
degrees of freedom and a scoring function for evaluation. The docking algorithm
tends to find the global energy minimum by searching the potential energy landscape
extensively. Only translational and rotational degrees of freedom are allowed for the
ligand interacting with receptor active site in rigid docking. In contrast, flexible
ligand docking involves torsional degrees of freedom for the ligand by allowing
variation in ligand dihedral angles. The scoring function, in general, evaluates not
only the steric complementarity between the ligand and its receptor but their
chemical complementarity as well. A popular docking program Autodock
implements simulated annealing Monte Carlo algorithms having tens of thousands
of steps during each cycle. First, the ligand is placed randomly onto the binding site
of the receptor and allowed to move towards a global energy minimum. The
structure is minimized after each move and subsequently the energy of the new
structure is measured. The simulation may consist of several cycles with decreasing
temperature in each cycle. The lowest energy in the previous cycle becomes the
starting point of the next cycle.
The protein-ligand binding affinity is measured in terms of free energy gains. The
largest contributor to this free energy gain emerges from the displacement of water
from the hydrophobic regions of the receptor. Another important source of free
energy after ligand binding is the hydrogen bonds between the ligand and receptor
protein. This free energy gain is achieved by displacing the water bound to the
receptor protein by the ligand. The purpose of a docking program is to use a global
scoring function that can be equally applied to all protein-ligand complexes. The
5.10 Structure-Based Drug Design 77

scoring function, in general, consists of three force field terms for van der Waals
interactions, hydrogen bonding and electrostatic interactions and a term representing
for desolvation effects. The strength of interaction between a ligand and a protein in
terms of binding free energy ΔG bind is as follows.

ΔGbind ¼ E MM  TΔSSolute þ ΔGSolvent

where E MM is the internal energy of the molecule determined from the bond lengths,
the bond angles and the dihedral angles followed by enthalpic contributions to free
energy from the van der Waals interactions and the electrostatic interactions. The
solute entropy consists of terms for translational, rotational, vibrational and confor-
mational entropy. The solvent free energy includes both polar and nonpolar terms
and represents both entropic and enthalpic effects of the solvent. The accuracy of a
docking program is usually evaluated by comparison of the predicted ligand-
receptor complex structure with the experimental structure obtained by X-ray crys-
tallography. During such comparison, each ligand is docked to its own cognate
receptor conformation (self-docking). Sometimes accuracy of the program is also
tested by docking a ligand to non-cognate receptor conformations such as apo
structure of receptor or co-crystallized structure with a different ligand. This method
of testing accuracy of a docking program is known as cross-docking.

5.10 Structure-Based Drug Design

A small-molecule ligand is tailored to fit the three-dimensional topology of the

binding pocket of a target protein during structure-based drug design. This compu-
tational method has a very important role in the overall drug discovery process. The
availability of a three-dimensional structure of target protein at atomic resolution
preferably with a bound ligand is prerequisite for a structure-based drug design. The
first and foremost challenge of this process is to quantify the binding affinity of a
ligand-protein complex. Quantitative structure-activity relationships (QSAR) is the
most common and an efficient approach to determine quantitatively the binding
affinity of a complex using linear or multiple regression methods. This technique is
widely used not only for lead identification and optimization but also to predict
ADMET (adsorption, distribution, metabolism, elimination, toxicity) properties in
pharmaceutical industries. QSARs attempt to find a correlation between the struc-
tural properties of potential drug candidates and their binding affinity towards a
macromolecular target. The structure-based drug design has become more powerful
with ever increasing availability of three-dimensional structures from X-ray diffrac-
tion analysis. The structure-activity relationships may also depend upon the three-
dimensional structure of a ligand (3D-QSAR) using comparative molecular field
analysis (CoMFA). Here, a grid or a surface is used as a surrogate of the binding site
of the protein receptor (pharmacophore) having all its steric and electrostatic
properties. A series of compounds in their bioactive conformations are then
superimposed around this surface or grid. The quality of this predictive model
78 5 Structural Bioinformatics

depends upon the correct superposition of ligands and available structural informa-
tion regarding the target protein. The 4D-QSAR is an advanced form of multidimen-
sional QSAR (mQSAR) where energetically feasible binding modes are included in
a 4D-dataset such as different ligand conformations, ligand poses and protonation
states. Using this dataset, the actual bioactive conformation or true binding mode of
the ligand is identified based on the QSAR algorithm. Each ligand molecule is
represented as a single entity in the 3D-QSAR, whereas the ligand molecule is
represented in an ensemble of conformations, poses, protonation states, tautomeric
forms and stereoisomers in the 4D-QSAR. The binding pocket of a protein structure
is not only modified by a bound ligand (induced-fit effect) but other biochemical
properties of a receptor protein such as hydrophobicity, hydrophilicity and solvent
accessibility are also altered. In some proteins like GPCR proteins, the structure of
target protein is not known and realistic induced-fit effect cannot be determined.
Thus, a fifth dimension has been added in the 5D-QSAR where we can identify
realist induced-fit protocols in addition to 3D topology of the binding pocket. The
protein target is strongly affected by solvation phenomena (ligand desolvation,
solvent stripping and proton transfer) when bound to a small-molecule ligand.
Thus, 6D-QSAR accounts for different solvation models simultaneously and thereby
allows more realistic simulation of the ligand-protein binding process. 6D-QSAR is
implemented in a receptor modelling program known as Quasar allowing optimiza-
tion based on genetic algorithm. The binding energy in Quasar is determined as:

E binding ¼ E ligandreceptor  E ligand desolvation  Eligand strain  TΔS  E induced fit

where Eligand-receptor is composed of Eelectrostatic, Evan der Waals, Ehydrogen bonding and
Epolarization.
Raptor is another popular program for 6D-QSAR based on a different scoring
function. The underlying scoring function is composed of directional terms for
hydrogen bonding and hydrophobicity considering solvation effect implicitly. The
binding energy in Raptor is calculated as:

Ebinding ¼ E ligandreceptor  TΔS  Einduced fit

where E ligand-receptor is composed of Ehydrogen bonding (shell 1), Ehydrophobic (shell

1), E hydrogen bonding (shell 2) and Ehydrophobic (shell 2).

5.11 Exercises

1. Malate dehydrogenase is homodimer enzyme catalysing conversion of oxaloace-

tate and malate in citric acid cycle. Read the human malate dehydrogenase PDB
structure (ID:2DFD) in the R environment and find the total number of chains,
atoms and residues present in the structure. Compute the binding sites and plot the
B-factors of residues present in the enzyme.
5.11 Exercises 79

Solution (Fig. 5.7)

We will load an R package “bio3d” for reading and finding different properties of
PDB structure using following scripts:

> library(bio3d)
> help(package=bio3d)
> MDH<-read.pdb("2dfd")
> MDH
Call: read.pdb(file = "2dfd")
Total Models#: 1
Total Atoms#: 10213, XYZs#: 30639 Chains#: 4 (values: A B C D)
Protein Atoms#: 9195 (residues/Calpha atoms#: 1260)
Nucleic acid Atoms#: 0 (residues/phosphate atoms#: 0)
Non-protein/nucleic Atoms#: 1018 (residues: 814)
Non-protein/nucleic resid values: [ CL (7), HOH (799), MLT (4), NAD (4) ]
> bs<-binding.site(MDH)
> bs
$inds
Call: NULL
Atom Indices#: 1310 ($atom)
XYZ Indices#: 3930 ($xyz)
+ attr: atom, xyz
$resnames
[1] "LEU 12 (A)" "GLY 13 (A)" "SER 15 (A)" "GLY 16 (A)" "GLY 17 (A)"
[6] "ILE 18 (A)" "GLY 19 (A)" "TYR 38 (A)" "ASP 39 (A)" "ILE 40 (A)"
[11] "ALA 41 (A)" "PRO 81 (A)" "ALA 82 (A)" "GLY 83 (A)" "VAL 84 (A)"
[16] "PRO 85 (A)" "ARG 86 (A)" "THR 91 (A)" "ARG 92 (A)" "ASP 93 (A)"
[21] "LEU 95 (A)" "ASN 99 (A)" "ILE 102 (A)" "LEU 106 (A)" "ILE 122 (A)"
[26] "ALA 123 (A)" "ASN 124 (A)" "VAL 126 (A)" "PRO 145 (A)" "ASN 146 (A)"
[31] "VAL 151 (A)" "LEU 154 (A)" "ASP 155 (A)" "ARG 158 (A)" "HIS 182 (A)"
[36] "ALA 183 (A)" "GLY 184 (A)" "GLY 216 (A)" "THR 217 (A)" "VAL 220 (A)"
[41] "SER 228 (A)" "ALA 229 (A)" "THR 230 (A)" "MET 233 (A)" "LYS 261 (A)"
[46] "ILE 281 (A)" "LEU 12 (B)" "GLY 13 (B)" "SER 15 (B)" "GLY 16 (B)"
[51] "GLY 17 (B)" "ILE 18 (B)" "GLY 19 (B)" "TYR 38 (B)" "ASP 39 (B)"
[56] "ILE 40 (B)" "ALA 41 (B)" "PRO 81 (B)" "ALA 82 (B)" "GLY 83 (B)"
[61] "VAL 84 (B)" "PRO 85 (B)" "ARG 86 (B)" "ARG 92 (B)" "LEU 95 (B)"
[66] "ASN 99 (B)" "ILE 102 (B)" "LEU 106 (B)" "ILE 122 (B)" "ALA 123 (B)"
[71] "ASN 124 (B)" "VAL 126 (B)" "VAL 151 (B)" "LEU 154 (B)" "ASP 155 (B)"
[76] "ARG 158 (B)" "HIS 182 (B)" "ILE 191 (B)" "VAL 198 (B)" "ASP 199 (B)"
[81] "PHE 200 (B)" "LEU 205 (B)" "GLY 216 (B)" "THR 217 (B)" "VAL 220 (B)"
[86] "SER 228 (B)" "ALA 229 (B)" "THR 230 (B)" "MET 233 (B)" "LEU 12 (C)"
[91] "GLY 13 (C)" "SER 15 (C)" "GLY 16 (C)" "GLY 17 (C)" "ILE 18 (C)"
[96] "GLY 19 (C)" "TYR 38 (C)" "ASP 39 (C)" "ILE 40 (C)" "ALA 41 (C)"
[101] "PRO 81 (C)" "ALA 82 (C)" "GLY 83 (C)" "VAL 84 (C)" "PRO 85 (C)"
[106] "ARG 86 (C)" "THR 91 (C)" "ARG 92 (C)" "ASP 93 (C)" "LEU 95 (C)"
80 5 Structural Bioinformatics

[111] "ASN 99 (C)" "ILE 102 (C)" "THR 105 (C)" "LEU 106 (C)" "ILE 122 (C)"
[116] "ALA 123 (C)" "ASN 124 (C)" "VAL 126 (C)" "VAL 151 (C)" "LEU 154 (C)"
[121] "ASP 155 (C)" "ARG 158 (C)" "HIS 182 (C)" "ALA 183 (C)" "GLY 184 (C)"
[126] "GLY 216 (C)" "THR 217 (C)" "VAL 220 (C)" "SER 228 (C)" "ALA 229 (C)"
[131] "THR 230 (C)" "MET 233 (C)" "LYS 261 (C)" "SER 262 (C)" "GLN 263 (C)"
[136] "LEU 12 (D)" "GLY 13 (D)" "SER 15 (D)" "GLY 16 (D)" "GLY 17 (D)"
[141] "ILE 18 (D)" "GLY 19 (D)" "TYR 38 (D)" "ASP 39 (D)" "ILE 40 (D)"
[146] "ALA 41 (D)" "PRO 81 (D)" "ALA 82 (D)" "GLY 83 (D)" "VAL 84 (D)"
[151] "PRO 85 (D)" "ARG 86 (D)" "THR 91 (D)" "ARG 92 (D)" "ASP 93 (D)"
[156] "LEU 95 (D)" "ASN 99 (D)" "ILE 102 (D)" "LEU 106 (D)" "VAL 121 (D)"
[161] "ILE 122 (D)" "ALA 123 (D)" "ASN 124 (D)" "VAL 126 (D)" "VAL 151 (D)"
[166] "LEU 154 (D)" "ASP 155 (D)" "ARG 158 (D)" "HIS 182 (D)" "ALA 183 (D)"
[171] "GLY 184 (D)" "ILE 191 (D)" "VAL 198 (D)" "ASP 199 (D)" "PHE 200 (D)"
[176] "LEU 205 (D)" "GLY 216 (D)" "THR 217 (D)" "VAL 220 (D)" "SER 228 (D)"
[181] "ALA 229 (D)" "THR 230 (D)" "MET 233 (D)"
$resno
[1] 12 13 15 16 17 18 19 38 39 40 41 81 82 83 84 85 86 91
[19] 92 93 95 99 102 106 122 123 124 126 145 146 151 154 155 158 182 183
[37] 184 216 217 220 228 229 230 233 261 281 12 13 15 16 17 18 19 38
[55] 39 40 41 81 82 83 84 85 86 92 95 99 102 106 122 123 124 126
[73] 151 154 155 158 182 191 198 199 200 205 216 217 220 228 229 230 233 12
[91] 13 15 16 17 18 19 38 39 40 41 81 82 83 84 85 86 91 92
[109] 93 95 99 102 105 106 122 123 124 126 151 154 155 158 182 183 184 216
[127] 217 220 228 229 230 233 261 262 263 12 13 15 16 17 18 19 38 39
[145] 40 41 81 82 83 84 85 86 91 92 93 95 99 102 106 121 122 123
[163] 124 126 151 154 155 158 182 183 184 191 198 199 200 205 216 217 220 228
[181] 229 230 233
$chain
[1] "A" "A" "A" "A" "A" "A" "A" "A" "A" "A" "A" "A" "A" "A" "A" "A" "A" "A"
[19] "A" "A" "A" "A" "A" "A" "A" "A" "A" "A" "A" "A" "A" "A" "A" "A" "A" "A"
[37] "A" "A" "A" "A" "A" "A" "A" "A" "A" "A" "B" "B" "B" "B" "B" "B" "B" "B"
[55] "B" "B" "B" "B" "B" "B" "B" "B" "B" "B" "B" "B" "B" "B" "B" "B" "B" "B"
[73] "B" "B" "B" "B" "B" "B" "B" "B" "B" "B" "B" "B" "B" "B" "B" "B" "B" "C"
[91] "C" "C" "C" "C" "C" "C" "C" "C" "C" "C" "C" "C" "C" "C" "C" "C" "C" "C"
[109] "C" "C" "C" "C" "C" "C" "C" "C" "C" "C" "C" "C" "C" "C" "C" "C" "C" "C"
[127] "C" "C" "C" "C" "C" "C" "C" "C" "C" "D" "D" "D" "D" "D" "D" "D" "D" "D"
[145] "D" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D"
[163] "D" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D"
[181] "D" "D" "D"
> plot.bio3d(MDH$atom$b[MDH$calpha], sse=MDH, typ="l", ylab="B-
factor")
5.11 Exercises 81

Fig. 5.7 Residue B-factors for malate dehydrogenase (PDB ID 2DFD) with secondary structure
annotations in the marginal regions

2. Malate dehydrogenase and lactate dehydrogenase are homologous metabolic

enzymes. Download the PDB structures of human malate dehydrogenase (ID:
2DFD) and human lactate dehydrogenase (ID: 5ZJE) in the R environment.
Perform following tasks in the R environment using Package “bio3d”:
(a) Split and align the chains present in both dehydrogenase structures and
compute the sequence identity and root mean square deviation (RMSD)
between different chains of the protein structure.
(b) Plot the histogram of RMSD and generate a dendrogram of RMSD clusters.

Solution (Figs. 5.8, 5.9, 5.10, and 5.11)

> library(bio3d)
> ids <- c("2dfd","5zje")
> raw.files <- get.pdb(ids)
trying URL 'https://files.rcsb.org/download/2dfd.pdb'
Content type 'application/octet-stream' length unknown
downloaded 920 KB
trying URL 'https://files.rcsb.org/download/5zje.pdb'
Content type 'application/octet-stream' length unknown
82 5 Structural Bioinformatics

downloaded 2.5 MB
> files <- pdbsplit(raw.files, ids)
> pdbs <- pdbaln(files,exefile="msa")
> pdbs$id <- substr(basename(pdbs$id),1,6)
> seqidentity(pdbs)

Fig. 5.8 Sequence identity between different chains of human malate dehydrogenase and lactate
dehydrogenase

> rmsd(pdbs, fit=TRUE)

Fig. 5.9 Root mean square deviation (RMSD) between different chains of human malate dehydro-
genase and lactate dehydrogenase

> core <- core.find(pdbs)

core.inds <- print(core, vol=1.0)
> xyz <- pdbfit( pdbs, core.inds )
> rd <- rmsd(xyz)
> hist(rd, breaks=40, xlab="RMSD (Å)", main="Histogram of RMSD")
5.11 Exercises 83

Fig. 5.10 A histogram of RMSD between different chains of human malate dehydrogenase and
lactate dehydrogenase

>hc <- hclust(as.dist(rd))

> hclustplot(hc, k=3, ylab="RMSD (Å)", main="RMSD Cluster
Dendrogram")

Fig. 5.11 A RMSD cluster dendrogram based on RMSD between different chains of malate
dehydrogenase and lactate dehydrogenase
84 5 Structural Bioinformatics

5.12 Multiple Choice Questions

1. Which of the following is NOT a molecular visualization application?

(a) RasMol
(b) PyMol
(c) Modeller
(d) jmol
2. Which of the following structure prediction method does NOT rely on a
template structure?
(a) Homology modelling
(b) Ab initio modelling
(c) Threading
(d) Integrative modelling
3. The minimum percent identity required between a template structure and a target
sequence for accurate homology modelling is:
(a) 20%
(b) 30%
(c) 40%
(d) 50%
4. Which of the following structure prediction software is based on threading?
(a) I-TASSER
(b) Modeller
(c) SwissModel
(d) HHPred
5. Which of the following interaction is modelled using Lennard-Jones potential
during MD simulation?
(a) van der Waals interaction
(b) Electrostatic interaction
(c) Both van der Waals and electrostatic interactions
(d) None of the above
6. The potential energy function equation consists of following force(s):
(a) Dihedral angle potential
(b) Electrostatic interaction
(c) van der Waals interaction
(d) All the above
7. The global scoring function used by a docking program consists of:
(a) van der Waals interactions
(b) Hydrogen bonding
(c) Electrostatic interactions
(d) All the above
8. Which of the following statement is true for QSAR?
(a) QSAR is useful in lead identification and optimization
(b) QSAR predicts the ADMET properties
(c) QSAR is based on linear or multiple regression method
(d) All the above
5.12 Multiple Choice Questions 85

9. A popular program for 6D-QSAR is:

(a) Quasar
(b) HQSAR
(c) CoMFA
(d) CoMSA
10. Which of the following is NOT a docking program?
(a) Autodock
(b) Glide
(c) Bio3d
(d) GOLD

Answer: 1. c 2. b 3. c 4. a 5. a 6. d 7. d 8. d 9. a 10. c

Box 5.1 B-Factors of a Protein Structure

The B-factors of a protein structure are also known as temperature structures
and represent the flexibility of a protein due to fluctuation of C alpha atoms
around their average positions. Due to this flexibility, there is a constant
movement of polypeptide backbones and side chains of a protein molecule.
The high B-factors indicate that the residue positions in the protein have a
higher flexibility than the average value, whereas low B-factors are the
reflections of rigid positions in the protein molecule. Buried residues in the
core of the protein molecule are likely to have low B-values and therefore are
more rigid in comparison to the residues present on the surface of protein.

Summary
• Homology modelling of a protein structure is performed using evolutionary
related target protein structure from PDB.
• Ab initio modelling depends on energy functions or information obtained from
other protein structures.
• Threading assigns folds in the target protein based common folds in the template
PDB structures.
• Integrative modelling is an iterative process starting with a proposed initial model
followed by several rounds of refinement and validation.
• Large-scale determination of protein structures for a systems level view of protein
structural world is known as structural genomics.
• The CATH and SCOP databases describe protein structural classification cover-
ing the entire protein structures deposited in the PDB.
• Molecular dynamics simulations provide useful insights into mechanistic basis of
protein folding, protein ligand interactions and protein dynamics.
• Molecular docking is useful in identification and optimization of lead compounds
during a drug discovery process.
86 5 Structural Bioinformatics

• Quantitative structure-activity relationships (QSAR) determine quantitatively the

binding affinity of a drug-macromolecular target complex using linear or multiple
regression methods.

Suggested Reading
Schwede T, Peitsch M (2008) Computational structural biology: methods and applications. World
Scientific, Hackensack
Buxbaum E (2007) Fundamentals of protein structure and function. Springer, Cham
Becker OM, Karplus M (2006) A guide to biomolecular simulations. Springer, Dordrecht
Rigden DJ (2009) From protein structure to function with bioinformatics. Springer, Dordrecht
Lesk AM (2001) Introduction to protein architecture: the structural biology of proteins. Oxford
University Press, Oxford
Molecular Evolution
6

Learning Objectives
You will be able to understand the following after reading this chapter:

• A general understanding of molecular evolutionary processes.

• Neutral and nearly neutral theories of evolution.
• The concepts of homology, molecular clock and genetic distances.
• A basic understanding of nucleotide substitution models.
• Phylogenetic tree and various methods of its reconstruction.
• Estimation of molecular signatures of natural selection.
• Evolutionary understanding of protein structures.
• Evolution of pathogenic viruses including human SARS-CoV-2.

6.1 Introduction

Evolution is a continuous process in natural populations in terms of regular

fluctuations in gene (allele) frequencies. Some alleles show an increasing trend in
frequency and reach up to 100% in order to be fixed in a population. On the other
hand, other alleles tend to decrease in frequency and ultimately lost in a population.
The molecular evolution is the result of accumulation of mutations in genes in form
of substitution, insertion/deletion, recombination and gene conversion. Mutations
generate many genetic variants in a population and this phenomenon is known as
genetic polymorphism. The genetic polymorphism of a population is not maintained
for indefinite period of time and new mutations are subjected to filtration by two
evolutionary forces, namely natural selection and genetic drift, and ultimately fixed
or eliminated in a population. The new morphological and physiological characters
in a new species appear due to fixed mutations in its genome. The fixed mutations in

# Springer Nature Singapore Pte Ltd. 2022 87

B. K. Tiwary, Bioinformatics and Computational Biology,
https://doi.org/10.1007/978-981-16-4241-8_6
88 6 Molecular Evolution

a species are known as substitutions. The substitution rate is computed in terms of

accumulation of genetic differences between individuals in an evolutionary time-
scale. Consequently, substitution is the manifestation of differences in nucleotide or
amino acid sequences between two taxa. Overall, the genetic divergence between
two populations are generated by various factors such as underlying mutation rate,
generation time and multiple evolutionary forces. When a particular allele is more fit
than other alleles in an environment, it is subjected to positive selection; lower fitness
of alleles exposes them to negative selection. In certain traits like sickle cell trait,
heterozygote is more advantageous than the homozygotes in terms of protection
against malaria. A genetic polymorphism is maintained in the population under the
influence of balancing selection. In addition, rate of fixation of a mutation depends
on the effective population size (Ne) producing offspring which is substantially
smaller than the overall population size (N). The overall evolutionary process is a
blend of both deterministic and stochastic processes. Natural selection operates on
the gene frequency through differential reproduction and survival of some variants in
an environment. Thus, evolutionary trajectory of selection can be predicted if
prevailing environmental condition and fitness of the variant is known. In contrast,
random genetic drift is a result of statistical fluctuations in gene frequency in a
population with a small effective population size and ultimately determines the
mutation rate. The change in gene frequency does not have a particular direction
for genetic drift unlike modus operandi of natural selection. Nonsynonymous
mutations are likely to be subjected to selective pressure and cause phenotypic
changes in an organism. In contrast, synonymous mutations are largely neutral and
are fixed in a population under the impact of genetic drift. The comparison of
nonsynonymous and synonymous substitution rate gives us an assessment of posi-
tive and negative selection operating on genes.

6.2 Neutral and Nearly Neutral Theories of Evolution

Motoo Kimura proposed his landmark neutral theory of molecular evolution in 1968
which is central to the principles of molecular evolution. The rate of molecular
evolution among proteins varies widely from the slowest evolving histone IV (10
11
substitutions per site per year) to the fastest evolving fibrinopeptide (9  10
9
substitutions per site per year). Based on this observation, Kimura and Ohta pro-
posed that functionally less important molecules or parts of such molecules are likely
to evolve faster than more important proteins. In other words, functionally important
molecules or their parts are selectively constrained due to their detrimental effect on
the fitness of the organism. The major evolutionary force involved in genetic
substitutions is stochastic fixation of mutations. The majority of these substitutions
are product of random fixation of neutral and nearly neutral mutations. The synony-
mous nucleotide changes and changes in nucleotides in introns, pseudogenes, other
non-coding and nonregulatory regions are examples of neutral mutations. Some
nonsynonymous mutations having no effect on the protein functions are also treated
as neutral mutations. Since effective population size is usually very small in
6.3 The Concept of Homology 89

comparison to the magnitude of selection, positive selection plays an insignificant

role in shaping the genome. Consequently, a minority of mutations are fixed under
the influence of positive selection. Organisms are well adapted in their environment
and many nonsynonymous mutations are deleterious to the organism. As a result,
deleterious nonsynonymous mutations are readily removed by the purifying selec-
tion. Thus, Kimura advocated that evolution is largely a stochastic process operating
on neutral mutations where substitutions are fixed by the random genetic drift.
Tomoko Ohta proposed some modifications in the neutral theory as the nearly
neutral theory of molecular evolution in 1973 and subsequently published two
seminal papers in 1990 and 1991. Ohta emphasized on those borderline mutations
which are neither strictly neutral nor are strongly selected and called them as nearly
neutral mutations. In contrast, Kimura focussed only on the strictly neutral
mutations. A mutation with selection coefficient (s) value much lower than the
reciprocal of the population size behaves as a neutral mutation (s ¼ 0) in a popula-
tion. She further proposed that only neutral mutations are fixed in the population
when s is less than or close to the reciprocal of the population size (s<<1/2Ne). On
the other hand, natural selection rejects the mutations in the functionally constrained
regions due to their detrimental effect on the fitness (s>>1/N). There is a negative
correlation between generation time and effective population size. Species with short
generation time are likely to have large effective population size. The mutations vary
with generation time and as a result, the differences in generation time and mutation
rates cancel each other in order to make mutation rate effectively neutral. Thus, rate
of molecular evolution appears to be proportional to absolute time rather than the
generation time.

6.3 The Concept of Homology

The term homology refers to evolutionary relationship between traits in various

organisms. Different kinds of traits such as structure and function of molecular
sequences, genomes, chromosomes, cells, organs and behaviour may be evaluated
for homology. The origin of all living beings on this earth can be traced back to a
single common ancestor. Genes are treated as homologous (i.e. descent from com-
mon ancestor) due to their sequence similarity. However, the overall similarity
between two genes gradually disappear as nucleotide or amino acid variation
accumulates in both genes with passage of time. Thus, adequate similarity to declare
them as homologous genes is retained in the sequence if they share a most recent
common ancestor. Genes are either completely homologous or completely
non-homologous. Thus, a statement that there is 70% homology between two
genes/proteins is incorrect; instead, a statement that there is 70% similarity between
two genes/proteins is correct. The higher similarity between two genes is an indica-
tor of their likelihood of being homologous. There are different forms of homology
based on evolutionary processes involved in the origin of a gene (Table 6.1;
Fig. 6.1). Homologous genes in each species evolve independently due to a specia-
tion event (cladogenesis) and are known as orthologous genes. For example,
90 6 Molecular Evolution

Table 6.1 Different forms of homology and evolutionary processes involved

Homology Evolutionary processes involved
Orthology Speciation (Cladogenesis)
Pro-orthology Speciation followed by gene duplication
Paralogy Gene duplication
Inparalogy Speciation and lineage-specific duplication
Outparalogy Gene duplication and speciation
Xenology Lateral gene transfer
Partial homology Exon shuffling or other recombination events leading to a composite gene

Fig. 6.1 Different forms of homology between genes based on speciation and duplication events

cytochrome C is an ortholog between human and chimpanzee. Moreover, the

homology between a singleton gene and its duplicate genes in another lineage is
known as pro-orthology. For instance, a homeotic gene AmphiOtx is
pro-orthologous to Otx1 and Otx2 in bony fishes and tetrapods. In reverse, Otx1
and Otx2 are semi-orthologous to AmphiOtx. In contrast, genes originating from
duplication event in a species are known as paralogous genes. Haemoglobin and
myoglobin genes are well-known examples of paralogous genes in a vertebrate
species. When gene duplication occurs after a speciation event, the paralogs are
known as inparalogs or symparalogs. In contrast, gene duplication before the
speciation event gives rise to outparalogs or alloparalogs. In some cases, homology
appears due to lateral gene transfer in bacteria and is known as xenology. The
homology between two genes may not be for the entire length of the gene but
restricted to a specific part of the gene. This partial homology may be due to exon
shuffling and other recombination processes resulting in generation of a composite
gene. Fibronectin and tissue plasminogen activator genes in human show partial
6.5 Nucleotide Substitution Models 91

homology due to exon shuffling. Thus, orthologous and paralogous genes should be
carefully selected for phylogenetic analysis in order to understand speciation and
duplication events, respectively. Two non-homologous enzymes may secondarily
acquire similarity at the active site during evolution due to similar functional
constraint which is often mistaken for homology. This evolutionary process is
known as convergent or parallel evolution.

6.4 Genetic Distances

The genetic or evolutionary distances between all pairs of aligned sequences are
computed in form of a distance matrix. A pair of sequences derived from a common
ancestor are likely to diverge in the course of time under the influence of various
evolutionary forces. This divergence between two sequences is measured by genetic
distance. Genetic distance is an indicator of similarity between sequences and
provides useful information for inference of a phylogenetic tree. The substitution
process is generally modelled as a stochastic or random event. One needs to specify
the appropriate statistical model of substitution prior to computing genetic distances
between two nucleotide or protein sequences. The simplest measure of genetic
distance (d) is the total number of nucleotide differences per site between two
sequences. This measure is also known as the p-distance or the observed distance.
However, it is not informative about the actual number of substitutions occurred in
case of high level of divergence between two sequences. Multiple nucleotide
substitutions at a particular site in an evolutionary timescale usually leads to the
saturation of such sites in the DNA sequences.
Another approach to measure expected genetic distances between a pair of
sequences is to apply a likelihood function L (d). This likelihood is defined as the
probability to observe two sequences given the distance d. The value of the distance
d that maximizes L (d) is known as the maximum likelihood estimate (MLE) of the
genetic distance.

6.5 Nucleotide Substitution Models

There is a variety of nucleotide substitution rate matrices from a very simple model
like Jukes–Cantor (JC69) model to many highly complex models. In JC69 model,
the equilibrium frequencies of four nucleotides are equal (25%) and each nucleotide
has the same probability to replace other one. In the general time reversible (GTR)
model, all eight free parameters of a reversible nucleotide rate matrix are clearly
stated. The free parameters are usually unknown and required to be estimated from
the data. Thus, it is always desirable to reduce the number of free parameters by
including some constraints in the underlying substitution process. The Tamura–Nei
(TN93) model considers only two independent rate parameters, the ratio (κ) between
transition and transversion rates and the ratio (γ) between two types of transition
rates. If we assume that the purine and pyrimidine rates are equal (γ ¼ 1), The
92 6 Molecular Evolution

Tamura–Nei model becomes a simpler model and the reduced model is known as
HKY85 model. This HKY85 model can be further reduced to the Kimura
two-parameter (K80) model assuming uniform nucleotide frequencies of 25% for
each nucleotide base. The HKY85 model is reduced to F81 model and K80 model is
reduced to the simplest model, Jukes and Cantor model in case the κ ¼1. It is always
advisable to choose a simple model like Jukes–Cantor model or Kimura
two-parameter model than a complex model like GTR model for phylogenetic
analysis of closely related nucleotide sequences. The best-fit evolutionary model
of a particular dataset is selected using different statistical approaches such as
hierarchical likelihood ratio test, Akaike Information Criterion (AIC) and Bayesian
Information Criterion (BIC). The likelihood ratio test (LRT) compares the log
likelihoods of two contrasting nested models to find the best fit and is defined by
LRT ¼ 2 (l1
l0), where l1 and l0 are the maximum likelihood values under complex
(parameter-rich) and simple models (less parameter rich), respectively. However,
AIC and BIC are applicable to both nested and non-nested models. The fully
automated model selection process to find best-fit model is implemented for nucleo-
tide sequence and protein sequence in programs jModelTest and protTest, respec-
tively. The protTest selects the best model of amino acid replacement. These
evolutionary models are based on the amino acid substitution matrices computed
from a large dataset of protein families. The common amino acid replacement
models used for protein sequence analysis are Jones–Taylor–Thornton (JTT),
Dayhoff, BLOCKS Substitution Matrices (BLOSUM62) and Whelan and Goldman
(WAG) matrices.
It is well known that nucleotide substitution rates vary greatly at different
positions in DNA sequences. For instance, the third codon position mutates faster
than the first and second codon positions. This rate heterogeneity at different codon
positions may influence the measurement of genetic distance. A gamma (Γ) distri-
bution model with variance 1/α considers varying degree of distribution of rates over
sites by adjusting the shape parameter α. The gamma distribution is bell shaped in
case the value of α is more than one and shows weak heterogeneity over sites. On the
other hand, the gamma distribution looks like L-shape if the value of α is less than
one and represents strong rate heterogeneity over the sites. Four to eight discrete
categories of gamma distribution are usually considered in phylogenetic analysis to
approximate the continuous gamma distribution.

6.6 Phylogenetic Analysis

The evolutionary relationship between organisms and their genes/proteins are

represented in form of a phylogenetic tree. A phylogenetic tree resembles the parts
of a tree such as root, branch, node and leaf (Fig. 6.2). The existing or extant taxa are
represented by the terminal leaves or nodes and commonly known as the operational
taxonomic unit (OTU). The internal nodes are known as hypothetical taxonomic
units (HTU) because these nodes are the hypothetical ancestors of OTU. A phylo-
genetic tree consists of two characteristic features, namely topology of a tree and its
6.7 Methods for Reconstruction of Phylogenetic Tree 93

Fig. 6.2 A rooted phylogenetic tree with its constituent parts

branch lengths. Topology of a phylogenetic tree is represented by its branching

pattern. There may be the presence of root (i.e. rooted) or absence of root
(i.e. unrooted) in a phylogenetic tree. An unrooted phylogenetic tree only provides
the details regarding the topology and the branch lengths. However, it lacks very
vital information about the evolutionary history of sequences under study. An
unrooted phylogenetic tree is usually transformed to a rooted phylogenetic adding
additional evolutionary information such as a reference sequence or an outgroup,
that is, the most distant species than the rest of the species (ingroup). A phylogenetic
tree is, in general, bifurcating in nature. But, multifurcating phylogenetic trees are
also reconstructed because of some statistical errors or presence of a zero-branch
length. A species tree describes the phylogenetic history of a species, whereas the
gene tree is generated from a particular gene sequence and represents the evolution-
ary history of that particular gene. Thus, the overall topology of an individual gene
tree is likely to differ from a species tree. The branch lengths of a phylogenetic tree
are represented as the number of substitutions per site. In a phylogram, the branch
lengths are scaled based on the evolutionary distances between taxa. On the other
hand, a chronogram demonstrates the complete evolutionary timescale from the
common ancestor to the extant species.

6.7 Methods for Reconstruction of Phylogenetic Tree

Various statistical methods are available to reconstruct a phylogenetic tree from

nucleotide and protein sequence data (Tables 6.2 and 6.3). Most popular methods
may be classified into four major categories: distance-based methods, maximum
parsimony methods, maximum likelihood methods and Bayesian inference methods.
94 6 Molecular Evolution

Table 6.2 Software for common evolutionary analyses

Method Name of the program URL
Phylogenetic analysis MEGA www.megasoftware.net
PHYLIP evolution.genetics.washington.edu/
phylip.html
DAMBE http://dambe.bio.uottawa.ca/
DAMBE/dambe.aspx
phyML http://www.atgc-montpellier.fr/
phyml
RaxML cme.h-its.org/exelixis/web/software/
raxml
MrBayes nbisweden.github.io/MrBayes/
download.html
Recombination RDP4 http://web.cbio.uct.ac.za/?darren/rdp.
detection html
GARD http://www.datamonkey.org/GARD
Positive selection PAML http://abacus.gene.ucl.ac.uk/
software/paml.html
HyPhy (BUSTED, www.hyphy.org
aBSREL, FADE)
Relaxed selection HyPhy (RELAX) www.hyphy.org
Substitution saturation DAMBE http://dambe.bio.uottawa.ca/
DAMBE/dambe.aspx
Synonymous constraint FRESCo A batch script for HyPhy
McDonald–Kreitman iMKT https://imkt.uab.cat
test
Positive selection in DnaSP http://www.ub.edu/dnasp
populations

Table 6.3 Popular R packages for phylogenetic analysis

Package Phylogenetic application
Ape A core package for phylogenetic analysis and evolution
Phangorn A package suitable for maximum likelihood analysis
Treedist A package to measure similarity between phylogenetic trees
Ggtree A package for visualization and annotation of phylogenetic trees
Treeman An R package for manipulation of phylogenetic trees
Phytools A package for phylogenetic comparative biology

6.7.1 Distance-Based Methods

The distance-based methods make an attempt to fit a phylogenetic tree to a matrix of

pairwise genetic distances computed from sequence data. Here, the number of
substitutions occurred truly in the sequences are estimated using a specific evolu-
tionary model best fitting the data. Distance-based methods are implemented in two
steps. First, the pairwise evolutionary distances are computed and stored in a matrix
6.7 Methods for Reconstruction of Phylogenetic Tree 95

of pairwise distances. The second step involves clustering algorithms to infer the tree
topology based on evolutionary distances. Clustering methods such as UPGMA
(unweighted-pair group method with arithmetic means) usually generate an
ultrametric tree assuming molecular clock. The ultrametric trees are the rooted
trees with all the end nodes having equal distance from the root. The tree is
constructed in sequential order by clustering those OTUs closely related to each
other followed by more distant OTUs. However, the UPGMA method is highly
sensitive to errors due to unequal rates of substitutions in different lineages. Since the
ordinary clustering methods have serious limitations, additive distance trees are the
better alternatives to ultrametric trees. Additive distance trees are unrooted trees in
which the genetic distance between a pair of OTUs is equal to sum of the lengths of
connecting branches. This kind of tree is always a better fit to the genetic distances in
the absence of clock-like behaviour in the sequence data. An additive tree can be
constructed using minimum evolution (ME) method that minimizes the length of the
tree which is the sum of the lengths of its branches. The length of the tree is
computed from a matrix of genetic distances. The main drawback of this method
is to search all possible topologies to find the minimum tree which appears to be an
impractical computational task with more than ten sequences. A better solution to
this computational challenge is to use a heuristic approach to find the ME tree using
clustering without assuming a clock-like behaviour. This improved method is known
as the neighbour joining (NJ) which is most popular distance-based method used in
phylogenetics. The distance-based method can be best adopted for phylogenetics
combining the NJ approach with bootstrapping. The reliability of an inferred tree
especially a specific clade of a tree is usually evaluated using bootstrap analysis. It is
very efficient statistical method to approximate underlying sampling distribution by
resampling from the original data. In phylogenetics, bootstrapping is implemented
on the original alignment by sampling of nucleotides sites with replacement,
reconstructing the phylogenetic tree and checking the presence of the original
nodes in all resampled set of phylogenetic trees (Fig. 6.3). This process is usually
repeated from 200 to 2000 times and the proportion of each node among all
bootstrap replicates is known as the bootstrap value. Bootstrap value indicates the
statistical confidence supporting a monophyletic clade and usually labelled on the
top of each clade. A bootstrap value of 70 or more demonstrates a good confidence
for a specific clade and the clades having lesser bootstrap values are treated with
caution and usually collapsed into a multifurcated phylogenetic tree. Jackknifing is
an alternative approach to bootstrapping in evaluating specific clades in a phyloge-
netic tree. Here, half of the sites are randomly purged from the original sequences to
generate numerous new samples. The proportion of clades in the resampled tree is
known as the jackknife value and 70% or more jackknife value of a clade is treated
with good confidence. The NJ method is very fast even for phylogenetic analysis of
hundreds of sequences and is an appropriate method for reconstructing large trees.
96 6 Molecular Evolution

Fig. 6.3 The steps involved in bootstrapping

6.7.2 Maximum Parsimony

It is one of the best and useful method for molecular phylogenetics based on discrete
character states and is not based on an explicit model of evolution. This method
searches for tree or a collection of a tree assuming minimum number of genetic
changes from a common ancestor to its descendants. This method is equally effective
in case the rate of evolution is either fast or slow. This method changes randomly the
topology of a tree in stepwise manner until a parsimony score is obtained which
cannot be improved further. The exhaustive search for the best tree is done by
evaluating all possible trees but this algorithm is feasible up to 11 sequences. An
alternative search method known as the branch-and-bound method is effective for
alignments containing 12 to 25 sequences. This method can be improved further
adding a heuristic method or neighbour joining. Heuristics is an approximate method
which can make an effort to find optimal solutions. First, an initial tree is generated
using a greedy algorithm and then subjected to several rounds of perturbations to
yield the most optimal tree. The stepwise addition is the most widely used greedy
algorithm to generate an initial tree but rarely yields a globally optimum tree. Thus,
three common methods of tree-rearrangement perturbations or branch swapping are
6.7 Methods for Reconstruction of Phylogenetic Tree 97

nearest-neighbour interchange (NNI), subtree pruning and regrafting (SPR) and tree
bisection and reconnection (TBR). These methods are effective for an alignment up
to 100 sequences. However, a hill climbing algorithm like branch swapping is likely
to be trapped in local minima.
The major advantage associated with this method is that it is free from unrealistic
model assumptions. The limitation of this method is high computational cost,
oversimplification of model of evolution and is likely to be inaccurate in case of
substantial evolution. Parsimony method takes into account only the informative
sites during phylogenetic analysis and does not consider information from the
non-informative sites. Sometimes, maximum parsimony method is likely to generate
an incorrect topology due to artificial clustering of long or short branches forming a
common clade. This phenomenon is known as long branch attraction and short
branch attraction, respectively.

6.7.3 Maximum Likelihood Methods

Likelihood is a statistical measure of goodness of fit of a specific model to some

datasets under given values of unknown parameters. In phylogenetics, maximum
likelihood finds for the best tree among a set of competing hypotheses such as
possible tree topologies, branch lengths, parameters describing the model of
sequence evolution, etc. The probability of the data is computed under assigned
values of the parameter and a decision is made regarding their likelihood. It is
expected that some hypotheses generate the observed data with a higher probability
than other alternatives. When maximum likelihood estimate (b θ ) is chosen, the
observed data are produced with the highest likelihood. However, likelihood func-
tion differs from probability, although it is defined in terms of probability. Likeli-
hood is the probability of the observed event instead of unknown parameters and is
not linked to the probability that the specified model is correct.
Since points mutations are considered as random events, the probability of
finding a point mutation along a lineage in a phylogenetic tree can be computed in
maximum likelihood framework. Here, we determine the tree topology, branch
lengths and parameters of the substitution model that maximizes the probability of
observing the sequence data under study. In fact, the likelihood L is the conditional
probability of a sequence data given a specific model of substitution with a set of
parameters θ and the tree τ with branch lengths:

L ðθ, τÞ ¼ Probability ðSequence alignmentj tree, evolutionary modelÞ

It is computationally impossible to test each tree from numerous possible trees

and estimate their model parameters. Therefore, different greedy heuristics using hill
climbing are commonly used to search for the maximum likelihood of a tree.
Stepwise addition is one of the early heuristics used and is available as DNAml in
PHYLIP package. Other popular heuristics commonly used are full-tree rearrange-
ment operations using three approaches: nearest-neighbour interchange (NNI), tree
98 6 Molecular Evolution

bisection and reconnection (TBR) and subtree pruning and regrafting (SPR). The
certainty of a maximum likelihood tree can be assessed by comparing maximum
likelihood values using likelihood ratio test (LRT). Similarly, statistical support to a
specific branch is assessed by either nonparametric bootstrapping or jackknifing.

6.7.4 Bayesian Phylogenetic Inference

The Bayesian phylogenetic inference is based on our prior beliefs about the topology
of a phylogenetic tree. In case the background information regarding the topology of
a tree is lacking, equal probability is assigned to each possible tree topology and such
prior probability is known as uninformative prior. We need a molecular sequence
alignment data and an evolutionary model in order to update this prior probability to
a posterior probability distribution. Bayes’ rule is implemented to compute the
posterior probability distribution. The posterior probability stipulates the probability
of an individual tree given the prior, model and the data. The most of the posterior
probability focusses on a single tree or few possible trees out of a large tree space in
case the sequence data is informative. It is an impossible task to compute posterior
probability analytically or drawing random sampling because posterior probability is
restricted to a very small part of the huge parameter space. Therefore, the posterior
probability distribution (Fig. 6.4) is usually estimated using the Markov Chain
Monte Carlo (MCMC) sampling. Markov chains have a tendency to converge
towards an equilibrium state irrespective of starting value. Thus, our focus is to set
up a Markov chain converging on our posterior probability distribution. All plausible
trees are sampled in order to find the range of a parameter, e.g. substitution rates in
MCMC analysis. The MCMC algorithm samples the entire distribution avoiding any
suboptimal solution. It proposes a new set of parameter values including tree
topology in every step (generation) and the newly proposed state is either accepted
in case better than previous state or rejected if worse than the previous state. The
present state is retained in case of rejection and the process is repeated. The
simulation process continues for a long duration changing one tree topology to
another topology. The early phase of the simulation run is known as burn in and
samples of this period is discarded to avoid the influence of starting value. The
posterior clade probabilities are used as a statistical support to a clade in Bayesian
phylogenetic inference in contrast to the bootstrap value used in other phylogenetic
methods.

6.8 Molecular Clock

The molecular clock hypothesis is based on the assumption that the rate of molecular
evolution is approximately uniform over evolutionary timescale. This hypothesis is
fully consistent with the arguments of neutral theory of evolution. Emile
Zuckerkandl and Linus Pauling proposed this concept in 1962 by comparing the
amino acid changes in the various protein sequences including haemoglobin in
6.8 Molecular Clock 99

Fig. 6.4 A multimodal posterior probability distribution showing one-dimensional representation

of three different phylogenetic tree topologies. The area under each topology is the posterior
probability of respective topology

different species with their age estimated from fossil records. Overall, there was a
linear trend in amino acid differences in respect to evolutionary time. Although there
was a constant rate of molecular evolution across the species, each protein had a
characteristic rate of molecular evolution (Fig. 6.5). The fibrinopeptides had a faster
rate of molecular evolution in comparison to extremely slow rates in cytochrome c
and histones. The differences in characteristic rates of evolution of these proteins can
be explained by the proportion of neutral sites in the protein sequences. The proteins
showing faster rate of evolution are likely to have a greater proportion of neutral
sites. The molecular clock ticks at irregular intervals and is probabilistic in nature.
The molecular clock estimates are prone to two types of errors. First type of error is
caused by the sloppy nature of the rate of substitution process leading to a huge
imprecision in the molecular date estimation. The variable mutation rate, population
size and proportion of sites under differential selection pressure are the underlying
causes of the second type of error. The presence of global molecular clock in a gene
is usually tested by relative rate test, Tajima test and the likelihood ratio test. Since
the global molecular clock is usually lacking in majority of genomic sequences, the
presence of local molecular clock may be tested in specific lineages allowing rate
variation across the lineages (relaxed clock). Molecular clock is useful in estimating
the date of origin of a viral disease in absence of any fossil record. The evolutionary
history of an epidemic disease is reconstructed from available extant virus samples.
The viral molecular clocks are usually calibrated using the stored viral samples.
100 6 Molecular Evolution

Fig. 6.5 The molecular clock of three different proteins ticking at variable rates

6.9 Ohno Hypothesis

The evolution of organismal complexity and the origin of novel gene function are
likely to be linked with gene duplication and whole genome duplication. Susumo
Ohno proposed his hypothesis that duplication of genes and whole genomes have
played an important role in the evolution of organismal complexity. He further
suggested that the duplication of a gene is followed by retention of original function
by one copy and adoption of new function by another copy. The paralogs that
emerged in the vertebrate genome through whole genome duplication are known
as Ohnologs. The expansion in the number of genes in genomes is caused by various
processes such as whole genome duplication, tandem gene duplication and segmen-
tal duplication. Consequently, large genomes allow more functional diversification
of genes and creation of gene families leading to formation of complex gene
networks in an organism. For example, amphioxus, ancestral cephalochordate line-
age, usually have small number of gene families even singleton in many cases. On
the other hand, advanced vertebrates such as mammals have three to four copies of
genes per family. The two-rounds of genome duplication (2R) hypothesis was
proposed as an explanatory model for the presence of large number of gene families
in vertebrates. This model suggests that early genome duplication occurred in the
genome of a cephalochordate-like ancestor (1R) followed by a second round of
genome duplication (2R) leading to the formation of four sets of genomes. This 2R
model is based on the empirical evidences obtained during genomic analysis of
6.10 Molecular Signatures of Selection 101

deuterostomes. Evolution of Hox clusters from one cluster in amphioxus to four

clusters in tetrapods is a very strong evidence in support of this model. In addition,
the synteny (i.e. same gene order) of gene clusters across evolutionary distant
vertebrates also supports this hypothesis. Interestingly, an additional fish-specific
whole genome duplication (3R) occurred after 2R duplication. For example, teleost
fishes have multiple copies of some genes and gene clusters, whereas tetrapods have
only one copy of the same in their genomes.

6.10 Molecular Signatures of Selection

Natural selection frequently acts on the genes and genomic sequences in an organism
and thereby causes some genomic variations that are likely to enhance the overall
fitness of individuals carrying these variations. These molecular footprints left on the
nucleotide and protein sequences by the natural selection in the past can be identified
using various statistical approaches. The relative rate of silent and replacement
fixations in the molecular sequences is an important information to infer the past
action of natural selection. A mutation in the gene, if confers a fitness benefit to the
species is likely to be under positive selection that includes different processes.
Positive selection is usually restricted to a small region of a gene and may point to a
functionally important region or sites of a selected gene. There are two primary
forms of positive selection: directional and diversifying selection. The action of
directional selection on a particular gene is manifested in form of concerted substi-
tution in the direction of a specific amino acid residue which is ultimately fixed in a
population after a long span of time (selective sweep). The development of a specific
mutation conferring drug resistance in HIV-1 against retroviral drugs is an example
of selective sweep. On the other hand, diversifying selection maintains an amino
acid diversity at a specific site in a population. This type of natural selection operates
on the codon positions of HIV-1 that are the targets of the human immune response.
There are two major approaches to identify natural selection on a gene. In order to
perform selection analysis on a certain gene, we need a multiple sequence alignment
of orthologs from different species or strains along with underlying phylogenetic tree
of these species or strains. In case of recombination, each non-recombinant segment
of the gene is represented by a separate phylogenetic tree. One needs to be very
careful during multiple sequence alignment to preserve all codons and avoid any
frameshift. There are two methods to identify positive selection on a gene based on
the predictions of Kimura’s neutral theory of evolution. The first approach deals with
comparison of nonsynonymous and synonymous changes in a gene. The synony-
mous (silent) changes are likely to be neutral and a greater rate of nonsynonymous
(amino acid replacement) changes relative to synonymous changes indicate the
action of positive selection. It is estimated as ω or dN/dS or Ka/Ks ratio for the
coding sequences. The ω is measured as the observed number of nonsynonymous
substitutions per nonsynonymous site (dN) over the synonymous substitutions per
synonymous site (dS) (Fig. 6.6). The ω is expected to be equal to one under the
assumption of neutral evolution due to the fact that the synonymous changes are
102 6 Molecular Evolution

Fig. 6.6 The directional shift in ω values away from or towards the neutral value (ω ¼ 1) indicates
the impact of natural selection on a gene. Positive selection operates in case of ω value more than
one, whereas purifying selection shift the ω value towards zero. Relaxed selection is identified in a
lineage by a shift of ω towards the neutral value in comparison to another related lineage

more or less equal to amino acid changes. The essential protein sequences allow only
minor replacement changes in the amino acid sequences and majority of amino acid
changes are eliminated by natural selection. This kind of natural selection is known
as the negative or purifying selection and estimated as an ω value significantly less
than one. An ω value close to zero indicates an exceptional selective constraint on
the protein sequence. However, amino acid replacements for some proteins are also
beneficial for the organism under selection. The genes encoding these proteins show
an ω value significantly more than one and the selection operating is inferred as
positive or directional selection. A high value of ω indicates the recurrent rounds of
selection on a gene rather than a single selective sweep. The genes under positive
selection are usually involved in male reproduction (e.g. human protamine gene),
host’s immune response to pathogen (e.g. human major histocompatibility complex
(MHC) locus) and neural developments in primates. The estimation of ω varies from
site to site and from branch to branch in a phylogenetic tree during a selective event.
The positive selection is estimated based on the variation of ω across the sites and/or
lineages. In site models, two maximum likelihood models, the first model (positive
selection) allowing a class of codons in the alignment to evolve with ω more than one
and the second model (neutral model) not allowing the same, are compared using
likelihood ratio tests. In case, two models are found to be significantly different
based on the chi-square distribution, the neutral model is rejected in favour of
positive selection model. In branch-site models, episodic positive selection is
detected on specific branches of a phylogenetic tree. Relaxed selection or relaxation
of positive or purifying selection is also identified on a specific lineage as a shifting
of ω value towards one in comparison to a closely related lineage.
When molecular sequences are sampled from different individuals in a popula-
tion, the ω value reflects the ratio of replacement to silent polymorphism in a
population. The ω value is likely to be one in the absence of positive and negative
selection and any significant statistical deviations in this value indicates the action of
6.11 Evolution of Protein Structure 103

natural selection. Natural selection might not have fixed the beneficial mutation or
removed the deleterious mutation in a population during recent evolutionary history.
Thus, the McDonald–Kreitman test, which compares the distribution of
nonsynonymous and synonymous polymorphisms in a particular lineage with the
ratio of nonsynonymous to synonymous differences fixed between lineages or
species, is widely used to detect recent positive selection. However, saturation of
substitution rates especially at the synonymous site for deep tree branches may affect
the estimation of selection in case of fast evolving sequences. However, branch-site
models are insensitive to the biases introduced by synonymous site saturation and
are suitable for analysis of distant species. In addition, there are certain indexes to
detect the saturation indexes and third codon position is often removed from analysis
in case of synonymous site saturation. Moreover, this test is extremely conservative
in the assessment of positive selection as it is an average estimate over the entire
gene. This limitation can be overcome using the entire phylogeny instead of
sequence pairs and additional computational cost. The purifying selection or positive
selection are relaxed on certain genes by shifting the ω values towards the value of
one (neutral value). The relaxed selection is determined based on whether intensity
of natural selection is relaxed or intensified in two subsets of branches of a phyloge-
netic tree. The selection pressure either significantly increases or decreases in a
foreground lineage with respect to a background lineage and is measured as a model
parameter K (relaxation parameter).
The second approach is based on the predictions made by the neutral theory on
allele or haplotype frequency in a particular population or between two populations.
For example, if positive selection has swept a mutation to high frequency in a certain
genomic region, it is expected that the targeted genomic region is likely to have low
sequence diversity, a surplus of rare alleles and a greater amount of linkage disequi-
librium than the expectation of neutral theory. Selection operating on a particular
population in comparison to other populations eventually leads to a greater degree of
population differentiation than expected in neutral evolution. Most common test
statistics are Tajima’s D, Wright’s FST and Fu’s W to measure these patterns in a
population. However, demographic history, population bottleneck and expanding
human population may generate spurious signatures of positive selection using this
approach.

6.11 Evolution of Protein Structure

The evolution of a protein structure can be explained through a sequence space. Each
possible amino acid sequence of a protein is treated as node and a single amino acid
mutation between each sequence is represented by an edge in a sequence space.
Since each node has a specific physical and functional feature, a genotype-phenotype
space is created having all possible protein sequences. The protein sequence follows
the edges in evolutionary trajectories through genotype-phenotype space. This
approach helps us in understanding the underlying evolutionary mechanism across
genotype-phenotype space in emergence of diversity and novel functions of
104 6 Molecular Evolution

proteins. These evolutionary trajectories of proteins can be traced using combined

computational and experimental approaches.
The reconstruction of an ancestral state of an extant protein or a group of proteins
is one of the powerful methods of choice. Firstly, the ancestral protein reconstruction
was performed at the internal node of a phylogenetic tree using maximum likelihood
approach. The maximum likelihood sequence at the internal node has a high
probability of generating present world protein sequences at the terminal ends of a
tree. Finally, the genes encoding the ancestral state of protein are synthesized and
expressed in vitro cultured cells. Thus, the structural and functional analysis of each
resurrected protein will allow us to understand the impact of historical mutations on
the biophysical property of a protein structure.
Another powerful strategy is the directed evolution of a protein towards a desired
function. Here, a library of random variants of a desired protein is generated
followed by screening those variants with the specific property. The screened
variants are subjected to repeated mutations to select the protein with the optimal
function. The underlying mutations present in each variant of a protein and their
impact on the overall evolution of a protein structure can be well understood using
this approach.
The third strategy is focussed deeply on a part of the sequence space without any
knowledge about the ancestral state. Here, random mutagenesis is performed on a
protein followed by weak selection on the mutants for a desirable property. Conse-
quently, the library of clones with desired property is enriched and the clones
without this property are depleted. Finally, we can quantitate the direct and epistatic
effects of each mutation based on the degree of enrichment of each clone after deep
sequencing. Thus, this approach will not only unveil the distribution of a desirable
sequence in the sequence space but also decipher the role of key evolutionary forces
in directing trajectories across the sequence space.
Most protein structures are found to be marginally stable. The stability of a
protein structure is described by the difference in free energy between its unfolded
and folded states. A majority of proteins are marginally above the energetic thresh-
old of unfolding and their stability can be improved further by simple point
mutations. It is assumed that marginal stability of a protein structure is a trade-off
between stability and function. However, it is possible to create an enzyme through
directed evolution which is hyperstable as well as hyperactive refuting this trade-off
hypothesis. Recent studies have shown that the marginal stability of a protein
structure is a product of neutral evolution through balance between mutation and
selection. In case, the hyperstability of a protein has neither beneficial nor deleterious
effect on the function, the natural selection will not be able to discriminate between
marginal and hyperstable ones. Thus, both mutation and random genetic drift are
likely to drive the proteins neutrally towards lower stable states compatible with their
function. Protein structures are maintained by some physicochemical constraints like
correct folding, solubility, maintaining a particular function and thermodynamic
stability.
Parallel evolution of phenotypes in different lineages occurs very frequently in
nature under similar selective constraints. Repeated occurrence of similar mutations
6.11 Evolution of Protein Structure 105

as a determinant of protein structure throws some light on the fundamental question

in evolution: Is molecular evolution repeatable, predictable and deterministic in
nature? Interestingly, parallel evolution occurs in independent proteins by
accumulating same mutations possibly following same order. Rat and mice across
different continents developed same mutations in the vitamin K epoxide reductase
complex when exposed to warfarin. The large effect parallel mutations often occur at
the key functional sites of the protein such as catalytic site of enzyme and binding
site of a ligand. Moreover, these mutations are accompanied by non-parallel second-
ary mutations with small effect located far from the functional site. The non-parallel
nature of secondary mutations indicates a weaker selective constraint on these
mutations.
The emergence of unique physical and biological properties in a protein such as
thermodynamic stability and substrate specificity is a result of interactions among
amino acid residues. Single permissive mutation does not have any impact on the
protein function per se but need one or more mutations for altered function in the
protein. Thus, it allows a mutation to change the function of a protein by adopting
contingency in the evolutionary process. A single mutation (H274Y) in the neur-
aminidase of influenza virus, although, confers resistance against the antiviral drug
oseltamivir but adversely affects the viral fitness. Two permissive mutations shortly
before the appearance of deleterious mutation counteract the adverse effect of this
mutation. Permissive mutations also act as a buffer against the deleterious function-
switching mutations. Usually, these new functional variants changing the protein
function are removed frequently by the purifying selection. These permissive
mutations are also observed in the evolution of new enzyme functions using directed
evolution.
Permissive mutations operate using two physical mechanisms: non-specific epis-
tasis and specific epistasis. The improvement of influenza virus fitness is an example
of non-specific epistasis. The function-switching mutations are found to adversely
affect the stability of a protein. But, permissive stabilizing mutations improve the
stability of protein and shift the evolutionary trajectory towards new function. For
example, the active site of beta lactamase is distended in bacteria to degrade
cephalosporin antibiotics. Although, the expansion of active site increases the
activity of enzyme but decreases its stability leading to weak resistance. Moreover,
a high resistance variant of enzyme is developed by a compensatory mutation far
away from active site. The second class of permissive mutations is based on specific
epistasis where two and more mutations collaborate to produce new functional
protein sometimes through conformational changes. For example, glucocorticoid
receptor has undergone two historical mutations to develop specificity to glucocorti-
coid hormones in its evolutionary history. First mutation introduced a hydrogen
bond on the helix in the ancestral structure with no change in the function. Subse-
quently, second mutation further shifted the position of the first mutated residue for
the formation of novel ligand-specific hydrogen bond. Thus, combined effect of both
mutations is essential for adoption of new function in the glucocorticoid receptor.
Both chance events and determinism play important role in the evolution of
protein structure. Most proteins show parallel evolutionary trend accumulating
106 6 Molecular Evolution

same mutations under similar selective pressure. Thus, evolution of a protein

structure appears to be predictable and deterministic. However, some mutations
which do not change the protein function are not selected by the natural selection.
Moreover, these mutations in future are exposed to chance events like genetic drift
and ultimately shape the evolutionary trajectory. Overall, the evolution of a protein
seems to be unpredictable and unrepeatable as well.

6.12 Evolution of Viruses

The recent global pandemic of coronavirus disease 2019 (COVID-19) has reinforced
our interest in understanding the mechanisms involved in initiation and exponential
spread of viral epidemics in human population. The evolutionary properties of
viruses involved in invading a new host species and evading new vaccines are
largely unknown. The emergence of novel viruses is a microevolutionary process
often involving cross-species transmission. Natural selection is a powerful force in
virus evolution and operates on these viruses during replication and production of
progeny in the host cell. The rates of mutation and subsequent fixation of mutation in
the population (i.e. substitution) of viruses are many folds higher than the cellular
genes of the host species. The sexual reproduction in viruses occurs through two
distinct mechanisms: recombination and reassortment. Recombination process in
viruses involves coinfection of a single host cell by two viruses and generation of a
hybrid molecule during replication. On the other hand, two or more segmented
viruses sometimes infect a single cell and the resulting progeny virus has segments
from multiple ancestries. Although recombination in a virus often leads to a new
beneficial genetic combination, mutation plays a crucial role in adaptive evolution in
viruses. For example, the resistance to antiviral drugs in human immunodeficiency
virus (HIV) can be attributed to mutation alone rather than recombination. The intra-
host population size of viruses is very large considering both the huge number of
viruses per host cell and the large number of infected host cells. For instance, HIV-1
can infect 107 to 108 cells in a patient producing 1010 virions in a single day. Viruses
lack functionless regions such as introns and pseudogenes in their noncoding
genome suggesting no significant role of genetic drift in viral evolution.
RNA viruses are fast evolving pathogens and have ability to accumulate rich
genetic diversity in a short period of time. The underlying evolutionary pattern can
be understood using phylogenetic analysis of homologous sequences from diverse
types of extant viral species or strains. Since recombination is a common phenome-
non in viruses, multiple phylogenetic trees describe the complete model of a viral
evolution, one tree representing each non-recombinant fragment of homologous
sequences. The failure to account for recombination may not only lead to a distorted
phylogenetic analysis but produce an incorrect estimation of natural selection. The
recombination breakpoints can be detected using GARD program in a sequence
alignment based on the phylogenetic incongruence among segments. Alternatively,
RDP is a popular program to identify breakpoints in a sequence alignment using
different methodologies. Natural selection acts on the RNA viruses with great
6.12 Evolution of Viruses 107

efficiency due to their extremely large and well-mixed populations. Thus, positive
selection is a very common mode of adaptive evolution in RNA viruses and the host
genes responding to viral infections as well. The current methods of the dN/dS
analysis need repeated fixation of nonsynonymous changes at a particular site in
order to infer a positive selection. This type of positive selection is known as
diversifying selection. In contrast, a single amino acid change in a specific lineage
is known as directional selection which is hard to detect using dN/dS method.
Directional selection is a rampant adaptive process in the emergence of new viruses.
Most mutations especially at the nonsynonymous sites in RNA viruses like HIV are
deleterious and consequently removed by the purifying selection. Although synony-
mous sites are generally treated as selectively neutral, synonymous sites in RNA
viruses may not be selectively neutral. For instance, polymerase basic 2 (PB2)
protein encoding gene is highly conserved in influenza A due to its crucial role in
viral packaging. A major change in the nucleotide composition of a virus is expected
when a virus jumps from one species to another species. For example, influenza A
virus underwent a major change in the nucleotide composition during transmission
from bird to human. Nucleotide composition of viruses can vary widely in the same
families due to different hosts and vectors utilized during the transmission. There is a
prevalence of epistatic interactions although mostly antagonistic in viral evolution
may be due to very common use of overlapping reading frames.
Coronaviruses are zoonotic pathogens with a positive-sense and single-stranded
RNA genome having origin in wild animals. The genome size of coronaviruses is
extraordinarily large with size up to 32 kb. The genomic organization of all
coronaviruses are similar with two large overlapping open reading frames, ORF1a
and ORF1b coding for two polyproteins pp1a and pp1b, respectively. Further
processing of these polyproteins produces 16 non-structural proteins (nsp1-nsp16).
In addition, four types of structural proteins, namely spike, envelope, membrane and
nucleoproteins, are coded by some ORFs in the remaining part of the genome. Spike
proteins are key surface glycoproteins having very crucial role in interacting with
distinct host cell receptors. A wide number of accessory proteins are also encoded by
the genome of different viruses. Coronaviruses have lower mutation rates than other
RNA viruses due to some proofreading activity of 30 -to-50 exoribonuclease. How-
ever, this lower mutation rate in this virus is well compensated by a high rate of virus
replication in hosts. In addition, coronaviruses expand their genome through fre-
quent gene gains and losses and the genome expansion in the coronaviruses is
largely contributed by their exceptionally high replication fidelity. The genome
expansion has facilitated acquisition and maintenance of novel genes encoding a
large number of accessory proteins performing crucial functions such as enhanced
virulence, suppression of immune responses and adaptation in a specific host. Thus,
the gene losses and gains in coronaviruses are key evolutionary phenomena in the
emergence of novel viral phenotypes. Coronavirus genome codes for an enzyme
phosphodiesterase (PDE) which is responsible for blocking the interferon-induced
antiviral immune responses in the host and consequently increases the overall fitness
of viruses. Interestingly, coronaviruses frequently steal some additional genes from
their hosts like other viruses. For example, a viral enzyme, hemagglutinin esterase
108 6 Molecular Evolution

and the N-terminal domain of the spike protein are derived from cellular lectins of
the host. The structural components of the coronavirus genome have received much
evolutionary attention because these proteins are exposed on the viral surface to
induce the immune responses in the human host. The spike proteins of coronaviruses
can adapt to exploit diverse types of cellular receptors in hosts. This key adaptive
feature in coronaviruses might play an important role in frequent host jump. It is still
a mystery how these binding affinities of the spike proteins to cellular receptors have
evolved in coronaviruses. The ongoing pandemic of novel pneumonia widely known
as coronavirus disease 2019 (COVID-19) is caused by a new human
betacoronavirus, SARS-CoV-2, a close relative of SARS (severe acute respiratory
syndrome)-like viruses. There is about 79% similarity between SARS-CoV-2 and
SARS-CoV at the nucleotide level. However, this sequence similarity is little less of
about 72% in the spike protein. This virus has a zoonotic origin and appears to be
more infectious than any other coronavirus known till date. Most of the closely
related viruses to SARS-CoV-2 are found in bats as bats are known to be a reservoir
for a wide variety of coronaviruses. The genome of this novel virus has approxi-
mately 96% similarity with the genome of a related virus (RaTG13) from horseshoe
bat. In spite of very high similarity at the nucleotide level, a number of key genomic
features differ between two viruses. For example, SARS-CoV-2 has a polybasic
S1/S2 cleavage site insertion which might account for its high infectivity and
pathogenicity in human. Interestingly, the receptor binding domain (RBD) of these
two viruses has approximately 85% similarity and only one amino acid is common
out of six critical amino acid residues at this domain. It appears that the receptor
binding domain of SARS-CoV-2 is well optimized to bind the angiotensin
converting enzyme 2 (ACE2) receptor in human like other SARS-CoV. It is possible
that SARS-CoV-2 might have acquired the key mutations needed for efficient
human transmission in an intermediate host like pangolins. For example, the pango-
lin viruses have not only amino acid sequence similarity of 97% with SARS-CoV-
2 but contain all human-specific six mutations at the receptor binding domain
optimized for binding ACE2 receptor. However, another possibility of acquiring
some of its key mutations during a period of cryptic spread in human before its first
detection cannot be excluded. There are different variants of this virus across the
world (Fig. 6.7) and some amino acid sites in the spike glycoprotein accumulated not
only synonymous and missense mutations but also gained few stop mutations
(Fig. 6.8). We can study the genetic diversity in viruses using distinct phylogenetic
clusters of genomic sequences. However, it is difficult to determine the phenotypi-
cally important mutations using genomic sequences only; it can only be validated
using clinical samples in laboratories.

6.13 Exercises

1. GnRH is a neuropeptide present in the hypothalamus regulating reproductive

functions. Retrieve nine orthologous coding sequences of GnRH1 from eight
primates and tree shrew using Ensembl database. Perform the multiple sequence
6.13 Exercises 109

Fig. 6.7 A view of sequence alignment of 11 strains of SARS-CoV-2 on the Ensembl COVID-19
database showing variability at some nucleotide sites. The nucleotides highlighted in green and
yellow colour are synonymous and missense variants

Fig. 6.8 A view of various variants of SARS-CoV-2 spike glycoprotein on the Ensembl COVID-
19 database

alignment of these sequences, compute the pairwise distances between nucleotide

sequences, reconstruct an unrooted neighbour joining (NJ) tree and write the tree
in Newick format in the R environment. Using tree shrew an outgroup, convert
the unrooted tree into rooted tree. Plot both unrooted and rooted tree in using an R
package.

Solution (Figs. 6.9, 6.10, 6.11, 6.12, and 6.13)

We will use three R packages, namely “Biostrings”, “msa” and “ape”, for this
exercise. The R commands with their output are as follows:
110 6 Molecular Evolution

Fig. 6.9 An overview of GnRH coding sequences under study

Fig. 6.10 Multiple sequence alignment of GnRH coding sequences

Fig. 6.11 Pairwise distance between sequences

> library(Biostrings)
> GnRH<-readDNAStringSet("GnRH.fasta")
DNAStringSet object of length 9:

#Fig. 6.9
6.13 Exercises 111

Fig. 6.12 Unrooted NJ tree

Fig. 6.13 Rooted NJ tree

> library(msa)
> GnRH.alignment<-msa(GnRH)
> GnRH.alignment
CLUSTAL 2.1
Call:
msa(GnRH)
MsaDNAMultipleAlignment with 9 rows and 291 columns

#Fig. 6.10
112 6 Molecular Evolution

Fig. 6.14 Top ten best models for the GnRH coding sequences. The model (TPM1 + I) showing
lowest AICc value is the best model

Fig. 6.15 A maximum likelihood phylogenetic tree showing bootstrap values more than 50 at
respective nodes

> library(ape)
> GnRH<-as.DNAbin(GnRH.alignment)
> distance<-dist.dna(GnRH)
> distance

#Fig. 6.11
> tree<-nj(distance)
> write.tree(tree)
6.13 Exercises 113

[1] "(Gorilla:0.003763678232, Chimpanzee:0.003519046866,

(((((Lemur:0.1006666062, Tree_shrew:0.1045425764):0.04148715176,
Gibbon:0.01975429486):0.004614720588, (Monkey:0.000578976943,
Baboon:0.003057402722):0.0191423532):0.008601196208, Orangutan:
0.005888531322):0.01325500637, Human:0.001950267732):0.001699432124);"
> plot(tree)

#Fig. 6.12
> rooted<-root (tree, outgroup=9)
> plot(rooted)

#Fig. 6.13

2. Using same set of coding sequences as in Example 1, reconstruct a maximum

likelihood phylogenetic tree in the R environment in following steps:
(a) Compute a neighbour joining initial tree.
(b) Compute the best ten models fitting to the dataset based on AICc criteria
using model test.
(c) Optimize the initial tree using maximum likelihood method using bootstrapping.
(d) Plot the maximum likelihood tree with bootstrap values more than 50.

Solution (Figs. 6.14 and 6.15)

We will use four R packages, namely Biostrings, msa, ape and phangorn, for this
exercise.

> library(phangorn)
Loading required package: ape
> library(msa)
Loading required package: Biostrings
Loading required package: BiocGenerics
Loading required package: parallel

> GnRH<-readDNAStringSet("GnRH.fasta")
> GnRH.alignment<-msa(GnRH)
use default substitution matrix
> GnRH<-msaConvert(GnRH.alignment,type="phangorn::phyDat")
> distance.matrix <-dist.ml(GnRH, "F81")
> tree <- NJ (distance.matrix)
> GnRH.model <- modelTest(GnRH, tree=tree, multicore=F,model="all")
> GnRH.model[order(GnRH.model$AICc),]

#Fig. 6.14
> env <- attr(GnRH.model, "env")
> fitStart <- eval(get("TPM1+I", env), env)
> fit <- optim.pml(fitStart, rearrangement = "stochastic",
114 6 Molecular Evolution

+ optGamma=FALSE, optInv=TRUE, model="TPM1")

optimize edge weights: -812.2572 --> -812.2572
optimize rate matrix: -812.2572 --> -812.2572
optimize invariant sites: -812.2572 --> -812.2572
optimize edge weights: -812.2572 --> -812.2572
optimize topology: -812.2572 --> -812.2572
> bs =<-bootstrap.pml(fit, bs=100, optNni=TRUE, multicore=F)
> plotBS(midpoint(fit$tree), bs, p = 50, type="p")

#Fig. 6.15

6.14 Multiple Choice Questions

1. Which of the following statement(s) is/are true for genetic drift?

(a) It is a stochastic process
(b) It operates in a small population
(c) It is a dispersive process
(d) All the above
2. The slowest evolving protein molecule is:
(a) Haemoglobin
(b) Cytochrome C
(c) Histone IV
(d) Fibrinopeptide
3. Neutral theory of molecular evolution advocated that:
(a) Neutral mutations are fixed by genetic drift
(b) Nearly neutral mutations are fixed by genetic drift
(c) Neutral mutations are fixed by natural selection
(d) Nearly neutral mutations are fixed by natural selection
4. The genes subjected to positive selection are usually involved in:
(a) Male reproduction
(b) Immune response of host against pathogens
(c) Neural development in primates
(d) All the above
5. Most simple nucleotide substitution model is:
(a) Jukes–Cantor model
(b) GTR model
(c) HKY model
(d) Kimura two-parameter model.
6. The statistical test(s) to find best-fit evolutionary model is/are:
(a) Likelihood ratio test
(b) Akaike Information Criterion
(c) Bayesian Information Criterion
(d) All the above
7. The minimum bootstrap value indicating good confidence of a clade is:
(a) 50%
(b) 60%
6.14 Multiple Choice Questions 115

(c) 70%
(d) 80%
8. Which of the following is not true about maximum parsimony method?
(a) An assumption of evolutionary model is necessary
(b) It considers only informative sites
(c) It is prone to long branch attraction
(d) It is prone to short branch attraction
9. The confidence of a clade in a Bayesian inference tree is indicated by:
(a) Bootstrap
(b) Posterior probability
(c) Prior Probability
(d) None of the above
10. Discarding the early phase of MCMC simulation run in Bayesian inference is
known as:
(a) burn in
(b) burn out
(c) burning
(d) sampling Answer: 1. d 2. c 3. a 4. d 5. a 6. d 7. c 8. a 9. b 10. a

Summary
• The substitution rate is computed in terms of accumulation of genetic differences
between individuals in an evolutionary timescale.
• Nonsynonymous mutations are subjected to selective pressure and cause pheno-
typic changes.
• Synonymous mutations are neutral and are fixed in a population under the impact
of genetic drift.
• Neutral theory suggests that evolution is a stochastic process operating on neutral
mutations where substitutions are fixed by the random genetic drift.
• Nearly neutral theory emphasized on those borderline mutations which are
neither strictly neutral nor are strongly selected and called them as nearly neutral
mutations.
• The higher similarity between two genes is an indicator of their being homolo-
gous (having a common ancestry).
• The best-fit evolutionary model of a particular dataset is selected using different
statistical approaches such as hierarchical likelihood ratio test, Akaike Informa-
tion Criterion (AIC) and Bayesian Information Criterion (BIC).
• A phylogenetic tree consists of two characteristic features, namely topology of a
tree and its branch lengths.
• The reliability of a specific clade of a phylogenetic tree is evaluated using
bootstrap analysis.
• In Bayesian inference of phylogeny, posterior probability distribution is usually
estimated using the Markov Chain Monte Carlo (MCMC) sampling.
• The molecular clock ticks at irregular intervals and is probabilistic in nature.
116 6 Molecular Evolution

• The duplication of genes and whole genomes have played an important role in the
evolution of organismal complexity.
• The omega (dN/dS) is measured as the observed number of nonsynonymous
substitutions per nonsynonymous site (dN) over the synonymous substitutions
per synonymous site (dS).
• The omega is expected to be equal to one under the assumption of neutral
evolution.
• An omega value significantly less than one indicates negative or purifying
selection.
• An omega value significantly more than one indicates operation of positive or
directional selection.
• The marginal stability of a protein structure is a product of neutral evolution
through balance between mutation and selection.
• The spike proteins of coronaviruses can adapt to exploit diverse types of cellular
receptors in hosts.

Suggested Reading
Li W-H (1997) Molecular Evolution. Sinauer Associates Inc, Sunderland
Graur D, Li W-H (2000) Fundamentals of molecular evolution. Sinauer Associates Inc, Sunderland
Nei M, Kumar S (2000) Molecular evolution and Phylogenetics. Oxford University Press, Oxford
Bromham L (2016) An introduction to molecular evolution and Phylogenetics. Oxford University
Press, Oxford
Next-Generation Sequencing
7

Learning Objectives
You will be able to understand the following after reading this chapter:

• The principles of second-generation, third-generation and fourth-

generation sequencing platforms.
• Next-generation sequencing data formats.
• Visualization of next-generation sequencing data.
• Methods involved in reference-based assembly and de novo assembly of
short reads.
• Application of next-generation sequencing technology.

7.1 Introduction

Although Sanger sequencing technique was instrumental in completing the first

human genome sequence in 2004, a worldwide effort to develop a high-throughput,
cheaper and faster next-generation sequencing (NGS) technology started in the same
year. This initiative has led to the development of novel next-generation
technologies generating enormous number of short reads at an unprecedented
speed. The first NGS technology developed in 2005 was based on the
pyrosequencing method and is popularly known as the 454-genome sequencer.
This technology was soon followed by two new variants of the NGS: the Solexa/
Illumina sequencer and the SOLiD sequencer in 2007. Both of these new
technologies could generate large number of reads in comparison to 454 but initial
size of the reads generated were very short (35 bp). Next addition to the series of
NGS technology was the Ion Torrent technology in 2010 based on semiconductor
technology with a small instrument size at a lower cost. All these next-generation

# Springer Nature Singapore Pte Ltd. 2022 117

B. K. Tiwary, Bioinformatics and Computational Biology,
https://doi.org/10.1007/978-981-16-4241-8_7
118 7 Next-Generation Sequencing

technologies require prior amplification of the template DNA. The necessity of prior
DNA amplification was circumvented in the third-generation sequencing technology
known as Pacific Biosciences (PacBio) platform. Moreover, it allows the detection
of single molecules in real time and generates thousands of several kilo-base long
reads suitable for de novo assembly. The latest addition to sequencing technology,
Oxford nanopore is treated as the fourth-generation sequencing technology and is
capable of producing ultra-long reads at a cheaper cost. Overall, the second-
generation technologies have high throughput with a lower error rate and cost per
base. On the other hand, the third- and fourth-generation technologies offer long read
length with short running time. Thus, advantages of different generation of
technologies are complementary and a hybrid-sequencing approach having a mix-
ture of different generations of sequencing offers a better solution to whole-genome
sequencing. The rapid development of next-generation technology was
complemented by ever increasing computing power and development of efficient
algorithms for assembly of short reads.

7.2 Sequencing Platforms

Most popular second-generation sequencing platforms are 454, Illumina, SOLiD and
Ion Torrent. The PacBio and the Oxford nanopore technologies are the third-
generation and the fourth-generation technologies, respectively. A comparison of
different generations of sequencing platforms is briefly described in Table 7.1. The
principles of each sequencing method with its advantages and limitations are
described as follows:

7.2.1 454 (Pyrosequencing)

The preparation of sample starts with capturing each DNA single-stranded fragment
of the DNA libraries on an agarose bead using adaptor sequences. The fragment-
bead complex is secluded in water-in-oil micelles containing PCR reactants. These
micelles act as a microreactor for thermal cycling (emulsion PCR) of each DNA

Table 7.1 A comparative account of different generations of sequencing platforms

Read No. of reads Single pass Time per
Generation length (bp) per run error rate (%) run
First generation (Sanger ABI 600–1000 96 0.001 0.5–3 hours
3730 X 1)
Second generation (Illumina 2 X 250 8 X 109 0.1 7–60 hours
HiSeq 2500)
Third generation 10,000- 3.5–7.5 X 13 0.5–4 hours
(PacBio RS II) 20,000 104
Fourth generation (Oxford 10,000- 1.1–4.7 X 10 48 hours
Nanopore MinION) 20,000 104
7.2 Sequencing Platforms 119

fragment into about one million copies. Finally, microreactor is broken and each
28 μm DNA-coated bead containing one DNA fragment is fitted into a well of 44 μm
average diameter on a fibre-optic slide. The sequencing enzyme and primer are
loaded into the well. The synthesis of complementary DNA strand is performed by
passing one unlabelled nucleotide through the well at a time. The incorporation of a
nucleotide leads to release of pyrophosphate with a light emission by firefly enzyme
luciferase in the well. The amount of light produced by the luciferase is proportional
to the number of nucleotides incorporated by the DNA polymerase. The signal for
nucleotide incorporation is detected by an underlying charged couple device (CCD)
sensor in a real-time basis. Finally, 400,000 short reads (400–500 bp length) are
produced in parallel sequencing reactions using millions of wells. The raw reads are
processed by the 454-analysis software and produce about 100 Mb data consisting of
quality reads.

7.2.2 Illumina (Sequencing-by-Synthesis)

Denatured adapter-linked DNA libraries are attached at one end to the complemen-
tary oligonucleotides attached onto a flow cell. The flow cell is a closed glass
microfabricated device having eight channels. The open end of the DNA fragment
bends down and further hybridizes to a complementary adaptor attached on the solid
surface. The DNA strands bound to the flow cell undergo solid phase PCR to
produce clusters of clonal populations. This process is known as bridge amplifica-
tion. The solid phase amplification process undergoes several cycles and the resulted
amplified DNA is denatured to produce a cluster of around 1000 single-stranded
DNA molecules. The DNA polymerase along with primers and four reversible
terminator nucleotides labelled in different way are added for synthesis. The
incorporation of new nucleotides at 30 terminator base during synthesis are detected
by a CCD sensor based on its colour followed by enzymatic removal of fluorophore.
This cycle is repeated several times during the synthesis process. First machine using
this process was known as Genome analyser (GA) and it was capable of producing
paired-end reads with length up to 35 bp. Now, two new variants of the GA, HiSeq
and MiSeq are being used by various laboratories with a longer read length and
depth, although MiSeq machine has a lower throughput with a cheaper cost than the
HiSeq machine.

7.2.3 SOLiD (Sequencing by Ligation)

The sample preparation of the SOLiD platform is highly similar to the 454 platform.
The sequencing process in the SOLiD machine is performed by ligation using a
DNA ligase rather than synthesis using DNA polymerase in the Illumina platform.
First, the genomic DNA is sheared into 200 bp fragments and oligonucleotide
adaptors are added at the both ends of the fragment. Single-molecule templates are
attached to the magnetic beads and are amplified on the beads using emulsion PCR.
120 7 Next-Generation Sequencing

The beads containing one amplified DNA fragment are spread onto a glass slide
surface which are subsequently loaded into a flow cell. First, a sequencing oligonu-
cleotide primer is hybridized to an adaptor. The 50 end of the primer ligates with an
oligonucleotide hybridizing to the adjoining sequence. The ligation to the primer is
subjected to competition between a combination of octamers. The ligated octamer is
cleaved after detection of colour and this ligation cycle is repeated several times.

7.2.4 Ion Torrent (Semiconductor Sequencing)

The sample preparation and sequencing methodology are similar to the 454 sequenc-
ing. The sample preparation is based on the emulsion PCR, similar to sample
preparation in 454 machines. The incorporation of each nucleotide is detected by a
minute semiconductor pH sensor implanted in each well of the picotitre plate. The
formation of phosphodiester bond during nucleotide incorporation leads to release of
a proton (H+ ion) during the polymerization process in the Ion Torrent machine
instead of pyrophosphate in 454 platform. The difference in pH caused by proton
release is detected by a complementary metal-oxide-semiconductor (CMOS) sensor
technology.

7.2.5 PacBio

This third-generation technology has a potential to sequence single molecules with

longer read length without any need for prior DNA amplification. This single-
molecule real-time (SMRT) sequencing is the most widely used third-generation
sequencing technology. A target double-stranded DNA molecule is ligated with
hairpin adaptors at both ends to create a single-stranded closed circular DNA known
as SMRTbell. The SMRTbell samples are loaded into a SMRT cell consisting of
1,50,000-minute sequencing units known as zero-mode waveguides (ZMWs). A
single DNA polymerase is imbedded in the bottom of ZMW and further binds to
either adaptor of the SMRTbell. The DNA polymerization process takes place in the
presence of four fluorescent labelled nucleotides in the ZMW. Single fluorophore
molecules are visualized near the bottom of the ZMW. The incorporation of fluores-
cent molecule is monitored in real time followed by termination of signal after
cleavage of dye. The replication process occurring in all ZMWs of a SMRT cell is
recorded in form of light pulses that are interpreted as the continuous long read
(CLR). A single CLR may contain multiple passes of both strands in case DNA
polymerase activity lasts longer. This sequencing process is much faster requiring a
short running time up to several hours. This machine provides additional advantage
of producing extremely long reads of over 10 kb suitable for de novo assembly.
7.2 Sequencing Platforms 121

7.2.6 Oxford Nanopore

Biological nanopores are widely used for the single-molecule detection and single-
and double-stranded DNA sequencing. α-hemolysin is most common biological
nanopore suitable for single-stranded DNA sequencing. Nanopore sequencing
provides high-throughput label-free and ultra-long reads at a cheaper cost. Trans-
membrane nanometre-sized pores are usually embedded in a biological membrane or
in a solid-state film (Fig. 7.1). These membranes separate two compartments
containing conductive electrolytes. The electrolyte ions can pass through the
nanopores when subjected to a voltage difference and consequently generate signal
in form of ionic current. The ionic current signal is stopped when nanopores are
blocked by a negatively charged DNA molecule. The ionic current signals generated
during the sequencing are segmented into discrete events. A machine learning
algorithm either recurrent neural network (RNN) or hidden Markov model (HMM)
transforms these segmented events to consensus sequence of template and comple-
ment strands of the double-stranded DNA molecules. The Oxford nanopore
technologies have developed two nanopore sequencing platforms, namely GridION
and MinION. MinION is a small USB portable device producing cheaper and faster
ultra-long read (average length 10 kb) data without DNA amplification. The current
MinION flow cells consist of 2048 protein nanopores arranged in 512 channels
allowing parallel processing of 512 DNA molecules. It is useful in generating
bacterial genome sequences in a couple of days. Unfortunately, the sequencing
error rate of nanopore technology is very high in the range of 15–40% at present.

Fig. 7.1 Oxford nanopore sequencing technology is based on the changes in electrical pulses due
to blocking of ionic current in a nanopore clogged by a DNA strand
122 7 Next-Generation Sequencing

7.3 Data Formats

FASTA format is a text file with a single header line starting with the character “>”,
which is widely used to store the nucleic acid and protein sequences. On the other
hand, FASTQ format is a standard format for the NGS data with a single header line
starting with the character “@”. In addition, a Phred quality score for each nucleotide
base is also included in the FASTQ format. The Phred quality score assigns error
probability for each nucleotide base (Fig. 7.2). These error scores are the log
transformed value of the error probability and computed using the formula which
is as follows:

Phred Quality score ¼ 10 log 10 P, where P is the estimated error probability:

Thus, a Phred score of 20 and 30 denotes an error probability of 1/100 and

1/1000, respectively. The average Phred score is used as an overall quality metric for
a set of sequence. An average Phred score of 20 or more is treated as a good score for
quality sequences. Moreover, high-quality sequences usually have a Phred score
around 30.
SAM (Sequence Alignment/Map) format has a complex structure and is com-
monly used to store the reference sequence and the sequence reads aligned on the
reference sequence. The SAM file has 11 mandatory fields in different columns,
namely QNAME, FLAG, RNAME, POS, MAPQ, CIGAR, RNEXT, PNEXT,
TLEN, SEQ and QUAL appearing in same order. All the fields are present in the
SAM format even if some information is unavailable (indicated by “0” or “*”). Each
field is briefly described in Table 7.2. BAM format is a binary file with compressed
and indexed SAM data. The indexing is done for the BAM file in order to extract the

Phred quantiles
10% 25% 50% 75% 90%
50

40
Phred score

30

20

10

0
0 20 40 60 80 100
Sequence position

Fig. 7.2 The Phred quality score across each nucleotide base in the reads. Phred quantiles are
indicated at quantiles 10%, 25%, 50%, 75% and 90% using different lines
7.5 Alignment of NGS Data 123

Table 7.2 Description of various fields in the SAM format

Column
Field No. Brief description
QNAME 1 A string for query template name. Indicated by “*” if information is not
available.
FLAG 2 An integer for combination of bitwise flags
RNAME 3 Reference sequence name
POS 4 1-based leftmost position mapping of the first matching base. POS is
indicated as 0 for unmapped read.
MAPQ 5 Integer indicating mapping quality computed as
10 log10 Pr (mapping position is wrong)
CIGAR 6 CIGAR string
RNEXT 7 Reference name of the next/mate read
PNEXT 8 Position of the next/mate read
TLEN 9 Observed template length
SEQ 10 Segment sequence
QUAL 11 ASCII of Phred-scaled base quality+33

required information without uncompressing the whole file. SAMtools are used to
convert the SAM files to the BAM files and vice versa.

7.4 Visualization of NGS Data

The UCSC Genome browser is a common web-based genome browser for visuali-
zation of NGS data. It uses two compressed binary file formats: BigBed and BigWig
for transfer of voluminous NGS data. The GenomeStudio is a commercial software
developed by Illumina and is capable of reading and exporting all types of NGS data
formats. Mauve is a C package available for all kinds of platforms and is widely used
for visualization of multiple genome sequences. The Integrative Genomics Viewer
(IGV) is an open-source JAVA-based visualization software for common NGS data
formats such as FASTA, FASTQ, SAM and BAM formats. Another common
visualization package is the GenomeView for reading the BAM files with some
advanced visualization features, for example, zooming from chromosome level to
nucleotide level. The Generic Genome Browser (GBrowse) is a web-based genome
visualization tool written in Perl which can be customized based on the need of the
user. The preferable format for the GBrowse is GFF3.

7.5 Alignment of NGS Data

A single NGS machine is capable of producing millions of short reads. The align-
ment of short reads to a long reference genome sequence (reference-based assembly)
or alignment of a sequence read to another sequencing read (de novo assembly) is a
computationally demanding step in the overall NGS data analysis. The traditional
124 7 Next-Generation Sequencing

Table 7.3 Popular programs for next-generation sequencing data analysis

Program Description
FastQC A quality control JAVA-based tool for raw reads
seqTools An R package for analysing FASTQ files
Bowtie 2 A reference-based aligner with FM-index
BWA A reference-based aligner with FM-index
SOAP2 A reference-based aligner with FM-index
MUMmer A reference-based aligner with suffix tree
Newbler A de novo aligner based on OLC
Phusion A de novo aligner based on OLC
Euler A de novo aligner based on DBG
SOAPdenovo A de novo aligner based on DBG
DESeq An R package for differential gene expression analysis of RNA-seq data

methods of alignment such as Smith Waterman algorithm or BLAST have a high

computational cost in terms of processing and memory usage and therefore are not
suitable for alignment of short reads. There is a need for high-performance compu-
tational clusters with a high-speed storage system and large amounts of RAM for the
alignment process especially in de novo assembly. The popular programs of analysis
and assembly of NGS data analysis are listed in Table 7.3.

7.6 Reference-Based Assembly

The short reads are aligned to a long reference sequence using modifications of
existing alignment methods such as seed-and-extend method and hash table method.
In addition, new alignment method such as Burrows–Wheeler transform has been
introduced for alignment of short reads. These methods are briefly described as
follows:

7.6.1 Seed-and-Extend Method

The local alignment in the BLAST is based on the significant exact matches of the
k-mers or seeds known as hits and comparing the matches in a hash table. The
BLAST method is optimized for short read alignment by including spaced seeds in
order to improve sensitivity. A spaced seed contains few internal mismatches
especially at the third codon position for protein coding sequences. The alignment
method consists of three steps: finding the seeds, first performing gap-free extension
and rejecting the alignment with a low score and finally extension with gaps. If the
seed size is very large, it is very difficult to get large exact matches and consequently
the method has a low sensitivity. On the other hand, we get many random small
matches with a very small size of seed increasing the running time. Moreover, this
method is unsuitable for human genome with abundant repeat sequences due to lots
7.6 Reference-Based Assembly 125

of hits during alignment. The popular programs for alignment based on spaced seed
are Short Oligonucleotide Alignment Program (SOAP), Eland, MAQ, RMAP
and ZOOM.

7.6.2 Suffix Array and Suffix Tree-Based Alignment

Suffix array is a simple data structure having a list of alphabetically sorted starting
positions of the suffixes in the sequence. During this alignment, a substring along a
long reference sequence is searched using all suffixes starting with the substring. All
these suffixes are ordered alphabetically and resulted sorted suffixes are searched for
query substring using an efficient binary search. The reference string is represented
by a data structure known as trie (derived from the word retrieval) allowing faster
string matching. A trie is composed of all possible substrings representing a refer-
ence string. The terminator end of a substring is indicated by $ in suffix tries. The $
character is treated as smaller character than all other characters and is alphabetically
or lexicographically first to all other characters in the substring. The lexicographic
order will be $ < A < C < G < T for a DNA string. For instance, suffix arrays are
created for a DNA sequence using following steps:

DNA sequence: A T G A T A G C A $
Position: 0 1 2 3 4 5 6 7 8 9
Suffix arrays: 9 8 5 3 0 7 2 6 4 1 (ordered alphabetically)

Suffix tree is a more complex data structure related to suffix array. A suffix tree of
a string is a compressed trie for all suffixes of that string. A suffix tree is made from a
string in three successive steps. First, all suffixes of a string are generated and these
suffixes are treated as individual words. Finally, a compressed trie is built based on
the suffixes. However, suffix array is a better space-efficient data structure than
suffix tree requiring 15–20 bytes per character. On the other hand, five bytes are
sufficient to represent a character in suffix array. Overall, a suffix array can store a
3.15 billion base pairs human genome in 12 GB RAM. Segemehl and Vmatch are the
common alignment software based on the enhanced suffix arrays. MUMmer and
OASIS are popular software using suffix tree algorithm for alignment.

7.6.3 Burrows–Wheeler Transformation (BWT)-Based Alignment

The BWT of the reference sequence is a compressed version of suffix array for
compression and indexing leading to enhanced speed and reduction in the memory
footprint. It is closely related to the creation of a suffix arrays. For example, T is a
DNA sequence string having a length of m. We can make rotations of T by taking a
character from one end of string and sticking it on the other end repetitively. As a
result, a m x m matrix is created by assembling them vertically followed by sorting
them alphabetically into Burrows–Wheeler matrix of T. Finally, the last column of
126 7 Next-Generation Sequencing

Fig. 7.3 LF mapping during

Burrows–Wheeler
transformation

the BWM (T) is the Burrows–Wheeler transform of T: BWT (T). For example, a
DNA string ACAGCA undergoes Burrows–Wheeler transformation in following
three steps.
ACAGCA$ $ACAGCA
CAGCA$A A$ACAGC
AGCA$AC ACAGCA$
ACAGCA$ GCA$ACA AGCA$AC AC$CGAA
T CA$ACAG CA$ACAG BWT (T)
A$ACAGC CAGCA$A
$ACAGCA GCA$ACA
Rotations BWM (T)

The Burrows–Wheeler has an important property known as the LF (last first)

mapping. The LF mapping states that the rank of a character remains same in the first
column and the last column. Burrows–Wheeler transform of a string can be reversed
to reconstruct T using LF mapping (Fig. 7.3). Therefore, this transformation is useful
in both compression and decompression. FM-index is a data structure containing
only sampled positions which are small fractions of all string positions. The sampled
positions are directly retrieved using FM-index and subsequently all non-sampled
positions are obtained using LF mapping. FM-index needs small memory footprints
and can store a human genome with 3.15 billion base pairs in less than 3GB RAM.
7.7 De Novo Assembly 127

The common alignment software based on the FM-index are BWA, SOAP2 and
Bowtie.

7.7 De Novo Assembly

De novo assembly of short reads are performed without any prior information about
the genome. This kind of assembly is necessary for novel genomes and even for a
human genome with a large-scale genomic alteration induced by cancer. There are
two algorithms employed for de novo assembly of short reads: overlap-layout-
consensus (OLC) and de-bruijn-graph (DBG).

7.7.1 Overlap-Layout-Consensus (OLC)

The OLC algorithm is performed in three successive steps (Fig. 7.4). First, the
overlapping regions between short reads are identified explicitly by aligning each
read against the other read. An OLC read graph is constructed based on the layout of
all the reads along with overlap information. The read graph treats each read as a
node and two nodes are connected if overlap region is larger than a certain cut-off
length. Thus, the total number of nodes is equal to the number of reads. The layout
step involves determination of Hamiltonian path which is a NP-hard problem.
Finally, the reads are aligned along the assembly path and a consensus sequence is
inferred from multiple sequence alignments. This algorithm is suitable for short
reads with low sequencing depth or base coverage depth (i.e. total amount of
sequencing data) and is less sensitive to read errors and repeats in the genome.
The sequencing depth depends upon the size of the genome and a species with large
genome size needs more sequencing depth. For example, a human genome needs a
minimum sequencing depth of 22. Newbler, Arachne, Phrap and Phusion are most
widely used programs based on OLC algorithm. The OLC method is a better choice

Fig. 7.4 Overlap-layout-consensus (OLC) algorithm for de novo assembly of reads

128 7 Next-Generation Sequencing

than the DBG for assembly of long and error-prone reads generated using PacBio
and Oxford nanopore platforms due to its sensitive overlapping step.

7.7.2 de Bruijn-Graph (DBG)

This algorithm involves chopping of short reads into strings of specified length
known as the k-mers and concomitant detection of neighbouring relationship implic-
itly among the k-mers (Fig. 7.5). A k-mer graph consists of individual nodes
representing all unique k-mers and the edges representing exact overlap between
the k-mers in the genome. Both node numbers and edge numbers are expected to be
equal to the genome size and are independent of sequencing depths. However, the
number of DBG nodes are likely to be much higher than the overall genome size due
to inclusion of many false k-mers introduced by sequencing errors. The contig
sequence is assembled from the k-mer graph through the Eulerian path using linear
time algorithm which is an easier computational approach than finding Hamiltonian
path. There is no need to call the consensus sequence separately in the DBG
algorithm because the consensus information is already available with the k-mers.
This algorithm achieves a higher memory and CPU efficiency than the OLC
algorithm with a genome size with high coverage. Since the second-generation
sequencing platforms usually have a high sequencing depth (>30X) in order to
compensate for a short read length, the DBG is a better algorithm than the OLC for
large genome assembly using short reads. However, the DBG is more sensitive to
read errors and repeats unlike the OLC. Euler, Velvet and SOAPdenovo are common
alignment programs developed based on the DBG algorithm.

Fig. 7.5 De-Bruijn-graph (DBG) algorithm for de novo assembly of reads

7.9 Applications of Next-Generation Sequencing 129

7.8 Scaffolding

Short reads are assembled into large number of long and accurate contigs using both
OLC and DBG algorithms. Scaffolding is process of ordering and orienting contigs
similar to their position in the genome. Scaffold is a collection of contigs positioned
in a linear order and correct orientation. However, there may be some gaps between
contigs representing unknown sequences. This process is performed in four succes-
sive steps: contig orientation, contig ordering, contig distancing and gap closing. The
scaffolding problem is approximately solved using graph theory where each contig
is a node and linking read pairs are treated as edges. The read pairs are generated
during paired-end sequencing and also known as mates. Mates may overlap with
each other in the middle of the fragment based on their lengths. Some of the mates at
one end of a contig are likely to pair with the mates in another neighbouring contig
and are known as spanning pairs. The presence of spanning pairs indicates the
proximity of two contigs. The distance between the contigs is estimated from the
expected distance between the linking reads obtained from the fragment length
distribution. This distance is included in the scaffolding graph as lengths to the
edges. Finally, one scaffold per chromosome of the genome including gaps of
correct lengths between the contigs is obtained. Some of the most popular contem-
porary scaffolders are Bambus, Opera, ABySS, SGA and SOAPdenovo2.

7.9 Applications of Next-Generation Sequencing

There is a variety of applications of next-generation sequencing in genomics. Some

of the areas of applications are as follows:

7.9.1 Whole-Genome Sequencing

Whole-genome sequencing is useful for understanding the genome of novel

organisms. The short reads from a novel genome is assembled using de novo
assembly methods. The genetic variations present in the regulatory sequences and
non-coding regions have been implicated in genetic susceptibility to diseases. These
variations along with variations in the protein coding sequences are determined
using whole-genome sequencing. However, the whole-genome sequencing incurs
about ten-fold more cost in comparison to exome sequencing. The genetic variation
present in a species can be unveiled and measured using existing reference genome
of that species. This method helps us in understanding the single nucleotide poly-
morphism (SNP) and copy number variations (CNV) in relation to the biology of
genetic diseases and developing potential therapeutic drugs. The causative variants
of cancer, cardiovascular diseases and neurological diseases have been successfully
determined using whole-genome sequencing. There are several ongoing whole-
genome sequencing projects across the world such as Human Genome Project-
Write, 100,000 genome project and GenomeAsia 100 K (GA 100 K).
130 7 Next-Generation Sequencing

7.9.2 Transcriptomics

The transcriptome refers to complete set of all RNA molecules encoded by the active
genes present in a cell. Thus, the transcriptome of a cell is likely to change with
respect to its functional and temporal state. The transcriptome of a cell can be
analysed for gene expression studies and identify all upregulated and downregulated
genes under a particular condition. Although microarray technology which is based
on hybridization is routinely used for global gene expression studies on thousands of
target genes, NGS technology is technically superior to microarray technology
because it can detect novel transcripts and splice-events. Moreover, it is free from
unspecific hybridization. The RNA-seq is an RNA-specific next-generation sequenc-
ing protocol to analyse the RNA sequences and their expression levels without any
knowledge about the target genes. It can detect the accurate gene expression levels
quantitatively and identify tissue-specific transcript isoforms as well. The exact
location of exon boundaries and polymorphism in the sequences are also detected
during RNA-seq experiment. Complementary DNA is synthesized from RNA
extracted from the cell. RNA sequencing allows us to understand not only differen-
tial gene expression but also mutation, gene fusion and splicing of RNA. Moreover,
it provides us additional information regarding non-coding RNAs such as
microRNA (miRNA) and long non-coding RNA (ncRNA) linked to the pathogene-
sis of various diseases. The abundance of a transcript is measured in terms of read
counts or RPKM (reads per thousand nucleotides in transcripts per million reads) or
FPKM (fragment per thousand nucleotides per million map reads) or TPM
(Transcripts per million). The RPKM and FPKM are applicable for single-end and
paired-end RNA-seq, respectively. Thus, FPKM takes into account the fact that two
reads can map a single fragment. TPM measures the frequency of a transcript within
a sample.

7.9.3 Epigenomics

Epigenetics is concerned with DNA modification through methylation or

hydroxymethylation affecting both DNA and associated histone proteins in the
nucleosome. A high-resolution genome-wide epigenomic profile can be developed
using next-generation sequencing platforms. The genome-wide epigenetic
modifications between the healthy and the disease status are compared. The
N-terminal ends of histones are usually modified under environmental or develop-
mental signals and this post-translational genome-wide modification can be mapped
using chromatin immunoprecipitation (ChIP). The histone-modification profiling
includes cross-linking DNA to histones using formaldehyde, digesting the cross-
linked DNA followed by immunoprecipitation using appropriate antibodies raised
against a particular post-translational N-tail modification. Finally, the DNA
fragments associated with the histone proteins are purified and purified DNA is
sequenced. This method is known as chromatin immunoprecipitation sequencing
(ChIP-seq). Alternatively, chromatin immunoprecipitation-exonuclease (ChIP-exo)
7.10 Examples 131

method involves digestion of the precipitated DNA fragment at either end using
exonuclease. The most dominant sequencing platform used for ChIP-seq is Illumina
platform followed by SOLID platform. On the other hand, a wide range of methods
such as methylation-dependent enzymatic restriction, methyl DNA enrichment and
direct bisulphite conversion are used for profiling DNA methylation. Moreover,
PacBio and Oxford nanopore platforms can directly detect DNA methylation at the
level of single molecule.

7.9.4 Exome Sequencing

The whole-exome analysis is a cost-effective approach to determine the genome-

wide variations. The whole-exome sequencing involves the sequencing of all protein
encoding genes in the human genome in order to identify all disease-causing
variants. Only targeted regions of the genome are sequenced in exome sequencing
with a lesser cost than the whole-genome sequencing. The sequencing data is
processed in three successive steps: read quality control, read mapping and post
mapping control. The first step involves removal of all low-quality reads followed by
mapping of remaining reads to the reference genome. The post-mapping processing
includes measuring mapping statistics, merging BAM files, marking duplicates and
recalibration of base quality scores. The exome sequencing has been used success-
fully to identify many novel causal genes associated with various diseases. More-
over, exome sequencing can detect some variants often missed by the low-coverage
whole-genome sequencing.

7.10 Examples

1. Construct a suffix array for a DNA string “ATCGATAACGA”.

Solution
The suffixes of the DNA string are as follows:

[0] ATCGATAACGA
[1] TCGATAACGA
[2] CGATAACGA
[3] GATAACGA
[4] ATAACGA
[5] TAACGA
[6] AACGA
[7] ACGA
[8] CGA
[9] GA
[10] A
132 7 Next-Generation Sequencing

All suffixes of the DNA string are then ordered alphabetically:

[10] A
[6] AACGA
[7] ACGA
[4] ATAACGA
[0] ATCGATAACGA
[8] CGA
[2] CGATAACGA
[9] GA
[3] GATAACGA
[5] TAACGA
[1] TCGATAACGA

2. Construct a Burrows–Wheeler Transform of a DNA string “AACTGACAGCAT

AAGCAT”.

Solution
DNA string:
AACTGACAGCATAAGCAT$
Rotations of the string:
AACTGACAGCATAAGCAT$
ACTGACAGCATAAGCAT$A
CTGACAGCATAAGCAT$AA
TGACAGCATAAGCAT$AAC
GACAGCATAAGCAT$AACT
ACAGCATAAGCAT$AACTG
CAGCATAAGCAT$AACTGA
AGCATAAGCAT$AACTGAC
GCATAAGCAT$AACTGACA
CATAAGCAT$AACTGACAG
ATAAGCAT$AACTGACAGC
TAAGCAT$AACTGACAGCA
AAGCAT$AACTGACAGCAT
AGCAT$AACTGACAGCATA
GCAT$AACTGACAGCATAA
CAT$AACTGACAGCATAAG
AT$AACTGACAGCATAAGC
T$AACTGACAGCATAAGCA
$AACTGACAGCATAAGCAT

Sorted Rotations:
$AACTGACAGCATAAGCAT
AACTGACAGCATAAGCAT$
7.11 Multiple Choice Questions 133

AAGCAT$AACTGACAGCAT
ACAGCATAAGCAT$AACTG
ACTGACAGCATAAGCAT$A
AGCAT$AACTGACAGCATA
AGCATAAGCAT$AACTGAC
AT$AACTGACAGCATAAGC
ATAAGCAT$AACTGACAGC
CAGCATAAGCAT$AACTGA
CAT$AACTGACAGCATAAG
CATAAGCAT$AACTGACAG
CTGACAGCATAAGCAT$AA
GACAGCATAAGCAT$AACT
GCAT$AACTGACAGCATAA
GCATAAGCAT$AACTGACA
T$AACTGACAGCATAAGCA
TAAGCAT$AACTGACAGCA
TGACAGCATAAGCAT$AAC

Transform of the DNA string:

T$TGAACCCAGGATAAAAC

7.11 Multiple Choice Questions

1. The solid phase PCR in a flow cell is known as:

(a) Emulsion PCR
(b) RT-PCR
(c) Bridge amplification
(d) Multiplex PCR
2. In the PacBio platform, DNA polymerization occurs in:
(a) Flow cell
(b) ZMW
(c) Microreactor
(d) Well
3. The average length of reads produced by the MinION is:
(a) 1 kb
(b) 5 kb
(c) 10 kb
(d) 20 kb.
4. The Phred score of high-quality reads is 30 and it indicates an error probability
of:
(a) 1/100
(b) 1/1000
(c) 1/10000
(d) 1/100000.
134 7 Next-Generation Sequencing

5. A SAM file is converted to a BAM file and vice versa using:

(a) SAMtools
(b) BAMtools
(c) CONVERTtools
(d) None of the above
6. In a suffix trie, the terminator end of a DNA substring contains:
(a) $
(b) A
(c) C
(d) T
7. LF mapping is a property of:
(a) Suffix array
(b) Suffix tree
(c) Burrows–Wheeler transformation
(d) Spaced seed
8. The layout step of the OLC algorithm is based on:
(a) Eulerian path
(b) Hamiltonian path
(c) Shortest path
(d) Longest path
9. The k-mer graph in a DBG algorithm is assembled into contig using:
(a) Eulerian path
(b) Hamiltonian path
(c) Shortest path
(d) Longest path
10. The abundance of a transcript in RNA-seq data is measured using:
(a) RPKM
(b) FPKM
(c) TPM
(d) All the above

Answer: 1. c 2. b 3. c 4. b 5. a 6. a 7. c 8. b 9. a 10. d

Summary
• The PacBio technology has a potential to sequence single molecules with longer
read length without any need for prior DNA amplification.
• FASTQ format is a standard format for the next-generation sequencing data
including a Phred quality score for each nucleotide base.
• BAM format is a binary file with compressed and indexed SAM data.
• Suffix array is a simple data structure having a list of alphabetically sorted starting
positions of the suffixes in the sequence.
• The Burrows–Wheeler transformation (BWT) of the reference sequence is a
compressed version of suffix array for compression and indexing.
Suggested Reading 135

• FM-index is a data structure containing only sampled positions which are small
fractions of all string positions.
• The OLC method is a better choice than the DBG for assembly of long and error-
prone reads generated using PacBio and Oxford-nanopore platforms.
• The DBG is a better algorithm than the OLC for large genome assembly using
short reads.
• Scaffold is a collection of contigs positioned in a linear order and correct
orientation.
• RNA-seq detects the accurate gene expression levels quantitatively and identifies
tissue-specific transcript isoforms.
• The abundance of a transcript is measured in terms of read counts or RPKM or
FPKM or TPM.
• The whole exome sequencing involves the sequencing of all protein encoding
genes in the human genome.

Suggested Reading
Ye SQ (2016) Big data analysis for bioinformatics and biomedical discoveries. CRC Press, Boca
Raton
Brown SM (2013) Next generation DNA sequencing informatics. Cold Spring Harbor Laboratory
Press, Cold Spring Harbor
Janitz M (2008) Next generation genome sequencing: towards personalized medicine. Wiley-
Blackwell, Somerset
Masoudi-Nejad A, Narimani Z, Hosseinkhan N (2013) Next generation sequencing and sequence
assembly: methodologies and algorithms. Springer, New York
Systems Biology
8

Learning Objectives
You will be able to understand the following after reading this chapter:

• Definition of systems biology and complex biological systems.

• Description of various types of computational models.
• Common motifs present in biological systems.
• The mechanistic basis of robustness in biological systems.

8.1 Introduction

The molecular biology has provided us the finer details of the structure and function
of genes and proteins in a cell. As a result, we can explain the molecular basis of
many biological processes such as reproduction, development and initiation of a
disease. The whole-genome sequencing of human and other organisms has
contributed significantly in understanding the molecular basis of a biological process
with a better understanding of genes and proteins involved in the process. However,
this molecular level understanding is not sufficient for systems-level understanding
of the biological systems. Systems biology is an emerging approach to understand
the holistic biology at the systems level. Here, we study the components of a system
and their inter-relationship, behaviour of a system on temporal and spatial scales and
the properties controlling the behaviour of a system. There are four fundamental
steps in systems biology. First, we identify the components of a system and model
the system. Then, we perturb the system and monitor the changes in the behaviour of
the system. Finally, we refine and test the model in iterative manner under different
conditions. Thus, systems biology model provides new insights into the mechanistic
link between genotype and phenotype.

# Springer Nature Singapore Pte Ltd. 2022 137

B. K. Tiwary, Bioinformatics and Computational Biology,
https://doi.org/10.1007/978-981-16-4241-8_8
138 8 Systems Biology

8.2 Complex Biological System

Complex biological systems consist of simple components such as genes, proteins

and cells but exhibit complex dynamic behaviour at different scales of biological
organization. This property is known as the emergent property of a system and this
phenomenon is called as emergence. The components of a biological system are
highly heterogeneous. Although all organisms are made of cells, the cells are diverse
in structure and function across the organisms. Even, the behaviour of each cell
cannot be approximated as an average behaviour of whole system. Each component
of the system shows complexity in their structure and function. For example, each
protein has its own primary, secondary and tertiary structure with functional dynam-
ics in a cell. The interaction among the components in a complex system is highly
selective. This selective nature of interactions finally leads to a stable complex
macromolecular structure and maintenance of functional stability. Although evolu-
tion is a stochastic process, the design principles in a complex biological system are
not random. Natural selection has preferred some design principles that seem to be
optimal solutions under a given environment.

8.3 Computational Model

Computational model is a simple representation of a real complex system having

only essential aspects of the system. The computational models are generally defined
based on the question in mind. For example, how can we engineer a metabolic
pathway in a yeast cell to maximize the production of ethanol? For building a
computational model, model structure of a system is first specified in the form of
mathematical equations and the set of parameters under biologically feasible
variations are also identified. Large complex models are generally more accurate
due to a large number of parameters. However, many parameters are unknown in the
model and numerical simulation of large models is very hard. Therefore, a trade-off
between simplicity and accuracy needs to be maintained in computational
modelling. Thus, a good model only captures the fundamental features of a system
ignoring other irrelevant details. Overall, a good computational model should be
simple, accurate, reusable and capable to predict the behaviour of a system.

8.4 Network Models

We need to understand first the interactions between components of a system in

order to fully understand the overall behaviour of a complex system. A biological
network consists of several nodes (N ) represented by various genes/proteins/
metabolites in a system and connected by edges (L ) showing interaction between
two and more nodes. Highly connected nodes in a network are known as the hubs.
Transcription regulatory network, protein–protein interaction network, signalling
network, neural network and metabolic network are different types of biological
8.4 Network Models 139

Fig. 8.1 A bipartite network

showing connections between
two sets of nodes: diseases
and drugs

networks. The number of nodes and edges varies widely in these biological
networks. For example, there are 297 nodes (neurons) and 2345 edges (synapses)
in the neural network of small worm C. elegans. On other hand, a human brain is
composed of 1011 neurons and each neuron is connected to 7000 synapses on
average.
A biological network is mathematically represented by a graph. The interaction
between two nodes in some cases represented by a directed edge when there is a
particular direction in the interaction. All nodes are connected to each other in a
complete graph. However, the interaction may be two sided in many cases
represented by undirected edges between two nodes. The most important property
of a node in a network is its degree (k) (i.e. the number of links to other nodes). The
average degree (<k>) of a network is a global property reflecting the overall
connections between nodes. The overall connection in the network is mathematically
represented in the form of an adjacency matrix having two elements 0 and
1 indicating either absence or presence of an interaction. The adjacency matrix of
an undirected network is symmetric and is typically sparse in case of real networks.
Sometimes, a network consists of two disjoint sets U and V where node of a
particular set can connect to the node of other set. This kind of network is known
as bipartite network and best exemplified by a disease-drug network (Fig. 8.1). In a
network, the distance between two nodes is measured by the path length where a
path is the total number of links between two nodes. The geodesic path (d) is the
shortest path (i.e. path with minimum number of links) between two nodes. The
140 8 Systems Biology

diameter (dmax) is the longest geodesic path between two connected nodes in a
network. The average path length (<d>) of a network is the average of shortest paths
between all pairs of nodes.
Clustering coefficient is the measurement of well-connected clusters in the
network. The local clustering coefficient (C) is a measure of local density which
indicates the degree to which the neighbours of a particular node are connected to
each other. Moreover, the global density of the whole network is represented by the
average clustering coefficient (<C>) taking the average value of all nodes in the
network. The clustering coefficient of small degree nodes is significantly higher than
the clustering coefficient of the hubs. Thus, the local neighbourhood of a hub node
and a small degree node is sparse and dense, respectively.
The concept of centrality is based on detection of most important node in the
network. The degree of node is the simplest measure of centrality and commonly
known as degree centrality. The eigenvector centrality is high for nodes having
connections with other important nodes. Between centrality measures the frequent
position of a node between the paths of other nodes in the network. Closeness
centrality indicates fast communication of a node with other nodes. The graph
density is an indicator of number of connections in a subset of nodes. Most of the
biological networks are not dense but found to be sparse with graph density less than
0.1. Clique is a sub-graph where all nodes are connected to all other nodes and a
clique has a clustering coefficient of 1.
There are two network models proposed to represent the real properties of
biological networks: random network and scale-free network. A random network
G (N, p) is defined as the network with N nodes connecting with each other
randomly with probability p (Fig. 8.2). When we increase the value of p from 0 to
1, the network increasingly becomes denser with linear increase in average number
of edges. The random network model is also known as the Erdos and Renyi model in
honour of their valuable contributions to this theory. Most nodes in this model are
likely to have average degree and the degree distribution approximates Poisson
distribution in a random network. The distance between two random nodes in a
network is unexpectedly small. This is known as small-world phenomena. The
average distance between two random nodes <d> is directly proportional to log
N/ log <k>. Although real networks are not random, the random model is used as a
reference guide to explain the characteristic features of real networks.
The degree distribution of real biological networks is approximated by a power
law degree distribution which is as follows:

Pk
k γ

The exponent γ is known as the exponent of power law equation. The overall
behaviour of a scale-free network depends on the value of this exponent. The
networks with a power law degree distribution are known as scale-free networks
(Fig. 8.3). Power law distribution has a higher peak and fat tail with a characteristic
straight slope on a log-log scale (Fig. 8.4). The probability of finding a high degree
or hub node is many folds higher in a scale-free network than a random network. The
8.4 Network Models 141

Fig. 8.2 A random network consisting of 500 nodes based on Erdos–Renyi model showing
random connections between nodes

hubs have a tendency to grow polynomially and consequently occupy a large size in
large networks. The value of the exponent γ varies from 2–3 in a scale-free biological
network such as the protein interaction network. The scale-free networks are usually
ultra-small due to the presence of hubs connecting numerous small nodes. However,
the networks show the small-world property and look like a random network in case
the value of exponent γ exceeds 3. On the other hand, the value of exponent less than
1 in a scale-free network is not expected unless hubs have many self-loops and there
are multiple connections between same pair of nodes. Hubs in a biological network
prefer to connect to nodes with less degrees rather than connecting other hubs
showing disassortative nature of biological network. Biological network consists
of modules and each module is characterized by a group of physically or functionally
connected nodes performing a particular function.
The interaction between coexpressing genes may be inferred using quantitative
measures such as correlation coefficient and mutual information approach on global
gene expression data. The correlation coefficient measures linear relationship
between two variables, whereas mutual information is a measure of non-linear and
non-continuous dependency among two variables. Weighted correlation network
142 8 Systems Biology

Fig. 8.3 A scale-free network consisting of 500 nodes based on Barabasi–Albert model showing
hub nodes with many connections

analysis (WGCNA) is popular program for inferring correlation network among

highly correlated genes from gene expression data and to identify significant
modules in the network. Moreover, this method is useful in detecting correlation
between image data, genetic marker data, proteomics data and other high-throughput
data. Mutual information is an information theoretic approach which captures those
interactions which are non-linear in nature. When the mutual information between a
pair of nodes exceeds a given threshold, the nodes are connected by an edge. All
pairwise connections between nodes result in a coexpression network. There are
various algorithms such as CLR, ARACNE and MRNET based on mutual informa-
tion approach. All these algorithms are available in a package known as minet in the
R language and environment.
8.5 Genome-Scale Metabolic Model (GSMM) 143

Fig. 8.4 The degree distribution of scale-free network on a log-log scale showing characteristic
straight slope. k is the degree of a node in log scale and Pk represents probability of degree in a log
scale. Red straight line indicates the power law fit

8.5 Genome-Scale Metabolic Model (GSMM)

The metabolic function of a cell can be represented by a mathematical relationship

between metabolites, reactions, enzymes and the encoding genes. This is popularly
known as gene-protein-reaction (GPR) relationship. This kind of model evaluates
the complex interactions between genotype, phenotype and environment. The math-
ematical framework of GSMM is based on two fundamental assumptions. The first
assumption is the conservation of total charge and mass in the system (i.e. total
metabolites produced equal the total metabolites consumed). The second assumption
is that the system maintains steady state (i.e. the concentration of internal metabolites
does not change with respect to time). Flux balance analysis is the most common
method to estimate the flow of metabolites in a metabolic network. It helps in the
prediction of growth rate of an organism or rate of production of a useful metabolite.
All metabolites are balanced in a cellular system by a combination of enzymes. This
balanced and stable metabolite flux is known as flux mode. The values of metabolic
fluxes are determined in a typical constraint-based model. Sometimes, thermody-
namic constraints are also added to the model in addition to steady-state constraint to
reduce the solution space. A set of elementary modes (EM) are all possible routes
(pathways) in cellular metabolic network that convert some reactants to some
144 8 Systems Biology

Fig. 8.5 Reconstruction of a genome-scale metabolic model

products. The total number of EMs in the cellular network indicates the overall
redundancy and robustness to perform a certain function. Extreme pathways are the
subsets of elementary modes. When both internal and external reactions are irrevers-
ible in a metabolic network, a set of elementary modes becomes identical to a set of
extreme pathways. A GSMM of an organism is built in two major steps (Fig. 8.5).
The first step is automatic reconstruction of a gap-filled draft model based on the
genome annotation of a species. The second step is the manual refinement of the
draft model using available experimental biochemical data. The building of a
GSMM starts with downloading the genome sequence of a particular species.
Then, the genes are annotated using bidirectional BLAST using high matching
length of the query sequence (70%), moderate amino acid sequence identity (40%)
and a very low e-value (<1  1030) in the first model. In the second model, KEGG
Automatic Annotation server (KAAS) is used for functional annotation of all amino
acid query sequences. A draft model is developed based on both annotated models.
This draft model is further refined using biochemical information from public
databases such as KEGG (filling of missing genes), MetaCyc (direction of reaction),
CELLO (position subsystems) and TCDB (transport reaction). Subsequently, man-
ual corrections like gap filling, deletion of error reactions, a check for mass-charge
balance and addition of species-specific information are performed in sequential
phases. Further, a software environment like COBRA Toolbox or sybil package in R
simulates the SBML (System Biology Markup Language) model with flux balance
analysis (FBA) using a desired objective function such as growth rate or product
formation (Fig. 8.6). First, all metabolic reactions are mathematically represented in
the form of a stoichiometric matrix having stoichiometric coefficient of each reac-
tion. Each reaction is defined by a minimum allowable flux (lower bound) and
maximum allowable flux (upper bound). The second step is to define a phenotype
in form of an objective function. For example, if the objective function is biomass
production (i.e. the conversion of metabolites into biomass), an artificial biomass
reaction based on experimental measures is added as an extra column in the
stoichiometric matrix. As a result, we can predict the maximum growth rate of an
organism by computing the conditions allowing maximum flux through the biomass
reaction. Thus, objective function is a quantitative indicator of relative contribution
of each reaction to the phenotype. The optimization problem is solved in flux balance
analysis using linear programming where an objective function (Z ) is either
minimized or maximized. The objective function (Z ) is computed by multiplying
each reaction flux (vi) with known constant (ci) and then adding all the values. FBA
finds a solution for steady state (Sv ¼ 0) where each reaction is bound by an upper
8.5 Genome-Scale Metabolic Model (GSMM) 145

Fig. 8.6 The metabolic reconstruction has a list of stoichiometrically balanced biochemical
reactions. This reconstruction is converted into a stoichiometric matrix of size m x n, where each
row and column represent metabolite and reaction, respectively. Here, two additional reactions are
added into the matrix to represent growth (biomass reaction) as objective function and exchange
reaction of glucose inside and outside the cell. The VBiomass is the objective function to be
maximized during optimization. The flux through each reaction under steady state is SV ¼ 0
which is defined by a system of linear equations

bound value and a lower bound value. A solution space is defined by the set of all
possible points of an optimization problem satisfied by the mass balanced constraints
having an upper and a lower bound for each reaction. The optimization for a
biological objective function such as ATP utilization or biomass production finds
an optimal flux within the solution space (Fig. 8.7). The flux value obtained by the
FBA is not a unique optimal solution. A biological system can achieve same
objective value using alternative pathway. The identification of alternate pathways
using FBA is known as flux variability analysis (FVA) which measures maximum
and minimum possible fluxes through a particular reaction constrained by keeping
the objective function at its optimal value. Finally, the model is validated by
comparing in silico result with experimentally observed phenotype. All GSSM
models have some missing reactions and thus are incomplete due to our knowledge
gaps.
Since all enzymes are not active in each type of a cell or under a variable culture
condition, specific algorithms are necessary to simulate a particular condition in
silico. Sometimes biological, physical or chemical constraints are derived from
experimental data obtained under a particular condition to build a biological mean-
ingful model known as context-specific or condition-specific GSMM. Here, the
omics (i.e. transcriptomics, proteomics and metabolomics) data under a particular
condition is mapped onto the template framework of a general purpose model. There
are specific algorithms to reconstruct a cell or a strain specific model (Table 8.1).
146 8 Systems Biology

Fig. 8.7 The flux distribution of a metabolic network may lie at any point in an unconstrained
space. After defining constraints in the model, any flux distribution is feasible only within the
allowable space. Further, the objective function is optimized using linear programming to find a
single optimal flux distribution lying on the edge of the allowable space

Table 8.1 Algorithms used for context-specific genome-scale models

Method Description of the algorithm
MBA Any omics data can be used to identify high and medium confidence reactions.
Only high confidence reactions are retained in the model and medium confidence
reactions are included in certain conditions
mCADRE Transcriptomic data is used to identify core reactions and zero expression
reactions. All core reactions are retained in the model unless supported by a
number of zero expression reactions
GIMME Transcriptomic data is used to define low-expressed reactions. The low-expressed
reactions are used at minimal level keeping the objective function above a certain
level
iMAT High and low expression reactions are identified using any omics data. An
optimal trade-off needs to be obtained between including high expression
reactions and excluding low expression reactions
INIT Weights are defined using any omics data and an optimal trade-off is obtained
between including and excluding reactions based on their weights
FASTCORE Any omics data can be used to define one set of core reactions that are active in
the extracted model. It also finds the minimum number of supporting reactions to
the core reaction
Metabo Extracellular metabolome, transcriptome and proteome data can be integrated in
tools the model

Thus, we can generate a condition-specific GSMMs reflecting specific properties of

a specific cell like a cancer cell or a microbial strain producing a particular
metabolite.
8.7 Motifs 147

8.6 Kinetic Models

These models do not maintain steady state unlike GSMM and allows us to study the
system as a dynamic system. It requires knowledge about initial concentration of
metabolites and kinetic reaction coefficients. Therefore, this framework can capture
the metabolic state of a small system, whereas GSMM reflects the entire metabolic
profile of a cell. The overall size of a kinetic model is much smaller than a genome-
scale model. Here, the focus is to model few reactions in deeper mechanistic details.
For instance, we can develop a kinetic model describing the rate and affinities of
multiple reactions. A set of reactants participating in metabolic reactions in a cell is
defined by deterministic methods such as non-linear differential equations or partial
differential equations. However, some biological processes with few participating
molecules such as cell signalling and gene expression are better explained using
stochastic kinetic models rather than conventional deterministic models. The deter-
ministic equations describe the alterations in intracellular metabolite concentration
over a defined time period. The computational solutions obtained by kinetic
modelling helps us in understanding the dynamics of a biological process. However,
the biological information obtained is very limited in kinetic models due to small
number of reactions (often less than 20) included in the model. These models include
various reaction kinetic parameters like enzyme concentration (E0), turnover number
(Kcat) and Michaelis constant (Km), which are measured experimentally using bio-
chemical methods. In case the estimation of kinetic parameters is not possible,
computational methods such as parametric estimation methods and Monte Carlo
methods are implemented to obtain the value of the kinetic parameters. The kinetic
model can be integrated with the genome-scale models forming large kinetic
genome-scale models. Overall, a kinetic model often has a large number of
parameters. The kinetic models of metabolic reactions are constructed using various
computational approaches such as mechanistic modelling, thermokinetic modelling,
modular and convenience kinetic rate law modelling, ensemble modelling, optimi-
zation and risk analysis of complex living entities (ORACLE), structural kinetic
modelling (SKM) and mass action stoichiometric simulation (MASS). Most of these
approaches excluding mechanistic and ensemble models can be applied to develop a
large-scale genome model. However, both mechanistic and ensemble methods have
a potential to build a kinetic model with a maximum of 50 reactions. For example, a
whole cell kinetic erythrocyte model of many healthy individuals was developed
recently using the MASS approach. This kinetic model is useful in predicting
susceptibility to the hemolytic anaemia often induced by an antiviral drug.

8.7 Motifs

Motifs are recurring patterns found in significantly higher frequency in the real
network in comparison to an ensemble of randomized networks (Fig. 8.8).
Randomized networks have all the characteristic features of a real network such as
total number of nodes and edges except the connection between nodes made on a
148 8 Systems Biology

Fig. 8.8 Common network motifs in the biological networks: (a) Positive and negative
autoregulation (b) Coherent feed-forward loop (c) Incoherent feed-forward loop (d) Single-input
module (e) Multi-output feed-forward loop (f) Bi-fan and (g) Dense overlapping regulons (DOR)

random basis. The regulation of a gene by its own product is known as

autoregulation and about 59% of transcription factors in E. coli are autoregulated.
This is the simplest example of motif concerned with a single node. Presence of self-
edges on a node is an example of positive and negative autoregulation in a transcrip-
tional network. The negative autoregulation (auto-repression) is more frequent
network motif than the positive autoregulation (auto-activation). It not only speeds
up the response time of the gene circuits but also provides robustness to fluctuations
in the production rate of a gene. On the other hand, positive autoregulation is less
common network motif and constitutes about 10% of transcription factor in E. coli.
The positive autoregulation has an opposite effect to negative autoregulation by
slowing down the response time of a gene circuit which may be useful in stage-
specific production of gene products during prolonged developmental processes.
Moreover, the feed-forward loop (FFL), a very strong motif with three nodes is more
abundant in the transcriptional networks of E. coli, yeast and higher organisms. A
feed-forward loop consists of three transcription factors X, Y and Z. Z is regulated by
X directly and is also regulated indirectly by X through Y. The three edges of the
FFL may indicate regulation by activation (positive sign) or repression (negative
sign). The FFL may be classified into coherent and incoherent types based on the
positive and negative signs of the edges. Both direct path and indirect path to Z from
X has same signs in the coherent FFL, whereas sign of the direct path from X is
opposite to the sign of indirect path from X in incoherent FFL. Most abundant type
of FFL is coherent FFL in the biological networks, where all edges are positive
followed by incoherent FFL where direct positive regulation is opposed by indirect
negative regulation. Coherent FFL usually filters out slight fluctuations in the input
8.8 Robustness 149

signals and thereby stabilizes cellular gene expression under ever-changing stimuli.
On the other hand, incoherent FFL helps in the generation of pulse and speeds up the
response time. It seems that same FFL is rediscovered by convergent evolution again
and again in the biological networks of different organisms due to its vital cellular
function.
In addition, there are two larger families of motifs known as single-input module
(SIM) and dense overlapping regulons (DOR). In SIM network motif, a single
master transcription factor performs dynamical function by controlling several target
genes concomitantly. It regulates the temporal expression of genes in a defined order
based on their protein product requirement in the cell. In SIM circuit, the gene
activated first is deactivated in the last and this temporal order is known as the last-
in-first-out (LIFO). This order is observed experimentally in the arginine biosynthe-
sis of E. coli where individual genes are expressed at an interval of 5–10 min.
However, same activation and deactivation order of genes is desirable in some cases
and is known as the first-in-first-out (FIFO) order. The multi-output FFL can
generate the temporal FIFO order. For example, the multi-output FFL regulates
the expression of motor proteins of flagella in E. coli. Bi-fan is a four-node
overlapping pattern where two transcription factors X1 and X2 control jointly the
expression of two target genes Z1 and Z2. The dense overlapping regulons (DOR) is a
set of input transcription factors regulating through a dense overlapping wiring to
another set of target genes. It is a combinatorial decision-making device having
multiple input functions in order to regulate a series of target genes. There is a large
number of DORs regulating hundreds of genes in the transcriptional networks of
E. coli and yeast. The DORs are often composed of other common motifs like SIMs
and FFLs.
The neuronal networks of C. elegans have similar FFLs akin to transcriptional
networks although the first one has a much higher spatial scale than the latter one.
Another interesting similarity is the presence of multi-input FFLs in neuronal
network instead of multi-output FFLs in the transcriptional network.

8.8 Robustness

Robustness is the inherent property of a complex biological system which protects

an organism from environmental and genetic perturbations. The traits contributing to
the robustness of a system and thereby enhancing overall fitness in an organism are
favoured by major evolutionary forces like natural selection and genetic drift.
Robustness does not mean staying unchanged under the perturbations. Instead, it
allows certain changes in the structure and components of the system in order to
maintain the specific function of the system. There are four basic mechanisms
involved in the maintenance of robustness in the system: system control, modularity,
decoupling and alternative mechanisms. A system is controlled by both positive and
negative feedback mechanisms as a robust response to perturbations. The negative
feedback mechanism often plays a predominating role in enabling robust responses.
For example, bacterial chemotaxis uses negative feedback in response to a variety of
150 8 Systems Biology

external stimuli. Modularity plays an important role in containing perturbations

locally in order to minimize its impact on the whole system. For example, cell is a
physical module of a multicellular organism. Decoupling separates the low-level
variations such as mutations from its high-level expressions such as phenotypes. For
example, Hsp90 decouples mutations from the overall phenotype of an organism and
not only provides a rich genetic diversity but also buffers against harmful mutations.
An alternative or fail-safe mechanism provides several means to accomplish a
specific function. This mechanism includes redundancy, diversity and overlapping
functions. Redundancy is the presence of several identical or similar interchangeable
components in a system. If one component fails under a perturbation, the second one
takes over the function to achieve the functional integrity of the system. The
presence of glycolysis and oxidative phosphorylation in a cell metabolism is a
common example of redundancy. Glycolysis and oxidative phosphorylation are
well known for production of ATP under anaerobic and aerobic conditions, respec-
tively. Thus, a yeast cell undergoes drastic changes in the energy metabolism based
on the availability of either glucose or ethanol as a nutrient in culture medium.
A highly robust biological system is likely to have a point of fragility which turns
out to be the vulnerable point for initiation and progression of a disease in an
organism. Diabetes mellitus and cancer are two major complex diseases manifested
by failure of robustness in a biological system. Human being evolved under several
environmental constraints such as starvation, high energy demanding hunting life-
style and always exposed to ever-changing pathogen through ages. Thus, physiolog-
ical and biochemical system in human evolved in such a way to develop robustness
against these environmental perturbations. However, the drastic changes in the
lifestyle of contemporary human such as abundant high energy food and low energy
requiring lifestyle have paved the way for fragility of this robust system in form of
Diabetes mellitus. Similarly, cancer is a highly robust disease against anti-cancer
drugs and frequently relapse even after chemotherapy. It co-opts the intrinsic
mechanisms of robustness in the body such as genetic heterogeneity of tumour
cells and inherent responses against hypoxia. Thus, the mechanistic basis of
biological robustness can be better understood using a quantitative measure of the
robustness such as network robustness measure (NRM) and response ability measure
(RAM). The network of cancer cell exhibited better robustness than the healthy cell
in terms of NRM, whereas the response ability measure (RAM) was lower for the
cancer cell in respect to healthy cell.

8.9 Exercises

1. Scale-free networks are frequently found in biological world. Create a scale-free

network and its adjacency matrix with 25 nodes in the R environment and
measure degree, degree distribution, global clustering coefficient (transitivity)
and diameter of the network.
8.9 Exercises 151

Fig. 8.9 A scale-free network with 25 nodes. Node 2 is the hub node in the network

Solution (Figs. 8.9 and 8.10)

We will load R package igraph and run following scripts for construction of network
and analysis.

> scale.free <- barabasi.game(25, directed=F,algorithm ="psumtree",

start.graph = NULL)
> plot(scale.free, vertex.label.cex= 0.5, edge.arrow.size=0.5,vertex.
size = 10, layout=layout.fruchterman.reingold)

#Fig. 8.9
> scale.free.adjacency<-get.adjacency(scale.free)
> scale.free.adjacency
25 x 25 sparse Matrix of class "dgCMatrix"

#Fig. 8.10
> degree(scale.free)
[1] 1 12 3 2 1 1 2 1 1 1 3 3 1 1 1 2 3 2 1 1 1 1 1 1 1
> degree_distribution(scale.free)
[1] 0.00 0.64 0.16 0.16 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.04
152 8 Systems Biology

Fig. 8.10 The adjacency matrix of the network is a binary matrix (0, 1) where each entry in the
diagonal is zero. Each dot indicates zero in the matrix

#Degree distribution shows the relative frequency of all degrees such as

0,1, 2, 3,4 etc.
> transitivity(scale.free)
[1] 0
> diameter(scale.free)
[1] 7

2. An R package sybil is widely used in the constraint-based modelling of metabolic

networks. A core metabolic model of E. coli is available as a data file with this
package. Perform following tasks in the R environment using sybil package:
(a) Read the basic features of this model, total number of reactions and their
reactions IDs using sybil. Plot the stoichiometric matrix of this model in the
R environment.
(b) FBA estimates a single flux distribution in order to solve the optimization
problem. Find the total number of exchange reactions with their upper and
lower bounds and perform flux balance analysis of the model using the
objective function of biomass production.
8.9 Exercises 153

Fig. 8.11 A view of all reactants present in the model

(c) In silico gene knockout involves disabling the reactions catalysed by the gene
product through setting the upper and lower bound to zero. Find the effects of
single-gene knockouts on the objective function of biomass production.
(d) FVA estimates the range of feasible fluxes through each metabolic reaction.
Compute and plot flux variability in order to find the flux values for an
objective function setting at 80% of maximal biomass production.

Solution (Figs. 8.11, 8.12, 8.13, 8.14, and 8.15)

> library(sybil)
Loading required package: Matrix
Loading required package: lattice
> library(glpkAPI)
using GLPK version 4.47
> mp <- system.file(package = "sybil", "extdata")
> model <- readTSVmod(prefix = "Ec_core", fpath = mp, quoteChar = "\"")
reading model description, ... OK
reading metabolite list ... OK
parsing reaction list ... OK
GPR mapping ... OK
154 8 Systems Biology

Fig. 8.12 A stoichiometric matrix showing stoichiometric coefficient of each reaction (row) and
metabolite (column) pair

sub systems ... OK

prepare modelorg object ... OK
validating object ... OK
> model
model name: Ecoli_core_model
number of compartments 2
C_c
C_e
number of reactions: 95
number of metabolites: 72
number of unique genes: 137
objective function: +1 Biomass_Ecoli_core_w_GAM
> react_num(model)
[1] 95
> react_id(model)

#Fig. 8.11
> cg <- gray(0,8/8)
> image(S(model), col.regions = c(cg, rev(cg)))

#Fig. 8.12
8.9 Exercises 155

Fig. 8.13 Exchange reactions present in the model

>findExchReact(model)

#Fig. 8.13
> opt <- optimizeProb(mod, algorithm = "fba", retOptSol = TRUE)
> opt
solver: glpkAPI
method: simplex
algorithm: fba
number of variables: 95
number of constraints: 72
return value of solver: solution process was successful
solution status: solution is optimal
value of objective function (fba): 0.873922
value of objective function (model): 0.873922
> opt<-oneGeneDel(model)
156 8 Systems Biology

Fig. 8.14 Objective function of each reaction after single-gene deletion

compute affected fluxes ... OK

calculating 137 optimizations ...
> lp_obj(opt)

#Fig. 8.14
> opt <- fluxVar(mod, percentage = 80, verboseMode = 0)
calculating 190 optimizations ... OK
>summaryOptsol(opt, model)
Min. 1st Qu. Median Mean 3rd Qu. Max.
-35.34 0.00 0.00 18.82 10.13 1000.00
> plot(opt)

#Fig. 8.15
8.10 Multiple Choice Questions 157

Fig. 8.15 Flux variability analysis showing minimum and maximum flux values of each metabolic
reaction

8.10 Multiple Choice Questions

1. A complete graph consists of:

(a) Multiple nodes without any connection
(b) Multiple nodes with connection to each other
(c) Multiple nodes with few connections
(d) Multiple nodes with some connections
2. Which of the following is/are measure(s) of centrality in a network?
(a) degree
(b) eigenvector
(c) closeness
(d) All the above
3. The degree distribution of a random network follows:
(a) Poisson distribution
(b) Exponential distribution
(c) Power law distribution
(d) None of the above
158 8 Systems Biology

4. The range of values of exponent of a power law distribution in a scale-free

network is:
(a) 1–2
(b) 2–3
(c) 3–4
(d) 4–5
5. Which of the following statement is true regarding objective function?
(a) Relative contribution of each reaction to the phenotype
(b) A phenotype of a cell is designated as objective function
(c) Biomass production or ATP utilization may be an objective function
(d) All the above
6. The maximum and minimum fluxes through a reaction are measured using:
(a) Flux mode
(b) Elementary mode
(c) Extreme pathway
(d) Flux variability
7. An abundant motif with three nodes in living organism is:
(a) Autoregulation.
(b) Feed-forward loop
(c) Feed-backward loop
(d) Single-input module
8. The neuronal network of C. elegans consists of:
(a) Multi-output FFL
(b) Multi-input FFL
(c) LIFO
(d) FIFO
9. Linear programming during flux balance analysis involves:
(a) Linear constrained optimization
(b) Non-linear optimization
(c) Unconstrained optimization
(d) Non-linear constrained optimization
10. The optimum expression value of Lac Z protein in E. coli cell is:
(a) 40,000
(b) 50,000
(c) 60,000
(d) 80,000
Answer: 1. b 2. d 3. a 4. b 5. d 6. d 7. b 8. b 9. a 10. c

Box 8.1 Biochemical Reaction to Algebraic Equation

A metabolic system consists of series of biochemical reactions catalysed by
specific enzymes. Thus, computational modelling of these biochemical
reactions can only be performed after transforming them into mathematical

(continued)
8.10 Multiple Choice Questions 159

Box 8.1 (continued)

expressions. For example, we have a pool of metabolite, M with concentration
m in a biochemical system. The rate of change in m depends upon the
individual rates of reactions involved in either producing or consuming M.

V1 þ V2 þ V3 ! M ! V4 þ V5 þ V6

Thus, the rate of change in c at any time point may be written as

dm=dt ¼ V 1 þ V 2 þ V 3  V 4  V 5  V 6

where V1 is the net rate of the first reaction and V2 is the rate of the second
reaction, etc.
For example, we have another reaction 3X + Y ! 4Z.
Alternatively, this reaction may be written as 0 ¼ 3X  Y + 4Z.where 3,
1, and + 4 are stoichiometric coefficients. All reactants have negative signs
and all products have positive signs.
We can write a general equation using the symbol c for stoichiometric
coefficient:

0 ¼ cx X þ c y Y þ cz Z þ . . . :

Box 8.2 Stoichiometric Matrix

Stoichiometric matrix is a mathematical expression of metabolic reactions of
size m x n. Each row (m) and column (n) of this matrix represents one
metabolite and one reaction, respectively. Each element of the matrix is the
stoichiometric coefficient of the metabolite participating in a particular reac-
tion. The coefficient is negative for a metabolite consumed in a reaction and
positive for a metabolite produced in a reaction. The stoichiometric coefficient
is zero if a metabolite is not participating in a reaction. Stoichiometric matrix is
usually a sparse matrix because a few metabolites participate in majority of
biochemical reactions.

Box 8.3 Linear Programming

Linear programming is a form of constrained optimization where all mathe-
matical expressions are linear. The main elements of a linear constrained
optimization process are decision variables, objective function, constraints

(continued)
160 8 Systems Biology

Box 8.3 (continued)

and variable bounds. The values of the decision variables may not be available
initially but can be adjusted to find the best value of the objective function. The
objective function is the mathematical expression of a cost function combining
all the variables to be treated as an ultimate goal of the optimization process
such as biomass or ATP production. The objective function is either
minimized or maximized under certain number of constraints. Constraints
are mathematical expressions to represent limits in finding the possible
solutions. Each variable has an upper and lower bound during the optimization
process. When variables have only integer values, solving the linear program-
ming problem is more difficult and is known as integer programming.

Box 8.4 Optimal Design of Gene Circuits

The optimal expression of a protein in a cell is determined by a fitness function
that tends to maximize its value. This fitness function in a bacterial population
is reflected in the growth rate of cells. An E. coli cell produces about 60,000
copies of Lac Z protein which is utilized for breaking down the lactose, a
source of energy. These cells utilize this energy to maximize the growth rate.
The optimum expression value of Lac Z which is well defined in a particular
environment is obtained by the difference between the benefit function and the
cost function of producing a protein. Here, benefit is measured in terms of
relative increase in the growth rate of the bacteria due to action of Lac Z. On
the other hand, the cost function is measured experimentally as the relative
decrease in growth rate of bacteria due to expression of Lac Z protein in a
lactose-deficient environment induced by a chemical analogue of lactose
known as IPTG. High lactose level in the environment allows more production
of Lac Z enzyme in E. coli than the optimal level of 60,000 per cell because it
favours benefit over cost and thus supported by the selection pressure. On the
other hand, low level of lactose in the environment reduces the production of
enzyme less than the optimal level. Interestingly, the optimal level of enzyme
production becomes zero in the absence of lactose in the environment as it
incurs only cost without any benefit to the organism. Consequently, the
organism will lose the Lac Z gene after few generations in the prolonged
absence of lactose in the media.
8.10 Multiple Choice Questions 161

Box 8.5 LINCS

The Library of Integrated Network-based Cellular Signatures (LINCS) is a
NIH-supported systems-level program providing a reference library of
signatures related to drugs, genetic perturbation, antibody and disease
perturbed human cells. The response to perturbation is measured by various
omics technologies such as transcript profiling, cell imaging, mass spectrome-
try and biochemical methods. The LINCS Data Portal (LDP) is a unified
interface to access LINCS Dataset and associated metadata and it contains
more than 350 datasets along with 42,000 drugs and other small molecules and
1200 cells. The LDR is a relational database and the central data repository of
LINCS datasets and metadata. The main focus of this program is to understand
the holistic cellular physiology of various cells and tissues in connection to
health and common diseases such as cancer, neurodegenerative disorders and
cardiovascular diseases.

Box 8.6 Stochastic Gene Expression

The gene expression is a stochastic phenomenon due to random fluctuations in
transcription and translation in spite of a constant environment. These
fluctuations may arise due to both intrinsic and extrinsic noises in a biological
system. The molecular intrinsic noise is caused by random promoter activation
and deactivation, mRNA and protein production and their decay. Although the
stochastic gene expression appears to be detrimental to cellular physiology, it
provides potential benefits in form of phenotypic cellular diversity. Microbes
adapt to a fast changing environment or to a sudden stress due to presence of
heterogeneous distinct physiological states without any genetic mutations.
This adaptation may play a crucial role in the survival of few microbial cells
during antibiotic treatment. DNA methylation and alterations in chromatin
structure in eukaryotes can have a profound effect on stochastic transcriptional
activity in a particular cell or across a population of cells. The heterochromatin,
a dynamic structure, is an inactive chromatin state and also shows variable
transcriptional activity.

Box 8.7 Recon3D

Recon3D is the most comprehensive human metabolic network model
endowed with three-dimensional structural information of both metabolites
and proteins along with large-scale phenotype data. This computational model

(continued)
162 8 Systems Biology

Box 8.7 (continued)

allows us to characterize various mutations associated with a disease. More-
over, the metabolic response to certain drugs can also be identified using this
model. A total of 3288 open reading frames, 13,543 metabolic reactions, 4140
unique metabolites and 12,890 protein structures are included in this model.
This computational resource is useful in investigating the molecular basis of
human metabolism in health and diseases.

Summary
• Systems biology is an emerging discipline to understand the holistic biology at
the systems level.
• Complex biological systems exhibit complex dynamic behaviour at different
scales of biological organization.
• A good computational model only captures the fundamental features of a system
ignoring other irrelevant details.
• The biological networks with a power-law degree distribution are known as scale-
free networks.
• Flux-balance analysis estimates the flow of metabolites in a metabolic network
and predicts the growth rate of an organism or rate of production of a useful
metabolite.
• Motifs are recurring patterns found in higher frequency in the real network in
comparison to an ensemble of randomized networks.
• Coherent FFL is the most abundant type of feed-forward loop in the biological
networks.
• Robustness allows certain changes in the structure and components of the system
in order to maintain the specific function of the system.

Suggested Reading
Newman MEJ (2010) Networks: An introduction. Oxford University Press, Oxford
Kovert MW (2015) Fundamentals of Systems Biology: From synthetic circuits to whole cell model.
CRC Press, Boca Raton
Alon U (2019) An introduction to systems biology: Design principles of biological circuits, 2nd
edn. Chapman and Hall/CRC, Boca Raton
Marian Walhout AJ, Vidal M, Dekker J (2013) Handbook of Systems Biology. Academic Press,
Cambridge
Clinical Bioinformatics
9

Learning Objectives
You will be able to understand the following after reading this chapter:

• Role of bioinformatics in clinical research.

• Big data in medicine.
• Different approaches of next-generation sequencing to identify genomic
variations associated with a disease.
• Computational models of complex diseases.
• Computational discovery of network biomarkers.
• Discovery of multi-target drugs.
• Application of artificial intelligence in medicine.
• Genome-wide association mapping.

9.1 Introduction

The clinical bioinformatics is an integrative multidisciplinary approach to combine

natural sciences, computer science and medicine in such a way that datasets are
transformed into knowledge useful in clinical practices. Biological systems are
largely complex and non-linear in nature. A complex system consists of numerous
components such as genes, proteins and metabolites at each scale. For example,
genomic scale consists of thousands of genes and proteomic scale has numerous
protein molecules as components. These components in each scale usually interact
and function in a coordinated manner. Moreover, there is an interaction across
different scales in a multi-scale biological system. Thus, no single omics data like
genomics, transcriptomics or proteomics can provide a clear understanding of
regulatory networks involved in a disease aetiology. The common diseases like

# Springer Nature Singapore Pte Ltd. 2022 163

B. K. Tiwary, Bioinformatics and Computational Biology,
https://doi.org/10.1007/978-981-16-4241-8_9
164 9 Clinical Bioinformatics

cancer, diabetes and mental disorders have not only a genetic basis but also induced
by various epigenetic and environmental factors. The symptoms associated with a
certain disease is a result of perturbations in the molecular interaction networks in a
multi-scale system. Computational models have a property to reflect the overall
complexity associated with a disease. A comparative study of a disease model along
with a normal healthy model helps us in understanding the mechanistic details of
pathogenesis at a deeper level. These models are generally developed in form of
biological networks, genome-scale models and kinetic models.
The clinical bioinformatics has emerged as a powerful discipline due to availabil-
ity of big data. Most advanced technologies such as NGS, MRI and MS produce a
flood of big data which needs to processed and analysed computationally on a
regular basis. For example, metabolic profiling gives a holistic picture of all
metabolites present in a sample. There is a need for integrating all the data from
different sources in order to provide a unified understanding of the data. There are
four kinds of data types: structured data, unstructured data, semi-structured data and
time-series data. Structured data is stored in a relational database or spreadsheet and
accounts for almost 20% of an AI project. Majority of the data used in AI is in form
of unstructured data without any predefined formatting like text files and images.
However, some data is a hybrid of structured and unstructured data known as semi-
structured data such as XML and JSON formats and constitutes about 5–10% of data
used in AI project. The fourth type of data is a time-series data which is a combina-
tion of structured, unstructured and semi-structured data. Big data is a critical part of
an AI project. A big data is characterized by three V-features: volume, variety and
velocity. The volume of a big data is often in the scale of the terabytes and usually
unstructured. Moreover, big data is highly diverse consisting of mainly unstructured
data along with structured and semi-structured data. These data are created at an
extremely high velocity and processed in the memory instead of disks. Other Vs
associated with big data are veracity (accuracy of data), value (data from reliable
sources), variability (data changing over time) and visualization (graphical represen-
tation of the data).

9.2 NGS Applications in Clinical Research

The detection of genomic mutations or variations is one of the major applications of

NGS in clinical research. These genomic mutations are classified into several types
according to their sizes: single nucleotide polymorphism (SNP), insertion and
deletion (indel), copy number variation (CNV) and structural variation (SV). The
SNP is the smallest mutation in the genome involving only one nucleotide. The
detection of SNP in the NGS data is a simple process and involves alignment of test
reads to the reference genome. If most of the reads show a mismatch at the same
locus, it is likely that a SNP is present in the test genome. Indel is either an insertion
or deletion or a combination of both in the genome. The size of indel varies from one
nucleotide base to 1 kbp. The detection of indel in a genome is a more complicated
process. The short reads covering the indel region do not align with the reference
9.2 NGS Applications in Clinical Research 165

genome and therefore labelled as unmapped reads. These unmapped reads are
broken into two to three smaller parts which are finally aligned to the reference
sequence using Smith–Waterman algorithm in order to identify indels in the test
genome. Structural variations are either insertions or deletions having a size more
than 100 kbp. There are different kinds of structural variations in the human genome
known as deletion, duplication, insertion, inversion and translocation. Methods to
detect structural variations are based on paired-end mapping, split read, de novo
assembly of genome and read depth. Copy number variation (CNV) is a kind of
structural variation involving either a duplication or deletion of a genomic segment
more than 1 kbp size. There are many diseases known to be associated with the CNV
in human such as cancer, autism, schizophrenia and Alzheimer disease. Methods for
CNV detection are largely based on either significant increase or significant decrease
in read depth signal. The popular R packages for CNV detection are CNV-seq,
readDepth and SeqCBS.
Although whole-genome sequencing is most comprehensive method to identify
different kinds of genomic variations, targeted sequencing is the most cost-effective
method to detect Mendelian disorders and complex diseases. The first step in
targeted sequencing is known as target enrichment which is based either on
PCR-amplicons or hybridization capture. Targeted sequencing consists of two
forms: whole-exome sequencing and targeted deep sequencing. Targeted deep
sequencing is further classified into sequencing with multi-gene panel and designed
regions. Whole-exome sequencing (WES) covers only 2% of the genome encoding
protein sequences but about 85% of genomic variations associated with diseases are
localized in this region. It has successfully identified causal gene for Miller syn-
drome in a very small population consisting of only four individuals and mutations
for autism in a large population consisting of 2500 families. However, regulatory
regions of the genome located in introns cannot be explored using WES approach.
Targeted deep sequencing requires the prior knowledge of the disease and is a better
technology than WES in dealing with regulatory regions. Multi-gene panel sequenc-
ing is based on array hybridization and covers large number of candidate genes along
with introns. The number of genes varies between 70 and 377 in genes panels for
diseases such as muscular dystrophy, cardiomyopathy and epilepsy. Similarly, a
desired gene or a region can be captured on custom basis using PCR-amplicons or
hybridization. Targeted sequencing is often used as a validation step after WES
because an average coverage of 1200-fold can be achieved using this approach. The
size of NGS sample varies from small number of cases to a large population. A small
size of NGS sample is adequate for a common variant but a large population-based
study is required for rare variants of complex diseases. Family-based whole-genome
or whole-exome sequencing needs about 200 pedigrees for efficient detection of
Mendelian disorders. However, this family-based approach is not suitable for com-
plex diseases. Case-control deep sequencing needs 125 to 500 case-control pairs for
high detection power but is not sufficient for a novel gene discovery. However, a
case-control whole-genome or whole-exome sequencing using 500–1000 case-
control pairs is capable for discovery of novel genes. Most cost-efficient approach
166 9 Clinical Bioinformatics

is to perform case-only exome sequencing of 100–1000 subjects but identification of

new susceptibility loci is difficult using this approach.

9.3 Network Medicine

The network models are very useful in understanding the overall pathobiology of a
disease. The initiation and progression of a disease can be explicitly captured in form
of network module. Each disease is represented by a well-defined disease module.
The genes involved in a disease are likely to be connected in a disease module. This
module consists of an ensemble of directly connected genes performing a certain
biological function. Any disruption in this disease module may lead to manifestation
of disease. The neighbouring directly interacting gene of a disease-causing gene is
likely to play some role in the pathogenesis as well. These network models are
developed using the known interactions from literature, RNA-Seq, microarray and
yeast two-hybrid data. If the candidate genes of a disease are known, their inclusion
in the model makes the modelling process more accurate and effective. There are
four types of statistical methods employed in reconstruction of a gene network from
microarray gene expression data. These are probabilistic network-based methods
such as Bayesian networks, correlation-based methods, partial correlation-based
methods and information theory-based methods.
We can identify a disease module including the neighbouring genes of a candi-
date gene of a network. The average degree of disease genes is higher than the
average degree of control genes. Moreover, disease genes involved in a certain
disease are likely to interact with the disease genes of other diseases. The average
shortest path of disease proteins is lower in protein–protein interaction network. A
disease gene does not encode a hub protein and is localized in the periphery of the
network. However, a static disease network cannot capture the dynamic changes
occurring during the progression of a disease. The dynamic rewiring in the molecular
interactions during different stages of a disease is compared with the healthy state
using differential network analysis (Fig. 9.1). It is expected that different networks

Fig. 9.1 Gene coexpression network of cancer biomarker genes in healthy ovary and ovarian
cancer (stage II and IV) showing differential connectivity among genes
9.4 Biomarker Discovery 167

representing various disease states show differential connectivity in a timescale

barring some highly conserved housekeeping interactions. Even, a particular func-
tional module may be associated with a certain disease and have a potential to be
developed as a network biomarker.
Many computational models have been developed for a better understanding of
complex diseases such as cancer, type 2 diabetes and various forms of psychiatric
diseases. The initiation and progression of various types of cancer is a stepwise
process in a timescale and can be modelled as a gene regulatory network from gene
expression data from healthy and cancer cells. Gene coexpression analysis has
demonstrated that upregulated genes in the lung cancer are highly connected with
high degree of centrality. Similarly, the global gene expression in the liver of
diabetic patients has established the role of thyroid hormone in alteration in gene
expression during type 2 diabetes. Different types of mental disorders such as
schizophrenia, bipolar disorder and depression have shown differential connectivity
among candidate genes when gene coexpression networks were reconstructed from
post-mortem brain. In addition, a genome-scale metabolic model representing a
specific type of cancer cell may help in the identification of a drug target for various
types of cancer cells. For example, a renal cancer cell having deficiency of fumarate
hydratase was modelled in silico and computational models of six hepatocellular
carcinoma patients were developed to predict 101 antimetabolites having anti-
tumour properties.

9.4 Biomarker Discovery

The development of molecular biomarkers for diagnosis of a disease is one of the

major challenges in medicine. Moreover, this task becomes more challenging in case
of complex diseases due to overlapping symptoms. Network biomarkers are useful
in diagnosis of complex diseases like cancer. EdgeMarker is a computational
approach to identify differentially expressed gene pairs known as edge biomarkers.
There are successful attempts to develop network biomarkers for diagnosis of early
stages of a disease. Dynamical network biomarker (DNB) is an algorithm to detect
early warning signals from both single sample and multiple samples before actual
onset of a disease. It has been successfully implemented in diagnosis of early stages
of lymphoma cancer and liver cancer using high-throughput data. However, it is yet
to be applied in a clinical setup with small number of samples. A machine learning
method like support vector machine is usually implemented in classification of
disease samples from healthy samples using genomic biomarkers. A SVM model
can discriminate cancer patients before and after chemotherapy based on the blood
levels of plasma nitrite and von Willebrand factor.
168 9 Clinical Bioinformatics

9.5 Multi-Target Drugs

Computational modelling is playing a major role in the identification of novel drug

targets. It is well known that cancer cells develop resistance to multiple drugs after
treatment. Drugs are also having multifarious effect in a cellular environment. Thus,
network model provides an effective framework to understand not only the effect of
a drug on the target but also side effects associated with the drug. A new paradigm
has emerged recently in drug discovery process known as polypharmacology. It
emphasizes on the development of network models including multiple drugs and
their interaction with multiple targets. A bipartite network shows direct connection
between a drug network and its target protein network. Interestingly, high degree
nodes in the drug network are connected with high degree nodes in the target protein
network. Even the drugs acting on same target protein or direct neighbours in the
network model share their side effects. Moreover, the side effects of a drug are
caused by corresponding target protein with high degree and betweenness. Cancer
cells always find an alternative pathway of proliferation when exposed to a single
drug. Therefore, drug combination therapy is an essential approach in cancer
treatment. But it is not feasible to test a combination of more than 200 cancer
drugs. Combenefit is a computational method to circumvent this challenge and
computes a synergy score between a pair of drugs. Majority of drugs act in coherent
manner on the target protein and have synergistic effect. However, few drugs are
known to function in an incoherent fashion, simultaneously having both positive and
negative effect on the target protein. The positive synergistic coefficient measures
the pharmacological effect of a combination of drugs, whereas positive side effect
coefficient is a measure of side effects of a combination of drugs.

9.6 Artificial Intelligence in Medicine

Artificial intelligence (AI) is a buzz word today in biomedical research including

health diagnostics. It has two components: machine learning and deep learning. The
rapid development of AI is attributable to better computational infrastructure, avail-
ability of big data, and advent of high-speed GPU in last two decades.

9.6.1 Machine Learning

Machine learning was first developed by Arthur L. Samuel in 1959 and its ultimate
goal is to develop a predictive model based on training the dataset using one or more
algorithms. Here, we train the model rather than explicitly program a computational
task. First, the order of the data is randomized and appropriate algorithm is selected
by trial and error. The training data which constitutes about 70% of the complete
dataset is used to create relationships in the algorithm. Training phase is followed by
the evaluation of the model’s accuracy using remaining 30% of the dataset. The
parameters in the algorithm are adjusted to fine-tune the model for improving the
9.6 Artificial Intelligence in Medicine 169

results. Some hyperparameters which cannot be learned directly from the training
process are also adjusted during fine tuning. The algorithms used are usually
complex but there is no need to compute these algorithms by the practitioner as
plenty of programs already available in the R and the Python. Although there are
hundreds of machine learning algorithms available, they can be categorized into four
major classes which are as follows:

9.6.1.1 Supervised Learning

The labelled data are generally used in large amount in order to get more accurate
results. The common types of supervised learning are classification and regression.
Classification divides the datasets based on common features. For example, support
vector machine and Naïve Bayes classifier are two methods for classification.
Regression methods such as linear regression, decision trees and ensemble
modelling find a continuous pattern in the dataset. Linear regression finds relation-
ship between some variables to predict the outcome based on inputs of quality data.
Decision trees are suitable for nonnumerical data and make decision based on a
probability. Ensemble modelling uses more than one model for prediction for more
accurate results.

9.6.1.2 Unsupervised Learning

Unlabelled data are used in unsupervised learning to find a common pattern. For
example, clustering like k-Means clustering takes unsupervised data to put similar
items in one group. It is suitable method for large datasets. Closely related data are
identified based on some metrics like Euclidian distance, cosine similarity and
Manhattan distance.

9.6.1.3 Reinforcement Learning

The learning process is improved by trial and error using positive and negative
reinforcement. It is widely used in games and robotics.

9.6.1.4 Semi-Supervised Learning

This hybrid approach uses both supervised and unsupervised learning. When small
amount of unlabelled data are available, unsupervised data can be labelled using
deep learning and converted to supervised data. This process is known as pseudo-
labelling. MRI data is usually processed using semi-supervised learning.

9.6.1.5 Support Vector Machines (SVM)

Support vector machine is a supervised kernel-based classification method that has
been extensively used to develop biomarkers to discriminate disease samples from
healthy samples. This algorithm finds a decision boundary between datasets belong-
ing to two different groups. These models can map both healthy and disease data
points in a space and then classify them into two groups based on a hyperplane
(Fig. 9.2). Only the hyperplane separating the datasets is learned in the SVM. A
hyperplane is a flat one-dimensional line in two dimensions. On the other hand, it is a
flat two-dimensional plane in three dimensions. But, we cannot visualize a
170 9 Clinical Bioinformatics

Fig. 9.2 A SVM model classifying healthy (circles) and disease (triangles) samples using radial
basis function (RBF) kernel. The maximum margin hyperplane divides both samples and solid data
points close to the hyperplane are support vectors

hyperplane in more than three dimensions. The optimal separating hyperplane

always maintains extreme minimum distance from the training dataset. A test dataset
is evaluated based on its situation on both sides of a maximal margin hyperplane.
Any data situated on the incorrect side of hyperplane is treated as misclassification
by the algorithm. Thus, the classification is optimized based on the support vectors
lying on margin of the hyperplane or on the incorrect side of the margin of their
group. The tuning parameter C, the cost of violation to hyperplane margin, is taken
for the optimization process using cross-validation. If the tuning parameter C is very
small, many data points may violate the hyperplane margin and in turn create
numerous support vectors, whereas a large C explores a narrow margin of the
hyperplane which is not prone to any violation of hyperplane. In case two groups
(i.e. healthy and disease samples) cannot be separated using a linear hyperplane, a
polynomial or radial kernel is a best choice to map the two-dimensional data into a
higher dimensional space. Support vector machines (SVM) have been widely used
for the development of molecular biomarkers as health diagnostics. For example, a
support vector machine model has successfully discriminated cancer patients before
and after chemotherapy using two blood biomarkers, namely plasma nitrite and von
Willebrand factor. However, it is difficult to apply SVM on large datasets and is not
9.6 Artificial Intelligence in Medicine 171

suitable for perceptual problems which needs feature engineering like an image
classification.

9.6.2 Deep Learning

Deep learning is a subfield of machine learning which analyses the datasets using
neural networks mimicking the human brain. The word “deep” refers to the inherent
hidden layers in the neural networks directly related to the learning power of the
algorithm. Even a single hidden layer neural network is not treated as deep due to
presence of a single hidden layer. Here, the goal of learning is to predict a response
or classify a response based on certain attributes. An Artificial Neural Networks
(ANN) is a function consisting of units called neurons (also known as perceptrons or
nodes). Typical feed-forward neural networks have three layers, namely an input
layer, a hidden layer and an output layer (Fig. 9.3). The number of neurons in an
input layer corresponds to the number of features or variables to be fed in the
networks. Similarly, the number of neurons in the output layer corresponds to the
number of items predicted or classified. The hidden layer neurons perform non-linear
transformation of the input attributes. The relative importance of each neuron in a
network is indicated by its value and weight. Each neuron has an activation function
and a threshold value to activate the neuron. Thus, a neuron performs summation of
weights of all inputs and subsequent activation function in order to pass the output to
the next layer. The net input is calculated by the formula which is as follows:

Fig. 9.3 The architecture of an artificial neural network (ANN) consisting of input layer, hidden
layers and output layer
172 9 Clinical Bioinformatics

yin ¼ x1 :w1 þ x2 :w2 þ x3 :w3 þ x4 :w4 . . . . . . :xm :wm

The value and weight of each neuron is passed to a hidden layer of neurons which
uses a function to produce a final output. The activation function is a non-linear
function which varies between 1 and + 1 considering the non-linear nature of real
world. Each neuron in the networks usually has same activation function. So, the
output is computed by applying the activation function over the net input which is as
follows:

Y ¼ F ðyin Þ

Most common function is the sigmoid one which produces an output value
between 0 and 1. The highest accuracy achieved is 1 for a perfect model. Bias is
another constant value used in the calculation of the function and is similar to an
intercept in linear regression. It facilitates the activation function of the networks to
move either upwards or downwards. This type of training is known as feed-forward
neural network. However, this model is too simple and can be improved further by
multiple hidden layers resulting in multilayered perceptron (MLP). MLP is endowed
with a unique property of backpropagation. The loss function or objective function
compares the output prediction and the true target using a distance score. This score
gives a feedback signal to adjust the value of weights to reduce this loss score. The
adjustment is done by an optimizer using backpropagation algorithm.
Backpropagation involves adjustment of weights in a neural network after getting
the errors followed by iterating the new values to optimize the model. For example,
one of the input value has an output value of 0.7 indicating an error value of 0.3
(1–0.7). After backpropagation, the output value may improve to 0.75. Thus, the
training will continue till output value reaches close to 1. Initially, errors are large
due to large weights of the input nodes. After few iterations, the error gradually
decreases to find an optimum mid-point value at the bottom of the curve. Most
common types of neural networks are recurrent neural networks (RNN),
convolutional neural network (CNN) and generative adversarial network (GAN).
The recurrent neural networks have a function which processes the input and prior
inputs across time as well. The convolutional neural network analyses a complex
data like an image data section by section. In GAN, two neural networks compete
with each other using a feedback loop and create new object.
The scope of deep learning applications is still narrow as it requires large amount
of input data and high computational power. It is a challenging task to select the
number of hidden layers and hyperparameters to develop a right model. Usually,
comparatively simple machine learning models are found to be more effective using
less amount of input data than the complex deep learning models.
9.7 Genome-Wide Association Mapping 173

9.7 Genome-Wide Association Mapping

Humans have only 0.1% variation of nucleotide bases in their genome. A single
nucleotide varies between two individuals in a population and this variation is
known as single-nucleotide polymorphism (SNP). The presence of one SNP in a
gene can lead to a monogenic disease like sickle cell anaemia. Moreover, some
complex diseases such as diabetes and cardiovascular disease are caused by the
epistatic interaction among multiple genes and the interaction between the genes and
the environment. However, there are two major challenges in understanding the
genetic basis of these complex diseases. The discovery of large number of genetic
variants along with their modified phenotypic manifestations under different envi-
ronmental conditions is the first challenge. The second challenge is to define a
correct phenotype for a study and proper statistical analysis of the data. There are
ten million SNPs estimated in human population. Current commercially available
SNP arrays provide a comprehensive genomic coverage of about 660,000 SNPs in
different populations. The study design of GWAS involves cases having the disease
and controls without the disease and other traits contributing to confounding effect.
The SNPs and samples with low quality are usually detected through low call rates
and subsequently removed from analysis. In addition, the SNPs with low minor
allele frequency and the SNPs deviating from Hardy–Weinberg equilibrium are
removed from study in order to avoid false positive findings. It is expected that
both allele and genotype frequencies remain stable in a particular population as per
requirement of Hardy–Weinberg equilibrium. However, it is advisable to first
analyse all SNPs and then examine the validity of Hardy–Weinberg equilibrium
for those SNPs associated with the phenotype. A single SNP analysis is conducted
using a statistical test to compare the null hypothesis (i.e. there is no association
between the SNPs and the phenotype) with the alternative hypothesis (i.e. there is an
association between the SNPs and the phenotype). A small p-value obtained during
the test leads to the rejection of the null hypothesis and indicates that there is a
significant association between the SNP and the phenotype. The genetic effect of the
SNP is modelled on a continuous phenotype trait using a linear regression model.
The result of the statistical analysis can be visualized using a Manhattan plot
highlighting the genomic regions showing log10 (p-value) of the association
(Fig. 9.4). Principle components analysis (PCA) is usually used in summarizing
the genome-wide variability of SNPs creating principal components of all SNPs in
the genome. The first principal component generally captures the population sub-
structure based on ethnicity. There are many programs publicly available for
genome-wide association mapping. Some R packages are available on bioconductor
for GWAS analysis and analysis of structural variations (Tables 9.1 and 9.2). PLINK
is another popular program for whole genome-wide association mapping including
various basic analysis such as haplotype analysis and LD estimation. The GWAS
Central is a comprehensive database for comparing genotype and phenotype from
various genome-wide association studies and appropriate data sets can be retrieved
from this database for analysis.
174 9 Clinical Bioinformatics

Fig. 9.4 Manhattan plot of a genome-wide association analysis of a disease. X-axis and Y-axis
show chromosomal positions and -log10 p values, respectively. The horizontal red line indicates the
significance threshold of genome-wide association

Table 9.1 R packages for GWAS analysis

R package Description
GWASTools A popular package for quality control and analysis of genome-wide association
studies of extremely large datasets.
StatgenGWAS It performs single trait genome-wide association studies.
IntAssocPlot It provides integrated visual display of GWAS results with linkage
disequilibrium matrix and gene structure.
fastJT It is useful in feature selection from high-dimensional data in machine learning
and genome-wide association studies.
GWAF It is designed for genome-wide association studies with family data.
bGWAS It is designed to improve genome-wide association studies using Bayesian
approach based on informative priors.
R/fGWAS A package for functional analysis of genome-wide association studies using
Bayesian lasso model.
RAINBOW It is designed for haplotype-based GWAS without any prior haplotype
information.
BGData It is suite of four R packages, namely BEDMatrix, LinkedMatrix, symDMatrix
and BGData for big genomic data sets.
9.8 Exercises 175

Table 9.2 R packages for analysis of structural variations

R package Description
Intasv A package for integrative analysis, annotation and visualization of
structural variations
RSVSim A package for simulation of structural variations
StructuralVariantAnnotator A package for annotation of structural variants

9.8 Exercises

1. The genotypes of a SNP in a sample are AA, AT, TT, AA, AA, AA, AT, AA, AT,
AA, TT, AT. Using the R package genetics, compute the allele frequency,
genotype frequency and heterozygosity of the genotypes. Plot the frequency of
three types of genotypes. Test whether these genotypes are in Hardy–Weinberg
equilibrium.

Solution First, the data is converted to genotypes and frequency is computed using
the function “summary ()”

>library(genetics)
>data<-c("A/A","A/T","T/T","A/A","A/A","A/A","A/T","A/A","A/
T","A/A", "T/T", "A/T" )
> geno <- genotype(data)
> summary(geno)
Number of samples typed: 12 (100%)
Allele Frequency: (2 alleles)
Count Proportion
A 16 0.67
T 8 0.33
Genotype Frequency:
Count Proportion
A/A 6 0.50
A/T 4 0.33
T/T 2 0.17
Heterozygosity (Hu) = 0.4637681
Poly. Inf. Content = 0.345679
>plot(geno)

> HWE.test (geno)

---------------------------------------------
Test for Hardy-Weinberg-Equilibrium
-----------------------------------------------
Call:
176 9 Clinical Bioinformatics

HWE.test.genotype(x = geno)
Raw Disequlibrium for each allele pair (D)
A T
A -0.05555556
T -0.05555556
Scaled Disequlibrium for each allele pair (D')
A T
A -0.5
T -0.5
Correlation coefficient for each allele pair (r)
A T
A 0.25
T 0.25
Observed vs Expected Allele Frequencies
Obs Exp Obs-Exp
A/A 0.5000000 0.4444444 0.05555556
T/A 0.1666667 0.2222222 -0.05555556
A/T 0.1666667 0.2222222 -0.05555556
T/T 0.1666667 0.1111111 0.05555556
Overall Values
Value
D -0.05555556
D' -0.50000000
r 0.25000000
Confidence intervals computed via bootstrap using 1000 samples
Observed 95% CI NA's Contains Zero?
Overall D -5.555556e-02 (-1.875000e-01, 8.333333e-02) 0 YES
Overall D' -5.000000e-01 (-2.840000e+00, 3.333333e-01) 1 YES
Overall r 2.500000e-01 (-3.333333e-01, 8.222222e-01) 1 YES
Overall R^2 6.250000e-02 ( 7.061648e-05, 6.760494e-01) 1 *NO*
Significance Test:
Exact Test for Hardy-Weinberg Equilibrium
data: geno
N11 = 6, N12 = 4, N22 = 2, N1 = 16, N2 = 8, p-value = 0.5176

The p-value indicates that genotypes are in Hardy–Weinberg equilibrium.

2. The linkage disequilibrium (LD) is non-random association among different

alleles at different loci in a population. There are two measures of LD. The D0
is the normalized coefficient of linkage disequilibrium (D), whereas R2 is the
measure of correlation coefficient between two loci.

Using the example file in the R package snpStats, compute linkage disequilibrium
(LD) among the SNPs in European and African populations based on D0 and R2
9.8 Exercises 177

Fig. 9.5 Plot of three types of genotypes

measures. Compare the LD map in both populations based on D0 plus R2 as well as

on R2 alone (Figs. 9.5, 9.6, 9.7, 9.8, and 9.9).

> library(snpStats)
Loading required package: survival
Loading required package: Matrix
> data(ld.example)
> ceph.1mb
A SnpMatrix with 90 rows and 603 columns
Row names: NA06985 ... NA12892
Col names: rs5993821 ... rs5747302
> head(support.ld)
dbSNPalleles Assignment Chromosome Position Strand
rs5993821 G/T G/T chr22 15516658 +
rs5993848 C/G C/G chr22 15529033 +
rs361944 C/G C/G chr22 15544372 +
rs361995 C/T C/T chr22 15544478 +
rs361799 C/T C/T chr22 15544773 +
rs361973 A/G A/G chr22 15549522 +
178 9 Clinical Bioinformatics

Fig. 9.6 Linkage disequilibrium in European population based on D0 and R2 measures

> ld.afro <- ld(yri.1mb, stats=c("D.prime", "R.squared"),

depth=50)
> ld.euro <- ld(ceph.1mb, stats=c("D.prime", "R.squared"),
depth=50)
> spectrum <- rainbow(10, start=0, end=1/6)[10:1]
> image(ld.euro$D.prime, lwd=0, cuts=9, col.regions=spectrum,
colorkey=TRUE)

> image(ld.afro$D.prime, lwd=0, cuts=9, col.regions=spectrum,

colorkey=TRUE)

> use <- 50:300

> image(ld.euro$R.squared[use,use], lwd=0)

> image(ld.afro$R.squared[use,use], lwd=0)

9.9 Multiple Choice Questions 179

Fig. 9.7 Linkage disequilibrium in African population based on D0 and R2 measures

Linkage disequilibrium is more noticeable in the European population than in the

African population in all comparisons.

9.9 Multiple Choice Questions

1. A big data is characterized by:

(a) Volume
(b) Variety
(c) Velocity
(d) All the above
2. Copy number variation (CNV) is a structural genomic variation involving more
than:
(a) 1 kbp
(b) 10 kbp
(c) 20 kbp
(d) 25 kbp
3. Whole-exome sequencing for discovery of a novel gene requires:
(a) 100–200 case-control pairs
(b) 300–400 case-control pairs
180 9 Clinical Bioinformatics

Fig. 9.8 Linkage disequilibrium in European population based on R2 measure

(c) 500–1000 case-control pairs

(d) 2000–3000 case-control pairs
4. The machine learning algorithm not based on supervised learning is:
(a) Support vector machine
(b) Naïve Bayes classifier
(c) k-mean clustering
(d) Linear regression
5. The activation function of artificial neural network is:
(a) Linear
(b) Non-linear
(c) Both linear and non-linear
(d) None of the above
6. The backpropagation is the property of:
(a) Multilayered perceptron
(b) Feed-forward neural network
(c) Feed-backward neural network
(d) None of the above
9.9 Multiple Choice Questions 181

Fig. 9.9 Linkage disequilibrium in African population based on R2 measure

7. The total number of SNPs estimated in human population is:

(a) one million
(b) five million
(c) ten million
(d) 15 million.
8. All SNPs associated with the phenotype are expected to follow:
(a) Hardy–Weinberg equilibrium
(b) Hardy–Weinberg disequilibrium
(c) Selection and migration
(d) Random genetic drift
9. The genome-wide association can be visualized through:
(a) Scatter plot
(b) Manhattan plot
(c) Euclidian plot
(d) Hamilton plot
182 9 Clinical Bioinformatics

10. Which of the following is/are a type (types) of neural network?

(a) Recurrent neural networks (RNN)
(b) Convolutional neural network (CNN)
(c) Generative adversarial network (GAN)
(d) All the above

Answer: 1. d 2. a 3. c 4. c 5. b 6. a 7. c 8. a 9. b 10. d

Summary
• A comparative study of a disease model and a normal healthy model provides the
mechanistic details of pathogenesis at a deeper level.
• A big data is characterized by three V-features: volume, variety and velocity.
• The human diseases associated with the copy number variation (CNV) are cancer,
autism, schizophrenia and Alzheimer disease.
• Whole-exome sequencing (WES) covers only 2% of the human genome encoding
protein sequences and detects 85% of the genomic variations associated with
various diseases.
• The initiation and progression of a disease can be explicitly captured in form of a
computational network module.
• A disease gene does not encode a hub protein and is localized in the periphery of
the network.
• A genome-scale metabolic model representing a specific-type of cancer cell may
help in the identification of a drug target.
• A bipartite network shows direct connection between a drug network and its
target protein network.
• Support vector machine (SVM) is a machine learning classification method for
development of molecular biomarkers to discriminate disease samples from
healthy samples.
• A single SNP analysis compares the null hypothesis (i.e. there is no association
between the SNPs and the phenotype) with the alternative hypothesis (i.e. there is
an association between the SNPs and the phenotype).

Suggested Readings
Wang X, Baumgartner C, Shields DC, Deng H-W, Beckmann JS (2016) Application of clinical
bioinformatics. Springer, Dordrecht
Trent RJA (2014) Clinical bioinformatics. Humana Press, Totowa
Liang K-H (2013) Bioinformatics for biomedical science and clinical applications. Woodhead
Publishing, Oxford
Raza K, Dey N (2021) Translational bioinformatics applications in healthcare. CRC Press, Boca
Raton
Agricultural Bioinformatics
10

Learning Objectives
You will be able to understand the following after reading this chapter:

• Role of bioinformatics in agriculture.

• Pan-genomic view of various crop species.
• Genome assembly of crop plants.
• Identification of homeologs in crop species.
• Principles and methods of genomic selection for improvement of crops.
• Crop phenomics.
• Computational modelling of agricultural systems.

10.1 Introduction

The estimated population of this world would be about nine billion by the year 2050
and consequently there is a need for a very steep increase in the food production to
feed the projected population. So, there is a necessity for large-scale innovations in
agriculture to boost its productivity and sustainability as well. Crops have a vital role
in global economy being a major source of nutrients for the ever increasing world
population. In fact, crop genomics is expected to play a significant role in the second
green revolution. The available genomic databases useful in crop genomics research
are listed in Table 10.1. The agricultural system is a complex system where the
biology of a crop interacts with the environment and management practices. Bioin-
formatics can play a pivotal role in enhancing the agricultural productivity due to
rapid progress in both computing power and next-generation sequencing technol-
ogy. The interaction and behaviour of the overall agricultural system is better
modelled through integrating the data with computational analysis. We can capture

# Springer Nature Singapore Pte Ltd. 2022 183

B. K. Tiwary, Bioinformatics and Computational Biology,
https://doi.org/10.1007/978-981-16-4241-8_10
184 10 Agricultural Bioinformatics

Table 10.1 Genomic databases of crops

Database Description URL
Gramene A knowledgebase of genomic and pathway data of major www.gramene.org
crops
LegumeIP An integrative comparative genomics and transcriptomics plantgrn.noble.org/
database of legume crops LegumeIP
SoyBase A database of soybean genetic and genomic data soybase.org
GreenPhyID A database of comparative pan-genomics of plants www.greenphyl.
including major crops org

Fig. 10.1 Structural variants like CNV and PAV cause variations in the functional gene content of
crop plants. Here, CNV in the genome is represented by three copies of Gene 2 and two copies of
Gene 3. The PAV indicates loss of one copy of each gene with gain of another gene, Gene 4 in the
genome

a wide variety of real-time big data of different genomic and environmental variables
using existing sensor technologies. Some of the promising areas of agriculture
having significant attention of bioinformatics research are pan-genome of crops,
assembly of plant genomes, identification of homeologs, genomic selection and
modelling of agricultural systems.

10.2 Pan-Genome of Crops

Modern crops are endowed with a wide variety of small to large continuum of
genomic variations in the form of single-nucleotide polymorphisms (SNPs)-
insertions and deletions (indels)-structural variants (SV). SNPs have a major effect
on the functional gene content of a crop genome either through premature stop
codons or changes in the key functional sites. Similarly, small indels can have a
major impact on the genome variation caused by frameshift changes and premature
stop codons in the genes. The structural variants such as copy number variants
(CNVs) and presence/absence variants (PAVs) are crucial determinants of useful
agronomical traits (Fig. 10.1). Both natural and artificial selection operate on these
genomic variations to increase the overall fitness of the genotype and crop
10.2 Pan-Genome of Crops 185

Fig. 10.2 The pan-genome of crop plants exhibits variable portions of core genes and dispensable
genes. The pan-genome size of each species is indicated by the number of pan-genes in a particular
species at the centre of the ring

productivity in an agricultural field, respectively. We depend on a single reference

genome for understanding the existing genomic variations in a single species and
this approach has a limited scope. The pan-genome concept was developed to
capture large variations in gene content among individuals of a particular species.
Thus, the entire gene repertoire of a species can only be described by the sequences
obtained from multiple individuals belonging to a species. Recent pan-genomic
studies on crop species like rice, maize, wheat, soybean, etc. have brought a
paradigm shift in our understanding of crop biodiversity and their genetic
improvement.
Pan-genome refers to the full set of genes in a taxonomic clade like a species. The
pan-genome can be divided into two groups: core genes shared by all individuals in a
species and dispensable genes only present in few individuals or present as singleton
gene in a specific individual. The proportion of dispensable gene varies widely
between different species of crops (Fig. 10.2). For example, rice plant has 8% of
dispensable genes, whereas maize plant has 61% of dispensable genes. The percent-
age of dispensable genes in a species reflects the diversity among individuals of a
particular species. A species with higher ploidy and outcrossing rate is likely to have
larger pan-genome with a higher percentage of dispensable genes. Rice has a large
pan-genome (48,098 genes) with 41% of dispensable genes. However, the number
of new genes discovered in a pan-genome is likely to decrease with inclusion of new
individual samples of a species. The sample size and selection of samples are key
factors in a pan-genome study. Ideally, sample size should be large enough to realize
the full expansion of the pan-genome. For example, 3010 rice accessions were
sequenced to obtain the large pan-genome of rice. A genome browser, Rice
Pan-genome Browser (RPAN) can be used as a search tool for the rice
pan-genome derived from this study and for visualization of all genomic variations
186 10 Agricultural Bioinformatics

as well. However, the expansion of the pan-genome in a species reaches a plateau

after incorporation of finite number of genomes. The selection of sample has a
significant impact on the size of the pan-genome. Thus, a pan-genome study should
include both cultivable crop and its wild relatives to generate a large pan-genome
with high content of dispensable genes. Crop species with a high ploidy level and a
high outcrossing rate are likely to have a larger pan-genome with higher percentage
of dispensable genes.
The core genes in a crop species are highly conserved and likely to play a critical
role in essential cellular functions. On the other hand, dispensable gene contributes
to the phenotypic diversity of agronomic traits in a crop species and has a critical role
in enhancing its agricultural productivity. The rate of dispensable gene evolution is
faster than that of core genes. Dispensable genes are enriched with adaptive
functions such as gene regulation, signal transduction and responses to environmen-
tal and defence stimuli. Interestingly, dispensable genes have a high frequency of
both synonymous and nonsynonymous substitutions and a high proportion of them
is not functionally annotated in many crop species like soybean.
Structural variation is a major causative factor in generating dispensable genes.
Structural variations are defined as the large genomic alterations more than 50 bp in
length including duplications, deletions, insertions, inversions and translocations.
Transposable elements (TEs) or jumping genes are one of the most dynamic parts of
the genome and generate structural variations in the form of PAVs in many crop
species. Similarly, structural variations are also generated through recombination of
non-allelic non-homologous sequences. The structural variations have an adverse
impact on the fitness of a crop species. Moreover, both polyploidy and outcrossing
rate provide a capacity in crops to tolerate the structural variations in the genome. A
polyploid species has a multiple copies of essential genes that provide a buffer
against the deleterious effects of structural variations. For example, a tetraploid
species of mustard plants has a higher percentage (38%) of dispensable genes in
comparison to a diploid species (19%). An outcrossing species is likely to have
abundant structural variations conferring heterozygosity to tolerate the adverse
effects of deleterious mutations. For example, maize crops which were outbred
extensively have 61% of dispensable genes in their genome.
The genetic basis of agronomic traits is controlled by structural variations in crop
plants. The underlying genomic variation in a crop species may be linked to an
agronomic trait using quantitative trait loci (QTL) or genome-wide association
studies (GWASs). The dispensable genome of a crop is enriched with genes
associated with biotic and abiotic stress resistance. The structural variations in the
genes involved in the abiotic and the biotic stress in crop plant are given in
Tables 10.2 and 10.3. In addition, structural variations in some genes in crop plants
are known to regulate the photoperiod sensitivity and flowering time. For example, a
copy number variation at the HvFT1 locus is associated with the variation of
flowering time in barley. Even the whole plant architecture such as height of a
crop plant can be reduced drastically by structural variations. In wheat plant,
duplication of a gene (Rht-D1b) leads to reduction of plant height by more than
70%. Structural variations also influence the overall yield of the crop and other
10.2 Pan-Genome of Crops 187

Table 10.2 Structural variations in the genes involved in the abiotic stress
Type of SV Gene Crop Trait
PAV Sub1A Rice Tolerance to submergence
PAV Pup1 Rice Tolerance to phosphorus starvation
CNV MATE1 Maize Tolerance to aluminium
CNV FR-2 Wheat Tolerance to cold climate
CNV FR-H2 Barley Tolerance to frost
PAV GmCHX1 Soybean Tolerance to salinity
CNV Ppd-B1 Wheat Sensitivity to photoperiod

Table 10.3 Structural variations in the genes involved in the biotic stress
Type of SV Gene Crop Trait
PAV Pikm1-TS Rice Resistance to blast disease
PAV Pi21 Rice Resistance to blast disease
PAV qHSR1 Maize Resistance to head smut
PAV Lr10 Wheat Resistance to leaf rust
CNV Rhg1 Soybean Resistance to cyst nematode
PAV R1, ELR Potato Resistance to late blight
PAV Yr36 Wheat Resistance to stripe rust

associated traits such as grain quality or fruit quality. For example, a 1212-bp
deletion in GW5 gene can alter the grain weight and width in rice plants. Elongated
fruit shape in tomato is due to copy number variation of SUN gene. Since structural
variations create differences in the gene content between different individual lines,
the heterosis effect in hybrids may be due to passage of complementary genes from
the individual parental lines. Structural variations might have played a significant
role in the domestication of crop plants. For example, during domestication of maize
plant, both increase in apical dominance and decrease in tiller number occurred due
to insertion of transposable elements at the tb1 locus.
Pan-genomic study of crop wild relatives will reveal the full landscape of
biodiversity in the species and this untapped resource can be utilized for boosting
the crop productivity. In fact, wild relatives of crop plant are being used for
backcrossing during conventional breeding. This pan-genomic studies in the crop
species can be linked to the QTL/GWAS to identify useful elite genes for further
breeding strategies. Thus, a comprehensive genome resource of crop plants will be
created through pan-genome studies in near future which can be further exploited by
plant breeders for enhanced agricultural productivity.
Computational analysis of the pan-genome is prerequisite for fully understanding
the genomic landscape of a crop species. The pan-genome of a species is computa-
tionally represented by a data structure known as coordinate system where all genetic
variants are explicitly represented in form of a sequence coordinate graph. Although
there are many available software for pan-genome analysis in prokaryotes, recent
development of a software like Pangloss for pan-genome analysis in eukaryotes will
be highly useful for a crop species. Pangloss is written in the Python language to
188 10 Agricultural Bioinformatics

characterize the pan-genome of eukaryotes for gene prediction, gene annotation and
functional analyses. The total diversity in a crop species in form of a pan-genome can
be visualized in form of a phylogenetic tree. A phylogenetic tree from whole-
genome data can be reconstructed using both DNA sequences and the gene content.
In a sequence-based phylogenetic tree reconstruction, genomic sequences from all
variants are first aligned using a multiple sequence alignment and a phylogenetic tree
is then reconstructed using the evolutionary distances. In gene content approach, the
presence and absence of a gene in the genome is scored in binary numbers 1 and
0, respectively, followed by construction of a distance matrix consisting of 1 and 0 to
represent total pan-genomic profile of a crop species. The distances between differ-
ent variants are measured in terms of Jaccard distance and Manhattan distance.
Manhattan distance refers to sum of absolute gene-wise differences between two
genomes, whereas Jaccard distance between two genomes measure the amount of
intersection in terms of presence or absence of a gene. If two genomic sequences are
identical, both Manhattan distance and Jaccard distance are 0.0. Conversely, if there
is no overlap between two genomic sequences, both measurements are 1.0. A
distance-based phylogenetic approach such as neighbour joining or UPGMA will
generate an evolutionary tree exhibiting relationship among different variants of the
pan-genome in a crop species based on presence or absence of individual genes.

10.3 Assembly of Crop Genomes

Plants have evolved a large and complex genome for their survival in the terrestrial
environment. The genome of the first plant species, A. thaliana followed by genome
of some economically important species like rice, maize and papaya were sequenced
using the Sanger method. The advent of next-generation sequencing technology has
speeded up the pace of genome sequencing with a reduced cost. The second-
generation deep sequencing of crops is challenging due to large genome size,
duplications and repeat contents. The genome assembly process consists of a
combination of sequencing and computation. The reads generated by sequencing
is combined using a computer program called assembler. Therefore, assembly and
analysis of short reads in plant genome needs substantial bioinformatics skill. The
short reads generated through second-generation technology needs high coverage
from 50X to 100X; even this high coverage of 100X may be inefficient in
deciphering complex plant genomes. On the other hand, long-read technologies do
not need such high coverage and a coverage of 20X to 30X may be sufficient to get a
good assembly of the genome. Thus, second-generation deep sequencing technology
has produced a draft genome of many plant species lacking almost 20% of the vital
genomic information about the species. The resulted draft genome cannot be used
reliably to understand the repetitive part of the genome and to discriminate between
functional genes and pseudogenes as well.
In fact, a plant genome appears like a gene island surrounded by more than 80%
of the repeat sequences. Transposons have played a profound evolutionary role in
the evolution of plant species and transposable elements (TE) constitute the major
10.4 Identification of Homeologs 189

portion of repetitive sequence in a plant genome. The small genome crops like rice
have only 17% of transposons, whereas a large genome of maize have almost 85% of
transposable elements. The abundant presence of repetitive transposons in a crop
species poses a critical challenge in assembly of their genome. The short reads have
less power to resolve the repeats in the plant genome. Thus, longer reads generated
by single-molecule sequencing must be combined with short reads to resolve the
repetitive DNA. The longer repeats than the read length create gaps during the
assembly process which in turn can be resolved by paired-ends during sequencing.
Polyploidy or fusion of two or more genomes in a species has played a significant
role in the evolution of wide diversity in plants. It is a result of either whole-genome
duplication known as autopolyploidy or crossing of two species followed by dupli-
cation known as allopolyploidy. Genome duplications produce new genes with new
functions and thereby generate novel phenotypes. The polyploid plants have better
adaptive capability in an ever-changing environment due to their genomic plasticity.
A majority of crop plants like wheat, potato, sugarcane and banana are polyploids.
The redundancy due to presence of two or more copies of genes can adversely affect
the accuracy of the genome assembly. Gene duplication is another major force for
creating new genes in a genome in form of paralogs. It is difficult to distinguish
alleles from paralogs in the genome assembly of natural heterozygotes. We always
look for lineage-specific genes in a crop species for functional studies. Sometimes,
these lineage-specific genes are simply artefacts of misassembles. These artefacts
can be avoided by developing some novel algorithms to identify real lineage-specific
genes. The development of de novo approaches in assembly such as de Bruijn graphs
combined with Eulerian paths has facilitated assembly of plant genomes with long
repeats. With the advent of third-generation single-molecule sequencing such as
PacBio, the assembly of plant genomes with long reads having an average length of
10,000 bp can circumvent the problem of long repeats. However, the high error rate
(5–15%) of this newer technology is a major constraint in the application to crop
genomes.

10.4 Identification of Homeologs

The homeologous genes in plants are the products of allopolyploidy and have a
common ancestry like other homologous genes. Homoeologs are the gene pairs
derived by speciation and again recombined in a genome of a species by allopoly-
ploidy. In simple words, we can define homeologs as the orthologs between the
subgenomes of a species. The subgenomes are individual genomes in an allopoly-
ploid species each inherited from a different ancestral species. Thus, homeologs
have a combined evolutionary and functional features of orthologs (derived by
speciation) and ohnologs plus paralogs (derived by duplication) (Fig. 10.3). The
homeologous genes are expected to follow same order (collinear) between the
ancestral genome and the descendant genomes. This is known as positional
homeology like positional orthology and the homeologous genes maintain one-to-
one relationship. The best bidirectional hits (BBH) approach works well in inferring
190 10 Agricultural Bioinformatics

Fig. 10.3 Venn diagram showing relationship of homeologs with orthologs, ohnologs and
paralogs. Homeologs are located at the intersection of orthologs and ohnologs plus paralogs

one-to-one relationship between orthologs. However, this approach fails to identify

those homeologs underwent genomic rearrangements through single-gene duplica-
tion or translocation. The sequencing of complex crop genome is challenging due to
difficulty in correct identification of homeologous sequences.
Computational analysis of the genome of an allopolyploid crop species such as
wheat, cotton and coffee is necessary for accurate identification of homeologous
genes. Evolution-based computational algorithms are used to infer homoeologs from
the genome sequences. These algorithms are based on two approaches: the phyloge-
netic tree-based and the graph-based. The phylogenetic tree-based approach
reconciliates gene tree with species tree to distinguish orthologs (speciation event)
and paralogs (duplication event). It distinguishes each node of a phylogenetic tree
either as a speciation or duplication event using species tree as a reference. Conse-
quently, the gene pairs coalescing or joining at the speciation node and the duplica-
tion node are inferred as the orthologs and the paralogs, respectively. Ensembl
Genomes has applied this algorithm to identify homoeologs in a polyploid species
like wheat. Here, each subgenome is treated as a different species and orthologs
between subgenomes are detected using phylogenetic tree-based orthology pipeline.
Finally, the identified orthologs among different subgenomes are relabelled as
homeologs in an allopolyploid species.
Graph-based orthology detection method is based on the similarity between
homologous sequences. The best reciprocal hits are identified among two
subgenomes using the BLAST to identify homoeologs. This approach was used
successfully to identify homeologs in wheat. However, this method only detects one-
to-one homeology but often fails to identify one-to-many and many-to-many
homeologs. The limitation of this approach is more obvious for highly duplicated
crop genomes. It generates many false-negatives because it does not account for the
differential gene loss among the subgenomes. In addition, there are many fragmen-
tary genes and sequencing errors in the crop genome leading to suboptimal score
10.5 Genomic Selection 191

during the BLAST. Orthologous matrix (OMA) database has implemented another
graph-centric approach for identification of homeologs in wheat. First, mutually
closest homologs are chosen based on the evolutionary distance taking into account
both differential genes loss and many-to-many relationship among genes. In parallel
to earlier phylogenetic-based approach, orthologs between different subgenomes are
detected using the standard pipeline followed by relabelling of orthologs as
homeologs. Thus, the OMA approach is a better graph-based approach because it
depends on the evolutionary distances in comparison to alignment score. However,
the quality of assembly and annotation of allopolyploid crop genomes need to be
improved for the high accuracy of homeolog inference.

10.5 Genomic Selection

Crop breeding depends upon the repetitive cycles of phenotypic selection followed
by crossing in each generation to produce a superior genotype. The marker-assisted
selection (MAS) was used for improvement of common crop in the recent past by
detecting underlying major genes but failed to detect minor gene effects in the
breeding population. With the availability of whole genomic sequences of various
crop species, the minor-effect genetic variants associated with agronomic traits can
be identified across the genome. The genomic selection operates on these genome-
wide genetic variants circumventing the need for repeated cycles of phenotype
selection. It selects the best candidates as parents for the next breeding cycle using
predicted breeding value which includes their genotypes, their phenotypes and the
genotypes of their relatives (Fig. 10.4). The breeding value of an individual is
measured by the average performance of its progenies. Single-nucleotide
polymorphisms (SNPs) are the variations at the nucleotide level in a population
and are extensively used for identification of more than 10,000 quantitative trait loci
(QTL) having economic importance. In a typical genomic selection program, there

Fig. 10.4 Schematic diagram showing development of prediction equations from phenotype and
genotype in form of thousands of SNPs in the reference population and subsequent application of
prediction equations in the selection of candidates using thousands of SNPs data for computing
genomic estimated breeding value
192 10 Agricultural Bioinformatics

are two kinds of distinct but related population: the training population and the
breeding population. The genotype with all genome-wide markers and phenotype of
individuals in the training population are known, whereas only genotypes including
all genome-wide markers of individuals in the breeding population is known without
any knowledge of their phenotype. The genetic values of individuals in the breeding
population need to be predicted during genomic selection. A prediction model is
developed from the training population to predict the genomic estimated breeding
values (GEBV) of individuals in the breeding population. This approach has a
greater power in capturing the effects of small-effect loci as well. Overall, this
model captures all the additive genetic variance of a particular trait of economic
importance. Genomic selection has a high accuracy of genomic-enabled prediction
for simple traits with high heritability. Moreover, it also has a potential to improve
the complex traits with low heritability. Thus, genomic selection can play a signifi-
cant role in enhancing genetic gain per unit time and cost in a breeding population
through accelerated breeding cycles.
Since the variance of a complex trait is assumed in various forms in each type of
prediction model, the field response of different types of prediction models varies
widely. The general model for the whole-genome regression analysis can be
formulated as:

y ¼ Xb þ Wam þ e

where y is the vector representing phenotypic value,

X is an incidence matrix for the fixed effects,
b is a vector of fixed effects including the overall mean (μ),
W is a matrix of genotypic score of each SNP marker,
am is the vector representing the additive effects of the SNP marker
and e is the vector of random residual effects with distribution e ~ N (0, Iσ 2e).
The SNP genotypes at each locus are represented as a matrix of minor allele
frequency (0, 1 and 2 copies) or deviation of minor allele counts from heterozygous
genotype (1, 0 and 1) for diploid genotypes of AA, AB and BB, respectively.
The most common predictive models for quantitative traits are the genomic best
linear unbiased prediction (G-BLUP) and the ridge-regression BLUP (RR-BLUP).
The best linear unbiased predictor (BLUP) is the best model as it finds the maximum
correlation between the predicted value and the true value. Further, it is linear
because of linear combinations of estimation of model effects. It is also called as
unbiased because both expected value and true value are equal. A crop breeder
typically uses BLUPS for marginal effect predictions. In simple words, the BLUP is
a linear combination of effect estimates from mixed model with random effects. The
G-BLUP model is a mixed linear model estimated from genome-wide markers and a
modified form of conventional BLUP model emphasizing on the pedigree relation-
ship. This model estimates random effects and utilizes the genomic information from
relatives giving more weights to closest relatives. The G-BLUP model can be
formulated as:
10.6 Crop Phenomics 193

y ¼ Xb þ Zu þ e

where y is the vector of phenotypic observations,

b is the n x 1 vector of fixed effects,
X is the n x p design matrix relating observations to fixed effects,
Z is an n x n identity matrix connecting the n observations to the n individual
effects,
u is the n x 1 vector of breeding values of each individual,
and e is the n x 1 vector of residual errors having the variance Iσ 2e.
More complex models can be developed in case of a plant breeding experiment.
For example, y may represent phenotypes of crops and the g families or lines that are
replicated over s environments. The Z matrix will be transformed to n X (g + s) in
such a case. The u represents the G-BLUPS or genomic estimated breeding value
(GEBV) of all individuals in the matrix. G-BLUP model can be fitted with the
package regress and synbreed in the R environment.
Since the number of SNP predictors is larger than the number of phenotypic
observations during genomic selection, the regression coefficients are constrained by
adding a small constant lambda. This process of shrinkage or regularization is
known as ridge regression. The ridge regression (RR) is like least square method
which shrinks the estimated coefficients towards zero. Here, lambda is a tuning
parameter for penalty which varies between 0 and infinity. The bias in the model
increases with the amount of shrinkage (large lambda) but there is a corresponding
decrease in the variance. RR-BLUP is also a mixed linear model where variance is
treated as equal for all the markers with small effects. RR-BLUP model can be
developed in the R environment using a package called solving mixed model
equations in R (sommer) and rrBLUP.
Bayesian analysis is often conducted using Markov Chain Monte Carlo (MCMC)
methods. Gibbs sampler is one of the most common Bayesian method for genomic
selection. Bayesian whole-genome regression prediction models can be developed
using the BGLR package in the R environment. It implements a variety of Bayesian
approaches including Bayesian ridge, Bayesian Lasso, BayesA, BayesB,
BayesC, etc.

10.6 Crop Phenomics

The plant phenotyping is a core concept of plant breeding where some elite plants are
selected for subsequent crossing and genetic gain. This traditional phenotyping
method focusses on one or few traits in a particular environment and often leads to
a marginal annual genetic gain (0.5–1%) in the productivity of major cereal crops.
The crop phenomics deals with collection of multi-dimensional phenotypic data at
various scales such as cell level, organ level, plant level and population
level (Table 10.4). There are automated phenotyping platforms such as for
generating high-throughput data from individual plants. The most common
phenotyping indices of individual plants are leaf length, leaf area and fruit volume.
194 10 Agricultural Bioinformatics

Table 10.4 Phenomics traits in crops

Morphological traits Structural traits Physiological traits Performance traits
Plant volume Cell division rate Cell turgor Biomass/hectare
Stalk shape No. of vascular bundle Water transportation Seed yield
Coverage fraction Vessel size Stalk lodging
Internode diameter Cross section thickness Interception of PAR

A genomic selection model can be applied for genetic gain under variable season or
field condition. However, combination of genomics data with phenomics data
obtained from repeated experiment in variable environment has a better potential
for genetic gain in crop breeding. In phenome to genome approach, the SNPs and
genomic regions are associated with agronomical phenotypic traits using high-
throughput phenotyping tools. The final statistical model consists of three
components: genotype, environment and management (G X E X M).
Image-based phenotyping is used frequently in both laboratory and field
environments. Plant phenotypes are divided into four groups: colour, texture, quan-
tity and geometry in image-based phenotyping. Each plant is exposed to top and side
cameras at three different wavelength bands: visible, near-infrared and fluorescence
to get thousands of images during whole phenotyping period. The visible range
provides colour images regarding nutrition, health, growth and biomass status of an
individual plant. Moreover, infrared range detects a quantitative measure of the
water content and fluorescence imaging senses chlorophyll and other fluorophores
in a plant. Machine learning and deep learning techniques are widely used in the
image-based phenotyping such as object detection and classification. The phenotype
image analysis consists of four successive steps: pre-processing, segmentation,
feature extraction and post-processing. The first step involves preparation of the
image for analysis like detection of outlier, reproducibility of trait and normalization
of phenotypic profile. It is followed by the second step of segmentation dividing the
image in foreground (the plant) and background (the imaging chamber) parts.
Feature extraction selects the optimal set of explanatory variables in a stepwise
manner using variance inflation factors to get a list phenotypic traits. In general,
about 400 phenotypic traits are extracted from the image of each crop plant. The final
step of post-processing is summarization of all computed results.

10.7 Crop Systems Modelling

Agricultural systems science is an interdisciplinary area to understand the overall

behaviour of complex agricultural systems in different time and space. An agricul-
tural system model consists of components and their interactions with respect to
natural resources, agricultural production and human factors. These models can
predict the overall performance of agro-ecosystem in near future. They are also
helpful to land managers and policymakers in taking decisions regarding financial
planning, crop management, land management and pest management. These
10.7 Crop Systems Modelling 195

software programs are known as decision support systems. First crop systems
models were statistical models to predict the response of a system such as crop
yield based on the past data sets. However, statistical models are not suitable to
predict the impact of unseen climate changes such as increase in CO2 concentration
in the atmosphere. Statistical models are generally useful in case sufficient historical
data sets are available for prediction.
The dynamic system simulation models are widely used to understand crop and
farming system models in response to certain external changes such as weather or
management practices. The typical output of a dynamic model is daily outputs of a
specific crop over a period of time. The models are highly complex having numerous
descriptive variables and parameters with long run times. Thus, summary models are
derived from a complex model suitable in certain situations. A common model of
crop system, Agricultural Production System Simulator (APSM) is a complex
dynamic model, which predicts the yield of a crop as a function of time and space
based on several inputs such as weather and soil properties. This model can be
implemented in the R environment using a package called apsmr. Similarly, World
Food Studies (WOFOST) crop growth simulation model and Light interception and
utilization (LINTUL) can also be implemented in the R using the packages Rwofost
and LINTUL-package, respectively. AquaCrop is a mechanistic crop growth generic
model from emergence to maturity developed by Food and Agriculture organization.
It is most widely used generic model to simulate growth of different crops under
variable climates and geographical locations. AquaCropR is an R package built upon
the AquaCrop software with some additional functions. An R package for agricul-
tural data sets known as agridat is available in the R environment.

Box 10.1 Heritability

The heritability in broad sense is a measure of relationship between phenotypic
value and the breeding value for a particular trait in a plant or animal popula-
tion. Breeding values are likely to have more impact on a phenotype if
heritability of the trait is high. On the other hand, heritability in the narrow
sense is the square of correlation between breeding value and phenotypic
value. The value of heritability varies from zero to 100% and traits having
heritability more than 0.4 are treated as highly heritable. For example, the traits
for percent fat and percent protein in milk of dairy cattle have a heritability of
0.5 and are highly heritable. It is a measure of overall population and varies
between different populations and environments.

Box 10.2 Breeding Value

The plants or animals with best genes are believed to have the best breeding
value. In artificial selection, the parents with best breeding value are chosen to

(continued)
196 10 Agricultural Bioinformatics

Box 10.2 (continued)

contribute genes to the next generation. The relationship between a phenotype
and breeding value is measured in terms of heritability. The phenotype is the
correct manifestation of underlying breeding value in case the heritability of
the trait is high. Conversely, the low heritability indicates little connection
between a phenotype and its breeding value. The breeding value is predicted
from objective and numerical phenotypic and genomic data. Sometimes, we
rely on the information from relatives to achieve high accuracy in prediction.
Accuracy measures the parity between true breeding value and its predicted
value.

10.8 Exercises

1. The R package ZeBook contains a dynamic model of crop growth for maize
cultivated in potential conditions. Three state variables, namely leaf area index
(LAI), total biomass (B) and cumulative thermal time since plant emergence
(TT) indicate the overall crop growth. Find the parameters and compute the
growth of crop in terms of three state variables from day 100 to 150. Plot the
increase in total biomass during this period.

Solution (Figs. 10.5 and 10.6)

library(ZeBook)
> weather = maize.weather(working.year=2010, working.site=30,weather_
all= weather_ EuropeEU)
> maize.define.param()
Tbase RUE K alpha LAImax TTM TTL
nominal 7 1.85 0.7 0.00243 7 1200 700
binf 6 1.50 0.6 0.00200 6 1100 600
bsup 8 2.50 0.8 0.00300 8 1400 850

> model<-maize.model2(maize.define.param()["nominal",], weather, sdate=

100, ldate=150)
>model

>plot(model$B,type="o",pch=16)

2. The alpha lattice design of spring oats is an available dataset with the R package
agridat. This dataset contains 72 observations on five variables, namely plot
number (plot), replicate (rep), incomplete block (block), genotype (gen) and dry
matter yield (yield). Estimate the genetic effects using the best linear unbiased
prediction (BLUPs) and heritability for yield in the R environment.
10.8 Exercises 197

Fig. 10.5 Growth of maize crop in terms of three state variables

198 10 Agricultural Bioinformatics

Fig. 10.6 Increase in total biomass from day 100 to day 150

> library(agridat)
> library(lme4)
> library(emmeans)
> data(john.alpha)
> dat <- john.alpha
> model <- lm(yield ~ 0 + gen + rep, data=dat) # Randomized Complete Block
(RCB) design
> model1 <- lm(yield ~ 0 + gen + rep + rep:block, dat) # Intra-block analysis
> model2<-lmer(yield~0 + gen+rep+(1|rep:block),dat)# Combined inter-intra
block analysis
> anova(model2)
Analysis of Variance Table
npar Sum Sq Mean Sq F value
gen 24 380.44 15.8515 185.9959
rep 2 1.57 0.7851 9.2124
> means <- data.frame(rcb=coef(model)[1:24],
+ ib=coef(model1)[1:24],
+ intra=fixef(model2)[1:24]) # Variety means
> head(means)
rcb ib intra
genG01 5.201233 5.268742 5.146433
10.9 Multiple Choice Questions 199

genG02 4.552933 4.665389 4.517265

genG03 3.381800 3.803790 3.537933
genG04 4.439400 4.728175 4.528828
genG05 5.103100 5.225708 5.075944
genG06 4.749067 4.618234 4.575395
> covtosed <- function(x){
+ n <- nrow(x)
+ vars <- diag(x)
+ sed <- sqrt( matrix(vars, n, n, byrow=TRUE) +
+ matrix(vars, n, n, byrow=FALSE) - 2*x )
+ diag(sed) <- 0
+ return(sed)
+ } # conversion of variance-covariance matrix to SED matrix
> model5blue <- lmer(yield ~ 0 + gen + rep + (1|rep:block), dat) # BLUE
> ls5blue <- emmeans(model5blue, "gen")
> con <- ls5blue@linfct[,1:24] # contrast matrix for genotypes
> tmp <- tcrossprod( crossprod(t(con), vcov(model5blue)[1:24,1:24]), con)
> sed5blue <- covtosed(tmp)
> tmp <- tcrossprod( crossprod(t(con), vcov(m5blue)[1:24,1:24]), con)
> sed5blue <- covtosed(tmp)
> vblue <- mean(sed5blue[upper.tri(sed5blue)]^2) #average variance of
difference between genotypes
> model5blup<-lmer(yield~0+(1|gen)+rep +(1|rep:block), dat) #best linear
unbiased prediction (BLUP) for various effects in the model
> re5 <- lme4::ranef(model5blup,condVar=TRUE)
> vv1 <- attr(re5$gen,"postVar")
> vblup <- 2*mean(vv1)
> vblup
[1] 0.0577334
> sg2 <- c(lme4::VarCorr(model5blup)[["gen"]]) #h2 (heritability)
> h2<-sg2 / (sg2 + vblue/2)
>h2
[1] 0.8030173

10.9 Multiple Choice Questions

1. The structural variants associated with agronomic traits are:

(a) CNVs
(b) PAVs
(c) CNVs and PAVs
(d) None of the above
200 10 Agricultural Bioinformatics

2. The percentage of dispensable genes in maize is:

(a) 51%
(b) 61%
(c) 71%
(d) 81%
3. The phenotypic diversity of crop plants are due to the presence of:
(a) core genes
(b) dispensable genes
(c) pan-genes
(d) All the above
4. The amount of intersection in terms of presence and absence of a gene between
two genomes is measured by:
(a) Jaccard distance
(b) Manhattan distance
(c) Robinson–Foulds distance
(d) None of the above
5. The next-generation sequencing in crops is challenging due to:
(a) Large genome size
(b) Duplication
(c) Repeat content
(d) All the above
6. The percentage of transposable elements in maize crop is:
(a) 75%
(b) 85%
(c) 90%
(d) 95%
7. Homeologous genes in crops are the product of:
(a) Autopolyploidy
(b) Allopolyploidy
(c) All the above
(d) None of the above
8. The predicted breeding value of a parent crop consists of:
(a) Genotype
(b) Phenotype
(c) Genotype of relatives
(d) All the above
9. The genomic estimated breeding values of an individual capture:
(a) Additive genetic variance
(b) Dominance variance
(c) Epistatic variance
(d) None of the above
Suggested Readings 201

10. The common predictive model for quantitative traits is:

(a) BLUP
(b) BLPP
(c) BLAP
(d) BPLU

Answer: 1. c 2. b 3. b 4. a 5. d 6. b 7. b 8. d 9. a 10. a

Summary
• The structural variants in crops such as copy number variants (CNVs) and
presence/absence variants (PAVs) are crucial determinants of useful agronomical
traits.
• The percentage of dispensable genes in a crop species reflects the diversity among
individuals of a particular species.
• Structural variation is a major causative factor in generating dispensable genes in
crops.
• The dispensable genome of a crop is enriched with genes associated with biotic
and abiotic stress resistance.
• The second-generation deep sequencing of crops is challenging due to large
genome size, duplications and repeat contents.
• The homeologous genes in plants are the products of allopolyploidy.
• Genomic selection has a high accuracy of genomic-enabled prediction for simple
traits with high heritability.
• The most common predictive models for quantitative traits are the genomic best
linear unbiased prediction (G-BLUP) and the ridge-regression BLUP
(RR-BLUP).
• The crop phenomics deals with collection of multi-dimensional phenotypic data
of crops at various scales using automated platforms.
• The dynamic system simulation models are widely used to understand crop and
farming system models in response to certain external changes.

Suggested Readings
Wallach D, Makowski D, Jones J, Brun F (2013) Working with dynamic crop models. Academic
Press, Amsterdam
Isik F, Holland J, Maltecca C (2017) Genetic data analysis for plant and animal breeding. Springer,
Cham
Normanly J (2012) High-throughput phenotyping in plants: methods and protocols. Humana Press,
Totowa
Busch W (2017) Plant genomics: methods and protocols. Humana Press, Totowa
Farm Animal Informatics
11

Learning Objectives
You will be able to understand the following after reading this chapter:

• Role of bioinformatics in animal husbandry and veterinary science.

• Modelling of farm animal systems.
• Nutritional models of important farm animals.
• Computational model of laying hen.
• Lactation model of a dairy cow.
• Livestock genomics.
• Livestock phenomics.
• Methods of genomic selection for improvement of farm animals.

11.1 Introduction

Farm animal production system is a complex process consisting of three interacting

components: biology of the animal, environment and management practices. Thus
mathematical modelling of this system using automated real-time big data may
provide us new clues for enhanced efficiency in animal productivity. With the
advanced sensor technology, a large amount of data on individual animals can be
generated in a short span of time. The recent advancements in big data analytics and
precision agriculture, new algorithms and data structures are needed to develop next-
generation farm animal production models taking advantage of sensor data and
artificial intelligence. Mathematical modelling of this complex farm system has
been done from individual farm animals including many ruminant and
non-ruminant animals to entire farming system. There are two kinds of common
modelling approaches in the animal production system. The first model is an

# Springer Nature Singapore Pte Ltd. 2022 203

B. K. Tiwary, Bioinformatics and Computational Biology,
https://doi.org/10.1007/978-981-16-4241-8_11
204 11 Farm Animal Informatics

empirical model which is derived from experimental data to infer relationships

among different components at a single level. For example, an empirical model
can describe the relationship of enteric methane production with other explanatory
variable such as daily feed intake and level of milk production. Moreover, these
empirical models are merely based on observation and experiment but do not explain
the underlying biological processes. With the advent of big data and advanced
analytics methods in the animal production system, empirical methods can be a
powerful tool to understand the physiological and metabolic processes involved in
an animal production system. On the other hand, the mechanistic model is a process-
based model with several individual components and their specific interactions with
explaining the underlying structure of the system. A dynamic mechanistic model of
rumen physiology, for example, consists of 19 state variables such as microbial
biomass, carbohydrates, and fatty acids and is driven by inputs of various nutrients.
These models need meticulous construction and in turn provide accurate estimates of
system functions. Even, the system functions can be further extrapolated beyond the
data points used for the model construction. Mechanistic models have been proved
as a better model than the empirical model in predicting methane emission in cattle.
However, this modelling approach is usually improved by incorporating other
variables such as genetics, behaviour, environment and management of individual
animals.

11.2 Whole-Farm Systems Model

The whole-farm systems model is an interdisciplinary collaborative effort between

experimental biologists and computational biologists. The model development can
be developed in five steps which are as follows:

1. Description of the underlying biological system using existing models and

literature.
2. Information flowcharts and development of pseudocode.
3. Translation of pseudocode to codes by the programmers.
4. Submodel functions are evaluated by existing experimental datasets.
5. Submodules are incorporated in larger modules and in turn modules are
incorporated into systems model.
6. Evaluation of systems model and outputs for user application in the animal farm.

Thus, this modular format consists of several modules and submodules with clear
documentation, readable and readily adaptable in a new setup. Individual modules
can be adapted and improved further in a particular farm condition without
disrupting other modules of a system. For example, a whole-farm dairy systems
model consists of four integrated biophysical modules of Animal, Manure, Soil and
Crop, and Feed storage. In addition, there are three system balance modules of water,
energy and economics. The Animal module consists of three primary submodules:
Animal Life cycle, Nutrition & Production and Management & Facilities (Fig. 11.1).
11.3 Nutritional Models of Farm Animals 205

Fig. 11.1 Schematic diagram of an animal module with three submodules

The Animal Life Cycle submodule is a stochastic Monte Carlo model of events in
life of a dairy cow from birth to culling or death. The Animal module can simulate
stochastically the daily growth, production and reproduction of individual cows
based on the inputs regarding feed, weather and management. The stochastic
modelling simulates the probabilistic distribution of events for each individual
cow in a herd accommodating interactions among cows. The Nutrition and Produc-
tion submodule can predict the optimal diet of a cow for a desired milk production
using linear programming. The Management and Facility submodule is a barn-level
simulation including interaction of human intervention and environment with the
cow herd. The Soil and Crop module simulates soil temperature, hydrology, nutrient
dynamics, pasture growth and animal feeding using various climate inputs.
Interestingly, the feed storage module simulates the nitrogen and carbon loss
during harvesting and storage of animal feeds. Finally, the Water Balance module is
an all-encompassing module gathering information from four biophysical modules
with water use and generation in cattle management.

11.3 Nutritional Models of Farm Animals

11.3.1 Dairy Cattle Model

The efficiency and robustness of farm animals has been a focus of attention in the
field of animal nutrition. Productive animals are known to have better feed efficiency
and as a result have margin over feed cost. Although selection can improve signifi-
cantly the genotype of a farm animal in respect of some selected traits, other
non-selected traits remain prone to be fragile under fluctuating environmental
perturbations. The efficiency of a system is defined as the ratio between the fluxes
of outputs and inputs. For example, feed efficiency of a cow refers to production of
206 11 Farm Animal Informatics

milk and corresponding supply of nutrients. Robustness is the ability of a living

system to maintain its life trajectories under the influence of external and internal
perturbations. The efficiency and robustness of a ruminant animal is regulated by
two types of interacting regulations: homeorhetic regulations and homeostatic
regulations. Homeorhetic regulations control the basic life processes associated
with reproduction and growth of an animal such as gestation and lactation period
of a cow. On the other hand, homeostatic regulations operate on the adaptive
changes in animals under altered nutritional environment. The intake of dry matter
and feeding behaviour of animal was studied using regression models to assess the
individual variations in feed intake or residual feed. Most of the mechanistic models
developed, so far, have been focussed on the rate of feed intake and bite mass. Feed
cost constitutes about 60–70% of the overall production costs in cattle production.
The rumen provides a complex ecosystem for the rapid growth and proliferation
of billions of microbes. Many mechanistic models of ruminant digestion have been
developed to simulate the major alterations in digestive efficiency. An empirical
model of digestion showed that endogenous protein faecal losses is predominant in
dairy cows in comparison to basic maintenance requirements in terms of body
protein turnover. Another model measures the non-productive energy losses in
dairy cows which adversely affect the milk yield. The rumen protein balance is the
difference between the crude protein fluxes ingested and delivered to the duodenum.
This balance reflects an equilibrium state between the recycling of nitrogen both
through rumen wall and saliva and absorption of excess ammonia in the rumen. The
variations in rumen protein balance reflect the overall protein efficiency of the
individual animal. However, the genetic control of these variations and the range
of individual variation are yet to be known.
The efficiency of a farm animal is measured experimentally by the metabolic
partitioning of the total energy between maintenance energy and other energy
expenditures. For example, in a milk producing animal like dairy cow or goat,
partition of carbon occurs in favour of milk and in disfavour of carbon related to
carbon dioxide, faecal, urinary and methane. Thus mechanistic modelling of farm
animals needs a multidisciplinary team consisting of computational biologists,
nutritionists, geneticists, reproductive physiologists and ethologists.

11.3.2 Pig Model

The precision livestock farming is a management principle based on the processes

and technology of process engineering. The precision feeding is a part of this system
and requires proper amount of feed with appropriate composition in a timely manner
to an individual animal to enhance overall farm productivity. Like cattle farming, pig
farming has a major share of feed cost. But, the efficiency of conversion of nutrients
to animal product is very low in farm animals. For example, the nitrogen (protein)
retention efficiency varies from 15% to 33% in growing pigs. The nitrogen which is
not retained by the animal is excreted in the external environment and causes nitrate
pollution resulting in algal blooms in the water bodies. The tailoring of diet for
11.3 Nutritional Models of Farm Animals 207

individual animals as per production objective is a key principle in precision feeding.

The precision feeding of pigs can be implemented by collecting data automatically
from individual animals, data processing and actions required to control the system.
The feed intake and body weight data of growing pigs is estimated on real-time basis
for individual animals. A mathematical model can be developed to estimate the real-
time pig individual nutrients requirement. For example, the daily concentration of
lysine in the diet is estimated from the feed intake and body weight of an individual
pig. This data is used to develop both empirical and mechanistic models. Empirical
model predicts the expected feed intake, weight gain and body weight for the
subsequent day, whereas mechanistic model computes optimal concentration of
lysine in feed each day using all three variables estimated using empirical model.
Based on the results of mathematical modelling, right amount and composition of
feed is provided to each growing pig for optimum productivity.

11.3.3 Sheep and Goat Model

Although sheep and goat are small ruminants, different feeding strategies are needed
in these small ruminants due to their specific physiological requirements. For
example, the primary production of dairy sheep is wool in addition to milk produc-
tion. Sheep and goat eat more food as a percentage of body weight for regular
maintenance. Thus, a high producing dairy sheep and goat have a feed intake of 4%
to 7% of body weight but it never exceeds 4% of body weight in case of dairy cow.
Production efficiency of an animal is maximization of its product such as milk or
meat in comparison to inputs used and usually represented by the feed efficiency.
Feed conversion ratio is generally reciprocal of the feed efficiency. The higher value
of feed efficiency is desirable for an animal. On the other hand, lower feed conver-
sion efficiency is preferred during an animal production process. Various nutritional
models have been developed to enhance the production efficiency of goats and
sheep. The current nutritional models are more comprehensive having many mecha-
nistic components such as animal, dietary and environmental variables. The energy
and nutrient requirement of sheep and goats can be predicted more accurately under
diverse environmental and management practices. These nutrient models can be
further improved by real-time monitoring of small ruminants along with environ-
mental and production variables using modern sensor technology.

11.3.4 Laying Hen Model

A laying hen has a potential to produce at least 300 eggs annually. The weight of an
egg is determined not only by the age and genetic potential but on the nutrients fed
during laying period. The amino acid and energy requirement of a laying hen is
necessary to predict the potential body weight at first laying and subsequent potential
egg output over a time period. Energy constitutes about 70% of the costs incurred on
the poultry feed. Methionine is the primary limiting amino acid in the feed of a
208 11 Farm Animal Informatics

Fig. 11.2 Determination of optimum economic nutrient level for laying hen using optimization

laying hen. The age and body weight of an individual first laying bird determines its
future laying performance in terms of egg number and egg weight. This characteris-
tic feature of growth can be manipulated using different lighting and nutritional
regimes. The change in photoperiod especially the initial and final photoperiod has a
strong influence on the gonadal maturity of a laying hen. The daily intake of amino
acids and energy is largely used by a laying hen for maintenance. The body protein
content of a laying hen is found to be maximal at the age of sexual maturity and
remains comparatively stable throughout the entire laying period. A mathematical
model can predict food intake of a hen based on its body protein weight and potential
daily egg output. The model computes the energy requirements of maintenance and
egg output. However, the potential growth and egg output differ markedly between
growing and reproducing birds. An individual hen responds linearly to the increas-
ing amount of a limiting nutrient to its maximum genetic potential. The response to
nutrients are represented in terms of an economically important output such as egg
output. However, such responses are generally curvilinear when applied to a group
of birds. A simulation model can provide the answer for optimum economic nutrient
level for a group of laying birds considering marginal costs incurred and revenue
generated (Fig. 11.2). First, a feed formulation is passed to a laying hen model to
simulate the performance of hen. The costs and revenues are also calculated. During
the optimization process, feed formulations are altered following certain rules. This
optimization process iterates several times till it improves the value of objective
function using linear programming. Thus, egg producers can maximize their profits
by relying on this optimization process. In addition, it excludes the necessity for an
expensive and tedious long-term laying trial to measure the response of laying hens
to various feeds with differential nutrients.
11.3 Nutritional Models of Farm Animals 209

11.3.5 Lactation Model

Lactation in a cow is a product of specialized epithelial cells of the mammary gland.

These cells undergo variation in number and their secretory activity. The prolifera-
tion of cells starts at very beginning of pregnancy and reaches its maximum before
parturition. The shape of a lactation curve is characterized by an ascending phase
after parturition reaching its maximum peak followed by a declining slope till the
dry-off of the dairy cattle (Fig. 11.3). A lactation curve in dairy cattle consists of two
distinct phases: first phase of increasing milk production up to the peak level
followed by second phase of decline of milk production. Primarily, lactation is a
deterministic, continuous and regular process. However, stochastic components of
lactation curve are genetic variability among animals and their health status, feeding
and environmental practices, and overall farming system. The mathematical model
of lactation curve is described by analytical function of time, where yt is the daily
milk production recorded at time t:

yt ¼ f ð t Þ

There are many empirical models used for fitting the lactation curve data. Early
models emphasized on the deterministic component of the lactation curve. The

Fig. 11.3 The lactation curve of a dairy cattle showing ascending, peak and declining phases
210 11 Farm Animal Informatics

Wood incomplete gamma function is a popular empirical model of the lactation

curve. The function of this model is yt ¼ atbect with three parameters a (a scale tends
to increase with parity), b (controls the rate of increase to lactation peak) and c (rate
of decline after the peak).

11.4 Livestock Genomics

A high-quality annotated reference genome of a farm animal is not only prerequisite

for discovery and analysis of genetic variation but also useful in connecting geno-
type to function. The discovery of molecular genetic variants in the genome
sequences and subsequent development of single-nucleotide polymorphism (SNP)
chips are necessary to understand the genetic basis of complex traits such as growth,
body composition, feed conversion, reproduction and response to a microbial
infection. The genomes of farm animals have undergone a change in terms of
removal of disease genes under the influence of purifying selection during succes-
sive selective breeding of superior animals. The genomics-enabled genetic improve-
ment of farm animals is known as genomic selection. An overview of genomic
resources available for farm animals is listed in Table 11.1. The reference sequence
of majority of farm animals can be accessed on the Ensembl genomic database.
Chicken is an important food animal constituting 41% of meat produced in the
world. The draft genome of chicken was released in 2011 by the International
Chicken Genome Consortium. It consists of 31 chromosomes and 14,093 unplaced
scaffolds. The polymorphisms related to quantitative traits present in the chicken
genome can be investigated and analysed for directed evolution of superior breeds.
A draft reference genome of pig was developed in 2017 by the Swine Genome
Sequencing consortium. It consists of 20 chromosomes including two sex
chromosomes along with 583 unplaced scaffolds. Since a reference genome does
not represent complete genetic diversity in the pig genome due to divergence in
different geographic regions, genome sequences of 12 pig breeds were also
sequenced. The bovine genome consists of 30 chromosomes built from 2211

Table 11.1 Genomic databases of farm animals

Database Description URL
Ensembl genome Genomic database of all farm animals www.ensembl.org
browser
National Animal A comprehensive resource of USDA www.
Genome Research containing animal quantitative trait loci animalgenome.org
Program database (AnimalQTLdb) and animal trait
correlation database (corrDB)
Bovine genome A database of bovine genome based on bovinegenome.
database Hereford cow. elsiklab.missouri.
edu
Chickspress A database for chicken gene expression http://geneatlas.arl.
arizona.edu
11.5 Livestock Phenomics 211

scaffolds. Bovine Genome Database (BGD) is a web-based resource providing

access to bovine genome assemblies and annotations. The domestic goat genome
sequence was developed by the Bangladesh Goat Genome Consortium in 2019. It
consists of 29 chromosomes assembled from 3972 scaffolds. The reference genome
of sheep consists of 27 chromosomes assembled from 2641 scaffolds.
AnimalQTLdb is a database having QTL, candidate gene, GWAS data and copy
number variations (CNV) mapped to various farm animal genomes. Online Mende-
lian Inheritance in Animals (OMIA) has a manually curated collection of inherited
disorders and genes in various species of farm animals.

11.5 Livestock Phenomics

The phenotype is the measurement of some features in an animal and has been a
basis of all genetic improvement. High-throughput phenotyping is very crucial in
closing the gap between genotype and phenotype. The real-time acquisition of high
dimensional phenotypic data such as physiological or behavioural associated with
production on individual animal scale is known as livestock phenomics. The live-
stock phenomics is more challenging than the plant phenomics because animals have
a longer generation interval than the crops and they change their location frequently
unlike crops in a field. Phenomics data can be measured at different levels from
molecular level to morphological level. Molecular measurements include transcripts,
proteins and metabolites expressed in different cells or tissues at different time
points. Morphometric, behavioural and physiological data are other higher levels
of phenomics data. Recent developments in sensor technology have given us
opportunity to measure difficult or previously unmeasurable phenotyping traits in
farm animals. The knowledge of most appropriate phenotype such as feed efficiency
in a grazing livestock system will give deep insights into the biology involved in
manifestation of this phenotype. For example, interactions between genotypes and
environment and pleiotropic effects of genes can be well understood using
phenomics approaches. A sensor mounted on an animal allows comprehensive
two- and three-dimensional measurements or images of animal behaviour in a
pasture. Physiological traits such as body temperature, heart rate, respiratory rate
and rumen function can be monitored at a real-time basis on each individual animal.
Sometimes, radiant temperature in the farm environment are monitored using ther-
mal cameras. Many phenotypes of economic importance in animals such as feed
conversion efficiency, disease resistance and reproductive potential are extremely
difficult to measure in a large pastoral environment. These complex traits are derived
from individual components and their interactions. The rapid generation of multi-
sensor high-throughput phenotypic data in an animal farm over time poses a grand
challenge in big data management. Further, the computational analysis of these big
phenotypic data for biological interpretation is necessary to enhance the farm
productivity. In dairy industry, the performance of a Holstein breed animal in the
USA is measured by 42 traits of economic importance. The success of a genomic
selection program solely depends upon large number of phenotypic measurements.
212 11 Farm Animal Informatics

11.6 Genomic Selection in Farm Animals

Recent advances in genome sequencing technology has resulted in availability of

reference genome sequences for most livestock species. The single base pair
variations of individual animals from the reference genome called single-nucleotide
polymorphisms (SNPs) are genotyped as genetic markers using SNP-chip based
genotyping technology. SNPs are abundant markers usually located outside the
genic regions across the entire genome. They are linked to the genes or genic regions
known as quantitative trait loci (QTL) through linkage disequilibrium (LD). The LD
measures the non-random association between the alleles across loci and is a ratio
between expected and observed allele frequencies. In fact, it reflects the physical
distance between two loci; close distance between loci indicates a strong LD
justifying use of one locus as a proxy for the other one. The LD is strong for dairy
cattle due to small effective population size (Ne) and often broken down by recom-
bination events. A bovine genome contains almost 3 billion nucleotides with 30 mil-
lion SNPs. Thus, one SNP is present for every 100 nucleotides in the bovine
genome. Each low-cost chip has almost 50,000 genome-wide SNPs for cattle, pig,
sheep and chickens. The genomic selection is implemented on selection candidates
by measuring the traits and genotyping the markers in a training population. The
genomic selection has been adopted extensively for dairy cattle breeding. The
genomic estimated breeding values (GEBV) are predicted in selection candidate
animals from a prediction model. The genomic selection is based on the LD between
SNP and QTL (Fig. 11.4). Since we do not observe the direct association between
QTL and the phenotype, the effectiveness of genomic selection is predicted based on
how much variance of the trait is explained by the SNP. There are two classes of
genomic selection methods: SNP effect-based method and genomic

Fig. 11.4 Genomic selection in farm animals is based on linkage disequilibrium (LD) between
single-nucleotide polymorphism (SNP) and quantitative trait loci (QTL)
11.7 Exercises 213

relationship-based method. The RR-BLUP and SNP-BLUP are popular SNP-effect-

based methods. Genomic BLUP (GBLUP) is a common genomic-relationship
method widely used in animal breeding. The single-step genomic BLUP (ssGBLUP)
is another method which combines the phenotype, genotype and pedigree informa-
tion in single estimate. The accuracy of GEBV has been evaluated in cattle, sheep
and pig for many economically useful traits. The reliability of GEBV has a much
higher range of 44–49% for milk yield traits in dairy bulls when compared with
traditional selection (15–28%). There was also an improvement in accuracies of
GEBV for various traits in dairy, meat and wool sheep in comparison to pedigree-
based selection. However, accuracy of GEBV for growth and carcass traits improved
up to 0.42 and 0.65 respectively in beef cattle. The economic traits like feed
conversion ratio were improved significantly in pigs using 60,000 SNPs during
genomic selection. The genomic selection is highly useful for those traits expressing
after attaining a minimum age for breeding or expressing in one sex. For example,
milk yield is a sex-limited trait which only expresses in females.

11.7 Exercises

1. A model simulating the dynamics of lactating mammary glands of cattle is

available in the R package ZeBook. This model is explained by six state
variables, namely H (changes in level of hormones), CS (production and loss of
milk secreting cells), M (removing the secretion of milk), Mmean (the average
quantity of milk in the animal), RM (the amount of milk removed) and Y (yield of
the milk). Find the fourteen parameters involved in the production process and
simulate the model with a time step dt ¼ 0.1 assuming regular consumption of
milk by a calf. Find the weekly changes in the state variables of the milk
production model.

Solution (Fig. 11.5)

> library(ZeBook)
> lactation.define.param(type = "calf")
cu kdiv kdl kdh km ksl kr ks ksm mh mm p mum rc
nominal 1000 0.2 4.43 0.01 0.005 3 0.048 0.1 0.2 27 30 10 1 40
binf NA NA NA NA NA NA NA NA NA NA NA NA NA NA
bsup NA NA NA NA NA NA NA NA NA NA NA NA NA NA

> X<-lactation.calf.model2(lactation.define.param()["nominal",],300,0.1)
>X
214 11 Farm Animal Informatics

Fig. 11.5 Weekly changes in state variables during lactation

11.7 Exercises 215

2. An R package agridat has a dataset of mating crosses of chicken consisting of

45 observations on three variables, namely male parent, female parent and weight
of the progeny at 8 weeks. Plot the dataset using R package lattice. Estimate the
variance component fitting linear mixed model and heritability of both parents as
well as combined heritability.

Solution (Fig. 11.6)

> library(agridat)
>library(lattice)
> data(becker.chicken)
> chickdata <- becker.chicken
>dotplot(weight ~ female, data=chickdata, group=male,
main="Body weight at 8 weeks", xlab="Female dams",ylab="Progeny
weight",pch=16,auto.key=list(columns=5))

Fig. 11.6 Body weights of chicken at weight weeks linked with male and female parents
216 11 Farm Animal Informatics

#Estimation of variance component fitting linear mixed model#

> library(lme4)
> model <- lmer(weight ~ (1|male) + (1|female), data=chickdata) # a model
that incorporates both fixed- and random-effects term
> model
Linear mixed model fit by REML ['lmerMod']
Formula: weight ~ (1 | male) + (1 | female)
Data: chickdata
REML criterion at convergence: 516.6906
Random effects:
Groups Name Std.Dev.
female (Intercept) 33.10
male (Intercept) 27.87
Residual 74.33
Number of obs: 45, groups: female, 15; male, 5
Fixed Effects:
(Intercept)
808.5
#Extract the variance components from a fitted model#
>library(lucid)
> vc(model)
grp var1 var2 vcov sdcor
female (Intercept) <NA> 1096 33.1
male (Intercept) <NA> 776.8 27.87
Residual <NA> <NA> 5524 74.33
> Male.var <- 776
> Female.var <- 1095
> Withincross.var<- 5524
> Vp <- Male.var+ Female.var + Withincross.var
> 4*Male.var/Vp
[1] 0.4197431 # Heritability of male sires
> 4*Female.var/Vp
[1] 0.5922921 # Heritability of female dams
>2*(Male.var+Female.var)/Vp
[1] 0.5060176 # combined heritability

11.8 Multiple Choice Questions

1. The farm animal production system consists of:

(a) Biology of farm animal
(b) Environment of animal
11.8 Multiple Choice Questions 217

(c) Management of the farm

(d) All the above
2. The modelling of experimental data in farm animal system is known as:
(a) Mechanistic model
(b) Empirical model
(c) Data driven model
(d) None of the above
3. The adaptation of farm animal to a new nutritional regime is under control of:
(a) Homeorhetic regulation
(b) Homeostatic regulation
(c) Haemorrhagic regulation
(d) Nutritional regulation
4. The gonadal maturity of a laying hen is primarily determined by:
(a) Feeding regime
(b) Genetic merit
(c) Photoperiod
(d) Temperature
5. The optimization of laying hen model is performed by:
(a) Non-linear programming
(b) Linear programming
(c) Exponential programming
(d) Functional programming
6. The stochastic component (s) of a lactation curve is/are:
(a) Genetic variability among dairy cattle
(b) Health status of dairy cattle
(c) Feeding practices
(d) All the above
7. Linkage disequilibrium measures:
(a) Random association between the alleles at different loci
(b) Stochastic association between alleles at different loci
(c) Non-random association between alleles at different loci
(d) None of the above
8. A bovine genome contains:
(a) ten million SNPs
(b) 20 million SNPs
(c) 30 million SNPs
(d) 40 million SNPs
9. The genomic selection method (s) implemented in dairy cattle is/are:
(a) SNP-BLUP
(b) RR-BLUP
(c) G-BLUP
(d) All the above
10. The phenomics data of a dairy cattle consist of:
(a) Morphometric measures
(b) Physiological measures
218 11 Farm Animal Informatics

(c) Behavioural measures

(d) All the above

Answer: 1. d 2. b 3. b 4. c 5. b 6. d 7. c 8. c 9. d 10. d

Summary
• Empirical model is derived from experimental data to infer relationships among
different components at a single level.
• Mechanistic model is a process-based model with several individual components
and their specific interactions.
• Homeorhetic regulations control the basic life processes associated with repro-
duction and growth of an animal.
• Homeostatic regulations operate on the adaptive changes in animals under altered
nutritional environment.
• A mathematical model can predict food intake of a hen based on its body protein
weight and potential daily egg output.
• The shape of a lactation curve is characterized by an ascending phase after
parturition reaching its maximum peak followed by a declining slope till the
dry-off of the dairy cattle.
• Bovine Genome Database (BGD) is a web-based resource providing access to
bovine genome assemblies and annotations.
• Livestock phenomics data can be measured at different levels from molecular
level to morphological level.
• The genomic selection is based on the LD between SNP and QTL.
• Genomic BLUP (GBLUP) is a common genomic-relationship method widely
used in animal breeding.

Suggested Reading
Khatib H (2015) Molecular and quantitative animal genetics. Wiley-Blackwell, Hoboken
Mondal S, Singh RL (2020) Advances in animal genomics. Academic Press, London
Isik F, Holland J, Maltecca C (2017) Genetic data analysis for plant and animal breeding. Springer,
Cham
Mrode RA (2014) Linear models for the prediction of animal breeding values, 3rd edn. CABI
Publishing, Wallingford
Computational Bioengineering
12

Learning Objectives
You will be able to understand the following after reading this chapter:

• Application of bioinformatics in bioengineering.

• Control and systems theory in biological systems.
• Metabolic engineering of biological systems.
• Evolutionary engineering of microbes and directed evolution of proteins.
• Principles and methods of synthetic biology.

12.1 Introduction

Cell is a biochemical factory consisting of small molecules involved in the signalling

and energy transfer processes. The availability of high-throughput data and the
advent of systems approach in biology have paved the way for understanding
complex dynamic interactions among molecules in the cell. Thus, engineering
approaches are necessary to understand the overall functioning of the cell systems.
A cell can be perceived as a dynamic system characterized by a set of state variables.
The transcriptome or proteome at a particular time point represents state variables of
a cell. Moreover, a cell can be described as an input/output (I/O) system where an
input may be a hormone or drug and a secreted metabolite may represent a measur-
able output. Similar to engineering systems, cell consists of subsystems with differ-
ent functions such as cell growth, cell division and apoptosis. In addition to being a
dynamical system, a cell can also be treated as a controlled environment where all
state variables are interconnected to maintain the robustness of the cell. Thus,
computational model of the cell using high-throughput data will help in unravelling
the control mechanisms in cell regulating its specific functions.

# Springer Nature Singapore Pte Ltd. 2022 219

B. K. Tiwary, Bioinformatics and Computational Biology,
https://doi.org/10.1007/978-981-16-4241-8_12
220 12 Computational Bioengineering

12.2 Control and Systems Theory in Biology

In engineering, a control system consists of interconnected components acting in

concert to produce a desired behaviour in presence of multiple external
perturbations. Majority of biological systems like their engineering counterparts
are controlled by closed loop. The control system of human thermoregulation is
illustrated in Fig. 12.1 as an example of closed loop. Here, the output value of the
system is fed back in the system in order to adjust the input value and is known as
feedback control. The control system maintains the time dynamics of a particular
control variable in a system with a predetermined reference variable in spite of
regular external disturbances. A feedback sensor measures the output and subse-
quently compares this feedback signal with the desired reference value. Finally, the
error or deviation obtained between two values is used to compute control variable.
The feedback signal is subtracted from the input reference value in the negative
feedback control. Here, the controller tends to increase the output in case resulting
error is positive, whereas it always decreases the output if resulting error is negative.
In contract, the feedback signal is added to the input reference value in the positive
feedback control. The maintenance of constant body temperature in human is an
example of negative feedback control in a biological system. On the other hand, the
propagation of action potential in neuron is regulated by the positive feedback
control.
Control theory is an integral part of a dynamical system. The interaction of a
dynamical system with surrounding environment occurs through two vectors of
time-dependent variables: input variables and output variables. The knowledge of
an input variable is not enough to describe the output variables in a dynamical
system; a minimum state of variables known as state variables needs to be
characterized to predict the future output of a dynamical system. Thus, the knowl-
edge of both input variables and state variables is necessary to predict the overall
output of a system. Most of the biological systems like artificial systems are
non-linear dynamical systems. Dynamical systems are modelled computationally

Fig. 12.1 The control system of human thermoregulation using closed-loop feedback
12.4 Metabolic Engineering 221

either using two deterministic approaches: continuous time series and integer time
points. State variables represent the real value of the time variable (t) in a continuous-
time system, whereas integer time points (i.e. t ¼ 1, 2, 3...) define a discrete-time
system. However, deterministic models, although useful, often fail to capture the
inherent complexity of a biological system. Therefore, stochastic models like hidden
Markov models are developed to represent a biological system.

12.3 Strategies in Bioengineering

There is a huge demand for bio-based products such as biofuels, solvents, organic
acids, polymers, and food supplements across the world. In order to produce them in
large quantity at low cost, three performance metrics of industrial production,
namely titre (the final concentration of the product after bioprocess), yield and
productivity, should be very high. The minimum target for titre is about
100 grams per litre in most cases but a higher target of more than 200 grams per
litre is feasible in some cases. The yield of a product is defined as mole or gram of the
product obtained from mole or gram of the consumed substrate. On the other hand,
productivity is defined in two forms: specific productivity and volumetric productiv-
ity. The specific productivity is the amount of the product produced in terms of mole
or gram per cell per unit time, whereas the volumetric productivity is the amount of
product produced per volume per unit time. The importance of any parameter in
decision-making depends on the nature of bio-product and the bioprocess. A high
titre is not only a useful parameter for cost-effectiveness, it also facilitates separation
and purification of the product. A higher yield of a product is vital for production of
bulk chemicals as the major cost is incurred in procuring carbon substrates. Produc-
tivity covers the overall production cost incurred in a bioprocess such as fermenter
size and equipment depreciation cost per year.

12.4 Metabolic Engineering

Biological systems produce a wide range of natural products from precursor

molecules available in the surrounding environment. These natural products can
be synthesized in large quantities by culturing biological organisms in artificial
culture media. However, in majority of cases, the actual yield of the metabolic
product is very low resulting in high cost of industrial production. Thus, the primary
goal of metabolic engineering is to reduce this cost substantially by producing large
quantities of useful metabolites such as pharmaceuticals, enzymes and antibodies. A
biological system is akin to a complex circuit composed of specific interactions
among various biomolecules such as DNA, RNA, proteins and metabolites. The
common computational software for metabolic engineering are listed in Table 12.1.
The modularity in a biological system not only prevails at the circuit level but
manifests at each level of biological complexity. Biological pathways can be
separated into a series of different parts coordinating a particular biological function.
222 12 Computational Bioengineering

Table 12.1 Common computational tools in metabolic engineering

Computational
tool Description
BNICE Discovery of novel metabolic pathways based on generalized reaction rules
RetroPath A tool for pathway prediction based on the generalized reaction rules
CAMEO A tool for in silico modelling including algorithms for detecting gene
knockout and overexpression target

Fig. 12.2 A simple

representation of a metabolic
network. X is a precursor
molecule transported inside a
cell via active transport. X is
converted to Y and Z through
enzymatic reaction catalysed
by E1 and E2, respectively

A simple biological module can be represented as promoter with its expressed gene
and regulatory proteins with its DNA binding sites. These simple modules in turn
give rise to a gene regulatory network which produces an output signal after complex
computation in a biological organism when challenged by various inputs. This
computation is based on various engineering principles such as positive feedback,
negative feedback and feed-forward loop.
The performance of a biological system can be predicted in silico using compu-
tational modelling. First, the possible enzyme pathways involved in the biosynthesis
of a desired metabolite are identified. A set of reactions in individual paths are
stoichiometrically balanced and added to get the net chemical reaction for a path. A
metabolic engineer redesigns the existing network pathways in order to change the
metabolic flux rates. If one wishes to increase the production of a particular metabo-
lite, an attempt will be made to increase the flux of the metabolite from its precursor
without affecting the remainder of the metabolic pathways. The membrane flux of a
precursor molecule in a cell is increased by increasing the concentration of that
molecule in culture media in case of passive diffusion. In contrast, the concentration
or efficiency of a transporter protein is increased for active transport of precursor
across the cell membrane. The precursor molecule in the cell is converted to a
desired metabolite through multi-step enzymatic reactions. For an enzymatic reac-
tion, if substrate concentration is much lower than the Michaelis constant, the
reaction flux will be proportional to the substrate concentration. For example,
consider an example where a precursor molecule X enters a cell through transport
process and is converted to either Y or Z inside a cell via enzymatic reactions as
illustrated in Fig. 12.2. If we want to overproduce metabolite Y, the mass flux from
X to Y needs to be increased through enhanced activity of enzyme catalysing this
reaction. This enhanced enzyme activity can be achieved by alteration of specific
activity of the enzyme (kcat / Km) or alteration of the concentration of the intracellular
12.5 Evolutionary Engineering 223

enzyme. However, this kind of engineering approach is too simplistic to be accepted

by nature. The modified cell is shifted away from optimal state achieved through
evolution and shows adverse effects of modifications in form of stunted growth and
reduced robustness. There is no specific single rate limiting step in a metabolic
network, therefore, perturbations are needed at multiple steps to achieve the desired
output of a metabolite. This kind of engineered changes generally are not accepted
by the regulatory network of a cell. Sometimes, the enzyme synthesis/activity is
repressed by the product metabolite; otherwise, the product metabolite potentiates
the competing enzyme activity. Therefore, the best strategy for metabolic engineer-
ing is to perform minimum changes in the overall metabolic network of the cell so
that the resulting cell is akin to the unengineered system already optimized for
growth and reproduction under evolutionary selective processes. Typically, only
fluxes to a desired metabolite product are changed keeping the concentrations of
other intermediate metabolites unchanged. However, this kind of differential flux
balance is difficult to achieve in practice.
The metabolic systems are used as constraints to optimize an objective function,
for example, how to maximize the biomass of cells? The prediction of flux balance
under a primary objective function will predict the changes in a secondary objective
such as metabolite overproduction after perturbations in the metabolic network. The
genetic changes in the network such as addition of a gene or deletion of a gene can be
tested computationally by suitable modifications in the stoichiometric matrix. The
optimal fluxes are recomputed in order to achieve a greater flux towards a desired
metabolite. Here, absolute growth rate may not be appropriate objective function;
instead, minimal flux disturbance criterion performs better as an alternative
objective.

12.5 Evolutionary Engineering

Evolutionary engineering is a discipline of engineering to develop phenotypes of

microbial strains or a biological component like a protein molecule simulating
natural evolutionary processes. Evolutionary engineering operating on the cells for
improvement of strains is known as adaptive laboratory evolution (ALE). When
evolutionary engineering is applied on a protein molecule for better efficiency such
as enhanced catalytic activity, stability or tolerance to high temperature, it is known
as directed evolution. In both processes, first variants are randomly generated using
natural evolution and then desired protein or cells are selected from large number of
naturally derived variants. In ALE, microbial cells are cultured in defined culture
media under controlled conditions for long period of time (100–2000 generations) in
order to achieve improved phenotypes. Most common mutations observed during
adaptive laboratory evolution are single-nucleotide polymorphisms (SNPs), small
insertions and deletions (indels), transposition and deletion of large genomic regions
contributing towards genetic and gene regulatory changes in the improved strains.
The increased fitness of improved strain in culture is manifested by increasing
frequency of this strain among total population. Microbial cells are exposed to
224 12 Computational Bioengineering

various nutrients such as glycerol and glucose in chemostat cultures for improved
growth. Similarly, the effects of environmental stresses such as high temperature and
osmotic pressure or tolerance to high ethanol concentration on the adaptive labora-
tory evolution of microbial cells have tremendous industrial applications. Computa-
tional models are useful in understanding the evolution of laboratory strains.
However, the major challenges are the high computational cost of simulation and
unknown functions of a significant proportions of genes (~30%) in a model species
like E. coli or S. cerevisiae. The effects of various environmental stressors on
microbial cell are also very difficult to integrate in the model. In spite of limitations,
computational models can play a significant role in improving laboratory strains
through adaptive laboratory evolution.
The directed evolution has been applied on the artificial evolution of enzymes to
catalyse the industrial reactions. Frances Arnold received the Nobel prize in chemis-
try in 2018 for her pioneering work on the directed evolution. Darwinian evolution
favours the fit organisms by accumulating beneficial mutations. In directed evolu-
tion, macromolecules are purposefully evolved towards new desired properties. The
evolutionary processes are implemented in three successive steps in directed evolu-
tion: mutation, selection and amplification. First, a natural macromolecule whose
properties are similar to the desired one is taken as a precursor molecule for
redesigning. Alternatively, a new macromolecule is created de novo from a collec-
tion of random sequences and the process is known as protein design. During
redesigning, catalytic activity of an enzyme can be transplanted on another enzyme
with same folds and mechanistic functions. Even a few mutations at the active site of
the enzyme facilitates emergence of novel functions. Computational redesign of an
extant protein is generally associated with the experimental evolution of the protein.
For example, the computational design starts with simulated docking of a target
ligand into the active site of a protein retrieved from the protein data bank (PDB).
The residues around this pocket are randomized and the design algorithm repeatedly
searches for a conformational space of the side chains and ligands at the minimum
energy level for each sequence. Finally, only a few proteins are experimentally
produced from a list of sequences with minimum energy predicted by the computa-
tional design algorithm. Novel receptors for proteins and small molecules have been
designed using this approach. A variety of experimental methods such as error-prone
PCR, degenerate codons, DNA shuffling and recombination are used in vitro to
create diversity in the molecule. Directed evolution is an established method for
enzyme engineering. The specificity, stability and reaction conditions of many
enzymes have been custom-made for commercial exploitation. Lipases, generated
through directed evolution, are produced commercially as an additive to laundry
detergents to break down lipids. An enzyme becomes several hundred or thousand
times more effective than the template enzyme after a few iterative cycles of directed
evolution.
12.6 Synthetic Biology 225

12.6 Synthetic Biology

Synthetic biology is an emerging discipline integrating engineering and computa-

tional biology. It mainly deals with either construction of new artificial system or
redesigning a natural biological system. The basic principle of this discipline is to
simplify a complex biological system into much simpler molecular factories
consisting of living parts and devices to produce a biological product. Biological
circuits are assembled from the genes and regulatory DNA (similar to gates in
electronic circuits) and the proteins are akin to the wires in the electronic circuits.
The parts, devices and systems are defined in the framework of synthetic biology
which is as follows:

Parts The parts or bioparts is a portion of DNA molecule encoding a specific

biological function. For example, promoter, ribosome binding site, protein coding
sequence and terminator sequence are four separate parts of a genetic circuit
producing a polypeptide. The standardized and interchangeable biological parts
are available in the Registry of Standard Biological Parts and Biobricks.

Devices Devices are made of different parts performing a user-defined function. For
example, a simple device protein generator consists of promoter, ribosome binding
site, protein coding sequence and terminator sequence. The devices have logic gates
having binary states 1 (condition satisfied) and 0 (condition not satisfied). A logic
gate takes multiple binary inputs and produces a single output in form of either
0 (OFF) or 1 (ON). There are different types of logic gates in a device such as AND
gate (all input conditions must be satisfied for ON state), OR gate (any input
condition must be satisfied for ON state), NOT gate (single input is inverted with
compatible promoter/repressor pair) and negated gate (a combination of NOT gate
with AND/OR gate).

Systems Simple systems are developed from one or more devices. The examples of
simple biological systems are feedback loops, genetic toggle switches, oscillators
and repressilators. The effect of output on the future behaviour of a system is known
as feedback effect. When the future output of a system is increased by the present
output, it is known as positive feedback. Conversely, when a present output
decreases the future output of the system, it is known as negative feedback.
Activators and repressors are used for creating positive and negative feedback
respectively in synthetic biology. The switch is a device which can be turned ON
or OFF in response to an input signal. Toggle switches are having only two stable
steady states ON and OFF without any intermediate state. Toggle switches in
synthetic biology are designed by integrating two repressors and two constitutive
promoters, where each promoter transcribe a repressor which in turn inhibit the
opposite promoter. An oscillator is a device showing cyclical pattern (oscillations) in
the output nearby an unstable steady state. There are two important elements in a
biological oscillator: an inhibitory feedback loop and a source of delay in the
feedback loop. A repressilator is a kind of oscillator made by combining multiple
226 12 Computational Bioengineering

inverters. For example, three genes may be combined in feedback loop in such a way
that each gene represses the subsequent gene in the loop and itself repressed by the
preceding gene.
The synthetic organisms assembled from genetic parts and devices do not always
exhibit predictable behaviour unlike its electronic counterparts due to stochastic
noise and uncertain environment in a new chassis. Sometimes, engineered cells
show low productivity due to metabolic burden that is the part of the cellular
resources used in engineering the metabolic pathways. The user-defined function
such as synthesis of a new product in an engineered cell may jeopardize the overall
metabolic robustness optimized by natural selection in an evolutionary timescale.

12.7 Computational Design of Synthetic Systems

The rapid development of synthetic biology is a result of collaborative efforts of both

computational and experimental biologists. The rational design of a synthetic system
needs a prior design and optimization of the computational model in standard system
before translating or compiling into DNA sequences in vitro or in vivo. Synthetic
Biology Open Language (SBOL) was developed as a standard format to support the
specification and exchange of information regarding biological design. Systems
Biology Markup language (SBML) can only represent a model of a biological
process but lacks information regarding associated nucleic acids and amino acid
sequences. SBOL framework have a modular and hierarchical graph based structural
and functional description of genetic design components, their sequences and inter-
component interactions. A design repository, SynBioHub, is an open-source com-
munity-based effort to share information regarding synthetic systems. It uses SBOL
as a core storage data format and allows a user to upload, browse and share designs
on the web interface. There is a harmonized data exchange between multiple existing
SynBioHub repositories. We can transform the textual information of a biological
design to visualization of the design using the web interface of a software tools like
Pigeon and TinkerCell. There are many computer-aided circuit design applications
for synthetic biology (Table 12.2). BioJADE was the first drag and drop JAVA-
based tool for circuit design which allows use of parts from BioBrick repositories.
The list of parts is visible on the workspace and they are connected through
hypothetical wires. Unlike electronic circuits, biological parts and devices are

Table 12.2 Computational tools for circuit design

Tool URL
BioJADE http://web.mit.edu/jagoler/www/biojade/
TinkerCell http://www.tinkercell.com/
ProMoT https://www2.mpi-magdeburg.mpg.de/projects/promot
GenoCAD https://design.genofab.com/
GEC https://omictools.com/gec-tool
Cello www.cellocad.org
12.7 Computational Design of Synthetic Systems 227

connected through a variety of signalling molecules. For example, RNA

polymerases and ribosomes act as common signalling carriers for exchanging
information between the parts. Thus, the biological current between the parts are
represented by their fluxes and measured in terms of Polymerase per second (PoPS)
and ribosomes per second (RiPS). BioJADE and TinkerCell consider only polymer-
ase as a carrier of signal between the parts. TinkerCell is capable of running both
stochastic and deterministic simulations in addition to other analyses such as flux
balance and sensitivity analysis. Asmparts is another command line tool which treats
each part as independent SBML module and can run model kinetics of transcription
and translation. Process Modelling Tool (ProMoT) is a graphical tool for process
engineering which has additional signal carriers like transcriptional factors,
chemicals and small RNAs in addition to polymerases and ribosomes. It is used
for modelling both kinetic models and Boolean models and resulted models can be
exported to SBML and MATLAB format. GenoCAD and GEC are two other tools
connected with databases for circuit design.
The hardware description language like Verilog allows an electronics engineer to
design an integrated circuit using textual commands. This computational design is
subsequently transformed into silicon circuits in an electronic system. This concept
is recently translated to genetic circuits where a Verilog design was performed in a
computational environment known as Cellular Logic (Cello) and subsequently
translated to a linear DNA sequence. Cello builds genetic circuits using transcrip-
tional gates with RNA polymerase as signal carrier and also simulates the perfor-
mance of the genetic circuit.
The designed circuits are optimized to evaluate the performance efficiency. Two
classes of optimization methods have been applied till date in synthetic biology:
stochastic and deterministic. Stochastic optimization process such as evolutionary
algorithm is computationally expensive and finds solution through random search. In
fact, it often fails to find global optimal solution. In contrast, deterministic methods
such as gradient descent have local search algorithms with lower computational cost.
It often fails to get many better solutions during the optimization process. Evolu-
tionary algorithm requires the initial definition of basic components, biochemical
reactions, ensemble of independent networks and objective function. The best
networks are obtained as a final solution after several rounds of iterations. Genetdes
is a stochastic method for finding a single solution from random circuit configuration
but often get stuck in local minima. Conversely, OptCircuit is a deterministic method
for finding a local optimal solution very close to the best one. However, these
methods can only be implemented on simplified models of small genetic circuits.
RoVerGeNe is a software specially designed to adjust the performance of a synthetic
network and to measure its robustness.
228 12 Computational Bioengineering

Box 12.1 Risks Associated with Synthetic Biology

A synthetic microbe, algae, plants or animals may become pathogenic to other
organisms in the ecosystem and degrade the natural environment. Further, it
may outcompete with wild native populations and adapt even in an adverse
environment. The modified algal strain may cause algal blooms that creates
harmful effect on human health. The ecological impact of intentional release or
accidental escape of a synthetic organism in the environment is a decline in
biodiversity and genomic diversity of a species. The horizontal gene transfer
of engineered genes from modified organisms to wild-type organisms may
have catastrophic consequences. The designer protein toxins developed using
protein engineering for therapeutic purpose may be used as a bioweapon for
military purpose or bioterrorism. A synthetic organism may act as an allergen
or generate an allergen protein to human. The products generated during
biofuel production may be carcinogenic to human and other animals.

Box 12.2 iGEM

The International Genetically Engineered Machine (iGEM) is a competition
for high-school, undergraduate and postgraduate students to develop a project
in synthetic biology. This program was started in 2005 by the MIT. Each team
consists of 5–15 students under supervision of a qualified researcher. The
newly registered team gets a set of thousand standard DNA parts stored in the
BioBrick to start their project. The best new synthetic biology ideas in
different category are rewarded in this competition. Many projects of the
iGEM competition have been published in reputed international journals like
Nature, Science and Cell. Some of the noted examples of the iGEM project are
bacterial photofilm producing light-induced pigments and different smells
produced by bacteria during exponential phase and stationary phase of their
growth.

Box 12.3 Engineered Riboswitches

Riboswitches are natural regulatory cis-acting RNAs that bind small
molecules in order to perform their regulatory functions on the target genes.
They have specific ligand binding pockets containing aptamer domain sites in
the mRNA. The selective binding of a ligand at an aptamer site initiates a
conformational change in the RNA molecule leading to altered gene expres-
sion. Synthetic transacting riboregulator tools known as antiswitches have

(continued)
12.8 Exercises 229

Box 12.3 (continued)

been created by fusion of antisense-effector domains with riboswitch-like
aptamer domains in yeast. The binding of a small-molecule theophylline to
the aptamer domain leads to a conformational change in the synthetic molecule
allowing antisense domain to bind the target mRNA including reporter green
fluorescent protein (gfp)gene encoding the GFP protein. Moreover, the ther-
modynamic properties of antiswitches and their overall system response can
be changed by introducing targeted mutations. These antiswitches including
various combination of aptamer and antisense domains have a tremendous
potential in the post-transcriptional modulation of gene expression in
eukaryotes. They provide us a novel synthetic tool to activate or repress any
target gene in response to a small molecule.

12.8 Exercises

1. An R package BiGGR contains the BIGG model (iND750) of a yeast cell. Yeast
performs alcoholic fermentation using glycolysis to produce pyruvic acid from
glucose followed by production of ethanol from pyruvic acid. Plot the metabolites
and reactions involved in glycolysis in the yeast model using BiGGR package.

Solution (Fig. 12.3)

library(BiGGR)
> data(S.cerevisiae_iND750)
> model <- S.cerevisiae_iND750@model
> relevant.species <- c("M_glc_DASH_D_c", "M_g6p_c", "M_f6p_c",
+ "M_fdp_c", "M_dhap_c", "M_g3p_c",
+ "M_13dpg_c", "M_3pg_c", "M_2pg_c",
+ "M_pep_c", "M_pyr_c",
+ "M_6pgl_c", "M_6pgc_c", "M_ru5p_DASH_D_c",
+ "M_xu5p_DASH_D_c", "M_r5p_c", "M_g3p_c", "M_s7p_c")
> relevant.reactions <- c("R_HEX1", "R_PGI", "R_PFK", "R_FBA",
"R_TPI",
+ "R_GAPD", "R_PGK", "R_PGM", "R_ENO", "R_PYK",
+ "R_G6PDH2r", "R_PGL", "R_GND", "R_RPE", "R_RPI", "R_TKT1")
> hd <- sbml2hyperdraw(model,
+ relevant.species=relevant.species,
+ relevant.reactions=relevant.reactions,
+ layoutType="dot", plt.margins=c(20, 0, 20, 80))
> plot(hd)
230 12 Computational Bioengineering

Fig. 12.3 Glycolytic metabolites and reactions in the yeast model

2. Fluxer (https://fluxer.umbc.edu) is a web application for computation of

genomic-scale flux networks . It can compute and visualize the k-shortest path
between metabolites or reactions in order to find main metabolic route. Using this
web application, perform the following tasks on the yeast model (iND750):
(a) Visualize the complete yeast model in force-based layout.
(b) Find the k-shortest path between extracellular glucose and intracellular
glucose along with flux.
(c) Find the k-shortest path from extracellular glucose to the objective function
of maximum biomass. Knock out the glucose transport uniport and observe
the effect of this knockout on this metabolic pathway.

Solution
(a) The complete graph of genome-scale metabolic model of yeast is as follows
(Fig. 12.4):
(b) The k-shortest path between extracellular glucose and intracellular glucose is as
follows. The flux of this exchange reaction is 1 unit (Fig. 12.5).
(c) The k-shortest path from extracellular glucose to the objective function of
maximum biomass along with fluxes is as follows (Fig. 12.6):
12.8 Exercises 231

Fig. 12.4 The complete graph of a genome-scale metabolic model of yeast

Fig. 12.5 The k-shortest path between extracellular and intracellular glucose
232 12 Computational Bioengineering

Fig. 12.6 The k-shortest path from extracellular glucose to maximum biomass of yeast along with
fluxes

Fig. 12.7 The k-shortest path from extracellular glucose to maximum biomass of yeast showing
zero fluxes after knocking down glucose transport uniport

The knockout effect of glucose transport uniport on the metabolic pathway is as

follows. Interestingly, all reactions are having zero flux due to non-availability of
glucose as a sole source of energy (Fig. 12.7).
12.9 Multiple Choice Questions 233

12.9 Multiple Choice Questions

1. Propagation of action potential is an example of:

(a) Negative feedback control
(b) Positive feedback control
(c) Positive feed-forward control
(d) None of the above
2. Majority of biological systems are:
(a) Non-linear dynamical systems
(b) Linear dynamical systems
(c) Linear static systems
(d) None of the above
3. The stochastic models of biological systems are represented by:
(a) Hidden Markov models
(b) Continuous time series models
(c) Integer time points models
(d) None of the above
4. An improvement of laboratory microbial strains is achieved through:
(a) Continuous laboratory evolution
(b) Random laboratory evolution
(c) Adaptive laboratory evolution
(d) Stochastic laboratory evolution
5. Directed evolution of proteins is performed using:
(a) Error-prone PCR
(b) Degenerate codons
(c) DNA shuffling
(d) All the above
6. Toggle switches in synthetic biology consist of:
(a) Two repressors and two constitutive promoters
(b) Two activators and two constitutive promoters
(c) One activator and one constitutive promoter
(d) One repressor and one constitutive promoter
7. A repressilator is an oscillator made of:
(a) A combination of multiple inverters
(b) A combination of multiple activators
(c) A combination of multiple promoters
(d) A combination of multiple repressors
8. The standard format for specification of biological design is known as:
(a) SBML
(b) SBOL
(c) SBAL
(d) SBCL
234 12 Computational Bioengineering

9. The JAVA-based tool for biological circuit design is known as:

(a) BioJAVA
(b) BioJADE
(c) BioCAD
(d) None of the above
10. The deterministic method to find a local optimal solution is known as:
(a) OptCircuit
(b) OptSol
(c) OptSolution
(d) OptOutput

Answer: 1. b 2. a 3. a 4. c 5. d 6. a 7. a 8. b 9. b 10. a

Summary
• A cell can be perceived as a dynamic system characterized by a set of state
variables.
• Both biological systems and artificial systems are non-linear dynamical systems.
• A metabolic engineer redesigns the existing network pathways in order to change
the metabolic flux rates.
• Best strategy for metabolic engineering is to perform minimum changes in the
overall metabolic network of the cell.
• Computational models are useful in understanding the evolution of laboratory
strains.
• Directed evolution has been applied on the artificial evolution of industrial
enzymes.
• Synthetic biology deals with either construction of new artificial system or
redesigning a natural biological system.
• Common examples of simple biological systems are feedback loops, genetic
toggle switches, oscillators and repressilators.
• Toggle switches have only two stable steady states ON and OFF without any
intermediate state.
• Synthetic Biology Open Language (SBOL) is a standard format to support the
specification and exchange of information regarding biological design.

Suggested Reading
Filipovic N (2020) Computational bioengineering and bioinformatics: computer modelling in
bioengineering. Springer, Cham
Filipovic N (2019) Computational modeling in bioengineering and bioinformatics. Academic Press,
London
Zhang G (2017) Computational bioengineering. CRC Press, Boca Raton
Smolke C (2018) Synthetic biology: parts, devices and applications. Wiley-Blackwell, Weinheim