Chemical Library Design

Edited by

Joe Zhongxiang Zhou

Department of Pharmacology, University of California, San Diego, CA, USA

Editor
Joe Zhongxiang Zhou
Department of Pharmacology
University of California
La Jolla, CA 92093, USA
[email protected]

ISSN 1064-3745
e-ISSN 1940-6029
ISBN 978-1-60761-930-7
e-ISBN 978-1-60761-931-4
DOI 10.1007/978-1-60761-931-4
Springer New York Dordrecht Heidelberg London
Library of Congress Control Number: 2010937983
Springer Science+Business Media, LLC 2011
All rights reserved. This work may not be translated or copied in whole or in part without the written permission of
the publisher (Humana Press, c/o Springer Science+Business Media, LLC, 233 Spring Street, New York, NY 10013,
USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of
information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed is forbidden.
The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not identified
as such, is not to be taken as an expression of opinion as to whether or not they are subject to proprietary rights.
Printed on acid-free paper
Humana Press is part of Springer Science+Business Media (www.springer.com)

Preface
Over the last two decades we have seen a dramatic change in the drug discovery process
brought about by chemical library technologies and high-throughput screening, along
with other equally remarkable advances in biomedical research. Though still evolving,
chemical library technologies have become an integral part of the core drug discovery
technologies. This volume primarily focuses on the design aspects of the chemical library
technologies. Library design is a process of selecting useful compounds from a potentially
very large pool of synthesizable candidates. For drug discovery, the selected compounds
have to be biologically relevant. Given the enormous number of compounds accessible
to the contemporary synthesis and purification technologies, powerful tools are indispensible for uncovering those few useful ones. This book includes chapters on historical
overviews, state-of-the-art methodologies, practical software tools, and successful applications of chemical library design written by the best expert practitioners.
The book is divided into five section. Section I covers general topics. Chapter 1 highlights the key events in the history of high-throughput chemistry and offers a historical
perspective on the design of screening, targeted, and optimization libraries. Chapter 2 is
a short introduction to the basics of chemoinformatics necessary for library design. Chapter 3 describes a practical algorithm for multiobjective library design. Chapter 4 discusses
a scalable approach to designing lead generation libraries that emphasize both diversity
and representativeness along with other objectives. Chapter 5 explains how FreeWilson
selectivity analysis can be used to aid combinatorial library design. Chapter 6 shows how
predictive QSAR and shape pharmacophore models can be successfully applied to targeted library design. Chapter 7 describes a combinatorial library design method based
on reagent pharmacophore fingerprints to achieve optimal coverage of pharmacophoric
features for a given scaffold.
Three chapters in Section II focus on the methods and applications of structure-based
library design. Chapter 8 reviews the docking methods for structure-based library design.
Chapters 9 and 10 contain two detailed protocols illustrating how to apply structurebased library design to the successful optimization of lead matters in the real drug discovery projects.
Section III consists of three chapters on fragment-based library design. Chapter 11
describes the key factors that define a good fragment library for successful fragment-based
drug discovery. It also provides a summary view of the fragment libraries published so far
by various pharmaceutical companies. Chapter 12 shows how a fragment library is used
in fragment-based drug design. Chapter 13 introduces a new chemical structure mining
method that searches into a huge virtual library of combinatorial origin. The method uses
fragmental (or partial) mappings between the query structure and the target molecules in
its initial search algorithms.
Chapter 14 in Section IV describes a workflow for designing a kinase targeted library.
It illustrates how to assemble a lead generation library for a target family using known
ligandtarget family interaction data from various sources.
Section V contains four chapters on library design tools. PGVL Hub described
in Chapter 15 is an integrated desktop tool for molecular design including library
design. It streamlines the design workflow from product structure formation to property

vi

Preface

calculations, to filtering, to interfaces with other software tools, and to library production
management. An application of PGVL Hub to the optimization of human CHK1 kinase
inhibitors is presented in Chapter 16. Chapter 17 is a detailed protocol on how to use
library design tool GLARE to perform product-oriented design of combinatorial libraries.
Finally, Chapter 18 is a detailed protocol on how to use the library design tool CLEVER
to perform library design and visualization.
Joe Zhongxiang Zhou

Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Contributors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

ix

SECTION I

GENERAL TOPICS

1.

Historical Overview of Chemical Library Design . . . . . . . . . . . . . . . . .

Roland E. Dolle

2.

Chemoinformatics and Library Design . . . . . . . . . . . . . . . . . . . . . .

Joe Zhongxiang Zhou

27

3.

Molecular Library Design Using Multi-Objective Optimization Methods . . . .

Christos A. Nicolaou and Christos C. Kannas

53

4.

A Scalable Approach to Combinatorial Library Design . . . . . . . . . . . . . .

Puneet Sharma, Srinivasa Salapaka, and Carolyn Beck

71

5.

Application of FreeWilson Selectivity Analysis for Combinatorial

Library Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Simone Sciabola, Robert V. Stanton, Theresa L. Johnson, and Hualin Xi

91

6.

Application of QSAR and Shape Pharmacophore Modeling Approaches

for Targeted Chemical Library Design . . . . . . . . . . . . . . . . . . . . . . 111
Jerry O. Ebalunode, Weifan Zheng, and Alexander Tropsha

7.

Combinatorial Library Design from Reagent Pharmacophore Fingerprints . . . . 135

Hongming Chen, Ola Engkvist, and Niklas Blomberg

SECTION II

STRUCTURE-BASED LIBRARY DESIGN

8.

Docking Methods for Structure-Based Library Design . . . . . . . . . . . . . . 155

Claudio N. Cavasotto and Sharangdhar S. Phatak

9.

Structure-Based Library Design in Efficient Discovery

of Novel Inhibitors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
Shunqi Yan and Robert Selliah

10.

Structure-Based and Property-Compliant Library Design

of 11-HSD1 Adamantyl Amide Inhibitors . . . . . . . . . . . . . . . . . . . . 191
Genevieve D. Paderes, Klaus Dress, Buwen Huang, Jeff Elleraas,
Paul A. Rejto, and Tom Pauly

SECTION III
11.

FRAGMENT-BASED LIBRARY DESIGN

Design of Screening Collections for Successful Fragment-Based

Lead Discovery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
James Na and Qiyue Hu

vii

viii

Contents

12.

Fragment-Based Drug Design . . . . . . . . . . . . . . . . . . . . . . . . . . . 241

Eric Feyfant, Jason B. Cross, Kevin Paris, and Dsire H.H. Tsao

13.

LEAP into the Pfizer Global Virtual Library (PGVL) Space: Creation
of Readily Synthesizable Design Ideas Automatically . . . . . . . . . . . . . . . 253
Qiyue Hu, Zhengwei Peng, Jaroslav Kostrowicki, and Atsuo Kuki

SECTION IV
14.

LIBRARY DESIGN FOR KINASE FAMILY

The Design, Annotation, and Application of a Kinase-Targeted Library . . . . . 279

Hualin Xi and Elizabeth A. Lunney

SECTION V

LIBRARY DESIGN TOOLS

15.

PGVL Hub: An Integrated Desktop Tool for Medicinal Chemists

to Streamline Design and Synthesis of Chemical Libraries
and Singleton Compounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
Zhengwei Peng, Bo Yang, Sarathy Mattaparti, Thom Shulok,
Thomas Thacher, James Kong, Jaroslav Kostrowicki, Qiyue Hu,
James Na, Joe Zhongxiang Zhou, David Klatte, Bo Chao, Shogo Ito,
John Clark, Nunzio Sciammetta, Bob Coner, Chris Waller,
and Atsuo Kuki

16.

Design of Targeted Libraries Against the Human Chk1 Kinase

Using PGVL Hub . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
Zhengwei Peng and Qiyue Hu

17.

GLARE: A Tool for Product-Oriented Design of Combinatorial Libraries . . . . 337

Jean-Franois Truchon

18.

CLEVER: A General Design Tool for Combinatorial Libraries . . . . . . . . . . 347

Tze Hau Lam, Paul H. Bernardo, Christina L. L. Chai,
and Joo Chuan Tong

Subject Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357

Contributors
CAROLYN BECK Department of Industrial and Enterprise Systems Engineering,
University of Illinois at Urbana Champaign, Urbana, IL, USA
PAUL H. BERNARDO Institute of Chemical and Engineering Sciences, Singapore,
Singapore
NIKLAS BLOMBERG DECS GCS Computational Chemistry, AstraZeneca R&D
Mlndal, Mlndal, Sweden
CLAUDIO N. CAVASOTTO School of Biomedical Informatics, The University of Texas
Health Science Center at Houston, Houston, TX, USA
CHRISTINA L.L. CHAI Institute of Chemical and Engineering Sciences, Singapore,
Singapore
BO CHAO PGRD-La Jolla, Pfizer Inc., San Diego, CA, USA
HONGMING CHEN DECS GCS Computational Chemistry, AstraZeneca R&D
Mlndal, Mlndal, Sweden
JOHN CLARK PGRD-La Jolla, Pfizer Inc., San Diego, CA, USA
BOB CONER PGRD-La Jolla, Pfizer Inc., San Diego, CA, USA
JASON B. CROSS Cubist pharmaceuticals, Inc., Lexington, MA, USA
ROLAND E. DOLLE Department of Chemistry, Adolor Corporation, Exton, PA, USA
KLAUS DRESS Oncology Medicinal Chemistry, La Jolla Laboratories, Pfizer Inc., San
Diego, CA, USA
JERRY O. EBALUNODE Department of Pharmaceutical Sciences, BRITE Institute,
North Carolina Center University, Durham, NC, USA
JEFF ELLERAAS Oncology Medicinal Chemistry, La Jolla Laboratories, Pfizer Inc., San
Diego, CA, USA
OLA ENGKVIST DECS GCS Computational Chemistry, AstraZeneca R&D Mlndal,
Mlndal, Sweden
ERIC FEYFANT Pfizer Global R&D, Cambridge, MA, USA
QIYUE HU Pfizer Global Research and Development, La Jolla Laboratories, San Diego,
CA, USA
BUWEN HUANG Oncology Medicinal Chemistry, La Jolla Laboratories, Pfizer Inc., San
Diego, CA, USA
SHOGO ITO PGRD-La Jolla, Pfizer Inc., San Diego, CA, USA
THERESA L. JOHNSON Pfizer Research Technology Center, Cambridge, MA, USA
CHRISTOS C. KANNAS Department of Computer Science, University Of Cyprus, Nicosia,
Cyprus; Noesis Chemoinformatics, Nicosia, Cyprus
DAVID KLATTE PGRD-La Jolla, Pfizer Inc., San Diego, CA, USA
JAMES KONG PGRD-La Jolla, Pfizer Inc., San Diego, CA, USA
JAROSLAV KOSTROWICKI Pfizer Global Research and Development, La Jolla
Laboratories, San Diego, CA, USA
ATSUO KUKI Pfizer Global Research and Development, La Jolla Laboratories, San Diego,
CA, USA
TZE HAU LAM Data Mining Department, Institute for Infocomm Research, Singapore,
Singapore

ix

Contributors

ELIZABETH A. LUNNEY PGRD-La Jolla, Pfizer Inc., San Diego, CA, USA
SARATHY MATTAPARTI PGRD-La Jolla, Pfizer Inc., San Diego, CA, USA
JAMES NA Pfizer Global Research and Development, La Jolla Laboratories, San Diego,
CA, USA
CHRISTOS A. NICOLAOU Noesis Chemoinformatics, Nicosia, Cyprus
GENEVIEVE D. PADERES Cancer Crystallography & Computational Chemistry, La
Jolla Laboratories, Pfizer Inc., San Diego, CA, USA
KEVIN PARIS Pfizer Global R&D, Cambridge, MA, USA
TOM PAULY Oncology Medicinal Chemistry, La Jolla Laboratories, Pfizer Inc., San
Diego, CA, USA
ZHENGWEI PENG Pfizer Global Research and Development, La Jolla Laboratories, San
Diego, CA, USA
SHARANGDHAR S. PHATAK School of Biomedical Informatics, The University of Texas
Health Science Center at Houston, Houston, TX, USA
PAUL A. REJTO Oncology, La Jolla Laboratories, Pfizer Inc., San Diego, CA, USA
SRINIVASA SALAPAKA Department of Mechanical Science and Engineering, University
of Illinois at Urbana Champaign, Urbana, IL, USA
SIMONE SCIABOLA Pfizer Research Technology Center, Cambridge, MA, USA
NUNZIO SCIAMMETTA PGRD-La Jolla, Pfizer Inc., San Diego, CA, USA
ROBERT SELLIAH Drug Design Consulting, Irvine, CA, USA
PUNEET SHARMA Integrated Data Systems Department, Siemens Corporate Research,
Princeton, NJ, USA
THOM SHULOK PGRD-La Jolla, Pfizer Inc., San Diego, CA, USA
ROBERT V. STANTON Pfizer Research Technology Center, Cambridge, MA, USA
THOMAS THACHER PGRD-La Jolla, Pfizer Inc., San Diego, CA, USA
JOO CHUAN TONG Data Mining Department, Institute for Infocomm Research,
Singapore, Singapore; Department of Biochemistry, Yong Loo School of Medicine,
National University of Singapore, Singapore, Singapore
ALEXANDER TROPSHA Laboratory for Molecular Modeling and Carolina Center
for Exploratory Cheminformatics Research, School of Pharmacy, University of North
Carolina at Chapel Hill, Chapel Hill, NC, USA
JEAN-FRANOIS TRUCHON Chemical Modeling and Informatics, Merck Frosst
Canada, Kirkland, QC, Canada
DSIRE H.H. TSAO Pfizer Global R&D, Cambridge, MA, USA
CHRIS WALLER PGRD-La Jolla, Pfizer Inc., San Diego, CA, USA
HUALIN XI Pfizer Research Technology Center, Cambridge, MA, USA
SHUNQI YAN Drug Design Consulting, Irvine, CA, USA
BO YANG PGRD-La Jolla, Pfizer Inc., San Diego, CA, USA
WEIFAN ZHENG Department of Pharmaceutical Sciences, BRITE Institute, North
Carolina Center University, Durham, NC, USA
JOE ZHONGXIANG ZHOU PGRD-La Jolla, Pfizer Inc., San Diego, CA, USA;
Department of Pharmacology, University of California, San Diego, CA, USA

Section I
General Topics

Chapter 1
Historical Overview of Chemical Library Design
Roland E. Dolle
Abstract
High-throughput chemistry (HTC) is approaching its 20-year anniversary. Since 1992, some 5,000
chemical libraries, prepared for the purpose of biological intestigation and drug discovery, have been
published in the scientific literature. This review highlights the key events in the history of HTC with
emphasis on library design. A historical perspective on the design of screening, targeted, and optimization libraries and their application is presented. Design strategies pioneered in the 1990s remain viable in
the twenty-first century.
Key words: High-throughput chemistry, chemical library, random library, targeted library,
optimization library, library design, biological activity, drug discovery.

1. Milestones in
High-Throughput
Chemistry

High-throughput chemistry (HTC) is a widely used technology

for accelerating the synthesis of chemical compounds, in particular the synthesis of biologically active compounds. HTC originated in the early 1990s. Its development and application was
largely driven by the pharmaceutical industry. In the years leading up to the introduction of HTC, the pharmaceutical industry
had been transformed by advances in molecular biology. Routine cloning and expression of molecular targets enabled medicinal chemists to optimize the potency of chemical leads directly
against an enzyme or receptor prior to in vivo testing. Brimming with molecular targets and nascent high-throughput screening technology, there was a demand to access large compound
collections to discover new drug leads. Vintage industrial compound collections generated over many decades amounted to

J.Z. Zhou (ed.), Chemical Library Design, Methods in Molecular Biology 685,
DOI 10.1007/978-1-60761-931-4_1, Springer Science+Business Media, LLC 2011

Dolle

less than a few hundred thousand molecules and the perceived

diversity of such collections was low. Accelerating the synthesis of new analogs during lead optimization was desired. The
lack of medicinal chemistry resource was a frequent bottleneck in
drug discovery programs. The benchmark at the time was that a
chemist required on average 2 weeks to synthesize a single analog
at an estimated cost of $5,000$7,000 per compound. Hence,
the prospect of HTC potentially creating chemical libraries of
hundreds of thousands of structurally diverse compounds formatted for high-throughput screening and the potential to prepare
analogs in half the time at half the cost had overwhelming appeal.
As such, HTC promised to revolutionize medicinal chemistry just
as molecular biology ushered in the era of molecular-based drug
discovery. The amalgamation of these technologies was thought
to dramatically reduce the cost and time to bring a drug to market, increasing the overall efficiency of the drug discovery process.
For these reasons, the pharmaceutical industry invested heavily
in HTC.
Figure 1.1 offers a perspective on selected major events in
HTC. Most of the innovations in HTC were made during the
1990s. In 1992, Ellman published a report in the Journal of
the American Chemical Society describing the solid-phase-assisted
synthesis of benzodiazepinones (Fig. 1.2) (1). This was hailed as
the first example of accelerated synthesis of small molecule, nonpeptide drug-like compounds. Within a year, DeWitt and coworkR
(2). The paper,
ers at Parke-Davis introduced Diversomer
appearing in the Proceedings of the National Academy of Sciences,
described the first apparatus specifically designed to carry out
HTC (Fig. 1.3). It was a rather simple device consisting of eight
gas dispersion tubes for loading solid-phase resin. It was used to
prepare parallel arrays of hydantoins and benzodiazepines. In retrospect, these HTC milestones seem insignificant relative to the
advances made in the field over the past 20 years. At the time
they served to fuel the excitement of HTC. Today, they serve
as an early example of what would become one of the recurring
themes in library design: chemical libraries modeled after known
biologically active scaffolds. Solid-phase and solution-phase synthesis techniques are used to prepare libraries (3). In solid-phase
synthesis, building blocks are immobilized on resin through a
cleavable linker. Reactants and reagents are used in excess to
speed synthesis and then simply rinsed away from resin eliminating tedious purification of intermediates. Target compounds are
detached from the linker and eluted from the resin and tested
for biological activity. The utility of solid-phase synthesis was
greatly enhanced when electrophoric tags were invented to index
the reaction history on a single resin bead (4). This advance
enabled binary encoded split-pool synthesis, i.e., the combining of building blocks in true combinatorial fashion to give tens

Historical Overview of Chemical Library Design

organization
inflection
Ellman's
solid-phase
synthesis
AFMX
IPO
$85M

IRORI
(a)

(b)

fragmentbased
discovery

flow
Chem
through
genetics dynamic
method
PCOP
DOS
Broad
CC
inflection
6M
SAR
Institute 10th
1st
NMR
Lipinski ARQL
GRC
GRC
Chem
Ro5
peak
Bank
FTI

Glaxo
buys
AFMX
$539 M

PCOP
IPO

(d)
(c)

(f)

(g)

(e)

1992

(h)

(j)
(i)

ARQL
IPO

(k)

(m)
(l)

(n)

(o)
(p)

(q)

(s) (t) (u)

MD

(w)
(v) (x)

(r)

JCC
Human
industry/solid
AGPH genone
phase
Phase I
synthesis

1995

Diversomer

(y) (z)

(ab)

2000
MW
DNA
template

(ac) (ae)

(aa)

(ad)

2005
NIH
CGC
CMLD

pubs

Fig. 1.1. Time chart showing selected events in the history of HTC. Key: (a) Affymax is the first combinatorial chemistry company to go public. (b) Ellmans solid-phase parallel synthesis of benzodiazepines fuels HTC. (c) Parke-Davis
R
, apparatus for solid-phase synthesis of small molecules. (d) Pharmacopeia licenses Columbia
introduces Diversomer
Universitys encoded split synthesis technology and company goes public a year later (NASDAQ symbol: PCOP). (e) ArQule
goes public (NASDAQ symbol: ARQL) with its industrialized solution-phase synthesis of discrete purified compounds.
(f) IRORI introduces radio frequency (Rf) encoding technology for solid-phase synthesis in cans containing reusable Rf
chips. (g) Glaxo Wellcome buys Affymax for $539 M in cash. (h) Lipinski publishes landmark correlation of physiochemical properties of drugs Rule of 5 (Ro5) has profound impact on library design. (i) 19921996: 80% of published
libraries are from industry; 75% using solid-phase synthesis. (j) Pharmacopeia generates 6 M encoded compounds.
(k) ArQule has the largest number of collaborations (27) reported for a combichem company. (l) Inaugural issue of
Molecular Diversity, the first journal dedicated to HTC. (m) SAR by NMR compounds binding to proximal subsites of
a protein are linked and optimized using HTC. (n) Agouron Pharm. moves human rhinovirus 3C protease inhibitor into
the clinical trials; HTC played a key role in its discovery. (o) S. Schreiber introduces the concept of chemical genetics
and diversity-oriented synthesis (DOS). (p) A. Czarnik editor of a new ACS journal: Journal of Combinatorial Chemistry.
(q) Academia overtakes industry library synthesis publications for the first time. (r) Human genome sequence is published
in Science. (s) Dynamic combinatorial chemistry. (t) First Gordon Research Conference entitled combinatorial chemistry:
High Throughput Chemistry & Chemical Biology. (u) D. Curran develops fluorous reagents and tags and launches Fluorous Technology Inc. (FTI). (v) DNA-templated synthesis. (w) Solution-phase overtakes solid-phase in library synthesis.
(x) Microwave-assisted synthesis gains momentum in HTC. (y) ChemBank public database established. (z) First reports
of fragment-based drug discovery. (aa) NIH Roadmap defined. NIH funds the Chemical Genomics Center and Molecular
Library Initiative, establishing 10 chemical methodology and library design centers throughout the US. (ab) Broad Institute established, furthers application of DOS in chemical biology. (ac) Flow through synthesis for HTC gains in popularity.
(ad) Of the 497 library publications reported in 2008, 90% originated from academic labs; >80% were made by solutionphase chemistry. (ae) HTC Gordon Research Conference celebrates tenth anniversary and revises conference title: High
Throughput Chemistry & Chemical Biology.

of thousands of compounds per library with a minimal number

of synthetic steps. Encoding technology was honed at Pharmacopeia, Inc., one of the early HTC startups. Within just a few
years the company had amassed over six million compounds.
Simultaneous with these developments were advances in solutionphase synthesis. Resin-bound reagents were developed to assist
in common reaction transformations. Spent resins are filtered
from reaction mixtures aiding in product isolation. Similarly, scavenger resins were invented to clean up reaction mixtures also aiding in the isolation of target molecules. ArQule, Inc. embraced

Dolle
NHFmoc
O

RB

NH2

RB

RB

NHFmoc

RC
NH

b, c

Suppor t

Support

RA

RA

RA

b, d

RD
RB

RD

RB
RC

H
N

RB
RC

RA

RC
N
Support

RA
5

N
Suppor t

RA
4

Fig. 1.2. One of the first nonpeptide library synthesis (reprinted (adapted or in part) with permission from Journal of
the American Chemical Society. Copyright 1992 American Chemical Society).

Fig. 1.3. One of the first devices for HTC (copyright (1993) National Academy of
Sciences, USA).

solution-phase parallel synthesis on a massive scale. Table 1.1

shows the number (27) of collaborations ArQule enjoyed in the
mid-1990s as companies flocked to design and purchase parallel libraries (5). ArQules solution-phase approach made available
milligram quantities of discrete purified compounds for screening
and immediate resupply.

Historical Overview of Chemical Library Design

Table 1.1
ArQule collaborations 19961997
Pharmaceutical companies
Abbott Laboratories

ACADIA Pharmaceuticals

Fibrogen

Monsanto Company

Aurora Biosciences

Genome Therapeutics

Pharmacia Biotech AB

Cadus Pharm. Corp.

GenQuest

Roche Biosciences

Cubist Pharm., Inc.

Genzyme

Solvay Duphar B.V.

DGI Biotechnologies

Immunex Corp.

Amersham Pharmacia Biotech

ICAgen, Inc.

Ontogeny

American Home Products

Scriptgen Pharm., Inc.

Ribogene

Sankyo Company

Signal Pharm., Inc.

Sepracor, Inc.

T Cell Sciences, Inc.

SUGEN, Inc.

ViroPharma

Library design was less important than library size and

>3-point scaffold diversification was a common practice invariably producing physicochemically-challenged compound arrays.
However, refocus on design occurred in 1996 when Lipinski linked certain physicochemical properties with orally active
drugs (6). Lipinskis Rule of 5 (Ro5; molecular weight (MW)
<500, clogP <5, total number of hydrogen bond (H-bond) acceptors <5, total number of H-bond donors and rotatable bonds each
<10) was rapidly adopted into library design. Ro5 put an abrupt
end to the practice of numbers inflation. A similar analysis of the
physicochemical properties yielding productive leads, i.e., those
that led to marketed drugs, gave rise to the Rule of 4 (7). This
correlation underscored the concept that the preferred leads are
those in which MW, H-bond donor/acceptor counts, and rotatable bonds can be increased during optimization as opposed to
trimming these parameters from ligands. These and related correlations had a profound impact on chemical library design. During
its first decade, while development of HTC was driven by the
pharmaceutical industry, interest of academic researchers in HTC
was bolstered in 2004 by the creation of the Chemical Genomics
Center and allied high-throughput academic screening centers
[Combinatorial Molecular Design Centers (CMLDs)]. Under the
auspices of the National Institutes of Health (NIH), their mission
is to identify small-molecule probes to establish the function of all
proteins in the proteome. Also in 2004, the Broad Institute was
established, strengthening the resolve to apply diversity-oriented
synthesis (DOS) in chemical biology. Over 5,000 libraries have
been reported in the literature from 1992 to 2008 (8).

Dolle

2. Historical
Library Designs
The objective of creating a chemical library for drug discovery,
regardless of its size or method of synthesis, is to supply biologically active compounds. For the purpose of this text, chemical
libraries can be classified into one of two categories: screening
libraries and optimization libraries. The screening library category is further subdivided into (a) random libraries, collections
with a unique design theme that has a distant, if any, relation to
known biologically active agents, and (b) targeted libraries where
the link with other biologically active structures is clearly evident.
Targeted libraries generally contain a known pharmacophore, i.e.,
a set of structural features in a molecule that is recognized at
the molecular target (enzyme, receptor, etc.) and is responsible
for that molecules biological activity (9). They may also contain
structural scaffolds that interact with a variety of molecular targets, commonly referred to as privileged scaffolds (10). Optimization libraries, on the other hand, function primarily to enhance
the biological activity of an existing lead. Potency, selectivity, and
metabolic stability are examples of deficits in leads which can be
addressed using optimization libraries. The term lead is defined
here as a biologically active molecule that has emerged from a
high-throughput screen or reported in the scientific or patent
literature.
2.1. Historical
Designs: Random
Screening Libraries

Peptide libraries. The very first examples of random screening

libraries were massive collections of peptides. Although amino
acid monomers and peptides are endowed with biological activity
and therefore may be thought of as privileged structures, it is the
scale and extensive screening of these libraries that they are considered random libraries. Researchers at Affymax developed a process for generating and screening peptide libraries on microchips
(11). Houghten conceived the technique of positional scanning to create synthetic peptide combinatorial libraries (SPCLs;
Fig. 1.4) (12). In positional scanning, each amino acid in a
given peptide sequence is sequentially held constant while the
other amino acid positions are randomized. In this way peptide mixtures are formed and screened in solution for biological activity. Deconvolution and resynthesis of single peptides are
necessary to confirm the activity of the screening results. Peptide coupling reactions were initially carried out in hand-labeled
tea bags. Libraries of several hundred thousand to millions of
members are attainable. Naturally occurring L-amino acids and
unnatural D-amino acids are employed in SPCLs. In the example

Historical Overview of Chemical Library Design

Iterative positional scanning SPCL

"Libraries f rom libraries"

H-O1-X-X-X-NH2

H-X-O2-X-X-NH2

N
R1

H-X-X-O3-X-NH2

R3

H
N
2

H-X-X-X-O4-NH2

O
N
H

R4

N
H

R4

reduction

library composed of 4 sublibraries,

each defined by a single amino acid
(O1 to O4) and X is a mixture of 50
different L-, D-, and unnatural
amino acids
6,250,000 tetrapeptides per
sublibrary
solid-phase synthesis: C-terminus
attachment point
free N-terminus; C-terminal amide

N
R1

H
N

R3

R2

cleavage

HN
R1

H
N
R2

Active peptides f rom opioid receptor screen

R3

(COIm)2

N
H

R1
N

R2

H-phe-phe-nle-arg-NH2
K i = 1.2 (selective kappa opioid agonist)

Optimized tetrapeptide: clinical candidate

O
N
R3

H-Tyr-tyr-Gly-Trp-NH2
K i = 3.0 nM (selective delta opioid agonist)
H-Tyr-nve-Gly-Nal-NH2
K i = 0.4 nM (selective mu opioid agonist)

O
N

R1
HN

R2
N
N
3

Ph
H2N

H
N
O
Ph

O
N
H

R4

reduction
cleavage

R4

H
N

N
H

NH
H2N

NH

FE 2000665
H-phe-phe-nle-arg-NHCH2-(4-pyridyl)
K i = 0.24 nM (kappa opioid agonist)
K i (mu) = 4,050 nM
K i (delta) = 20,300 nM

Fig. 1.4. Synthetic peptide combinatorial libraries (SPCLs).

of Fig. 1.4, a ca. 25 million member tetrapeptide SPCL was

screened against the mu (), kappa (), and delta () opioid
receptors (13). Peptide sequences with high affinity and selectivity for each of the receptor types were found. One of the
all D-amino acid-containing peptides, HphephenleargNH2 ,
identified as a selective receptor agonist, was further optimized to the C-terminal-modified analog, Hphephenlearg
NHCH2 (4-pyridyl). This agent, also known as FE 2000665, is
currently undergoing evaluation in human clinical trials as an analgesic. Chemical modification of SPCLs, for example, through a
borane-mediated reduction reaction (amide bond CH2 NH
bioisostere) affords libraries from libraries. These new random
derivative libraries are useful in the discovery of biologically active

10

Dolle

compounds (14). The SPCLs and their derivative libraries have

provided ligands for numerous molecular targets.
Peptoid libraries. A variant of peptide libraries, known as peptoids, was designed at Chiron (15) and then explored by many
other research groups (Fig. 1.5). In peptoids, the amino acid
side chains are relocated from the -carbon to the nitrogen atom;
hence, N-substituted glycines are monomeric building blocks.
Peptoid sequences are synthesized on solid support from immobilized -bromoacetic acid and primary amines thus giving rise
to structural diversity. In contrast to amino acids, peptoids are
not recognized as substrates for proteolytic-type metabolizing
enzymes. Peptoids were thought to be superior to peptides as
drug leads because of their perceived metabolic stability in vivo.
Nonoligomeric libraries. Peptide and peptoid libraries are
examples of oligomeric (polymeric) libraries made up of repeating
monomers (-amino acids, N-substituted glycines). Random
libraries composed of nonoligomeric compounds have been
extensively explored. One illustration comes from the former
laboratories at Organon (Fig. 1.6) (16). Thirteen different
secondary amino-phenol inputs were attached to solid support
by reaction with REM resin yielding resin-bound -amino propionates. Two-site derivatization was then used to drive library
diversity. The free phenolic OH was subjected to O-alkylation,
O-acylation, O-sulfanylation, O-triflation/Suzuki coupling
followed by N-quaternization (six inputs) and Hofmann
elimination to release a 3,042-member library of tertiary amino
aryls. One advantage of small-molecule nonoligomeric libraries

Fig. 1.5. Peptoid libraries.

Historical Overview of Chemical Library Design

11

Library design and synthesis

OH

O
N
H

REM resin
(4 commercial and 9
custom amino phenol
inputs)

R = 3-OH
R = 4-OH
R = 5-OH

OH
N

OH

HN

R1

R2

acylation,
sulfonylation,
carbamoylation then
cleavage

Mitsunobu then
cleavage

alkylation
then cleavage
OH
N

O-XR3

R2

R1
R2

R2

triflation
Suzuki coupling
then cleavage
Ar
N

R1

OR
N

R1
Physicochemical
properties
MW
ClogP
No. H-bond donors
No. H-bond acceptors
No. rotatable bonds

Lipinski
Ro5
< 500
<5
< 10
<5
< 10

>75%
members
<
<
<
<
<

450
6
1
3
1

Fig. 1.6. Nonoligomeric library (reprinted (adapted or in part) with permission from
Journal of the American Chemical Society. Copyright 2002 American Chemical Society).

versus oligomeric libraries is the control over design and physicochemical properties. In the example of Fig. 1.6, >75% of the
library members fell well within the Ro5 and successfully targeted
central nervous system (CNS) property space. Screening the
library against a variety of biological targets revealed a ca. 1 M
lead against the glycine-2 transporter.
Diversity-oriented synthesis (DOS) libraries. DOS libraries are
a special class of nonpolymeric libraries distinguished by their
synthetic design. Emphasis is placed on complexity-generating
reactions to drive structural complexity in combination with
branching pathways to drive structural diversity (Fig. 1.7) (17).
A single library will contain multiple stereochemically rich molecular frameworks incorporating multiple building blocks and
functional groups. Less emphasis is placed on physicochemical
properties. They are intended for application in chemical biology. Originally prepared using encoded split-pool synthesis, DOS
libraries are now prepared as discrete compounds on multimilligram scale. Build-couple-pair is the current paradigm for
constructing DOS libraries (18).

12

Dolle
Achmatowicz r eaction (1260 members)
Br

HO
O

R1

R4

R2
R4

HO

R4

HO

OAc

R4

O
O
Ar

Br

R1

R1

OR3

HO

R4

O
OH

HO

R1

Ar

HO

R4

R1
O

OAc

R4
OAc

Lewis-acid-catalyzed 3-component reaction (3520 members)

Ph
O

R1

CHO

Ph

HO R1
O

Ph

Ph

N
H

HN

N
O
O

NR2R3

R4

R1

Medium rings (1412 members)

O

CHO
amino alcohols

Br
CHO

N
H

HO

(2-bromo)bromomethyl aryls

Ar

HO

aryl
M

aryl

HO
P

N
H

Fig. 1.7. Diversity-oriented synthesis (DOS) libraries (reprinted (adapted or in part) with permission from Journal of
the American Chemical Society. Copyright 2005 American Chemical Society).

2.2. Historical
Designs: Targeted
Screening Libraries

The design, synthesis, and evaluation of targeted libraries have

been described much more frequently in the literature than random libraries. As mentioned above, the distinguishing feature of
targeted libraries is the intentional inclusion of a known pharmacophore or privileged structure. Targeted library design revolving
around these motifs can dramatically increase the odds of finding
valuable leads. Examples of just a few such motifs used in library
generation are shown in Fig. 1.8.
Mercaptoacyl pharmacophore library. Zinc metalloproteases
are inhibited by small molecules that contain mercaptans (thiols; CH2 SH), carboxylic acids (CO2 H), and hydroxamic acids
(CONHOH). These functional groups chelate the active-site
metal disrupting normal enzyme function. The angiotensinR
is an example of
converting enzyme (ACE) inhibitor Captopril
a thiol-based metalloprotease inhibitor. Thiols, carboxylic acids,
and hydroxamic acids are consequently affirmed pharmacophores
for this protease family. A historical example of a pharmacophore-

Historical Overview of Chemical Library Design

Biaryl

Y
N
R
Indole

Spirocycles

Z
N

O
Benzopyran

Y
1,4-Benzodiazepinone

X
NRR

X
N

N R
N

R
N

13

N
R

RRN

N
R
Arylpiperazine

Purine

Benzyhydryl

Diarylethyl

Fig. 1.8. Privileged scaffolds and pharmacophores found in libraries.

based library was described by Affymax (Fig. 1.9) (19). An

encoded pool of highly substituted prolines was prepared utilizing the 1,3-dipolar cycloaddition reaction of resin-bound azomethine ylides and electron-deficient olefins. A mercapto pharmacophore was then introduced via N-acylation with a series of
S-acetyl protected mercaptoacyl chlorides. S-deprotection and
cleavage from resin afforded a ca. 500-member library of substituted prolines bearing a COYCH2 SH functional group. The
library was assayed against ACE. Several inhibitors were found
with one possessing extraordinary potency: Ki = 160 pM. A
closely related diastereomer was >1,000-fold less active indicating
a preferred stereochemical display of pyrrolidine ring substituents
for high-affinity binding. The CH2 SH pharmacophore was similarly introduced as a final step in a dipeptide amide library from
which potent matrix metalloprotease-1 inhibitors were discovered
(20).
Statine pharmacophore library. Aspartic acid proteasemediated peptide bond hydrolysis occurs via the addition of water
to the amide carbonyl. The newly formed high-energy tetrahedral intermediate, tightly bound to enzyme, is stabilized through
hydrogen bonding with aspartic acid residues in the active site.
Collapse of the tetrahedral intermediate completes hydrolysis
releasing the corresponding C-terminal acid and N-terminal
amine peptide fragments. Statine, (3S,4S)-4-amino-3-hydroxy-6methylheptanoic acid, may be considered a mimic of the putative high-energy tetrahedral intermediate, and when embellished
with appropriate functionality, potently inhibits aspartic acid
proteases. Therefore, 4-substituted-4-amino-3-hydroxybutanoic
acids are pharmacophores for this class of protease. Researchers
at Pharmacopeia designed a library using statine and an analog,

14

Dolle
Example 1: Library design
Ar

X
1

4 x dienophiles

R2

Ar
HN

metal salt
(1,3-dipolar
cycloaddition
reaction)

R1
O

Ar
3 x mecaptoacyl
chlorides
N
Y
deprotection
HS
cleavage
O

CO2Me
N

HS
CO2H

R1
OH

CO2Me

HS

R2

O
angiotensin converting enzyme (ACE)
Ki = 0.16 nM
(purified diastereomer)

CO2H
O
ACE
Ki >100 nM
(purified diastereomer)

Example 2: Library design

R2

O
H2N
R1

N
H

H
N

HS
O

mecaptoacyl
chlorides
deprotection
cleavage

H
N
O

HS
O

R2

H
N
R1

N
H

NH2
O

O
N
H

NH2
O

matrix metalloprotease-1
(MMP-1): IC50 = 50 nM

Fig. 1.9. Targeted library containing the mercaptoacyl pharmacophore.

4-amino-3-hydroxy-5-phenylpentanoic acid (Fig. 1.10) (21).

Encoded split-pool synthesis was utilized in its construction,
generating all possible combinations of compounds from the
2 statines, 10 N-terminal capping groups, 63 C-terminal
amino acids and 40 C-terminal capping groups. The 25,200member library was screened for aspartic acid protease inhibition, in particular for inhibitory action against plasmepsin II (plm
II) and human cathepsin D (cat D). Plm II is a protease found
in the malaria (Plasmodium) parasite and functions to degrade
hemoglobin, an energy source for the maturing organism. It is a
potential molecular target for malaria intervention. A large number of active compounds were found. Following bead decoding and compound resynthesis, agents with balanced inhibitory

Historical Overview of Chemical Library Design

15

Library design
R2

R3
R4-CO2H

CO2H

FmocHN

[10 acids

OH
2 statines

R3

O
4

N
H

63 amino acids

(H)
N

H2N-R1

CO2H

(R)HN

40 amines]

OH O
R2
25,200 members

N
H

R1

Plasmepsin II (plm II) and cathepsin D (cat D) screening results

Ph
Z-Val

N
H

H
N
OH

Ph

O
Z-Val

N
H

N
H

Ph

Ph
N
H

Z-Val

N
OH O

OH

N
H

Ki = 15 nM, plm II
Ki = 140 nM, cat D

Ki = 29 nM, plm II
Ki = 44 nM, cat D

Z-Val

H
N

N
H

H
N
OH

O
N
H

N
H

Ki = 210 nM, plm II

Ki = 3 nM, cat D

Ki = 7.0 nM, plm II

Ki = 530 nM, cat D

Fig. 1.10. Statine pharmacophore library targeting aspartic acid proteases (reprinted (adapted or in part) with permission from Journal of the American Chemical Society. Copyright 2001 American Chemical Society).

activity at the two enzymes were identified as well as agents showing up to 75-fold selectivity for plm II versus cat D.
Affymaxs thiolacyl library (Fig. 1.9) and Pharmacopeias
statine library (Fig. 1.10) are pharmacophore-based libraries;
however, their design is different. In the former library, a
pool of advanced library intermediates are derivatized with
the pharmacophore (thiolacylation) as the final step in library
construction, while in the latter library the pharmacophore
(statine) is derivatized with synthons as part of library
construction.
2-Aryl indole as a G-protein-coupled receptor (GPCR) privileged scaffold. The indole ring is a premier example of a privileged scaffold. The heterocycle is present in a profusion of medicinally important natural products and pharmaceutical substances,
and it is associated with an extraordinary manifold of biological

16

Dolle

activity. Indoles have been extensively modified to exploit their

inherent therapeutic properties. In many instances, these properties are manifested through interaction with GPCRs. The privileged nature of the heterocyclic system was adroitly demonstrated in a library of 2-aryl indoles (Fig. 1.11) (22). The
library was prepared using combinatorial mixture and deconvolution techniques. Twenty arylalkyl keto acids were anchored
to Kenners safety catch resin. These resin pools were subjected to Fisher indole synthesis with a selection of 20 aryl
hydrazines. Upon activation of the sulfonamide linker, the indoles
were cleaved from resin with 80 amines yielding an indole
amide library. Half of the library was further treated with a
reducing reagent to furnish the corresponding amine indole
library. In total, 128,000 compounds were generated. The judicious choice of synthons introduced substitutes at the indole
4-, 5-, 6-, and 7-positions as well as variation of aryl substitution
at the indole 2-position. Evaluation of the compound collection
was conducted across an array of GPCR binding assays. Remarkably, potent ligands were found for many of the receptors. The
0.8 nM human neurokinin-1 (hNK1 ) ligand proved to be receptor subtype selective, devoid of affinity for hNK2 or hNK3 . Several
selective serotonin receptor ligands were uncovered representing
potential leads for medicinal chemistry. Interestingly, with the
exception of the NK1 ligand emerging from the amide library,
all of the other reported active compounds were 2-aryl indoles
bearing a 3-alkylamine.
Purine as a privileged scaffold. The purine ring is another
example of a ubiquitous, biologically-active heterocycle. It is
readily identified as a substructure in adenine, one of the base
units in DNA/RNA, and the nucleotide adenosine triphosphate
(ATP). Among its many roles, ATP is the high-energy phosphate
donor in phosphorylation reactions mediated by a large number
of kinases. As a result, functionalized purines interact with a vast
number of enzymes, receptors, and other biomolecules and satisfy the definition of a privileged structure. A purine derivative
library was designed by Schlutz and coworkers (Fig. 1.12) (23).
A series of N-9-substituted 2,4-dichloropurines were prepared by
the direct alkylation (R1 -halogen) or Mitsunobu reaction (R1 OH) of 2,6-dichloropurine. Reaction of these custom inputs with
a selection of acid-labile amine resins resulted in selective displacement of the C-6 chlorine atom and simultaneous anchoring of
the purine inputs to resin. This avails the C-2 position to a range
of derivatization chemistries including nucleophilic displacement
with amines, alcohols, phenols, thiols, and Pd-catalyzed Suzuki
coupling with aryl boronic acids (carboncarbon bond formation). Treatment of the penultimate resin intermediates with trifluoroacetic acid releases the final products from solid support
for biological testing. This chemistry is sufficiently versatile to be

Historical Overview of Chemical Library Design

O O
S
NH2 + HO2CH2

Ar

i) C6F5CH2OH,
DIAD, Ph3P
ii) R1R2NH2
(Z-subunits)
iii) amine scavenge
n

O
R1

iv) split in half for

amide reduction

NH
Ar

i) ArNHNH2
ii) ZnCl2, HOAc
iii) archive, mix/split

Ar
n

aryl ketone subunits

Kenner safety
catch resin

O O O
S
N
H

O O O
S
N
H

DIC,
THF/DCM

17

N
R2

NH
Ar

i) BH3-DMS,
dioxane, 50 oC
ii) HCl/MeOH, 50 oC
then azeotrope 3x

R1
R1

iii) split in half for

amide reduction

N
R2

N
R2

NH
Ar

NH
Ar

128,000 member library

(320 pools of 400
compounds each)

Selected (-NR1 R2) pools and biological activity:

(Numbers in columns are % inhibition values at the given screening concentration)
(R1R2N-)-subunit
OH

NH2

Assay
(concn, uM)
5-HT6 (5)
MCR-4 (2)
5-HT2a (0.1)
GnRH (1)
NPY5 (2)
CCR5 (8)
NK1 (1)

NH2

Ph

H
N

NH

76
62
14
7
89
21
23

97
10
81
4
82
1
7

NH

95
17
82
4
85
4
17

Ph

N
NH2

HO

44
-54
66
-10
--

87
5
45
6
98
0
2

68
23
63
6
96
62
42

NH

42
-0
-23
-92

Ph
N
H

N
H

N
NH

HO

NH

Br

5-HT2a
Ki = 10 nM
CO2Et

N
H

MCR-4
Ki = 612 nM

NH

Br

HO
Ph

NH

NH

Br

Br

5-HT6
Ki = 0.7 nM

Br

NPY5
Ki = 0.8 nM
Ph

N
H

NH

NH

GnRH
Ki = 52 nM

NK1
Ki = 0.8 nM

Br

CCR5
Ki = 1190 nM

Fig. 1.11. Privileged GPCR pharmacophore library (reprinted (adapted or in part) with permission from Journal of
the American Chemical Society. Copyright 2003 American Chemical Society).

18

Dolle
Library design
Cl
N
H

N
R1
custom inputs
N

Cl

R2

R2

i Pr2EtN
nBuOH
80 C

Cl

N-, S-, Onucleophiles

and Suzuki
couplings

resin
cleavage

R1

HN

R2

N
R3

N
R1

Additional representative heterocyclic inputs

Cl

Cl
N

Cl
N
H

Cl

Cl
N
N

Cl

Cl

N
Cl

Cl

N
N

Cl

Cl

Cl
Biological activity
NH2
HO2C
Cl

NH

Cl
N

N
HO

N
H

N
HO

CDK1: IC50 = 28 nM
CDK2: IC50 = 33 nM

NH

NH

N
H

CDK2-cyclin A
IC50 = 6 nM

O
N

estrogen sulfotransferase
IC50 = 500 nM

O
N

N
H

N
N

N
H

CF3 self renewal assay: EC50 = 1 M

ERK1 Kd = 98 nM
RasGAP Kd = 212 nM

Fig. 1.12. Privileged purine library.

applied to structurally related halogenated heterocycles expanding chemotype diversity beyond the purine scaffold. Approximately 45,000 substituted purines and related derivatives were
synthesized in total. Screening the library afforded potent cyclindependent kinase-1 (CDK1; IC50 = 28 nM) and CDK2 (IC50 =
6 nM) inhibitors. These kinases utilize ATP to phosphorylate proteins on serine and threonine amino acid residues regulating cell
division. Inhibitors of estrogen sulfotransferase (IC50 = 500 nM)
(24) and enzymes involved in cell regeneration were found (25).
Estrogen sulfotransferase catalyzes the transfer of a sulfuryl group
from 3 -phosphoadenosine 5 -phosphosulfate (PAPS) to estrogen
regulating hormone homeostasis.

Historical Overview of Chemical Library Design

2.3. Historical
Designs:
Optimization
Libraries

19

In the previous examples, random and targeted libraries are used

to discover leads. Optimization libraries are employed when lead
structures have already been identified, serving to improve the
potency, selectivity, or other characteristics of the molecule.
Human rhinovirus 3C protease inhibitor. The former Agouron
Pharmaceuticals research laboratories identified a tripeptidyl
Michael acceptor as a lead structure in a rhinovirus 3C protease inhibitor program (Fig. 1.13) (26). This agent was
an irreversible inhibitor (second-order rate constant: kobs /
I = 280,000 M1 s1 ) of the enzyme which is essential
for viral replication. One issue with the compound was the
N-benzylthiocarbamate, an N-terminal capping group potentially undergoing metabolism leading to a short half-life of the
agent in vivo. An optimization library was designed to find an
N-terminal amide to replace the benzylthiocarbamate that would
provide the necessary metabolic stability. The lead possessed a
modified glutamate residue which proved to be a strategic asset
for crafting an optimization library. A modified N-Fmoc glutamic acid bearing an ,-unsaturated ethyl ester was attached
to solid support through a Rink amide linker. A multistep elaboration completed the assembly of the penultimate tripeptide
intermediate with a free N-terminus. This resin-bound intermediate was then acylated with ca. 500 carboxylic acids and acid
chlorides. Cleavage of the analogs from the resin and evaluation in a high-throughput enzyme assay led to the discovery
of 5-methylisoxazole-3-carboxamide as an ideal surrogate for
the benzylthiocarbamate. The 5-methylisoxazole-3-carboxamide
analog was essentially equipotent (kobs /I = 260,000 M1 s1 )
with the original lead. This compound showed antiviral activity without cytotoxicity in cell culture and also exhibited broadspectrum antiviral activity. The advanced lead was subjected
to traditional optimization ultimately giving rise to AG7088
(kobs /I = 1,470,000 M1 s1 ). AG7088 was nominated for
development and subsequently advanced into human clinical trials (27). The N-(5-methylisoxazole-3-carboxamido) group was
retained in the clinical candidate.
Kappa opioid receptor antagonist. There are three opioid receptors, mu (), kappa (), and delta (). 4-(3Hydroxyphenyl)-trans-3,4-dimethylpiperidine is an opioid receptor antagonist pharmacophore originally discovered in the 1970s.
All piperidine nitrogen analogs of this scaffold reported in
the literature displayed no receptor selectivity with the exception of the selective agent shown in Fig. 1.14. This suggested to Carroll and coworkers that given the appropriate
N-substituent, it may be possible to obtain selective antagonists
for the and opioid receptors (28). In a program to investigate selective antagonists for the treatment of drug abuse,

20

Dolle
Lead enhancement
CONH2
O
Ph

N
H

H
N

Optimization libr ar y
O

O
N
H

OH

CO2Et
CO2Et

FmocHN

Ph
lead
kobs/I= 280,000 M1s1

CONH2
O

O N

N
H

H
N

N
H

CO2Et

O
O
R
O

O N

NH

O
N
H

Ph
RCOCl

Ph
advanced lead
kobs/I = 260,000 M1s1

CO2Et

N
H

H
N

H2N

NH

N
H

CO2Et

N
H

H
N
O

NH2

O
N
H

CO2Et

Ph
ca. 500 member optimization library
to discover a replacement for the
metabolically labile N-benzylthiocarbamate in the lead inhibitor

F
AG7088: clinical candidate
kobs/I = 1,470,000 M1s1

Fig. 1.13. Optimization library for human rhinovirus 3C protease (reprinted (adapted or in part) with permission from
Journal of the American Chemical Society. Copyright 2001 American Chemical Society).

an optimization library was conceived to search for such agents.

(+)-4-(3-Hydroxyphenyl)-trans-3,4-dimethylpiperidine was coupled to 11 amino acids and reduced to give a series of piperidines
with a newly appended primary or secondary amine. These
inputs were subjected to solution-phase acylation reactions using
substituted benzoic, phenylacetic, phenyl cinnamic, and (3phenyl)propionic acids. The acylation reaction was sufficiently
clean that following aqueous workup, the 288 library products
were screened directly, without purification, against the three opioid receptors. A selective agent was discovered (Ki = 7 nM;
57-fold and >825-fold selective versus and , respectively) and
its binding and antagonist activity confirmed upon retesting the
purified compound. A remarkable range of potency and selectivity

Historical Overview of Chemical Library Design

21

Library design
OH

OH
i) Boc-Aa
ii) TFA
iii) BH3-Me2S

selective kappa
antagonists?

N
H

R3COOH

R1

R1

(+)-enantiomer
as starting
material

Ph
lead structure
Ki = 0.74 nM ( )
Ki = 322 nM ( )

OH

OH

HN

R2

R3

R2

O
Library
288 members

Screening results
OH

OH

OH

HO

HO

HO

NH
O
K i = 7 nM ( )
= 57;
>824
(functional antagonist)

Ph

NH
O
no binding

NH
O
Ki = 54 nM ( )
Ki = 10 nM ( )

Fig. 1.14. Kappa () opioid receptor antagonist optimization library (reprinted (adapted or in part) with permission
from Journal of the American Chemical Society. Copyright 1999 American Chemical Society).

was also observed. Structureactivity relationship (SAR) data

obtained from the library accentuates the critical role of the isopropyl group. This is corroborated both in terms of stereochemistry, (S)-configuration necessary for affinity, and selectivity as the
isopropyl benzyl exchange resulted in selective antagonists.
Using a molecule described in the literature as a starting point
for library (analog) synthesis is an example of a knowledge-based
approach to lead optimization.
Raf kinase inhibitor. High-throughput screening of a chemical collection at the former Bayer Research Center turned up
3-thienyl urea as a modestly potent inhibitor of p38 kinase
(IC50 = 290 nM) possessing comparatively weak activity at raf
kinase (IC50 = 17 M; p38/raf = 0.017; Fig. 1.15) (29).
Because of its low activity, many laboratories would have discounted this compound as a raf kinase inhibitor lead. Smith and
coworkers applied HTC techniques in an attempt to improve

22

Dolle
Dual approach to library design
O

O
H
N

H
N

S
O

Raf kinase lead

IC50 = 17000 nM, raf kinase
IC50 = 290 nM, p38 kinase
sequential
optimization
strategy

combinatorial
optimization
strategy

Part 1
Y

NH2

+ R-Ph-NCO

H
N

screening
Z

H
N

H
N

O
H
N

Z
X

R
N

Z Y

R
N
R

S
O
ca. 1000 member library
IC50 = 1700 nM

screening

Part 2
N

NH2

Y
Z

H
N
O

+ 4-Me-Ph-NCO

H
N

H
N

O
O
IC50 > 25,000 nM

advanced lead
IC50 = 54 nM, raf kinase
IC50 = 360 nM, p38 MAP kinase

screening

H
N

F3C
Cl

H
N

H
N
O

H
N

BAY 43-9006: clinical candidate

IC50 = 12 nM

Fig. 1.15. Contrasting raf kinase inhibitor optimization strategies (reprinted (adapted
or in part) with permission from Journal of the American Chemical Society. Copyright
2002 American Chemical Society).

both the inhibitory potency and selectivity of the urea against raf
kinase. A two-part sequential optimization strategy was devised.
In part one, coupling conservatively altered 3-aminothienyls with
phenyl-substituted isocyanates was carried out. A ca. 10-fold
improvement in activity over the original lead was obtained with
a 4-methyl group in the phenyl ring. In part 2, the optimized

Historical Overview of Chemical Library Design

23

4-methylphenyl portion of the molecule was held constant and

a broad range of heterocycles was explored to optimize the 3thienyl moiety. This resulted in no further improvement in activity. The sequential two-part optimization strategy failed to meet
the objective. This was followed by a combinatorial strategy in
which 300 anilines/heterocyclic amines were combined with 75
aryl/heteroaryl isocyanates to produce an array of a ca. 1,000
compounds. Evaluation of these compounds resulted in the identification of the advanced lead, 1-(5-tert-butylisoxazol-3-yl)-3-(4phenoxyphenyl)urea: IC50 = 54 nM) possessing 7-fold selectivity over p38 kinase. This agent represented a significant 314-fold
increase in raf kinase potency versus the original lead. The result
was unanticipated. The 5-tert-butyl-3-aminoisoxazole present in
the advanced lead was considered an inactive heterocycle based
on the SAR data generated from the original sequential optimization strategy. Further optimization of the advanced lead was
achieved, identifying a clinical candidate [IC50 (raf kinase) = 12
nM] displaying sufficient potency and favorable kinase enzyme
selectivity. Key structural elements present in the advanced lead
are retained in the clinical candidate. This library design example
beautifully underscores the advantage of combinatorial versus the
traditional step-wise approach to lead optimization.

3. Summary
HTC originated in the early 1990s in response to unprecedented
access to molecular targets, advances in high-throughput screening technology, and the demand for new chemical compound
collections. Approaching two decades of application, there are
over 5,000 chemical libraries reported in the literature (8). Initial
design strategies based on oligomeric and nonoligomeric libraries
with multiple (>3) points of diversity have progressed toward
more carefully crafted molecules with attention paid to physicochemical and toxiphoric properties. Today, library compounds
are typically synthesized on a milligram scale (10100 mg), purified, and evaluated not only against the primary target but also in
selectivity assays including (a) in vitro drug metabolism pharmacokinetic (DMPK) assays which measure a compounds metabolic
stability and interaction with cytochrome P450 metabolizing
enzymes, and (b) ion channels associated with cardiac function. Libraries are being used to generate multiple SARs to efficiently identify and simultaneously address compound liabilities.
Library designs incorporating pharmacophores (19, 21) and privileged structures (22, 23) have historically been successful in lead
finding. New chemotypes are needed to investigate previously

24

Dolle

unexplored diversity space to discover fresh leads. Identifying a

metabolically stable surrogate for the N-benzylthiocarbamate in
the rhinovirus 3C protease inhibitor (25), generating a series of
selective kappa opioid receptor antagonists starting with a nonselective opioid ligand (27), and enhancing the potency and
selectivity of a marginally active raf kinase inhibitor by combinatorializing synthons when traditional medicinal chemistry
failed (28) serve as historical references to the successful application of HTC in lead optimization. Such references are valuable lessons in library design that can still be considered in
contemporary HTC.
References
1. Bunin, B. A., Ellman, J. A. (1992) A general and expedient method for the solid phase
synthesis of 1,4-benzodiazepine derivatives. J
Am Chem Soc 114, 1099710998.
2. DeWitt, S. H., Kiely, J. S., Stankovic, C. J.,
Schroeder, M. C., Cody, D. M. R., Pavia,
M. R. (1993) Diversomers: an approach to
nonpeptide, nonoligomeric chemical diversity. Proc Natl Acad Sci USA 90, 69096913.
3. Terrett, N. (1998) Combinatorial Chemistry.
Oxford University Press, Oxford, UK.
4. Ohlmeyer, M. H. J., Swanson, R. N., Dillard,
L., Reader, J. C., Asouline, G., Kobayashi,
R., Wigler, M., Still, W. C. (1993) Complex synthetic chemical libraries indexed with
molecular tags. Proc Natl Acad Sci USA 90,
1092210926.
5. Data taken from ArQules 10 K annual
reports for years ending 19961997.
http://www.sec.gov/Archives/edgar/data.
6. Lipinski, C. A., Lombardo, F., Dominy, B.
W., Feeney, P. J. (1997) Experimental and
computational approaches to estimate solubility and permeability in drug discovery
and development settings. Adv Drug Delivery Rev 23, 325.
7. Teague, S. J., Davis, A. M., Leeson, P. D.,
Oprea, T. (1999) The design of leadlike combinatorial libraries. Angew Chem, Int Ed 38,
37433748.
8. Dolle, R. E., Le Bourdonnec, B., Goodman,
A. J., Morales, G. A., Thomas, C. J., Zhang,
W. (2009) Comprehensive survey of chemical libraries for drug discovery and chemical
biology: 2008. J Comb Chem 11, 755802.
9. Gund, P. (1977) Three-dimensional pharmacophoric pattern searching. Prog Mol Subcell
Biol 5, 117143.
10. Hajduk, P. J., Bures, M., Praestgaard, J.,
Fesik, S. W. (2000) Privileged molecules for
protein binding identified from NMR-based
screening. J Med Chem 43, 34433447.

11. Fodor, S. P. A., Read, J. L., Pirrung,

M. C., Stryer, L., Lu, A. T., Solas, D.
(1991) Light-directed, spatially addressable
parallel chemical synthesis. Science 251,
767773.
12. Dooley, C., Houghten, R. (1993) The use
of positional scanning synthetic peptide combinatorial libraries for the rapid determination of opioid receptor ligands. Life Sci 52,
15091517.
13. Dooley, C. T., Ny, P., Bidlack, J. M.,
Houghten, R. A. (1998) Selective ligands
for the , , and opioid receptors identified from a single mixture based tetrapeptide
positional scanning combinatorial library. J
Biol Chem 273, 1884818856.
14. Ostresh, J. M., Husar, G. M., Blondelle, S.,
Dorner, B., Weber, P. A., Houghten, R. A.
(1994) Libraries from libraries: chemical
transformation of combinatorial libraries to
extend the range and repertoire of chemical diversity. Proc Natl Acad Sci USA 91,
1113811142.
15. Zuckermann, R. N., Martin, E. J.,
Spellmeyer, D. C., Stauber, G. B., Shoemaker, K. R., Kerr, J. M., Figliozzi, G. M.,
Goff, D. A., Siani, M. A., Simon, R., Banville,
S. C., Brown, E. G., Wang, L., Richter,
L. S., Moos, W. H. (1994) Discovery of
nanomolar ligands for 7-transmembrane Gprotein-coupled receptors from a diverse N(substituted)glycine peptoid library. J Med
Chem 37, 26782685.
16. Barn, D., Caulfield, W., Cowley, P., Dickins, R., Bakker, W. I., McGuire, R., Morphy, J. R., Rankovic, Z., Thorn, M.
(2001) Design and synthesis of a maximally
diverse and druglike screening library using
REM resin methodology. J Comb Chem 3,
534541.
17. Burke, M. D., Berger, E. M., Schreiber, S.
L. (2004) A synthesis strategy yielding skele-

Historical Overview of Chemical Library Design

18.

19.

20.

21.

22.

23.

24.

tally diverse small molecules combinatorially.

J Am Chem Soc 126, 1409514104.
Nielsen, T. E., Schreiber, S. L. (2008)
Towards the optimal screening collection. A
synthesis strategy. Angew Chem, Int Ed 47,
4856.
Murphy, M. M., Schullek, J. R., Gordon,
E. M., Gallop, M. A. (1995) Combinatorial organic synthesis of highly functionalized pyrrolidines: identification of a potent
angiotensin converting enzyme inhibitor
from a mercaptoacyl proline library. J Am
Chem Soc 117, 70297030.
Lynas, J. F., Martin, S. L., Walker, B.,
Baxter, A. D., Bird, J., Bhogal, R., Montana, J. G., Owen, D. A. (2000) Solidphase synthesis and biological screening of
N--mercaptoamide template-based matrix
metalloprotease inhibitors. Comb Chem High
Throughput Screening 3, 3741.
Dolle, R. E., Guo, J., OBrien, L., Jin, Y.,
Piznik, M., Bowman, K. J., Li, W., Egan,
W. J., Cavallaro, C. L., Roughton, A. L.,
Zhao, W., Reader, J. C., Orlowski, M., JacobSamuel, B., DiIanni Carroll, C. (2000) A
statistical-based approach to assessing the
fidelity of combinatorial libraries encoded
with electrophoric molecular tags. Development and application of tag decode-assisted
single bead LC/MS analysis. J Comb Chem
2, 716731.
Willoughby, C. A., Hutchins, S. M., Rosauer,
K. G., Dhar, M. J., Chapman, K. T., Chicchi,
G. G., Sadowski, S., Weinberg, D. H., Patel,
S., Malkowitz, L., Di Salvo, J., Pacholok,
S. G., Cheng, K. (2001) Combinatorial synthesis of 3-(amidoalkyl) and 3-(aminoalkyl)2-arylindole derivatives: discovery of potent
ligands for a variety of G-protein-coupled
receptors. Bioorg Med Chem Lett 12, 9396.
(a) Ding, S., Gray, N. S., Ding, Q., Wu,
X., Schultz, P. G. (2002) Resin-capture
and release strategy toward combinatorial
libraries of 2,6,9-substituted purines. J Comb
Chem 4, 183186. (b) Ding, S., Gray, N.
S., Wu, X., Ding, Q., Schultz, P. G. (2002)
A combinatorial scaffold approach toward
kinase-directed heterocycle libraries. J Am
Chem Soc 124, 15941596.
Verdugo, D. E., Cancilla, M. T., Ge, X., Gray,
N. S., Chang, Y. -T., Schultz, P. G., Negishi,

25.

26.

27.

28.

29.

25

M., Leary, J. A., Bertozzi, C. R. (2001) Discovery of estrogen sulfotransferase inhibitors

from a purine library screen. J Med Chem 44,
26832686.
Chen, S., Do, J. T., Zhang, Q., Yao, Q.,
Yao, S., Yan, F., Peters, E. C., Schoeler, H.
R., Schultz, P. G., Ding, S. (2006) Selfrenewal of embryonic stem cells by a small
molecule. Proc Natl Acad Sci USA 103,
1726617271.
Dragovich, P. S., Zhou, R., Skalitzky, D. J.,
Fuhrman, S. A., Patick, A. K., Ford, C. E.,
Meador, J. W., III, Worland, S. T. (1999)
Solid-phase synthesis of irreversible human
rhinovirus 3C protease inhibitors. Part 1:
optimization of tripeptides incorporating
N-terminal amides. Bioorg Med Chem 7,
589598.
Matthews, D. A., Dragovich, P. S., Webber,
S. E., Fuhrman, S. A., Patick, A. K., Zalman, L. S., Hendrickson, T. F., Love, R.
A., Prins, T. J., Marakovits, J. T., Zhou,
R., Tikhe, J., Ford, C. E., Meador, J. W.,
Ferre, R. A., Brown, E. L., Binford, S.
L., Brothers, M. A., Delisle, D. M., Worland, S. T. (1999) Structure-assisted design
of mechanism-based irreversible inhibitors of
human rhinovirus 3C protease with potent
antiviral activity against multiple rhinovirus
serotypes. Proc Natl Acad Sci USA 96,
1100011007.
Thomas, J. B., Fall, M. J., Cooper, J. B.,
Rothman, R. B., Mascarella, S. W., Xu,
H., Partilla, J. S., Dersch, C. M., McCullough, K. B., Cantrell, B. E., Zimmerman,
D. M., Carroll, F. I. (1998) Identification
of an opioid receptor subtype-selective Nsubstituent for (+)-(3R,4R)-dimethyl-4-(3hydroxyphenyl)piperidine. J Med Chem 41,
51885197.
Smith, R. A., Barbosa, J., Blum, C. L.,
Bobko, M. A., Caringal, Y. V., Dally, R.,
Johnson, J. S., Katz, M. E., Kennure, N.,
Kingery-Wood, J., Lee, W., Lowinger, T. B.,
Lyons, J., Marsh, V., Rogers, D. H., Swartz,
S., Walling, T., Wild, H. (2001) Discovery
of heterocyclic ureas as a new class of raf
kinase inhibitors: identification of a second
generation lead by a combinatorial chemistry approach. Bioorg Med Chem Lett 11,
27752778.

Chapter 2
Chemoinformatics and Library Design
Joe Zhongxiang Zhou
Abstract
This chapter provides a brief overview of chemoinformatics and its applications to chemical library design.
It is meant to be a quick starter and to serve as an invitation to readers for more in-depth exploration of
the field. The topics covered in this chapter are chemical representation, chemical data and data mining,
molecular descriptors, chemical space and dimension reduction, quantitative structureactivity relationship, similarity, diversity, and multiobjective optimization.
Key words: Chemoinformatics, QSAR, QSPR, similarity, diversity, library design, chemical
representation, chemical space, virtual screening, multiobjective optimization.

1. Introduction
Library design is essentially a selection process, selecting a useful subset of compounds from a candidate pool. How to select
this subset depends on the purpose of the library. For a simple
probe of a local structureactivity relationship (SAR), medicinal
chemists may be able to choose an excellent subset of representatives from a small pool of synthesizable compounds to achieve the
goal without resorting to any sophisticated design tools. For complex applications of library though, design tools are indispensable
for obtaining optimal results. Majority of the design tools used
for library design fall into a field called chemoinformatics, a discipline that studies the transformation of data into information
and information into knowledge for better decision making (1).
Actually, the recent explosive development in chemoinformatics
has mainly been stimulated by the ever-increasing applications of
chemical library technologies in pharmaceutical industry.
J.Z. Zhou (ed.), Chemical Library Design, Methods in Molecular Biology 685,
DOI 10.1007/978-1-60761-931-4_2, Springer-Science+Business Media, LLC 2011

27

28

Zhou

Theoretically, there are 1060 10100 compounds available to

a small-molecule drug discovery program of any given drug target (2, 3). The purpose of a drug discovery program is to find
a good compound that can modulate the function of the target
while avoiding harmful side effects. It is not a trivial task to navigate even a small portion of this huge chemical space and locate
a few optimal candidates with desirable properties. Therefore, a
drug discovery program usually starts with the discovery of lead
compounds followed by their optimizations, instead of the impossible task of sifting through the entire chemical space directly
for a drug compound. Even this two-step divide-and-conquer
approach cannot divide the chemical space small enough for manual identification of desirable compounds. Library design as a
drug discovery technology faces the same finding-a-needle-in-ahaystack issues as the drug discovery itself. Computational tools
are necessary for efficient navigations in the chemical space. Thus,
chemoinformatic methods are developed to allow chemical data
manipulations, chemoinformatic transformations, easy navigation
in chemical space, predictive model building, etc. Chemoinformatics has played a very important role in the rapid development
and widespread applications of chemical library technologies.
In this chapter, we will give a brief introduction to the basic
concepts of chemoinformatics and their relevance to chemical
library design. In Section 2, we will describe chemical representation, molecular data, and molecular data mining in computer;
we will introduce some of the chemoinformatics concepts such as
molecular descriptors, chemical space, dimension reduction, similarity and diversity; and we will review the most useful methods
and applications of chemoinformatics, the quantitative structure
activity relationship (QSAR), the quantitative structureproperty
relationship (QSPR), multiobjective optimization, and virtual
screening. In Section 3, we will outline some of the elements
of library design and connect chemoinformatics tools, such as
molecular similarity, molecular diversity, and multiple objective
optimizations, with designing optimal libraries. Finally, we will
put library design into perspective in Section 4.

2. Chemoinformatics
Although still rapidly evolving, chemoinformatics as a scientific
discipline is relatively mature. This section is meant to be introductory only. Interested readers are referred to various monographs on chemoinformatics for a deep understanding of the field
(48).

Chemoinformatics and Library Design

29

The first task of chemoinformatics is to transform chemical

knowledge, such as molecular structures and chemical reactions,
into computer-legible digital information. The digital representations of chemical information are the foundation for all chemoinformatic manipulations in computer. There are many file formats
for molecular information to be imported into and exported from
computer. Some formats contain more information than others.
Usually, intended applications will dictate which format is more
suitable. For example, in a quantum chemistry calculation the
molecular input file usually includes atomic symbols with threedimensional (3D) atomic coordinates as the atomic positions,
while a molecular dynamics simulation needs, in addition, atom
types, bond status, and other relevant information for defining a
force field.
Chemical representation can be rule-based or descriptive.
Here we will give a short description of two popular file formats
for molecular structures, MOLfiles (9) and SMILES (1013), to
illustrate how molecules are represented in computer. SMILES is
a rule-based format while MOLfile is a more descriptive one.
A MOLfile usually contains a header block and a connection table (see Fig. 2.1). The header block consists of three lines

2.1. Chemical
Representation

(a)

(b)
Header block

SMMXDraw12120917342D
11 11 0
12.3082
13.0242
13.7402
14.4562
15.1722
15.1722
15.8882
13.7402
13.0242
12.3082
11.5922
2 1 2
3 2 1
3 4 1
4 5 1
5 6 2
5 7 1
8 3 2
9 8 1
10 9 2
10 11 1
1 10 1
M END

0
0
0
0
0
0
0
0
0
0
0

0 0 0
-7.1882
-7.6016
-7.1882
-7.6016
-7.1882
-6.3615
-7.6016
-6.3615
-5.9481
-6.3615
-5.9481
0 0 0
0 0 0
0 0 0
0 0 0
0 0 0
0 0 0
0 0 0
0 0 0
0 0 0
0 0 0
0 0 0

0 0999 V2000
0.0000 C
0 0
0 0
0.0000 C
0 0
0.0000 C
0.0000 N
0 0
0.0000 C
0 0
0 0
0.0000 O
0 0
0.0000 C
0 0
0.0000 C
0.0000 C
0 0
0 0
0.0000 C
0.0000 O
0 0

Counts line

0
0
0
3
0
0
0
0
0
0
0

0
0
0
0
0
0
0
0
0
0
0

0
0
0
0
0
0
0
0
0
0
0

0
0
0
0
0
0
0
0
0
0
0

0
0
0
0
0
0
0
0
0
0
0

0
0
0
0
0
0
0
0
0
0
0

0
0
0
0
0
0
0
0
0
0
0

0
0
0
0
0
0
0
0
0
0
0

0
0
0
0
0
0
0
0
0
0
0

0
0
0
0
0
0
0
0
0
0
0

Atom block
Connection
Table (Ctab)

Bond block

Fig. 2.1. Illustrative example of a MOLFile for acetaminophen (also known as paracetamol). (a) Molecular structure of
acetaminophen, commonly known as Tylenol. Tylenol is a widely used medicine for reducing fever and pain. (b) MOLFile
for acetaminophen.

30

Zhou

containing such information as molecular IDs, owner of the

record, dates, and other miscellaneous information and comments. The connection table (CTab) contains the actual molecular structure information in several sections: a count line, an
atom block, a bond block, and a property block. The count line
includes number of atoms, number of bonds, number of atom
lists, chiral flag for the molecule, and number of lines of additional
property information in the property block. The atom block is
made up of atom lines with each line containing atomic coordinates, atomic symbol, relative mass, charge, atomic stereo parity, valence, and other information. The bond block consists of
bond lines for all bonds. Each bond line contains information
about bond type, bond stereo, bond topology, and reacting center status. The property block consists of property lines. Most of
the property lines start with a letter M followed by a property
identifier. The usual properties appearing in property blocks are
charges, radical status, isotope, Rgroup properties, 3D features,
and other properties. The property block ends with an M END
line.
The MOLfile format belongs to a general format definition
for Chemical Table Files (CTFiles). CTFiles define file formats
for various purposes. Particularly, multiple molecular entries can
be stored in an SDFile format. Each molecular entry in an SDFile
may consist of the MOLfile as described above and other data
records associated with the molecule. Other important file formats of CTFiles definitions are RGFile for Rgroup files, rxnfile
for reaction files, RDFile for multiple records of molecules and/or
reactions along with their associated data, and XDFile for XMLbased records of molecules and/or reactions along with their
associated data. Interested readers are referred to Symyxs MDL
white paper for a complete coverage of the CTFile formats in general and Molfile format in particular (9).
SMILES (Simplified Molecular Input Line Entry Systems) is
a line notation system based on principles of molecular graph
theory for entering and representing molecules and reactions
in computer (1013). It uses a set of simple specification rules
to derive a SMILES string for a given molecular structure (or
more precisely, a molecular graph). A simplified set of rules is as
follows:
Atoms are represented by their atomic symbols enclosed by
square bracket, [ ], which can be dropped for the organic
subset B, C, N, O, P, S, F, Cl, Br, and I. Hydrogen atoms
are usually implicit.
Bonds between adjacent atoms are assumed to be single unless specified otherwise; double and triple bonds are
denoted as = and #.
Branches are specified by enclosing them in parentheses,
which can be nested and stacked. The implicit connection

Chemoinformatics and Library Design

31

of a branch in a parenthesized expression is to the left of the

string.
Rings in cyclic structures are broken with a unique number
attached to the two atoms at each break point. A single atom
may involve in multiple ring breakages. In this case, it will
have multiple numbers attached to it with each number corresponding to a single break point.
Atoms in aromatic rings are denoted by lower case letters.
Disconnected structures are separated by a period (.).
There are also rules specifying chiral centers, configurations
around double bonds, charges, isotopes, etc. A complete list of
specification rules can be found in the SMILES document at Daylights web site (13). Even with this simplified subset of rules,
SMILES strings can be derived for a lot of molecules. Table 2.1
illustrates just a few of them.

Table 2.1
Illustrative SMILES: molecular structures and the corresponding SMILES strings are paired vertically. The numbered
arrows on the three cyclic molecular structures are not part
of the molecules. They are used to indicate the break points
for deriving the corresponding SMILES strings (see text)
1

N
CCC

CC = C

CC#C

c1ccncc1

N
O
CCC(C)N

CC(C)C(C(C)N)C(C)
O

c1cc2c(cc[nH]2)nc
1

CC(=O)Nc1ccc(cc1)O

Note that a single molecule may correspond to many different, but equivalent, SMILES strings. For example, for a given
asymmetric molecule, starting from a different asymmetric atom
will lead to a different, but equally valid, SMILES string. These
various SMILES are called isomeric SMILES. They can be converted to a unique form called canonical SMILES (11).
Daylight has extended SMILES rules to accommodate
general descriptions of molecular patterns and chemical reactions (13). These SMILES extensions are called SMARTS and
SMIRKS. SMARTS is a language for describing molecular patterns while SMIRKS defines rules for chemical reaction transformations.

32

Zhou

SMILES strings are very concise and hence are suitable for
storing and transporting a large number of molecular structures,
while MOLfiles and its extension SDFiles have the option to store
more complicated molecular data such as 3D molecular conformational information and biological data associated with the
molecules. There are many other file formats not discussed here.
Interested readers can find a list of file types at the following web
site: http://www.ch.ic.ac.uk/chemime/.
2.2. Data, Databases,
and Data Mining

Modern drug discovery is largely a data-driven process. There

are tremendous amounts of data collected to facilitate decision
making at almost every stage of the drug discovery process.
Majority of the data are associated with molecules. These molecular data can be classified into two broad categories: physicochemical properties and biological assay data. Typical physicochemical properties for a molecule include molecular weight,
number of heavy atoms, number of rings, number of hydrogen
bond donors/acceptors, number of oxygen or nitrogen atoms,
polar/nonpolar surface area, volume, water solubility, 1-octanol
water partition coefficient (CLogP), pKa , and molecular stability.
Most of these properties can be calculated while some are measured experimentally.
Biological data associated with small molecules come from
a heterogeneous array of assays. Typical biological assay data
include percentage inhibitions from high-throughput screening
of binding assays against specific biological targets, biochemical
binding constants, activity IC50 constants in cell-based assays,
percentage inhibitions or binding constants against various CYP
450 proteins as first screening for metabolic liabilities, compound
stabilities in human/animal microsome and hepatocytes, transmembrane permeabilities (such as Caco-2 or PAMPA), dofetilide
binding constants for finding potential hERG blockers (may
cause prolongation of QT interval), genotoxicity data from assays
like AMES tests, and various pharmacokinetic and pharmacodynamic data. Different biological assays vary greatly in experimental modes (biochemical, in vitro, in vivo, etc.), readout accuracies, and throughputs. Therefore, some types of data are abundant while others are only available very scarcely.
Computational models can be built based on experimental results for both physicochemical properties and biological
assays. Thereby predicted physicochemical properties and biological assay data become available to compounds before their syntheses or to compounds without the data because of various experimental limitations such as cost or throughput. These computed
data become an integral part of the molecular data.
Molecular data are usually stored in databases along with
their corresponding molecular structures. Database is the central part of a typical chemoinformatics system that further-

Chemoinformatics and Library Design

33

more consists of interfaces and programs for capturing, storing,

manipulating, and retrieving data. Careful data modeling for
designing a robust chemoinformatics system integrating various
heterogeneous molecular data is essential for the chemoinformatics system to deliver its designed functions with acceptable performances (14).
Data mining is to seek patterns among a given set of data.
Mining molecular data to aid molecular design is one of the most
important functions of a chemoinformatics system. Typical data
mining tasks in drug discovery include subsetting libraries; identifying lead chemical series from HTS data (HTS hit triage); querying databases for similar compounds in terms of structural patterns, activity profiles across various biological targets, or property
profiles across various physicochemical properties; and establishing quantitative structureactivity relationships (QSAR) or quantitative structureproperty relationships (QSPR). In a general
sense, drug design is an ideal field of applications for chemical
data mining. Therefore, most of the drug design tools are actually chemical data mining tools.
2.3. Molecular
Descriptors

To distinguish one molecule from another in computer and to

establish various predictive QSAR/QSPR models for design purposes, molecules need to be projected into a chemical space of
molecular characteristics. This projection is usually done through
molecular descriptors. Given the diverse molecular characterizations, it is not an easy task to give a simple definition for all molecular descriptors. A formal definition of the molecular descriptor is
given by Todeschini and Consonni as follows: molecular descriptor is the final result of a logic and mathematical procedure which
transforms chemical information encoded within a symbolic representation of a molecule into a useful number or the result of
some standardized experiment (15). Here the term useful has
two meanings: the number can give more insight into the interpretation of the molecular properties and/or it is able to take
part in a model for the prediction of some interesting property of
other molecules.
Molecular descriptors vary greatly in both their origins and
their applications. They come from both experimental measurements and theoretical computations. Typical molecular descriptors from experimental measurements include logP, aqueous
solubility, molar refractivity, dipole moment, polarizability, Hammett substituent constants, and other empirical physicochemical
properties. Notice that the majority of experimental descriptors are for entire molecules and come directly from experimental measurements. A few of them, such as various substituent
constants, are for molecular fragments attached to certain
molecular templates and they are derived from experimental
results.

34

Zhou

Theoretical molecular descriptors cover much broader varieties and usually are more readily available even though the complexity of their computational procedures may vary widely. Major
classes of computed molecular descriptors include the following:
(i) Constitutional counts such as molecular weight, number
of heavy atoms, number of rotatable bonds, number of
rings, and number of aromatic rings.
(ii) 2D molecular properties such as number of hydrogen
bond donor/acceptor and their strengths, number of
polar atoms.
(iii) Topological descriptors from graph theory such as various
graph-theoretic invariants of molecular graphs, 2D and
3D autocorrelations, various property-weighted graphtheoretic quantities.
(iv) Geometrical descriptors such as shape, radius of gyration, moments of inertia, volume, polar/nonpolar surface
areas.
(v) Electrostatic properties such as dipole moment, partial
atomic charges.
(vi) Fingerprints such as 2D fingerprints like Daylight fingerprints and UNITY fingerprints and 3D fingerprints like
pharmacophore fingerprints.
(vii) Quantum chemical descriptors such as HOMO/LUMO
energies, E-state values.
(viii) Predicted physicochemical properties such as calculated
solubility, calculated logP, and various molecular properties from QSPR predictive models.
There are also various hybrid descriptors. For example, electrotopological descriptors are a hybrid of topological and electronic descriptors.
Applications of molecular descriptors are as diverse as their
definitions. The important classes of applications include QSAR
and/or QSPR, similarity, diversity, predictive models for virtual
screening and/or data mining, data visualization. We will discuss
briefly some of these applications in the next sections.
There are literally thousands of molecular descriptors available
for various applications. We have only mentioned a few of them
in previous paragraphs. Interested readers can find a more complete coverage of molecular descriptors in reference (15), which
gives definitions for 3,300 molecular descriptors. Many software,
or subroutines as an integral part of other programs, are available
to generate various types of molecular descriptors. Table 2.2 lists
a few of these software.
2.4. Chemical Space
and Dimension
Reduction

Molecular descriptors for a given molecule can be considered

as its coordinates in a multidimensional chemical space. Since

Type of descriptors

Topological, electronic, geometric, and some combination

Constitutional, functional group counts, topological, Estate, Moriguchi descriptors, Meylan flags, molecular
patterns, electronic properties, 3D descriptors, hydrogen bonding, acidbase ionization, empirical estimates
of quantum descriptors

Global physicochemical descriptors, size and shape

descriptors, atom property-weighted 2D and 3D
autocorrelations and RDF, surface property-weighted
autocorrelations

Constitutional, topological, geometrical, electrostatic,

surface property, quantum chemical, and thermodynamic descriptors

Constitutional descriptors, topological descriptors, walk

and path counts, connectivity indices, information indices, 2D autocorrelations, edge adjacency
indices, BCUT descriptors, topological charge indices,
eigenvalue-based indices, Randic molecular profiles,
geometrical descriptors, RDF descriptors, 3D-MoRSE
descriptors, WHIM descriptors, GETAWAY descriptors, functional group counts, atom-lefted fragments,
charge descriptors, molecular properties, 2D binary
fingerprints, 2D frequency fingerprints

Software name

ADAPT

ADMET
Predictor

ADRIANA.code

CODESSA

DRAGON

Table 2.2
A selected list of software for computing molecular descriptorsa

3,224

1,500

1,244

297

>264

Number of
descriptors

Talete srl, Milano, Italy

Alan R. Katritzky, Mati

Karelson, and Ruslan
Petrukhin, University of
Florida

Molecular Networks

Simulations Plus

Peter Jurs, Penn State University

Distributor (and/or
author)

Reference for its web version,

E-DRAGON: Tetko, I. V.,
et al. (2005) J Comput Aid
Mol Des 19, 453463.
http://www.talete.mi.it/
products/dragon_
description.htm

http://www.codessapro.com/index.htm

http://www.molecularnetworks.com/products/
adrianacode

http://www.simulationsplus.com/

http://research.chem.psu.
edu/pcjgroup/
desccode.html

Reference

Chemoinformatics and Library Design

35

Constitutional, topological, and geometrical

Constitutional,
BCUT, etc.

Constitutional, topological,
physicochemical, etc.

Constitutional, property-based, 2D topological, and 3D conformational descriptors

Topological

Constitutional and topological

Molgen-QSPR

PowerMV

PreADMET

Sarchitect

TAM

TOPIX

130

>20

1,084

1,081

>1,000

708

>40

>600

68

Number of
descriptors

a See complete list at http://www.moleculardescriptors.eu/softwares/softwares.htm

geometrical,

fingerprints,

Topological and E-state

Molconn-Z

pairs,

Topological, structural keys, E-state, physical

properties, surface area descriptors including CCGs VSA descriptors, etc.

MOE

atom

Counting, topological, geometrical, properties, etc.

Type of
descriptors

JOELib

Software name

Table 2.2
(continued)

D. Svozil and H. Lohninger, Epina

Software Labs, Austria

M. aric-Medic. et al., University

of Zagreb, Croatia

Strand Life Sciences, India

Bioinformatics
&
Molecular
Design Research Center, South
Korea

J. Liu, J. Feng, A. Brooks, and S.

Young, National Institute of Statistical Sciences, USA

J. Braun, M. Meringer, and C.

Rcker, University of Bayreuth,
Germany

eduSoft

Chemical Computing Group

J.K. Wegner, University of Tbingen, Germany

Distributor (and/or
author)

http://www.lohninger.com/
topix.html

Vedrina, M., et al. (1997) Computers & Chem 21, 355361

http://www.strandls.com/
sarchitect/index.html

http://preadmet.bmdrc.org/
index.php?option=com_
content&view=
frontpage&Itemid=1

http://nisla05.niss.org/
PowerMV/

http://www.edusoftlc.com/molconn/
http://www.molgen.de/
?src=documents/
molgenqspr.html

http://www.ra.cs.unituebingen.de/software/
joelib/index.html
http://www.chemcomp.com/

Reference

36
Zhou

Chemoinformatics and Library Design

37

value ranges for different descriptors may substantially differ for a

given data set, it is desirable to scale (or normalize) descriptors
selected before any mathematical manipulations. Another scenario for rescaling descriptors is to use weighting factors to differentiate important descriptors from unimportant ones. Therefore,
a scaled individual descriptor is represented by one dimension in
this multidimensional space and each molecule is represented by a
single point in such a space. The distance between two molecules
is often defined as their Euclidean distance in this space.
Chemical space so defined is highly degenerate because of the
high redundancy of various molecular descriptors. For example,
molecular weight is highly correlated with the number of heavy
atoms. The high degeneracy along with the high dimensionality of the molecular descriptor space poses a real challenge to
many applications of molecular descriptors. Therefore, dimension
reduction of a chemical space is not only important to identifying key factors affecting the trends in various predictive models
but also necessary for efficient mathematical manipulations during model developments.
It is evidently beneficial and easy to remove those trivial descriptors with constant or near-constant values across all
molecules. To further eliminate duplication and redundancy of
descriptors for a given data set, statistical methods, such as principal component analysis (PCA) (16), multidimensional scaling
(MDS) (17), or nonlinear mapping (NLM) (18), can be very
helpful for dimensionality reduction. PCA is a method of identifying patterns in a data set and expressing the data in such a
way as to highlight their similarities and differences. It is able
to find linear combinations of the variables (the so-called principal components) that correspond to directions of the maximal
spread in the data. On the other hand, MDS is a method that
represents measurements of similarity (or dissimilarity) among
pairs of objects as distances between points of a low-dimensional
multidimensional space. It preserves the original pairwise interrelationships as closely as possible. Finally, NLM tries to preserve distances between points as similar as possible to the actual
distances in the original space. The NLM procedure for performing this transformation is as follows: compute interpoint
distances in the original space; choose an initial configuration
(generally random) in the display space (i.e., the target and lowdimensional space); calculate a mapping error from the distances
in the two spaces; and modify iteratively the coordinates of points
in the display space by means of a nonlinear procedure so as
to minimize the mapping error. PCA is a linear method while
both MDS and NLM are nonlinear methods. All these methods
endeavor to optimally preserve information while reducing the
dimensionality of the descriptor space (hence the mathematical
complexity).

38

Zhou

Reducing the dimensionality of the descriptor space not only

facilitates model building with molecular descriptors but also
makes data visualization and identification of key variables in various models possible. Notice that while a low dimension mathematically simplifies a problem such as model development or
data visualization, it is usually more difficult to correlate trends
directly with physical descriptors, and hence the data become
less interpretable, after the dimension transformation. Trends
directly linked with physical descriptors provide simple guidance
for molecular modifications during potency/property optimizations.
2.5. Similarity
and Diversity

Molecules with similar structures should behave similarly while

it is more efficient to use a diverse set of compounds to cover a
broad range of chemical space. Chemical similarity and diversity
are interesting because even a fuzzy understanding of these concepts can aid the design of useful molecules. For example, similarity probe is essential to analogue designs during lead optimization while enough diversity of a chemical collection is critical to
the successful lead generation through high-throughput screening (HTS) (19).
The quantification of molecular similarity generally involves
three components: molecular descriptors to characterize the
molecules, weighting factors to differentiate more important
characteristics from less important ones, and the similarity coefficient to quantify the degree of similarity between pairs of
molecules (20, 21). The first two components are related to the
definition of chemical space as discussed in Section 2.4. Therefore, it is natural to assume that structurally similar molecules
should cluster together in a chemical space, and to define the similarity coefficient of a pair of molecules to be the distance between
them in the chemical space. The shorter the distance is the more
similar the pair is.
Because of the numerous choices for molecular descriptors,
weighting factors, and similarity coefficients, there are many ways
in which the similarities between pairs of molecules can be calculated. The most used molecular descriptors for defining similarity are probably the 2D fingerprints (22). The bit strings of the
molecular fingerprints are used to calculate similarity coefficients.
Table 2.3 lists several selected similarity coefficients that can be
used with various 2D fingerprints (23). The Tanimoto coefficient
is the most popular one (22).
A related concept to similarity is dissimilarity. Dissimilarity
can be considered as the opposite of similarity. It is also defined
by the distance between two molecules in a chemical space. The
larger the distance between the two molecules is the more dissimilar the pair is. Sometimes, dissimilarity is used interchangeably with diversity in literature even though there are subtle

Chemoinformatics and Library Design

39

Table 2.3
Selected similarity coefficients to be used with 2D fingerprints for
molecule pair (A, B)
Coefficient

Expressiona

Value range

Tanimoto

c
a+b+c

0.01.0

Cosine

0.01.0

Hamming

a+b

0.0

RussellRao

c
a+b+c+d

0.01.0

Forbes

(a+b+c+d)c
(a+c)(b+c)

0.0

Pearson

1.01.0

Simpson

c
min{(a+c), (b+c)}


Euclid

c
(a+b)(b+c)

cdab
(a+c)(b+c)(a+d)(d+d)

c+d
a+b+c+d

Notes

This is a dissimilarity
coefficient

0.01.0
0.01.0

a a is the count of bits that is on in A string but off in B string; b is the count of bits that is off
in A string buton in B string; c is the count of bits that is on in both A string and B string; d is
the count of bits that is off in both A string and B string.

differences between diversity and dissimilarity. Diversity is a property of a molecular collection while dissimilarity can be defined
for pairs of molecules as well.
Since diversity is a collective property, its precise quantification requires a mathematical description of the distribution of the
molecular collection in a chemical space. When a set of molecules
are considered to be more diverse than another, the molecules
in this set cover more chemical space and/or the molecules distribute more evenly in chemical space. Historically, diversity analysis is closely linked to compound selection and combinatorial
library design. In reality, library design is also a selection process,
selecting compounds from a virtual library before synthesis. There
are three main categories of selection procedures for building a
diverse set of compounds: cluster-based selection, partition-based
selection, and dissimilarity-based selection.
The cluster-based selection procedure starts with classifying
compounds into clusters of similar molecules with a clustering
algorithm followed by selection of representative(s) from each
cluster (24). On the other hand, the partition-based selection
procedure partitions chemical space into cells by dividing values
of each dimension into various intervals and selects representative

40

Zhou

compounds from each cell (25). Because of the exponential

dependence of cell numbers on dimensions of the chemical space,
the partition-based selection procedure is only suitable for applications in a low-dimensional chemical space. Hence, most representative molecular descriptors need to be identified to form
the chemical space, or the dimension reduction as described in
Section 2.4 needs to be performed before the partition-based
selection procedure can be used. In addition to the cell-based
partitioning, statistical partitioning methods, such as decision
tree method (26), are also used for classification. Finally, the
dissimilarity-based selection procedure iteratively selects compounds that are as dissimilar as possible to those already selected
(27). This method tends to select molecules with more complexity as well as a diverse set of chemical cores. For combinatorial
library design, there is also an optimization-based selection procedure to select compounds from virtual libraries. It formulates the
compound selection as an optimization problem with some quantitative measures of diversity (see, for example, reference (28)).
2.6. QSAR and QSPR

Building predictive QSAR and QSPR models is a cost-effective

way to estimate biological activities, physicochemical properties
such as partition coefficients and solubility, and more complicated
pharmaceutical endpoints such as metabolic stability and volume
of distribution. It seems to be reasonable to assume that structurally similar molecules should behave similarly. That is, similar molecules should have similar biological activities and physicochemical properties. This is the (Q)SAR/(Q)SPR hypothesis.
Qualitatively, both molecular interactions and molecular properties are determined by, and therefore are functions of, molecular
structures. Or
Activity = f1 (mol structure/descriptors)

[1]

Property = f2 (mol structure/descriptors)

[2]

and

There is a long history of efforts to find simple and interpretable f1 and f2 functions for various activities and properties
(29, 30). The quest for predictive QSAR models started with
Hammetts pioneer work to correlate molecular structures with
chemical reactivities (3032). However, the widespread applications of modern predictive QSAR and QSPR actually started
with the seminal work of Hansch and coworkers on pesticides
(29, 33, 34) and the developments of various powerful analysis tools, such as PLS (partial least squares) and neural networks,
for multivariate analysis have fueled these widespread applications.
Nowadays, numerous publications on guidelines, workflows, and

Chemoinformatics and Library Design

41

common errors for building predictive QSAR and QSPR models, not to mention the countless papers of applications, are well
documented in literature (3541).
In principle, a valid QSAR/QSPR model should contain the
following information (39): (i) a defined endpoint; (ii) an unambiguous algorithm; (iii) a defined domain of applicability; (iv)
appropriate measures of goodness of fit, robustness, and predictivity; and (v) a mechanistic interpretation, if possible. Building
predictive QSAR/QSPR models is a process from experimental data to model and to predictions. Collecting reliable experimental data (and subdividing the data into training set and
testing set) is the first step of the model-building process. The
second step of the process is usually to select relevant parameters
(or molecular descriptors) that are most responsive to the variation of activities (or properties) in the data set. The third step is
QSAR/QSPR modeling and model validation. Finally, the validated models are applied to make predictions. Usually, the second and the third, and sometimes the first, steps are repeated
to select the best combination of parameter set and models (see,
for example, reference (40)). Although majority of QSAR/QSPR
models are built with molecular descriptors, there are parameterfree models. For example, the FreeWilson method builds predictive QSAR/QSPR models for a series of substituted compounds
without any molecular descriptors (42). Its drawback is that the
FreeWilson method requires a data set for almost all combinations of substituents at all substituted sites and the method is not
applicable to molecular set of noncongeners.
It is interesting to note that various QSAR/QSPR models
from an array of methods can be very different in both complexity and predictivity. For example, a simple QSPR equation with
three parameters can predict logP within one unit of measured
values (43) while a complex hybrid mixture discriminant analysis
random forest model with 31 computed descriptors can only predict the volume of distribution of drugs in humans within about
twofolds of experimental values (44). The volume of distribution
is a more complex property than partition coefficient. The former is a physiological property and has a much higher uncertainty
in its experimental measurements while logP is a much simpler
physicochemical property and can be measured more accurately.
These and other factors can dictate whether a good predictive
model can be built.
2.7. Multiobjective
Optimization

The ultimate goal of a small-molecule drug discovery program

is to establish an acceptable pharmacological profile for a drug
candidate. To achieve this goal, usually many pharmacological
attributes, or their numerous surrogates, of individual lead compounds need to be optimized either sequentially or in parallel.
That is, drug discovery itself is a multiobjective optimization

42

Zhou

process (45). Furthermore, the multiobjective optimization is

also involved in both various stages of the drug discovery process and many drug discovery enabling technologies. For example, to design libraries for lead generation or lead optimization,
multiple physicochemical properties need to be optimized along
with diversity and similarity (4649). It is also a common practice to test multiple hypotheses in a single SAR/SPR run during lead optimization. The algorithms for solving these various
multiobjective optimization problems can be quite similar even
though the properties to be optimized are evaluated very differently, ranging from simple computations to complex in vivo
experiments.
When optimizing multiple objectives, usually there is no best
solution that has optimal values for all, and oftentimes competing,
objectives. Instead, some compromises need to be made among
various objectives. If a solution A is better than another solution B for every objective, then solution B is dominated
by A. If a solution is not dominated by any other solution,
then it is a nondominated solution. These nondominated solutions are called Pareto-optimal solutions, and very good compromises for a multiobjective optimization problem can be chosen among this set of solutions. Many methods have been developed and continue to be developed to find Pareto-optimal solutions and/or their approximations (see, for example, references
(5052)). Notice that solutions in the Pareto-optimal set cannot be improved on one objective without compromising another
objective.
Searching for Pareto-optimal solutions can be computationally very expensive, especially when too many objectives are to be
optimized. Therefore, it is very appealing to convert a multiobjective optimization problem into a much simpler single-objective
optimization problem by combining the multiple objectives into
a single objective function as follows (5355):

F (Obj1 , Obj2 , . . . , Objn ) =

n


wi fi (Obj1 )

[3]

i=1

where wi is a weighting factor that reflects the relative importance

of ith objective among all objectives. With this conversion, all
algorithms used for single-objective optimizations can be applied
to find optimal solutions as prescribed by equation [3]. Notice
that both functional forms for {fi } and weighting factors wi in
equation [3] may be attenuated to achieve optimal results, when
enough data are available for testing and validating (55).
It is a common practice in early drug discovery to select
compounds by some very simple filters such as rule-of-five
and drug-likeness (56, 57). For example, multiobjective

Chemoinformatics and Library Design

43

optimization methods have been applied to design combinatorial

libraries of drug-like compounds (53, 54).
2.8. Virtual Screening

Virtual screening (VS) has emerged as an important tool for drug

discovery (5867). The goal of VS is to separate active from inactive molecules in a compound collection and/or a virtual library
through rapid in silico evaluations of their activities against a
biological target. A full VS process generally involves three components: a library to be screened, an in silico assay to test the
activities of molecules in the library, and a hit follow-up plan to
experimentally verify the activities (see Fig. 2.2). The in silico
assay is the core component of VS. The other two components
are also very important for a successful VS campaign.

Library
For VS

Library: compound collection or

virtual library
Pre-VS filtering:
Druglikeness/leadlikeness
Target-specific criteria, etc

In-silico
Assay

Virtual Hit
Follow-up

Structure-based:
Docking

Synthesis
if needed

Ligand-based:
Similarity
clustering
Pharmacophore
QSAR models
etc

Experimental
validation of activity
etc

Fig. 2.2. Three components of a typical VS process: compound library, virtual assay,
and hit follow-up for virtual hits.

A compound library for VS could be a corporate compound

collection, a public compound collection such as NCIs compound library (68), a collection of commercially available compounds, or a virtual library of synthesizable compounds. Nowadays, a corporate compound collection has a typical size of 106
compounds. More often than not, various filters are applied, for
example, to filter out non-drug-like/non-lead-like compounds,
thereby to reduce the library size for VS (56, 57, 6467). Prefiltering becomes imperative for screening large virtual libraries
within a reasonable period of time. It is obviously crucial to have
target-relevant molecules in the library. Actually, library design in
large part is to cover enough chemical space of these biologically
relevant molecules.

44

Zhou

Computational methods acting as in silico assay can be

roughly classified into two major categories: structure based and
ligand based (5867). The structure-based methods require the
knowledge of target structures. The most common structurebased approach is to dock each small molecule onto the active site
of a target structure to determine its binding affinity (or docking score) to the target (69). A wide array of docking methods
and their associated scoring functions are available for screening
large libraries (70). The less-used methods in structure-based virtual screening include VS with pharmacophores built from target structures and the low-throughput free energy computations
for ligandreceptor complexes via molecular dynamics or Monte
Carlo simulations.
On the other hand, starting from knowledge about active ligands and optionally inactive compounds, various computational
methods can be used to find related active compounds. These
methods include the following:
(i) Nearest-neighbor methods such as similarity methods and
clustering methods (assuming chemically similar compounds behave similarly biologically)
(ii) Predictive QSAR model built from actives and optionally
inactives (40)
(iii) Pharmacophore models built from actives and inactives
(71)
(iv) Machine learning methods such as classification, decision
tree, support vector machine, and neural networks (72)
Virtual hits need to be synthesized for hits from virtual
libraries and their bioactivities experimentally verified for VS to
have any real impact. More importantly, these virtual hit followup steps can act as a validation stage for the computational models
and the associated VS protocol. The results of experimental verification can be fed back to the in silico assay stage for building
better predictive in silico models.

3. Library and
Library Design
3.1. Compound
Library for Drug
Discovery

There are two major classes of libraries for drug discovery: diverse
libraries for lead discovery and focused libraries for lead optimization. Lead discovery libraries emphasize diversity while lead
optimization libraries prefer similar compounds. The purpose of
lead discovery libraries is to find lead matter and to provide
potential active compounds for further optimization. Without any
prior knowledge about the active compounds for a given target, it is reasonable to start with a library of enough chemical
space coverage to demarcate the biologically relevant chemical

Chemoinformatics and Library Design

45

space for the target. Therefore, libraries for lead discovery often
comprise diverse compounds with drug-like/lead-like properties.
Lead matters without proper drug-likeness/lead-likeness properties might be trapped in a local and unoptimizable zone of
the chemical space during lead optimization stage. On the other
hand, the purpose of lead optimization libraries is to improve the
activity and the property profile of the lead matter. With a lead
compound, searching for better and optimized compounds is usually performed among similar compounds with limited diversity
around the lead molecule in the chemical space.
There are three major sources for a typical corporate compound collection: project-specific compounds accumulated over
a long period of time through medicinal chemistry efforts for various therapeutic projects, individual compounds from commercial sources, and compounds from combinatorial chemistry. In
practice, compound collections are often divided into subsets, for
example, the diverse subsets for general HTS and target-focused
subsets (such as kinase libraries or GPCR libraries). For library
design, diversity and similarity are generally built into the libraries
of compounds to be synthesized and/or purchased (73).
Stimulated by the widespread applications of HTS technologies, combinatorial chemistry has provided a powerful tool for
rapidly adding large number of compounds to corporate collections for many pharmaceutical companies. Virtual combinatorial
library consists of libraries from individual reactions and compounds from a single reaction share a common product core
(see Fig. 2.3). The number of compounds in a combinatorial
library can grow rapidly with number of reaction components
and numbers of reactants for individual components. For example, a full combinatorial library from a three-component reaction

Virtual
Combinatorial
Library

Compounds of
core 1 from
Reaction 1
R2

R1

Compounds of
core 2 from
Reaction 2
R2

R1

Compounds of
core N from
Reaction N
R2

R1

R3

Fig. 2.3. Virtual combinatorial library is the start point for any combinatorial library
design. It consists of libraries from individual reactions. Compounds from a given
reaction share a unique product core.

46

Zhou

with 200 reactants for each component would contain 8 million products. Virtual library can also be represented as a template with R-groups attached at various variation sites. This representation is also called Markush structure. Markush structure is
the standard chemical structure often used in chemical patents.
Template-based libraries can be considered as a generalization
of the scenario shown in Fig. 2.3 where the product cores of
individual reactions are the templates. Notice that reaction-based
virtual libraries have explicit chemistries for compound syntheses
and therefore may include only those synthesizable compounds
through careful selections of reactants while general templatebased virtual libraries usually do not indicate chemistry accessibilities of the compounds.
3.2. Library
Enumeration

The product structures of a combinatorial library can be formed

from product core and structures of reactants or by attaching
R-groups to the various variation sites of a template (see Fig. 2.4).
Product formation is conventionally called product enumeration.
It is accomplished by removing the leaving groups of reactants,
a process also called clipping, followed by pasting the retained
fragments at the variation sites of the product core or the template. For template-based enumerations, the R-groups, generated
either by molecular fragmentation programs or from molecular
clipping, are usually listed as part of the library definition. There
are many automatic tools for library enumeration, either as standalone software or as subroutines of other application packages (see,
for example, references (7478)).
With product structures, many chemoinformatics tools can
be applied to filter the virtual libraries and to select a few use-

Reaction-based enumeration runs through

independent reaction components
List runs through
all reactant B
R3
R3
R1

R2

R1

List runs through

all reactant A

R2

Product

R3
R3

Template-based enumeration runs

through independent R-groups
R1
R1

R2
R2

Fig. 2.4. Product enumerations of a combinatorial library. For reaction-based enumeration, individual groups of N(R1)(R2) and (CO)-R3 are replaced by corresponding molecular fragments from reactants A and B. For template-based enumeration,
the R-groups R1, R2, and R3 are replaced by independent lists of molecular fragments. Note that some combinations of R1 and R2 may not exist in component A for
reaction-based enumerations. The template-based product structure with R-groups is
also called Markush structure and its enumeration is called Markush enumeration or
Markush exemplification.

Chemoinformatics and Library Design

47

ful compounds for syntheses. These filtering tools include the

various tools and methods discussed in the previous section (see
Sections 2.52.8). Multiobjective optimization algorithms can
be used to design combinatorial libraries with optimal diversity/similarity, cost efficiency, and physical properties (4647).
Nevertheless, library design can be performed without the full
enumeration of the entire virtual combinatorial libraries (7980).
3.3. Library Design

Library design is a compound selection process that maximizes

the number of compounds with desirable attributes while minimizing the number of compounds with undesirable characteristics. The desirability of compounds in a library is defined
by the ultimate usage of the library and the cost efficiency
for producing the library. Therefore, libraries for lead discovery demand sufficient diversity among compounds selected while
lead optimization libraries usually contain compounds similar
to those lead compounds. The diversity of a compound collection can be improved through inclusion of more diverse chemotypes/scaffolds and side chains (8182). Diversity in chemotypes
and scaffolds is usually derived from more reactions with novel
chemistries. Diversity in side chains can be achieved by selecting
more diverse reagents for a given reaction. It is well recognized
that probability of finding effective ligandreceptor interactions
decreases as a molecule becomes more complex (83). That is, relatively simple molecules from diverse chemotypes/scaffolds have a
better chance than those complex molecules derived from diverse
side chains to generate lead matter with more specific ligand
receptor interactions. Therefore, the current practice of building a
diverse compound collection prefers more small libraries of many
diverse novel chemistries to less large libraries of a few chemistries.
Another important consideration in designing a library is cost
efficiency. Inexpensive reagents should always be favored as reactants for library production. Producing a library of a full combinatorial array is much more cost-effective than synthesizing cherrypicked singletons. Selections in a library design can be product
based or reactant based. In a product-based design library, compounds are chosen purely based on their own properties. On the
other hand, the reactant-based design chooses reactants, instead
of library products directly, based on the collective properties of
the associated products (47, 84). Reactant-based design generally leads to libraries of full combinatorial arrays. While costly,
product-based design is more effective than reactant-based design
in achieving optimal design objectives other than cost (47, 84).
This seems to be obvious since limiting product choices to a subarray of a full combinatorial library will compromise other design
objectives, unless the selected reactants are so dominant that the
products derived from their combinations are superior to other
products with respect to all objectives.

48

Zhou

Frequently, library design involves simultaneous optimization

of multiple objectives, among which diversity, similarity, and cost
efficiency are three examples. Other typical properties include the
rule-of-five properties (molecular weight, logP value, number
of hydrogen bond donors, and total number of N and O
atoms), polar surface area, and solubility. Complicated properties from predictive models can also be included. Library design
in large part is actually a multiobjective optimization problem.
Therefore, all methods discussed in Section 2.7 can be applied to
library design.
To summarize, library design involves choices of diversity vs.
similarity, product based vs. reactant based, and single objective
vs. multiobjective optimizations. Chemoinformatics tools, such
as various predictive models and chemoinformatics infrastructures, can be utilized to facilitate the selection process of library
design.

4. Concluding
Remarks
Library design has become an integral part of drug discovery process. Chemical library design underwent a transformation from a
pure tool for supplying vast number of compounds to a power
tool for generating quality leads and drug candidates. Although
the controversy of how to define a best set of compounds for lead
generation is not completely resolved, tremendous progress has
been made to find biologically relevant subregions of the chemical space, particularly when confined to a target or a target family
(see, for example, references (85, 86)). Providing biologically relevant compounds will continue to be one of the main goals of
library design.
Since modern drug discovery is mainly a data-driven process and chemoinformatics is at the center of data integration
and utilization, it is natural that majority of library design tools
are chemoinformatics tools. Therefore, a deep understanding of
chemoinformatics is necessary for taking full advantage of library
technologies.
Though relatively mature, chemoinformatics is still an active
field of intensive research. Numerous new methods and tools continue to be developed. Here we have selectively covered, without
giving too many details, a few topics important to library design.
Actually the interplays and costimulations of chemoinformatics
with library design have been well documented in literature. We
hope that the brief introduction in this chapter can serve as a
guide for you to enter into the exciting field of chemoinformatics
and its applications to chemical library design.

Chemoinformatics and Library Design

49

Acknowledgment
The chapter was prepared when the author was visiting with professor Andy McCammons group. The author is very grateful to
Professor Andy McCammon and his group for the exciting and
stimulating scientific environment during the preparation of the
chapter.
References
1. Brown, F. B. (1998) Chemoinformatics: what is it and how does it impact
drug discovery. Annu Rep Med Chem 33,
375384.
2. Bohacek, R. S., McMartin, C., Guida, W.
C. (1996) The art and practice of structurebased drug design: a molecular modeling perspective. Med Res Rev 16, 350.
3. Walters, W. P., Stahl, M. T., Murcho, M. A.
(1998) Virtual screeningan overview. Drug
Discov Today 3, 160178.
4. Gasteiger, J. (ed.) (2003) Handbook of
Chemoinformatics: From Data to Knowledge,
Wiley-VCH, Weinhiem.
5. Bajorath, J. (ed.) (2004) Chemoinformatics:
Concepts, Methods, and Tools for Drug Discovery, Humana Press, Totowa, NJ.
6. Oprea, T. I. (ed.) (2005) Chemoinformatics
in Drug Discovery, Wiley-VCH, Weinheim.
7. Leach, A. R. and Gillet, V. J. (2007) An
Introduction to Chemoinformatics, Springer,
London.
8. Bunin, B. A., Siesel, B., Morales, G. A., Bajorath, J. (2007) Chemoinformatics: Theory,
Practice, & Products, Springer, The Netherlands.
9. http://www.symyx.com/solutions/white_
papers/ctfile_formats.jsp,
last
accessed
February, 2010.
10. Weininger, D. (1988) SMILES, a chemical
language and information system. 1. Introduction to methodology and encoding rules.
J Chem Inf Comput Sci 28, 3136.
11. Weininger, D. (1989) SMILES, 2. Algorithm
for generation of unique SMILES notation. J
Chem Inf Comput Sci 29, 97101.
12. Weininger, D. (1990) SMILES, 3. Depict.
Graphical depiction of chemical structures. J
Chem Inf Comput Sci 30, 237243.
13. http://www.daylight.com/dayhtml/doc/
theory/theory.smiles.html, last accessed
February, 2010.
14. Simsion, G. C., Witt, G. C. (2001) Data
Modeling Essentials, 2nd ed. Coriolis, Scottsdale, USA.

15. Todeschini, R., Consonni, V. (2009) Molecular Descriptors for Chemoinformatics Vol. 1,
2nd ed. Wiley-VCH, Weinheim, Germany.
16. Jolliffe, I. T. (2002) Principal Component
Analysis, 2nd ed. Springer, New York.
17. Borg, I. and Groenen, P. J. F. (2005) Modern
Multidimensional Scaling: Theory and Applications, 2nd ed. Springer, New York.
18. Domine, D., Devillers, J., Chastrette, M.,
Karcher, W. (1993) Non-linear mapping
for structure-activity and structure-property
modeling. J Chemometrics 7, 227242.
19. Wermuth, C. G. (2006) Similarity in drugs:
reflections on analogue design. Drug Discov
Today 11, 348354.
20. Willett, P. (2000) Chemoinformatics
similarity and diversity in chemical libraries.
Curr Opin Biotech 11, 8588.
21. Maldonado, A. G., Doucet, J. P., Petitjean,
M., Fan, B. -T. (2006) Molecular similarity
and diversity in chemoinformatics: from theory to applications. Mol Divers 10, 3979.
22. Willett, P. (2006) Similarity-based virtual
screening using 2D fingerprints. Drug Discov
Today 11, 10461053.
23. Holliday, J. D., Hu, C. -Y., Willett, P. (2002)
Grouping of coefficients for the calculation
of inter-molecular similarity and dissimilarity using 2D fragment bitstrings. Comb Chem
High Throughput Screening 5, 155166.
24. Dunbar J. B. (1997) Cluster-based selection.
Perspect Drug Discov Des 7/8, 5163.
25. Mason J. S., Pickett S. D. (1997) Partitionbased selection Perspect Drug Discov Des
7/8, 85114.
26. Rusinko, A. III, Farmen, M. W., Lambert,
C. G. et al. (1999) Analysis of a large structure/biological activity dataset using recursive partitioning. J Chem Inf Comput Sci 39,
10171026.
27. Lajiness, M. S. (1997) Dissimilarity-based
compound selection techniques. Perspect
Drug Discov Des 7/8, 6584.
28. Pickett, S. D., Luttman, C., Guerin, V.,
Laoui, A., James, E. (1998) DIVSEL and

50

29.

30.
31.
32.

33.

34.
35.

36.

37.

38.
39.

40.

41.

Zhou
COMPLIBstrategies for the design and
comparison of combinatorial libraries using
pharmacophore descriptors. J Chem Inf Comput Sci 38, 144150.
Hansch, C., Hoekman, D., Gao, H. (1996)
Comparative QSAR: toward a deeper understanding of chemicobiological interactions.
Chem Rev 96, 10451074.
Jaff, H. H. (1953) A reexamination of
the Hammett equation. Chem Rev 53,
191261.
Hammett, L. P. (1935) Some relations
between reaction rates and equilibrium.
Chem Rev 17, 125136.
Hammett, L. P. (1937) The effect of structure upon the reactions of organic compounds. Benzene derivatives. J Am Chem Soc
59, 96103.
Hansch, C., Maloney, P. P., Fujita, T., Muir,
R. M. (1962) Correlation of biological activity of phenoxyacetic acids with Hammett
substituent constants and partition coefficients. Nature 194, 178180.
Hansch, C. (1993) Quantitative structureactivity relationships and the unnamed science. Acc Chem Res 26, 147153.
Livingstone, D. J. (2004) Building QSAR
models: a practical guide, in (Cronin, M.
T. D., Livingstone, D. J. eds.) Predicting
Chemical Toxicity and Fate. CRC Press, Boca
Raton, FL, 2004, pp. 151170.
Walker, J. D., Dearden, J. C., Schultz, T.
W., Jaworska, J., Comber M. H. I. (2003)
in (Walker, J. D. ed.) QSARs for New Practitioners, in QSARs for Pollution Prevention,
Toxicity Screening, Risk Assessment, and Web
Applications. SETAC Press, Pensacola, FL,
pp. 318.
Walker, J. D., Jaworska, J., Comber, M. H. I.,
Schultz, T. W., Dearden, J. C. (2003) Guidelines for developing and using quantitative
structureactivity relationships. Environ Toxicol Chem 22, 16531665.
Cronin, M. T. D., Schultz, T. W. (2003)
Pitfalls in QSAR J Theoret Chem (Theochem)
622, 3951.
OECD Principles for the Validation of
(Q)SARs, http://www.oecd.org/dataoecd/
33/37/37849783.pdf, last accessed February, 2010.
Tropsha, A., Golbraikh, A. (2007) Predictive QSAR modeling workflow, model applicability domains, and virtual screening. Curr
Pharmaceut Design 13, 34943504.
Dearden, J. C., Cronin, M. T. D., Kaiser,
K. L. E. (2009) How not to develop a
quantitative structure-activity or structureproperty relationship (QSAR/QSPR). SAR
and QSAR in Environ Res 20, 241266.

42. Free, S. M., Wilson, J. W. (1964) A mathematical contribution to structure-activity

studies. J Med Chem 7, 395399.
43. Xing, L., Glen, R. C. (2002) Novel methods
for the prediction of logP, pKa , and logD. J
Chem Inf Comput Sci 42, 796805.
44. Lombardo, F., Obach, R. S., et al. (2006) A
hybrid mixture discriminant analysis-random
forest computational model for the prediction of volume of distribution of drugs in
human. J Med Chem 49, 22622267.
45. Nicolaou, C. A., Brown, N., Pattichis, C.
S. (2007) Molecular optimization using
computational multi-objective methods Curr
Opin Drug Discov Develop 10, 316324.
46. Gillet, V. J., Willett, P., Bradshaw, J., Green,
D. V. S. (1999) Selecting combinatorial
libraries to optimize diversity and physical properties. J Chem Inf Comput Sci 39,
169177.
47. Brown, R.D., Hassan, M., Waldman, M.
(2000) Combinatorial library design for
diversity, cost efficiency, and drug-like characters. J Mol Graph Model 18, 427437.
48. Gillet, V. J., Khatib, W., Willett, P., Fleming,
P. J., Green, D. V. S. (2002) Combinatorial
library design using a multiobjective genetic
algorithm. J Chem Inf Comput Sci 42,
375385.
49. Chen, G., Zheng, S., Luo, X., Shen, J.,
Zhu, W., Liu, H., Gui, C., Zhang, J.,
Zheng, M., Puah, C.M., Chen, K., Jiang, H.
(2005) Focused combinatorial library design
based on structural diversity, drug likeness
and binding affinity score. J Comb Chem 7,
398406.
50. Eichfelder, G. (2008) Adaptive Scalarization Methods in Multiobjective Optimization,
Springer-Verlag, Berlin, Germany.
51. Abraham, A., Jain, L., Goldberg, R. (eds.)
(2005) Evolutionary Multiobjective Optimization: Theoretical Advances and Applications, Springer-Verlag, London, UK.
52. Van Veldhurizen, D. A., Lamont, G. B.
(2000) Multiobjective evolutionary algorithms: analyzing the state-of-the-art. Evol
Comput 8, 125147.
53. Gillet, V. J., Willett, P., Bradshaw, J., Green,
D. V. S. (1999) Selecting combinatorial
libraries to optimize diversity and physical properties. J Chem Inf Comput Sci 39,
169177.
54. Zheng, W., Hung, S. T., Saunders, J. T.,
Seibel, G. L. (2000) PICCOLO: a tool for
combinatorial library design via multicriterion optimization. Pac Symp Biocomput 5,
585596.
55. A multi-endpoint optimization tool with a
graphics user interface developed at PfizerLa

Chemoinformatics and Library Design

56.

57.

58.
59.
60.

61.

62.
63.
64.
65.

66.
67.
68.
69.
70.

71.

Jolla by Zhou, J. Z., Kong, X., Mattaparti, S,

et al. (unpublished).
Lipinski, C. A., Lombardo, F., Dominy, B.
W., Feeney, P. J. (1997) Experimental and
computational approaches to estimate solubility and permeability in drug discovery and
development settings. Adv Drug Deliv Rev
23, 325.
Gillet, V. J., Willett, P., Bradshaw, J. (1998)
Identification of biological activity profiles
using substructural analysis and genetic algorithms. J Chem Inf Comput Sci 38, 165179.
Walter, W. P., Stahl, M. T., Murcko, M. A.
(1998) Virtual screeningan overview. Drug
Discov Today 3, 160178.
Bajorath, J. (2002) Integration of virtual and
high-throughput screening. Nat Rev Drug
Discov 1, 882894.
Reddy, A. S., Pati, S. P., Kumar, P. P.,
Pradeep, H. N., Sastry, G. N. (2007) Virtual
screening in drug discoverya computational
perspective. Curr Prot Pept Sci 8, 329351.
Klebe, G. (ed.) (2000) Virtual Screening: An
Alternative or Complement to High Throughput Screening? Kluwer Academic Publishers,
Boston.
Alvarez, J., Shoichet, B. (ed.) (2005) Virtual
Screening in Drug Discovery, Taylor & Francis, Boca Raton, USA.
Varnek, A., Tropsha, A. (ed.) (2008)
Chemoinformatics: An Approach to Virtual
Screening, RSC, Cambridge, UK.
Rishton, G. M. (1997) Reactive compounds
and in vitro false positives in HTS. Drug Discov Today 2, 382384.
Walters, W. P., et al. (1998) Can we
learn to distinguish between druglike and
nondrug-like molecules? J Med Chem 41,
33143324.
Sadowski, J., Kubinyi, H. A. (1998) A scoring scheme for discriminating between drugs
and nondrugs. J Med Chem 41, 33253329.
Rishton, G. M. (2003) Nonleadlikeness and
leadlikeness in biochemical screening. Drug
Discov Today 8, 8696.
http://dtp.nci.nih.gov/docs/3d_database/
Structural_information/structural_data.html,
last accessed February, 2010.
Kuntz, I. D. (1992) Structure-based strategies for drug design and discovery. Science
257, 10781082.
Kitchen, D. B., Decornez, H., Furr, J. R.,
Bajorath, J. (2004) Docking and scoring in
virtual screening for drug discovery: methods and applications. Nat Rev Drug Discov
3, 935949.
Sun, H. (2008) Pharmacophore-based
virtual screening. Curr Med Chem 15,
10181024.

51

72. Melville, J. L., Burke, E. K., Hirst, J. D.

(2009) Machine learning in virtual screening.
Comb Chem High Throughput Screening 12,
332343.
73. Harper, G., Pickett, S. D., Green, D. V. S.
(2004) Design of a compound screening collection for use in high throughput screening.
Comb Chem High Throughput Screening 7,
6370.
74. Schller, A., Hhnke, V., Schneider, G.
(2007) SmiLib v2.0: a Java-based tool
for rapid combinatorial library enumeration
QSAR. Comb Sci 26, 407410.
75. Pipeline Pilot distributed by Accelrys Inc.
can be used to enumerate libraries defined
either by reactions or by Markush structures:
http://accelrys.com/resource-center/casestudies/enumeration.html, last accessed
February, 2010.
76. CombiLibMaker is software distributed by
Tripos Inc.: http://tripos.com/data/SYBYL/
combilibmaker_072505.pdf, last accessed
February, 2010.
77. Yasri, A., Berthelot, D., Gijsen, H., Thielemans, T., Marichal, P., Engels, M., Hoflack,
J. (2004) REALISIS: a medicinal chemistryoriented reagent selection, library design,
and profiling platform. J Chem Inf Comput
Sci 44, 21992206.
78. (a) Peng, Z., Yang, B., Mattaparti, S., Shulok,
T., Thacher, T., Kong, J., Kostrowicki,
J., Hu, Q., Na, J., Zhou, J. Z., Klatte,
K., Chao, B., Ito, S., Clark, J., Coner,
C., Waller, C., Kuki, A. PGVL Hub:
an integrated desktop tool for medicinal
chemists to streamline design and synthesis of chemical libraries and singleton compounds, in (Zhou, J. Z., ed.) Chemical
Library Design. Humana Press, New York,
Chapter 15.
78. (b) Truchon, J. -F. GLARE: a tool for
product-oriented design of combinatorial
libraries, in (Zhou, J. Z., ed.) Chemical
Library Design. Humana Press, New York,
Chapter 17.
78. (c) Lam, T. H., Bernardo, P. H.,
Chai, C. L. L., Tong, J. C. CLEVER a general design tool for combinatorial libraries, in
(Zhou, J. Z., ed.) Chemical Library Design.
Humana Press, New York, Chapter 18.
79. Shi, S., Peng, Z., Kostrowicki, J., Paderes,
G., Kuki, A. (2000) Efficient combinatorial
filtering for desired molecular properties of
reaction products. J Mol Graph Model 18,
478496.
80. Zhou, J. Z., Shi, S., Na, J., Peng, Z.,
Thacher, T. (2009) Combinatorial librarybased design with basis products. J Comput
Aided Mol Des 23, 725736.

52

Zhou

81. Grabowski, K., Baringhaus, K. -H., Schneider, G. (2008) Scaffold diversity of natural products: inspiration for combinatorial
library design. Nat Prod Rep 25, 892904.
82. Stocks, M. J., Wilden, G. R. H, Pairaudeau,
G., Perry, M. W. D, Steele, J., Stonehous, J. P. (2009) A practical method for
targeted library design balancing lead-like
properties with diversity. ChemMedChem 4,
800808.
83. Hann, M. M., Leach, A. R., Harper, G.
(2001) Molecular complexity and its impact
on the probability of finding leads for
drug discovery. J Chem Inf Comput Sci 41,
856864.

84. Gillet, V. J. (2002) Reactant- and productbased approaches to the design of combinatorial libraries. J Comput Aided Mol Des
16:371380.
85. Balakin, K. V., Ivanenkov, Y. A., Savchuk,
N. P. (2009) Compound library design for
targeted families, in (Jacoby, E. ed.)
Chemogenomics. Humana Press, New York,
pp 2146.
86. Xi, H., Lunney, E. A. (2010) The design,
annotation and application of a kinasetargeted-library, in (Zhou, J. Z. ed.) Chemical Library Design. Humana Press, New
York, Chapter 14.

Chapter 3
Molecular Library Design Using Multi-Objective
Optimization Methods
Christos A. Nicolaou and Christos C. Kannas
Abstract
Advancements in combinatorial chemistry and high-throughput screening technology have enabled the
synthesis and screening of large molecular libraries for the purposes of drug discovery. Contrary to initial
expectations, the increase in screening library size, typically combined with an emphasis on compound
structural diversity, did not result in a comparable increase in the number of promising hits found. In
an effort to improve the likelihood of discovering hits with greater optimization potential, more recent
approaches attempt to incorporate additional knowledge to the library design process to effectively guide
the search. Multi-objective optimization methods capable of taking into account several chemical and
biological criteria have been used to design collections of compounds satisfying simultaneously multiple pharmaceutically relevant objectives. In this chapter, we present our efforts to implement a multiobjective optimization method, MEGALib, custom-designed to the library design problem. The method
exploits existing knowledge, e.g. from previous biological screening experiments, to identify and profile
molecular fragments used subsequently to design compounds compromising the various objectives.
Key words: Multi-objective molecular library design, multi-objective evolutionary algorithm,
selective library design, MEGALib.

1. Introduction
Drug discovery can be seen as the quest to design small molecules
exhibiting favourable biological effects in vivo. Such molecules
need to balance a combination of multiple properties including
binding affinity to the pharmaceutical target, appropriate pharmacokinetics, limited (or no) toxicity (1, 2). The lack of consideration of the multitude of properties in the early stages of lead identification and optimization frequently hinders subsequent efforts
J.Z. Zhou (ed.), Chemical Library Design, Methods in Molecular Biology 685,
DOI 10.1007/978-1-60761-931-4_3, Springer Science+Business Media, LLC 2011

53

54

Nicolaou and Kannas

for drug discovery (3). Indeed, one of the common causes for
lead compounds to fail in the later stages of drug discovery is the
lack of consideration of multiple objectives at the early stage of
optimization of candidate compounds (4).
Traditional molecular library design (MLD) methods, modelled after the standard experimental drug discovery procedures,
ignored the multi-objective nature of drug discovery and focussed
on the design of libraries taking into account a single criterion.
Often, the focus has been on maximizing library diversity in an
effort to select compounds representative of the entire possible
population (5) or in designing compound collections exploring
a well-defined region of the chemical space defined by similarity to known ligands (6). The resulting molecular libraries, typically synthesized using combinatorial chemistry which enables the
synthesis of large numbers of compounds and screened via highthroughput screening systems, revealed that simply synthesizing
and screening large numbers of diverse (or similar) compounds
may not increase the probability of discovering promising hits
(7). Instead, due to the multi-objective nature of drug discovery,
other factors, such as absorption, distribution, metabolism, excretion, toxicity (ADMET), selectivity and cost, molecular screening libraries need to be carefully planned and a number of design
objectives must be taken into account (8). In recent times, MLD
efforts have been exploring the use of multi-objective optimization (MOOP) techniques capable of designing libraries based on
a number of properties simultaneously (9).
1.1. Multi-objective
Optimization Basics

Problems that require the accommodation of multiple objectives,

such as molecular library design, are widely known as multiobjective problems (MOP) or vector optimization problems
(10). In contrast to single-objective problems where optimization methods explore the feasible search space to find the single
best solution, in multi-objective settings, no best solution can be
found that outperforms all others in every criterion (3). Instead,
multiple best solutions exist representing the range of possible
compromises of the objectives (11). These solutions, known as
non-dominated, have no other solutions that are better than them
in all of the objectives considered. The set of non-dominated solutions is also known as the Pareto-front or the trade-off surface.
Figure 3.1 illustrates the concept of non-dominated solutions
and the Pareto-front in a bi-objective minimization problem.
MOPs are often characterized by vast, complex search spaces
with various local optima that are difficult to explore exhaustively,
largely due to the competition among the various objectives. In
order to decrease the complexity of the search landscape, MOPs
have traditionally been simplified, either by ignoring all objectives but one or by aggregating them. Multi-objective optimization (MOOP) methods enable the simultaneous optimization of

Molecular Library Design

55

Fig. 3.1. A MOP with two minimization objectives and a set of solutions represented
as circles. The rank of each solution (number next to circle) is based on the number of
solutions that dominate it (i.e. are better) in both objectives. The area defined by the
dashed lines of each solution contains the solutions that dominate it. Non-dominated
solutions are labelled 0. Point (0, 0) represents the ideal solution to this problem.

several objectives by considering numerous dependent properties

to guide the search. Pareto-based MOOP methods produce a set
of solutions representing various compromises among the objectives and allow the user to choose the solutions that are most
suitable for the task. The challenge facing these methods is to
ensure the convergence of well-dispersed solutions to guarantee
the effective coverage of the true optimal front (11). The major
benefit of MOOP methods is that local optima corresponding to
one objective can be avoided by consideration of all the objectives, thereby escaping single objective dead ends.
1.2. Evolutionary
Algorithms

Evolutionary algorithms (EAs) have been used extensively for

MOPs with several multi-objective optimization EAs (MOEA)
cited in the literature (12, 13). EA-based algorithms use populations of individuals evolved through a set of genetic operators such as reproduction, mutation, crossover and selection of
the fittest for further evolution (11). In the case of single objectives, selection of solutions involves ranking the individual solutions according to their fitness and choosing a subset. MOEAs
extend traditional EAs by adding a Pareto-ranking component
to enable the algorithm to handle multiple objectives simultaneously. MOEAs are particularly attractive since their populationbased approach enables the exploration of multiple search space
regions and thus the identification of numerous Pareto-solutions
in a single run. EAs impose no constraints on the morphology of
the search space, and thus, are suitable for complex, multi-modal
search spaces with various local optima such as the ones typically
found in MOPs (9). Figure 3.2 outlines the main steps of an
MOEA algorithm.

56

Nicolaou and Kannas

Generate initial population P

Evaluate solutions in P against objectives O1-n
Assign Pareto-rank to solutions
Assign efficiency value to solutions based on Pareto-rank
While Not Stop Condition:
Select parents Pparents in proportion to efficiency values
Generate population Poffspring by reproduction of Pparents
Mutation on individual parents
Crossover on pairs of parents
Evaluate solutions in Poffspring against objectives O1-n
Merge P, Poffspring to create Pnew
Assign Pareto-rank to solutions
Assign efficiency value to solutions based on Pareto-rank

Fig. 3.2. The MOEA algorithm.

1.3. Multi-objective
Molecular Library
Design Applications

Typical multi-objective molecular library design approaches use

the weighted-sum-of-objective-functions method that combines
the multiple objectives into a composite one via a weightedaverage transformation (14). Representative methods include
SELECT (15) which combines diversity and drug-likeness criteria to design libraries via an EA-based optimization method
and PICCOLO (16) which combines various objectives including reagent diversity, product novelty, similarity to known ligands
and pharmacokinetics into a single one and uses simulated annealing (11) to search for optimal solutions. Alternatively, the method
described by Bemis and Murcko enumerated a large virtual library
of compounds and applied a set of filters, including predictive
models for target-specific activity and drug-likeness thresholds on
chemical properties, to generate a compound library satisfying
multiple objectives (17).
In more recent years, Pareto-based methods have also been
used for molecular library design. MoSELECT employs the
multi-objective genetic algorithm (MOGA) (12) to simultaneously handle multiple objectives such as diversity, physicochemical
properties and ease of synthesis (7), and MoSELECT II incorporates library size (i.e. number of compounds) and configuration
(i.e. number of reagents at each position) as additional objectives
(18). A multi-objective incremental construction method, generating libraries based on a supplied scaffold and a set of reagents,
was proposed in reference (5). The method relies on the selection of appropriate reagents based on the similarity of the virtual
molecules they produce to the set of query molecules. The multiple similarities calculated for the virtual products are subjected to
Pareto-ranking that is subsequently used for reagent selection.
This chapter describes our work in developing an MOEA
algorithm specifically designed to address the problem of multiobjective library design given available knowledge, including
results from initial rounds of screening. The next sections describe
the algorithm in detail and present the software implemented.

Molecular Library Design

57

A sample application of the method focussing on designing a

selective library of compounds for secondary screening is also presented. The chapter concludes with a set of notes for a user to
avoid common mistakes and make better use of the method.

2. Materials
2.1. Multi-objective
Optimization
Software

1. NSisDesign: A molecular library design application program,

part of the NSisApps0.8 software suite (19), was developed
and used. The program is capable of generating a collection
of chemical designs of a given size produced by combining building blocks from a fragment collection supplied at
run time. The designs produced represent compromises of a
number of objectives also supplied at run time.
2. Molecular Fitness Assessment Software:
a. Fuzzee (20), a molecular similarity method based on a fuzzy,
property-based molecular representation.
b. OEChem (21), a chemoinformatics toolkit used to calculate chemical structure properties such as molecular weight,
hydrogen bond donors and acceptors.

2.2. Molecular
Building Block
Preparation Software

1. NSisFragment, a molecular fragmentation and substructure

mining tool part of the NSisUtilities0.9 software suite (19)
was used. The tool is able to extract fragments from molecular graphs in a variety of ways including frequent subgraph
mining (22) and the RECAP chemical bond type identification and cleaving technique (23). The fragments contain
information about their attachment points and the type of
bond cleaved at each attachment point.
2. NSisProfile, a chemical fragment characterization and profiling tool from the NSisUtilities0.9 software suite. The
tool characterizes supplied molecular fragments with respect
to chemical structure characteristics, e.g. molecular weight,
hydrogen bond donors and acceptors, complexity (24) and
number of rotatable bonds. When supplied with molecular libraries annotated with biological screening information,
the tool matches fragments and molecules, prepares lists
of molecules containing each fragment and annotates fragments with properties related to the molecules containing
them, for example, average IC50 values for a specific assay.

2.3. Datasets

1. Dataset 1, a set of well-known estrogen receptor (ER) ligands, contains five compounds, three with increased selectivity to ER- and two with increased selectivity to ER-.

58

Nicolaou and Kannas

LIGAND

RBA RBA Selectivity

ER- ER-

0.17

13

76

32.2

6.4

0.2

Fig. 3.3. Ligands and their relative binding affinity (RBA) to estrogen receptors
and (25).

Figure 3.3 shows two of the molecules used, representative

of the two sets used.
2. Dataset 2 is an ER-inhibitor dataset obtained from PubChem (26). The dataset consists of 86,098 compounds
tested on both ER- (Bioassay 629: HTS of Estrogen Receptor-alpha Coactivator Binding inhibitors, Primary Screening) and ER- (Bioassay 633: HTS for Estrogen Receptor-beta Coactivator Binding inhibitors, Primary
Screening).

3. Methods
Recently, we proposed the Multi-objective Evolutionary Graph
Algorithm (MEGA), an optimization algorithm designed for the
evolution of chemical structures satisfying multiple constraints
(9). The technique combines evolutionary techniques with graph
data structures to directly manipulate graphs and perform a global
search for promising molecule designs. MEGA supports the use
of problem-specific knowledge and local search techniques with
an aim to improve both performance and scalability. Initial applications of the algorithm to the problem of de novo design showed
that the technique is able to produce a diverse collection of equivalent solutions and, thus, support the drug discovery process (9).
Based on our experiences we have designed a custom version of

Molecular Library Design

59

the original algorithm, termed MEGALib, to meet the requirements of multi-objective library design. The method focusses on
designing the best possible products, i.e. chemical structures, for
the problem under investigation and makes no attempt to minimize the number of reagents used; its main applications to date
have been in designing small, focussed molecular libraries for secondary screening. This section initially describes MEGALib followed by a detailed overview of the methodology used to prepare the fragment collection and the computational objectives
required by the algorithm. The later part of the section thoroughly describes an application of MEGALib to the problem of
designing a selective library of compounds.
3.1. Multi-objective
Library Design
Algorithm
Description

1. MEGALib input. MEGALib requires the supply of a set

of molecular building blocks, the implemented objectives
to be used for scoring molecules, a set of attributes controlling evolutionary operations, including mutation and
crossover methods and probabilities, and hard filters for
solution elimination. User input indicating the size of the
designed library is also supplied. MEGALib operates on
two population sets, the normal, working population and
the secondary population or the Pareto-archive. The size
of the two populations is also supplied by the user.
2. Initial working population generation. The first phase of
the algorithm generates the initial population by combining pairs of building blocks from the collection supplied
by the user and initiates the external archive of solutions
intended to store the secondary population. The virtual
synthesis step operates by taking into account the weight
associated with each building block, if one is provided.
Specifically, a roulette-like method selects building blocks
via a probabilistic mechanism that assigns higher selection
probability to those having a higher weight (11). To synthesize a member of the initial population the algorithm
selects a core building block and attaches to each of its
attachment points a building block with matching attachment point bond type. The algorithm repeats the above
process until the number of initial population members
reaches a multiple of the user-defined working population
size, by default five times more, in order to avoid problems with insufficient working population size resulting
from the elimination of solutions by the application of filtering in step 4 below. It is worth noting that the algorithm uses graph-based chromosomes corresponding to
chemical structures to avoid the information loss associated
with the encoding of more complex structures into simpler
ones (9).

60

Nicolaou and Kannas

3. Solution fitness assessment. The population is then

subjected to fitness assessment through application of the
available objectives, a process that results in the generation
of a list of scores for each individual.
4. Hard filter elimination. The list of solution scores is used
for the elimination of solutions with values outside the
range allowed by the corresponding active hard filters
defined by the user.
5. Working population update. This step combines the two
populations, working and secondary, to update the working population pool. This step is eliminated from the first
iteration of the algorithm since the secondary, archive population is empty.
6. Pareto-ranking. The individuals list of scores is subjected
to a Pareto-ranking procedure to set the rank of each individual. According to this procedure the rank of an individual is set to the number of individuals that dominate
it incremented by 1, thus, non-dominated individuals are
assigned rank order 1.
7. Efficiency score calculation. The algorithm then proceeds
to calculate an efficiency score for each individual using a
methodology that operates both in parameter and objective space. The methodology employs an elaborate niching
mechanism that performs diversity analysis of the population based on the genotype, i.e. the chromosome graph
structure, and subsequently assigns an efficiency score that
takes into account both the Pareto-rank and the diversity
analysis outcome (9). The current implementation of the
diversity analysis uses the Wards agglomerative clustering
technique (27) and atom-type descriptors (28). The resulting Wards cluster tree is processed with the Kelley cluster
level selection method (27) to produce a set of natural clusters. The results from clustering are subsequently used in
the preparation of the efficiency score of individuals, which
consists of its Pareto-rank and the cluster assignment.
8. Secondary population update. Efficiency scores are initially
used to update the Pareto-archive. The current Paretoarchive is erased and a subset of the current working population that favours individuals with high efficiency score, i.e.
low domination rank and high chromosome graph diversity, takes its place. Note that the size of the secondary
population selected is limited by a user-supplied parameter.
The secondary population mechanism has been designed
specifically to preserve good solutions, non-dominated or
dominated but substantially structurally unique, from all

Molecular Library Design

61

iterations from getting lost due to working population size

limitations.
9. Parent selection. Following the update of the Paretoarchive, MEGALib checks for the termination conditions
and terminates if they have been satisfied. If this is not the
case the process moves to select the parent subset population from the combined population set using a variation
of the roulette method (11) operating on the dual-valued
efficiency scores of the candidate solutions. Specifically, the
selection method is applied on the clusters rather than the
entire population. The process picks one solution from
each cluster starting from the largest cluster and proceeding
to clusters containing fewer compounds (9). The process
traverses the set of clusters until the number of parents is
selected. The parent selection method can be fine-tuned via
user-supplied parameters to favour the parameter space or
the objective space. Favouring the objective space amounts
to selecting non-dominated solutions from each cluster.
This method only proceeds to select dominated solutions
when all non-dominated have been selected. Favouring the
parameter space focusses on selecting solutions from all
clusters by applying the roulette-like, weighted selection
method to each cluster.
10. Offspring generation. The parents are then subjected to
mutation and crossover according to the probabilities indicated by the user. MEGALib evolves solutions through a
set of fragment-based operations inspired by mutation and
crossover techniques. Mutation processes include insertion, removal and exchange of fragments. For fragment
insertion, an attachment point is first chosen and a fragment from the weighted fragment collection is chosen and
attached. For the fragment removal and exchange operations RECAP (23) is used to break the molecule into two
disconnected parts and either remove or replace one of
them with a fragment from the fragment collection. Note
that fragment weights influence the probability of selection
of a fragment for the insertion and exchange operations.
Also note that the exchange fragment operation involves
building blocks with attachment points of compatible bond
types. Crossover takes place by identifying and cleaving a
RECAP-type bond in each of two parents and recombining the resulting fragments to generate offspring. In a manner similar to the exchange fragment operation described
above, this type of crossover is restricted to breaking specific bond types and combining fragments with compatible bond types in order to produce reasonable chemical
designs.

62

Nicolaou and Kannas

11. New working population generation. The new working

population is formed by merging the original working population and the newly produced mutants and crossover
children. The process then iterates as shown in Fig. 3.4.

Fig. 3.4. The MEGALib algorithm.

Upon termination of the process the algorithm selects a compound set from the working population equal to the user-supplied
library size as the library proposed. The selection of the library
members is performed in a manner identical to the parent selection method described previously.
The algorithm exploits existing knowledge through the inclusion of multiple, problem-specific objectives, the use of bondtype information when evolving molecules and the exploitation of
the weights associated with the building blocks provided which
result in favouring those with an increased weight, i.e. having
privileged status.
3.2. Fragment
Collection
Preparation

The building block collection required by MEGALib consists

of information-rich reagents, e.g. chemical fragments annotated
with information on attachment points and bond types, as well
as weights that designate their privileged, or not, status. The
building blocks may be prepared via the application of the NSisFragment and NSisProfile tools, described previously, on a set
of compounds with biological property information. The building blocks may also be obtained using other means by following the detailed advanced programmer interface (API) provided by the toolkit. For example, commercially available reagents
may be appropriately annotated with information about attachment points, reaction types and privileged status by expert
chemists.

3.3. Computational
Objective Encoding

Fitness scores required for the application of MEGALib rely on

the encoding of computational objective scorers that measure, or
predict, molecular attributes.The main use of such scorers is to
guide the optimization process, i.e. to direct the search towards
interesting regions of the chemical space. Additionally, objective
scorers may be used as hard filters to remove solutions with fitness values outside a predefined allowed range provided by the

Molecular Library Design

63

user. Objectives used in this manner are typically referred to as

secondary while objectives used to guide optimization are considered primary.
MEGALib can use a wide range of molecular scorers provided
that they have been encoded inline with a well-defined available
API that allows smooth integration with the algorithm. The set
of scorers available by the current implementation includes the
following:
(a) Binding affinity scorers: MEGALib provides an interface
that facilitates the encoding of objectives based on the predicted binding affinity of a designed molecule to a target
protein. The implementation uses the docking program
Glamdock and the ChillScore scoring method recently
developed by Tietze and Apostolakis (29) to dock the
designed molecules into the binding site of a receptor site
provided by the user interactively, in real time. The interaction score of the best solution is used as an objective function. Settings for docking correspond to the slow settings
described in Tietze and Apostolakis (29).
(b) Molecular similarity scorers: MEGALib encodes molecular
similarity to a collection of user-supplied molecules as a distinct objective. The method uses the Fuzzee tool from the
Chil2 molecular modelling platform (20) which operates
on abstractions of molecular graphs that replace atoms with
molecular features to produce the so-called feature graphs.
The actual similarity is calculated in a pair-wise manner by
first aligning the feature graphs of two molecules, identifying common features, and then applying the Tanimoto
similarity measure (30). In the event of similarity to a set of
compounds, the average value of the pair-wise similarities is
used.
(c) Chemical structure scorers: A list of chemical structure
objectives, including molecular weight, number of hydrogen bond donors and acceptors, rotatable bonds and complexity is also available in the current implementation of
MEGALib. Typically, chemical structure scorers are used
as secondary objectives to constrain the search space by
filtering out solutions such as those not conforming to
the Rule-of-Five (31) or those estimated to be highly
complex (24).
3.4. Selective Library
Design Case Study

Designing selective libraries implies taking into consideration

more objectives than just collecting compounds from various
structural classes (32). The sample case study described in this section involves the application of MEGALib to design a library of
compounds potentially exhibiting selectivity to one of two related
but distinct pharmaceutical targets, namely ER- over ER-. The

64

Nicolaou and Kannas

example given is meant to highlight the steps to be followed to

produce a library satisfying multiple criteria.
A single collection of 51,123 building blocks was used for all
the tests performed. The building blocks were obtained via fragmentation of Dataset 2 described previously with the fragmentation tool NSisFragment. The resulting fragments were profiled
using the NSisProfile tool against the properties of the molecules
that contain them as found in the Pubchem Assays 629 and 633
and weights corresponding to the values of the properties have
been recorded. For the purposes of this application, a propertyspecific weight of a fragment is the average value of the property
for the molecules that contain the fragment. Note that known ligands were not included in the fragmentation and building block
generation process to favour the design of structurally different
chemical designs.
The experimental settings used population size 100 and
1,000 generations. Runs were performed using both mutation
and crossover. Mutation probability was set at 0.25 and crossover
at 1.0. The maximum Pareto-archive set was set to 1,000. The
desired library size was set at 250. Parent selection was set to balance between the diversity in parameter and objective space.
Two ligand-based objectives that measured the average similarity of a query molecule to known ligands were used. Similarity was calculated using the tool Fuzzee (20). The two objectives measured shape and property-based similarity of a given
query molecule to the set of ER--selective and ER--selective
ligands in Dataset 1. The experiments aimed at designing a library
of molecules exploring the selectivity potential between the two
ERs, with an emphasis on designs more similar to compounds
selective to ER-, and so the algorithm was set so as to maximize
average similarity to the ER- ligand set and minimize the average similarity to the ER- ligand set. The search was constrained
by imposing limits on the acceptable similarity values of the new
designs to the two objectives. Specifically, minimum acceptable
similarity to ER- was set to 0.5 and maximum acceptable similarity to ER- was also set to 0.5. Additionally, a set of hard filters based on chemical structure objectives was applied in order
to remove potentially problematic designs from further consideration in line with step 4 of the MEGALib algorithm. The set
of hard filters included limitations in the number of hydrogen
bond donors and acceptors, and molecular weight, in line with
the Rule-of-Five (31).
Progress monitoring of the MEGALib execution was performed by calculating the quality of the Pareto-approximation
using quantitative measures in a post-processing step taking place
after each generation. Specifically, the performance measures
encoded included the calculation of the Pareto-approximation
set hypervolume (13), the spacing measure (11) and the

Molecular Library Design

65

chromosome/structural diversity. The latter was calculated by

averaging the Euclidean distances of each solution to all other
solutions in the proposed set, using atom-pair descriptors (28) of
the molecules involved. All three measures were calculated using
code developed in-house for this reason. To avoid the extraction
of misleading conclusions obtained through chance results a total
of five runs were performed with identical input parameter settings but different initial population sets resulting from alternative initial population generation settings. The assessment of the
results obtained from the five runs indicated similar performance
with respect to the hypervolume, spacing and chromosome diversity and no major deviations. The results presented in the figures
below correspond to one of the five runs and are representative
of the set of results produced. Time requirements for the execution of the runs were sufficiently reasonable. A typical run of
MEGALib executed, with population 250 and 1,000 iterations,
took approximately 6 h on a normal PC.
The resulting library consisted of 250 compounds representing different compromises between the two conflicting objectives
supplied. Figure 3.5 presents a plot of the Pareto-approximation
proposed by the software library (circles connected by line). Each
of the remaining circles represents a solution from the initial population set after the hard filtering process. The x-axis represents
similarity to ER- ligands and the y-axis dissimilarity (1-similarity)

Fig. 3.5. Pareto-approximation formed by the designed library. The non-connected

circles represent the initial population set. The x-axis represents shape similarity to
ER- ligands and the y-axis shape dissimilarity to ER- ligands. Both objectives were
minimized.

66

Nicolaou and Kannas

Fig. 3.6. Scaffolds representative of the compounds in the library designed using MEGALib.

to ER- ligands; thus, the problem has been transformed to a biobjective minimization problem with the ideal solution at point
(0, 0).
Figure 3.6 presents a small subset from the collection of the
scaffolds found in the compounds of the designed library. Each
scaffold gave rise to one or more compounds of the designed
library with varying performance to the objectives of the experiment through different substitutions on the various attachment
points indicated as R groups. Consequently, the resulting library
was sufficiently diverse indicating that MEGALib has been successful in identifying and preserving the structural diversity of the
designed compounds.

4. Notes
1. Designing focussed vs diverse libraries. The scope and diversity of the library designed by MEGALib can be controlled
using the user-supplied parameters required by the algorithm primarily by the choice of objectives and building
block pool. Diverse libraries may be designed by formulating
population diversity as one of the objectives of MEGALib.
To this end the Wards clustering method combined with the
Kelley cluster level selection described in Section 3 may be
used. Additional objectives ensure that the set of diverse
molecules produced will meet, for example, drug-likeness
criteria. Focussed libraries are meant for a specific target
(or related targets) and therefore objectives encoding targetspecific information must be used (17). The use of a carefully

Molecular Library Design

67

selected building block set consisting of fragments privileged

for the specific target as well as objectives based on similarity
to one or more known ligands can guide the search to generate a custom library for the target. The sample application
presented in this chapter belongs to the latter library design
category.
2. Types of objectives. MEGALib is agnostic to the type of
objectives used. It is sufficient to prepare a computational
method implementing a specific objective with an interface
strictly in line with the NSisDesign API to enable its use by
MEGALib during execution time. While this provides great
flexibility to the user it is worth noting that special consideration must be given when preparing objectives to ensure
their quality and reliability to facilitate the search. Typically,
objectives based on noisy data or models of questionable
quality may impede the algorithmic search and should only
be used to provide general guidance to the search or as loose
hard filters. Similarly, the use of highly correlated objectives
should be avoided since their presence is not beneficial and
may instead result in degraded computational performance.
3. Hard filtering. The use of multiple and/or strict sets of hard
filters may cause problems especially in the initial iterations
of the execution of MEGALib since they may reduce the
population below the size required for subsequent operations and/or decrease greatly the working population diversity. The current algorithm implementation checks whether
the solutions passing through the hard filters satisfy the population size indicated by the user. In the event that this is not
the case eliminated solutions are selected and added to the
working population. The solution recovery step sorts the
eliminated solutions according to the number of filters they
failed and selects a large enough subset to add to the working
population in a quasi-random fashion favouring each time
the least problematic individuals.
4. Performance issues. The performance of MEGALib is largely
dependent on the computational cost of the objectives used
for the fitness assessment of the population. Certain objectives, such as those based on docking, require substantial
execution time while others, such as those based on chemical
structure or comparisons to known ligands, are less costly.
5. Pareto-archive size. MEGALib, as well as other MOEAs,
has the ability to generate a large number of equivalent
solutions for a given MOP. Consequently, the size of the
Pareto-archive may increase to several thousands or even
more depending on the number of iterations, the size of
the working population, the number of building blocks, etc.
An overly large archive, even though theoretically able to
hold all promising solutions from all iterations, in practice

68

Nicolaou and Kannas

imposes a significant performance cost during execution

time mostly due to the clustering step invoked by the niching
mechanism. Extensive experimentation has shown that limiting the size of the archive using a user-supplied parameter
available in the current implementation and a cluster-based
elimination of solutions is able to maintain population diversity and reduce the computational cost to reasonable times.
6. Niching mechanism. Care must be exercised when sampling from clusters to accommodate the likely presence of
singleton and under-represented clusters often found when
the population size is small or particularly diverse. Such
clusters may cause problems during selection, for example, when attempting to sample from singleton clusters. To
avoid this type of problem MEGA implements appropriate
rules, such as allowing only simple selection from singleton
clusters (9).
7. Repair mechanism. Following the virtual synthesis step that
takes place during parent solution evolution a repair mechanism is applied to ensure that the resulting offspring are
valid molecules with respect to valences. Briefly, in its current
implementation the mechanism identifies atoms with valence
problems and attempts to repair them either by removing
hydrogens attached to the atom or by downgrading atom
bonds to a lower order, i.e. converts a double bond to single or a triple to double. If such action is not possible or
sufficient to fix the problem, the offspring is discarded (9).
8. Parent selection method. Typical settings of the MEGALib
algorithm use the parent selection method favouring the
parameter space, i.e. selecting solutions from clusters using
the roulette-like method described in Section 3. This setting
has been experimentally proven to preserve graph chromosome diversity and ensure that a variety of different promising subgraphs (scaffolds/chemotypes) survive long enough
in the evolution cycle to contribute to the solution search.

References
1. Ekins, S., Boulanger, B., Swaan, P. W.,
Hupcey, M. A. (2002) Towards a new age of
virtual ADME/TOX and multidimensional
drug discovery. J Comput Aided Mol Des 16,
381401.
2. Agrafiotis, D. K., Lobanov, V. S., Salemme, F.
R. (2002) Combinatorial informatics in the
post-genomics era. Nat Rev Drug Discov 1,
337346.
3. Baringhaus, K. H., Matter, H. (2004)
Efficient strategies for lead optimization by

simultaneously addressing affinity, selectivity

and pharmacokinetic parameters, in (Oprea,
T., ed.) Chemoinformatics in Drug Discovery. Wiley-VCH, Weinheim, Germany, pp.
333379.
4. Nicolaou, C. A., Brown, N., Pattichis, C. S.
(2007) Molecular optimization using computational multi-objective methods. Curr
Opin Drug Discov Dev 10, 316324.
5. Soltanshahi, F., Mansley, T. E., Choi, S.,
Clark, R. D. (2006) Balancing focused

Molecular Library Design

6.

7.

8.
9.

10.

11.
12.

13.

14.

15.

16.

17.
18.

combinatorial libraries based on multiple

GPCR ligands. J Comput Aided Mol Des 20,
529538.
Gillet, V. J., Willet, P., Fleming, P. J., Green,
D. V. (2002) Designing focused libraries
using MoSELECT. J Mol Graph Model 20,
491498.
Gillet, V. J., Khatib, W., Willett, P., Fleming, P. J., Green, D. V. (2002) Combinatorial
library design using a multiobjective genetic
algorithm. J Chem Inf Comput Sci 42,
375385.
Agrafiotis, D. K. (2000) Multiobjective optimization of combinatorial libraries. Mol
Divers 5, 209230.
Nicolaou, C. A., Apostolakis, J., Pattichis, C.
S. (2009) De novo drug design using multiobjective evolutionary graphs. J Chem Inf
Model 49, 295307.
Coello Coello, C. A. (2002) Evolutionary
multiobjective optimization: a critical review,
in (Sarker, R., Mohammadian, M., Yao, X.
eds.) Evolutionary Optimization. New York:
Springer 48, pp. 117146.
Yann, C., Siarry, P. (eds.) (2004) Multiobjective Optimization: Principles and Case Studies, Springer, Berlin, Germany.
Fonseca, C. M., Fleming, P. J. (1998) Multiobjective optimization and multiple constraint handling with evolutionary algorithms. I: a unified formulation. IEEE Trans
Syst Man Cybernet 28, 2637.
Zitzler, E., Thiele, L. (1999) Multiobjective evolutionary algorithms: a comparative
case study and the strength Pareto approach.
IEEE Trans Evol Comput 3, 257271.
Gillet, V. J. (2004) Designing combinatorial libraries optimized on multiple objectives in methods in molecular biology, in
(Bajorath, J., ed.) Chemoinformatics: Concepts, Methods, and Tools for Drug Discovery. Humana Press, Totowa, NJ, 275, pp.
335354.
Gillet, V. J., Willett, P., Bradshaw, J., Green,
D. V. S. (1999) Selecting combinatorial
libraries to optimize diversity and physical properties. J Chem Inf Comput Sci 39,
169177.
Zheng, W., Hung, S. T., Saunders, J. T.,
Seibel, G. L. (2000) PICCOLO: a tool for
combinatorial library design via multicriterion optimization. Pac Symp Biocomput 5,
585596.
Bemis, A. G. W., Murcko, M. A. (1999)
Designing libraries with CNS activity. J Med
Chem 42, 49424951.
Wright, T., Gillet, V. J., Green, D. V., Pickett,
S. D. (2003) Optimizing the size and configuration of combinatorial libraries. J Chem Inf
Comput Sci 43, 381390.

69

19. Noesis Chemoinformatics, Ltd. http://www.

noesisinformatics.com (accessed August 12,
2009).
20. MoDest. http://www.chil2.de (accessed
June 30, 2009).
21. OpenEye, Inc. http://www.eyesopen.com
(accessed July 3, 2009).
22. Nicolaou, C. A., Pattichis, C. S. (2006)
Molecular substructure mining approaches
for computer-aided drug discovery: a review.
Proceedings of the 2006 ITAB Conference,
October 2628, Ioannina, Greece.
23. Lewell, X. O., Budd, D. B., Watson, S. P.,
Hann, M. M. (1998) RECAP Retrosynthetic Combinatorial Analysis Procedure: a
powerful new technique for identifying privileged molecular fragments with useful applications in combinatorial chemistry. J Chem
Inf Comput Sci 38, 511522.
24. Barone R., Chanon, M. (2001) A new
and simple approach to chemical complexity.
Application to the synthesis of natural products. J Chem Inf Comput Sci 41, 269272.
25. Angelis, M. D., Stossi F., Waibel M.,
Katzenellenbogen, B. S., Katzenellenbogen,
J. A. (2005) Isocoumarins as estrogen
receptor beta selective ligands: isomers of
isoflavone phytoestrogens and their metabolites. Bioorg Med Chem 13, 65296542.
26. Wheeler, D. L., et al. (2006) Database
resources of the National Center for Biotechnology Information. Nucleic Acids Res 34,
173180.
27. Wild, D. J., Blankley, C. J. (2000) Comparison of 2D fingerprint types and hierarchy
level selection methods for structural grouping using wards clustering. J Chem Inf Comput Sci 40, 155162.
28. Kearsley, S. K., Sallamack, S., Fluder, E. M.,
Andose, J. D., Mosley, R. T., Sheridan, R.
P. (1996) Chemical similarity using physiochemical property descriptors. J Chem Inf
Comput Sci 36, 118127.
29. Tietze, S., Apostolakis, J. (2007) GlamDock:
development and validation of a new docking
tool on several thousand protein-ligand complexes. J Chem Inf Model 47, 16571672.
30. Willet, P., Barnard, J. M., Downs, G. M.
(1998) Chemical similarity searching. J Chem
Inf Comput Sci 39, 983996.
31. Lipinski, C. A., Lombardo, F., Dominy, B.
W., Feeney, P. J. (1997) Experimental and
computational approaches to estimate solubility and permeability. Drug discovery and
development settings. Adv Drug Discovery
Rev 23, 325.
32. Prien, O. (2005) Target-family-oriented
focused libraries for kinases conceptual
design aspects and commercial availability.
ChemBioChem 6, 500505.

Chapter 4
A Scalable Approach to Combinatorial Library Design
Puneet Sharma, Srinivasa Salapaka, and Carolyn Beck
Abstract
In this chapter, we describe an algorithm for the design of lead-generation libraries required in
combinatorial drug discovery. This algorithm addresses simultaneously the two key criteria of diversity and representativeness of compounds in the resulting library and is computationally efficient when
applied to a large class of lead-generation design problems. At the same time, additional constraints on
experimental resources are also incorporated in the framework presented in this chapter. A computationally efficient scalable algorithm is developed, where the ability of the deterministic annealing algorithm to
identify clusters is exploited to truncate computations over the entire dataset to computations over individual clusters. An analysis of this algorithm quantifies the trade-off between the error due to truncation
and computational effort. Results applied on test datasets corroborate the analysis and show improvement
by factors as large as ten or more depending on the datasets.
Key words: Library design, combinatorial optimization, deterministic annealing.

1. Introduction
In recent years, combinatorial chemistry techniques have provided important tools for the discovery of new pharmaceutical
agents. Lead-generation library design, the process of screening
and then selecting a subset of potential drug candidates from a
vast library of similar or distinct compounds, forms a fundamental step in combinatorial drug discovery (1). Recent advances in
high-throughput screening such as using micro/nanoarrays have
given further impetus to large-scale investigation of compounds.
However, combinatorial libraries often consist of extremely large
collections of chemical compounds, typically several million. The
time and cost of associated experiments makes it practically
J.Z. Zhou (ed.), Chemical Library Design, Methods in Molecular Biology 685,
DOI 10.1007/978-1-60761-931-4_4, Springer Science+Business Media, LLC 2011

71

72

Sharma, Salapaka, and Beck

impossible to synthesize each and every combination from such a

library of compounds. To overcome this problem, chemists often
work with virtual combinatorial libraries (VCLs), which are combinatorial databases containing enumeration of all possible structures of a given pharmacophore with all available reactants. A subset of lead compounds from this VCL is selected which is used
for physical synthesis and biological target testing. The selection
of this subset is based on a complex interplay between various
objectives, which is cast as a combinatorial optimization problem. The main goal of this optimization problem is to identify
a subset of compounds that is representative of the underlying
vast library as well as manageable, where these lead compounds
can be synthesized and subsequently tested for relevant properties, such as activity and bioaffinity. The combinatorial nature of
the selection problem makes it impractical to exhaustively enumerate each and every possible subset of obtaining the optimal
solution. For example, to select 30 lead compounds from a set of
1,000, there are approximately 3 1025 different possible combinations. Selection based on enumeration is thus impractical and
requires numerically efficient algorithms to solve the constrained
combinatorial optimization problem.

2. Issues
in LeadGeneration
Library Design

In addition to the computational complexity that arises due to the

combinatorial nature of the problem, any algorithm that aims to
address the lead-generation library design problem must address
the following key issues:
Diversity versus representativeness: The most widely used
method to obtain a lead-generation library involves maximizing
the diversity of the overall selection (2, 3), based on the premise
that the more diverse the set of compounds, the better the chance
to obtain a lead compound with desired characteristics. Such a
design strategy suffers from an inherent problem that using diversity as the sole criterion may result in a set where a large number
of lead compounds disproportionately represent outliers or singletons (4, 5). However, from a drug discovery point of view, it
is desirable for the lead-generation library to more proportionally
represent all the compounds, or at least to quantify how representative each lead compound is in order to allot experimental
resources. A maximally diverse subset is of little practical significance because of its limited pharmaceutical applications. Therefore, representativeness should be considered as a lead-generation
library design criterion along with diversity (6, 7).

A Scalable Approach to Combinatorial Library Design

73

Design constraints: In addition to diversity and representativeness, other design criteria include confinement, which quantifies the degree to which the properties of a set of compounds
lie in a prescribed range (8), and maximizing the activity of
the set of compounds against some predefined targets. Activity is usually measured in terms of the quantitative structure
of the given set. Additionally, the cost of chemical compounds
and experimental resources is significant and presents one of the
main impediments in combinatorial diagnostics and drug synthesis. Different compounds require different experimental supplies
which are typically available in limited quantities. The presence
of these multiple (and often conflicting) design objectives makes
the library design a multiobjective optimization problem with
constraints.

3. Basic Problem
Formulation and
Modifications

Basic formulation: The problem of selecting lead compounds for

lead-generation library design can be stated in general as follows:
Given a distribution of N compounds, xi , in a descriptor space
, find the set of M lead compounds, rj, that solves the following
minimization problem:

min

N


rj ,1jM


p(xi )

i=1


min d(xi , rj )

1jM

[1]

Here,  represents the chemical property space corresponding to the VCL, d(xi , rj ) represents an appropriate distance metric
between the lead compound rj and the compound xi , p(xi ) is the
relative weight that can be attached to compound xi (if all compounds are of equal importance, then the weights p(xi ) = N1 for
each i), and M is typically much smaller than N. That is, this problem seeks a subset of M lead compounds rj in a descriptor space
such that the average distance of a compound xi from its nearest
lead compound is minimized. Alternatively, this problem can also
be formulated as finding
 an optimal partition of the descriptor
space into M clusters Rj and assigning to each cluster a lead
compound rj such that the following cost function is minimized:
M 


d(xi , rj )p(xi )

j=1 xi Rj

Incorporating diversity and representativeness: One drawback

of the basic formulation is that all the lead compounds are

74

Sharma, Salapaka, and Beck

weighted equally. However, design constraints often require distinguishing them from one another to reflect different aspects of
the clusters. For example, when addressing the issue of representativeness in the lead-generation library, the lead compounds that
represent larger clusters need to be distinguished from those that
represent outliers.
We incorporate representativeness into the problem formulation by specifying an additional relative weight parameter j , 1
j M for each lead compound. This parameter j quantifies the
size of the cluster represented by the compound rj , and it is
proportional to the number of the compounds in that cluster.
Thus, the resulting library design will associate lead compounds
that represent outliers with low values of and the lead compounds that represent the majority members with corresponding
high values. In this way, the algorithm can be used to identify
distinct compounds through property vectors rj in the descriptor space  that denote the jth lead compound and at the same
time determine how representative each lead compound is. For
instance, j = 0.2 implies that lead compound rj represents 20%
of all compounds in the VCL. The following modified optimization problem adequately describes the diversity goals in the basic
formulation as well as the representativeness through the relative
weights j :


min

rj ,j ,1jM

such that

M



p(xi )


min d(xi , rj )

1jM

[2]

j = 1

j=1

where j is the fraction of compounds in VCL that are nearest to

(represented by) the lead compound rj .
Incorporating constraints on experimental resources: Experiments associated with compounds with different properties often
require different experimental resources. The constraints on availability of these resources can vary depending on their respective
handling costs and time. These constraints can be incorporated
in the selection problem by associating appropriate weights to
lead compounds. For instance, consider a VCL that is classified
into q types of compounds corresponding to q types of experimental supplies required for testing. More specifically, the jth lead
compound can avail only Wjn amount of the nth experimental
resource (1 n q). The modified optimization problem is then
given by (9, 10)
min D =
rj


n



pn (xin ) min d(xin , rj )
j

[3]

A Scalable Approach to Combinatorial Library Design

75

such that jn = Wjn 1 j M , 1 n q

where pn (xin ) is the weight of the compound location xin , which
requires the nth type of supply.

4. Computational
Issues
Problem formulations [13] for designing lead-generation library
under different constraints belong to a class of combinatorial
resource allocation problems, which have been widely studied.
They arise in many different applications such as minimum distortion problems in data compression (11), facility location problems (12), optimal quadrature rules and discretization of partial
differential equations (13), locational optimization problems in
control theory (9), pattern recognition (14), and neural networks
(15). Combinatorial resource allocation problems are nonconvex
and computationally complex and it is well documented (16) that
most of them have many local minima that riddle the cost surface.
Therefore, the main computational issue is developing an efficient algorithm that avoids local minima. Due to the large size of
VCLs, and the combinatorial nature of the problem, the issue of
algorithm scalability takes central importance. Since the number
of computations to be performed by the lead-generation library
design algorithm scales up exponentially with an increase in the
amount of data, most algorithms become prohibitively slow and
expensive (computationally) for large datasets.
4.1. Deterministic
Annealing Algorithm

The main drawback of most popular algorithms that address

the basic combinatorial resource allocation problem [1], such as
Lloyds or K-means algorithms (11, 17), is that they are extremely
sensitive to initialization step in their procedures and typically
get trapped in local minima. Other algorithms such as simulated
annealing that actively try to avoid local minima are often computationally inefficient. Other drawbacks of these algorithms mainly
stem from the lack of flexibility to incorporate various constraints
on the resource locations discussed in Section 3. The deterministic annealing (DA) algorithm (18) overcomes these drawbacks; this algorithm is heuristically based on law of minimum
free energy in statistical chemistry that models similar combinatorial problems occurring in nature. The DA algorithm is versatile in
terms of accommodating constraints on resource locations while
simultaneously it is designed to be insensitive to the initialization
step and to avoid local minima.
The central concept of the DA algorithm is based on developing a homotopy from an appropriate convex function to the nonconvex cost function; the local minima of cost function at every

76

Sharma, Salapaka, and Beck

step of homotopy serves as the initialization for the subsequent

step. Since minimization of the initial convex function yields a
global minimum, this procedure is independent of initialization.
The heuristic is that the global minimum is tracked as the initial
convex function deforms into the actual nonconvex cost function
via the homotopy. Accordingly, the DA algorithm solves the following multiobjective optimization problem:
min min D Tk H

 
rj p(rj |xi )
:=F

over iterations indexed by k, where Tk is a parameter called temperature which tends to zero as k tends to infinity. The cost function F is called free energy, where this terminology is motivated
by statistical chemistry (18). Here the distortion
D=

N


p(xi )

i=1

M


d(xi , rj )p(rj |xi )

j=1

which is similar to the cost function in equation [1] is the

weighted average distance of a lead compound rj from a compound xi in the VCL. This formulation associates each xi to every
rj through the weighting parameter p(rj |xi ) and thus diminishes
the sensitivity of algorithm to initialization of locations rj . The
more uniformly (or randomly) these weights are distributed, the
more insensitive is the 
algorithm with respect to the initialization. The term H = i,j p(yj |xi ) logp(yj |xi ) is the entropy of


the weights p(yj |xi ) that quantifies their uniformity (or randomness). The annealing parameter Tk defines the homotopy from the
convex function H to the nonconvex function D. Clearly, for
large values of Tk , we mainly attempt to maximize the entropy. As
Tk is lowered, we trade entropy for the reduction in distortion,
and as Tk approaches zero, we minimize D directly to obtain a
hard (nonrandom) solution, where p(yj |xi ) is either 0 or 1 for
each pair (i,j). Minimizing the free energy term F with respect to
the weighting parameter p(rj |xi ) is straightforward and gives the
Gibbs distribution
p(rj |xi ) =


e d(xi ,rj )/Tk
, where Zi :=
e d(xi ,rj )/Tk
Zi

[4]

Note that the weighting parameters p(rj |xi ) are simply radial
basis functions, which clearly decrease in value exponentially as rj
and xi move farther apart. The corresponding minimum of F is
obtained by substituting for p(rj |xi ) from equation [4]:

F

= Tk


i

p(xi ) log Zi

[5]

A Scalable Approach to Combinatorial Library Design

To minimize


F

77

 
with respect to the lead compounds rj , we


set the corresponding gradients equal to zero, i.e., rF j = 0; this

yields the corresponding implicit equations for the locations of
lead compounds:
rj =


i

p(xi |rj )xi , 1 j M , where

p(xi )p(rj |xi )

p(xi |rj ) = 
k p(xk )p(rj |xk )

[6]

Note that p(xi |rj ) denotes the posterior probability calculated

using Bayes rule and the above equations clearly convey the centroid aspect of the solution.

 The DA algorithm consists of minimizing F with respect to
rj starting at high values of Tk and then tracking the minimum
Tk . At each k,
of 
F while

 lowering
and
use
equation [4] to compute the new weights
1. Fix
r
j


p(rj |xi ) .


2. Fix p(rj |xi ) and use
 equation [6] to compute the lead
compound locations rj .
4.2. A Scalable
Algorithm

As noted earlier, one of the major problems with combinatorial optimization algorithms is that of scalability, i.e., the number
of computations scales up exponentially with an increase in the
amount of data. In the DA algorithm, the computational complexity can be addressed in two steps first by reducing the number of iterations and second by reducing the number of computations at every iteration. The DA algorithm, as described earlier,
exploits the phase transition feature (18) in its process to decrease
the number of iterations (in fact in the DA algorithm, typically the
temperature variable is decreased exponentially which results in
few iterations). The number of computations per iteration in the
DA algorithm is O(M 2 N ), where M is the number of lead compounds and N is the total number of compounds in the underlying VCL. In this section, we present an algorithm that requires
fewer computations per iteration. This amendment becomes necessary in the context of the selection problem in combinatorial
chemistry as the sizes of the dataset are so large that DA is typically too slow and often fails to handle the computational complexity. We exploit the features inherent in the DA algorithm that,
for a given temperature, the farther an individual compound is
from a cluster, the lower is its influence on the cluster (as is evident from equation [4]). That is, if two clusters are far apart,
then they have very small interaction between them. Thus, if we
ignore the effect of a separated cluster on the remaining compound locations, the resulting error will not be significant (see
Fig. 4.1). Ignoring the effects of separated regions (i.e., groups

78

Sharma, Salapaka, and Beck

Fig. 4.1. (a) Illustration depicting the different clusters in the dataset, together with the
interaction between each pair of points (and clusters). (b) Separated regions determined
after characterizing intercluster interaction and separation.

of clusters) on one another will result in a considerable reduction

in the number of computations since the points that constitute a
separated region will not contribute to the distortion and entropy
computations for the rest. This computational saving increases as
the temperature decreases since the number of separated regions,
which are now smaller, increases as the temperature decreases.
4.2.1. Cluster Interaction
and Separation

In order to characterize the interaction between different clusters,

it is necessary to consider the mechanism of cluster identification during the process of the DA algorithm. As the temperature (Tk ) is reduced after every iteration, the system undergoes a series of phase transitions (see (18) for details). In this
annealing process, at high temperatures that are above a precomputable critical value, all the lead compounds are located at
the centroid of the entire descriptor space, thereby there is only
one distinct location for the lead compounds. As the temperature is decreased, a critical temperature value is reached where a
phase transition occurs, which results in a greater number of distinct locations for lead compounds and consequently finer clusters
are formed. This provides us with a tool to control the number
of clusters we want in our final selection. It is shown (18) for
2

a square Euclidean distance d(xi , rj ) = xi rj that a cluster
Ri splits at a critical temperature Tc when twice the maximum
eigenvalue of the posterior covariance matrix, defined by Cx|rj =

p(xi )p(xi |rj )(xi rj )(xi rj )T , becomes greater than the temi


perature value, i.e., when Tc 2max Cx|ri . This is exploited in
the DA algorithm to reduce the number of iterations by jumping from one critical temperature to the next without significant
loss in performance. In the DA algorithm, the lead location rj
is primarily determined by the compounds near it since far-away
points exert small influence, especially at low temperatures. The
association probabilities p(rj |xi ) determine the level of interaction
between the cluster Rj and the data-point xi . This interaction
decays exponentially with the increase in the distance between
rj and xi . The total interaction exerted by all the data-points in a

A Scalable Approach to Combinatorial Library Design

79

given space determines the relative weight of each cluster, p(r)j :=

N
N
i p(xi , rj ) =
i p(rj |xi )p(xi ), where p(rj ) denotes the weight
that data-points
of cluster Rj . We define the level of interaction

in cluster Ri exert on cluster Rj by ji = xRi p(rj |x)p(x). The
higher this value is, the more interaction exists between clusters
Ri and Rj . This gives us an effective way to characterize the interaction between various clusters in a dataset. In a probabilistic
framework, this interaction can also be interpreted as the probability of transition from Ri to Rj . Consider the m n matrix
mn



p(r1 |x)p(x)
p(r1 |x)p(x)
xR1

xRm


p(r2 |x)p(x)
p(r1 |x)p(x)

xRm
A = xR1

..
..
..

.
.
.




p(rm |x)p(x)
p(r1 |x)p(x)
xR1

xRm

In a probabilistic framework, this matrix is a finitedimensional Markov operator, with the term Aj,i denoting the
transition probability from region Ri to Rj . The higher the transition probability, the greater is the amount of interaction between
the two regions. Once the transition matrix is formed, the next
step is to identify regions, that is, groups of clusters, which are
separate from the rest of the data. The separation is characterized by a quantity which we denote by . We say a cluster (Rj ) is
-separate if the level of its interaction with each of the other clusters (Aj,i , i = 1, 2, . . . , n, i = j) is less than . The value is used
to partition the descriptor space into separate regions for reduced
and scalable computational effort, and it quantifies the increase
in the distortion cost function of the proposed scalable algorithm
with respect to the DA algorithm.
4.2.2. Trade-Off
Between Error in Lead
Compound Location and
Computation Time

As was discussed in Section 4.2, the greater the number of separate regions we use, the smaller the computation time for the
scalable algorithm. At the same time, a greater number of separate regions results in a higher deviation in the distortion term
of the proposed algorithm from the original DA algorithm. This
trade-off between reduction in computation time and increase in
distortion error is systematically addressed in the following. For
any pair (rj , V ), where rj is a lead compound and V is a subset of
the descriptor space , we define
Gj (V ) : =

xi p(xi )p(rj |xi ),

xi V

Hj (V ) : =

xi V

p(xi )p(rj |xi )

[7]

80

Sharma, Salapaka, and Beck

Then, from the DA algorithm, the location of the lead comG ()
pound (rj ) is determined by rj = Hj () . Since the cluster j is
j
separated from all the other clusters, the lead compound location
r j will be determined in the scalable algorithm by


xi j

r j = 

xi p(xi )p(rj |xi )

xi j

p(xi )p(rj |xi )

Gj (j )
Hj (j )

[8]

We obtain the component-wise difference between rj and r j

by subtracting terms. Note that we use the symbols and  for
component-wise operations. On simplifying, we have


 

rj rj  =
where

cj



max Gj (cj )Hj (j ),Gj (j )Hj (cj )
Hj (j )Hj ()

[9]

= \j

Denoting the cardinality of  by N and Mjc =

1
N

xi cj

xi ,

we note that


Gj (cj )
xi Hj (cj ) = NMjc Hj (cj )

[10]

xi cj

We have assumed x = 0 without any loss of generality since

the problem definition is independent of translation or scaling
factors. Thus,


max NMjc Hj (j ), Gj (j ) Hj (cj )

 
r j rj  =
Hj (j )Hj ()
  H (c ) 

j
(
)
G
j
j
j
= max NMjc ,
Hj (j )
Hj ()
then dividing through by N and using M =

1
N

xi  xi


 c

kj
 

r j rj 
Mj M j
k =j
= max
,
j , where j = 
MN
M M
kj

[11]

gives

[12]

and kj is the level of interaction between cluster j and k . For

a given dataset, the quantities M , Mj ,and Mjc are known a priori.


For the error in lead compound location r j rj  /M to be less
than a given value j (where j > 0), we must choose j such that

A Scalable Approach to Combinatorial Library Design

j


j
N max

4.2.3. Scalable
Algorithm

Mjc Mj
M , M

81

[13]

1. Initiate the DA algorithm and determine lead compound

locations together with the weighting parameters.
2. When a split occurs (phase transition),
identify individual

clusters and use the weights p(rj |x) to construct the transition matrix.
3. Use the transition matrix to identify separated clusters and
group them to form separated regions. k will be separated
from j if the entries Aj,k and Ak,j are less than a chosen jk .
4. Apply the DA to each region, neglecting the effect of separate regions on one another.
5. Stop if the terminating criterion (such as maximum number
of lead compounds (M) or maximum computation time) is
met, otherwise go to 2.
Identification of separate regions in the underlying data provides us with a tool to efficiently scale the DA algorithm. In the
DA algorithm, at any iteration, the number of computations is
M 2 N . In the proposed scalable algorithm, thenumber of computations at a given iteration is proportional to sk=1 Mk2 Nk , where



Nk N = sk=1 Nk is the number of compounds and Mk is the
number of clusters in the kth region. Thus, the scalable algorithm
saves computations at each iteration. This savings increases as
temperature decreases since corresponding values of Nk decrease.
Moreover, since the scalable algorithm can run these s DA algorithms in parallel, it will result in additional potential savings in
computational time.

5. Simulation
Results
5.1. Design for
Diversity and
Representativeness

As a first step, a fictitious dataset (VCL) was created to present the

proof of concept for the proposed optimization algorithm. The
VCL was specifically designed to simultaneously address the issue
of diversity and representativeness in the lead-generation library
design. This dataset consists of few points that are outliers while
most of the points are in a single cluster. Simulations were carried
out in MATLAB. The results for dataset 1 are shown in Fig. 4.2.
The pie chart in Fig. 4.2 shows the relative weight of each lead
compound. As was required, the algorithm gave larger weights at
locations which had larger numbers of similar compounds. At the
same time, it should be noted that the key issue of diversity is not

82

Sharma, Salapaka, and Beck

Fig. 4.2. Simulation results for dataset 1. (a) The locations xi , 1 i 200, of compounds (circles) and rj , 1 j
10, of lead compounds (crosses) in the 2-d descriptor space. (b) The weights j associated with different locations of
lead compounds. (c) The given weight distribution p(xi ) of the different compounds in the dataset. Reprinted (adapted
or in part) with permission from Journal of Chemical Information and Modeling. Copyright 2008 American Chemical
Society.

compromised. This is due to the fact that the algorithm inherently recognizes the natural clusters in the VCL. As is seen from
the figure, the algorithm identifies all clusters. The two clusters
which were quite distinct from the rest of the compounds are also
identified albeit with a smaller weight. As can be seen from the
pie chart, the outlier cluster was assigned a weight of 2%, while
the central cluster was assigned a significant weight of 22%.
5.2. Scalability and
Computation Time

In order to demonstrate the computational savings, the algorithm

was tested on a suite of synthesized datasets. The first set was
obtained by identifying ten random locations in a square region
of size 400 400. These locations were then chosen as the cluster centers. Next, the size of each of these clusters was chosen
and all points in the cluster were generated by a normal distribution of randomly chosen variance. A total of 5,000 points comprised this dataset. All the points were assigned equal weights (i.e.,
p(xi ) = N1 for all xi ). Figure 4.3 shows the dataset and the
lead compound locations obtained by the original DA algorithm.
The crosses denote the lead compound locations (rj ) and the pie
chart gives the relative weight of each lead compound (j ).
The algorithm starts with one lead compound at the centroid of the dataset. As the temperature is reduced, the cluster
is split and separate regions are determined at each such split.

A Scalable Approach to Combinatorial Library Design

83

Fig. 4.3. (a) Locations xi , 1 i 5, 000, of compounds (circles) and rj , 1 j

12, of lead compounds (crosses) in the 2-d descriptor space determined from the original algorithm. (b) Relative weights j associated with different locations of lead compounds. Reprinted (adapted or in part) with permission from Journal of Chemical
Information and Modeling. Copyright 2008 American Chemical Society.

Figure 4.4a shows the four separate regions identified by the

algorithm (as described in Section 4.2.1) at the instant when
12 lead compound locations have been identified. Figure 4.4b
shows a comparison between the two algorithms. Here the crosses
represent the lead compound locations (rj ) determined by the
original DA algorithm and the circles represent the locations (r j )
determined by the proposed scalable algorithm. As can be seen
from the figure, there is little difference between the locations
obtained by the two algorithms. The main advantage of the scalable algorithm is in terms of computation time and its ability to

84

Sharma, Salapaka, and Beck

Fig. 4.4. (a) Separated regions R1 , R2 , R3 , and R4 as determined by the proposed algorithm. (b) Comparison of lead
compound locations rj and r j . Reprinted (adapted or in part) with permission from Journal of Chemical Information
and Modeling. Copyright 2008 American Chemical Society.

Table 4.1
Comparison between the original and proposed algorithm
Algorithm

Distortion

Computation time (s)

The original DA

300.80

129.41

Proposed algorithm

316.51

21.53

Reprinted (adapted or in part) with permission from Journal of Chemical Information and Modeling. Copyright 2008 American Chemical Society

handle larger datasets. The results from the two algorithms are
presented in Table 4.1. As can be seen, the proposed scalable
algorithm takes just about 17% of the time used by the original (nonscalable) algorithm and results in only a 5.2% increase in
distortion; this was obtained for = 0.005. Both the algorithms
were terminated when the number of lead compounds reached
12. The computation time for the scalable algorithm can be further reduced (by changing ), but at the expense of increased
distortion.
5.2.1. Further Examples

The scalable algorithm was applied to a number of different

datasets. Results for three such cases have been presented in
Fig. 4.5. The dataset in Case 2 is comprised of six randomly
chosen cluster centers with 1,000 points each. All the points
were assigned equal weights (i.e., p(xi ) = N1 for all xi ).
Figure 4.5a shows the dataset and the eight lead compound locations obtained by the proposed scalable algorithm. The dataset in
Case 3 is also comprised of eight randomly chosen cluster locations with 1,000 points each. Both the algorithms were executed
till they identified eight lead compound locations in the underlying dataset. Case 4 is comprised of two cluster centers with 2,000

A Scalable Approach to Combinatorial Library Design

85

(b)

(a)

(c)

Fig. 4.5. (a, b, c) Simulated dataset with locations xi of compounds (circles) and lead compound locations rj (crosses)
determined by the algorithm. Reprinted (adapted or in part) with permission from Journal of Chemical Information
and Modeling. Copyright 2008 American Chemical Society.

Table 4.2
Distortion and computation times for different datasets
Computation
time (s)

Case

Algorithm

Distortion

Case 2

The original DA
Proposed algorithm

290.06
302.98

44.19
11.98

Case 3

The original DA
Proposed algorithm

672.31
717.52

60.43
39.77

Case 4

The original DA
Proposed algorithm

808.83
848.79

127.05
41.85

Reprinted (adapted or in part) with permission from Journal of Chemical Information and Modeling. Copyright 2008 American Chemical Society

points each. Both the algorithms were executed till they identified 16 lead compound locations. Results for the three cases have
been presented in Table 4.2.

86

Sharma, Salapaka, and Beck

It should be noted that both the algorithms were terminated

after a specific number of lead compound locations had been identified. The proposed algorithm took far less computation time
when compared to the original algorithm while maintaining less
than 5% error in distortion.
5.3. Drug Discovery
Dataset

This dataset is a modified version of the test library set (19). Each
of the 50,000 members in this set is represented by 47 descriptors which include topological, geometric, hybrid, constitutional,
and electronic descriptors. These molecular descriptors are computed using the Chemistry Development Kit (CDK) Descriptor
Calculator (20, 21). These 47-dimensional data were then normalized and projected onto a two-dimensional space. The projection was carried out using Principal Component Analysis. Simulations were completed on this two-dimensional dataset. The
proposed scalable algorithm was used to identify 25 lead compound locations from this dataset (see Fig. 4.6). The algorithm
gave higher weights at locations which had larger numbers of similar compounds. Maximally diverse compounds are identified with
a very small weight. The original version of the algorithm could
not complete the computations for this dataset (on a 512 MB
RAM 1.5 GHz Intel Centrino processor).

5.4. Additional
Constraints on Lead
Compounds

As was discussed in Section 3, the multiobjective framework

of the proposed algorithm allows us to incorporate additional constraints in the selection problem. In this section, we
have addressed two such constraints, namely the experimental
resources constraint and the exclusion/inclusion constraint.

Fig. 4.6. Choosing 25 lead compound locations from the drug discovery dataset.
Reprinted (adapted or in part) with permission from Journal of Chemical Information
and Modeling. Copyright 2008 American Chemical Society.

A Scalable Approach to Combinatorial Library Design

5.4.1. Constraints on
Experimental Resources

87

In this dataset, the VCL is divided into three classes based on the
experimental supplies required by the compounds for testing, as
shown in Fig. 4.7a by different symbols. It contains a total of
280 compounds with 120 of the first class (denoted by circles),
40 of the second class (denoted by squares), and 120 of the third
class (denoted by triangles). We incorporate experimental supply
constraints into the algorithm by translating them into direct constraints on each of the lead compounds. With these experimental
supply constraints, the algorithm was used to select 15 lead compound locations (rj ) in this dataset with capacities (Wjn ) fixed for

(a)

(b)

Fig. 4.7. (a) Simulation results with constraints on experimental resources. (b) Simulation results with exclusion constraint. The locations xi , 1 i 90, of compounds (circles) and rj , 1 j 6, of lead compounds (crosses). Dotted circles represent undesirable properties. Reprinted (adapted or in part) with permission from Journal of
Chemical Information and Modeling. Copyright 2008 American Chemical Society.

88

Sharma, Salapaka, and Beck

each class of resource. The crosses in Fig. 4.7a represent the selection from the algorithm in the wake of the capacity constraints for
different types of compounds. As can be seen from the selection,
the algorithm successfully addressed the key issues of diversity and
representativeness together with the constraints that were placed
due to experimental resources.
5.4.2. Constraints on
Exclusion and Inclusion
of Certain Properties

There may arise scenarios where we would like to inhibit selection of compounds exhibiting properties within certain prespecified ranges. This constraint can be easily incorporated in the
cost function by modifying the distance metric used in the problem formulation. Consider a case in a 2-d dataset where each
point xi has an associated radius (denoted by ij ). The selection problem is the same, but with the added constraint that all
the selected lead compounds (rj ) must be at least ij distance
removed from xi . The proposed algorithm can be modified to
solve this problem by defining the distance function, given by


2
d(xi , rj ) = xi rj ij , which penalizes any selection (rj )
which is in close proximity to the compounds in the VCL. For
the purpose of simulation, a dataset was created with 90 compounds (xi , i = 1, . . . , 90). The dotted circle around the locations
xi denotes the region in the property space that is to be avoided
by the selection algorithm. The objective was to select six lead
compounds from this dataset such that the criterion of diversity and representativeness is optimally addressed in the selected
subset. The selected locations are represented by crosses. From
Fig. 4.7b, note that the algorithm identifies the six clusters under
the constraint that none of the cluster centers are located in the
undesirable property space (denoted by dotted circles).

6. Conclusions
In this chapter, we proposed an algorithm for the design of leadgeneration libraries. The problem was formulated in a constrained
multiobjective optimization setting and posed as a resource allocation problem with multiple constraints. As a result, we successfully tackled the key issues of diversity and representativeness of
compounds in the resulting library. Another distinguishing feature of the algorithm is its scalability, thus making it computationally efficient as compared to other such optimization techniques.
We characterized the level of interaction between various clusters and used it to divide the clustering problem with huge data
size into manageable subproblems with small size. This resulted
in significant improvements in the computation time and enabled
the algorithm to be used on larger sized datasets. The trade-off
between computation effort and error due to truncation is also
characterized, thereby giving an option to the end user.

A Scalable Approach to Combinatorial Library Design

89

References
1. Gordon, E. M., Barrett, R. W., Dower,
W. J., Fodor, S. P. A., Gallop, M. A.
(1994) Applications of combinatorial technologies to drug discovery. 2. Combinatorial
organic synthesis, library screening strategies,
and future directions. J Med Chem 37(10),
13851401.
2. Blaney, J., Martin, E. (1997) Computational
approaches for combinatorial library design
and molecular diversity analysis. Curr Opin
Chem Biol 1, 5459.
3. Willett, P. (1997) Computational tools for
the analysis of molecular diversity. Perspect
Drug Discov Design, 7/8, 111.
4. Rassokhin, D. N., Agrafiotis, D. K. (2000)
Kolmogorov-Smirnov statistic and its applications in library design. J Mol Graph Model
18(45), 370384.
5. Lipinski, C. A., Lomabardo, F., Dominy, B.
W., Feeny, P. J. (1997) Experimental and
computational approaches to estimate solubility and permeability in drug discovery and
development setting. Adv Drug Del Review
23, 225.
6. Higgs, R. E., Bemis, K. G., Watson, I. A.,
Wikel, J. H. (1997) Experimental designs
for selecting molecules from large chemical databases. J Chem Inf Comput Sci 37,
861870.
7. Clark, R. D. (1997) Optisim: an extended
dissimilarity selection method for finding
diverse representative subsets. J Chem Inf
Comput Sci 37(6), 11811188.
8. Agrafiotis, D. K., Lobanov, V. S. (2000)
Ultrafast algorithm for designing focussed
combinatorial arrays. J Chem Inf Comput Sci
40, 10301038.
9. Salapaka, S., Khalak, A. (2003) Constraints
on locational optimization problems. Proceedings of the IEEE Control and Decisions
Conference. Maui, HI, 912 December 2003,
pp. 17411746.
10. Sharma, P., Salapaka, S., Beck, C. (2008)
A scalable approach to combinatorial library

11.
12.
13.
14.

15.
16.
17.
18.

19.

20.

21.

design for drug discovery. J Chem Inf Model

48(1), 2741.
Gersho, A., Gray, R. (1991) Vector Quantization and Signal Compression. Kluwer, Boston,
Massachusetts.
Drezner, Z. (1995) Facility location: a survey
of applications and methods. Springer Series
in Operations Research, Springer, New York.
Du, Q., Faber, V., Gunzburger, M. (1999)
Centroidal Voronoi tessellations: applications
and algorithms. SIAM Rev 41(4), 637676.
Therrien, C. W. (1989) Decision, Estimation
and Classification: An Introduction to Pattern Recognition and Related Topics, 1st ed.
Wiley, New York.
Haykin, S. (1998) Neural Networks: A Comprehensive Foundation, Prentice Hall, Englewoods Cliffs, NJ.
Gray, R., Karnin, E. D. (1982) Multiple local
minima in vector quantizers. IEEE Trans
Inform Theor 28, 256361.
Lloyd, S. P. (1982) Least squares quantization in PCM. IEEE Trans Inform Theory
28(2), 129137.
Rose, K. (1998) Deterministic annealing
for clustering, compression, classification,
regression and related optimization problems. Proc IEEE 86(11), 22102239.
Mcmaster hts lab competition. HTS data
mining and docking competition. http://
hts.mcmaster.ca/downloads/82bfbeb4f2a4-4934-b6a8-804cad8e25a0.html
(accessed June 2006).
Guha, R. (2006) Chemistry Development Kit (CDK) descriptor calculator
GUI (v 0.46). http://cheminfo.informatics.
indiana.edu/rguha/code/java/cdkdesc.html
(accessed October 2006).
Steinbeck, C., Hoppe, C., Kuhn, S., Floris,
M., Guha, R., Willighagen, E. L. (2006)
Recent developments of the Chemistry
Development Kit (CDK) an open-source
JAVA library for chemo and bioinformatics.
Curr Pharm Des 12(17), 21102120.

Chapter 5
Application of FreeWilson Selectivity Analysis
for Combinatorial Library Design
Simone Sciabola, Robert V. Stanton, Theresa L. Johnson,
and Hualin Xi
Abstract
In this chapter we present an application of in silico quantitative structureactivity relationship (QSAR)
models to establish a new ligand-based computational approach for generating virtual libraries. The Free
Wilson methodology was applied to extract rules from two data sets containing compounds which were
screened against either kinase or PDE gene family panels. The rules were used to make predictions for all
compounds enumerated from their respective virtual libraries. We also demonstrate the construction of
R-group selectivity profiles by deriving activity contributions against each protein target using the QSAR
models. Such selectivity profiles were used together with protein structural information from X-ray data
to provide a better understanding of the subtle selectivity relationships between kinase and PDE family
members.
Key words: QSAR, FreeWilson, MLR, virtual libraries, combinatorial chemistry, protein kinase,
PDE, enzyme inhibition, enzyme selectivity, docking.

1. Introduction
Combinatorial chemistry has become an essential tool in the
pharmaceutical industry for identifying new leads and optimizing the potency of potential lead candidates while reducing the
time and costs associated with producing effective and competitive new drugs. By speeding up the process of chemical synthesis,
it is now possible to generate large diverse compound libraries
to screen for novel bioactivities. At the same time improvements
in high-throughput screening (HTS) allow selectivity panels for
J.Z. Zhou (ed.) , Chemical Library Design, Methods in Molecular Biology 685,
DOI 10.1007/978-1-60761-931-4_5, Springer Science+Business Media, LLC 2011

91

92

Sciabola et al.

gene families or diverse off-target activity to be regularly run

against all compounds of interest. Unfortunately, despite these
significant synthetic and screening efforts, only few novel lead
candidates have been identified for optimization, resulting in
increased interest in the use of computational techniques for
the design of focused combinatorial libraries rather than simply
diverse ones. An additional benefit of these libraries is that they
can be used to probe enzyme specificity by analyzing the activity of diverse groups of intrafamily proteins using in silico methods. In this respect, protein kinases (PKs) and phosphodiesterases
(PDEs) represent two well-known examples of enzyme superfamilies which have been heavily pursued by both pharmaceutical
companies and academic groups because of their mechanistic role
in many diseases, thus providing us with a large amount of structural and biological data to be used for developing and validating
new in silico methodologies.
Protein kinases (1, 2) catalyze the transfer of the terminal
phosphoryl group of ATP to specific hydroxyl groups of serine,
threonine, or tyrosine residues of their protein substrates. Because
protein kinases have profound effects on a cell, their activity is
highly regulated by the binding of activator/inhibitor proteins or
small molecules or by controlling their location in the cell relative to their substrates. Intracellular phosphorylation by protein
kinases, triggered in response to extracellular signals, provides a
mechanism for the cell to switch on or off many diverse processes
(3). Deregulated kinase activity is a frequent cause of disease, particularly cancer, since kinases regulate many aspects that control
cell growth, movement, and death. Drugs which inhibit specific
kinases are being developed to treat many diseases and several are
currently in clinical use. These include (1) Gleevec (Imatinib) (4)
for chronic myeloid leukemia (CML), (2) Sutent (Sunitinib) (5),
a multitargeted receptor tyrosine kinase for the treatment of renal
cell carcinoma (RCC) as well as imatinab-resistant gastrointestinal stromal tumor (GIST), (3) Iressa (Gefitinib) (6), and Erlotinib
(Tarceva) (7) for non-small cell lung cancer (NSCLC). Previously,
studies have shown how molecular specificity varies widely among
known inhibitors (8), and this variation is not dictated by the
general chemical scaffold of an inhibitor (e.g., EGFR inhibitors,
belonging to the quinazoline/quinoline class, range from highly
specific to quite promiscuous) or by the primary, intended kinase
target toward which the particular inhibitor was initially optimized (e.g., compounds considered tyrosine kinase inhibitors also
bind to Ser-Thr kinases and vice versa). Moreover, with over 500
kinases in the human genome, selectivity is a daunting task and
predicting selectivity based on the protein-binding site or ligand
pharmacophores is extremely challenging given the high degree
of homology across the kinase protein family, particularly in the
active site region.

Application of FreeWilson Selectivity Analysis

93

The second gene family we included in our study is the PDE

superfamily of enzymes that degrade cyclic adenosine monophosphate (cAMP) and cyclic guanosine monophosphate (cGMP)
(911). Both cAMP and cGMP are intracellular second messengers that play a key role in mediating cellular responses to various
hormones and neurotransmitters (12, 13), and their intracellular concentration is tightly regulated at the level of synthesis (by
the catalytic reaction of adenylyl cyclase and guanylyl cyclase) as
well as degradation (by binding to cyclic nucleotide phosphodiesterases). PDEs are involved in a wide array of pharmacological
processes, including proinflammatory mediator production and
action, ion channel function, muscle contraction, learning, differentiation, apoptosis, glycogenolysis, and gluconeogenesis (14),
and have become recognized as important drug targets for the
treatment of various diseases, such as heart failure, depression,
asthma, inflammation, and erectile dysfunction (13, 1517).
Since the early discovery of multiple phosphodiesterase isoforms and their potential use as therapeutic targets (18), the biological and functional understanding around PDEs has expanded
from what was understood to be a family of three isozymes (19)
toward a total of 21 human PDE genes falling into 11 families
with over 60 isoforms (10, 12, 15, 2024). Selective inhibitors
for each of the multiple PDE forms can offer an opportunity for
desired therapeutic intervention and would be an extremely useful
tool in drug discovery efforts for a medicinal chemist. Although
there are distinct differences in the full-length structure of the
PDEs, not surprisingly the catalytic domain that shares a common
function across different isoforms has a more conserved structure,
making the design of highly selective PDE inhibitors a difficult
challenge. As for kinases, PDE inhibition has potential therapeutic utility but care must be taken in the rational design of active
inhibitors to avoid unwanted off-target PDE inhibition.
Over the past few years, Pfizer has focused on developing selectivity screening platforms (25) to provide high-quality
data against a diverse range of PKs and PDEs, which has been
used to guide therapeutic projects by analyzing structureactivity
relationship (SAR) and identifying potential off-target liabilities
of compounds within a chemical series. The integration of this
highly valuable data together with appropriate computational
methods can speed up the overall lead discovery process by allowing the optimization of property-based design within a homologous series. However, the success of such studies depends on the
choice of an appropriate molecular characterization, through the
use of informative descriptors.
In the chapter, we report a successful application of
the FreeWilson (2630) methodology to model structure
activity/selectivity relationships. The FujitaBan (3134) modification of FreeWilson coupled with multiple linear regression

94

Sciabola et al.

analysis (MLR) was used to model the selectivity profiles of different chemical series in our in-house kinase and PDE screening
panel. Overall, reliable estimations for R-group activity contributions against each protein in the data set were observed and
used for enumerating focused virtual libraries to predict more
selective inhibitors. When an external test set of cherry-picked
compounds was used to test the validity of the in silico models, a strong correlation of experimental versus predicted inhibition values was found. Lastly, the availability of X-ray structures
in the public domain for both PKs and PDEs allowed us to further validate our QSAR models by combining the information
from the FreeWilson approach with the three-dimensional (3D)
structural knowledge of the target, providing more insight into
specific enzyme selectivity.

2. Methods
2.1. Assay Conditions

All of the kinase assays are performed in a 384-well format

using either a radioactive or Caliper protocol (25). In all assays,
5 L of 5 concentration compound in 3.75% DMSO is added
to the plates. 10 L of 2.5 enzyme in 1.25 kinase buffer
(optimized for each individual kinase) is then added, followed
by a 15-min preincubation at room temperature. 10 L of a
2.5 mixture of peptide substrate (optimized for each individual
kinase) and ATP in 1.25 kinase buffer are then added to initiate
the reaction. Each assay is run at the experimentally determined
MichaelisMenten constant (Km ) concentration of ATP for the
relevant kinase with an incubation time that was determined to
be within the linear reaction time. Reactions are stopped by the
addition of EDTA to a final concentration of 20 mM. Detection
of phosphorylated substrate is achieved using either a radioactive
method or a nonradioactive mobility shift assay format (Caliper).
In the radioactive assay, tracer amounts of -(33) P-labeled ATP
are included in the reaction, and biotinylated peptide substrates
are used. After the reactions are stopped, 25 L is transferred
to streptavidin-coated FlashplatesTM (Perkin Elmer). Plates are
washed with 50 mM Hepes and soaked for 1 h with 500 M
unlabeled ATP before reading in a TopCount. Alternatively, for
the mobility shift assay, reactions are stopped within the assay
plates followed by detection of fluorescently labeled substrates on
a Caliper LC3000 using a 12-sipper chip and conditions that were
optimized for each kinase.
The PDE assays are performed in a 384-well format using a
radioactive protocol where the enzymatic activities were assayed
by using 3 H-cAMP or 3 H-cGMP as substrates to a final

Application of FreeWilson Selectivity Analysis

95

concentration of 20 nM. The catalytic domain of PDEs was incubated with a reaction mixture of 50 mM TrisHCl, pH 7.5,
1.3 mM MgCl2 , 1 mM DTT, and 3 H-cAMP or 3 H-cGMP at
room temperature on an orbital shaker for 30 min. Compounds
to be tested are submitted to the assay at a concentration of
4 mM in 100% DMSO. Compounds are initially diluted in 50%
DMSO/water. Subsequent dilutions are in 15% DMSO/water to
achieve 5 the desired assay concentration. Each well receives
10 L drug or DMSO vehicle, 20 L 3 H-cAMP or 3 H-cGMP,
and 20 l enzyme (diluted 1:1,000 in assay buffer). The incubation is terminated by the addition of 25 L of PDE SPA beads
(0.2 mg/well). The reaction product 3 H-cAMP or 3 H-cGMP
was precipitated out by BaSO4 while unreacted 3 H-cAMP or
3 H-cGMP remained in supernatant. After centrifugation, the
radioactivity in the supernatant was measured in a liquid scintillation counter after a 500 min delay. The enzymatic properties were analyzed by the steady-state kinetics. The nonlinear
regression of the MichaelisMenten equation as well as Eadie
Hofstee plots was analyzed to obtain the values of KM , Vmax , and
kcat . For measurement of IC50 , ten concentrations of inhibitors
were used at the substrate concentration of <1/10 KM and the
suitable enzyme concentration. All measurements were repeated
three times.
2.2. Data Sets

In the kinase case study, 975 compounds based on chemical

series belonging to four different chemotypes (Table 5.1) were
screened against 45 protein kinases, selected to provide maximal coverage across subfamilies within the kinome (25). The data
set consists of 388 compounds with the Diaminopyrimidine core
(R1=77, R2=183), 312 with the Pyrrolopyrazole core (R1=124,
R2=87), 181 with the Pyrrolopyrimidine core (R1=8, R2=169),
and 94 sharing the Quinazoline core (R1=19, R2=5, R3=37,
R4=33). Due to changes in the panel over time, not all the compounds in the study were screened against each individual kinase,
giving rise to an incomplete combinatorial matrix. However, these
chemical series were selected trying to consistently meet the criteria of having a high number of compound per kinase assay (Fig.
5.1). Percent inhibition data at two compound concentrations,
1 M and 10 M, was first transformed into pIC50 (35) and then
combined to give a single pIC50 value by applying the following
equations:


100 (percent inhibition@10 M)
6
pIC50 @1 M = log 10
percent inhibition@10 M


100 (percent inhibition@10 M)
pIC50 @10 M = log 105
percent inhibition@10 M

96

Sciabola et al.

Fig. 5.1. Number of compounds tested against each of the 45 protein kinases in the
in-house selectivity panel. Histogram bars are subdivided according to the compounds
frequency in the four kinase chemical series.

pICCalc
50

pIC50 @1M,
= pIC50 @10M,

pIC50 @1M+pIC50 @10M

2

Inhib@10M > 99%

Inhib@1M < 5%
, 5% Inhib@1M 99%

The reported block function was adopted to improve the

overall correlation between calculated and experimental pIC50 at
the two different concentrations. As reported previously (25, 36),
in the lower range of inhibition, below 5%, a stronger correlation
between pICCalc
50 computed at 10 M concentration and experimental pICCalc
50 was found, when compared to 1 M. An opposite
trend was present in the upper range of inhibition (above 99%),
where pICCalc
50 computed at 1 M concentration tended to correlate better with experiment than that at 10 M. For inhibition
values between the previously defined cut-offs, we used the average pIC50 .
The second data set consists of 1,505 total compounds sharing a unique chemotype (Pyrazolopyrimidine) tested in two different PDE biochemical assays (PDE2 and PDE10). Four sites
of substitutions (R1=62, R2=157, R3=872, R4=339) were
allowed to change around the Pyrazolopyrimidine core substructure (Table 5.1). Although not all the compounds in the study
were tested against both PDEs (1,357 and 1,346 compounds
tested, respectively, in PDE2 and PDE10), a large number of
compounds (1,198) were in common between the two assays,
providing us with a wealth of data to be used for studying their
selectivity profiles. Different from the kinase data set, all the

Application of FreeWilson Selectivity Analysis

97

Table 5.1
2D depiction for the five chemical series. R-positions represent sites which were allowed to change within a given
library while X-positions indicate not changing chemical
matter whose structure cannot be disclosed
Protein
family

Chemical
series

2D
depiction

Number of
Compounds

R-groups

388

R1 = 77
R2 = 183

312

R1 = 124
R2 = 87

181

R1 = 8
R2 = 169

94

R1 = 19
R2 = 5
R3 = 37
R4 = 33

1505

R1 = 62
R2 = 157
R3 = 872
R4 = 339

X1
X2
N

Kinases

Diaminopyrimidine

R2

R1
N

N
H

N
H

X1

X2

H
N

O
N

Pyrrolopyrazole

R2

R1

X4

X3

NH

R2
X1

N
X2

Pyrrolopyrimidine

R1

N
X4

R4

X3

R1

Quinazoline

R3
X1

R2

R3

PDEs

R2

Pyrazolopyrimidine

N
R4

N
R1

PDE compounds were tested for IC50 , therefore, no data transformation was required in the case. The negative logarithm of
IC50 was used as a dependent variable in the model-building
process.
2.3. FreeWilson
(FW)

The FreeWilson approach was the first mathematical technique

to be developed for the quantitative prediction of the structure
activity relationships for a series of chemical analogs (26). The
basic idea behind this methodology is that the biological activity of a molecule can be described as the sum of the activity contributions of specific substructures (parent fragment and
the corresponding substituents). It does not require any substituent parameters or descriptors to be defined; only the activity is needed. The underlying assumption in FreeWilson modeling is that the contribution of each substituent to the biological

98

Sciabola et al.

activity is additive and constant, regardless of the structural variation on the other sites of substitution in the rest of the molecule.
The classical FreeWilson linear model is expressed by the following equation:
BioActivity =

ij Rij +

ij

where the constant term (activity value of the unsubstituted

compound) is the overall average of biological activities and ij
is the R-group contribution of substituent Ri in position j. If
substituent Ri is in position j, then Rij = 1, otherwise Rij = 0.
This gives rise to a set of equations that can be potentially solved
by MLR, where ij are the regression coefficients, Rij the independent variables, and the intercept. Unfortunately, MLR cannot be applied directly to the resulting structural matrix due to a
linear dependence on its columns (34). One way to get around
these dependencies is to use the FujitaBan modification where
the activity contribution of each substituent is relative to H and
the constant term , obtained by the least-squares method, is a
theoretically predicted activity value of the unsubstituted compound itself (all R-groups set to H) (31). Kubinyi et al. have
shown that the original FreeWilson and the FujitaBan modifications are linearly related, with the latter approach being a
linear transformation of the classical FreeWilson model (34).
Additionally, the FujitaBan model leads to a number of important advantages. First, no complex transformation of the structural matrix is required and only the removal of one column for
each site of substitution is necessary to move from the structural matrix to the FujitaBan matrix. Second, the matrix is not
changed by the addition or elimination of a compound. Third,
in the FujitaBan model the constant term in the linear equation is derived theoretically by applying the least-squares method
and therefore not markedly influenced by the addition or elimination of a compound. In consideration of these advantages, the
FujitaBan modification of the FreeWilson mathematical model
was implemented for the analysis reported here.
2.4. FW Model
Building and
Validation

The FujitaBan modification of the FreeWilson methodology

was applied to the structural matrices of descriptors corresponding to each chemical series/biochemical assay combination analyzed in this study and individual QSAR models were built. The
first step consisted of generating the R-groups by fragmenting
all the compounds within each series, thus obtaining the initial structural matrices. After that, compounds with correlated
R-groups and outlier compounds whose R-groups did not occur
in other compounds were removed from the data set as the activity contribution for these R-groups could not be estimated. Then

Application of FreeWilson Selectivity Analysis

99

the remaining structural matrix was rearranged into independent

blocks where R-groups from one block would not cross over with
other blocks, and statistical analysis was applied to each block separately to estimate activity contribution for each R-group. Furthermore, blocks whose R-group activity contributions could not
be estimated due to a lack in R-group crossovers were further
eliminated. This block separation and compound removal procedure maximized the total number of R-group activity contributions that could be estimated.
The relationship between the enzyme inhibition data and
the chemical structures was analyzed using MLR, a multivariate
regression method able to quantitatively model the relationship
between two or more explanatory variables and a response variable by fitting a linear equation to the observed data. An MLR
model was first built independently for each series/biochemical
assay combination. The quality of the models both in terms
of fitting the experimental data and predicting the activity for
new compounds through cross-validation techniques was assessed
2 )
by computing the squared Pearson correlation coefficient (rcorr
between predicted and actual activities together with the associated standard error of correlation (STE):
#
2
rcorr
=

$

 2
  pred
pred
yi
yiact yiact
yi

itest



2
  pred
pred 2 
yi
yiact yiact
yi

itest

%
&
&
&
&
& 1
STE = &
&
&n 2
'

itest

$

 2
  pred
pred
act
act

yi
yi yi
yi

 

itest
pred
pred 2

yi
yi

2
  act

act

itest
yi yi

pred

Here, yi

itest

is the predicted activity for the ith test set compred

pound,
is its measured activity, yi
and yiact are the average
of the predicted and measured activity values, respectively, and n
is the sample size.
The squared Pearson correlation coefficients for the linear models built upon the Diaminopyrimidine, Pyrrolopyrazole,
Pyrrolopyrimidine, and Quinazoline series across the 45 pro2
= 0.82 0.95
tein kinases are, respectively, in the range of rfitting
yiact

2
2
2
= 0.87), rfitting
= 0.73 0.93 (average rfitting
=
(average rfitting
2
2
2
= 0.36 0.99 (average rfitting
= 0.80), and rfitting
=
0.85), rfitting
2
= 0.76). For the PDE case study, the
0.46 0.97 (average rfitting
correlation coefficients for the Pyrazolopyrimidine series when
tested in the PDE2 and PDE10 biochemical assays are, respec2
2
= 0.94 0.17 and rfitting
= 0.92 0.18. The highly
tively, rfitting
significant correlation between experimental and calculated pIC50

100

Sciabola et al.

confirmed the basic assumption of the FreeWilson method for

this set of biological data, which is the additivity of R-group
effects.
The models predictivity was evaluated using standard LeaveOne-Out (LOO) analysis as internal validation technique.
LOO is a cross-validation procedure that works by building
reduced models (models for which one object at a time is
removed) and using them to predict the Y-variables of the object
held out. Results obtained by applying LOO validation to the
kinase and PDE data sets are shown in Figs. 5.2 and 5.3,
respectively. In general, the predicted pIC50 is in agreement
with the calculated pIC50 derived from experimental data. In the
Diaminopyrimidine series, taking all 45 kinase models together,
6,712 LOO estimations were carried out giving a global corre2
= 0.90 and a standard error of the prelation coefficient rcorr,CV
dicted pIC50 value in the regression STE = 0.35. Similar results

Fig. 5.2. Leave-One-Out cross-validation results reported as predicted vs. experimental pIC50 values for the four kinase
chemical series. In general, model prediction of pIC50 is in good agreement with experimental pIC50 derived from percent
2
2
= 0.90 for Diaminopyrimidine (a), rcorr,CV
= 0.84 for the
of inhibition, with a global correlation coefficient rcorr,CV
2
2
Pyrrolopyrazole (b), rcorr,CV
= 0.77 for the Pyrrolopyrimidine (c), and rcorr,CV
= 0.73 for the Quinazoline (d) series.

Application of FreeWilson Selectivity Analysis

101

Fig. 5.3. Leave-One-Out cross-validation results for the Pyrazolopyrimidine series tested in the biochemical assays PDE2
(a) and PDE10 (b).

were obtained for the Pyrrolopyrazole series, where LOO estimations of 5,413 objects gave an overall correlation coefficient
2
= 0.85 (STE = 0.47), the Pyrrolopyrimidine series with
rcorr,CV
2
rcorr,CV = 0.77 (STE = 0.53 based on 650 LOO estimations),
and the Quinoline series where 707 LOO estimations resulted in
2
= 0.73 (STE = 0.64). The same cross-validation protocol
rcorr,CV
was carried out in the case of Pyrazolopyrimidine series obtain2
= 0.78 (STE =
ing the following correlation coefficients: rcorr,CV
2
0.38, 485 LOO estimations) and rcorr,CV = 0.76 (STE = 0.46,
473 LOO estimations) when tested in the PDE2 and PDE10
assays, respectively (Fig. 5.3).
Since FreeWilson models use the presence or absence of
distinct R-group fragments as the basic variables in regression,
the derived model coefficients can be treated as a quantitative
estimate of the activity contribution of each R-group. Assuming
the additive assumption holds, then these R-group contributions
can be used to make reliable predictions for all the enumerated
compounds in a virtual library, where all R-group fragments are
crossed with each other.
2.5. Virtual Library
Space Analysis

After model building and validation, the R-groups within each

chemical series were exhaustively combined with each other and
their pIC50 contributions from the FW QSAR models used to
predict the final activity of the compounds enumerated in the
virtual library. This step represents one of the key advantages of
using FW methodology over standard descriptors-based QSAR
techniques that is the deconvolution of the biological activity of
a molecule into its components (parent fragment plus the corresponding substituents). Indeed, due to experimental and synthetic limitations, typically only a small number of compounds
can be synthesized and screened against a given biochemical assay.

102

Sciabola et al.

As a result, many compounds with desired potency and selectivity profiles could potentially be missed. By using high-quality
QSAR models, the activity and selectivity of compounds in the
virtual library can be reliably estimated, thus, greatly expanding
the chemical space coverage and increasing the chance of finding
compounds with attractive biological properties.
To demonstrate this, we enumerated the full virtual library
for the five chemical series shown in Table 5.1. We obtained
861 compounds for the Diaminopyrimidine series, 1,764 compounds for the Pyrrolopyrazole series, 598 for the Pyrrolopyrimidine series, 2,370 for the Quinazoline series, and 214,486 for
the Pyrazolopyrimidine series, using only those R-groups from
the existing compounds for which the activity contribution could
be estimated across the 45 protein kinase (first four chemical
series) and the two PDE assays (Pyrazolopyrimidine). We then
calculated their selectivity profile using the QSAR models derived
from FreeWilson analysis. Among the existing compounds in the
kinases series, 27 of them (17 Diaminopyrimidines, 1 Pyrrolopyrazole, 6 Pyrrolopyrimidine, 3 Quinazoline) met our selectivity
criteria (pIC50 > 5.3 against no more than 5 kinases on the panel).
In the full virtual library, however, 111 additional compounds
(57 Diaminopyrimidines, 8 Pyrrolopyrazoles, 31 Pyrrolopyrimidine, 15 Quinazoline) were predicted to be selective. In the PDE
series, the library expansion provided with a greater enrichment
in the number of compounds potentially selective, moving from
three selective compounds in the original library (pIC50 7 in
one assay and pIC50 5.3 in the second assay) to 4,103 selective
compounds in the virtual space.
We have also noticed an increase in the number of kinases
selectively targeted upon the expansion of the inhibitors chemical space, suggesting that such a procedure would also be suitable as a tool for exploring potential Target Hopping. Indeed,
when applied to our data set, existing selective compounds from
the Diaminopyrimidine, Pyrrolopyrazole, Pyrrolopyrimidine, and
Quinazoline series targeted 14, 5, 7, and 3 protein kinases,
respectively. However, after complete enumeration of the virtual
libraries, 28, 19, 31, and 12 protein kinases were predicted to
be selectively inhibited by compounds in the four series, respectively. This shows how series originally developed for a specific
kinase could be turned into selective inhibitors for other kinases
by exploiting different R-group combinations.
2.6. R-Group
Selectivity Profiles

The objective of this analysis was to gain knowledge from the

R-group contributions as determined by the FreeWilson
methodology. Only R-groups for which a coefficient could be
determined across the 45 kinases in the panel and the 2 PDE biochemical assays reported in this study were taken into account.
For the Diaminopyrimidine series, this resulted in 36 R1- and 26

Application of FreeWilson Selectivity Analysis

103

R2-group structures, giving rise to two different matrices containing 3645 R1- and 2645 R2-group contributions. In the
Pyrrolopyrazole series, a total of 60 R1- and 35 R2-group structures were available for analysis, leading to two coefficient matrices of 6045 R1- and 3545 R2-group contributions. Analysis
of the R-group structures for the Pyrrolopyrimidine and Quinazoline series resulted in two coefficient matrices of 345 R1- and
5745 R2-group contributions for the former series and four
coefficient matrices of 445 R1-, 245 R2-, 1545 R3-, and
1145 R4-group contributions for the latter. In the Pyrazolopyrimidine PDE series, a total of 5 R1-, 79 R2-, 543 R3-, and 3
R4-group structures were available for analysis, leading to four
coefficient matrices of 52 R1-, 792 R2-, 5432 R3-, and 32
R4-group contributions.
The main objective in this R-group selectivity analysis was to
detect whether small changes in structure could give rise to large
variations in activity. This was achieved by computing all pairwise structural similarities between R-groups at each substitution
site (using a combination of structural descriptors (37, 38) and
Tanimoto as similarity measure), then keeping only R-group pairs
with Tanimoto similarity greater than 0.8. Afterward, each surviving R-group pair was assigned a profile resulting from the difference in the original coefficients profiles for the R-groups being
compared. This produced one selectivity map for each R-group
position within each different chemical series. Figures 5.4 and
5.5 show a few snapshots of this data transformation for the
Diaminopyrimidine and Pyrazolopyrimidine series, reported as
heat maps where each R-group pair/assay combination is assigned
a color ranging from white (pIC50 = 0) to red (pIC50 2).

Fig. 5.4. Structural models for binding site interactions of Diaminopyrimidine series. Selectivity maps are shown next
to each binding site model. pIC50 for the specific R-group pair/assay combination is highlighted in yellow (R-group
combinations are reported as rows and protein kinase assays as columns within the heat map). (a) R-groupA (orange)
and R-groupB (violet) at site R1 of Diaminopyrimidine docked into the crystal structure of GSK3 (1O9U). The extra
methyl in R-groupB is responsible for its increased activity contribution. (b) Position R2 of Diaminopyrimidine in protein
kinase PAK4 (2CDZ). R-groupB (violet) undergoes a 45 rotation in order to orient the tert-butyloxy tail toward the buried
lipophilic pocket made up residues R586, M585, and L448.

104

Sciabola et al.

Fig. 5.5. Structural models for binding site interactions of Pyrazolopyrimidine series. Selectivity maps are shown next
to each binding site model. pIC50 for the specific R-group pair/assay combination is highlighted in yellow (R-group
combinations are reported as rows and protein PDE assays as columns within the heat map). (a) R-groupA (orange) and
R-groupB (violet) at site R2 of Pyrazolopyrimidine docked into the in-house PDE2 crystal structure. The extra phenethyl
moiety in R-groupB makes an extended hydrophobic interaction with residue L809 and it is responsible for the observed
increased in activity. (b) Position R3 of Pyrazolopyrimidine in the PDE2 crystal structure. The presence of two extra atoms
linker in R-groupB (violet) determines its different binding mode compared to R-groupA . The 1,3-dimethoxy benzene
portion of R-groupB undergoes a 90 rotation in order to orient itself toward a buried lipophilic pocket and interacting
directly with the side chain of residue L770.

To provide more insight into kinase/PDE selectivity and

to analyze the variations in pIC50 based upon small structural
changes at the R-group level, we combined the information from
the FreeWilson approach with the 3D structural knowledge of
the target. This analysis was made possible by the availability of
numerous in-house as well as public protein kinase and phosphodiesterase crystal structures. In this respect, a structure-based
study was carried out for each R-group/protein combination
using an internal core-docking workflow (39), which consists of a
protocol specifically designed for screening multiple combinatorial libraries against a family of proteins and relies on the common
alignment of all the available protein X-ray structures.
Although all the virtual compounds were docked into their
corresponding protein crystal structures, an exhaustive analysis of
these dockings and the interpretation of the R-group contributions contained in each of the individual selectivity heat maps is
beyond the scope of this study. Our objective here was spot checking the ligand-based results obtained through the FreeWilson
analysis to see if they were consistent with the known enzyme
crystal structure; therefore, only one example for each site of substitution for the Diaminopyrimidine kinase series and the Pyrazolopyrimidine PDE series is shown here (Figs. 5.4 and 5.5).
Starting with the R1 position of Diaminopyrimidine, structural poses for R-groupA and R-groupB , as described in Table
5.2, were analyzed after docking into protein active site of
kinase GSK3 (PDB entry: 1O9U). A variation in pIC50 of 1.8

Application of FreeWilson Selectivity Analysis

105

Core

Site

Diaminopyrimidine

R1

Pyrazolopyrimidine

Table 5.2
R-group/kinase contributions from FreeWilson selectivity maps

R2

R-groupA

R-groupB

Protein

gsk3

N
N

N
S

pIC50

(RBRA)

+1.8

(1O9U)

F-W

O
O

OH

pak4

R2

+2.5

(2CDZ)

pde2
N

+1.8

(in-house)

R3

pde2
O
N

+1.1

(in-house)

O
N

logarithmic units was found using FreeWilson calculations for

estimating the activity contributions of these R-groups. The only
structural difference between the two is a methyl at position 5 of
the pyridine ring. Although the docking study showed the same
binding mode, the methyl moiety in R-groupB is now buried
into the protein kinase active site and pointing toward a small
lipophilic pocket (F67, V70, K85, V87), explaining the increase
in activity predicted by the FreeWilson model (Fig. 5.4a). A
different combination of R-groups/protein kinase was examined
using the R2 position of Diaminopyrimidine. Figure 5.4b shows
the resulting poses for R-groupA and R-groupB (Table 5.2) when
docked into the PAK4 protein kinase-binding site (PDB entry:
2CDZ). Changing from the carboxy- to the tert-butyloxymoiety forces a different binding orientation of the
R-groups within the active site. The structure-based rationalization for pIC50 difference (pIC50 = 2.5) is the R-groupB
which undergoes a 45 rotation, around the CN single bond
linking the R-group to the Diaminopyrimidine core, allowing the
tert-butyloxy tail to orient in the direction of a buried lipophilic
pocket made by cavity-flanking residues L448, M585, and R586
(Fig. 5.4b).

106

Sciabola et al.

Similar conclusions can be derived when analyzing the coredocking results for the Pyrazolopyrimidine series. The in-house
X-ray structure of PDE2 was used to elucidate the differences
in activity (pIC50 = 1.8) when moving from R-groupA to
R-groupB (Table 5.2) at position R2 of the Pyrazolopyrimidine
core. Figure 5.5a highlights the structural explanation for that,
where the presence of the additional phenyl ring at this site is
not influencing the R-group binding mode, but is extending the
staked hydrophobic interaction toward residue L809. When position R3 of the Pyrazolopyrimidine series was examined, a pIC50
of 1.1 units was obtained by substituting two highly similar
R-groups in the PDE2 biochemical assay (R-groupA and
R-groupB in Table 5.2). Figure 5.5b shows how the variation
in R-group composition determines a different binding mode
for the two R-groups, with the 1,3-dimethoxy benzene portion
of R-groupB now filling a hydrophobic pocket in the active site
made up of a combination of lipophilic (L770, L809, I866, I870)
residues, and optimizing stacked hydrophobic interactions with
the isopropyl moiety of residue L770 (Fig. 5.5b).

3. Notes
1. The FreeWilson approach has proven to be a successful
strategy for the analysis of data sets where large library
collections of compounds obtained through combinatorial
chemistry have been screened against a panel of related proteins or target families, thus boosting the overall quest for
selective inhibitors.
2. A key advantage of the FreeWilson method over standard descriptors-based QSAR techniques is the estimation of activity contribution for individual R-group
structures that are readily interpretable to medicinal
chemists.
3. The possibility to expand the original chemical space of a
given chemical series into a complete virtual library provided us with the identification of compounds with desirable selectivity profiles.
4. The major disadvantage relies on the use of R-groups as
descriptors in model building which gives the models a
well-defined boundary of the chemical space that can be
predicted. It can only explore the chemical space defined by
the R-group combinations present in the training set compounds and cannot be applied, as it is, for predicting the
activity of new compounds with R-groups beyond those
used in the analysis.

Application of FreeWilson Selectivity Analysis

107

5. Data preparation and quality control is a key step in applying FreeWilson methodology to model biological data.
Care must be taken to make sure the underlying data complies with FW additive assumption.
6. Compounds with correlated R-groups and outlier compounds whose R-groups did not occur in other compounds
were removed from the data set as the activity contribution
for these R-groups could not be estimated.
7. In case of sparse structural matrices, these were normally
rearranged into independent blocks where R-groups from
one block would not cross over with other blocks, and statistical analysis was applied to each block separately to estimate the activity contribution for each R-group. Blocks
whose R-group activity contributions could not be estimated due to a lack in R-group crossovers were further
eliminated. The block separation and compound removal
procedure maximized the total number of R-group activity
contributions that could be estimated.
8. LOO cross-validation analysis of FW QSAR models
showed an overall agreement between predicted and experimental pIC50 for each individual combination of chemical
series and protein target.
9. The construction of R-group selectivity profiles based on
in silico R-group contributions allowed us to identify structural determinants for selectivity where a small modification
in the R-groups results in significant difference in selective
profiles.
10. The R-group selectivity knowledge coupled with the availability of X-ray data for many of the kinase/PDE structures
provides substrates for scientists to formulate novel lead
transformation ideas for inhibitor compounds with better
physicochemical properties.

Acknowledgment
This chapter is adapted in part with permission from Simone
Sciabola et al. (2008) J Chem Info Model 48, 18511867.
Copyright 2008 American Chemical Society.
References
1. Manning, G., Whyte, D. B., Martinez, R.,
Hunter, T., Sudarsanam, S. (2002) The
protein kinase complement of the human
genome. Science 298, 19121934.

2. Kostich, M., English, J., Madison, V., et al.

(2002) Human members of the eukaryotic protein kinase family. Genome Biol 3(9),
0043.10043.12.

108

Sciabola et al.

3. Johnson, L. N., Lewis, R. J. (2001) Structural basis for control by phosphorylation.

Chem Rev 101, 22092242.
4. Nagar, B., Bornmann, W. G., Pellicena, P.,
et al. (2002) Crystal structures of the kinase
domain of c-Abl in complex with the small
molecule inhibitors PD173955 and Imatinib
(STI-571). Cancer Res 62, 42364243.
5. George, S. (2007) Sunitinib, a multitargeted
tyrosine kinase inhibitor, in the management
of gastrointestinal stromal tumor. Curr Oncol
Rep 9(4), 323327.
6. Yun, C. -H., Boggon, T. J., Li, Y., et al.
(2007) Structures of lung cancer-derived
EGFR mutants and inhibitor complexes:
mechanism of activation and insights into
differential inhibitor sensitivity. Cancer Cell
11(3), 217227.
7. Stamos, J., Sliwkowski, M. X., Eigenbrot, C.
(2002) Structure of the epidermal growth
factor receptor kinase domain alone and
in complex with a 4-anilinoquinazoline
inhibitor.
J
Biol
Chem
277(48),
4626546272.
8. Fabian, M. A., Biggs, W. H., Treiber, D. K.,
et al. (2005) A small moleculekinase interaction map for clinical kinase inhibitors. Nat
Biotech 23(3), 329336.
9. Beavo, J. A. (1995) Cyclic nucleotide phosphodiesterases: functional implications of
multiple isoforms. Physiologic Rev 75(4),
725748.
10. Soderling, S. H., Beavo, J. A. (2000) Regulation of cAMP and cGMP signaling: new
phosphodiesterases and new functions. Curr
Opin Cell Biol 12(2), 174179.
11. Manallack, D. T., Hughes, R. A., Thompson, P. E. (2005) The next generation
of phosphodiesterase inhibitors: structural
clues to ligand and substrate selectivity of
phosphodiesterases. J Med Chem 48(10),
34493462.
12. Conti, M., Jin, S. L. (1999) The molecular biology of cyclic nucleotide phosphodiesterases. Prog Nucleic Acid Res Mol Biol 63,
138.
13. Mehats, C., Andersen, C. B., Filopanti, M.,
Jin, S. L. C., Conti, M. (2002) Cyclic
nucleotide phosphodiesterases and their role
in endocrine cell signaling. Trends Endocrinol
Metab 13(1), 2935.
14. Perry, M. J., Higgs, G. A. (1998) Chemotherapeutic potential of phosphodiesterase
inhibitors. Curr Opin Chem Biol 2(4),
472481.
15. Torphy, T. (1998) Phosphodiesterase
isozymes: molecular targets for novel antiasthma agents. Am J Respir Crit Care Med
157, 351.

16. Rotella, D. P. (2002) Phosphodiesterase

5 inhibitors: current status and potential
applications. Nat Rev Drug Discov 1(9),
674682.
17. Conti, M., Nemoz, G., Sette, C., Vicini, E.
(1995) Recent progress in understanding the
hormonal regulation of phosphodiesterases.
Endocr Rev 16(3), 370389.
18. Weishaar, R. E., Cain, M. H., Bristol, J. A.
(1985) A new generation of phosphodiesterase inhibitors: multiple molecular forms
of phosphodiesterase and the potential
for drug selectivity. J Med Chem 28(5),
537545.
19. Appleman, M. M., Thompson, W. J. (1971)
Multiple cyclic nucleotide phosphodiesterase
activities from rat brain. Biochemistry 10(2),
311316.
20. Manganiello, V. C., Degerman, E.
(1999) Cyclic nucleotide phosphodiesterases (PDEs): diverse regulators of cyclic
nucleotide signals and inviting molecular targets for novel therapeutic agents. Thromb
Haemostasis 82, 407.
21. Houslay, M. D., Adams, D. R. (2003) PDE4
cAMP phosphodiesterases: modular enzymes
that orchestrate signalling cross-talk,
desensitization and compartmentalization.
Biochem J 370, 1.
22. Corbin, J. D., Francis, S. H. (1999) Cyclic
GMP phosphodiesterase-5: target of sildenafil. J Biol Chem 274, 1372913732.
doi:10.1074/jbc.274.20.13729.
23. Francis, S. H. T. I., Corbin, J. D. (2001)
Cyclic nucleotide phosphodiesterases: relating structure and function. Prog Nucleic Acid
Res Mol Biol 65, 1.
24. Conti, M., Richter, W., Mehats, C., Livera,
G., Park, J. Y. (2003) Cyclic AMP-specific
PDE4 phosphodiesterases as critical components of cyclic AMP signaling. J Biol Chem
278, 5493.
25. Card, A., Caldwell, C., Min, H., et al. (2009)
High-throughput biochemical kinase selectivity assays: panel development and screening applications. J Biomol Screen 14(1),
3142.
26. Free, S. M., Wilson, J. W. (1964) A mathematical contribution to structure-activity
studies. J Med Chem 7(4), 395399.
27. Craig, P. N. (1972) Structure-activity correlations of antimalarial compounds. 1. FreeWilson analysis of 2-phenylquinoline-4carbinols. J Med Chem 15(2), 144149.
28. Nisato, D., Wagnon, J., Callet, G., et al.
(1987) Renin inhibitors. Free-Wilson and
correlation analysis of the inhibitory potency
of a series of pepstatin analogs on plasma
renin. J Med Chem 30(12), 22872291.

Application of FreeWilson Selectivity Analysis

29. Schaad, L. J., Hess, B. A., Purcell, W. P.,
Cammarata, A., Franke, R., Kubinyi, H.
(1981) Compatibility of the Free-Wilson and
Hansch quantitative structure-activity relations. J Med Chem 24(7), 900901.
30. Tomic, S., Nilsson, L., Wade, R. C. (2000)
Nuclear receptor-DNA binding specificity: a
COMBINE and Free-Wilson QSAR analysis.
J Med Chem 43(9), 17801792.
31. Fujita, T., Ban, T. (1971) Structure-activity
study of phenethylamines as substrates of
biosynthetic enzymes of sympathetic transmitters. J Med Chem 14(2), 148152.
32. Hernandez-Gallegos, Z., Lehmann, P. A.
(1990) A Free-Wilson/Fujita-Ban analysis
and prediction of the analgesic potency
of some 3-hydroxy- and 3-methoxy-Nalkylmorphinan-6-one opioids. J Med Chem
33(10), 28132817.
33. Kubinyi, H., Kehrhahn, O. H. (1976) Quantitative structure-activity relationships. 1. The
modified Free-Wilson approach. J Med Chem
19(5), 578586.
34. Kubinyi, H., Kehrhahn, O. H. (1976)
Quantitative structure-activity relationships.
3. A comparison of different Free-Wilson
models. J Med Chem 19(8), 10401049.

109

35. Ekins, S., Gao, F., Johnson, D. L., Kelly,

K. G., Meyer, R. D. inventors (2001) Single point interaction screen to predict IC50
patent EP 1 139 267 A2.26.03.2001.
36. Sciabola, S., Stanton, R. V., Wittkopp, S.,
et al. (2008) Predicting kinase selectivity
profiles using Free-Wilson QSAR analysis.
J Chem Inform Model 48(9), 18511867.
37. Rogers, D., Brown, R. D., Hahn, M. (2005)
Using extended-connectivity fingerprints
with Laplacian-modified Bayesian analysis
in high-throughput screening follow-up.
J Biomol Screeni 10, 682686. First published on September 16, 2005 doi:10.1177/
1087057105281365.
38. Durant, J. L., Leland, B. A., Henry, D.
R., Nourse, J. G. (2002) Reoptimization of
MDL keys for use in drug discovery. J Chem
Inform Comput Sci 42(6), 12731280.
39. Wittkopp, S., Penzotti, J. E., Stanton, R. V.,
Wildman, S. A. (2007) Knowledge-based
docking for kinases with minimal bias. In:
234th ACS National Meeting, Boston, MA,
United States.

Chapter 6
Application of QSAR and Shape Pharmacophore Modeling
Approaches for Targeted Chemical Library Design
Jerry O. Ebalunode, Weifan Zheng, and Alexander Tropsha
Abstract
Optimization of chemical library composition affords more efficient identification of hits from biological
screening experiments. The optimization could be achieved through rational selection of reagents used in
combinatorial library synthesis. However, with a rapid advent of parallel synthesis methods and availability
of millions of compounds synthesized by many vendors, it may be more efficient to design targeted
libraries by means of virtual screening of commercial compound collections. This chapter reviews the
application of advanced cheminformatics approaches such as quantitative structureactivity relationships
(QSAR) and pharmacophore modeling (both ligand and structure based) for virtual screening. Both
approaches rely on empirical SAR data to build models; thus, the emphasis is placed on achieving models
of the highest rigor and external predictive power. We present several examples of successful applications
of both approaches for virtual screening to illustrate their utility. We suggest that the expert use of both
QSAR and pharmacophore models, either independently or in combination, enables users to achieve
targeted libraries enriched with experimentally confirmed hit compounds.
Key words: QSAR modeling, pharmacophore modeling, model validation, virtual screening.

1. Introduction
There is an increased realization that rationally designed chemical libraries facilitate significantly the process of discovering new
drug candidates. The library is described as focused (or targeted)
when compounds selected into the library are optimized with
respect to at least one target property [the property(-ies) can be
specific biological activities and/or various desired parameters of
drug likeness, including drug safety, that are generally covered by
the optimal ADME/Tox paradigm].Naturally, rational design of
J.Z. Zhou (ed.), Chemical Library Design, Methods in Molecular Biology 685,
DOI 10.1007/978-1-60761-931-4_6, Springer Science+Business Media, LLC 2011

111

112

Ebalunode, Zheng, and Tropsha

such libraries is only enabled when sufficient amount of experimental data (e.g., results of biological testing for ligands and/or
target structural information) relevant to the target property(-ies)
is available.
In the early days of combinatorial chemistry, rational design
of chemical libraries frequently implied the selection of building
blocks (from a large available pool) that would produce a reduced
library enriched with potential hit compounds. For instance,
in one of our earlier studies we have developed an approach
termed FOCUS-2D (1, 2) for designing targeted libraries via
rational selection of building blocks. The approach was based
on a virtual combinatorial synthesis procedure where the products were assembled by combining reagents (or building blocks)
into virtual compounds. The building blocks were sampled using
stochastic optimization procedure, and the scoring function optimized in this process was either the similarity of products to
a known active compound(-s) or target activity predicted from
independently developed quantitative structureactivity relationship (QSAR) models. The virtual library of high scoring (i.e.,
predicted to be active) compounds was assembled and analyzed
in terms of building blocks found with the highest frequency
within selected compounds; thus, the ultimate goal of the study
was the rational selection of building blocks that would be used to
build a complete chemical library (as opposed to cherry-picking
selected compounds).
Although studies into rational building block selection such
as those described above were popular in the early days of computational combinatorial chemistry, the alternative approaches
looking into rational selection of compounds from commercial libraries of already synthesized or synthetically feasible compounds have gradually prevailed. In fact, in a popular review
Jamois (3) has compared reagent-based vs. product-based strategies for library design and concluded that several studies have
demonstrated the superiority of product-based designs in yielding diverse and representative subsets. Nowadays, large commercial libraries and services that provide integrated links to commercially available compounds are widely available (for instance,
ca. 10 M compounds have been compiled in publicly available ZINC database (4)); see a recent review (5) for a partial
list of additional chemical databases. Thus, most of the current approaches employ various virtual screening strategies to
select specific compound subsets for subsequent experimental
exploration.
This chapter discusses the application of popular cheminformatics approaches, such as rigorously built QSAR models and
shape pharmacophore models, to the problem of targeted library
design. QSAR models offer unique ability to rationalize existing experimental SAR data in the form of robust quantitative

Application of QSAR and Shape Pharmacophore Modeling Approaches

113

models that predict target property directly from structural chemical descriptors; thus, they can be used to screen an external chemical library to select compounds predicted to be active against
the target. Conversely, shape pharmacophore models utilize the
representative shape of active ligands or the negative image (or
pseudo molecule) extracted from the binding site of the target
protein to query 3D conformational databases of virtual or real
molecular libraries. With enough attention paid to critical issues of
model validation and applicability domain definition, both QSAR
and shape pharmacophore models could be used successfully (and
concurrently) to mine external virtual libraries to identify putative compounds with the desired target properties. The selected
compounds could be chosen as candidates for thereby rationally
designed compound library.
This chapter will initially discuss current algorithms for developing externally predictive QSAR models and present experimentally confirmed examples of identifying novel bioactive compounds by the means of QSAR model-based virtual screening. It
will also present a novel shape pharmacophore modeling method
and its validation through retrospective analysis of known biologically active compounds. Of course, many approaches, both
structure based and ligand based, have been used for virtual
screening.We have decided to focus on these specific methodologies, i.e., QSAR and pharmacophore modeling because both
approaches are well known to both computational and medicinal chemists as structure optimization tools used at later stages
of drug discovery after the lead compounds have been identified experimentally. However, in recent years these approaches,
among other cheminformatics methods (6), have found new
applications as virtual screening tools. The methods and applications discussed in this chapter should be of interest to both
computational and synthetic chemists and experimental biologists
working in the areas of biological screening of chemical libraries.

2. Predictive
QSAR Models as
Virtual Screening
Tools

QSAR modeling has been traditionally viewed as an evaluative

approach, i.e., with the focus on developing retrospective and
explanatory models of existing data. Model extrapolation has
been considered only in hypothetical sense in terms of potential modifications of known biologically active chemicals that
could improve compounds activity. Nevertheless recent studies
suggest that current QSAR methodologies may afford robust
and validated models capable of accurate prediction of compound properties for molecules not included in the training sets.

114

Ebalunode, Zheng, and Tropsha

Below, we discuss a data analytical modeling workflow developed

in our laboratory that incorporates modules for combinatorial
QSAR model development (i.e., using all possible binary combinations of available descriptor sets and statistical data modeling techniques), rigorous model validation, and virtual screening of available chemical databases to identify novel biologically
active compounds. Our approach places particular emphasis on
model validation as well as the need to define model applicability
domains in the chemistry space. We present examples of studies
where the application of rigorously validated QSAR models to virtual screening identified computational hits that were confirmed
by subsequent experimental investigations. This approach enables
to identify subsets of putative active compounds that form a targeted chemical library expected to be enriched with target-specific
bioactive compounds.
2.1. Basic QSAR
Modeling Concepts

Any QSAR method can be generally defined as an application

of mathematical and statistical methods to the problem of finding empirical relationships (QSAR models) of the form Pi =
1 , D2 , . . . , Dn ), where Pi are biological activities (or other
k(D
properties of interest) of molecules, D1 , D2 ,. . .,Dn are calculated
(or, sometimes, experimentally measured) structural properties
(molecular descriptors) of compounds, and k is some empirically
established mathematical transformation that should be applied
to descriptors to calculate the property values for all molecules
(Fig. 6.1). The goal of QSAR modeling is to establish a trend
in the descriptor values, which parallels the trend in biological
activity. In essence, all QSAR approaches imply, directly or indi-

Fig. 6.1. General framework of QSAR modeling.

Application of QSAR and Shape Pharmacophore Modeling Approaches

115

rectly, a simple similarity principle, which for a long time has

provided a foundation for the experimental medicinal chemistry:
compounds with similar structures are expected to have similar
biological activities. The detailed description of major tenets of
QSAR modeling is beyond the scope of this chapter; the overview
of popular QSAR modeling techniques could be found in multiple reviews, e.g., (7). Here, we comment on most critical general
aspects of model development and, most importantly, validation
that are especially important in the context of using QSAR models
for virtual screening.
2.1.1. Critical
Importance of Model
Validation

In our important paper titled Beware of q2 ! (8), we have

demonstrated the insufficiency of the training set statistics for
developing externally predictive QSAR models and formulated
the main principles of model validation. Despite earlier observations and warnings of several authors (911) that high crossvalidated correlation coefficient R2 (q2 ) is a necessary but insufficient condition for the model to have high predictive power,
many studies continue to consider q2 as the only parameter characterizing the predictive power of QSAR models. In reference (8)
we have shown that the predictive power of QSAR models can be
claimed only if the model was successfully applied for prediction
of the external test set compounds, which were not used in the
model development. We have demonstrated that the majority of
the models with high q2 values have poor predictive power when
applied for prediction of compounds in the external test set. In
the subsequent publication (12) the importance of rigorous validation was again emphasized as a crucial, integral component of
model development. Several examples of published QSAR models
with high fitted accuracy for the training sets, which failed rigorous validation tests, have been considered. We presented a set of
simple guidelines for developing validated and predictive QSAR
models and discussed several validation strategies such as the randomization of the response variable (Y-randomization) and external validation using rational division of a data set into training
and test sets. We highlighted the need to establish the domain
of model applicability in the chemical space to flag molecules for
which predictions may be unreliable, and discussed some algorithms that can be used for this purpose. We advocated the broad
use of these guidelines in the development of predictive QSPR
models (1214).
At the 37th Joint Meeting of Chemicals Committee and
Working Party on Chemicals, Pesticides and Biotechnology, held
in Paris on 1719 November 2004, the OECD (Organization for
Economic Co-operation and Development) member countries
adopted the following five principles that valid (Q)SAR models
should follow to allow their use in regulatory assessment of chemical safety: (i) a defined endpoint; (ii) an unambiguous algorithm;

116

Ebalunode, Zheng, and Tropsha

(iii) a defined domain of applicability; (iv) appropriate measures of

goodness-of-fit, robustness, and predictivity; and (v) a mechanistic interpretation, if possible. Since then, most of the European
authors publishing in QSAR area include a statement that their
models fully comply with OECD principles (e.g., see (1518)).
Validation of QSAR models is one of the most critical problems of QSAR. Recently, we have extended our requirements for
the validation of multiple QSAR models selected by acceptable
statistics criteria of prediction for the test set (19). Additional
studies on this critical component of QSAR modeling should
establish reliable and commonly accepted good practices for
model development, which should make models increasingly useful for virtual screening.
2.1.2. Applicability
Domains and QSAR
Model Acceptability
Criteria

One of the most important problems in QSAR analysis is establishing the domain of applicability for each model. In the absence
of the applicability domain restriction, each model can formally
predict the activity of any compound, even with a completely different structure from those included in the training set. Thus, the
absence of the model applicability domain as a mandatory component of any QSAR model would lead to the unjustified extrapolation of the model in the chemistry space and, as a result, a
high likelihood of inaccurate predictions. In our research we have
always paid particular attention to this issue (12, 2027). A good
overview of commonly used applicability domain definitions can
be found in reference (28).
In our earlier publications (8, 12) we have recommended
a set of statistical criteria which must be satisfied by a predictive model. For continuous QSAR, criteria that we will follow
in developing activity/property predictors are as follows: (i) correlation coefficient R between the predicted and the observed
activities; (ii) coefficients of determination (29) (predicted versus
observed activities R02 and observed versus predicted activities R02
for regressions through the origin); (iii) slopes k and k of regression lines through the origin. We consider a QSAR model predictive if the following conditions are satisfied (i) q2 >0.5; (ii) R2 >0.6;
(R2 R2 )

(R2 R2 )

0
0
< 0.1 and 0.85 k 1.15 or
< 0.1 and
(iii)
R2
R2


2
2

2


0.85 k 1.15; (iv) R0 R0 < 0.3 where q is the crossvalidated correlation coefficient calculated for the training set, but
all other criteria are calculated for the test set (for additional discussion, see (30)).

2.1.3. Predictive QSAR

Modeling Workflow

Our experience in QSAR model development and validation has

led us to establishing a complex strategy that is summarized
in Fig. 6.2. It describes the predictive QSAR modeling workflow, which focuses on delivering validated models and ultimately,
computational hits confirmed by the experimental validation. We

Application of QSAR and Shape Pharmacophore Modeling Approaches

117

Fig. 6.2. General workflow for predictive QSAR modeling.

start by randomly selecting a fraction of compounds (typically,

1015%) as an external validation set. The remaining compounds
are then divided rationally (e.g., using the Sphere Exclusion protocol implemented in our laboratory (14)) into multiple training
and test sets that are used for model development and validation, respectively using criteria discussed in more detail below.
We employ multiple QSAR techniques based on the combinatorial exploration of all possible pairs of descriptor sets coupled with
various statistical data mining techniques (termed combi-QSAR)
and select models characterized by high accuracy in predicting
both training and test sets data. Validated models are finally tested
using the evaluation set. The critical step of the external validation
is the use of applicability domains. If external validation demonstrates the significant predictive power of the models we use all
such models for virtual screening of available chemical databases
(e.g., ZINC (4)) to identify putative active compounds and work
with collaborators who could validate such hits experimentally.
The entire approach is described in detail in several recent papers
and reviews (e.g., (7, 12, 30, 31))
2.2. Application of
QSAR Models to
Virtual Screening

In our recent studies we were fortunate to recruit experimental collaborators who have validated computational hits identified
by virtual screening of commercially available compound libraries
using rigorously validated QSAR models. Examples include anticonvulsants (25), HIV-1 reverse transcriptase inhibitors (32),
D1 antagonists (33), antitumor compounds (34), beta-lactamase
inhibitors (35), human histone deacetylase (HDAC) inhibitors

118

Ebalunode, Zheng, and Tropsha

(36), and geranylgeranyltransferase-I inhibitors (37). Thus, models resulting from predictive QSAR workflow could be used to
prioritize the selection of chemicals for the experimental validation. To illustrate the power of validated QSAR models as virtual screening tools, we shall discuss the examples of studies that
resulted in experimentally confirmed hits. We note that such studies could only be done if there are sufficient data available for
a series of tested compounds such that robust validated models
could be developed using the workflow described in Fig. 6.2.
The following examples illustrate the use of QSAR models developed with predictive QSAR modeling and validation workflow
(Fig. 6.2) for virtual screening of commercial libraries to identify
experimentally confirmed hits.
2.2.1. Discovery of
Novel Anticancer Agents

A combined approach of validated QSAR modeling and virtual screening was successfully applied to the discovery of novel
tylophorine derivatives as anticancer agents (34). QSAR models have been initially developed for 52 chemically diverse
phenanthrine-based tylophorine derivatives (PBTs) with known
experimental EC50 using chemical topological descriptors (calculated with the MolConnZ program) and variable selection
knearest neighbor (kNN) method. Several validation protocols
have been applied to achieve robust QSAR models. The original data set was divided into multiple training and test sets, and
the models were considered acceptable only if the leave-one-out
cross-validated R2 (q2 ) values were greater than 0.5 for the training sets and the correlation coefficient R2 values were greater
than 0.6 for the test sets. Furthermore, the q2 values for the
actual data set were shown to be significantly higher than those
obtained for the same data set with randomized target properties
(Y-randomization test), indicating that models were statistically
significant. Ten best models were then employed to mine a commercially available ChemDiv Database (ca. 500K compounds)
resulting in 34 consensus hits with moderate to high predicted
activities. Ten structurally diverse hits were experimentally tested
and eight were confirmed active with the highest experimental
EC50 of 1.8 M implying an exceptionally high hit rate (80%).
The same 10 models were further applied to predict EC50 for
four new PBTs, and the correlation coefficient (R2 )between the
experimental and the predicted EC50 for these compounds plus
eight active consensus hits was shown to be as high as 0.57.

2.2.2. Discovery of
Novel Histone
Deacetylase (HDAC)
Inhibitors

Histone deacetylases (HDACs) play a critical role in transcription regulation. Small molecule HDAC inhibitors have become
an emerging target for the treatment of cancer and other cell proliferation diseases. We have employed variable selection k nearest
neighbor approach (kNN)and support vector machines (SVM)
approach to generate QSAR models for 59 chemically diverse

Application of QSAR and Shape Pharmacophore Modeling Approaches

119

compounds with inhibition activity on class I HDAC. MOE (38)and MolConnZ (39)-based 2D descriptors were combined with
knearest neighbor (kNN) and support vector machines (SVM)
approaches independently to improve the predictive power of
models. Rigorous model validation approaches were employed
including randomization of target activity (Y-randomization test)
and assessment of model predictability by consensus prediction
on two external data sets. Highly predictive QSAR models were
generated with leave-one-out cross-validation R2 (q2 ) values for
the training set and R2 values for the test set as high as 0.81 and
0.80, respectively, with MolconnZ/kNN approach and 0.94 and
0.81, respectiveley, with MolconnZ/SVM approach. Validated
QSAR models were then used to mine four chemical databases:
National Cancer Institute (NCI) database, Maybridge database,
ChemDiv database, and one ZINC database, including a total of
over 3 million compounds. The searches resulted in 48 consensus hits, including two reported HDAC inhibitors that were not
included in the original data set. Four hits with novel structural
features were purchased and tested using the same biological assay
that was employed to assess the inhibition activity of the training
set compounds. Three of these four compounds were confirmed
active with the best inhibitory activity (IC50 ) of 1 M. The overall
workflow for model development, validation, and virtual screening is illustrated in Fig. 6.3.
2.2.3. Discovery of
Novel Histone
Deacetylase (HDAC)
Inhibitors

In another recent study (37), we employed our standard QSAR

modeling workflow (Fig. 6.2) to discover novel geranylgeranyltransferase type I (GGTase-I) inhibitors. Geranylgeranylation is
critical to the function of several proteins including Rho, Rap1,
Rac, Cdc42, and G protein gamma subunits. GGTase-I inhibitors

Fig. 6.3. Application of predictive QSAR workflow including virtual screening to discover
novel HDAC inhibitors.

120

Ebalunode, Zheng, and Tropsha

(GGTIs) have therapeutic potential to treat inflammation, multiple sclerosis, atherosclerosis, and many other diseases. Following
our standard QSAR modeling workflow, we have developed and
rigorously validated models for 48 GGTIs using variable selectionk nearest neighbor (40) and automated lazy learning (26)
and genetic algorithm-partial least square (41) QSAR methods.
The QSAR models were employed for virtual screening of 9.5
million commercially available chemicals yielding 47 diverse computational hits. Seven of these compounds with novel scaffolds
and high predicted GGTase-I inhibitory activities were tested in
vitro, and all were found to be bona fide and selective micromolar
inhibitors.
Figure 6.4 shows the structures of both representative training set compounds and confirmed computational hits. We should
emphasize that QSAR models have been traditionally viewed
as lead optimization tools capable of predicting compounds
with chemical structure similar to the structure of molecules
used for the training set. However, this study clearly indicates
(Fig. 6.4) that with enough attention given to the model development process and using chemical descriptors characterizing
whole molecules (as opposed to, e.g., chemical fragments), it is
indeed possible to discover compounds with novel chemical scaffolds. Furthermore, in our study we have additionally demonstrated that these novel hits could not be identified using tradi-

Training Set Scaffolds

Peptidomimetics

Major Hits with Novel Scaffolds

Sigma: IC50 = 8 M

Asinex: IC50 = 35 M
Pyrazoles
Mean IC50

5 M

Enamine: IC50 = 43 M
Two similar hits

Fig. 6.4. Discovery of GGTase-I inhibitors with novel chemical scaffolds using a combination of QSAR modeling and
virtual screening.

Application of QSAR and Shape Pharmacophore Modeling Approaches

121

tional chemical similarity search (37), which highlights the power

of robust QSAR models as the drug discovery tool.
In summary, our studies have established that QSAR models could be used successfully as virtual screening tools to discover compounds with the desired biological activity in chemical
databases or virtual libraries (25, 31, 33, 34, 42). It should be
stressed that the total number of compounds selected for virtual
screening based on QSAR model predictions is typically relatively
small, only a few dozen. Obviously, the total number of computational hits is controlled by the value of applicability domain.
In most published cases, because we were limited in both time
and resources, we chose a very conservative applicability domain
leading to the selection of a small library of computational hits
with an expectation that a large fraction of these would be confirmed as active compounds. In the industrial size projects it may
be more reasonable to loosen the applicability domain requirement and increase the size of virtual hit library. One may expect
that the increase in the library size will result in lower relative
accuracy of prediction but the absolute number of confirmed hits
may actually increase. Thus, scientists using QSAR models that
incorporate the applicability domain should always be aware of
the interplay between the size of the domain, the coverage of
the virtual screening library, and the prediction accuracy so they
should use the applicability domain as a tunable parameter to control this interplay. The discovery of novel bioactive chemical entities is the primary goal of computational drug discovery, and the
development of validated and predictive QSAR models is critical
to achieve this goal.

3. Shape
Pharmacophore
Modeling as
Virtual Screening
Tool

Shape complementarity plays an important role in the process of

molecular recognition (43). In a typical 3D structure of ligand
receptor complex, one can observe tight van der Waals contacts
between the ligand atoms and the receptor atoms of the binding pocket. Grant et al. (44) pointed out the fundamental reasons for such shape complementarity. They argued that the intermolecular interactions that stabilize the receptorligand complex
are enthalpically weak, and they become effective only when the
chemical groups involved can approach each other closely, which
is favored by the shape complementarity. They further argued
that the entropic contributions advantageous to binding, which
involve the loss of bound water of both the host and the guest,
are also favored by shape complementarity. Thus, the concept of
shape complementarity is widely adopted by medicinal chemists

122

Ebalunode, Zheng, and Tropsha

in structure-based drug design. When the constraints of critical functional groups and their spatial orientation are taken into
account, together with shape complementarity, one can create a
shape pharmacophore model. This latter model has proved to be
more effective in virtual screening experiments. In the following
sections, we first describe the basic concept of shape and shape
pharmacophore modeling and then present some recent literature
examples.
3.1. Basic Concept of
Molecular Shape
Analysis

Molecular shape analysis tools can be broadly categorized into

two groups. In terms of the methodology employed, they are
either superposition based or superposition free. The former calculates a shape-matching measure only after an optimal superposition of the two objects has been obtained. The second category
of methods calculates shape similarity score based on rotationand-translation independent descriptors that are computed from
different representations of molecular objects, and thus, it does
not depend on the orientation or alignment of the two molecular
objects. Zauhars shape signatures (45) and the more recent USR
method (42, 43) belong to this category.
The following two categories of methods can be identified, in
terms of the input information for shape analysis tools: (1) ligandbased analysis, where receptors structural information is not
included in the analysis and (2) receptor-based methods, where
the structural information of the receptor is an integral part of
the analysis process and is essential in formulating the models.

3.1.1. Alignment-Based
and Alignment-Free
Methods
3.1.1.1.
Alignment-Based
Algorithms

In alignment-based algorithms, the shape similarity calculation

is conducted after an optimal superposition of two molecular
objects is achieved. One of the earliest methods, studied by Meyer
and Richards (46), performed the alignment and then counted
common points between the two objects as a way to quantify the similarity between two molecular objects. The optimization process was slow, which limited its use. The shape similarity concept was further developed by Good and Richards, by
employing Gaussian functions as the basis for similarity calculation (47). Grant et al. also employed Gaussian functions to calculate shape similarity (44), based on the calculation of volume
overlap between two superposed molecular objects. This latter
method has further been modified and implemented in the program ROCS (Rapid Overlay of Compound Structures) (48) and
the OE Shape toolkit (49).
Gaussian Shape Similarity by Good and Richards. This method
introduced the use of Gaussians for molecular shape matching for

Application of QSAR and Shape Pharmacophore Modeling Approaches

123

the first time (47). The shape of each atom was described as a
suitable electron density function, and then three Gaussian functions were fitted to each of the atomic electron density functions.
An analytical shape similarity index was formulated according to
the Carbo index (50, 51). Molecular superposition was achieved
via the optimization of the similarity index.
Shape-Matching Method by Grant and Pickup (40, 51). This
method defines a Gaussian density for each atom to replace the
hard sphere representation of atoms (52). The molecular volume
is expressed as a series of integration terms, representing the intersection volumes between the atoms in a molecule. The Gaussian
description was used to compare the shapes of two molecules by
optimizing their volume overlap using analytical derivatives with
respect to rotations and translations. This idea was later implemented in the ROCS program. However, it has been pointed
out (53) that ROCS, by default, gives the same radius value to
all heavy atoms in the molecule. This approximation led to the
conclusion that the volume calculation in ROCS might not be
as accurate as expected from the original theory of Grant and
Pickup. Nonetheless, ROCS has been shown to be very successful in many validation studies and actual applications.
3.1.1.2. Alignment-Free
Algorithms

The basic idea of alignment-free shape matching is that a set of

rotation- and translation-free descriptors are calculated for conformers under consideration, and then some similarity measure is
devised to quantify the similarity between two molecular objects.
Zauhars shape signatures (45), Brenemans PEST and PESD
methods (5456), the USR (ultrafast shape recognition) method
(53), the atom triplets method (57), and Schlossers recent TrixX
BMI approach (58) are a few examples. One advantage of these
algorithms is that they offer much faster computational speed and,
thus, are suitable for screening large molecular databases and virtual compound libraries.
The Shape Signatures Method. This method was reported by
Zauhar et al. for shape description and comparison (45). Solventaccessible molecular surface is triangulated using the smooth
molecular surface triangulator algorithm (59) (SMART). The
molecular surface is divided into regular triangular area elements.
The volume defined by the molecular surface is explored using ray
tracing, which starts each ray from a randomly selected point on
the molecular surface and then allows the ray to propagate by the
rules of optical reflection. The tracing and reflection of light stop
until some preset conditions are met. The result is a collection of
line segments that connect two successive reflection points. The
simplest shape signature is the distribution of the lengths of these
segments, stored as histogram for each molecule. The similarity
between molecular shapes is simply the similarity between their
histograms.

124

Ebalunode, Zheng, and Tropsha

The PEST and PESD Method These methods were developed by Brenemans group. The PEST (property-encoded surface translator) method is based on the combination of the
TAE descriptors (54) and the shape signatures idea by Zauhar
(45). It uses the TAE molecular surface representations to define
property-encoded boundaries. It first computes the molecular surface property distributions and then collects ray-tracing
path information and lastly generates the shape descriptors. The
2D histograms are generated to represent surface shape profile, encoding both shape and surface properties. Similarly, the
property-encoded shape distributions (PESD) descriptors have
recently been reported and employed to study ligandprotein
binding affinities (56). The PESD algorithm is different from
PEST in that it is based on a fixed number of randomly sampled point pairs on the molecular surface that does not require
ray tracing. Both PEST and PESD descriptors should account for
the distribution of both the polar and non-polar regions and electrostatic potential on the molecular surface.
The USR (Ultrafast Shape Recognition) Method. This method
was reported by Ballester and Richards (53) for compound
database search on the basis of molecular shape similarity. It was
reportedly capable of screening billions of compounds for similar shapes on a single computer. The method is based on the
notion that the relative position of the atoms in a molecule is
completely determined by inter-atomic distances. Instead of using
all inter-atomic distances, USR uses a subset of distances, reducing the computational costs. Specifically, the distances between
all atoms of a molecule to each of four strategic points are calculated. Each set of distances forms a distribution, and the three
moments (mean, variance, and skewness) of the four distributions
are calculated. Thus, for each molecule, 12 USR descriptors are
calculated. The inverse of the translated and scaled Manhattan
distance between two shape descriptors is used to measure the
similarity between the two molecules. A value of 1 corresponds
to maximum similarity and a value of 0 corresponds to minimum similarity.
3.2. Examples of
Application of Shape
and Pharmacophore
Models for Virtual
Screening
3.2.1. Ligand-Based
Studies

When a few ligands are known for a particular target, one can
use ligand-based shape-matching technology to search for potential ligands via virtual screening. A ligand-based application of
shape-matching methods starts with a ligand with known biological activity. Its 3D conformations are often pregenerated by a

Application of QSAR and Shape Pharmacophore Modeling Approaches

125

conformer generator. Also, a multiconformer database of potential drug molecules is pregenerated to be used by the shapematching program. For alignment-based methods, the conformers of the known ligand (i.e., the query) will be directly aligned
with those of the database molecules. Molecules that align well
with the query molecule will be selected for further consideration.
In the case of alignment-free methods, both the shape descriptors
of the query and those of the database molecules are first calculated, and a similarity value is calculated between the query and
each of the database molecules. Molecules with better similarity
values to the query are selected for further consideration.
In the validation study by Hawkins et al. (60), the shapematching method ROCS was compared to 7 well-known docking
tools, in terms of their abilities to recover known ligands for 21
different protein targets. The comparative study showed that the
3D shape method (ROCS) performed at least the same as, and
often better than, the docking tools studied. Their work indicated
that shape-based virtual screening method could be both efficient
(in terms of the computational speed) and effective (in terms of
hit enrichment) in virtual screening projects.
In a comparative study, McGaughey et al. (61) investigated
several 2D similarity methods (including Daylight fingerprint similarity (62) and TOPOSIM (63)), 3D shape similarity methods
(ROCS and SQ (64)), and several known docking tools (FLOG
(65), FRED (66), and Glide (67, 68)). Based on the performance on a benchmark set of 11 protein targets, they observed
that, on average, the ligand-based shape method with chemistry
constraints outperformed more sophisticated docking tools. Their
results also demonstrated that shape matching (including chemistry constraints) could select more diverse active compounds
than 2D similarity methods. This indicates that shape-matching
tools may offer a better scaffold hopping capability than 2D
methods.
Moffat et al. (69) also compared three ligand-based shape
similarity methods, including CatShape (70), FBSS (71), and
ROCS. These methods have been compared on the basis of
retrospective virtual screening experiments. All three methods
have demonstrated significant enrichment, but ROCS with CFF
option (CFF: chemical force field) gave the best performance.
They reported that shape matching, coupled with chemistry constraints, afforded better enrichment factors than shape-matching
alone, indicating the importance of including chemistry information in the search. This observation is consistent with the
recent validation study by Hawkins et al. (60) and by Ebalunode et al. (72). In general, flexible methods gave slightly better performance than the respective rigid search methods; however, the increased performance could not justify the increased

126

Ebalunode, Zheng, and Tropsha

computational cost. This observation is again consistent with the

finding by a different validation study by Ebalunode et al. (73).
Zauhar et al. reported an interesting application of the shape
signatures approach to shape matching and similarity search (45).
In a validation study (74), they found significant enrichment
of ligands for the serotonin receptor using the shape signatures
approach. A set of 825 agonists and 400 antagonists as well
as roughly 10,000 randomly chosen compounds from the NCI
database were used in that study.
Ballester et al. (75) evaluated a new algorithm (Ultrafast
Shape Recognition or USR) in the context of retrospective
ligand-based virtual screening. They showed that USR performed
better, on average, than a commercially available shape similarity method, while screening conformers at a rate that is >2500
times faster. This feature makes USR an ideal virtual screening
tool for searching extremely large molecular databases. However,
no atomic property information is encoded in this method.
When ROCS or any other ligand-based 3D shape-matching
method is used for virtual screening, the choice of the query conformation can have significant impact on the results of virtual
screening. This is especially true when no X-ray structure of a
bound ligand is available. In a recent study by Tawa et al. (76),
the authors developed a rational conformation selection protocol (named CORAL), which allows the selection of conformation that affords better enrichment than using simply the lowest energy conformation as the query. They have demonstrated
that this method can significantly improve the effectiveness of
ligand-based method (ROCS) for drug discovery. In a related
study, Kirchmair et al. (77) described ways to optimize shapebased virtual screening. They discussed how to choose the right
query together with chemical information. They have examined
various parameters that may improve the performance and offered
guidelines on how to achieve the optimum performance using
shape-matching techniques in virtual screening.
3.2.2. Receptor-Based
Studies

Various variants of the basic shape-matching algorithms have been

reported in the literature (69, 78). The general idea of these
tools is to extract the shape and pharmacophore information
from the binding site structure and represent such information
or constraints as pseudo-molecular shapes. Once the pseudomolecular shape is created, a regular shape-matching algorithm
can be employed to compare binding sites with small molecules.
Here, we review a few of the recent developments as follows.
To employ the shape-matching algorithm in a receptor-based
fashion, Ebalunode et al. developed a method that can be considered as a structure-based variant of ROCS. The method, SHAPE4
(72), utilizes a computational geometry method (the alpha-shape
algorithm) to extract and characterize the binding site of a given

Application of QSAR and Shape Pharmacophore Modeling Approaches

127

target. It then uses a grid to approximate the geometric volume

of the binding site, defined by the Delaunay simplices generated
from the alpha-shape analysis. The pharmacophore centers are
derived from the binding site atomic information, using either the
LigandScout program (79) or other equivalent approaches. As a
result, the extracted binding site shape and the pharmacophore
constraints reflect the nature of the binding site. In theory, this
approach can overcome the limit imposed by using the bound
ligand per se as the query, in that the query in SHAPE4 can cover
more diverse characteristics of the binding site than the bound
ligand itself. The effectiveness, in terms of enrichment factors
and diversity of the hits, has been demonstrated in the SHAPE4
article (72).
Similar to SHAPE4, Lee et al. developed the SLIM program
(80), another variant of the ROCS technology. It derives the
binding site shape and pharmacophore information based on the
X-ray structure of the target. It is different from SHAPE4 in that
a more straightforward method for extracting the binding site is
employed by SLIM, where a geometric box is defined based on
the knowledge of the binding site. Visualization by human expert
is often needed to help define the binding pocket, and it is harder
to use in cases where large number of protein targets are being
studied. However, as pointed out by the authors, their focus was
to test the effect and impact of multiple conformations of the target protein in order to address the conformational flexibility issue.
Thus, SLIM worked very well for their purpose.
3.3. Prospective
Applications of
Pharmacophore
Shape Technologies

Markt et al. (81) reported the discovery of PPAR ligands using

an integrated screening protocol. Using a combination of pharmacophore, 3D shape similarity, and electrostatic similarity, they
discovered 10 virtual screening hits, of which 5 tested positive
against the ligand-binding domain (LBD) of human PPAR in
transactivation assays and showed affinities for PPAR in a competitive binding assay. Therefore, this represents a successful application of multiple complementary technologies in drug discovery,
where the 3D shape technology was part of the workflow.
An application of the ROCS program has been reported
recently (82). New scaffolds for small molecule inhibitors of
the ZipA-FtsZ proteinprotein interaction have been found. The
shape comparisons are made relative to the bioactive conformation of a HTS hit, determined by X-ray crystallography. A followup X-ray crystallographic analysis also showed that ROCS accurately predicted the binding mode of the inhibitor. This result
offers the first experimental evidence that validates the use of
ROCS for scaffold hopping purposes.
Another successful application of a shape similarity method
was reported by Cramer et al. (83). Over 400 compounds were
synthesized and tested for their inhibition of angiotensin II. The

128

Ebalunode, Zheng, and Tropsha

63 compounds that were identified by topomer shape similarity

as most similar to one of the four query structures covered all the
compounds found to be highly active. None of the remaining 362
structures were highly active. Thus, this report is a nice demonstration of the ability of a shape similarity method for discovering new biologically active compounds. In another study, Cramer
et al. (84) reported the application of topomer shape similarity
for lead hopping. The hit rate averaged over all assays was 39%.
The average 2D fingerprint Tanimoto similarity between a query
and the newly found structures was 0.36, similar to the Tanimoto
similarity between random drug-like structures. Thus, this is a
good indication of the lead hopping ability of the topomer shape
method.
A successful application of the shape and electrostatic similarity methods to prospective drug discovery has been reported by
Muchmore et al. (85). To identify novel melanin-concentrating
hormone receptor 1 (MCHR1) antagonists, a library of virtual
molecules was designed. Over 3 million molecules were searched
using 3D shape similarity methods (in conjunction with an electrostatic similarity-matching algorithm). One of the top scoring hits was made and tested for MCHR1 activity. A threefold
improvement in binding affinity and cellular potency has been
achieved compared to the parent ligand. This example demonstrated the power of the ligand-based shape method for the discovery of new compounds from a large virtual library for targets
without crystallographic information.
In a study that combined a variety of ligand-based and
structure-based methods, Perez-Nueno et al. reported the success of a prospective virtual screening project (86). They first
established a screening protocol based on a retrospective virtual
screening, using a database of CXCR4 inhibitors and inactive
compounds compiled from the literature. A large virtual combinatorial library of molecules was designed. The virtual screening
protocol has been employed to select five compounds for synthesis and testing. Experimental binding assays of those compounds
confirmed that their mode of action was blocking the CXCR4
receptor. This represents another successful example of using a
shape similarity method for the discovery of new compounds via
virtual screening.
In a more recent virtual screening study, Ballester et al.
reported the successful identification of novel inhibitors of arylamine N-acetyltransferases using the USR algorithm (87). A
computational screening of 700 million molecular conformers
was conducted very efficiently. A small number of the predicted
hits were purchased and experimentally tested. An impressive hit
rate of 40% has been achieved. The authors also showed the
ability of USR to find biologically active compounds with different chemical structures (i.e., scaffold hopping), evidenced by

Application of QSAR and Shape Pharmacophore Modeling Approaches

129

low Tanimoto coefficients between the found hits and the query
molecule. Visual inspection also confirmed that none of the nine
actives found shared a common scaffold with the template. Thus,
this example demonstrated the power of a pure shape similarity
method for scaffold hopping projects.
Finally, Ebalunode et al. reported a structure-based shape
pharmacophore modeling for the discovery of novel anesthetic
compounds (88). The 3D structure of apoferritin, a surrogate
target for GABAA , was used as the basis for the development of
several shape pharmacophore models. They demonstrated that
(1) the method effectively recovered known anesthetic agents
from a diverse database of compounds; (2) the shape pharmacophore scores had a significant linear correlation with the measured binding data of several anesthetic compounds, without
prior calibration and fitting; and (3) the computed scores also
correctly predicted the trend of the EC50 values of a set of
anesthetics.

4. Summary
and Conclusions
We have discussed the application of cheminformatics approaches
such as QSAR and shape pharmacophore modeling to the problem of targeted library design by means of virtual screening. Both
approaches offer unique abilities to rationalize existing experimental SAR data in the form of models that could identify novel
compounds predicted to interact with the specific target. Pharmacophore models achieve this task by establishing that a compound contains specific chemical features characteristic of known
bioactive compounds, whereas QSAR models have the ability to
predict the target activity quantitatively from structural chemical
descriptors of compounds. As with any computational molecular
modeling approach, it is imperative that both QSAR and pharmacophore modeling approaches are used expertly. Therefore, this
chapter has focused on the discussion of critical components of
both approaches that should be studied and executed rigorously
to enable their successful application. We have shown that with
enough attention paid to critical issues of model validation and
(in the case of QSAR modeling) applicability domain definition,
the models could be indeed used successfully to mine external
virtual libraries, especially of commercially available chemicals, to
create targeted compound libraries with desired properties. The
methods and applications discussed in this chapter should be of
help to both computational and synthetic chemists and experimental biologists working in the areas of biological screening of
chemical libraries.

130

Ebalunode, Zheng, and Tropsha

Acknowledgments
AT acknowledges the support from NIH (grant R01GM066940).
J.E. and W.Z. acknowledge the financial support by the Golden
Leaf Foundation via the BRITE Institute, North Carolina Central University. W.Z. also acknowledges funding from NIH (grant
SC3GM086265).
References
1. Zheng, W., Cho, S. J., Tropsha, A. (1998)
Rational combinatorial library design. 1.
Focus-2D: a new approach to the design of
targeted combinatorial chemical libraries. J
Chem Inf Comput Sci 38, 251258.
2. Cho, S. J., Zheng, W., Tropsha, A. (1998)
Focus-2D: a new approach to the design of
targeted combinatorial chemical libraries, in
(Altman, R. B., Dunker, A. K., Hunter, L.,
Klein, T. E. eds.) Pacific Symposium on Biocomputing 98, Hawaii, Jan 49, 1998. World
Scientific, Singapore, pp. 305316.
3. Jamois, E. A. (2003) Reagent-based and
product-based computational approaches in
library design. Curr Opin Chem Biol 7,
326330.
4. Irwin, J. J., Shoichet, B. K. (2005) ZINCa
free database of commercially available compounds for virtual screening. J Chem Inf
Model 45, 177182.
5. Oprea, T., Tropsha, A. (2006) Target, chemical and bioactivity databases integration is
key. Drug Discov Today 3, 357365.
6. Varnek, A., Tropsha, A. (2008) Cheminformatics Approaches to Virtual Screening, RSC,
London.
7. Tropsha, A. (2006) I in (Martin, Y. C.,
ed.) Comprehensive Medicinal Chemistry I.
pp. 113126, Elsevier, Oxford.
8. Golbraikh, A. and Tropsha, A. (2002)Beware
of q2 ! J Mol Graph Model 20, 269276.
9. Novellino, E., Fattorusso, C., Greco, G.
(1995) Use of comparative molecular field
analysis and cluster analysis in series design.
Pharm Acta Helv 70, 149154.
10. Norinder, U. (1996) Single and domain
made variable selection in 3D QSAR applications. J Chemomet 10, 95105.
11. Tropsha, A., Cho, S. J. (1998) in (Kubinyi,
H., Folkers, G., and Martin, Y. C., eds.) 3D
QSAR in Drug Design. Kluwer, Dordrecht,
The Netherlands, pp. 5769.
12. Tropsha, A., Gramatica, P., Gombar, V. K.
(2003) The importance of being earnest: validation is the absolute essential for successful

13.

14.

15.

16.

17.

18.

19.

20.

application and interpretation of QSPR models. Quant Struct Act Relat Comb Sci 22,
6977.
Golbraikh, A., Tropsha, A. (2002) Predictive
QSAR modeling based on diversity sampling
of experimental datasets for the training and
test set selection. J Comput Aided Mol Des
16, 357369.
Golbraikh, A., Shen, M., Xiao, Z., Xiao, Y.
D., Lee, K. H., Tropsha, A. (2003) Rational selection of training and test sets for
the development of validated QSAR models.
J Comput Aided Mol Des 17, 241253.
Pavan, M., Netzeva, T. I., Worth, A. P.
(2006) Validation of a QSAR model for
acute toxicity. SAR QSAR Environ Res 17,
147171.
Vracko, M., Bandelj, V., Barbieri, P., Benfenati, E., Chaudhry, Q., Cronin, M., Devillers, J., Gallegos, A., Gini, G., Gramatica, P.,
Helma, C., Mazzatorta, P., Neagu, D., Netzeva, T., Pavan, M., Patlewicz, G., Randic,
M., Tsakovska, I., Worth, A. (2006) Validation of counter propagation neural network
models for predictive toxicology according
to the OECD principles: a case study. SAR
QSAR Environ Res 17, 265284.
Saliner, A. G., Netzeva, T. I., Worth, A. P.
(2006) Prediction of estrogenicity: validation
of a classification model. SAR QSAR Environ
Res 17, 195223.
Roberts, D. W., Aptula, A. O., Patlewicz, G.
(2006) Mechanistic applicability domains for
non-animal based prediction of toxicological endpoints. QSAR analysis of the Schiff
base applicability domain for skin sensitization. Chem Res Toxicol 19, 12281233.
Zhang, S., Golbraikh, A., Tropsha, A. (2006)
Development of quantitative structurebinding affinity relationship models based
on novel geometrical chemical descriptors of
the protein-ligand interfaces. J Med Chem 49,
27132724.
Golbraikh, A., Bonchev, D., Tropsha, A.
(2001) Novel chirality descriptors derived

Application of QSAR and Shape Pharmacophore Modeling Approaches

21.

22.

23.

24.

25.

26.

27.

28.

29.
30.

31.

from molecular topology. J Chem Inf Comput Sci 41, 147158.

Kovatcheva, A., Buchbauer, G., Golbraikh,
A., Wolschann, P. (2003) QSAR modeling
of alpha-campholenic derivatives with sandalwood odor. J Chem Inf Comput Sci 43,
259266.
Kovatcheva, A., Golbraikh, A., Oloff, S.,
Xiao, Y. D., Zheng, W., Wolschann, P., Buchbauer, G., Tropsha, A. (2004) Combinatorial
QSAR of ambergris fragrance compounds. J
Chem Inf Comput Sci 44, 582595.
Shen, M., Xiao, Y., Golbraikh, A., Gombar,
V. K., Tropsha, A. (2003) Development and
validation of k-nearest-neighbor QSPR models of metabolic stability of drug candidates.
J Med Chem 46, 30133020.
Shen, M., LeTiran, A., Xiao, Y., Golbraikh,
A., Kohn, H., Tropsha, A. (2002) Quantitative structure-activity relationship analysis
of functionalized amino acid anticonvulsant
agents using k nearest neighbor and simulated annealing PLS methods. J Med Chem
45, 28112823.
Shen, M., Beguin, C., Golbraikh, A., Stables,
J. P., Kohn, H., Tropsha, A. (2004) Application of predictive QSAR models to database
mining: identification and experimental validation of novel anticonvulsant compounds. J
Med Chem 47, 23562364.
Zhang, S., Golbraikh, A., Oloff, S., Kohn,
H., Tropsha, A. (2006) A Novel Automated Lazy Learning QSAR (ALL-QSAR)
approach: method development, applications, and virtual screening of chemical
databases using validated ALL-QSAR models. J Chem Inf Model 46, 19841995.
Golbraikh, A., Shen, M., Xiao, Z., Xiao,
Y. D., Lee, K. H., Tropsha, A. (2003)
Rational selection of training and test sets
for the development of validated QSAR
models. J Comput Aided Mol Des 17,
241253.
Eriksson, L., Jaworska, J., Worth, A. P.,
Cronin, M. T., McDowell, R. M., Gramatica, P. (2003) Methods for reliability and
uncertainty assessment and for applicability
evaluations of classification- and regressionbased QSARs. Environ Health Perspect 111,
13611375.
Sachs, L. (1984) Handbook of Statistics.
Springer, Heidelberg.
Tropsha, A., Golbraikh, A. (2007) Predictive QSAR modeling workflow, model applicability domains, and virtual screening. Curr
Pharm Des 13, 34943504.
Tropsha, A. (2005) in (Oprea, T., ed.) Cheminformatics in Drug Discovery., Wiley-VCH,
New York, pp. 437455.

131

32. Medina-Franco, J. L., Golbraikh, A., Oloff,

S., Castillo, R., Tropsha, A. (2005) Quantitative structure-activity relationship analysis of pyridinone HIV-1 reverse transcriptase
inhibitors using the k nearest neighbor
method and QSAR-based database mining. J
Comput Aided Mol Des 19, 229242.
33. Oloff, S., Mailman, R. B., Tropsha, A.
(2005) Application of validated QSAR
models of D1 dopaminergic antagonists
for database mining. J Med Chem 48,
73227332.
34. Zhang, S., Wei, L., Bastow, K., Zheng, W.,
Brossi, A., Lee, K. H., Tropsha, A. (2007)
Antitumor Agents 252. Application of validated QSAR models to database mining:
discovery of novel tylophorine derivatives as
potential anticancer agents. J Comput Aided
Mol Des 21, 97112.
35. Hsieh, J. H., Wang, X. S., Teotico, D., Golbraikh, A., Tropsha, A. (2008) Differentiation of AmpC beta-lactamase binders vs.
decoys using classification kNN QSAR modeling and application of the QSAR classifier
to virtual screening. J Comput Aided Mol Des
22, 593609.
36. Tang, H., Wang, X. S., Huang, X. P., Roth,
B. L., Butler, K. V., Kozikowski, A. P., Jung,
M., Tropsha, A. (2009) Novel inhibitors of
human histone deacetylase (HDAC) identified by QSAR modeling of known inhibitors,
virtual screening, and experimental validation. J Chem Inf Model 49, 461476.
37. Peterson, Y. K., Wang, X. S., Casey,
P. J., Tropsha, A. (2009) Discovery of
geranylgeranyltransferase-I inhibitors with
novel scaffolds by the means of quantitative
structure-activity relationship modeling, virtual screening, and experimental validation. J
Med Chem 52, 42104220.
38. CCG. Molecular Operation Environment.
2008.
39. MolconnZ.
http://www.edusoft-lc.com/
molconn/ . 2010.
40. Zheng, W., Tropsha, A. (2000) Novel variable selection quantitative structureproperty
relationship approach based on the k-nearestneighbor principle. J Chem Inf Comput Sci
40, 185194.
41. Cho, S. J., Zheng, W., Tropsha, A. (1998)
Rational combinatorial library design. 2.
Rational design of targeted combinatorial
peptide libraries using chemical similarity
probe and the inverse QSAR approaches. J
Chem Inf Comput Sci 38, 259268.
42. Tropsha, A., Zheng, W. (2001) Identification
of the descriptor pharmacophores using variable selection QSAR: applications to database
mining. Curr Pharm Des 7, 599612.

132

Ebalunode, Zheng, and Tropsha

43. DesJarlais, R. L., Sheridan, R. P., Seibel, G.

L., Dixon, J. S., Kuntz, I. D., Venkataraghavan, R. (1988) Using shape complementarity
as an initial screen in designing ligands for
a receptor binding site of known threedimensional structure. J Med Chem 31,
722729.
44. Grant, J. A., Gallardo, M. A., Pickup, B. T.
(1996) A fast method of molecular shape
comparison: a simple application of a Gaussian description of molecular shape. J Comput
Chem 17, 16531666.
45. Zauhar, R. J., Moyna, G., Tian, L., Li, Z.,
Welsh, W. J. (2003) Shape signatures: a
new approach to computer-aided ligand- and
receptor-based drug design. J Med Chem 46,
56745690.
46. Meyer, A. Y., Richards, W. G. (1991) Similarity of molecular shape. J Comput Aided Mol
Des 5, 427439.
47. Good, A. C., Richards, W. G. (1993) Rapid
evaluation of shape similarity using Gaussian functions. J Chem Inf Comput Sci 33,
112116.
48. ROCS. version 3.0.0. 2009. Santa Fe, NM,
USA, OpenEye Scientific Software.
49. OEShape Toolkit. version 1.7.2. 2009.
Santa Fe, NM, USA, OpenEye Scientific
Software.
50. Carbo, R., Domingo, L. (1987) Lcao-Mo
similarity measures and taxonomy. Int J
Quantum Chem 32, 517545.
51. Carbo, R., Leyda, L., Arnau, M. (1980)
An electron density measure of the similarity between two compounds. Int J Quantum
Chem 17, 11851189.
52. Masek, B. B., Merchant, A., Matthew, J.
B. (1993) Molecular shape comparison of
angiotensin II receptor antagonists. J Med
Chem 36, 12301238.
53. Ballester, P. J., Richards, W. G. (2007) Ultrafast shape recognition to search compound
databases for similar molecular shapes. J
Comput Chem 28, 17111723.
54. Breneman, C. M., Thompson, T. R., Rhem,
M., Dung, M. (1995) Electron-density modeling of large systems using the transferable atom equivalent method. Comput Chem
19, 161.
55. Breneman, C. M., Sundling, C. M., Sukumar, N., Shen, L., Katt, W. P., Embrechts,
M. J. (2003) New developments in PEST
shape/property hybrid descriptors. J Comput
Aided Mol Des 17, 231240.
56. Das, S., Kokardekar, A., Breneman, C. M.
(2009) Rapid comparison of protein binding site surfaces with property encoded
shape distributions. J Chem Inf Model 49,
28632872.

57. Nilakantan, R., Bauman, N., Venkataraghavan, R. (1993) New method for rapid characterization of molecular shapes: applications
in drug design. J Chem Inf Comput Sci 33,
7985.
58. Schlosser, J., Rarey, M. (2009) Beyond the
virtual screening paradigm: structure-based
searching for new lead compounds. J Chem
Inf Model 49, 800809.
59. Zauhar, R. J. (1995) SMART: a solventaccessible triangulated surface generator for
molecular graphics and boundary element
applications. J Comput Aided Mol Des 9,
149159.
60. Hawkins, P. C., Skillman, A. G., Nicholls,
A. (2007) Comparison of shape-matching
and docking as virtual screening tools. J Med
Chem 50, 7482.
61. McGaughey, G. B., Sheridan, R. P., Bayly, C.
I., Culberson, J. C., Kreatsoulas, C., Lindsley, S., Maiorov, V., Truchon, J. F., Cornell, W. D. (2007) Comparison of topological, shape, and docking methods in
virtual screening. J Chem Inf Model 47,
15041519.
62. Daylight. version 4.82. 2003. Aliso Viejo,
CA, USA, Daylight Chemical Information
Systems Inc.
63. Kearsley, S. K., Sallamack, S., Fluder, E. M.,
Andose, J. D., Mosley, R. T., Sheridan, R.
P. (1996) Chemical similarity using physiochemical property descriptors. J Chem Inf
Comput Sci 36, 118127.
64. Miller, M. D., Sheridan, R. P., Kearsley,
S. K. (1999) SQ: a program for rapidly
producing
pharmacophorically
relevant
molecular superpositions. J Med Chem 42,
15051514.
65. Miller, M. D., Kearsley, S. K., Underwood,
D. J., Sheridan, R. P. (1994) FLOG: a system to select quasi-flexible ligands complementary to a receptor of known threedimensional structure. J Comput Aided Mol
Des 8, 153174.
66. McGann, M. R., Almond, H. R., Nicholls,
A., Grant, J. A., aBrown, F. K. (2003)
Gaussian docking functions Biopolymers 68,
7690.
67. Friesner, R. A., Banks, J. L., Murphy, R. B.,
Halgren, T. A., Klicic, J. J., Mainz, D. T.,
Repasky, M. P., Knoll, E. H., Shelley, M.,
Perry, J. K., Shaw, D. E., Francis, P., Shenkin,
P. S. (2004) Glide: a new approach for rapid,
accurate docking and scoring. 1. Method and
assessment of docking accuracy. J Med Chem
47, 17391749.
68. Halgren, T. A., Murphy, R. B., Friesner, R.
A., Beard, H. S., Frye, L. L., Pollard, W. T.,
Banks, J. L. (2004) Glide: a new approach

Application of QSAR and Shape Pharmacophore Modeling Approaches

69.

70.

71.

72.

73.

74.

75.

76.

77.

78.

79.

80.

for rapid, accurate docking and scoring. 2.

Enrichment factors in database screening. J
Med Chem 47, 17501759.
Moffat, K., Gillet, V. J., Whittle, M., Bravi,
G., Leach, A. R. (2008) A comparison of
field-based similarity searching methods: CatShape, FBSS, and ROCS. J Chem Inf Model
48, 719729.
Hahn, M. (1997) Three-dimensional shapebased searching of conformationally flexible compounds. J Chem Inf Comput Sci 37,
8086.
Wild, D. J., Willett, P. (1996) Similarity
searching in files of three-dimensional chemical structures. Alignment of molecular electrostatic potential fields with a genetic algorithm. J Chem Inf Comput Sci 36, 159167.
Ebalunode, J. O., Ouyang, Z., Liang,
J., Zheng, W. (2008) Novel approach to
structure-based pharmacophore search using
computational geometry and shape matching
techniques. J Chem Inf Model 48, 889901.
Ebalunode, J. O., Zheng, W. (2009) Unconventional 2D shape similarity method affords
comparable enrichment as a 3D shape
method in virtual screening experiments. J
Chem Inf Model 49, 13131320.
Nagarajan, K., Zauhar, R., Welsh, W. J.
(2005) Enrichment of ligands for the serotonin receptor using the shape signatures
approach. J Chem Inf Model 45, 4957.
Ballester, P. J., Finn, P. W., Richards, W.
G. (2009) Ultrafast shape recognition: evaluating a new ligand-based virtual screening technology. J Mol Graph Model 27,
836845.
Tawa, G. J., Baber, J. C., Humblet, C.
(2009) Computation of 3D queries for
ROCS based virtual screens. J Comput Aided
Mol Des 23, 853868.
Kirchmair, J., Distinto, S., Markt, P., Schuster, D., Spitzer, G. M., Liedl, K. R., Wolber,
G. (2009) How to optimize shape-based virtual screening: choosing the right query and
including chemical information. J Chem Inf
Model 49, 678692.
Nagarajan, K., Zauhar, R., Welsh, W. J.
(2005) Enrichment of ligands for the serotonin receptor using the shape signatures
approach. J Chem Inf Model 45, 4957.
Wolber, G., Langer, T. (2005) LigandScout: 3-d pharmacophores derived from
protein-bound Ligands and their use as virtual screening filters. J Chem Inf Model 45,
160169.
Lee, H. S., Lee, C. S., Kim, J. S., Kim, D. H.,
Choe, H. (2009) Improving virtual screen-

81.

82.

83.

84.

85.

86.

87.

88.

133

ing performance against conformational variations of receptors by shape matching with

ligand binding pocket. J Chem Inf Model 49,
24192428.
Markt, P., Petersen, R. K., Flindt, E. N.,
Kristiansen, K., Kirchmair, J., Spitzer, G.,
Distinto, S., Schuster, D., Wolber, G., Laggner, C., Langer, T. (2008) Discovery of
novel PPAR ligands by a virtual screening
approach based on pharmacophore modeling, 3D Shape, and electrostatic similarity
screening. J Med Chem 51, 63036317.
Rush, T. S., Grant, J. A., Mosyak, L.,
Nicholls, A. (2005) A shape-based 3-D scaffold hopping method and its application to
a bacterial protein protein interaction. J Med
Chem 48, 14891495.
Cramer, R. D., Poss, M. A., ermsmeier, M.
A., Caulfield, T. J., Kowala, M. C., Valentine, M. T. (1999) Prospective identification
of biologically active structures by topomer
shape similarity searching. J Med Chem 42,
39193933.
Cramer, R. D., Jilek, R. J., Guessregen, S.,
Clark, S. J., Wendt, B., Clark, R. D. (2004)
Lead Hopping. Validation of topomer
similarity as a superior predictor of similar biological activities. J Med Chem 47,
67776791.
Muchmore, S. W., Souers, A. J.,
Akritopoulou-Zanze, I. (2006) The use of
three-dimensional shape and electrostatic
similarity searching in the identification of
a melanin-concentrating hormone receptor 1 antagonist. Chem Biol Drug Des 67,
174176.
Perez-Nueno, V. I., Ritchie, D. W., Rabal,
O., Pascual, R., Borrell, J. I., Teixido, J.
(2008) Comparison of ligand-based and
receptor-based virtual screening of HIV
entry inhibitors for the CXCR4 and CCR5
receptors using 3D ligand shape matching
and ligand-receptor docking. J Chem Inf
Model 48, 509533.
Ballester, P. J., Westwood, I., Laurieri, N.,
Sim, E., Richards, W. G. (2010) Prospective
virtual screening with Ultrafast shape recognition: the identification of novel inhibitors
of arylamine N-acetyltransferases. J R Soc
Interface 7, 335342.
Ebalunode, J. O., Dong, X., Ouyang,
Z., Liang, J., Eckenhoff, R. G., Zheng,
W. (2009) Structure-based shape pharmacophore modeling for the discovery of novel
anesthetic compounds. Bioorg Med Chem 17,
51335138.

Chapter 7
Combinatorial Library Design from Reagent Pharmacophore
Fingerprints
Hongming Chen, Ola Engkvist, and Niklas Blomberg
Abstract
Combinatorial and parallel chemical synthesis technologies are powerful tools in early drug discovery
projects. Over the past couple of years an increased emphasis on targeted lead generation libraries and
focussed screening libraries in the pharmaceutical industry has driven a surge in computational methods
to explore molecular frameworks to establish new chemical equity. In this chapter we describe a complementary technique in the library design process, termed ProSAR, to effectively cover the accessible
pharmacophore space around a given scaffold. With this method reagents are selected such that each
R-group on the scaffold has an optimal coverage of pharmacophoric features. This is achieved by optimising the Shannon entropy, i.e. the information content, of the topological pharmacophore distribution
for the reagents. As this method enumerates compounds with a systematic variation of user-defined pharmacophores to the attachment point on the scaffold, the enumerated compounds may serve as a good
starting point for deriving a structureactivity relationship (SAR).
Key words: ProSAR, combinatorial library design, topological pharmacophore, pharmacophore
fingerprint, genetic algorithm, Shannon entropy, multi-objective optimisation.

1. Introduction
Effective structureactivity relationship (SAR) generation is at the
centre of any medicinal chemistry campaign. Much work has been
done to devise effective methods to explain and explore SAR
data for medicinal chemistry teams to drive the design cycles
within drug discovery projects (1). Recent work on SAR generation highlights the commonly observed discontinuity of SAR
and bioactivity data, the so-called activity cliffs (2). This also
emphasises the need to empirically determine SAR for each lead
J.Z. Zhou (ed.), Chemical Library Design, Methods in Molecular Biology 685,
DOI 10.1007/978-1-60761-931-4_7, Springer Science+Business Media, LLC 2011

135

136

Chen, Engkvist, and Blomberg

series; indeed, it is often difficult to rationalise existing SAR data

even with access to high-resolution X-ray crystal structures of the
target-compound complex (3, 4). Another common challenge
for the medicinal chemistry teams is that many pharmacokinetic
properties are often inherent to the scaffold and breaking out of
this property space can be very difficult. Thus, the team needs
to quickly explore the chemical space around a novel scaffold
to establish SAR and make decisions on the medicinal chemistry
strategy.
The art and science of computational library design has been
reviewed extensively elsewhere (57), but it is interesting and
instructive to note the developments in library design over the
past 10 years showing the continued importance of the subject
for the industry. Today, there is less emphasis on the analysis
of molecular properties and diversity as the objectives of library
design have shifted towards focussed lead generation. Design of
target-directed libraries and the need to establish novel chemical
equity have driven the concept of scaffold diversity with a significant effort to identify novel methods for scaffold hopping.
A key enabler for this work is that direct structural descriptions
of molecules and common framework/substructure analysis have
been more computationally accessible (810).
Pharmacophore-based approaches are widely used in library
design. A pharmacophore refers to the topological (2D) or 3D
arrangement of functional groups that capture the key interactions of a ligand with its enzyme/receptor. The major attractiveness of pharmacophore-based methods is that they do not rely
on the 3D structural information of protein target and thus are
applicable for all target classes and therefore for all drug discovery projects. The concept of pharmacophore fingerprint (1112)
was introduced to describe the pharmacophore patterns present
in a molecule in a manner analogous to substructural fingerprints (13). A pharmacophore fingerprint is normally encoded
as a binary bit string where each bit refers to a pharmacophore
pattern, i.e. a set of pharmacophore points separated by a given
distance or distance range. The pharmacophore pattern can be
atom/pharmacophore pair, pharmacophore triplet (11, 14, 15)
or quartet (12, 16). The distance between a pair of pharmacophore points is usually binned to capture variations in conformation (3D) or bond distances (2D). The pharmacophore
types normally comprise hydrogen bond acceptor and donor, positive and negative charge centre and hydrophobic group, but
most softwares allow for user-defined types to capture targetspecific features such as metal-chelating groups. Pharmacophore
fingerprints are often used in diverse library design to cover a
broad pharmacophore space (12, 1416). Chem-Diverse (17)
was the first commercial software to exploit 3-point and 4-point
pharmacophore in diversity analysis. Since then many efforts

Reagent Pharmacophore Fingerprints

137

(1826) have been made in applying pharmacophore fingerprints

to combinatorial library design. For example, Good et al. (18)
reported their HARPick program which makes combinatorial
library design in reagent space. A Monte Carlo simulated annealing optimisation method was used to optimise the reagent selection to achieve maximal diversity fitness score. Chemical diversity
(2628) is often used as an optimisation function for combinatorial library design, either on the reagent side (29, 30) or on the
product side (31, 32). Such library design strategies are often very
efficient at selecting diverse compounds. However, they may lead
to libraries where it is hard to derive a clear structureactivity relationship (SAR) from the experimental data as the selected building blocks might have little or no relationship to one another.
Recently, we have reported (33) a reagent-based library
design strategy ProSAR to tackle these issues. The ProSAR
method relies on topological 2-point pharmacophores to enumerate and optimise a selection of reagents to systematically explore
novel scaffolds. Thus, the ProSAR method is complementary to
scaffold analysis and computational scaffold hopping tools and
addresses a separate step in the library design workflow.
In this chapter we will exemplify this method with selected
library design problems and also demonstrate how to apply
ProSAR designs with concurrent optimisation of product property profile to design libraries that will not only help to derive a
SAR, but also have an attractive property profile.

2. Materials
The 2-point pharmacophores were created by an in-house tool
TRUST (34) and a shell script was written to create reagent pharmacophore fingerprint based on TRUST output. The greedy
search algorithm (35) was implemented in Python (36) to read in
reagent pharmacophore fingerprint and optimise pharmacophore
entropy. An in-house genetic algorithm-driven library design tool
GALOP was used to optimise library under user-supplied multiple
constraints. Library product properties were calculated by various
in-house prediction tools. Tanimoto similarity for reagents was
calculated using FOYFI fingerprint which is an in-house developed fingerprint (37) and is similar in spirit to standard Daylight
fingerprint (38). Database similarity searches were done by using
an in-house 2D similarity search tool (34) with FOYFI fingerprint. An in-house program FLUSH (37) was used for the structure clustering.
Three sets of commercially available reagents are used in this
study: a set of 493 aliphatic primary amines for selecting 20

138

Chen, Engkvist, and Blomberg

reagents subset, 2,518 aldehydes and 634 amino acids as reagent

pool for making various 2020 2D libraries and 112 aliphatic
bromides together with 127 aliphatic amines as reagent collection for designing concurrent pharmacophore entropy and library
property profile optimised 2D libraries. All the reagents are from
ACD (39). In the second example, 139 known active compounds
which share the same scaffold and are used as validation set are
taken from GVKBio Medchem database (40).

3. Method
3.1. Methodology
3.1.1. Definition of the
Pharmacophore
Fingerprint

Three-point and 4-point pharmacophores (12, 1416) have been

widely used to represent chemical information of library products.
In ProSAR, we extend this concept to a 2-point pharmacophore
to encode the chemical information of a reagent. The ProSAR
reagent pharmacophore is composed of a single pharmacophore
point plus the attachment point of the reagent. Here, we use the
five common pharmacophore types: hydrogen bond donor (HD),
hydrogen bond acceptor (HA), positive charge centre (POS),
negative charge centre (NEG) and lipophilic groups (LIP). In our
in-house implementation, these are encoded as SMARTS strings
(38).
For each reagent the information of the pharmacophores and
their respective distance to the attachment point are incorporated
into a fingerprint (as shown in Fig. 7.1). Note that even rather
simple reagents will have multiple pharmacophores. As we normally would want to select low-complexity reagents and avoid
adding long side chains to the scaffold, the maximal topological distance (bond distance) between the pharmacophore element
and the attachment point is restricted to six bonds, and the sum of
donor, acceptor, positive and negative charge groups in a reagent
should be less than or equal to two. Thus, the total number of
unique 2-point pharmacophores in a reagent is 30 (56) and the

Fig. 7.1 Reagent pharmacophore fingerprint encoding (adapted from (33)).

Reagent Pharmacophore Fingerprints

139

reagent information is represented by a 30-bin pharmacophore

fingerprint. Each bin of the fingerprint refers to a specific 2-point
pharmacophore and the count of the specific pharmacophore in
the reagent is recorded into this bin. A pharmacophore fingerprint
for an amine reagent is exemplified in Fig. 7.1. Compared with
other pharmacophore fingerprints, this method explicitly captures
reagent pharmacophores where one endpoint for the fingerprint
is always the attachment point on the scaffold. The advantage of
such a fingerprint is that the pharmacophore variability in the fingerprint is always relative to the same position and thus provides
a common framework to compare pharmacophore variations for
different reagents to further derive SAR information.
3.1.2. Reagent Selection
Based on Optimisation in
Pharmacophore Space

Once reagent pharmacophore fingerprints are created for all

reagents, the next step in the ProSAR strategy is to do reagent
selection to optimally cover the pharmacophore fingerprint
space and keep the pharmacophore distribution as even as possible. Shannon entropy (SE) (41) has been shown to be an efficient way to characterise the variation of molecular descriptors in
compound databases (42). Groothhuis et al. (35, 43) and Miller
et al. (44) used SE to measure the chemical diversity of libraries;
here, we use SE to represent the pharmacophore distribution of a
selected reagent set. SE is defined as follows:
SE =

pi log2 pi

[1]

where pi is the probability of having a certain pharmacophore in

the whole reagent set. pi is calculated as follows:
pi = c i /

ci

[2]

where ci is the population of pharmacophore i in the whole

reagent set. A larger SE corresponds to a greater information content, i.e. a more even distribution of reagent pharmacophores.
Hence the optimal selection from the set of available reagents
will maximise the SE value after library optimisation. A Python
(36) program, in which a greedy search algorithm (35) is used as
the optimisation search engine, has been developed to make the
ProSAR reagent selection.
The general procedure for doing ProSAR library design is as
follows: first, all the reagents are collected in smiles file format,
and second, a shell script is run to do prefiltering on reagents
(remove too complex reagents as described in Section 3.1.1) and
create the 2-point pharmacophore fingerprints for the remaining
reagents; a greedy search optimisation is done by running the

140

Chen, Engkvist, and Blomberg

python script with the generated pharmacophore fingerprint to

select the desired number of reagents.
3.1.3. Concurrent
Pharmacophore Entropy
and Library Property
Profile Optimisation in
ProSAR Library Design

Physico-chemical properties and evaluation of potential safety

liabilities are important aspects of the library design process.
Predicted properties like hERG liability (45), compound aqueous solubility, etc. (4648) have been extensively studied and
included in various library design strategies (49, 50) as a part
of multiple constraints optimisation. We have therefore further
extended the ProSAR concept to take the library property profile into account in the design process. Several in-house calculated properties are considered; these include a compound novelty
check (that checks in in-house and external compound databases
to see if the compound is novel), predicted aqueous solubility
(51), predicted hERG liability (52) and an in-house developed
lead profile score (53, 54).
An in-house library design tool GALOP (33) was extended
to include ProSAR designs and it is used in the extended ProSAR
library design procedure to replace the greedy search optimisation. GALOP uses a genetic algorithm (GA) (55) to optimise the
reagent pharmacophore SE and the product properties simultaneously. In the GA algorithm, each chromosome corresponds to
a selected library and it consists of an array of binary bins. Each
bin refers to the presence of a reagent. The GA fitness function is
a linear combination of the reagent pharmacophore SE term and
the product property profile term (as shown in equation [3]):
Score = wp F + we

SEj

[3]

Here, F refers to the property profile term and is measured by

the fraction of good compounds in the designed library and SEj
refers to the SE for the reagent set which is used for side chain j.
A compound is regarded as good only if it meets all the specified
property criteria. wp and we are weighting factors for the properties and SE, respectively. In our experience, a weight ratio (we /wp )
of 2 works well and is used throughout the libraries designed in
this study.
3.2. Application
Examples
3.2.1. Selection of
Primary Amine Reagents

As the first test case, we selected 20 aliphatic primary amines from

a set of 493 commercially available ones by three different methods: random reagent selection, entropy-optimised ProSAR selection and an occupancy-optimised method which purely maximises
the occupancy of pharmacophore bins (56), i.e. ensures that as
many bins as possible are covered by the reagent selection regardless of the pharmacophore distribution. As the greedy algorithm

Reagent Pharmacophore Fingerprints

141

is a deterministic method in nature, ProSAR and occupancy optimisation give the optimal reagent selection (within the given constraints), while random selections was repeated ten times to get
ten different reagent sets.
The distributions of reagent pharmacophore bins for the
ProSAR reagent set, one random reagent set and the occupancyoptimised reagent set are compared in Fig. 7.2. The total SE of
the ProSAR selection is 4.15, average SE of ten random selections is 3.3 and the SE for occupancy-based selection is 3.4.
Entropy-driven ProSAR selections have the same pharmacophore
coverage as the occupancy-optimised set and both optimisation
techniques achieve better coverage than random selections. Additionally, entropy-optimised ProSAR has the most even bin distribution of the selections. As an example, for the lipophilic bins
from no. 14 to 17, the reagent count for random selection is 39,
the count for occupancy selection is 33, while the count of these
bins has been reduced to 15 in the ProSAR library. Entropy-based
optimisation achieves the same level of pharmacophore coverage
as occupancy optimisation but has a more even distribution of
pharmacophores in the reagent set and does not bias the selection towards reagents with lipophilic pharmacophores.

Fig. 7.2 Pharmacophore fingerprint distribution for 20 primary amines selected by using
the ProSAR strategy, random selection and occupancy optimisation of fingerprint bins,
respectively.

3.2.2. Affymax Library

Example

A pending question for ProSAR library design is how well the

design covers the pharmacophores from real active compounds.
Therefore, we compared the pharmacophores of a ProSAR library
with those of active compounds for a specific scaffold. A library
example from Affymax (5759) is selected as the test case here
(shown in Fig. 7.3). The library diversity is generated from aldehydes (R1) and amino acids (R2) and active compounds for several targets were identified by screening the library. A total of 139
known active compounds with this scaffold were retrieved from
the GVKBio MedChem database (40) and are used as validation
set.

142

Chen, Engkvist, and Blomberg

1. O = C(O)C(R2)NH2

O
R1

2. R1CHO
R

OH

HS
S(Trt)

N
HN
R2

3.

O
BocHN

Fig. 7.3 Combinatorial library example from Affymax (57).

In this study, 2,518 aldehydes and 634 amino acids were

selected from ACD (39) and used as reagent pool for the libraries
(2020). A ProSAR library was built using the greedy algorithm
with ten conventional diversity libraries and ten random reagent
selections as a comparison. The diversity libraries were built by
using GALOP with the average Tanimoto dissimilarity for the
reagent ensemble (based on the in-house FOYFI (37) structural
fingerprint) used as the GA fitness function. The pharmacophore
distributions of R1 and R2 for the different reagent collections
are compared in Fig. 7.4 and the results for the libraries from
different design strategies are summarised in Table 7.1. It can be
seen that the ProSAR reagent sets cover almost all of the pharmacophore bins (27 bins covered in both R1 and R2) while having an even reagent distribution in the covered bins (SE for R1
and R2 reagents are 4.61 and 4.65, respectively). Random and
diversity libraries have marked lower pharmacophore bin coverage (Fig. 7.4).
Comparing pharmacophores from known active compounds,
all the R1 pharmacophore bins in active set are covered by the
ProSAR library while two bins (no. 8 and no. 20) are missing
in the random and diversity libraries. For the R2 reagents, one
pharmacophore bin from the active molecules (no. 12) is not
found in the ProSAR reagents. For the random and the diversity
libraries there are ten and six bins not covered, respectively. In
this example, SE-driven optimisation of ProSAR pharmacophores
has a marked better coverage of potentially important pharmacophore elements present in the known active compounds set. In
addition to the pharmacophores present in the active molecules,
the ProSAR library also covers many more additional pharmacophores compared to the structural fingerprint diversity library
and random selections (Table 7.1).
To further estimate the likelihood of obtaining active
molecules from the compounds in the designed libraries, compounds in the designed libraries were used as queries and similarity searches against the GVKBio database with a high similarity cut-off were performed to investigate how many active compounds could be retrieved. Taking the observation that similar

Reagent Pharmacophore Fingerprints

143

(a)

(b)
Fig. 7.4 (a) Pharmacophore fingerprint distribution for the R1 reagents. (b) Pharmacophore fingerprint distribution for
the R2 reagents (adapted from (33)).

compounds tend to have similar bioactivity (60) as an axiom, a

high retrieval rate from the GVKBio database is taken as an indication that potentially active molecules are present in the library.
Library products are therefore used as query structures to search
against the GVKBio database to retrieve active compounds with
the conservative similarity cut-off (based on FOYFI fingerprint)
of 0.85. From these searches (Table 7.1) the ProSAR library
retrieves 20 compounds, while the random and diversity libraries
retrieve on average 11.7 and 1.1 compounds, respectively. The
ProSAR library clearly has the best retrieval rates for active compounds among all the designed libraries, and at the same time

144

Chen, Engkvist, and Blomberg

Table 7.1
Results of the designed libraries for the Affymax example
(adapted from (33))
ProSAR
libraries

Random
librariesa

Number of recovered active

compoundsb

20

11.7

1.1

Shannon entropy

R1

4.61

3.1

3.2

R2

4.65

2.9

3.6

R1

27

13.8

12.1

R2

27

13.3

17.2

Libraries

Number of
covered bins

Diversity
librariesa

a Average values based on ten library designs.

b Retrieved active compound from the GVKBio database in the similarity search with

a Tanimoto similarity cut-off of 0.85.

has the highest coverage of pharmacophores present in the active

compounds.
3.2.3. Concurrent
ProSAR and Property
Profile Optimisation

Optimisation of reagent pharmacophore space alone is not

enough for most pharmaceutical industry applications of library
design (61). A good compound property profile for the designed
libraries is required, so in practice the ProSAR strategy needs
to include the property profile of the products in the optimisation. Our in-house genetic algorithm optimiser GALOP (33) was
implemented specifically to design compound libraries with multiple constraints (62, 63).
In the extended ProSAR strategy, both the pharmacophore
SE and the compound property profile are included in the GA
fitness function as shown in equation [3]. Compound properties
considered in the algorithm implementation include (1) novelty
check, (2) in silico predicted aqueous solubility (51), (3) in silico
predicted hERG liability (52) and (4) in-house lead-like criteria
(53, 54). A good compound has to pass all the four criteria. One
library example (Fig. 7.5) is used to demonstrate this extended
ProSAR strategy. A set of 112 aliphatic bromides (R1 reagent)
and a set of 127 aliphatic amines (R2 reagent) are used as the
reagent pool. Ten ProSAR libraries, ten diversity combined with
property-optimised libraries and ten libraries only optimised by
property were created using GALOP with different fitness functions. As a reference, ten libraries are created with random reagent
selections. Each library was clustered using FOYFI structural fingerprints such that we can use a number of clusters as a simple
estimate of the structural diversity.
Property-optimised ProSAR libraries have the best pharmacophore Shannon entropy of all the libraries and 99.7% of the
compounds have good properties (Table 7.2). In terms of phar-

Reagent Pharmacophore Fingerprints

145

O
1. Br-R1
2. HCl
3. R2R3-NH

NH

R1
R3
N

O
O

R2

Fig. 7.5 Library example for concurrent reagent pharmacophore entropy and library
property profile optimisation.

macophore coverage, the ProSAR libraries cover on average 10.7

bins in R1, slightly lower than the coverage of random libraries.
This could be due to the limited variation in R1 for compounds
with a good property profile. In the R2 reagents, ProSAR libraries
cover on average 15.4 bins, markedly better than any other design
strategies. Diversity/property optimisation produces most diverse
libraries; this can be seen from its highest average FOYFI Tanimoto dissimilarity value and number of clusters. These libraries
have 99.7% good compounds. As expected, property-optimised
libraries have a perfect profile (100% good compounds) but low
SE and diversity (Table 7.2). The random libraries have the worst
property profile with medium entropy and diversity values.

Table 7.2
Results for the GA-optimised librariesa (adapted from (33))
Libraries

ProSAR+
propertyb

Diversity+
propertyc

Propertyd

Random
librarye

Full
library

% of good compounds

99.7

99.7

100

62.2

62

Number of clusters
Shannon entropy R1

21

46.1

23

NC

Dissimilarity
index
Number of
covered bins

14.1

3.03

2.86

2.38

2.71

2.83

R2

3.52

2.62

2.32

2.81

2.94

R1

0.74

0.80

0.64

0.72

0.74

R2

0.69

0.80

0.65

0.71

0.73

R1

10.5

10.3

R2

15.4

10.2

10.5

10.7

21

12

20

a The values listed in the table are averaged over ten library designs, except for the full library.
b Libraries obtained by optimising both the pharmacophore entropy and the property profile simultaneously.
c Libraries obtained by optimising both the diversity and the property profile simultaneously.
d Libraries obtained by only optimising the property profile.
e Libraries obtained by randomly selecting reagents.

As an illustration, one ProSAR library and one diversity

library were selected for a closer investigation. The R1 and R2
pharmacophore distributions are shown in Fig. 7.6 with the
structures of the selected R1 and R2 reagents shown in Figures
7.7, 7.8, 7.9 and 7.10. For the R1 reagents the diversity library

146

Chen, Engkvist, and Blomberg

lacks bin no. 5 (acceptor five bonds distant to the attachment

point) and 11 (donor five bonds distant to the attachment point)
while both of these pharmacophores are present in the ProSAR
library. For the R2 reagent set, bin no. 9, 10, 21, 22 and 27 are
missing in the diversity library while being present in the ProSAR
library. Again in this example the ProSAR library has a more balanced reagent set in terms of pharmacophoric features and pharmacophore variations than the diversity library. On examination
of the R1 and R2 reagents for the two libraries, one sees that the
ProSAR reagent set has more structurally related compounds. For
example, reagents 1, 2 and 3 of ProSAR R2 reagent set (Fig. 7.9)
are similar structures with variations on the alcohol functionality
and lipophilic bulk; this could potentially help to derive a SAR
around the HD functionality on the side chain. Similarly, structures 12 and 13 may provide SAR around the positive charge

(a)

(b)
Fig. 7.6 Comparison of pharmacophore fingerprint distribution for libraries with different design strategies. (a) Pharmacophore fingerprint distribution for R1 reagents. (b) Pharmacophore fingerprint distribution for R2 reagents (adapted from
(33)).

Reagent Pharmacophore Fingerprints

147

F
Br

Br

Br

Br

Br

Br

F
F

Br

Br

O
7

8
F

Br

10
S

OH

Br

Br

Br

N
Br

Br
F

O
N

11

12

13

Br

Br

Br

14

Cl
O

Br

O
15

16

17

18

Br
F
O
19

Br

F
20

Fig. 7.7 Selected R1 reagents for the ProSAR library (adapted from (33)).

functionality and structures 411 may show some SAR around

the piperazine ring. These structurally related reagents will have
less chance to be selected in the diversity-based design strategy
due to the low Tanimoto dissimilarity value (see Section 4). In
summary, the ProSAR libraries tend to include structurally related
reagents with systematic variation of side chain pharmacophore
elements. These designs are helpful to chemists attempting to
derive SAR.

4. Conclusion
The ProSAR strategy for library design selects reagents by optimising the reagent pharmacophore space to achieve a systematic
variation of the pharmacophores relative to a scaffold attachment.
We show that optimising the Shannon entropy of the reagent

148

Chen, Engkvist, and Blomberg

F

Br

Br

Br

Br

F Br

Br
1

F
2

Br

Br

Br

Br

Br

Br
F

O
8

11

10

Br

Cl Br

Br

Br

13

12
Br

Cl

14

O
16

15

17

F
Br
Br
N

Br

O
18

F
F

19

20

Fig. 7.8 Selected R1 reagents for the diversity library (adapted from (33)).

OH

OH

OH

O
N
N

S
O

10
N

N
N

S
11

N
N

15

13

12

14

N
N
N

17

O
16

O
N

18

19

Fig. 7.9 Selected R2 reagents for the ProSAR library (adapted from (33)).

20

Reagent Pharmacophore Fingerprints

O
2

Cl
6

5
N

N
N

N
8

11

10

N
N

14

13

149

15

S
O

16

12
N
N

17

18

19

20

Fig. 7.10 Selected R2 reagents for the diversity library (adapted from (33)).

pharmacophores effectively covers the available pharmacophores

among the reagents. It also reduces bias of over-represented pharmacophores and evens the distribution among the reagents, thus
potentially making it easier for medicinal chemists to derive SAR.
A ProSAR-derived library can also retrieve more bioactive compounds from a database than other design strategies evaluated. In
practice, the full ProSAR strategy includes compound properties
to obtain libraries which possess not only a wide pharmacophore
coverage from the reagents but also satisfactory physico-chemical
properties.
It should be borne in mind that diversity in pharmacophore
space is not equivalent to the structural diversity. As we can
see from the third application example, optimising the average Tanimoto dissimilarity will create a more structurally diverse
compound set with little relationship among the compounds,
while ProSAR-optimised reagent set tends to include several
clusters of structure-related compounds which have systematic variation on reagent pharmacophore. However, ultimately
the choice of library design strategy depends on the design
objective.

150

Chen, Engkvist, and Blomberg

Acknowledgements
The authors are grateful to the following colleagues at
AstraZeneca: Dr. David Cosgrove for providing the FOYFI fingerprint programs, Dr. Jens Sadowski for providing the tool to
extract the R-groups for the library compounds and Dr. Ulf
Brjesson for developing the GALOP program.
References
1. Bajorath, J., Peltason, L., Wawer, M.,
Guha, R., Lajiness, M. S., van Drie, J. H.
(2009) Navigating structure activity landscapes. Drug Discovery Today 14, 698705.
2. Maggiora, G. M. (2006) On outliers and
activity cliffs why QSAR often disappoints.
J Chem Inf Model 46, 1535.
3. Sisay, M. H., Peltason, L., Bajorath, J.
(2009) Structural interpretation of activity cliffs revealed by systematic analysis
of structureactivity relationships in analog
series. J Chem Inf Model 49, 21792189.
4. Bostrm, J., Hogner, A., Schmitt, S. (2006)
Do structurally similar ligands bind in a similar fashion? J Med Chem 49, 67166725.
5. Spellmeyer, D. C., Grootenhuis, P. D. J.
(1999) Recent developments in molecular
diversity: computational approaches to combinatorial chemistry. Annu Rep Med Chem
Rev 34, 287296.
6. Beno, B. R., Mason, J. S. (2001) The design
of combinatorial libraries using properties
and 3D pharmacophore fingerprints. Drug
Discovery Today 6, 251258.
7. Willett, P. (2000) Chemoinformatics similarity and diversity in chemical libraries. Curr
Opin Biotechnol 11, 8588.
8. Bemis, G. W., Murcko, M. A. (1996) The
properties of known drugs. 1. Molecular
frameworks. J Med Chem 39, 28872893.
9. Xu, Y. J., Johnson, M. (2002) Using molecular equivalence numbers to visually explore
structural features that distinguish chemical
libraries. J Chem Inf Comp Sci 42, 912926.
10. Pitt, W., Parry, D. M., Perry, B. G., Groom,
C. R. (2009) Heteroaromatic rings of the
future. J Med Chem 52, 29522963.
11. Good, A. C., Kuntz, I. D. (1995) Investigating the extension of pairwise distance
pharmacophore measures to triplet-based
descriptors. J Mol Comput Aided Mol Des 9,
373379.
12. Mason, J. S., Morize, I., Menard, P. R.,
Cheney, D. L., Hulme, C., Labaudiniere,
R. F. (1999) New 4-point pharmaophore

13.
14.

15.

16.

17.

18.

19.

20.

21.

method for molecular similarity and diversity applications: overview of the method and
applications, including a novel approach to
the design of combinatorial libraries containing privileged substructures. J Med Chem 42,
32513264.
Symyx, 2.5; Symyx Technologies Inc., Santa
Clara, CA 95051, USA.
McGregor, M. J., Muskal, S. M. (1999)
Pharmacophore fingerprinting. 1. Application to QSAR and focused library design. J
Chem Inf Comput Sci 39, 569574.
Pickett, S. D., Mason, J. S., Mclay, I. M.
(1996) Diversity profiling and design using
3D pharmacophore: pharmacophore-derived
queries (PDQ). J Chem Inf Comput Sci 36,
12141223.
Mason, J. S., Beno, B. R. (2000) Library
design using BCUT chemistry-space descriptors and multiple four-point pharmacophore
fingerprints: simultaneous optimization and
structure-based diversity. J Mol Graph Mod
18, 438451.
Cato, S. J. (2000) Exploring pharmacophores with Chem-X, in (Gner, O., ed.)
Pharmacophore Perception, Development, and
Use in Drug Designer. International University Line, La Jolla, CA, pp. 107125.
Good, A. C., Lewis, R. A. (1997) New
methodology for profiling combinatorial
libraries and screening sets: cleaning up the
design process with HARPick. J. Med Chem
40, 39263936.
Chen. X., Rusinko, A., III, Young, S. S.
(1998) Recursive partitioning analysis of a
large structure-activity data set using threedimensional descriptors. J Chem Inf Comput
Sci 38, 10541062.
Matter, H., Ptter, T. (1999) Comparing
3D pharmacophore triplets and 2D fingerprints for selecting diverse compound subsets. J Chem Inf Comput Sci 39, 12111225.
Eksterowicz, J. E., Evensen, E., Lemmen, C., Brady, G. P., Lanctot, J. K.,
Bradley, E. K., Saiah, E., Robinson, L. A.,

Reagent Pharmacophore Fingerprints

22.

23.

24.
25.

26.

27.

28.

29.

30.

31.

32.

33.

Grootenhuis, P. D. J., Blaney, J. M. (2002)

Coupling structure-based design with combinatorial chemistry: application of active
site derived pharmaophores with informative library design. J Mol Graph Model 20,
469477.
Good. A. C., Masson, J. S., Green, D. V. S.,
Leach, A. R. (2001) Pharmacophore-based
approaches to combinatorial library design
in (Ghose, A. K., Viswanadhan, V. N.,
eds.) Combinatorial Library Design and
Evaluation. Marcel Dekker, New York, pp.
399428.
McGregor, M. J., Muskal, S. M. (2000)
Pharmacophore fingerprinting. 2. Application to primary library design. J Chem Inf
Comput Sci 40, 117125.
SYBYL Pharmacophore triplet is distributed
by Tripos, Inc., 1699 S. Hanley Rd., St.
Louis, MO 63144, USA.
Schneider, G., Nettekoven, M. (2003)
Ligand-based combinatorial design of selective purinergic receptor (A2A ) antagonists
using self-organizing maps. J Comb Chem 5,
233337.
Turner, D. B., Tyrrell, S. M., Willett, P.
(1997) Rapid quantification of molecular
diversity for selective database acquisition. J
Chem Inf Comput Sci 37, 1822.
Jamois, E. A. (2003) Reagent-based and
product-based computational approaches in
library design. Curr Opin Chem Biol 7,
326330.
Potter, T., Matter, H. (1998) Random
or rational design? Evaluation of diverse
compound subsets from chemical structure
databases. J Med Chem 41, 478488.
McGregor, M. J., Muskal, S. M. (2000)
Pharmacophore fingerprinting. 2. Application to primary library design. J Chem Inf
Comput Sci 40, 117125.
Zheng, W., Cho, S. J., Tropsha, A. (1998)
Rational combinatorial library design. 1.
Focus-2D: a new approach to targeted combinatorial chemical libraries. J Chem Inf Comput Sci 38, 572584.
Leach, A. R., Green, D. V. S., Hann, M. M.,
Judd, D. B., Good, A. C. (2000) Where are
the gaps? A rational approach to monomer
acquisition and selection. J Chem Inf Comput
Sci 40, 12621269.
Gillet, V. J., Willett, P., Bradshaw, J. (1997)
The effectiveness of reactant pools for
generating structurally diverse combinatorial libraries. J Chem Inf Comput Sci 37,
731740.
Chen, H., Brjesson, U., Engkvist, O.,
Kogej, T., Svensson, M. A., Blomberg, N.,
Weigelt, D., Burrows, J. N., Lagne, T.

34.

35.

36.
37.

38.
39.
40.
41.
42.

43.

44.

45.
46.

47.

151

(2009) ProSAR: a new methodology for

combinatorial library design. J Chem Inf
Model 49, 603614.
Kogej, T., Engkvist, O., Blomberg, N.,
Muresan, S. (2006) Multifingerprint based
similarity searches for targeted class compound selection. J Chem Inf Model 46,
12011213.
Bradley, E. K., Miller, J. L., Saiah, E.,
Grootenhuis, P. D. J. (2003) Informative
library design as an efficient strategy to identify and optimize leads: application to cyclindependant kinase 2 antagonists. J Med Chem
46, 43604364.
Python Programming Language Official
Website, http://www.python.org/
Blomberg, N., Cosgrove, D. A., Kenny, P.
W., Kolmodin, K. (2009) Design of compound libraries for fragment screening. J
Comput Aided Mol Des 23, 513525.
Daylight Theory Manual; Daylight Chemical Information Systems, Inc. http://
www.daylight.com/dayhtml/doc/theory/
MDL Available Chemicals Directory
database 2007, Symyx Technologies, Inc.,
Santa Clara, CA 95051, USA.
GVKBio Medchem database 2007, GVK
Biosciences Private Ltd., Hyderabad 500016,
India.
Shannon, C. E., Weaver, W. (1963) The
Mathematical Theory of Communication,
University of Illinois Press, Urbana, IL, USA.
Godden, J. W., Stahura, F. L., Bajorath, J.
(2000) Variabilities of molecular descriptors
in compound databases revealed by Shannon
entropy calculations. J Chem Inf Comput Sci
40, 796800.
Lamb, M. L., Bradley, E. K., Beaton, G.,
Bondy, S. S., Castellino A. J., Gibbons, P.
A., Suto, M. J., Grootenhuis, P. D. J. (2004)
Design of a gene family screening library targeting G-protein coupled receptors. J Mol
Graph Model 23, 1521.
Miller, J. L., Bradley, E. K., Teig, S. L.
(2003) Luddite: an information-theoretic
library design tool. J Chem Inf Comput Sci
43, 4754.
Keating, M. T., Sanguinetti, M. C. (1996)
Molecular genetic insights into cardiovascular disease. Science 272, 681685.
Cavalli, A., Poluzzi, E., De Ponti, F.,
Recanatini, M. (2002) Toward a pharmacophore for drugs including the
long QT syndrome: insights from
a CoMFA study of HERG K(+)
channel blockers. J Med Chem 45,
38443853.
Pearlstein, R. A., Vaz, R. J., Kang, J.,
Chen, X. -L., Preobrazhenskaya, M.,

152

48.

49.

50.

51.

52.

53.

54.

55.

Chen, Engkvist, and Blomberg

Shchekotikhin, A. E., Korolev, A. M.,
Lysenkova, L. N., Miroshnikova, O. V.,
Hendrix,
J.,
Rampe,
D.
(2003)
Characterization
of
HERG
potassium channel inhibition using CoMSiA 3D QSAR and homology modeling
approaches. Bioorg Med Chem Lett 13,
18291835.
Jouyban, A., Soltanpour, S., Soltani, S.,
Chan, H. K., Acree, W. E. (2007) Solubility
prediction of drugs in water-cosolvent mixtures using Abraham solvation parameters. J
Pharm Sci 10, 263277.
Eagan, W. J., Merz, K. M., Baldwin, J.
J. (2000) Prediction of drug absorption
using multivariate statistics. J Med Chem 43,
38673877.
Darvas, F., Dorman, G., Papp, A. (2000)
Diversity measures for enhancing ADME
admissibility of combinatorial libraries. J
Chem Inf Comput Sci 40, 314322.
Bruneau, P. (2001) Search for predictive
generic model of aqueous solubility using
Bayesian neural nets. J Chem Inf Comput Sci
41, 16051616.
Gavaghan, C. L., Arnby, C. H., Blomberg,
N., Strandlund, G., Boyer, S. (2007) Development, interpretation and temporal evaluation of a global QSAR of hERG electrophysiology screening data. J Comput Aided Mol
Des 21, 189206.
Oprea, T. I., Davis, A. M., Teague, S. J.,
Leeson, P. D. (2001) Is there a difference between leads and drugs? A historical perspective. J Chem Inf Comput Sci 41,
13081335.
Oprea, T. I. (2002) Current trends in lead
discovery: are we looking for the appropriate properties? J Comp Aided Mol Des
16, 325.
Reynolds, C. H., Tropsha, A., Pfahler, D.
B., Druker, R., Chakravorty, S., Ethiraj, G.,
Zheng, W. (2001) Diversity and coverage
of structural sublibraries selected using the

56.

57.

58.

59.

60.

61.

62.

63.

SAGE and SCA algorithms. J Chem Inf Comput Sci 41, 14701477.
Jamois, E. J., Hassan, M., Waldman, M.
(2000) Evaluation of reagent-based and
product-based strategies in the design of
combinatorial library subsets. J Chem Inf
Comput Sci 40, 6370.
Szardenings, A. K., Antonenko, V.,
Campbell, D. A., DeFrancisco, N., Ida, S., Si,
L., Sharkov, N., Tien, D., Wang, Y., Navre,
M. (1999) Identification of highly selective
inhibitors of collagenase-1 from combinatorial libraries of diketopiperazines. J Med
Chem 42, 13481357.
Campbell, D. A., Look, G. C., Szardenings,
A. K., Patel, P. V. (2001) US6271232B1;
Campbell, D. A., Look, G. C., Szardenings,
A. K., Patel, P. V. (1999) US5932579A;
Campbell, D. A., Look, G. C., Szardenings,
A. K., Patel, P. V. (1997) WO97/48685A1.
Szardenings, A. K., Harris, D., Lam, S.,
Shi, L., Tien, D., Wang, Y., Patel, D. V.,
Navre, M., Campbell, D. A. (1998) Rational design and combinatorial evaluation of
enzyme inhibitor scaffolds: identification of
novel inhibitors of matrix metelloproteinases.
J Med Chem 41, 21942200.
Martin, Y. C., Kofron, J. L., Traphagen, L.
M. (2002) Do structurally similar molecules
have similar biological activity? J Med Chem
45, 43504358.
Pickett, S. D, McLay I. M., Clark, D. E.
(2000) Enhancing the hit-to-lead properties
of lead optimization libraries. J Chem Inf
Comput Sci 40, 263272.
Gillet, V. J., Khatlib, W., Willett, P., Fleming
P. J., Green, D. V. S. (2002) Combinatorial
library design using a multiobjective genetic
algorithm. J Chem Inf Comput Sci 42,
375385.
Brown, R. D., Hassan, M., Waldman, M.
(2000) Combinatorial library design for
diversity, cost efficiency, and drug-like character. J Mol Graph Model 18, 427437.

Section II
Structure-Based Library Design

Chapter 8
Docking Methods for Structure-Based Library Design
Claudio N. Cavasotto and Sharangdhar S. Phatak
Abstract
The drug discovery process mainly relies on the experimental high-throughput screening of huge
compound libraries in their pursuit of new active compounds. However, spiraling research and development costs and unimpressive success rates have driven the development of more rational, efficient, and
cost-effective methods. With the increasing availability of protein structural information, advancement in
computational algorithms, and faster computing resources, in silico docking-based methods are increasingly used to design smaller and focused compound libraries in order to reduce screening efforts and
costs and at the same time identify active compounds with a better chance of progressing through the
optimization stages. This chapter is a primer on the various docking-based methods developed for the
purpose of structure-based library design. Our aim is to elucidate some basic terms related to the docking technique and explain the methodology behind several docking-based library design methods. This
chapter also aims to guide the novice computational practitioner by laying out the general steps involved
for such an exercise. Selected successful case studies conclude this chapter.
Key words: Structure-based library design, drug discovery, docking, high-throughput screening,
combinatorial chemistry.

1. Introduction
The finding, optimization, and bioevaluation of small molecules
that can interact with therapeutically relevant targets to modulate biological processes is the core of the drug discovery process. So far, this has been mainly dominated by high-throughput
screening (HTS), a hardware technology that allows the rapid
screening of compound libraries to identify potentially active ones
(1, 2). HTS, however, requires a ready source of large and preferably diverse set of compounds to serve as starting points for the
J.Z. Zhou (ed.), Chemical Library Design, Methods in Molecular Biology 685,
DOI 10.1007/978-1-60761-931-4_8, Springer Science+Business Media, LLC 2011

155

156

Cavasotto and Phatak

screening process (3). In the pursuit of increasing the chemical

space for such molecules, combinatorial chemistry, a technology
that systematically mixes and matches various chemical building
blocks to generate chemical libraries was developed (4). Such
libraries were expected to cover the entire chemical space, which
is estimated to consist of 1060100 compounds (5, 6). This combination of HTS and combinatorial chemistry was expected to provide a large and diverse set of lead compounds, enhance shrinking
drug candidate pipelines, and reduce drug discovery time frames
(7, 8). Thus, over the past two decades, combinatorial chemistry
and HTS have been widely used in the modern drug discovery
process with reasonable success (9, 10). Notable improvements
in HTS technologies (e.g., robotics, automated liquid handling
devices, assay miniaturization techniques, signal detectors, and
data processing software) facilitated even further the rapid screening of these libraries to identify promising compounds against validated drug targets (7).
However, after a detailed inspection of the results of these
screening campaigns, it is now evident that neither hit rates (11)
nor hit quality (e.g., unsuitable functional groups, poor solubility of identified hits (12)) obtained from HTS experiments have
shown any significant improvements over time.
On the other hand, screening such huge libraries for every
target is impractical, uneconomical, and inefficient. Part of the
problem is attributed to the quality of compounds used for HTS
(13) (e.g., lack of drug-like properties, compounds with properties unsuitable for biological testing).
As HTS still remains the method of choice to discover novel
hit compounds, researchers focused their attention to the design
and development of appropriate tools to reduce the size of the
chemical libraries to be tested while increasing the quality of the
compounds, what could maximize the chances of identifying hit
compounds amenable to the subsequent lead optimization stages
(14). Such focused libraries were also expected to decrease synthesis, repository management, and screening costs (15). It has also
been observed that the number of drug-like compounds with relevant pharmacological profiles is smaller than the total chemical
space, and hits for a given target are clustered in a finite region
of the compound space (14). Compounds are considered to be
drug-like if they contain functional groups and possess physical
properties consistent with the majority of known drugs (6), along
with acceptable absorption, distribution, metabolism, excretion,
and toxicological (ADMET) profiles to pass through Phase 1 clinical trials (16).
Thus, it is but natural to incorporate the structural information from the target to bias compound selection prior to
the experimental testing (9, 17, 18). With the advent of highthroughput protein crystallography (19), structural genomics

Docking Methods for Structure-Based Library Design

157

consortium projects (20), and developments in homology modeling methods (21), an increasing number of 3D structures of targets are now available for several structure-based drug discovery
applications (e.g., virtual screening (2226), binding mode predictions (2731), and lead optimization (32)). Structural information coded in the characteristics of binding sites, such as
receptor:ligand interaction patterns, can be used to prioritize
compounds for experimental screening using docking methods
(3335), and such exercises have been successful in the past
(36, 37). With the continual development in docking algorithms
and computational resources, structure-based docking methods
will play an increasing important role in compound library
design.
It is timely, then, to review the use of docking methods
for structure-based library design and to understand how best
to implement them in drug discovery. First, we will define and
explain basic concepts and terminology related to structurebased drug design and docking. Next, docking methods for
library design will be presented with brief notes explaining the
practical considerations involved in such an exercise. The chapter will conclude with selected case studies highlighting the
recent successes of docking methods for structure-based library
design.

2. Requirements
The three major requirements for docking-based library design
are basically the same as for high-throughput docking (HTD):
a 3D representation of the target structure (experimental or modeled), a database of compounds in electronic format, and a suitable docking algorithm.
2.1. 3D Structure of
Target

Advancements in crystallography/NMR techniques have resulted

in an exponential increase in the number of protein structures in
publicly available structural databases (e.g., as of September 2009,
the Protein Database Bank PDB contains experimentally solved
3D structural data for 60,000 structures (www.pdb.org)). In
addition, several structural genomics consortiums aim to provide crystal structures across all protein families (38). In case
when experimental structures are not available, techniques such
as homology modeling are often used to build structural models
of other homologous proteins (21).
The structure is thoroughly analyzed to identify putative
binding sites, e.g., by the known location of co-crystallized active
compounds, or by applying in silico methods to identify such sites

158

Cavasotto and Phatak

(e.g., POCKET (39), LIGSITE (40), and SURFNET (41)). The

information obtained from these sites (receptor:ligand interactions (42), physiochemical characteristics of binding site residues,
nature, and size of the binding site) may then be used to restrict
the size of compound libraries by adequate filtering (43).
However, caution must be exercised in using crystal structures as is, since they may contain several inaccuracies (44). Low
resolution of electron density maps and crystallization conditions
different than those maintained in biological assays may introduce
errors in the final structure (45). Assumptions made by the crystallographer may result in errors in the orientation of side chains
(e.g., asparagine and glutamine flips, histidine tautomerization)
or proper location and conformation of the ligand (45). In addition, the crystal structure represents just one snapshot of a highly
dynamic conformational equilibrium ensemble. The impact of
protein flexibility in docking is not yet fully understood (46),
which further undermines the applicability of the crystal structure
as is for structure-based drug discovery.
To prepare the protein for a docking procedure, the following considerations are usually taken into account:
(a) Removal of the ligand or co-factors if any, from the cocrystallized protein complex.
(b) PDB structures may contain coordinates of several water
molecules. Water molecules play an important role in ligand
binding by mediating hydrogen bonds between the protein
and the ligand or by being displaced by the ligand upon
binding (47, 48). If available, several crystal structures of
the target are investigated to study water positions. Only
those which are highly conserved or tightly bound to the
receptor are retained (49, 50).
(c) Inspection and correction of any error in the crystal structures, such as incorrect bond orders and missing residues
(particularly in the binding site).
(d) Crystal structures lack hydrogens. It is necessary to check
the protonation state of receptor residues, add hydrogens,
and perform energy minimization.
(e) Check for asparagine and glutamine flips and for the correct
histidine tautomerization states.
Sequence identity and the quality of sequence alignment play
an important role in their accuracy of homology models. Exceptions to this rule exist, as in the case of Class A G protein-coupled
receptors where structural rather than sequence similarity drives
the modeling (21). As a word of caution, high sequence identities may mask the dissimilarities in certain flexible regions, what
may render the model less useful for drug discovery applications.
The choice of template and inefficient refinement methods are
the other sources of errors in homology modeling (21).

Docking Methods for Structure-Based Library Design

2.2. Compound
Collections

159

Many commercial vendors and academic labs provide collection of compounds or fragments (Sigma-Aldrige, Available Chemicals Directory (ACD) (http://www.symyx.com),
Maybridge (www.maybridge.com), TCI America (http://www.
tciamerica.com), ChemDB (51) (http://cdb.ics.uci.edu), ZINC
(52), ChemBridge (www.chembridge.com); cf (53) for a review
on public accessible chemical databases), in computationally
readably formats like sdf or mol2 (42). Along with the
experimental combinatorial library design process mentioned
earlier, several software tools, such as CombiLibMaker in
Sybyl (www.tripos.com) and QuaSAR-Combigen from CCG
(www.chemcomp.com), were developed to enumerate and predict the 2D/3D structures of compounds using chemical fragments without the expensive and tedious experimental part. On
the other hand, pharmaceutical companies have historically maintained huge compound libraries and continue to add compounds
with novel chemistry to these collections. The size of such compound collections is estimated to be 106 compounds (54).
These compound libraries or their constitutive parts (reagents,
fragments) form the source of inputs for docking methods for
structure-based library design.
However, these compounds and the fragments are not without their intrinsic problems and should not be used as is. Some
examples of potentially problematic compounds include those
with chemically reactive groups, dyes, and fluorescent compounds which interfere with assays, frequent hitters/promiscuous
binders, and inorganic complexes (55). It is important, then, to a
priori filter out such compounds or reagents which are practically
useless from a drug discovery point of view.
Some of the common steps toward preparing and filtering
chemical compound libraries include
(a) Removing compounds with salts, counter ions, chemically
reactive groups (e.g., metal chelators), undesirable atoms or
functional groups, inorganic compounds, and duplicates.
(b) Generation of correct tautomers, protonation, and
stereoisomeric states for each compound. Eventually, several states for a given compound could be generated and
kept in the final library.
(c) Filtering compounds based on drug-like or lead-like physiochemical properties or other in-house scoring schemes
(e.g., Lipinskis rule of 5 (56), lead-like filters suggested
by Oprea et al. (57)).
In cases where fragments, rather than compounds, are used
to design libraries, the fragments may be filtered based on
the nature of the binding site and their ability to adhere to
existing chemistry protocols with respect to their attachments
to templates or other fragments (58). Other filtering criteria

160

Cavasotto and Phatak

include excluding fragments with hydrolysable groups (e.g., sulfonyl halides, anhydride aliphatic ester) and potential cytotoxic
groups like thiourea and cyclohexanone (55).
2.3. Docking
Algorithms and
Methods

The third important requirement is the selection of an appropriate

docking algorithm. In brief, a docking-based virtual screening (or
HTD) consists of the following steps:
(a) Positioning compounds into the binding site of the target
via a process called docking.
(b) Assignment of a score to each compound which represents
the likelihood of binding to the target (scoring). Please refer
to (33, 34, 59, 60) for tutorials on docking and structurebased virtual screening.
(c) Prioritization of a subset of compounds based on scores and
other post-screening criteria, like the ability to mimic key
receptor:ligand interactions.
Docking programs routinely incorporate compound flexibility. However, incorporating receptor flexibility in docking procedures is still a major hurdle (61). Of late, several attempts have
been made for this purpose (cf (46, 62) for review).
One may choose from several docking methods, but it should
be noted that a thorough understanding of the principles underlying the program is important to achieve meaningful results (7).
A systematic review of docking methods or programs is not the
focus of this chapter, but their use in the context of structurebased library design. The interested reader may refer to (59, 63,
64) for reviews on docking programs. It should be stressed that
none of the current docking programs is universally applicable
(6567). Thus, instead of using the default settings of the programs, one could develop, test, and validate a protocol that optimizes the use of the program parameters for a given target.
After HTD, one may still be left with a large number of compounds. In such cases, and on top of filtering according to key
ligand:receptor interaction patterns if available, various datamining techniques may be applied to narrow down and identify
diverse compounds as possible (68). Clustering algorithms (e.g.,
exclusion sphere, k-nearest neighbor, JarvisPatrick) provide an
easy way to overview different chemical classes in the result set
and choose representative compounds within each class for experimental testing (3). Neural networks, support vector machinebased approaches are used to predict target-class likeliness (e.g.,
for GPCRs and kinases (69)).
Docking methods for library design can be broadly classified
into two strategies:
Sequential docking, where pre-enumerated compounds are
docked into the receptor binding site, scored, ranked, and
selected for further experimental testing; the conformational

Docking Methods for Structure-Based Library Design

161

space of the compounds is explored by flexible compound

docking or rigid docking of pre-generated conformers of
each compound.
Fragment-based design, where the constituents of the compounds (scaffolds and functional groups/substituents) are
docked in the binding site and then linked together to build
combinatorial libraries.
The latter strategy has two flavors (70).
(a) Seed and grow: A pre-selected scaffold is first docked into
the binding site. Each scaffold pose is scored and only topranking poses are considered for subsequent stages. Substituents are then attached to each selected scaffold pose,
optimized, and scored (71). The top-scoring substituents
are then used to build a combinatorial library. The advantage of this method is that it avoids the combinatorial
explosion problem by narrowing down the number of substituents used to build libraries and including knowledge
from the binding site of the target structure. This approach
is depicted in Fig. 8.1.

Fig. 8.1. Schematic depiction of the seed and grow docking approach for structurebased library design. The programs that use this approach include CombiDOCK and
PRO_SELECT.

162

Cavasotto and Phatak

(b) Dock and link: The substituent groups are docked to the
interacting sites in the binding pocket, scored, and then
linked to each other based on chemistry constraints (70).
This method, as illustrated in Fig. 8.2, attempts to take
advantage of the known significant interactions within the
binding site to bias the final compound library.

Fig. 8.2. Schematic depiction of the dock and link docking approach. The program
BUILDER v.2 is based on this approach.

Notes:
For fragment-based methods, one needs to consider some
important issues.
(a) For the seed and grow method, the orientation of the scaffold is highly critical. Any errors at this stage may render the
results at later steps irrelevant (71).
(b) There must be a ready-to-use synthetic protocol to build
these libraries based on the scaffold and fragments used.
(c) Although the seed and grow method reduces the number
of compounds as compared to a full library enumeration,
the availability of large number of fragments may still result
in a huge number of compounds. Further filtering steps,
diversity analysis exercises may be required to choose a final
subset of compounds.
(d) It is important to have diverse fragment libraries to maximize the chances of library diversity.
(e) In the case of the dock and link method, though it is likely to
have fragments satisfying key interactions within the binding
site, the final compound may not be amenable to synthesis
(71).

Docking Methods for Structure-Based Library Design

163

(f) It should be noted that docking programs may introduce

errors due to the inherent inaccuracies of force fields, sampling, and scoring functions (7, 33, 72).

3. Docking
Methods for
Structure-Based
Library Design

CombiDock is one of the first programs developed to design

structure-based combinatorial libraries (73). It is based on a simple variation of the original DOCK algorithm (74). In brief,
DOCK generates a negative image of the receptor binding site
which is represented by spheres. The algorithm searches for internal distance matches between subsets of ligand atoms and spheres
generated from the earlier stage. Based on a match, ligand atoms
are placed and scored using force field or empirical functions that
estimate the interaction energies. CombiDock tweaks this original algorithm such that only the scaffold atoms are used instead of
the ligand. The scaffold is oriented in different conformations in
the binding site and its atoms are matched against receptor binding site spheres. In the next step, all fragments/functional groups
are attached at every individual attachment point and interaction
scores are calculated for the scaffold and each attached fragment.
The fragments with higher scores are then combined to form individual compounds. The best combinations are scanned for any
intermolecular clashes with the receptor and saved. The method
reduces the combinatorial process to a simple numerical addition
of fragment scores to speed up library design (73).
Another tool, which combines combinatorial chemistry and
fragment-based docking methods to rationally restrict the size
of combinatorial libraries using structural restraints from binding site is PRO_SELECT (SELECT = Systematic Elaboration
of Libraries Enhanced by Computational Techniques) (75). The
underlying assumptions of this method are that the template
fragment and the receptor are considered rigid and each individual substituent can be assessed independently of each other.
PRO_SELECT also guides the library design process to build
compounds that are accessible to specified synthetic routes, eliminating the uncertainties associated with synthetic feasibility of virtual compound libraries.
The PRO_SELECT methodology consists of three major
parts which are explained in brief as follows:
I. Designing specifications for the target and the molecular
templates (scaffold)
a. The target is prepared on a protocol similar to the
general steps described earlier and analyzed for possible interaction features represented as vectors, which

164

Cavasotto and Phatak

denote favorable position and direction for hydrogen

bond interactions with the active site, and points, which
denote positions of favorable hydrophobic contacts with
the active site.
b. Using the structural knowledge of the receptor, template/s are chosen. It is desirable that these templates
have multiple attachment points to attach several substituent groups and restricted conformational freedom
to limit the number of alternative template positions
within the binding pocket.
c. A design model is then developed which contains the
vectors and points along with link sites, which are the
positions on the template where a potential substituent
group may be attached.
d. The templates are placed in the binding site using docking protocols based on molecular mechanics energy calculations (76, 77) or geometric positioning upon interaction sites (78).
II. Substituent/functional group selection
a. Databases of commercially available fragments (e.g.,
ACD) are used to search for possible substituent
groups.
b. The fragments are computationally screened using
PRO_LIGAND (79) and only those that can form good
molecular interactions based on the original template
position in the pocket (hydrogen bond interactions,
lipophilic interactions) are selected.
c. Possible bioisoteric (functional groups possessing similar chemical properties) replacements are searched in
the pursuit of novel compounds.
d. The substituents for each position are minimized using
a molecular mechanics energy function where the receptor and template are held rigid, scored using the function developed by Bohm (78), and ranked.
III. Combinatorial enumeration:
a. The shortlisted compounds from the earlier stage are
saved in a list.
b. It is recommended to reduce the size of the list by
excluding structures with high strain energies, bad
chemistries or geometries, and poor Bohm scores.
c. The structures may then be clustered based on 2D
chemical functionality.
d. Finally, via combinatorial enumeration, a final compound library is generated.

Docking Methods for Structure-Based Library Design

165

DREAM++ (Docking and Reaction programs using Efficient

seArch Methods) is a suite of programs (ORIENT++, REACT++,
SEARCH++) developed to design chemical libraries by incorporating information from known chemical reactions and receptor
active sites (80). The advantage of using well-studied organic
reactions is that only synthetically accessible product compounds
are generated in the final stage. The procedure begins by docking anchor parts or scaffolds into the binding site. These are
then minimized, scored, and analyzed based on binding modes
and other user-defined criteria. Functional groups from vendor
libraries are virtually reacted with reagents using knowledge from
a wide variety of organic reactions (e.g., amide bond formation,
urea formation, reductive animation, alkylation, and ester formation) and are systematically combined to generate compounds.
The conformational space of these compounds is explored and
these steps are repeated until a complete library is produced. The
generated library may then be visually inspected to study putative binding modes and offer further insights prior to selection
for experimental results.
The program BUILDER v.2 (81) belongs to the dock and link
category, where the importance is given to satisfying key interactions within the receptor binding site using fragments and then
linking these fragments to form product compounds. Prior to
the docking of fragments, the binding site is thoroughly investigated to identify hot spots or sites of potentially strong interaction with the receptor. The program DOCK (74) is used to
place fragments or functional groups in the hot spots. By using
a lattice around the protein, any two atoms of different fragments are connected via a set of lattice points. The set of such
points being termed as generic paths. These paths are generated using a modified breadth-first search algorithm (a graph
search algorithm which begins at the root node and explores all
the neighboring nodes). The points on these generic paths are
considered to be atoms. Using three atoms in the path and their
bond angles, the putative hybridization state of that atom is calculated. A pre-determined list, GOODLIST, contains a mapping of
chemically reasonable functional groups (e.g., carbonyl, amide,
thioester, and phenyl) for several of such three-atom combinations. Using the GOODLIST and the three-atom combinations,
specific atom types are added to the atoms. BUILDER uses the
SHAKE algorithm (82) to check for correct atom-type combinations, bond lengths, angles, and bump checks (steric clashes)
against the receptor. Finally the paths are then reexamined to
generate linker groups or bridges. Preference is given to embed
a ring structure; however, other simpler and chemically synthesizable connecting groups are also considered. The bridges are
expected to not have any strong contribution to binding. These

166

Cavasotto and Phatak

bridges along with the original fragments are then attached to

generate a product compound.
Another docking-based program developed by Sprous et al.,
OptiDock (83), attempts to exploit the common cores in a preenumerated combinatorial library. Instead of docking fragments
or scaffolds, a subset of compounds spanning the structural space
of the compound library are chosen and docked using the program FlexX (84). The binding mode for each compound is analyzed and distinctly different modes are shortlisted and the functional groups of these compounds are stripped. The core position
is held constant, functional groups are attached, and interaction
energies are calculated for each compound.
Recently, Zhou et al. developed a novel method termed as
basis products (BPs) (85), which exploits the redundancy of fragments in a combinatorial library. The premise of this method is
that all functional groups in a combinatorial library can be completely represented by a selected product subset of the library.
This subset of compounds is called as basis products (BPs),
which are formed by combining the smallest reactants (functional
groups) of all reaction components except one. The remaining
reactant is used against all viable reactants for a particular reaction while the other reactants are held constant. Thus for a two
component reaction A + B AB, the entire library will consist
of all the combinations of reactants A and B. In case of BPs, two
capping molecules As and Bs are pre-selected with the smallest A
and B, respectively. These capping molecules are then combined
by changing only one component on the other side to generate
two sub-libraries {AsB} and {ABs}. The sum of these libraries is
much smaller than the single set of the entire library {AB}. Thus,
every virtual library compound can be represented by a smaller
set of BPs. Given a target, BPs can be docked using various docking programs. Based on the scores, the BPs are selected for the
follow-up process, which involves designing libraries by using the
reactants corresponding to the variable components of the BP
hits among other strategies. To further improve the efficiency of
the method, the BPs themselves may be filtered based on physiochemical properties to reduce the number of BPs for the docking process. The algorithm was tested in a comparison-type study
(85) where an entire virtual library (34,000 compounds) and
a smaller subset (1225 compounds identified by BPs and hit
follow-up library) were both docked to the active site of dihydrofolate reductase and the top-ranked compounds were checked. In
both cases, the top 350 ranked compounds were the same. Thus,
in this case, it was shown that a smaller but focused library can
achieve comparable results as compared to docking entire virtual
libraries.
Several other programs like COMBISMOG (86),
CombiGlide (www.schrodinger.com), and COMBIBUILD

Docking Methods for Structure-Based Library Design

167

(http://mdi.ucsf.edu/CombiBUILD.html) have been developed for the purpose of library design. Please refer to Table 8.1
for a list of programs listed in this chapter. However, it should
be noted that though useful, most of the programs are neither
easy to implement nor use as is (87). As a result, these methods
have found limited applicability in the scientific community.
On the other hand, several commercial library vendors like
Cerep (www.cerep.fr), Asinex (www.asinex.com), and Enamine
(www.enamine.net) offer target-focused libraries using dockingbased protocols. A list of the docking methods/software tools
mentioned in this chapter can be found in Table 8.1.

Table 8.1
Docking-based programs for library design
Method/program

Description

Refs

CombiDock

Tweaks the DOCK algorithm to identify suitable scaffold orientations in the binding
pocket. Proceeds using the seed and grow
approach to design combinatorial libraries
Combines combinatorial chemistry and
fragment-based docking methods to design
structure-based libraries

(73)

PRO_SELECT

DREAM++

Builder v.2

Designs chemical libraries incorporating

information from known chemical reactions and receptor binding sites
Uses the dock and link strategy to link relevant
fragments, which satisfy key receptor:ligand
interactions to form product compounds

(75)

(80)

(81)

OptiDOCK

Uses the seed and grow strategy to first

dock representative compounds spanning
the chemical space of the library and subsequently use an optimal core for library enumeration

(83)

Basis products
(BPs)

Exploits the redundancy of fragments in a

combinatorial library and identifies a small
subset of compounds (BPs) which represent the entire virtual library. BPs are
docked, scored, and used for final library
enumeration

(85)

CombiGlide

Combines docking algorithms and core

hopping technologies to design focused
libraries

www.schrodinger.com

CombiSMoG

Uses a Monte Carlo ligand growth algorithm

and knowledge-based potentials to combine combinatorial and rational strategies
for generating biased compound libraries

(86)

168

Cavasotto and Phatak

4. Applications
This section will highlight a few applications of docking-based
methods for library design.
The CombiDOCK algorithm was applied to design a
structure-based library for cathepsin D protease using a hydroxyethylamine scaffold. This scaffold has three attachment points.
Ten fragments for each site were chosen and incorporated in the
final library design. The 1000 compounds were filtered to check
inaccuracies in bond geometries to give 750 compounds which
were synthesized and assayed for experimental testing. Results
indicated that this library had an enrichment factor (EF) of 2.5,
whereas a completely random ranking would result in an EF of
1.0. (73) The EF is the ratio between the probability of finding
a true ligand in a filtered sub-library compared to the probability
of finding a ligand at random.
The PRO_SELECT method was applied to design an
inhibitor library for thrombin, a key serine protease. The crystal structure of thrombin includes a covalently bound inhibitor,
tri-peptide PPACK. L-proline, the centrally located portion of
PPACK was chosen as the template and its alternate locations
were generated by docking/modeling a noncovalently bound
analogue of PPACK. Analysis of the binding site revealed the
requirements for a hydrogen bond donor and a hydrophobic
group at either ends of the template. A 3D database search of
potential fragment binders based on the analysis of the binding
site resulted in over 400,000 hits. PRO_SELECT method was
able to drastically reduce the number of fragments to 17, which
were then used to build a chemical library. Over 30 molecules
were then synthesized, of which at least 50% showed micromolar
activities (75).
In another study Head et al. used a docking-based method
to design a library of potentially novel inhibitors for caspases
3 and 8, a key regulator of apoptosis (88). The authors chose
thiomethylketone as a scaffold for this study, as it is a common
denominator of a class of compounds inhibiting caspases 3 and 8.
Two attachment points on the thiomethylketone scaffold were
identified. The ketone group is postulated to covalently bind with
the catalytic cysteine. Thus from Fig. 8.3 it is seen that the R
group points away from the S2 binding pocket. Hence a small
number of reagents (8) were fixed for R based on availability
and ease of synthesis. To identify potential functional groups for
the other attachment point (R), roughly 7000 monoacid reagents
from the ACD database were selected for combinatorial docking.
First, a simplified thiomethylketone with R and R set to methyl
was docked in the binding pocket to identify initial template

Docking Methods for Structure-Based Library Design

169

Fig. 8.3. (Bottom): Thiomethylketone D of (88) is used as an example of a caspase 3

inhibitor designed via a docking-based library generation protocol. S1 and S2 denote the
interaction sites within the binding pocket of caspase 3. (Top right): The thiomethylketone scaffold that is used as the starting point for library design. (Top left): The eight
R-groups used to attach to the R attachment point of the scaffold.

locations. Next, the eight reagents for the R point and 7000
monoacids for the R points were combinatorially attached to the
templates, docked, and scored. Two criteria were used to obtain
the final reagents for the R group: (1) docking scores and (2) distance filters based on the experimental data of isatin-based compounds and crystal structures of other caspases. Based on these
results approximately 150 reagents were selected per caspase and
roughly 10% of these reagents underwent full conformational
sampling. As the array size for synthesis was 96, only 12 reagents
for the R group (seven for caspase 3, three for caspase 8, and
three common for both) were selected based on visual inspection
of the predicted binding modes. Sixty-one compounds were synthesized and tested. Five of the 61 compounds tested against caspase 3 and two compounds against caspase 8 showed micromolar
activity. Interestingly, a homology model of caspase 8 was used
for this study, which clearly indicates the usefulness of homology
modeling in structure-based library design.
Decornez et al. used a generalized kinase model and a combination of 2D (fingerprint based similarity) and 3D methods
(docking) to develop a kinase family focused library (15). The
authors used 2800 kinase inhibitors compounds as a reference
for a 2D search of their in-house database of 260 K compounds

170

Cavasotto and Phatak

which resulted in 3135 compounds. As 2D methods are grossly

inadequate to incorporate receptor information, a docking protocol was developed using the crystal structure of PKA (PDB code
1BX6) and the software Glide (www.schrodinger.com). Since the
goal of the project was to design a generic kinase-specific library,
the authors mutated several residues of the crystal structure to
avoid any bias in the eventual compound library. The 3100
compounds were then docked, scored, and the top 170 compounds with significant 2D similarity to known inhibitors and
3D binding characteristics were submitted for biochemical screening. The identified hits were similar or analogues of p38, tyrosine
kinase, and PKC kinases.
Zhao et al. implemented a structure-based docking protocol
to narrow down 500 compounds from a database of 57 K compounds in their pursuit of FKBPs inhibitors (89). A novel scaffold
was designed using the information obtained from the binding
mode analysis of a known weak binder. To avoid any scoring function shortcomings, three scoring functions were used to select
the 500 compounds. Of these, 43 were synthesized and tested
to identify one potent inhibitor in a mouse peripheral synthetic
nerve model.

5. Conclusions
Despite the initial promise, advancements in HTS methods and
combinatorial chemistry have so far failed to improve the success
rates of drug discovery programs. Since the experimental screening of these gigantic libraries is costly and time consuming, it
is of utmost importance to rationally, efficiently, and economically explore the available chemical space of compounds in order
to design smaller and focused compound libraries for experimental evaluation. Several docking-based methods make use of the
increasing availability of structural information of drug targets to
a priori filter out those compounds that are unlikely to bind to the
target. This chapter highlights several of such docking methods
used in library design, together with their application to actual
cases.
References
1. Mayr, L. M., Fuerst, P. (2008) The future of
high-throughput screening. J Biomol Screen
13, 443448.
2. Entzeroth, M. (2003) Emerging trends in
high-throughput screening. Curr Opin Pharmacol 3, 522529.
3. Schnecke, V., Bostrom, J. (2006) Computational chemistry-driven decision making

in lead generation. Drug Discov Today 11,

4350.
4. Boldt, G. E., Dickerson, T. J., Janda, K. D.
(2006) Emerging chemical and biological
approaches for the preparation of discovery
libraries. Drug Discov Today 11, 143148.
5. Bohacek, R. S., McMartin, C., Guida, W. C.
(1996) The art and practice of structure-

Docking Methods for Structure-Based Library Design

6.
7.

8.
9.
10.

11.

12.

13.
14.
15.

16.

17.
18.
19.

20.

based drug design: a molecular modeling perspective. Med Res Rev 16, 350.
Walters, W. P., Stahl M. T., Murcko, M. A.
(1998) Virtual screening an overview. Drug
Discov Today 3, 160178.
Phatak, S. S., Stephan, C. C., Cavasotto,
C. N. (2009) High-throughput and in silico screenings in drug discovery. Expert Opin.
Drug Discov 4, 947959.
Keseru, G. M., Makara, G. M. (2006) Hit
discovery and hit-to-lead approaches. Drug
Discov Today 11, 741748.
Macarron, R. (2006) Critical review of the
role of HTS in drug discovery. Drug Discov
Today 11, 277279.
Fox, S., Farr-Jones, S., Sopchak, L., Boggs,
A., Nicely, H. W., Khoury, R., Biros, M.
(2006) High-throughput screening: update
on practices and success. J Biomol Screen 11,
864869.
Keseru, G. M., Makara, G. M. (2009) The
influence of lead discovery strategies on the
properties of drug candidates. Nat Rev Drug
Discov 8, 203212.
Lipkin, M. J., Stevens, A. P., Livingstone,
D. J., Harris, C. J. (2008) How large does
a compound screening collection need to be?
Comb Chem High Throughput Screening 11,
482493.
Nestler, H. P. (2005) Combinatorial chemistry and fragment screening Two unlike
siblings? Curr Drug Discov Tech 2, 112.
Diller, D. J., Merz, K. M., Jr. (2001) High
throughput docking for library design and
library prioritization. Proteins 43, 113124.
Decornez, H., Gulyas-Forro, A., Papp, A.,
Szabo, M., Sarmay, G., Hajdu, I., Cseh,
S., Dorman, G., Kitchen, D. B. (2009)
Design, selection, evaluation of a general
kinase-focused library. ChemMedChem 4,
12731278.
Lipinski, C. A. (2000) Drug-like properties
and the causes of poor solubility and poor
permeability. J Pharmacol Toxicol Methods 44,
235249.
Schnur, D. M. (2008) Recent trends in
library design: rational design revisited.
Curr Opin Drug Discov Devel 11, 375380.
Villar, H. O., Koehler, R. T. (2000) Comments on the design of chemical libraries for
screening. Mol Divers 5, 1324.
Manjasetty, B. A., Turnbull, A. P., Panjikar,
S., Bussow, K., Chance, M. R. (2008) Automated technologies and novel techniques to
accelerate protein crystallography for structural genomics. Proteomics 8, 612625.
Gileadi, O., Knapp, S., Lee, W. H., Marsden, B. D., Muller, S., Niesen, F. H.,
Kavanagh, K. L., Ball, L. J., von Delft, F.,

21.

22.

23.

24.

25.

26.

27.

28.

29.

30.

171

Doyle, D. A., Oppermann, U. C., Sundstrom, M. (2007) The scientific impact of

the Structural Genomics Consortium: a protein family and ligand-centered approach to
medically-relevant human proteins. J Struct
Funct Genomics 8, 107119.
Cavasotto, C. N., Phatak, S. S. (2009)
Homology modeling in drug discovery: current trends and applications. Drug Discov
Today 14, 676683.
Cavasotto, C. N., Orry, A. J., Murgolo, N.
J., Czarniecki, M. F., Kocsi, S. A., Hawes, B.
E., ONeill, K. A., Hine, H., Burton, M. S.,
Voigt, J. H., Abagyan, R. A., Bayne, M. L.,
Monsma, F. J., Jr. (2008) Discovery of novel
chemotypes to a G-protein-coupled receptor
through ligand-steered homology modeling
and structure-based virtual screening. J Med
Chem 51, 581588.
Hong, T. J., Park, H., Kim, Y. J., Jeong,
J. H., Hahn, J. S. (2009) Identification of
new Hsp90 inhibitors by structure-based virtual screening. Bioorg Med Chem Lett 19,
48394842.
Brozic, P., Turk, S., Lanisnik Rizner, T.,
Gobec, S. (2009) Discovery of new inhibitors
of aldo-keto reductase 1C1 by structurebased virtual screening. Mol Cell Endocrinol
301, 245250.
Park, H., Bhattarai, B. R., Ham, S. W., Cho,
H. (2009) Structure-based virtual screening
approach to identify novel classes of PTP1B
inhibitors. Eur J Med Chem 44, 32803284.
Heinke, R., Spannhoff, A., Meier, R., Trojer, P., Bauer, I., Jung, M., Sippl, W. (2009)
Virtual screening and biological characterization of novel histone arginine methyltransferase PRMT1 inhibitors. ChemMedChem 4,
6977.
Wang, Q., Wang, J., Cai, Z., Xu, W. (2008)
Prediction of the binding modes between
BB-83698 and peptide deformylase from
Bacillus stearothermophilus by docking and
molecular dynamics simulation. Biophys Chem
134, 178184.
Padgett, L. W., Howlett, A. C., Shim, J.
Y. (2008) Binding mode prediction of conformationally restricted anandamide analogs
within the CB1 receptor. J Mol Signal 3, 5.
Zampieri, D., Mamolo, M. G., Vio, L.,
Banfi, E., Scialino, G., Fermeglia, M., Ferrone, M., Pricl, S. (2007) Synthesis, antifungal and antimycobacterial activities of new
bis-imidazole derivatives, and prediction of
their binding to P450(14DM) by molecular docking and MM/PBSA method. Bioorg
Med Chem 15, 74447458.
Monti, M. C., Casapullo, A., Cavasotto,
C. N., Napolitano, A., Riccio, R. (2007)

172

31.

32.

33.

34.

35.

36.
37.

38.
39.

40.

41.

42.

Cavasotto and Phatak

Scalaradial, a dialdehyde-containing marine
metabolite that causes an unexpected noncovalent PLA2 Inactivation. Chembiochem 8,
15851591.
Diaz P., Phatak, S. S., Xu, J., Fronczek,
F. R., Astruc-Diaz, F., Thompson, C. M.,
Cavasotto, C. N., Naguib, M. (2009) 2,3Dihydro-1-benzofuran derivatives as a series
of potent selective cannabinoid receptor 2
agonists: design, synthesis, and binding mode
prediction through ligand-steered modeling.
ChemMedChem 4, 16151629.
Andricopulo, A. D., Salum, L. B., Abraham,
D. J. (2009) Structure-based drug design
strategies in medicinal chemistry. Curr Topics Med Chem 9, 777790.
Cavasotto, C. N., Orry, A. J. (2007) Ligand
docking and structure-based virtual screening in drug discovery. Curr Top Med Chem
7, 10061014.
Kitchen, D. B., Decornez, H., Furr, J. R.,
Bajorath, J. (2004) Docking and scoring in
virtual screening for drug discovery: methods and applications. Nat Rev Drug Discov
3, 935949.
Cavasotto, C. N., Ortiz, M. A., Abagyan, R.
A., Piedrafita, F. J. (2006) In silico identification of novel EGFR inhibitors with antiproliferative activity against cancer cells. Bioorg
Med Chem Lett 16, 19691974.
Klebe, G. (2006) Virtual ligand screening:
strategies, perspectives and limitations. Drug
Discov Today 11, 580594.
Zoete, V., Grosdidier, A., Michielin, O.
(2009) Docking, virtual high throughput
screening and in silico fragment-based drug
design. J Cell Mol Med 13, 238248.
Marsden, R. L., Orengo, C. A. (2008) Target selection for structural genomics: an
overview. Methods Mol Biol 426, 325.
Levitt, D. G., Banaszak, L. J. (1992)
POCKET: a computer graphics method for
identifying and displaying protein cavities
and their surrounding amino acids. J Mol
Graph 10, 229234.
Hendlich, M., Rippmann, F., Barnickel,
G. (1997) LIGSITE: automatic and efficient detection of potential small moleculebinding sites in proteins. J Mol Graph Model
15, 359363, 389.
Laskowski, R. A. (1995) SURFNET: a program for visualizing molecular surfaces, cavities, intermolecular interactions. J Mol Graph
13, 323330, 307328.
Balakin, K. V., Kozintsev, A. V., Kiselyov, A. S., Savchuk, N. P. (2006) Rational
design approaches to chemical libraries for hit
identification. Curr Drug Discov Technol 3,
4965.

43. Orry, A. J., Abagyan, R. A., Cavasotto,

C. N. (2006) Structure-based development
of target-specific compound libraries. Drug
Discov Today 11, 261266.
44. Brown, E. N., Ramaswamy, S. (2007) Quality of protein crystal structures. Acta Crystallogr D Biol Crystallogr 63, 941950.
45. Davis, A. M., St-Gallay, S. A., Kleywegt, G. J.
(2008) Limitations and lessons in the use of
X-ray structural information in drug design.
Drug Discov Today 13, 831841.
46. Cavasotto, C. N., Singh, N. (2008) Docking and high throughput docking: successes
and the challenge of protein flexibility. Curr
Comput Aided Drug Design 4, 221234.
47. Sousa, S. F., Fernandes, P. A., Ramos, M.
J. (2006) Protein-ligand docking: current
status and future challenges. Proteins 65,
1526.
48. Li, Z., Lazaridis, T. (2007) Water at
biomolecular binding interfaces. Phys Chem
Chem Phys 9, 573581.
49. Mancera, R. L. (2007) Molecular modeling
of hydration in drug design. Curr Opin Drug
Discov Devel 10, 275280.
50. Corbeil, C. R., Moitessier, N. (2009) Docking ligands into flexible and solvated macromolecules. 3. Impact of input ligand conformation, protein flexibility, and water
molecules on the accuracy of docking programs. J Chem Inf Model 49, 9971009.
51. Chen, J., Swamidass, S. J., Dou, Y., Bruand,
J., Baldi, P. (2005) ChemDB: a public
database of small molecules and related
chemoinformatics resources. Bioinformatics
21, 41334139.
52. Irwin, J. J., Shoichet, B. K. (2005) ZINCa
free database of commercially available compounds for virtual screening. J Chem Inf
Model 45, 177182.
53. Williams, A. J. (2008) Public chemical compound databases. Curr Opin Drug Discov
Develop 11, 393404.
54. Drie, J. H. (2005) Pharmacophore-based
virtual screening: a practical perspective, in
(Alvarez, J., Shoichet, B., eds.) Virtual
Screening in Drug Discovery. CRC Press,
Boca Raton, FL, pp. 157205.
55. Oprea, T. I., Bologa, C. G., Olah, M. M.
(2005) Compound selection for virtual
screening, in Virtual screening in Drug Discovery (Alvarez, J., Shoichet, B., eds.), CRC
Press, Boca Raton, FL, pp. 89106.
56. Lipinski, C. A., Lombardo, F., Dominy, B.
W., Feeney, P. J. (1997) Experimental and
computational approaches to estimate solubility and permeability in drug discovery and
development settings. Adv Drug Del Rev 23,
325.

Docking Methods for Structure-Based Library Design

57. Oprea, T. I. (2002) Current trends in lead
discovery: are we looking for the appropriate properties? J Comput Aided Mol Des 16,
325334.
58. Hubbard, R. E. (2008) Fragment
approaches
in
structure-based
drug
discovery. J Synchrotron Radiat 15,
227230.
59. Kroemer, R. T. (2007) Structure-based drug
design: docking and scoring. Curr Protein
Pept Sci 8, 312328.
60. Barril, X., Hubbard, R. E., Morley, S. D.
(2004) Virtual screening in structure-based
drug discovery. Mini Rev Med Chem 4, 779
791.
61. Teague, S. J. (2003) Implications of protein
flexibility for drug discovery. Nat Rev Drug
Discov 2, 527541.
62. B-Rao, C., Subramanian, J., Sharma, S. D.
(2009) Managing protein flexibility in docking and its applications. Drug Discov Today
14, 394400.
63. Dias, R., de Azevedo, W. F., Jr. (2008)
Molecular docking algorithms. Curr Drug
Targets 9, 10401047.
64. Sperandio, O., Miteva, M. A., Delfaud,
F., Villoutreix, B. O. (2006) Receptorbased computational screening of compound databases: the main dockingscoring engines. Curr Protein Pept Sci 7,
369393.
65. Stahl, M., Rarey, M. (2001) Detailed analysis of scoring functions for virtual screening.
J Med Chem 44, 10351042.
66. Perola, E., Walters, W. P., Charifson, P. S.
(2005) An analysis of critical factors affecting docking and scoring, in (Alvarez, J.,
Shoichet, B., eds.) Virtual screening in drug
discovery. CRC Press, Boca Raton, FL, pp.
4785.
67. Warren, G. L., Andrews, C. W., Capelli, A.
M., Clarke, B., LaLonde, J., Lambert, M.
H., Lindvall, M., Nevins, N., Semus, S. F.,
Senger, S., Tedesco, G., Wall, I. D., Woolven, J. M., Peishoff, C. E., Head, M. S.
(2006) A critical assessment of docking programs and scoring functions. J Med Chem 49,
59125931.
68. Waszkowycz, B. (2008) Towards improving compound selection in structure-based
virtual screening. Drug Discov Today 13,
219226.
69. Manallack, D. T., Pitt, W. R., Gancia,
E., Montana, J. G., Livingstone, D. J.,
Ford, M. G., Whitley, D. C. (2002) Selecting screening candidates for kinase and G
protein-coupled receptor targets using neural networks. J Chem Inf Comput Sci 42,
12561262.

173

70. Schneider, G. (2002) Trends in virtual combinatorial library design. Curr Med Chem 9,
20952101.
71. Beavers, M. P., Chen, X. (2002) Structurebased combinatorial library design: methodologies and applications. J Mol Graph Model
20, 463468.
72. Coupez, B., Lewis, R. A. (2006) Docking and scoringtheoretically easy, practically
impossible? Curr Med Chem 13, 29953003.
73. Sun, Y., Ewing, T. J., Skillman, A. G., Kuntz,
I. D. (1998) CombiDOCK: structure-based
combinatorial docking and library design.
J Comput Aided Mol Des 12, 597604.
74. Kuntz, I. D., Blaney, J. M., Oatley, S. J., Langridge, R., Ferrin, T. E. (1982) A geometric approach to macromolecule-ligand interactions. J Mol Biol 161, 269288.
75. Murray, C. W., Clark, D. E., Auton, T.
R., Firth, M. A., Li, J., Sykes, R. A.,
Waszkowycz, B., Westhead, D. R., Young,
S. C. (1997) PRO_SELECT: combining
structure-based drug design and combinatorial chemistry for rapid lead discovery. 1.
Technology. J Comput Aided Mol Des 11,
193207.
76. Blaney, J. M., Dixon, J. S. (1993) A good ligand is hard to find: automated docking methods. Perspect Drug Discov Des 1, 301319.
77. Kuntz, I. D., Meng, E. C., Shoichet, B.
(1994) Structure-based molecular design.
Acc Chem Res 27, 117123.
78. Bohm, H. J. (1994) The development of a
simple empirical scoring function to estimate
the binding constant for a protein-ligand
complex of known three-dimensional structure. J Comput Aided Mol Des 8, 243256.
79. Clark, D. E., Frenkel, D., Levy, S. A., Li,
J., Murray, C. W., Robson, B., Waszkowycz,
B., Westhead, D. R. (1995) PRO-LIGAND:
an approach to de novo molecular design.
1. Application to the design of organic
molecules. J Comput Aided Mol Des 9,
1332.
80. Makino, S., Ewing, T. J., Kuntz, I. D. (1999)
DREAM++: flexible docking program for virtual combinatorial libraries. J Comput Aided
Mol Des 13, 513532.
81. Roe, D. C., Kuntz, I. D. (1995) BUILDER
v.2: improving the chemistry of a de novo
design strategy. J Comput Aided Mol Des 9,
269282.
82. Van Gunsteren, W. F., Berendsen, H. J. C.
(1977) Algorithms for macromolecular
dynamics and constraint dynamics. Mol Phys
34, 13111327.
83. Sprous, D. G., Lowis, D. R., Leonard, J.
M., Heritage, T., Burkett, S. N., Baker, D.
S., Clark, R. D. (2004) OptiDock: virtual

174

Cavasotto and Phatak

HTS of combinatorial libraries by efficient

sampling of binding modes in product space.
J Comb Chem 6, 530539.
84. Rarey, M., Lengauer, T. (2000) A recursive algorithm for efficient combinatorial
library docking. Perspect Drug Discov Des 20,
6381.
85. Zhou, J. Z., Shi, S., Na, J., Peng, Z.,
Thacher, T. (2009) Combinatorial librarybased design with Basis Products. J Comput
Aided Mol Des DOI 10.1007/s10822-0099297-9.
86. Grzybowski, B. A., Ishchenko, A. V., Kim,
C. Y., Topalov, G., Chapman, R., Christianson, D. W., Whitesides, G. M., Shakhnovich,
E. I. (2002) Combinatorial computational
method gives new picomolar ligands for a
known enzyme. Proc Natl Acad Sci USA 99,
12701273.

87. Zhou, J. Z. (2008) Structure-directed combinatorial library design. Curr Opin Chem
Biol 12, 379385.
88. Head, M. S., Ryan, M. D., Lee, D., Feng,
Y., Janson, C. A., Concha, N. O., Keller,
P. M., deWolf, W. E., Jr. (2001) Structurebased combinatorial library design: discovery of non-peptidic inhibitors of caspases
3 and 8. J Comput Aided Mol Des 15,
11051117.
89. Zhao, L., Huang, W., Liu, H., Wang,
L., Zhong, W., Xiao, J., Hu, Y., Li,
S. (2006) FK506-binding protein ligands: structure-based design, synthesis,
and neurotrophic/neuroprotective properties of substituted 5,5-dimethyl-2-(4thiazolidine)carboxylates. J Med Chem 49,
40594071.

Chapter 9
Structure-Based Library Design in Efficient Discovery
of Novel Inhibitors
Shunqi Yan and Robert Selliah
Abstract
Structure-based library design employs both structure-based drug design (SBDD) and combinatorial
library design. Combinatorial library design concepts have evolved over the past decade, and this chapter
covers several novel aspects of structure-based library design together with successful case studies in the
anti-viral drug design HCV target area. Discussions include reagent selections, diversity library designs,
virtual screening, scoring/ranking, and post-docking pose filtering, in addition to the considerations
of chemistry synthesis. Validation criteria for a successful design include an X-ray co-crystal complex
structure, in vitro biological data, and the number of compounds to be made, and these are addressed in
this chapter as well.
Key words: Structure-based drug design, structure-based library design, library design, focused
library design, diversity library, combinatorial library, docking, reagent selections, HCV NS5B,
thiazolone.

1. Introduction
Structure-based library design engages in dual approaches of
structure-based drug design (SBDD) and combinatorial library
design (Fig. 9.1). Design concepts of combinatorial library have
been evolving since its conception more than a decade ago. Early
efforts mainly focused on the capability to synthesize large number of compounds through combinatorial chemistry with the confidence that high-throughput screening (HTS) (1) of every possible compound in a large library would lead to potential druggable
hits and leads, and eventually development of candidates after
subsequent lead optimizations. Needless to say, this approach
J.Z. Zhou (ed.), Chemical Library Design, Methods in Molecular Biology 685,
DOI 10.1007/978-1-60761-931-4_9, Springer Science+Business Media, LLC 2011

175

176

Yan and Selliah

Fig. 9.1. Structure-based combinatorial library design.

oversimplified the complex processes of drug discovery. Drugs

reported to originate solely from combinatorial chemistry thus
far are rare.
The advent of more accurate and rapid tools in chemoinformatics and virtual screening makes it possible to design and
synthesize a small subset of representative compounds (focused
library) of a larger library. Out of various improved methods these
two diversity- or structure-based approaches are frequently exercised in the design of a focused library. Once the 3D coordinates
of a protein target are determined by either X-ray crystal structures or NMR, a structure-based library design is a more productive and viable approach.
This chapter covers several aspects of structure-based library
designs coupled with the successful case studies in the anti-viral
HCV area. Discussions include reagent selections, diversity library
designs, virtual screening, scoring/ranking, and post-docking
pose filtering, in addition to the considerations of chemistry synthesis (Fig. 9.1). Validation criteria for a successful design include
an X-ray co-crystal complex structure and in vitro biological data,
and these are addressed in this chapter as well.

2. Materials
A number of computational methodologies have been used in this
experiment. Reagent selections for library designs were exported

Structure-Based Library Design

177

from ACD database (2). GOLD (3) and LigandFit (4) were used
for docking, where MOE software (5) was used for reagent filtering and library enumeration. Post-docking pose filters and reactive group filtering were carried out using a MOE SVL script
(5). Diversity analysis and physical property calculation were performed by Ceris2 (4). HKL package was applied for X-ray data
process (6) and the X-ray structure determination and refinement
were carried out by CNX (4). X-ray co-crystal complex structures
discussed in this chapter were deposited in PDB database with
codes 2O5D, 2HWH, 2HWI, and 2I1R.
The chemical synthesis of the compounds is applicable for
library production. The reactions start with a condensation reaction of a readily available reagent, rhodinine, with an aldehyde
to afford a thiazolonone intermediate, which can undergo a coupling reaction with an amine, amino acid, or a variety of other
amino-containing derivatives to give final products with good
yields (79).

3. Methods
3.1. Introduction

SBDD begins with designs of novel scaffolds based on the structural binding information of hit or lead compounds in complex
with a target which is usually an enzyme or a protein. Previously
hard-to-crystallized protein and their corresponding protein
ligand complex co-crystal structures have been routinely determined nowadays due to the significant technology improvement
in crystallography, molecular biology, and protein science in the
last decade. Given an X-ray complex structure of a proteinligand
co-crystal, various computational tools such as virtual screening
(10, 11), de novo design (11), and scaffold hopping are utilized to design brand new and better molecules with potentially
novel IP space coverage. Such designs are often accomplished
through exploring better or comparable ligandprotein interaction as predicted computationally. Alternatively, a 3D structure
of a target can be also approximated with reasonable confidence
from a homology model if identity level of amino acid sequences
between the target and a known structure from either in-house Xray determinations or PDB database is high. Recently more and
more protein structures have been solved with high resolutions
(<2.5 ) by NMR experiments. Structures from NMR are sometimes advantageous for SBDD in terms of the flexibility of protein
since the experiments are generally performed in solution at room
temperature and therefore, the dynamic states of protein resemble
more of the genuine states of protein in physiological conditions

178

Yan and Selliah

than a X-ray structure derived from a solid state of protein at

flash-frozen condition with liquid nitrogen.
Docking methods are commonly used to determine whether
the newly designed scaffold fits to the target in a desirable binding
mode similar to the one(s) found by an X-ray complex structure. A scaffold is rarely selected for further pursuit, for example, chemistry synthesis, if initial docking evaluations show unexpected results. Various docking programs, such as Glide (12),
GOLD (3), FlexX (1315), Surflex-Dock (16), Ligandfit (4),
AutoDock (17), and DOCK (18), are commercially available. It is
the specific users job to validate which docking method is applicable to their target (Fig. 9.2). A docking program is by and
large acceptable if it reproduces the X-ray complex structure with
a moderately low RMSD (<1.5 ) from the superimposition of
ligand structures between X-ray and a docking study (Fig. 9.2).
It is frequently practiced as well to apply two different docking
programs for cross-validation to increase the confidence level of
docking results.

Fig. 9.2. Validation of docking programs for scaffold designs.

3.2. Focused Library

Design

Rational design of a target-based focus library becomes critical

when a new scaffold carries different substituents, i.e., R1 and
R2 , which originate from functional handles and derivatives such
as amine, carboxyl acid, aldehyde, hydrazine (Fig. 9.3). As one
can imagine, the resulting virtual library for product C can easily
amount to millions of individual compounds based on availability of R groups. The critical questions for a medicinal chemist are
which compound sets are to be made first? Computational design
of focused library aims to trim a large virtual library to a manageable and viable realistic library that can be synthesized and
tested for desired activity of novel scaffolds (Fig. 9.3).
Computational library design process begins with reagent
selections, followed by diversity analysis and virtual library enumeration, and ends with selection of a final set of molecular
structures to be synthesized (Fig. 9.4). Two databases, Available Chemical Database (ACD) (2) and Chemicals Available for

Structure-Based Library Design

179

Fig. 9.3. A schematic reaction.

Fig. 9.4. Focused library design.

Purchase (CAP) (4), are commonly used by medicinal chemists to

select reagents for their chemical synthesis and these are also used
by computational chemists for the selection of building blocks;
both databases allow for export of all of the desirable reagents
into a single SDF structural file. Other non-structural information of the reagents such as the origins of countries, price, and
time for delivery are normally available in the databases and can
be exported into the same SDF file as well. With structures and
non-structural information together in one file, the task of weeding out some undesirable reagents and building blocks becomes
easy. Scripting language implemented in MOE software package
is used interactively for these tasks. Price of reagents, availability, delivery time are collectively classified as availability filters and
these filters are applied to select out non-optimal features. The
list of reagents is further narrowed down by applying (a) reactive
groups to remove building blocks that contain undesirable reactive functional groups which may interfere with desired chemical reactions and (b) physical property filters to remove the ones
that have undesirable physical properties such as high molecular
weight (MW > 300), too many rotational bonds (rot > 5), and
too many chiral centers (n > 2). Topological filter can further
reduce reagent list if it is necessary. Virtual library enumeration is

180

Yan and Selliah

enumerated using the refined building block list using commercial software such as MOE (5) or CombiLibMaker in Sybyl (16).
The virtual library thus obtained may be further trimmed using
ADME filters. The final focus library is thus prepared for virtual
screening against a protein target (Fig. 9.4) (4, 5, 16).
3.3. Virtual
Screening, Scoring,
and Ranking of
Focused Library

Virtual screening methods have been routinely and extensively

applied in the generation of lead compounds from commercially
available chemical libraries (1922). Conventional virtual screening programs for combinatorial libraries include CombiGlide (12)
and FlexXc (23). Both methods work similarly in a way by first
anchoring the core structure (or scaffold) in a predetermined ideal
location and then making side chain R groups flexible to identify
a focused set of molecules with most favorable R groups (12, 23).
Docking molecules in this way renders different R groups invisible to each other during docking and can thus generate a focused
library by eliminating energetically unfavorable R-group and conformations upon binding to a target.
This shortcoming of docking in a combinatorial fashion is
readily overcome by docking all molecules individually using conventional docking programs, i.e., GOLD (23), Glide (12), FlexX
(23), or Surflex-Dock (16), followed by post-docking pose mining. The post-docking filters are realized through some straightforward scripting language such as MOE SVL (5). A typical script
program allows users to identify all structures in a library that bind
to a target in a way as desired. Specifically, this program will read
the criteria definition parameters from a file for pharmacophore
matching (Table 9.1) and automatically select all molecules in
the docking pose database with desirable and anticipated poses.
For example, in Table 9.1, column 1 is label; column 2 represents SMARTS patterns of fragments, which is usually the core
structure of ligand, potentially for either hydrogen bonding or
hydrophobic contacts; column 3 is the coordinates of the core
interaction points in a receptor, and the last column denotes distance criteria between column 2 and column 3. Molecules in the

Table 9.1
Parameter definition for post-docking filter
Label,

smarts_pattern,

x_y_z_coordinates,

distance_
threshold

[d1_NH,

[NH]([CH2])cn,

[165.18, 27.16, 27.46],

2.5]

[d2_OC,

O(=C([NH])[CH2][NH]),

[165.68, 30.03, 29.63],

2.0]

[d3_nap,

[cX3][nX3]nc([NH])c[#6],

[162.06, 25.15, 28.78],

2.0]

[d4_Me,

[nX3]nc([NH])cc,

[162.6, 25.67, 29.98],

2.0]

[d5_nR,

[cX3]([cX3])[cX3]([NH])n[nX3],

[162.92, 25.61, 27.74],

2.0]

Structure-Based Library Design

181

initial virtual library are selected for further analysis only if all of
the distance criteria in column 4 of Table 9.1 are satisfied.
Post-docking pharmacophore-based filtering, followed by
various scoring functions (Fig. 9.1), is greatly advantageous in
comparison with results when only docking scoring functions are
used. Docking scoring functions (35, 12, 16, 17, 23) are in
reality met with considerable limitations in reasonably prioritizing compounds in accordance with their corresponding binding
affinities or enzymatic potency (2427). One of the reasons for
this lack of correlation is the artificially high ranking for an incorrect docking pose (28, 29). However, it is well documented that
docking methods are able to reproduce the bound conformation
of a ligand in a proteinligand complex determined by X-ray crystallography (2932). Therefore, once the molecules with the correct poses are identified by post-docking filters, the problem of
scoring wrong poses is avoided and multiple scoring functions
can thus be better suited to rank molecules in the focused library
with better chance of success (Fig. 9.1).
A set of molecules that rank high after this process would
be synthesized and subject to biological tests, i.e., in vitro enzymatic assay or binding affinity experiments, in order to confirm
design rationales. Simultaneously, X-ray co-crystal structures of
these ligands in complex with the target are to be determined to
further corroborate modeling results. Positive results from such
approaches are decisive for selection of next set of compounds for
synthesis and the future directions of lead optimization.
3.4. Structure-Based
Library Design in
Discovery of HCV
NS5B Polymerase
Inhibitors
3.4.1. Background

Hepatitis C virus (HCV) was discovered in 1989 and has been

regarded as the key causative agent for non-A, non-B virus hepatitis (3335). It is estimated that there are over 170 million people worldwide and about 4 million individuals in United States
with chronic HCV infection (36). Majority of the infected persons (80%) develop chronic hepatitis, where about 1025% of
them could advance to serious HCV-related liver diseases such as
fibrosis, cirrhosis, and hepatocellular carcinoma (37). Only a fraction of patients respond to current FDA-approved standard therapy with a sustained viral load reduction (38), and many of them
could not tolerate the treatment because of the various severe side
effects (39). Therefore, HCV still represents an unmet medical
need which requires discovery and development of more effective
and well-tolerated therapies.
HCV NS5B polymerase is a non-structural proteins encoded
in HCV genome. This polymerase plays a crucial role in replicating HCV virus and causing infectivity (38) and thus is a
key target for drug discovery against HCV (40). Various series
of non-nucleoside molecules with different scaffolds have been
published recently as HCV NS5B inhibitors (79, 41, 42). A
couple of scaffolds including Mercks indole scaffold, Pfizers

182

Yan and Selliah

dihydropyran-2-one derivatives, and Shires phenylalanine appear

to bind to allosteric sites of NS5b (4346). These binding sites
are located on the surface of the thumb sub-domains remote from
the NS5B active site. Inhibitors located in such binding sites are
believed to show more favorable on-target specific efficacy but
less unwanted side effects due to off-target binding.
3.4.2. SBDD of a Novel
Thiazolone Scaffold as
HCV NS5b Inhibitor

In our HCV programs, we aimed to discover novel scaffolds efficiently to explore the allosteric site of HCV NS5B by means of
structure-based approach including the focused library design (7).
Our main strategies for new scaffolds are to maintain key pharmacophores of the initial hit, establish sizable chemistry space,
and most importantly identify directions for future diversification
and optimization. We started with a hit 1 from high-throughput
screening which has an IC50 value of 2.0 M (Fig. 9.5). An
X-ray complex structure indicated that 1 binds to a location
in the allosteric site (Fig. 9.6). Key inhibitorprotein interactions include the following: (1) both C=O and N on the thiazolone ring hydrogen bond with backbone NHs of Tyr477
and Ser476, (2) sulfonamide oxygen atom engages in hydrogenbonding interaction with basic side chain NH3 + of Arg 501,
(3) the aromatic furan and phenyl rings interact with the protein
by hydrophobic contacts (Fig. 9.6). Furthermore, such binding
information enable us to envision that a novel structure 2 once
incorporated with a suitable (S) amino acid possesses not only
pharmacophore equivalent as 1 but additional chemistry opportunity for exploring more space in the pocket (Figs. 9.5 and 9.6).

Fig. 9.5. SBDD of a novel scaffold 2 as NS5B inhibitor.

The scaffold 2 was confirmed by GOLD docking to have a

binding mode which is similar to 1. Besides, the carboxyl group
on 2 picks up additional interaction with side chain of Lys 533
and can be further functionalized to explore more space in the
pocket (Figs. 9.6 and 9.7).
Starting with the desirable scaffold 2 at hand, we decided
to employ the approaches outlined in Fig. 9.1 for focus library

Structure-Based Library Design

183

Fig. 9.6. X-ray complex structure of 1 with NS5B.

Fig. 9.7. Alignment of 1 (in sticks) from X-ray with 2 (sticks and balls) from docking.

design and selecting compounds to synthesize, where reactions

to make final compounds such as 2 require amino acids as chemical reagents (7). A substructure search of amino acids in ACD
database (2) produced 2862 hits and the number was reduced to
1175 after application of a topological diversity selection using
MOE package (5). A virtual library was then enumerated and
underwent GOLD virtual screening with the standard parameters (3). During the virtual screening, the bound conformation
of 1 from X-ray was used as shape template similarity constraint
and the constraint weight was set to 10. Each molecule in the
focused library was allowed to have 10 docking poses and totally
11,750 poses were collected and filtered according to the predetermined pharmacophore-based criteria using an in-house MOE

184

Yan and Selliah

SVL script. Not surprisingly, only 60 molecules passed this filter

and were then re-ranked by GOLD scoring function (3). One of
the 10 top-scored molecules was proposed for synthesis and the
compound 3 was determined to have an IC50 value of 3.0 M.
Subsequent X-ray structure of 3 in complex with NS5B was solved
at 2.0 and this molecule shows a binding mode just as predicted
in the thumb domain (7) (Fig. 9.8).

Fig. 9.8. X-ray complex structure of 3 with NS5B at a 2.0 resolution.

3.4.3. Further SBDD of

Follow-Up Focused
Library

Structural analysis of the binding mode of 3 in the pocket further led to the identification of more new scaffolds 4 and 5
as HCVNS5B inhibitors (Fig. 9.9) (9). A small focused library
was enumerated and selected for synthesis after virtual screening.
In general, carboxyl compounds with more flexibility resulting
from addition of one methylene (CH2 ) group have comparable potency with original compound 3. The most potent compound 6 has an IC50 of 8.5 M (Table 9.2). Compound 11
shows similar enzymatic potency to 6, while mono-substituted
molecules 710, regardless of the chiral centers, showed much
weaker potency (Table 9.2). Molecules with tetrazole moiety, a
commonly used carboxyl group COOH bioisostere, were pre-

Fig. 9.9. New designs of novel scaffolds.

Structure-Based Library Design

185

Table 9.2
Enzymatic potency (IC50 in M) for new molecules. IC50 (M)
values of novel scaffolds
O

O
N

S
N

N
N

N
N

N
O

Entry

R:

IC50(M)

Entry

IC50(M)

R:

Entry

R:

IC50(M)

Cl

Cl

8.5

44.0

10

Me

27.0

12

13.0

13

9.0

14

9.7

19.0

Cl
Br

16.5

11

Cl

Me

14.0

dicted to fit well into target as well and a few of them were synthesized. As seen from Table 9.2, tetrazole compounds 1214 are
moderately potent with IC50 values of 9.7, 19.0, and 14.0 M,
respectively. Extending the tetrazole group by one more CH2
group is tolerated by protein. To prove the design rationale for
future structure-based designs, co-crystals structure of 12 in complex with HCV NS5B was successfully established at a resolution
of 2.2 . The electron density was clear for inhibitor 12, which
binds to the thumb sub-domain as expected (Fig. 9.10). Overall interactions of 12 with protein are comparable with those of 3
(Fig. 9.10).

Fig. 9.10. X-ray complex structure of 12 with HCV NS5B (3 in yellow sticks).

186

Yan and Selliah

3.4.4. Further Designs of

ThiazoloneAcylsulfonamide as
NS5B Inhibitors

All of the scaffolds discussed above make hydrogen-bonding

interaction with Ser476, Tyr477, and Arg501 and, in the same
region, engage in similar hydrophobic contacts with Met423,
Ile482, Val485, Leu489, Leu497, and Trp528 (Fig. 9.10). In the
vicinity of the inhibitorprotein interaction pocket there appears
to be more space open for additional interactions. In particular,
this new site has two basic residues, His475 and Lys533, as gatekeepers near its entrance (Fig. 9.10). A molecule with an appropriate moiety to interact with these two residues was predicted to
be able to reach this extra pocket. Our continued SBDD effort
was to design such a new scaffold. We envisioned that an acylsulfonamide 15 that has a comparable pKa with COOH could
serve as a candidate to hydrogen bond with the basic side chain of
Lys533 and additional aromatic moiety linked to sulfonyl group
picking up stacking with His475 (Fig. 9.11) (8).

Fig. 9.11. Design of acylsulfonamide scaffold.

To validate the design principle, a very small focused set of

library compounds, seven compounds in total, were synthesized
and subsequently evaluated for the inhibiting the activity of HCV
NS5B. All compounds were reasonably active with IC50 values in
the range of 620 M. One of the compounds was successfully
soaked into NS5B protein crystal and its complex structure with
protein was obtained at 2.2 . The selectron density was clear for
the inhibitor and the compound fits nicely to the same allosteric
site as the 3 and 15 with additional interactions with basic side
chains of Arg422 and Lys533 as predicted by GOLD (Fig. 9.12)
(8). It is also interesting to find that 4-NO2 -Ph makes a faceface
stacking with His475 (Fig. 9.12). New scaffolds like this open
fresh opportunity for SBDD targeting this allosteric site of HCV
NS5B.

Structure-Based Library Design

187

Fig. 9.12. Electron-density map and interactions of acylsulfonamide with NS5B allosteric site.

4. Notes
1. The key to a success is to diligently do various cycles of
filtering such as availability, reaction groups, and diversity
selection of reagents before library enumerations. It is also
necessary to carry out automatic pharmacophore-based
post-docking pose filtering prior to using any docking scoring functions.
2. Induced fit docking (IFD) should be carried out periodically
to check whether inclusion of flexibility of receptor improves
docking results or not. Regular docking, while very fast,
treats all amino acids rigid which does not reflect the true
nature of protein flexibility and consequently true positives
may be missed.
3. Molecular dynamics (MD) should be performed for binding pockets defined mostly by side chains of flexible protein residues to generate an ensemble of binding sites. Such
an ensemble can be used for subsequent docking or virtual
screening in a parallel fashion.
4. A SBDD design should be confirmed by a later X-ray complex structure which in turn serves to initiate a cycle of
iterative structural-based drug design (SBDD). SBDD starts
from X-ray or NMR complex structure of ligand with protein and a design, if synthesized, validated, and confirmed
by X-ray, creates a starting point for a new level of SBDD
efforts.
5. Do not use any scoring functions blindly without validation
in any specific drug targets. Most SBDD efforts involve both

188

Yan and Selliah

docking and scoring. Docking generates a number of poses

with different conformation of ligands in a binding site of a
receptor and subsequent scoring function is applied to rank
them energetically based on the interaction of a pose and a
given binding site. One can validate a scoring function by
performing a so-called enrichment ratio (ER) study, which
calculates the ratio of active compounds selected by scoring
function from a docking divided by the number of active
compounds if chosen randomly. While there is no specific
value for a good ER, a value of less than 1.0 unquestionably
suggests that the scoring does not do any better than a random selection. Thus, a greater value of ER corresponds to
the better performance of a scoring function in a docking
experiment.
6. Reagents with multiple reactive chemical groups should
be avoided in library enumeration because their presence
most likely requires specific protections of certain functional
groups which complicates chemical reactions and makes
library production unpractical.

References
1. Hertzberg, R. P., Pope, A. J. (2000) Highthroughput screening: new technology for
the 21st century. Curr Opin Chem Biol 4,
445451.
2. MDL Information Systems. http://www.
mdli.com.
3. Cambridge Crystallographic Data Centre,
UK. http://gold.ccdc.cam.ac.uk.
4. Accelyrs, San Diego, CA, USA. http://
www.accelrys.com.
5. Chemical Computing Group, Montreal,
Quebec, CA. http://www.chemcomp.com.
6. HKL Research, Inc. http://www.hklxray.com.
7. Yan, S., Appleby, T., Larson, G., Wu, J. Z.,
Hamatake, R., Hong, Z., Yao, N. (2006)
Structure-based design of a novel thiazolone
scaffold as HCV NS5B polymerase allosteric
inhibitors. Bioorg Med Chem Lett 16,
58885891.
8. Yan, S., Appleby, T., Larson, G., Wu, J.
Z., Hamatake, R. K., Hong, Z., Yao, N.
(2007) Thiazolone-acylsulfonamides as novel
HCV NS5B polymerase allosteric inhibitors:
convergence of structure-based drug design
and X-ray crystallographic study. Bioorg Med
Chem Lett 17, 19911995.
9. Yan, S., Larson, G., Wu, J. Z., Appleby,
T., Ding, Y., Hamatake, R., Hong, Z.,
Yao, N. (2007) Novel thiazolones as HCV

10.
11.
12.
13.

14.

15.

16.
17.
18.

NS5B polymerase allosteric inhibitors: Further designs, SAR, and X-ray complex structure. Bioorg Med Chem Lett 17, 6367.
Lyne, P. D. (2002) Structure-based virtual
screening: an overview. Drug Discov Today 7,
10471055.
Jain, S. K., Agrawal, A. (2004) De novo drug
design: an overview. India J Phar Sci 66,
721728.
Schrodinger, LLC., Portland, OR, USA.
http://www.schrodinger.com.
Rarey, M., Kramer, B., Lengauer, T.
(1999) Docking of hydrophobic ligands
with interaction-based matching algorithms.
Bioinformatics 15, 243250.
Kramer, B., Rarey, M., Lengauer, T. (1997)
CASP2 experiences with docking flexible
ligands using FlexX. Proteins Suppl 1,
221225.
Kramer, B., Rarey, M., Lengauer, T. (1999)
Evaluation of the FLEXX incremental construction algorithm for protein-ligand docking. Proteins 37, 228241.
Tripos,
St.
Louis,
MO,
USA.
http://www.tripos.com.
Molecular Graphics Laboratory, The Scripps
Research Institute, San Diego, CA, US.
http://autodock.scripps.edu.
DesJarlais, R. L., Sheridan, R. P., Dixon, J.
S., Kuntz, I. D., Venkataraghavan, R. (1986)

Structure-Based Library Design

19.

20.

21.
22.
23.
24.
25.

26.

27.

28.

29.
30.

31.

32.

Docking flexible ligands to macromolecular

receptors by molecular shape. J Med Chem
29, 21492153.
Aronov, A. M., Munagala, N. R., Kuntz, I.
D., Wang, C. C. (2001) Virtual screening
of combinatorial libraries across a gene family: in search of inhibitors of Giardia lamblia
guanine phosphoribosyltransferase. Antimicrob Agents Chemother 45, 25712576.
Ghosh, S., Nie, A., An, J., Huang, Z. (2006)
Structure-based virtual screening of chemical
libraries for drug discovery. Curr Opin Chem
Biol 10, 194202.
Green, D. V. (2003) Virtual screening of virtual libraries. Prog Med Chem 41, 6197.
Shoichet, B. K. (2004) Virtual screening of chemical libraries. Nature 432,
862865.
BioSolveIT GmhB, Germany. http://www.
biosolveit.de.
Coupez, B., Lewis, R. A. (2006) Docking and scoringtheoretically easy, practically
impossible? Curr Med Chem 13, 29953003.
Kontoyianni, M., Madhav, P., Suchanek, E.,
Seibel, W. (2008) Theoretical and practical
considerations in virtual screening: a beaten
field? Curr Med Chem 15, 107116.
Kontoyianni, M., Sokol, G. S., McClellan, L.
M. (2005) Evaluation of library ranking efficacy in virtual screening. J Comput Chem 26,
1122.
Kontoyianni, M., McClellan, L. M., Sokol,
G. S. (2004) Evaluation of docking performance: comparative data on docking algorithms. J Med Chem 47, 558565.
Verdonk, M. L., Berdini, V., Hartshorn, M.
J., Mooij, W. T., Murray, C. W., Taylor,
R. D., Watson, P. (2004) Virtual screening
using protein-ligand docking: avoiding artificial enrichment. J Chem Inf Comput Sci 44,
793806.
Stahl, M., Bohm, H. J. (1998) Development
of filter functions for protein-ligand docking.
J Mol Graph Model 16, 121132.
Stahura, F. L., Xue, L., Godden, J. W.,
Bajorath, J. (1999) Molecular scaffold-based
design and comparison of combinatorial
libraries focused on the ATP-binding site of
protein kinases. J Mol Graph Model 17, 19,
5152.
Godden, J. W., Stahura, F., Bajorath, J.
(1998) Evaluation of docking strategies for
virtual screening of compound databases:
cAMP-dependent serine/threonine kinase
as an example. J Mol Graph Model 16,
139143, 65.
Vigers, G. P., Rizzi, J. P. (2004) Multiple
active site corrections for docking and virtual
screening. J Med Chem 47, 8089.

189

33. Choo, Q. L., Weiner, A. J., Overby, L. R.,

Kuo, G., Houghton, M., Bradley, D. W.
(1990) Hepatitis C virus: the major causative
agent of viral non-A, non-B hepatitis. Br Med
Bull 46, 423441.
34. Choo, Q. L., Kuo, G., Weiner, A. J., Overby,
L. R., Bradley, D. W., Houghton, M. (1989)
Isolation of a cDNA clone derived from
a blood-borne non-A, non-B viral hepatitis
genome. Science 244, 359362.
35. Weiner, A. J., Kuo, G., Bradley, D. W.,
Bonino, F., Saracco, G., Lee, C., Rosenblatt, J., Choo, Q. L., Houghton, M. (1990)
Detection of hepatitis C viral sequences in
non-A, non-B hepatitis. Lancet 335, 13.
36. (2000) Hepatitis C-Global prevalence
(update). Weekly Epidemiol Rec 75, 1819.
37. Memon, M. I., Memon, M. A. (2002) Hepatitis C: an epidemiological review. J Viral
Hepat 9, 84100.
38. Kolykhalov, A. A., Mihalik, K., Feinstone, S.
M., Rice, C. M. (2000) Hepatitis C virusencoded enzymatic activities and conserved
RNA elements in the 3 nontranslated region
are essential for virus replication in vivo. J
Virol 74, 20462051.
39. Scott, L. J., Perry, C. M. (2002) Interferonalpha-2b plus ribavirin: a review of its use
in the management of chronic hepatitis C.
Drugs 62, 507556.
40. De Clercq, E. (2002) Strategies in the design
of antiviral drugs. Nat Rev Drug Discov 1,
1325.
41. Rong, F., Chow, S., Yan, S., Larson, G.,
Hong, Z., Wu, J. (2007) Structure-activity
relationship (SAR) studies of quinoxalines
as novel HCV NS5B RNA-dependent RNA
polymerase inhibitors. Bioorg Med Chem Lett
17, 16631666.
42. Yan, S., Appleby, T., Gunic, E., Shim, J.
H., Tasu, T., Kim, H., Rong, F., Chen, H.,
Hamatake, R., Wu, J. Z., Hong, Z., Yao, N.
(2007) Isothiazoles as active-site inhibitors of
HCV NS5B polymerase. Bioorg Med Chem
Lett 17, 2833.
43. Wang, M., Ng, K. K., Cherney, M. M., Chan,
L., Yannopoulos, C. G., Bedard, J., Morin,
N., Nguyen-Ba, N., Alaoui-Ismaili, M. H.,
Bethell, R. C., James, M. N. (2003) Nonnucleoside analogue inhibitors bind to an
allosteric site on HCV NS5B polymerase.
Crystal structures and mechanism of inhibition. J Biol Chem 278, 94899495.
44. Di Marco, S., Volpari, C., Tomei, L., Altamura, S., Harper, S., Narjes, F., Koch, U.,
Rowley, M., De Francesco, R., Migliaccio,
G., Carfi, A. (2005) Interdomain communication in hepatitis C virus polymerase abolished by small molecule inhibitors bound

190

Yan and Selliah

to a novel allosteric site. J Biol Chem 280,

2976529770.
45. Biswal, B. K., Cherney, M. M., Wang, M.,
Chan, L., Yannopoulos, C. G., Bilimoria, D.,
Nicolas, O., Bedard, J., James, M. N. (2005)
Crystal structures of the RNA-dependent
RNA polymerase genotype 2a of hepatitis C
virus reveal two conformations and suggest

mechanisms of inhibition by non-nucleoside

inhibitors. J Biol Chem 280, 1820218210.
46. Biswal, B. K., Wang, M., Cherney, M. M.,
Chan, L., Yannopoulos, C. G., Bilimoria,
D., Bedard, J., James, M. N. (2006) Nonnucleoside inhibitors binding to hepatitis C
virus NS5B polymerase reveal a novel mechanism of inhibition. J Mol Biol 361, 3345.

Chapter 10
Structure-Based and Property-Compliant Library Design
of 11-HSD1 Adamantyl Amide Inhibitors
Genevieve D. Paderes, Klaus Dress, Buwen Huang, Jeff Elleraas,
Paul A. Rejto, and Tom Pauly
Abstract
Multiproperty lead optimization that satisfies multiple biological endpoints remains a challenge in the
pursuit of viable drug candidates. Optimization of a given lead compound to one having a desired set
of molecular attributes often involves a lengthy iterative process that utilizes existing information, tests
hypotheses, and incorporates new data. Within the context of a data-rich corporate setting, computational
tools and predictive models have provided the chemists a means for facilitating and streamlining this
iterative design process. This chapter discloses an actual library design scenario for following up a lead
compound that inhibits 11-hydroxysteroid dehydrogenase type 1 (11-HSD1) enzyme. The application
of computational tools and predictive models in the targeted library design of adamantyl amide 11HSD1 inhibitors is described. Specifically, the multiproperty profiling using our proprietary PGVL (Pfizer
Global Virtual Library) Hub is discussed in conjunction with the structure-based component of the
library design using our in-house docking tool AGDOCK. The docking simulations were based on a
piecewise linear potential energy function in combination with an efficient evolutionary programming
search engine. The library production protocols and results are also presented.
Key words: Multiproperty lead optimization, library design, adamantyl amide, targeted library,
11-hydroxysteroid dehydrogenase type 1, 11-HSD1, PGVL, Pfizer Global Virtual Library,
structure-based, AGDOCK, piecewise linear, evolutionary programming.

1. Introduction
Glucocorticoids (GC) are steroid hormones that regulate various physiological processes via stimulation of the nuclear glucocorticoid receptors (1). Chronically elevated levels of active
GC hormones (e.g., cortisol) have been associated with many
J.Z. Zhou (ed.), Chemical Library Design, Methods in Molecular Biology 685,
DOI 10.1007/978-1-60761-931-4_10, Springer Science+Business Media, LLC 2011

191

192

Paderes et al.

diseases, including diabetes, obesity, dyslipidemia, and hypertension. In mammalian tissues, GC hormonal regulation is controlled by two isozymes of 11-hydroxysteroid dehydrogenase
that catalyze the interconversion of inert cortisone and active
cortisol, namely, 11-HSD1, which is present predominantly in
the liver, adipose tissue, and brain, and 11-HSD2, which is
mainly expressed in the kidney and placenta (2, 3). 11-HSD1
is a bidirectional, NADPH-dependent enzyme that catalyzes the
conversion of inactive 11-keto GCs (cortisone in humans and
11-dehydrocorticosterone in rodents) into hormonally active
11-hydroxy GCs (cortisol in human and corticosterone in
rodents), whereas 11-HSD2 is a unidirectional dehydrogenase
that catalyzes the reverse reaction (cortisol to cortisone) using
NAD+ solely as a cofactor (3, 4). In recent years, clinical studies
in animal models (57) and in humans (812) provided evidence
for the role of 11-HSD1 enzyme activity in obesity, diabetes, and
insulin insensitivity. In line with these findings, inhibition of 11HSD1 by the steroid carbenoxolone (CBX) showed improved
insulin sensitivity in human (1314). Thus, 11-HSD1 is considered a promising target for the treatment of glucocorticoidrelated diseases and has given rise to several classes of nonsteroidal
11-HSD1 inhibitors (1518), including the adamantyl triazoles
and amides (1921).
The identification of an adamantyl amide inhibitor of human
11-HSD1 (Fig. 10.1) in our laboratories has prompted us
to design a targeted library of close analogs using the Pfizer
Global Virtual Library (PGVL) Hub, a desktop tool for designing
libraries and accessing Pfizer internal tools, models, and resources.
With PGVL Hub, we were able to input our customized alcoholcontaining adamantyl amide template and select the appropriate
reaction protocol with its corresponding set of amine monomers
(Fig. 10.2). The reaction protocol involves the transformation
of alcohols to amines via mesylation followed by amine substitution (22, 23). The initial set of amine monomers from in-house
and commercial sources gave us 13,000 amines. Reduction in
the virtual chemistry space to 1,000 was achieved by selecting only available secondary amines having molecular weights

H
N

N
O
N

Fig. 10.1. Adamantyl amide inhibitor of human 11-HSD1 (hu11-HSD1 Ki(app) =

1.8 nM, EC50 = 171 nM, kinetic solubility = 376 M, HLM = 7.6 L/min/mg,
HHep = 3.0 L/min/million).

Library Design of 11-HSD1 Adamantyl Amide Inhibitors

193

Fig. 10.2. Reaction protocol for the transformation of alcohols to amines via mesylation and amine substitution, as
shown in PGVL Hub.

less than 200. Since the objective of the library design was to
improve the cellular potency with retention of good solubility,
stability in human liver microsomes (HLM) and human hepatocytes (HHep), in silico property calculation and profiling were
performed on the virtual enumerated products, resulting in 300
predicted property-compliant virtual products.
In order to ensure the retention of enzyme activity, the virtual products were subjected to fixed anchor docking using
AGDOCK, wherein the adamantyl amide moiety was fixed to
a specified crystal-bound coordinate during the docking simulations. At the time of our library design, the human 11HSD1 (hu11-HSD1) crystal structure was not available. Thus,
we utilized our available in-house guinea pig 11-HSD1 (gp11HSD1) protein crystal structure for docking our virtual library
and selected its bound adamantyl ligand, which showed activity in hu11-HSD1, for defining the coordinates of our fixed
anchor structure. The docking simulations were carried out
using a piecewise linear intermolecular function (2427) and
a stochastic search algorithm based on evolutionary programming (28, 29). Evaluation of the dock hits led to the selection of the top-ranking virtual compounds based on their estimated high-throughput docking scores. The resulting structurebased and property-compliant, 88-compound library design was
then submitted to production for combinatorial synthesis. Initial

194

Paderes et al.

screening at 0.2 M concentration followed by purification of the

37 selected best hits (>90% inhibition) yielded a compound with
improved cell potency and solubility, high stability in HLM and
HHep, and retention of enzyme activity. Subsequent elucidation
and publication of the X-ray crystal structures of guinea pig (30),
human (31), and murine (32) 11-HSD1 enabled us to crystallize later an adamantyl amide analog in hu11-HSD1, which confirmed the similarity in the binding modes of the adamantyl amide
anchor structure in human and in guinea pig (Fig. 10.3), thereby
lending credence to use of the gp11-HSD1 crystal complex as
reference for docking.

Ser-170
Tyr-231
Tyr-183

Fig. 10.3. Adamantyl amide analogs exhibit similar binding modes in guinea pig (green)
and in human (pink) 11-HSD1 cocrystal complexes, with Ser-170 and Tyr-183 forming
hydrogen bond interactions with the bound ligands. A nonconserved residue (Tyr-231 in
guinea pig and Asn-123 in human) differentiates the active sites for these analogs.

1.1. PGVL Overview

PGVL is defined as a set of virtual molecules that can be synthesized from the available monomers and existing templates using
validated reaction protocols at Pfizer. It covers a vast virtual chemistry space in the order of 1013 compounds. PGVL Hub is the
corresponding desktop interface used for a quick navigation of the
virtual chemistry space and contains the basic features of an earlier
library design tool called LiBrain (33). Searching PGVL for compounds similar to a given lead or HTS hit can be carried out using
a Lead Centric Mining tool within PGVL Hub or a desktop
application called the Bayesian Idea Generator (34). For library
designs, virtual searching and screening are simply conducted on
specific subsets of PGVL, as defined by reaction types that utilize a
set of registered chemistry protocols along with their specific sets
of mined reactant monomers. One of the most useful features in
PGVL Hub is its ability to access Pfizers internal computational
tools and models. Thus, calculation of physicochemical properties
(e.g., thermodynamic solubility) and use of predicted biological

Library Design of 11-HSD1 Adamantyl Amide Inhibitors

195

endpoints from these models (e.g., in silico HLM model (35))

become an integral part of the virtual screening process.
1.2. AGDOCK Theory
for Docking
Simulations

AGDOCK is a Pfizer application for rapid and automated computational prediction of the binding geometries (conformation
and orientation) of compounds in a given protein active site, as
defined by the input defining ligand. It operates in three modes,
namely noncovalent docking, covalent docking (25, 36), and partially fixed or fixed anchor docking (36, 37). The default mode
is noncovalent docking with full ligand conformational flexibility
that explores a large number of degrees of freedom. Significant
reduction in the number of degrees of freedom is achieved with
the latter two modes in which part of the ligand is fixed within the
active site of the protein, either through covalent bond formation
with the receptor (covalent docking) or by imposition of positional constraints on an anchor fragment (fixed anchor docking)
that is primarily responsible for molecular recognition. AGDOCK
employs two search engines, evolutionary programming (2427)
and simulated annealing (38), both of which allow for a full search
of the ligand conformation and orientation within the active site.
It also supports two intermolecular potentials, AMBER (39) and
piecewise linear potential (2427), and an intramolecular potential consisting of van der Waals and torsional terms derived from
the DREIDING force field (40). The intermolecular potential
developed for AGDOCK incorporates both steric and hydrogen
bond contributions which are calculated from the sum of pairwise interactions between the ligand and the protein heavy atoms
using piecewise linear potentials. This energy function along with
an evolutionary search technique enables the structure prediction
of the protein-ligand complex.

2. Materials
2.1. PGVL Monomers

1. Reactant A: R-1(2-hydroxy-ethyl)-pyrrolidine-2-carboxylic
acid adamantan-2-ylamide (Fig. 10.2)
2. Reactant B: 88 cyclic and acyclic secondary amines with
molecular weights ranging from 71.2 to 162.1 Da

2.2. Reagents and

Solvents

1. Anhydrous 1,2-dichloroethane (DCE)

2. Triethylamine (TEA)
3. 4-N,N-Dimethylaminopyridine (DMAP)
4. Methanesulfonyl chloride
5. Anhydrous dichloromethane (DCM)

196

Paderes et al.

6. Anhydrous N,N-dimethylformamide (DMF)

7. Dimethylsulfone (DMSO)
8. 95:5 Methanol/water mixture (MeOH/water)
2.3. Input Files for
Docking Simulations

1. Structure file (in SDF or PDB format) containing the crystalbound conformation of reference ligand (Reactant A) in the
gp11-HSD1 cocrystal complex
2. Structure file (in PDB format) containing the coordinates of
the protein crystal structure derived from the gp11-HSD1
cocrystal complex (Fig. 10.4a)
3. Structure file (in SDF or PDB format) of the anchor or core
structure (Fig. 10.4b) which will be used in specifying the
fixed coordinates of the common fragment in the virtual
library of adamantyl amide analogs
4. Structure file (in SDF format) of the virtual library compounds to be subjected to fixed anchor or partially fixed
docking

Ser-170

Tyr-183

a)

b)

Fig. 10.4. (a) Crystal structure of gp11-HSD1 complex with Reactant A which was used
as defining ligand in docking simulations. (b) Adamantyl amide core structure used in
fixed anchor docking.

2.4. Computational
Tools and Resources

1. PGVL Hub for reactant monomer retrieval, virtual product

enumeration, molecular property calculation, product property profiling, and exporting virtual product structures for
subsequent docking
2. Molecular property calculators and predictors (e.g., aqueous
solubility model)
3. AGDOCK tool for docking the virtual library
4. PLCALC tool for calculating the protein-ligand interaction
free energy (HT) scores

Library Design of 11-HSD1 Adamantyl Amide Inhibitors

197

5. A script for ranking and extracting the best docked poses

along with their HT scores and other parameters into an
Excel table
6. MoViT tool for viewing the dock poses

3. Methods
3.1. PGVL Library
Design

The library design was conducted with PGVL Hub which allows
the retrieval of the appropriate reaction protocol along with their
corresponding sets of reactant monomers (Fig. 10.2). There are
basically four monomer sources, a commercial domain (ACD),
and three in-house domains (AXL, MN, and PF). With the selection of the in-house and commercial domains, the virtual library
size is 26,404 (2 Reactant A 13,202 Reactant B). In
this design, we selected only the in-house monomers which gave
us a virtual library size of 11,664 (2A 5,832B). By specifying a single alcohol-containing template for Reactant A which
is needed for generating close analogs of the adamantyl amide
lead compound, the virtual library size was reduced to 5,832
(1A 5,832 B). Further reduction in chemistry space was
achieved through filtering done both at the monomer and virtual product levels, as outlined in the subsequent library design
steps.
1. Calculate the molecular weight and structural alerts (substructures containing undesirable or reactive functionalities)
for 5,832 amines (see Note 1).
2. Perform a substructure search for secondary amines.
3. Select only secondary amines with molecular weight (MW)
less than 200 and with no structural alerts. This step drastically reduced the number of amines from 5,832 to 1,019.
4. Enumerate the virtual products for the alcohol template and
the selected amines using the PGVL Hub virtual product
enumerator.
5. Calculate the following molecular properties within PGVL
Hub using global computational tools and models: (a) Ruleof-Five (RO5), MW, cLogP, number of hydrogen-bond
donors (HBD), number of N and O atoms (NO), and number of RO5 violations; (b) topological polar surface area
(TPSA); (c) number of rotatable bonds (NRB); (d) LogD
(see Note 2); and (e) aqueous solubility (see Note 3).
6. Impose the desired molecular property profile for the virtual
products by setting computed property thresholds using the
PGVL Hub Decision Maker feature, as shown in Fig. 10.5.

198

Paderes et al.

Fig. 10.5. Virtual product property profiling within PGVL Hub. The upper threshold for cLogD and the lower threshold
for c_LogS were determined from Spotfire analysis of 2-aminoacetamide lead series. The upper threshold values for
MW, number of rotatable bonds (NRB), and polar surface area (TPSA) were user-specified parameters. The rest of the
thresholds (e.g., lower threshold for calculated LogD at pH 7.4 or the upper threshold for calculated Log of solubility) are
either the lowest or the highest property values of the virtual products.

In this design, the cutoff values include MW <480, NRB

10, TPSA <95, cLogD <2.0, and cLogS >3.5 (for the
latter two thresholds, see Note 4).
7. Export the resulting 279 predicted property compliant
virtual products as a structure file in SDF format for docking
simulations. Our objective was to narrow down the products
to 88 compounds to fill a single screening plate at Pfizer.
3.2. Energy Function
for ProteinLigand
Interaction

The energy function (24) used to predict the structure and energy
of the proteinligand complex contains an intermolecular term
for the interaction between the ligand and the protein and an
intramolecular term for the internal energy of the ligand. The

Library Design of 11-HSD1 Adamantyl Amide Inhibitors

BC D

b)

Energy

Energy

a)

199

B
C
Interatomic Distance

Interatomic Distance

Fig. 10.6. (a) Functional form of the hydrogen bond interaction energy (A = 15.0,
B = 2.3, C = 2.6, D = 3.1, E = 3.4, F = 4.0) and nonpolar dispersion (A = 15.0,
B = 0.93 , C= , D = 1.25 , E = 1.5 , F = 0.4), where is the sum of the atomic
radii of the protein and the ligand atoms. (b) Functional form of the repulsive interaction
(A = 15.0, B = 3.2, C = 6.0, F = 1.5). A and F are in kcal/mol and BE are in Angstroms.

intramolecular term includes the torsional and the van der Waals
functions of the DREIDING force field (40) and is useful in
differentiating between low- and high-energy ligand geometries
and in preventing internal ligand collapse (overlap between ligand atoms). The intermolecular term is a simplified intermolecular potential, which is a pairwise sum of piecewise linear potentials over all ligand and protein heavy (nonhydrogen) atoms (Fig.
10.6). Both the hydrogen-bonding and repulsive terms are modulated by a scaling factor based on the relative orientation of the
protein and the ligand atoms (Fig. 10.7). The piecewise linear
potentials for the hydrogen bond and steric interactions have the
same functional form but different parameters. This functional
form has the advantage of having a finite value compared to the
very high energy value in LennardJones potential, when the
interatomic distance approaches zero, thereby allowing the ligand to come in close contact with the protein during the early

H-bond strength

a)

1.0

90 120 180

c)

b)

d)

A
H L

Fig. 10.7. (a) Hydrogen bond strength is a function of the angle, , determined by the
relative orientation of the protein and ligand atoms. (b) A protein donor atom D bound
to one hydrogen atom H makes an angle with the ligand atom L. (c) A protein donor
atom D bound to the two hydrogen atoms H makes an angle with the ligand atom L.
(d) A protein acceptor atom A makes an angle with the ligand atom L.

200

Paderes et al.

stages of docking simulations. The parameters used in the piecewise linear potentials depend on the type of interaction and the
size of the atom. There are three types of interaction that arise
from four different protein and ligand atom types, as follows: (a)
hydrogen bond interactions between donors and acceptors, (b)
repulsive interactions between pairs of donors or acceptors, and
(c) steric interactions between nonpolar atoms or one nonpolar
and another atom type (Table 10.1). Every pair of interacting
atoms is assigned one of these three types of interaction. Atoms
are also assigned the atomic radii of 1.4, 1.8, and 2.2 corresponding to small (F, metal ions), medium (C, O, N), and large
(S, P, Cl, Br) atoms, respectively (36). These parameters were
derived from optimized interatomic distances observed in highquality crystal structures.

Table 10.1
Three types of interaction between ligand and protein heavy
atoms arising from different atom types. Primary and secondary amines are defined to be donors while oxygen and
nitrogen atoms with no bound hydrogens are defined to be
acceptors. Crystallographic water molecules and hydroxyl
groups are defined to be both donor and acceptor. Carbon
and sulfur atoms are defined to be nonpolar
Ligand

Donor

Acceptor

Both

Nonpolar

Repulsive

H-bond

H-bond

Steric

Acceptor

H-bond

Repulsive

H-bond

Steric

Both

H-bond

H-bond

H-bond

Steric

Nonpolar

Steric

Steric

Steric

Steric

Donor

3.3. Evolutionary
Search for Ligand
Exploration

Protein

Evolutionary programming (28, 29), based on a natural selection

process whereby a population of solutions competes for survival,
has been adopted as a search technique for finding the optimal
binding conformation of the ligand within the protein active site.
In this optimization, the population consists of floating point vectors encoding dihedral angles about rotatable bonds, with each
vector representing a potential ligand conformation. These dihedral angles are initialized to random values and are allowed to
vary during the optimization process. The energy barrier (25) to
rotation about a given bond, as defined by the DREIDING force
field (40), determines whether this bond is allowed to rotate during optimization (Table 10.2). The search process consists of a
fixed number of generation cycles. In each cycle, members of a
population of ligand conformations are scored using the above

Library Design of 11-HSD1 Adamantyl Amide Inhibitors

201

Table 10.2
Rotatable bond types with common threshold energy values
Threshold
energy value
(kcal/mol)

Rotatable bond type

1.0

sp2 sp3 bond only

2.0

Add sp3 sp3 bonds

5.0

Add sp2 sp2 single bonds

10.0

Add exocyclic aromatic resonant single bonds

25.0

Add resonant bonds

50.0

Add double bonds

energy function. A subset of the population is selected to become

parents for the next generation, with the remainder of the population discarded. These parents are then used to produce offspring,
thereby restoring the population to its original size. Selection of
parents is based on a stochastic competition wherein the energy
of each member of the population is compared with the energies
of a fixed number of randomly selected subset of the population.
A win is assigned to the member with the lowest energy and the
number of wins for each member is used to determine whether
it survives into the next generation. All surviving members produce offspring by Gaussian mutations of the dihedral angles of
the parent vectors. Mutation sizes are allowed to vary as the simulation progresses (24). In the final generation cycle, the best
scoring member is minimized using a conjugate gradient method
(41). This minimized structure corresponds to the predicted ligand conformation in the active site.
3.4. Docking
Simulations of Virtual
Library

In the present work, docking of our virtual library was conducted

using the available in-house gp11-HSD1 protein crystal structure. This protein structure was derived from its cocrystal complex (Fig. 10.4a) with an adamantyl amide analog that exhibits
enzyme activity in hu11-HSD1 (Ki(app) = 3.6 nM). It was
hypothesized that both gp11-HSD1 and hu11-HSD1 interact
in the same way with the ligand, forming hydrogen bonds with
the conserved Tyr-183 and Ser-170 residues and, that, both share
a similar hydrophobic pocket for accommodating the adamantyl
group. Moreover, the bound ligand is the Reactant A monomer
which serves as the anchoring template in our library design.
Thus, we used this compound as our reference ligand for defining the active site. Input to docking consisted of a protein PDB
file with the gp11-HSD1 protein crystal structure, a ligand SDF
file containing the virtual enumerated products to be docked, a

202

Paderes et al.

defining ligand PDB file with the gp11-HSD1 bound ligand

for defining the active site in the protein structure, and a core
structure PDB file (Fig. 10.4b) derived from the bound ligand. Docking simulations were performed using a docking script
which contains the following steps:
1. Prepare the ligand SDF file for docking by titrating at pH 7.
2. Run AGDOCK using the titrated ligand SDF file and the
protein PDB file with the following options:
(a) dl will use the first entry in the defining ligand file
for defining the search area within the active site of the
protein where the ligand will be docked; the search area
is defined as the minimum bounding rectangle of the
defining ligand extended with the cushion site definition
(b) core will use the core structure PDB file to specify
the coordinates of the core structure that will be used
in aligning the virtual products during partially fixed or
fixed anchor docking (see Note 5)
(c) cushion 2 to be added as cushion to extend the
minimum bounding box defined by the defining ligand
(see Note 6)
(d) maxbarrier this value is used to indicate what bonds
will be considered rotatable during the ligand conformational search; a value of 25 will allow rotation of conjugated single bonds in addition to sp2 sp3 , sp3 sp3 , nonconjugated sp2 sp2 single bonds, and exocyclic aromatic
single bonds (Table 10.2)
3. Run PLCALC to calculate the receptor-ligand interaction
(HT) scores (42) for each docked ligand pose in the above
output file and output the scored dock poses to an SDF file.
4. Evaluate dock results by invoking in-house utility tools that
execute the following:
(a) Retrieval of the best ligand conformations from the output file with scored dock poses based upon specified criteria. In this work, the five best conformations with the
lowest HT scores were specified for retrieval
(b) Sorting the best dock poses by HT scores and storing
the sorted ligands to an output SDF file
(c) Converting the output file into a table with compound
ID, HT scores, and other optional parameters
5. View dock poses in a 3D molecular viewing tool and select
dock hits based on predicted binding modes and HT scores
(see Note 7).
3.5. Library
Synthesis and
Purification

In a glove box, the alcohol (320 L, 80.0 mol, 1.0 eq, 0.25 M
in anhydrous DCE), TEA (33 L, 240 mol, 3.0 eq, neat TEA),

Library Design of 11-HSD1 Adamantyl Amide Inhibitors

203

DMAP (40 L, 8.0 mol, 0.1 eq, 0.2 M in anhydrous DCE), and
methanesulfonyl chloride (320 L, 160 mol, 2.0 eq, 0.5 M in
anhydrous DCM) were added to a 1095 mm test tube. The
test tube was sealed with a test tube cap and stirred in glove
box for 3 h at ambient temperature. The solvent was evaporated (SpeedVac or GeneVac, vacuum, medium heat, 16 h) and
the residue was dissolved in anhydrous DMF (400 L). TEA (80
L, 80.0 mol, 1.0 eq, 1 M in anhydrous DMF) and the amine
(480 L, 240.0 mol, 3.0 eq, 0.5 M in anhydrous DMF) were
added. The test tube was sealed with a test tube cap. The reaction
was heated and stirred at 80 C for 5 h. The solvent was evaporated (SpeedVac or GeneVac, vacuum, medium heat, 16 h), and
the residue was dissolved in DMSO (1.340 mL). The reaction
mixtures were analyzed by LCMS and the products isolated by
automated mass-directed HPLC.
All chromatographic separations were at ambient temperature. Analytical-scale separations were achieved using Agilent HP
TM
R
MSD systems with a Phenomenex Gemini C18 column
1100
TM
R
(4.6 50 mm ID, 5.0 m) or Agilent Zorbax
Extend C18
column (4.6 50 mm ID, 3.5 m). The mobile phase consisted
of water and acetonitrile, each with 0.05% trifluoroacetic acid and
was applied as linear gradient 0100% organic solvent in 3.0 or
1.75 min, depending on the column used. The MSD utilized positive mode APCI with a scan range from 100 to 1,000 amu. The
TM
mass-directed preparative HPLC was a Waters Fractionlynx system operating at 50 mL/min using the Gemini C18 stationary
phase in a 20 50 mm ID column. The mobile-phase solvents
were the same as the analytical scale with a 1 min hold to allow
for 1,200 L injection of the crude sample. The gradient was
TM
R
ZQ sin0100% organic in 5.4 min. The Waters Micromass
gle quad MS utilized positive mode electrospray ionization with
a 1:10,000 split from the preparative flow to the MS using a
methanol carrier fluid.
The library synthesis steps are as follows:
(1) Prepare an 811 array of 1095 mm test tubes in a test
tube rack.
(2) Add one 63 mm stir bar into each of the test tubes.
(3) Dry the rack of test tubes, the vials, and caps needed to
make the stock solutions at 100 C for 16 h (overnight).
Predried vials and caps must be used in subsequent steps.
(4) Transfer the rack of test tubes, the vials, and the caps into
a glove box until future use.
(5) In the glove box, prepare a 0.25 M stock solution of each
alcohol (Reactant A) in anhydrous DCE. Note: In case of
salt, equal amount of equivalents of TEA should be added.
(6) In the glove box, prepare a 0.5 M stock solution
of methanesulfonyl chloride (MW = 114.55) in anhydrous DCE.

204

Paderes et al.

(7) In the glove box, prepare a 0.2 M stock solution of DMAP

(MW = 122.2) in anhydrous DCE.
(8) Outside the glove box, prepare a 1 M stock solution of
TEA in anhydrous DMF for use in Step 21.
(9) Outside the glove box, prepare a 0.5 M stock solution of
each amine (Reactant B) in anhydrous DMF for use in
Step 22. Note: In case of salt, equal amount of equivalents
of TEA should be added.
(10) In the glove box, add 320 L (80 mol, 1.0 eq) of the
appropriate alcohol (Reactant A) solutions into the appropriate predried test tubes. Note: The reaction is sensitive to
the order of the addition of reagents.
(11) In the glove box, add 33 L (240 mol, 3.0 eq) of neat
TEA into each test tube.
(12) In the glove box, add 40 L (8 mol, 0.1 eq) of the
DMAP solution into each test tube.
(13) In the glove box, add 320 L (160 mol, 2.0 eq)
of the methanesulfonyl chloride solution into each test
tube.
(14) In the glove box, cover each test tube with a test tube cap.
(15) In the glove box, stir the reactions at ambient temperature
for 3 h.
(16) Take the rack of test tubes out of the glove box.
(17) Remove the volatiles and solvents from the reactions until
TM
TM
dryness using a GeneVac or SpeedVac (medium heat,
6 h).
(18) Add 400 L of anhydrous DMF to each test tube.
(19) Cover the test tubes with Parafilm.
(20) Vortex and sonicate the covered test tubes until the
residues are dissolved.
(21) To each test tube, add 80 L (80 mol, 1.0 eq) of the
TEA/DMF stock solution from Step 8.
(22) To each test tube, add 480 L (240 mol, 3.0 eq) of the
appropriate amines (Reactant B) solution from Step 9.
(23) Cap each test tube with a test tube cap.
(24) Transfer the test tubes to a test tube heating block that
has been preheated to 80 C.
(25) Stir the reactions in the test tubes at 80 C for 5 h.
(26) Remove the volatiles and solvents from the reactions until
dryness using a GeneVacTM or SpeedVacTM (medium
heat, 16 h).

Library Design of 11-HSD1 Adamantyl Amide Inhibitors

205

(27) Dissolve the residue in each test tube in 1340 L DMSO

(containing 0.01% BHT) to reach a final concentration of
0.0572 M.
(28) Using a liquid handler, transfer the contents of each
test tube to its corresponding well in a 2-mL 96well polypropylene deep-well plate for purification by
HPLC.
(29) Using a liquid handler, remove 5 L of the solution from
each well, dilute the aliquot to 1.0 mL with MeOH/water
95:5 in a new deep-well plate for LC/MS analysis.
(30) Seal each plate with an aluminum foil lid.
(31) Deliver each plate to its appropriate destination.
3.6. Library Assay
Results

Initial screening for enzymatic inhibition of human 11-HSD1

was carried out on the 88 raw products of the library. This
screening resulted in 100% hit rate with activity ranging from
76 to 103% inhibition at 0.2 M concentration. Of these 88
crude products, 37 compounds were purified and submitted
for enzymatic and cellular assays in human 11-HSD1. The
enzymatic Ki(app) was determined using human 11-HSD1
and labeled 3H-cortisone substrate in a triethanolamine buffer
in the presence of NADPH, glucose-6-phosphate, glucose-6phosphate dehydrogenase, and MgCl2 . The ratio of the resulting radioactive cortisone and cortisol was determined by radioHPLC and used to calculate the Ki(app) values (43). The
cellular (HEK293T-11_HSD1 Cell Reporter) EC50 was determined using human kidney cells that had been stably transfected
with human11_HSD1 gene, a reporter plasmid containing DNA
sequences of glucocorticoid-activated glucocorticoid receptors,
and a luciferase reporter gene (Luc), allowing for quantification of
11_HSD1 enzyme modulation (44). Cellular activity and hepatocyte stability results for the 37 purified compounds are shown in
Fig. 10.8. In Fig. 10.8a, calculated lipophilic efficiency (cLipE)
was plotted against cellular EC50 in order to rapidly identify
potent compounds that are highly lipophilic efficient. Lipophilic
efficiency (45, 46) is defined as pIC50 (or pKi) minus cLogP (or
LogD) and is a measure of how efficient the ligand is in terms of
achieving a high binding affinity (Ki) or cellular potency (EC50 )
without increasing the ligand lipophilicity (cLogP or LogD).
Increasing lipophilicity tends to increase binding affinity through
greater van der Waals interactions with the protein target, but it
also tends to promote binding to unwanted drug targets, leading to toxicity and other side effects. Thus, designing a highaffinity ligand that engages in other types of interactions (e.g.,
hydrogen bonding), while keeping the LogP or LogD low, has

206

Paderes et al.

a)

b)

N
N
N

PF-03440142
PF-03440171
hKiapp = 1.6 nM
hHEK_EC50 = 0.03 nM
cLogD = 1.65
Kin_Solubility = 355 uM
GL_hHepCl = 25.6 ul/min/million
GL_HLM = 158 ul/min/mg

hKiapp = 5.1 nM
hHEK_EC50 = 67 nM
eLogD = 0.91 nM
Kin_Solubility = 459 uM
GL_hHepCl = 5.0 ul/min/Million
GL_hLM = 9.7 ul/min/mg

Fig. 10.8. Cellular activity and hepatocyte stability of the 37 purified library compounds. (a) cLipE vs. human 11HSD1 EC50 (hHEKEC50). Chart enables rapid identification of the active and lipophilic efficient compounds (in lower right
corner), e.g., PF-03440171 with cLipE = 8.7 and EC50 = 0.03 nM. (b) Human hepatocyte intrinsic apparent clearance
(GL_hHepCl) vs. hHEKEC50. Chart enables rapid identification of active and metabolically stable compounds, such as
PF-03440142 with GL_hHepCl = 5 L/min/million and EC50 = 67 nM.

always been the goal for optimization. In general, a LipE range

of 57 or higher is desired based on the average oral drug with
potency in the range of 110 nM and cLogP 2.5 (46). Since
the experimental LogD values for the purified products were not
available, calculated LogD (cLogD) was used in estimating LipE.
Hence, cLipE was calculated as the negative Log10 EC50 minus
cLogD and the values range from 3 to 8.9 with the most cellular potent compound having the highest cLipE value (cLipE
= 8.87, EC50 = 0.03 nM), albeit with a low stability in HHep
(Fig. 10.8a). A plot of the human hepatocyte clearance versus
cellular activity (Fig. 10.8b) shows four compounds with good
stability (GL_hHepCl < 10 L/min/million) and with high to
moderate activity. The library was able to achieve its objective
by giving rise to PF-03440142, with improved cellular activity
(EC50 = 67 nM) and retention of favorable solubility (kinetic
solubility = 459 M), while maintaining stability in human liver
microsomes (GL_HLM = 9.7 L/min/mg) and human hepatocytes (GL_hHepCl = 5 L/min/million).

Library Design of 11-HSD1 Adamantyl Amide Inhibitors

207

4. Notes
1. Structural alerts (STA) exemplify substructures which
have been associated with multiple examples of adverse in
vivo events (47). Information on the chemistry, pharmacology, and toxicology for these substructures has been collated to provide a strong rationale for their classification as
STA due to their intrinsic reactivity, ability to intercalate to
DNA, coordination with a metal, metabolic activation, or
transformation to a species capable of covalently binding to
biological macromolecules. A common example of STA is
the aniline functional group which can get oxidized either at
the aniline nitrogen or at the ortho and para aromatic carbon atoms to form the reactive nitroso and iminoquinone
intermediates, respectively, that can lead to increased risk
of mutagenicity, in vivo carcinogenicity, and hepatoxicity.
Hence, chemists who design libraries must be aware of all
these potential liabilities and understand the chemistry and
mechanism involved in the metabolism of the design compounds. If the purpose of the library is to probe the receptor pocket to find novel chemotypes, then compounds with
STA may be included in the design, provided they can be
modified later to eliminate the risks. In general, chemists are
advised to avoid compounds with STA in order to minimize
the risk of adverse outcome, thereby ensuring the safety and
efficacy of drugs.
2. PGVL Hub has access to a plethora of computational models for predicting molecular properties. In this library design,
the ACD LogD was used in profiling the virtual products.
LogD is an experimental measurement of lipophilicity, where
D is a pH-dependent parameter which changes with the
degree of ionization. The ACD LogD was selected after
careful comparison with alternative LogD predictors, such
as the Pallas LogD and the in-house Cubist models based
on Pfizer experimental LogD. Chemists are encouraged to
test different LogD models to see how well they correlate
with the experimental data of their lead series.
3. The legacy aqueous thermodynamic solubility model is an
in-house linear regression model consisting of 11 calculated
descriptors which include cLogP and polar surface area. This
is the model that was used to shape the property profile of
the virtual products. This model has now been superseded
by a Cubist (48) thermodynamic solubility model (R2 =
0.93 and RMSE = 0.44, N = 2794) based on a data set
of 3,075 compounds with intrinsic solubility. This model

208

Paderes et al.

calculates the intrinsic solubility and provides an estimate

of the apparent solubility at a given pH based on the computed intrinsic solubility and calculated pKa using the
ACD software. The accuracy of the model is highly dependent on the pKa prediction given by the ACD software.
This model has been validated using Cubist (R2 = 0.86 and
RMSE = 0.61, N = 281).
4. The multiproperty filtering criteria used in profiling the
virtual library were derived empirically from Spotfire analysis of biological and physicochemical data for existing
compounds belonging to the 2-aminoacetamide lead series
(Fig. 10.9). These compounds consisted mostly of
pyrrolidine-2-carboxylic acid N-(R1)-substituted amides
(Fig. 10.9a) and morpholine-3-carboxylic acid N-(R1)substituted amides (Fig. 10.9b), where R1 is an adamantyl,
a cycloalkyl, a benzyl, an aryl, or a heteroaryl group. For
our in-house 11-HSD1 inhibitors, correlations of biological stability endpoints with physicochemical parameters have
been shown to be lead-series specific. Hence, we selected
the 2-aminoacetamide lead series, which is closest to our
adamantyl amide virtual library, as our training set to guide
the selection of parameters to be used for filtering this
library. For example, a plot of the experimental LogD values (eLogD) versus human liver microsome stability measured as % remaining (HLM_%Rem@1M) shows that in
order to achieve our laboratory objective of designing compounds with HLM_%Rem@1M > 70%, an eLogD cutoff
value of <2.7 is desired since this includes the highly stable
compounds (Fig. 10.10a). Since we are using the calculated
LogD at pH 7.4 (LogD_pH74) for the virtual products, a
plot of the experimental versus calculated LogD translates
this cutoff to <2.0 (red horizontal line in Fig. 10.10b). For
stability in human hepatocytes, we have two types of biological endpoints, % remaining (HHEP_%Rem@1M) and
intrinsic clearance (HHEP_CL(int)_L/min/M). In order
to achieve the desired stability in HHep, i.e., >70% remaining or <3 L/min/million, a calculated Log of solubility
(c_LogS) threshold of >3 is desired (Fig. 10.11). Along
the same vein, a cLogS >3 based on HLM_%Rem@1M
plot is required to achieve >70%, while a cLogS >4 based
on the HLM_CL(int)_L/min/mg plot is needed to satisfy
the laboratory objective of < 30 L/min/mg (Fig. 10.12).
In the library design, we used c_LogS >3.6 as our threshold for filtering the virtual library for compounds with the
desired property profile.
5. Fixed anchor docking allows the docking of one part of a ligand while keeping the portion of the ligand that is primarily

Library Design of 11-HSD1 Adamantyl Amide Inhibitors

R3

H
R1

R2

H
R1

N
O

a)

209

R2

b)

Fig. 10.9. Aminoacetamide lead series used in establishing thresholds for solubility and
metabolic stability to guide the library design. R1 can be adamantyl, cycloalkyl, benzyl
or substituted benzyl, aryl, or heteroaryl. R2 can be alkyl, substituted alkyl, cycloalkyl,
benzyl, substituted benzyl, or acetyl. R3 can be H or OH.

eLogD vs. Calculated LogD at pH7.4

eLogD vs. HLM (%R)

b)

a)
6

100

80

(2.62, 68.8)
60

40

20
2
2

0
1

eLogD

eLogD

Fig. 10.10. (a) Experimental LogD (eLogD) vs. human liver microsome stability (HLM_%Rem@1M). A threshold value
of eLogD < 2.7 is required for >70% stability in HLM. (b) Experimental vs. calculated LogD. eLogD < 2.7 translates to
cLogD < 2.0 at pH 7.4.

responsible for molecular recognition fixed within the active

site (36, 37). In the current work, the adamantyl amide moiety acts as a molecular anchor, with the adamantyl group
occupying a specific lipophilic pocket within the enzyme
active site and with the amide carbonyl oxygen atom forming hydrogen bond interactions with the conserved Ser170 and Tyr-183 residues (Fig. 10.13). This computational
approach will work only if the binding mode of the anchor
fragment is not significantly affected by the different sub-

210

Paderes et al.
H HEP (%R) vs. c_LogS
0

CL(int) HHep vs. c_LogS

a)

b)

1
1
2
2
3

5
4
6
5
7

9
20

40

60

80

HHEP_%Rem@1uM

100

10

20

30

40

50

60

HHEP_CL(int)_uL/min/M

Fig. 10.11. (a) Experimental human hepatocyte stability (HHEP_%Rem@1M) vs. calculated Log of solubility (c_LogS).
(b) Experimental human hepatocyte stability expressed as apparent intrinsic clearance (HHEP_CL(int)_L/min/million)
vs. calculated Log of solubility (c_LogS). A threshold value of cLogS > 3.0 is required for stability in human
hepatocytes.

stituents in the analogs. One must be careful when selecting a ligand fragment to fix during docking since not all
ligand fragments can act as molecular anchors. A molecular anchor is characterized by a specific binding mode with
a dominant free energy minimum and a large stability gap,
defined as the free energy of the crystal binding mode relative to the free energy of alternative binding modes (26). An
advantage of fixed anchor docking is that the large number
of degrees of freedom due to ligand flexibility is drastically
reduced and that calculation of the free energy of binding for
close analogs containing the anchor fragment is significantly
facilitated.
6. During docking, the ligand is required to remain in a rectangular box that encompasses the active site. Ligand conformations and orientations are searched via an evolutionary
programming algorithm within this rectangular box. A constant energy penalty is added to every ligand atom outside
this box. If the virtual library of compounds contain a lot
of large substituents (Reactant B), it is advisable to increase
this cushion to a larger value in order to accommodate the

Library Design of 11-HSD1 Adamantyl Amide Inhibitors

CL(int) HLM vs. c_LogS

H LM (%R) vs. c_LogS

a)

211

b)

-1
1
2
2
3

5
4
6
5
7

9
0

20

40

60

80

10

100

20

30

40

50

60

70

80

90

HLM_CL(int)_uL/min/mg

HLM_%Rem@1uM

Fig. 10.12. (a) Experimental human liver microsome stability (HLM_%Rem@1M) vs. calculated Log of solubility (c_LogS). (b) Experimental human liver microsome stability expressed as apparent intrinsic clearance
(HLM_CL(int)_L/min/mg) vs. calculated Log of solubility (c_LogS). A threshold value of cLogS > 3 and 4.0 is
required for stability in HLM based on %R and intrinsic clearance, respectively.

Ser-A170

Tyr-A177
Tyr
O

O
N
N

Tyr-A183

Pro-A178

Fig. 10.13. Examples of dock poses from fixed anchor docking of the virtual library in
gp11-HSD1 crystal structure.

212

Paderes et al.

various conformations of the larger ligands and minimize the

energy penalty.
7. While the simplified piecewise linear potential energy function is able to reproduce the crystallographic bound complexes and predict the structure of the bound ligand in the
protein active site, the current high-throughput (HT) scoring function (42) is not sufficient to predict the free energy
of binding accurately. Hence, HT scores (42) must be interpreted with caution since these do not necessarily correlate
with binding affinities, especially for structurally diverse ligands. In the case of the current library design in which the
binding mode of the adamantyl amide anchor is likely to be
preserved in the docked analogs, the HT scores represent
the free energy differences in the substituents and can be
used in weeding out the least active compounds from the
virtual library. After visual inspection of the predicted binding modes, 82 virtual compounds with HT scores ranging
from 6 to 8 were selected along with six others for the
library design.

Acknowledgments
The authors would like to thank Simon Bailey, Martin Edwards,
and Michael McAllister for their valuable advice, encouragement,
and guidance. Specifically, the authors are grateful to Stanley
Kupchinsky for the synthesis of the starting adamantyl amide
lead and to the Discovery Computation group at PGRD La
Jolla for the development of PGVL and AGDOCK, under the
leadership of Atsuo Kuki and Peter Rose, respectively. Thanks
are especially due to the following colleagues who developed
and performed our project assays, specifically, Jacques Ermolieff (11-HSD1 enzyme assays); Andrea Fanjul (11-HSD1 cellular assays); Nora Wallace, Christine Taylor, and Rob Foti
(HLM assays); and Veronica Zelesky, Kevin Whalen, and Walter
Mitchell (HHEP assays). This work was supported by the 11HSD1 project team and the Pfizer Diabetes Therapeutic Area
management.
References
1. Charmandari, E., Kino, T., Chrousos, G.
P. (2004) Glucocorticoids and their actions:
an introduction. Ann N Y Acad Sci 1024,
18.

2. Tomlinson, J. W., Walker, E. A., Bujalska,

I. J., Draper, N., Lavery, G. G., Cooper,
M. S., Hewison, M., Stewart, P. M. (2004)
11-Hydroxysteroid dehydrogenase type 1:

Library Design of 11-HSD1 Adamantyl Amide Inhibitors

3.

4.

5.

6.

7.

8.

9.

10.

11.

a tissue-specific regulator of glucocorticoid

response. Endocr Rev 25(5), 831866.
Draper, N., Stewart, P. M. (2005)
11-Hydroxysteroid dehydrogenase and
the pre-receptor regulation of corticosteroid hormone action. J Endocrinol 186,
251271.
Walker, E., Stewart, P. M. (2003) 11Hydroxysteroid dehydrogenase: unexpected
connections. Trends Endocrinol Metab 14,
334339.
Masuzaki, H., Paterson, J., Shinyama, H.,
Morton, N. M., Mullins, J. J., Seckl J. R.,
Flier, J. S. (2001) A transgenic model of visceral obesity and the metabolic syndrome.
Science 294, 21662170.
Kotelevtsev, Y., Holmes, M. C., Burchell,
A., Houston, P. M., Schmoll, D., Jamieson,
P., Best, R., Brown, R., Edwards, C.
R. W., Seckl, J. R., Mullins, J. J.
(1997) 11-Hydroxysteroid dehydrogenase
type 1 knockout mice show attenuated
glucocorticoid-inducible responses and resist
hyperglycemia on obesity or stress. Proc Natl
Acad Sci USA 94, 1492414929.
Morton, N. M., Holmes, M. C., Fievet C.,
Staels, B., Tailleux, A., Mullins, J. J., Seckl,
J. R. (2001) Improved lipid and lipoprotein
profile, hepatic insulin sensitivity, and glucose tolerance in 11-hydroxysteroid dehydrogenase type 1 null mice. J Biol Chem 276,
4129341300.
Rask, E., Walker, B. R., Soderberg, S., Livingstone, D. E. W., Eliasson, M., Johnson, O., Andrew, R., Olsson, T. (2002)
Tissue-specific changes in peripheral cortisol metabolism in obese women: increased
adipose 11-hydroxysteroid dehydrogenase
type 1 activity. J Clin Endocrinol Metab 87,
33303336.
Paulmyer-Lacroix, O., Boullu, S., Oliver, C.,
Alessi, M. C., Grino, M. (2002) Expression
of the mRNA coding for 11-hydroxysteroid
dehydrogenase type 1 in adipose tissue from
obese patients: an in situ hybridization study.
J Clin Endocrinol Metab 87, 27012705.
Kannisto, K., Pietilainen, K. H., Ehrenborg,
E., Rissanen, A., Kaprio, J., Hamsten, A.,
Yki-Jarvinen, H. (2004) Overexpression of
11-hydroxysteroid dehydrogenase-1 in adipose tissue is associated with acquired obesity and features of insulin resistance: studies in young adult monozygotic twins. J Clin
Endocrinol Metab 89, 44144421.
Abdallah, B. M., Beck-Nielsen, H., Gaster,
M. (2005) Increased expression of 11hydroxysteroid dehydrogenase type 1 in type
2 diabetic myotubes. Eur J Clin Invest 35,
627634.

213

12. Valsamakis, G., Anwar, A., Tomlinson, J.

W., Shackleton, C. H. L. McTernan, P. G.,
Chetty, R., Wood, P. J., Banerjee, A. K.,
Holder, G., Barnett, A. H., Stewart, P. M.,
Kumar, S. (2004) 11-hydroxysteroid dehydrogenase type 1 activity in lean and obese
males with type 2 diabetes mellitus. J Clin
Endocrinol Metab 89, 47554761.
13. Walker, B. R., Connacher, A. A., Lindsay R.
M., Webb, D. J., Edwards, C. R. (1995) Carbenoxolone increases hepatic insulin sensitivity in man: a novel role for 11-oxosteroid
reductase in enhancing glucocorticoid receptor activation. J Clin Endocrinol Metab 80,
31553159.
14. Andrews, R. C., Rooyackers, O., Walker, B.
R. (2003) Effects of the 11-hydroxysteroid
dehydrogenase inhibitor carbenoxolone on
insulin sensitivity in men with type 2 diabetes.
J Clin Endocrinol Metab 88, 285291.
15. Barf, T., Williams, M. (2006) Recent
progress in 11b-hydroxysteroid dehydrogenase type 1 (11b-HSD1) inhibitor development. Drugs Future 31(3), 231243.
16. Barf, T., Vallgarda, J., Emond, R.,
Haggstrom, C., Kurz, G., Nygren, A., Larwood, V., Mosialou, E., Axelsson, K., Olsson,
R., Engblom, L., Edling, N., Ronquist-Nii,
Y., Ohman, B., Alberts, P., Abrahmsen, L.
(2002) Arylsulfonamidothiazoles as a new
class of potential antidiabetic drugs. Discovery of potent and selective inhibitors of the
11-hydroxysteroid dehydrogenase type 1.
J Med Chem 45, 38133815.
17. Hult, M., Shafqat, N., Elleby, B., Mitschke,
D., Svensson, S., Forsgren, M., Barf, T.,
Vallgarda, J., Abrahmsen, L., Oppermann,
U. (2006) Active site variability of type 1
11b-hydroxysteroid dehydrogenase revealed
by selective inhibitors and cross-species
comparisons. Mol Cell Endocrinol 248,
2633.
18. Xiang, J., Ipek, M., Suri, V., Massefski, W.,
Pan, N., Ge, Y., Tam, M., Xing, Y., Tobin,
J. F., Xu, X., Tam, S. (2005) Synthesis and
biological evaluation of sulfonamidooxazoles
and B-keto sulfones: selective inhibitors of
11-hydroxysteroid dehydrogenase type I.
Bioorg Med Chem Lett 15, 28652869.
19. Olson, S., Balkovec, J., Gao, Y. -D,
et al. (2004) Selective inhibitors of 11hydroxysteroid dehydrogenase type 1.
Adamantyl triazoles as pharmacological
agents for the treatment of metabolic syndrome. Keystone Symp Abst X2239.
20. Berwaer, M. (2004) Promising new targets.
The therapeutic potential of 11-HSD1 inhibition. 6th Annu Conf Diabetes (Oct. 1819,
London).

214

Paderes et al.

21. Webster, S. P., Ward, P., Binnie, M., Craigie,

E., McConnell, K. M. M., Sooy, K., Vinter,
A., Seckl, J. R., Walker, B. R. (2007) Discovery and biological evaluation of adamantyl
amide 11-HSD1 inhibitors. Bioorg Med
Chem Lett 17, 28382843.
22. Becker, D. P., Flynn, D. L., Villamil, C.
I. (2004) Bridgehead-methyl analog of SC53116 as a 5-HT4 agonist. Bioorg Med Chem
Lett 14(12), 30733075.
23. Reddy, P. G., Baskaran, S. (2004) Epoxideinitiated cationic cyclization of azides: a novel
method for the stereoselective construction
of 5-hydroxymethyl azabicyclic compounds
and application in the stereo- and enantioselective total synthesis of (+)- and ()
-indolizidine 167B and 209D. J Org Chem
69, 30933101.
24. Gehlhaar, D. K., Verkhivker, G. M., Rejto,
P. A., Sherman, C. J., Fogel, D. B., Fogel,
L. J., Freer, S. T. (1995) Molecular recognition of the inhibitor AG-1343 by HIV1 protease: conformationally flexible docking
by evolutionary programming. Chem Biol 2,
317324.
25 Gehlhaar, D. K., Bouzida, D., Rejto, P. A.
(1998) Fully automated and rapid flexible
docking of inhibitors covalently bound to
serine proteases. Proceedings of the 7th International Conference on Evolutionary Programming, MIT Press, Cambridge, MA., pp.
449461.
26. Rejto, P. A., Verkhivker, G. M. (1998)
Molecular anchors with large stability gaps
ensure linear binding free energy relationships for hydrophobic substituents. Pacific
Symp Biocomput 1998, 362373.
27. Bouzida, D., Rejto, P. A., Arthurs, S., Colson, A. B., Freer, S. T., Gehlhaar, D.
K., Larson, V., Luty, B. A., Rose, P W.,
Verkhivker, G. M. (1999) Computer simulations of ligand-protein binding with ensembles of protein conformations: a Monte
Carlo study of HIV-1 protease binding
energy landscapes. Intl J Quantum Chem 72,
7384.
28. Fogel, D. B. (1995) Evolutionary Computation: Toward a New Philosophy of Machine
Intelligence, IEEE Press, Piscataway.
29. Fogel, L. J., Owens, A. J., Walsh, M. J.
(1966) Artificial Intelligence Through Simulated Evolution, Wiley, New York.
30. Ogg, D., Elleby, B., Norstrom, C., Stefansson, K., Abrahmsen, L., Oppermann, U.,
Svensson, S. (2005) The crystal structure of
guinea pig 11-hydroxysteroid dehydrogenase type 1 provides a model for enzymelipid bilayer interactions. J Biol Chem 280,
37893794.

31. Hosfield, D. J., Wu, Y., Skene, R. J., Hilgers,

M., Jennings, A., Snell, G. P., Aertgeerts,
K. (2005) Conformational flexibility in crystal structures of human 11-hydroxysteroid
dehydrogenase type 1 provide insights into
glucocorticoid interconversion and enzyme
regulation. J Biol Chem 280, 46394648.
32. Zhang, J., Osslund, T. D., Plant, M. H.,
Clogston, C. L., Nybo, R. E., Xiong, F.,
Delaney, J. M., Jordan, S. R. (2005) Crystal structure of murine 11-hydroxysteroid
dehydrogenase 1: an important therapeutic target for diabetes. Biochemistry 44,
69486957.
33. Polinsky, A., Feinstein, R. D., Shi, S., Kuki,
A. (1996) LiBrain: software for automated
design of exploratory and targeted combinatorial libraries. Colorado Conf, Chapter 20,
219232.
34. Hoorn, W. P., Bell, A. S. (2009) Searching
chemical space with the Bayesian idea generator. J Chem Inf Model 49(10), 22112220.
35. Lee, P. H., Cucurull-Sanchez, L., Lu, J., Du,
Y. J. (2007) Development of in silico models
for human liver microsomal stability. J Comput Aided Mol Des 21(12), 665673.
36. Gehlhaar, D. K., Bouzida, D., Rejto,
P. A. (1999) Reduced dimensionality in
ligand-protein structure prediction: covalent
inhibitors of serine proteases and design of
site-directed combinatorial libraries. Proceedings of the Division of Computers in Chemistry,
ACS. Chapter 19, pp. 292310.
37. Bouzida, D., Gehlhaar, D. K., Rejto, P. A.
(1997) Application of partially fixed docking
towards automated design of site-directed
combinatorial libraries. ACS National Meeting, COMP 156.
38. Bouzida, D., Arthurs, S., Colson, A. B.,
Freer, S. T., Gehlhaar, D. K., Larson, V.,
Luty, B. A., Rejto, P. A., Rose, P. W.,
Verkhivker, G. M. (1999) Thermodynamics and kinetics of ligand-protein binding
studied with the weighted histogram analysis method and simulated annealing. Pacific
Symp Biocompu, pp. 426437.
39. Weiner, S. J., Kollman, P. A., Case, D. A.,
Singh, U. C., Ghio, C., Alagona, G., Profeta, S. Jr., Weiner, P. (1984) A new force
field for molecular mechanical simulation of
nucleic acids and proteins. J Am Chem Soc
106, 765784.
40. Mayo, S. L., Olafson, B. D., Goddard, W. A.
III (1990) DREIDING: a generic force field
for molecular simulations. J Phys Chem 94,
88978909.
41. Press, W. H., Teukolsky, S. A., Vettering,
W T., Flannery, B. P. (1992) Numerical
Recipes in C. The Art of Numerical Com-

Library Design of 11-HSD1 Adamantyl Amide Inhibitors

puting, 2nd ed. Cambridge University Press,
Cambridge.
42. Marrone, T. J., Luty, B. A., Rose,
P. W. (2000) Discovering high-affinity ligands from the computationally predicted
structures and affinities of small molecules
bound to a target: a virtual screening
approach. Perspect Drug Discovery Design,
20, 209230.
43. Castro, A., Zhu, J. X., Alton, G. R., Rejto, P.,
Ermolieff, J. (2007) Assay optimization and
kinetic profile of the human and the rabbit
isoforms of 11b-HSD1. Biochem Biophys Res
Commun 357(2),561566.
44. Bhat, B. G., Hosea, N., Fanjul, A., Herrera, J., Chapman, J., Thalacker, F., Stewart,
P. M., Rejto, P. A. (2008) Demonstration
of proof of mechanism and pharmacokinetics and pharmacodynamic relationship
with 4 -cyanobiphenyl-4-sulfonic acid(6amino-pyridin-2-yl)amide (PF-915275), an

45.

46.

47.
48.

215

inhibitor of 11b-hydroxysteroid dehydrogenase type 1, in cynomolgus monkeys.

J Pharm Exp Ther 324(1), 299305.
Ryckmans, T., Edwards, M. P., Horne,
V. A., Correia, A. M., Owen, D. R., Thompson, L. R., Tran, I., Tutt, M. F., Young, T.
(2009) Rapid assessment of a novel series of
selective CB2 agonists using parallel synthesis protocols: a lipophilic efficiency (LipE)
analysis. Bioorg Med Chem Lett 19(15),
44064409.
Leeson, P. D., Springthorpe, B. (2007) The
influence of drug-like concepts on decisionmaking in medicinal chemistry. Nat Rev
Drug Disc 6(11), 881890.
Blagg, J. (2006) Structure-activity relationships for in vitro and in vivo toxicity. Ann
Reps Med Chem 41, 353368.
Quinlan, J. R. (1992) Learning with continuous classes. In Proc. AI 92, Adams, Sterling,
Eds., 343348.

Section III
Fragment-Based Library Design

Chapter 11
Design of Screening Collections for Successful
Fragment-Based Lead Discovery
James Na and Qiyue Hu
Abstract
A successful fragment-based lead discovery (FBLD) campaign largely depends on the content of the
fragment collection being screened. To design a successful fragment collection, several factors must be
considered, including collection size, property filters, hit follow-up considerations, and screening methods. In this chapter, we will discuss each factor and how it was applied to the design and assembly of one
or more fragment collections in a major pharmaceutical company setting. We will also present examples
and statistics of screening results from such collections and how subsequent collections can be improved.
Lastly, we will provide a summary comparison of selected fragment collections from literature.
Key words: Fragment-based lead discovery, screening collection, library design, computational
filtering, NMR screening

1. Introduction

In the past decade, fragment-based drug discovery (FBDD), or

fragment-based lead discovery (FBLD), has become an exciting
way for the pharmaceutical industry to discover new medicines
(15). In addition to biochemical assays, fragment screening
takes advantage of several other screening technologies, including NMR (nuclear magnetic resonance), MS (mass spectroscopy),
SPR (surface plasmon resonance), X-ray crystallography, and various forms of calorimetry. Several clinical candidates can trace their
origin to FBLD from different screening methods; a recent review
by de Kloe et al. provided several interesting examples (6).
While most pharmaceutical and biotech companies utilize
high-throughput screening (HTS) as their primary assay, FBLD
offers numerous advantages. Compared with HTS, there is significantly fewer compounds to be screened. There are typically
thousands of compounds for a fragment screen versus hundreds

J.Z. Zhou (ed.), Chemical Library Design, Methods in Molecular Biology 685,
DOI 10.1007/978-1-60761-931-4_11, Springer Science+Business Media, LLC 2011

219

220

Na and Hu

of thousands or more for an HTS campaign. This reduced set

of compounds results in time and resources savings for an FBLD
screen as compared to an HTS screen. The smaller compound size
also means a relatively small set of fragments can cover a larger
chemical space than a typical HTS collection (7). Moreover, fragment screens generally result in much higher hit rates than HTS
campaigns, and often the hits are novel with respect to the HTSderived chemical series. Lastly, a distinct advantage of FBLD is
that often orthogonal screening methods are used as confirmation, e.g., NMR screen followed by SPR or MS, so chances of
false positives are diminished.
The drawbacks of FBLD are that not all targets are amenable
to FBLD, and that fragment hits are usually much weaker binders
than typical HTS hits, generally in the millimolar to high micromolar range. There are also concerns about whether the hit compounds are binding at the desired pocket or at a random hotspot
on the protein, although this can be resolved by competitive binding experiments or by crystallography.
Because fragment hits are smaller in size compared with a
more drug-like HTS hit, there is much more room for the
medicinal chemists to shape them into more novel leads and eventually lead series. This task is made easier if the fragment hits contain chemistry vectors for elaboration, which can be built in when
assembling a fragment screening collection.
A successful FBLD campaign requires a fragment library possessing several important characteristics, including proper collection size, good physicochemical properties, chemical diversity,
and chemistry follow-up considerations. There are a number of
publications describing the process of building a fragment collection, from a general collection to collection tailored for a specific
screening method (812). In this chapter, we will discuss some of
the factors to consider when assembling a fragment collection. In
addition, we will describe the process of how two fragment collections were assembled, and analyze the screening results from
multiple fragment screens.

2. Factors to
Consider in
Creating a
Fragment
Collection
2.1. Collection Size

Perhaps the biggest distinction between an HTS and a fragment

collection is the size of each collection. A typical HTS collection can contain 105 107 compounds, whereas a fragment collection can range in size from 102 to 104 compounds. While recent

Successful Fragment-Based Lead Discovery

221

advances have greatly increased the screening capabilities of HTS,

the cost factor can still be argued to favor fragment screening
in terms of assembling the collection and the compounds and
reagent resources consumed per screening campaign. The size of
a fragment collection can be created based on the targets being
pursued, e.g., a set of fragments most likely to be hits against
kinases. In such cases the collection can be relatively small, numbering in the hundred of compounds. Alternatively, the collection
size can also be dictated by the screening method. For example,
if the screening method is a high-concentration screening (HCS)
using biochemical assay with good throughput, then the size of
the collection can be relatively large with numbers in the thousands or even 1020 K compounds. In general, if the desire is to
build a generic collection, then the number of compounds can
typically number from 5 to 20 K, whereas a screening method or
a biased collection will be smaller by one order of magnitude in
size.
2.2. Physicochemical
Properties

Solubility is an important factor to consider in selecting fragments, especially when the assay of choice is high-concentration
screening. This is often the case since fragments are typically weak
binders with low millimolar to high micromolar activity. In most
instances the solvent used tends to be more polar, such as DMSO
or water. Therefore, the fragments have to be water soluble, and
compounds having ionizing groups or polar functions favor solubility. There are various methods to calculate the solubility of
a compound, see a recent review for details (13). These calculated values can be used to guide the selection of compounds for
a given collection.
Besides solubility, other physicochemical factors to be considered in building a fragment collection are molecular weight, number of heavy atoms, rotatable bonds, lipophilicity, and polar surface area. All or some of these factors can be accounted for when
building a collection, although molecular weight is one factor
that is almost always considered. In general the MW range for a
fragment collection should fall within 120300, with the median
MW at around 200250. Compounds with MW much less than
150 are undesirable due to higher risk of unspecific or undetectable binding, while compounds larger than 300 are becoming more full-size molecules rather than fragments. Attenuating the MW range would also affect the number of heavy atoms,
a closely related molecular property. For the number of rotatable bonds, typically a fragment with MW around 250 would
have 03 rotatable bonds, which in general is a desirable range
for a fragment collection. Another important factor to consider is
lipophilicity, usually measured as logP, which can have a big influence on binding affinity. The generally accepted range for logP is
03 (14).

222

Na and Hu

2.3. Hit Follow-Up

Consideration

One of the less common factors considered when building a fragment collection is synthetic attractiveness of the fragments, or put
differently, ease of hit follow-up for the hit-to-lead process. In
general, chemists like to have hits which present opportunities
for making analogs, but often fragments lack the synthetic handle(s) that a chemist desires. It would be relatively easy to assemble a fragment collection from a pool of reagent-type compounds
which contain one or multiple reactive centers. However, some
chemistry savvy must be exercised to avoid highly reactive functional groups such as sulfonyl halides or isocyanates, which can
react with the protein side chains. Because fragments are often
screened as a mixture, care also must be taken to avoid compounds which may react with each other when mixed together.
Another factor to consider is that most reactive functional groups
can also elicit binding interaction, and this interaction can be
altered or destroyed altogether when reaction occurs at the site.
For example, the bromine of an alkyl bromide can elicit a binding
interaction with the protein target which can result in the compound being a screening hit. However, if the bromine is used as
a chemistry vector for elaboration, the bromine is then displaced
upon reaction which eliminates the bromineprotein interaction,
which may be an important part of the overall binding interaction. Hence, the selection of compounds which contain reactive
functional groups to be included in a screening collection must
be done carefully, preferably with input from medicinal/synthetic
chemists.
Less reactive synthons which also facilitate the hit follow-up
process contain chemistry handles or functional groups on a
molecule which can be easily converted to grow or shape the
molecule. One of the more useful functional groups for this purpose are halo-aromatic compounds, in which the halogen (or even
the aromatic hydrogens) can often be the chemistry vector that
allows the chemist to explore other parts of the binding pocket.
Primary amines and carboxylic acids are other functional groups
that can be considered useful chemistry handles to be included in
a fragment collection. When screened against a protein target, the
binding interaction from these functional groups can mostly be
preserved even after they are used as chemistry vectors for elaboration. Figure 11.1 lists commonly used functional groups which
serve as chemistry handles in a fragment.
Sometimes, reactive functional groups are protected. For
instance, a novel primary amine can be protected by reacting
with a small acid (e.g., acetic acid), and the resulting amide would
still retain some of the characteristics of the original amine where
both the amino R-group and the resulting amide reaction product can elicit binding interaction with the protein. And if this pro-

Successful Fragment-Based Lead Discovery

O

O
Acid/Ester

R1

OH

R2

Ar

R1

NH2

NH

NH2

Amine

223

R2
O

O
Amide

O
Sulfonamide

Alcohol/Phenol

Thiol

R1

NH2

O
S

Aliph

N
H

NH2

O
S

R1

Ar

A
A

N
H
O

R2

H
H

Activated CH

R2

A
A

N
H

A = C, N
Aromatic halide

Nitrile

X
Arom
x = F, Cl, Br

N
R

Fig. 11.1. Functional groups that are useful as chemistry handles in a fragment.

tected compound becomes a screening hit, the amide moiety can

becomes a useful chemistry vector while preserving the possible
binding interactions from the amide itself. Note that the choice
of protecting groups for reactive monomers as fragments must
be carefully selected so as to limit MW of the resulting product to
stay within the desirable MW range for fragments.
Of course protecting the reactive functionality of a compound
alters its binding characteristic. Therefore, selecting the reactive
monomer types to be included in a fragment collection and selecting its protecting group must be carefully considered and chosen
wisely.

224

Na and Hu

3. Building
a Fragment
Collection

3.1. Pfizer Fragment

Screening
Collections

There are numerous ways to build a fragment collection. Consideration of factors described in the earlier sections can be used
to build a generic collection or to build a collection tailored to a
specific screening method or a particular target. One of the more
popular and rational approaches to assemble a fragment collection is to build the collection based on the screening technique.
Among the more popular fragment screening methods are NMR
techniques, beginning with the SAR by NMR method pioneered by Fesik et al. at Abbott (5). Other methods using NMR
include saturation transfer difference (STD) (15, 16) and waterLOGSY (17). There have been several efforts to assemble fragment collection based on NMR screening techniques (11, 18).
In the following sections, we will describe in detail efforts to
build fragment collections and the processes involved in their creation. Two such efforts were performed at a major pharmaceutical
company (Pfizer) while a third took place at a biotech company
(Vernalis).
At the Pfizer research site in La Jolla, the preferred primary
screening method for fragments is the STD NMR technique,
although other research sites have employed other screening
methods. Prior to 2006, Pfizer had a legacy fragment collection of
5 K compounds, but this collection had two major drawbacks:
many of the fragments lacked chemistry handles to facilitate hit
follow-up efforts and almost all of the fragments were purchased
as screening compounds and therefore were of insufficient quantity for chemistry efforts. Therefore, it was decided to build a
fragment collection for the La Jolla NMR screening campaigns.
The goal was to create a collection optimized for NMR screening
while being chemically attractive for hit follow-up efforts.
The approach taken to achieve these goals was to first select
a set of novel reagents, then react these compounds with simple
reagents to cap the reactive functionalities. Virtual products for
the selected compounds were created and are then passed through
an in silico filtering process. Finally, the filtered libraries were synthesized via combinatorial libraries.
Selection of the reagents was based on the Pfizer internal
compound collection which allowed speedy acquisition of any
selected compound. A set of primary amines, secondary amines,
and carboxylic acids which were not commercially available were
chosen for consideration. These acids and amines were designed
by medicinal chemists via a Pfizer internal screening file enrichment effort to be novel and diverse, and more importantly, were
not part of any existing Pfizer fragment collection. The MW

Successful Fragment-Based Lead Discovery

225

cutoff was set at an upper limit of 200, with most of the amines
having a MW range of 100150. All compounds in consideration
had at least 5 g quantity available to ensure that future follow-up
activities were enabled.
The combinatorial reactions chosen for the novel amines were
amide bond formation and sulfonamide formation. The novel carboxylic acids were derivatized to simple amides. For the amine
reactions, we chose two simple carboxylic acids (propionic acid
and benzoic acid) and two simple sulfonyl chlorides (methylsulfonyl chloride and benzenesulfonyl chloride) as the capping
groups. Propyl amine and benzylamine were chosen as the capping groups to react with the novel carboxylic acids. Because only
one reactant will be variable, these combinatorial libraries were
essentially 1 N libraries, where the one reactant was a simple
reactant and the N component is the novel amines or acids.
Next, we used an in-house library design software (see details
in Chapter 15) to enumerate the virtual libraries and then calculated various physical properties. Products were removed from
consideration if MW is > 300, number of rotatable bonds > 3, and
ClogP > 3. For solubility, two in-house model calculations were
applied as filters: turbidimetric 10 mg/mL and thermodynamic
solubility >100 M. The resulting cherry-picked library was then
reviewed by NMR spectroscopists to remove compounds with
possible artifacts, likely to be insoluble, or likely to be false positive. These included some conjugated systems and compounds
with likelihood of indistinct NMR spectra.
Approximately 1,200 amines and 300 carboxylic acids were
selected for inclusion in the fragment libraries, from which
approximately 20 fragment libraries were synthesized. These
libraries yielded 2,000 products with sufficient purity (>95%)
and quantity (1.2 mL of 30 M solution), and the product
structures were confirmed via 1D NMR. This fragment collection became known as the NMR Combicores to denote their
purpose and their combichem origin. It was distributed across
several major Pfizer research sites and used in multiple fragment
screens.
One of the lessons learned from previous fragment screening collections is that fragments which are enabled for chemical
expansion was a key factor in engaging chemists in performing hit
optimization. This was addressed in the Combicores collection
described above via capped functional groups. However, Combicores is a specialized collection since it was designed specifically
for NMR screening. To build a more generic fragment collection to accommodate protein target screening requirements such
as reagent stability, sensitivity of screening methods, and druggability (19) of binding sites, the Pfizer Global Fragment Initiative (GFI) (20) was initiated with the goal of assembling a fragment collection suitable for several screening methods, including

226

Na and Hu

NMR techniques, SPR, high-concentration bioassays, and fragment crystallography. The assembly process for the collection
involved several computational filters, chemical complexity analysis, diversity analysis, and manual review by chemists to ensure
chemical attractiveness for follow-up. Details of the complete process will be published elsewhere (20). The analysis of the screening results presented early in the 2009 fall ACS meeting showed
that fragment screening of GFI offered consistent high hit rate
across protein classes (21).
Figure 11.2 shows a typical screening and hit-to-lead cascade
utilized in an FBLD campaign at Pfizer. In the first stage, we perform a primary screen (STD) along with a confirmation screen
(HCS, MS, or SPR). For a fragment to be considered a primary
NMR hit, the STD values must be >10 but less than 40, and MS
confirms that at least one copy of the fragment is bound with the
protein (binding = YES). In the biophysical confirmation step,
we conduct competitive binding studies with MS or a biochemical
assay to see if the compound displaces known active site binders in
order to confirm that the fragment is bound at the active site. For
this purpose we also attempt crystallography on the more active
hits when the protein target is amenable. The biochemical assay
results also allow us to calculate ligand efficiencies (LE) (22) of
the fragment hits. In the second stage, the hits with LE 0.3
and are chemically attractive are selected to be progressed for hit
follow-up activities. These activities include database mining for
similar analogs which are submitted for biochemical screening,
synthesizing of analogs by chemists, designing core-based fragments to enable further elaboration of the hits, and designing
structural-based targeted library based on top selected hits. This
is an interactive and iterative process involving a project team

IC50
1 mM

100 M

Initial Fragment Screen

NMR, MS or SPR confirmation

Biophysical
Confirmation

Competitive binding studies,

crystallization, labelled NMR
Select hits based on LE, activity,
or known binding conf.

10 M

1 M

< 100 nM

DB Mining, SBDD,
Library Dgns,
Analoging
Lead
Series
to Lead
Dev.

Dev. lead
Opt.
leadseries
serieswith
withSAR
SAR
Series hand-off to medchem

# Compounds decreases at
each stage.

Fig. 11.2. Typical FBLD screening cascade.

Successful Fragment-Based Lead Discovery

227

consisting chemists, spectroscopists, biologists, and computational chemists. Once lead series are identified with good activity
and SAR, they are passed to the lead optimization stage.
3.2. Vernalis NMR
Screening Collection

Vernalis Ltd., a UK-based biotech company, has a drug discovery

platform based on fragment screening using NMR techniques.
The Vernalis FBLD strategy is called SeeDs (Selection of Experimentally Exploitable Drug Startpoints) (18) and an integral part
of this strategy involves the creation of their fragment libraries.
The following section will describe this effort and how it was
applied to iteratively create four separate fragment libraries.
The compound collections used to select the desired fragments were the 2001 version of ACD (Available Chemicals Directory) which contains 267 K compounds, and a database of
1.6 M compounds from 23 chemical vendors. Removing duplicates from both databases yields 1.79 M compounds.
Since solubility is an important factor to consider when
selecting fragment for NMR screening, the solubility calculation
was done using a cross-validated PLS regression algorithm fitting 49 2D descriptors trained with a 3,041 molecules training set (23). This model was shown to be predictive within
1 log unit for a small test set, which is on par with experimental
error.
Hit follow-up consideration was accounted for via two filters
to remove undesirable functional groups and to include molecules
with a desirable functional group that would act as chemistry handles to enhance compound elaboration. These filters were derived
from extended discussion with medicinal chemists. A total of 12
undesirable functional group sets were created and collectively
used as a negative filter:
Four aliphatic carbons except if also contains XC
CCX, XCCX, XCX with X = O or N
Any atom different from H, C, N, O, F, Cl, S
SH, SS, OO, SCl, Nhalogen
Sugars
Conjugated system: R=CC=O, with R different from O,
N, or S or aromatic ring
(C=O)halogen, O(C=O)halogen, SO2 halogen,
N=C=O, N=C=S, NC(=S)N
Acyclic C(=O)S, acyclic C(=S)O, acyclic N=C=N
Anhydride, aziridine, epoxide, ortho ester, nitroso
Quaternary amines, methylene, isonitrile
Acetals, thioacetal, NCO acetals
Nitro group, >1 chlorine atom

228

Na and Hu

For the desired functionalities, a set of 11 functional groups

was used to filter against the compound databases, and all
molecules that did not contain at least one of these desired
groups were removed. These functional groups are: RCO2 Me,
RCO2 H, RNHMe/RNH(Me)2 , RNH2 , RCONHMe/R
CON(ME)2 , RCONH2 , RSO2 NHMe/RSO2 N(Me)2 , R
SO2 NH2 , ROMe, ROH, and RSMe. In addition, all compounds also must contain at least one ring system or be removed
from consideration. This filtering process was done using 2D
SMILES strings.
Molecular complexity is defined by the number of pharmacophoric triangles, where compounds with more triangles are
more deemed complex. This is done by first identifying the pharmacophoric elements contained within each molecule, and then
a triangle is identified by three features and the shortest bond
path between each pair of features (Fig. 11.3). There are eight
pharmacophoric features used: H-bond donor, H-bond acceptor, polar, hydrophobic, pi donor, pi acceptor, pi polar, and pi
hydrophobic. The Molecular Operating Environment (MOE)
(24) was used to calculate the pharmacophoric features as well as
the pharmacophoric triangles. Each compound is then assigned a
set of integers representing the pharmacophoric triangles it contains, which becomes its fingerprint. For a given collection of
compounds, these fingerprints can be used to identify which features are present and which ones are missing, and the collective
fingerprints becomes a measure of the diversity for the compound
collection.
The last filtering step involves experimental quality control,
which validates whether a given compound is soluble to 2 mM
in buffer solution, has a consistent NMR spectrum with its structure, has 95% or greater purity, and is both stable and soluble for
24 h in buffer solution. In addition, a water-LOGSY spectrum is
taken and compounds with positive results are considered to have

Fig. 11.3. Pharmacophoric triangle detection. The dotted lines define a triangle comprising three features: piHydrophobic (H=, centroid of benzene), piAcceptor (A=, oxygen of carboxylic acid), and piPolar (P=, oxygen of hydroxyl), and the shortest bond
path between each pair of features is 2 (A= to H=), 1 (P= to H=), and 4 (A= to P=).
Reprinted (adapted or in part) with permission from Journal of Chemical Information
and Modeling. Copyright 2008 American Chemical Society.

Successful Fragment-Based Lead Discovery

229

self-association, which can lead to false results in the NMR screen

and are thus removed from consideration.
Four fragment libraries were generated with different combination of the compound databases and filtering processes.
The first library (SeeDs-1) was designed from a relatively small
database of 87 K compounds comprising compounds available
from the Aldrich and Maybridge companies. These compounds
were first passed through a MW criterion (110 MW 250;
350 for sulfonamides) and then filtered to remove compounds
containing a metal, five continuous carbon methylene units, or a
reactive functional group. The resulting 7,545 compounds were
visually inspected by a medicinal chemist who selected a set mostly
based on chemistry follow-up attractiveness. Note that this visual
filtering process was captured and became the undesired and
desired functional groups filtering described above. Of the 1,078
fragments that passed visual filtering and ordered, 723 passed the
QC filtering. The experimental solubility results for these 723
compounds were then used as a further test set for the aqueous
solubility prediction model, which showed an 88% correlation for
predicting solubility for both the soluble (636 out of 723 correctly predicted to be soluble) and insoluble (84 out of 95 correctly predicted to be insoluble) compounds.
SeeDs-2 library was generated from their in-house database
called rCat of 1,622,763 unique chemical compounds assembled
from 23 suppliers (25). The filtering cascade began with MW
(same as SeeDs-1), then the functional groups and solubility filters which resulted in 43 K unique compounds (no overlap with
SeeDs-1). These were then clustered by 2D, 3-point pharmacophoric features to provide 3 K clusters, and the centroids of
each cluster was submitted for chemist review. Of the 395 selected
compounds that were ordered, 357 passed QC to become the
SeeDs-2 fragment library.
SeeDs-3 library was designed as a kinase-specific fragment
collection. The filtering process began with selecting compounds
which had the potential to bind to the ATP-binding site of protein
kinases. This was achieved via pharmacophore queries to match
the donoracceptordonor motif present in the ATP-binding
site, which was used as a first pass filter. The compounds matching
these queries were then filtered for MW and predicted solubility,
as well as the wanted and unwanted functional groups filtering. In
the end, only 204 compounds were selected for purchasing, and
174 passed QC filtering.
The final library was designed with the purpose of adding
incremental diversity to the first three fragment libraries. The
main filtering criterion was novel pharmacophoric triangles not
found in the first three libraries. After clustering and visual inspection from a panel of medicinal chemists, only 65 compounds were
purchased and 61 compounds passed QC.

230

Na and Hu

Combining all four SeeDs libraries resulted in 1,315 fragments for the collection. Various properties were calculated, analyzed, and compared with a drug-like reference set created from
the WDI and a binding reference set created from PDB. These
results can be found in the key reference (18).

4. Screening
Results
There are various methods to conduct a fragment screening
campaign. The most commonly utilized methods include various NMR techniques, mass spectrometry, SPR, and biochemical
screens. X-ray crystallization is a preferred method since it provides a binding conformation, but can only be used when the target protein is well behaved. Various calorimetry techniques have
also been used for fragment screening, but these have been less
commonly utilized. The merits of each method have been discussed in the literature (26) and will not be outlined here.
An analysis of screening hits based on 12 NMR screens (Table
11.1) for a range of protein targets conducted over an 8-year
period at Vernalis was performed (27). Three main aspects of the
analysis were (1) the relationship of the fragment hit rates to the
druggability of the target; (2) comparison of hits, nonhits to the
entire fragment library; and (3) the specificity and ligand complexity of the fragment hits.
Composition of the Vernalis fragment library evolved over
the course of 4 years through changes in what was synthesized
in-house, available commercially, and removed from the collection through quality control process. Although the number of
compounds remains roughly the same, the content has changed
dramatically, which makes the analysis quite challenging and interesting as well.
4.1. Fragment
Screening
Campaigns

As mentioned above, all data in the analysis are from fragment

screens using NMR spectroscopy to detect fragment binding
(28). Three NMR spectra (STD, water-LOGSY, and CPMG
(29)) were recorded separately for the ligand, ligand + protein,
and ligand + protein + known binding ligand (for competitive
binding). This approach can identify hits which bind in the same
site as the known binding ligand used in the screens. The resulting spectra are then inspected and a hit is defined as a fragment
which binds to and can be displaced by the known binder from
the protein. Based on the screening results from the three NMR
experiments, hits are classified in three categories: Class 1 hit is
defined as a fragment which shows evidence for binding in all

Successful Fragment-Based Lead Discovery

231

three NMR experiments, Class 2 hit shows changes in two experiments, and a Class 3 hit in only one experiment.
4.2. Fragment Hit
Rates and
Druggability Index

One of the interesting observations is that the experimentally

observed hit rate for screening fragments can be related to a computationally defined druggability index for the target, which provided an interesting side usage of fragment screening.
Due to the evolutionary nature of the Vernalis fragment collections and the fact that various screens were performed over
a period of several years, it is difficult and perhaps unreasonable
to directly compare the hit rates across multiple screens. However, it is still helpful to notice that reasonable hit rates (compared to HTS) are obtained across a diverse group of targets
(Table 11.1).
A very intriguing aspect of the analysis is assessing and ranking the target druggability based on the NMR screenings. This
approach was first reported by Abbott in 2005 as a strategy
to quickly evaluate protein druggability by screening chemical
libraries with 2D heteronuclear-NMR (30). They observed that
NMR hit rates were shown to be correlated with a number of
surface properties calculated from the binding site. Inspired by
the Abbott findings, Vernalis took a similar approach using the
druggability score (DScore) calculated by SiteMap (31) from
Schrodinger. What they found was that they were able to reach
a similar conclusion correlating fragment binding hit rate by 1D
NMR with protein druggability. Three aspects of the binding
pocket are considered as major contributions to DScore: pocket
size, degree of enclosure, and hydrophilicity. The results shown
in Fig. 11.4 indicated that if using a hit rate of 2% as a cutoff, all targets which yielded high hit rates (<2%) have a DScore
greater than 0.8, with the only outlier being HSP70. Their internal data revealed significant plasticity of the HSP70 ATP-binding
site upon binding to different ligands, which currently cannot be
captured by the SiteMap calculation. Excluding HSP70, SiteMap
DScore appears to be a good indicator for the NMR hit rate
one should expect for a target. As the authors pointed out, few
data points are available in the DScore range between 0.6 and
0.8, but they are optimistic this gap will be filled from future
screens so that a more complete evaluation of DScore can be
achieved.

4.3. Comparison of
Hits, Nonhits, and
the Entire Fragment
Library

The importance of physicochemical properties and structural

diversity to the assembly of a successful fragment collection
has been described in earlier sections. In this section, we will
focus on an interesting analysis done on the distribution of
the physicochemical properties and 3D pharmacophore triplet in
three groups of molecules: hits, nonhits, and the whole library.
Hits are defined as fragments which have been identified as

39

10

24

58

20

23

54

53

42

51

39

35

10

Class 1
seriesc

1,351

1,068

1,064

1,351

1,260

1,351

1,351

1,351

868

855

1,250

308

Library size

0.7

2.2

3.2

0.4

4.5

4.0

4.4

0.4

7.3

4.9

3.1

3.6

Low

High

High

Low

High

High

High

Low

High

High

High

High

Category

Class 1 hit rate

No

Yes

Yes

No

Yes

Yes

Yes

No

Yes

Yes

Yes

Yes

High-affinity
ligandsd

a Total number of fragments identified by at least one NMR experiment to interfere with the binding of known competitor compound.
b Number of fragments identified by all three NMR experiments (STD, water-LOGSY, and CPMG) to interfere with the binding of known competitor compound.
c Total number of unique chemical series suggested by the clustering results of Class 1 fragment hits with a Tanimoto coefficient of 0.70 and MACCS keys.
d Reported affinities <300 nM. Please refer to the paper for references.

PPI-3

34

40

52

PPI-1

13

PIN-1

PPI-2

119

PDPK1

55

60

82

101

38

HSP70

63

44

JNK3

81

40

11

Class 1b

HSP90

54

FAAH

109

15

Totala

DNA gyrase

CDK2

AK

Protein

Number of hits

Table 11.1
SeeDs screening hit rates for 12 protein targets. With kind permission from Springer Science+Business Media: Journal
of Computer-Aided Molecular Design, 23, 2009, 603, I-Jen Chen and Roderick E. Hubbard, Table 4

232
Na and Hu

Successful Fragment-Based Lead Discovery

233

Fig. 11.4. Targets with observed high (>2%, light bars) and low (<2%, darker bars)
Class 1 hit rates compared to the druggability score (Dscore) calculated by SiteMap.
The red arrow indicates the minimum Dscore for targets yielding high hit rates for
the current data set. With kind permission from Springer Science+Business Media:
Journal of Computer-Aided Molecular Design, 23, 2009, 603, I-Jen Chen and Roderick
E. Hubbard, Fig. 4.

competitive binders by at least one NMR experiment in any one

of the 12 screens. The nonhits are the compounds which are not
recognized as hits by any of the 12 screens. The sum of hits (29%)
and nonhits (71%) equal the entire library.
As seen in Fig. 11.5, among the five properties plotted, the
distribution of MW, NRot (number of rotatable bonds), and
number of pharmacophore triangles show no clear differences
among hits and nonhits, while SlogP which represents ligand
lipophilicity and number of rings showed clear separation between
the hits and nonhits. The more hydrophobic nature of the hits
in comparison to the nonhits is in good agreement with a general observation that binding is largely driven by hydrophobicity
(32). Figure 11.5E shows that there are more two-member rings
in the hits than in the nonhits.
4.4. Specificity and
Ligand Complexity of
the Fragment Hits

The screening results also clearly indicate that small fragments can
be specific binders, even for proteins within the same family.
Given their relatively small size, there are natural concerns
regarding nonspecific binding of fragments. Based on the Vernalis study, 62% of the fragments were competitive binders with
just one target and another 24% were hits for just two targets.
This study shows that most fragments are in actuality quite target
specific. The pie chart in Fig. 11.6 focuses on the hits from three
kinase screenings, PDPK1, CDK2, and JNK3. It shows that at
least 52% of the fragment hits are unique to one kinase, and only
11% of the hits are shared among all three of proteins.

234

Na and Hu

Fig. 11.5. Distribution plots of (a) molecular weight (MW), (b) number of rotatable bonds (NRot), (c) SlogP, (d) number
of pharmacophore (ph4) triangles, and (e) number of rings for the whole library (VER_ref), all hits (Class 13), Class 1
hits, and nonhits. With kind permission from Springer Science+Business Media: Journal of Computer-Aided Molecular
Design, 23, 2009, 603, I-Jen Chen and Roderick E. Hubbard, Fig. 6.

Ligand complexity can be represented by the number of pharmacophore triangles in fragment structures. Figure 11.7 plots
the averaged pharmacophore complexity of both the hits and the
Class 1 hits (all three NMR spectra confirm binding) for each target. It would appear that the level of complexity required for a
fragment to be detected in binding varies from target to target.
HSP70 appears to be the most demanding target as it requires the

Successful Fragment-Based Lead Discovery

235

Fig. 11.6. Overlap of kinase fragment hits. The horizontal lines indicate the portion of
unique fragment hits to each kinase. The crossed area (11%) is the portion of common
hits to all three kinases. With kind permission from Springer Science+Business Media:
Journal of Computer-Aided Molecular Design, 23, 2009, 603, I-Jen Chen and Roderick
E. Hubbard, Fig. 9.

Fig. 11.7. Pharmacophore complexity observed for all fragment hits and Class 1 hits for 12 protein targets. With kind
permission from Springer Science+Business Media: Journal of Computer-Aided Molecular Design, 23, 2009, 603, I-Jen
Chen and Roderick E. Hubbard, Fig. 8.

most complex fragments (20 and 27 triangles for all hits and Class
1 hits) among all targets studied. Perhaps HSP70s hit rate was
among the lowest because fewer fragments have the complexity
required for HSP70 binding. On the contrary, both DNA gyrase
and FAAH showed low average ligand complexity, and they are
indeed top two targets among the 12 screens with highest Class
1 hit rates, 4.9 and 7.3%, respectively.

236

Na and Hu

5. Overview
of Published
Fragment
Libraries

Over the past several years, descriptions of fragment collections

have been published in journals as well as book chapters. In
two recent reviews, a chapter from Evotec (33) compared several collections based their origins, while a journal article from
Leiden University (26) summarized fragment collections based
on the intended screening methods. We would like to present an
overview by blending information from both reviews together to
provide a more complete and updated picture (see Table 11.2 for
details).

5.1. Origins and

Library Size

Table 11.2 is divided into three sections: the first section is a

group of large or mid-sized pharma/biotech; the second section
contains all small biotechs specializing in FBDD; and the last section contains the companies offering fragment screening as a service. This table demonstrates that FBDD has been adopted as a
drug discovery platform throughout the drug discovery industry regardless of company size or screening technology, and also
illustrates the fact that FBDD has now become a CRO service.
It is also interesting to see that within each group, the fragment
collection sizes vary from several hundreds to tens of thousands
of compounds. While there does not seem to be a consensus for
the collection size, most fragment libraries are assembled based
on some fundamental guiding principles which include sampling
of chemical space and alignment with a chosen screening technology. It is worthwhile to mention that both AstraZeneca and
Evotec have a relatively large fragment collection (>20 K), but
both companies stated they intend to screen only subsets of their
collection based on the nature of the intended target and some
practical considerations, such as the cost of the reagent.

5.2. Physical
Properties

Most of the fragment libraries are designed with physical properties within the rule-of-3 constraints (14), which were originally
proposed for molecules used for high-throughput fragment crystallography. For solubility, since there are no reliable methods to
predict aqueous solubility, clogP is often used as a guide. In the
case of AstraZeneca, a clogP value above 2.2 warrants enough risk
for a neutral compound that its solubility will be experimentally
determined (8).

5.3. Screening
Technologies

In recent years there have been encouraging advances in screening technologies suitable for fragment screening. Even for an
established method such as NMR, new techniques have been
developed and put into practice at various companies that
have expanded the scope of targets for FBDD. For example,

0.3
3

194

190

20,342

132

1,400

AstraZeneca

Vertex

Vernalis

1,000

7,000

10,000

20,063

21,869

2,000

SGXc

Sunesis

Graffinity

Evotec

Evotec

ZoBio/Pyxis
Discovery

1.2

1.6

1.3

2.7

1.9

1 to 3

NA

NA

NA

NA

NA

NA

2.3

NA

NA

NA

NA

Number of
rotatable
bonds

NA

NA

3.2

NA

NA

NA

NA

NA

H-bond
acceptors

a All single values in this table for properties are mean, except the values for the Pfizer collection which are median.
b Multiple means more than one screening technology, including NMR, SPR, biochemical assay, and X-ray.
c The property values reported for SGX applies to 90% of the molecules for the SGX collection.

218

247

2.2

300

276

NA

NA

NA

2.2 to 5.5
1.9

NA

NA

NA

1.6

1.5

<200

174

300

127 to 350

850

20,000

Astex

Plexxikon

NA

NA

1,200

AstraZeneca

1.0
3

205

300

2,792

2,000

1.5

Roche

220

10,000

Abbott

ClogP

Pfizera

MW

Size

Company

H-bond
donors

52.6

70

NA

NA

NA

NA

NA

60

NA

NA

NA

NA

60

56.9

NA

Polar
surface
area ()

NMR (target
immobilized)

Biochemical
assay

NMR

SPR (ligand
immobilized)

Tether

X-ray

X-ray

X-ray

NMR

(26)

(33)

(33)

(26, 33)

(37)

(36)

(35)

(34)

(18, 27)

(11, 12)

(8)
NMR

(8)
Multipleb

(26, 33)

(20, 21)

(26, 33)

References

NMR

SPR (target
immobilized)

Multipleb

SAR by NMR

Screening
technologies

Table 11.2
Overview of some key physical properties for selected fragment libraries and their associated screening methods

Successful Fragment-Based Lead Discovery

237

238

Na and Hu

membrane-bound protein can be made suitable for fragment

screening using the target-immobilized NMR screening (TINs)
(26). For an excellent comparison of all fragment screening technologies, please refer to the review from Siegal et al. (26).

6. Conclusion
The composition of a fragment collection can have a profound
effect on the success of an FBLD campaign. Consideration of the
screening method and ease of chemistry follow-up are two of the
more important factors in creating a fragment collection. It has
been shown that by using a combination of computational analysis and human expertise, a fragment collection can be created
to accommodate a single method or several screening methods
without being target or protein family specific. Further, a carefully
designed fragment collection can result in high hit rates across a
variety of targets to produce hits with novelty and good ligand
efficiency, thereby accelerating the lead discovery process.

Acknowledgments
We would like to thank Drs. Ben Burke and Zhongxiang (Joe)
Zhou for their valuable comments and insights throughout the
preparation of this manuscript.
References
1. Congreve, M., Chessari, C., Tisi, D.,
Woodhead, A. (2008) Recent advances in
fragment-based drug discovery. J Med Chem
51, 36613680.
2. Hesterkamp, T., Whittaker, M. (2008)
Fragment-based activity space: smaller is better. Curr Opin Chem Biol 12, 260268.
3. Hajduk, P. J., Greer, J. (2007) A decade
of fragment-based drug design: strategic
advances and lessons learned. Nat Rev Drug
Discov 6, 211219.
4. Albert, J. S., Blomberg, N., Breeze, A. L.,
Brown, A. J. H., Burrows, J. N., Edwards,
P. D., Folmer, R. H. A., Geschwindner, S.,
Griffen, E. J., Kenny, P. W., Nowak, T., Olsson, L. -L., Sanganee, H., Shapiro, A. B.

(2007) An integrated approach to fragmentbased lead generation philosophy, strategy

and case studies from AstraZenecas drug discovery programmes. Curr Top Med Chem 7,
16001629.
5. Shuker, S. B., Hajduk, P. J., Meadows, R. P.,
Fesik, S. W. (1996) Discovering high-affinity
ligands for proteins: SAR by NMR. Science
274, 15311534.
6. de Kloe, G. E., Bailey, D., Leurs, R., de
Esch, I. J. P. (2009) Transforming fragments
into candidates: small becomes big in medicinal chemistry. Drug Discovery Today 14,
630646.
7. Hann, M. M., Oprea, T. I. (2004) Pursuing the lead-likeness concept in pharma-

Successful Fragment-Based Lead Discovery

8.

9.

10.

11.

12.
13.

14.

15.

16.

17.

18.

19.
20.

ceutical research. Curr Opin Chem Biol, 8,

255263
Blomberg, N., Cosgrove, D. A., Kenny,
P. W., Kolmodin, K. (2009) Design of compound libraries for fragment screening. J
Comput Aided Mol Des 23, 513525.
Barker, J., Courtney, S., Hesterkamp, T., Ullmann, D., Whittaker, M. (2005) Fragment
screening by biochemical assay. Exp Opin
Drug Discov 1, 225236.
Schuffenhauer, A., Ruedisser, S., Marzinzik,
A., Jahnke, W., Selzer, P., Jacoby, E. (2005)
Library design for fragment based screening.
Curr Top Med Chem 5, 751762.
Fejzo, J., Lepre, C. A., Peng, J. W., Bemis,
G. W., Ajay, Murcko, M. A., Moore, J. M.
(1999) The SHAPES strategy: an NMRbased approach for lead generation in drug
discovery. Chem Biol 6, 755769.
Lepre, C. A. (2001) Library design for
NMR-based screening. Drug Discov Today 6,
133140.
Taskinen, J. Norinder, U. (2006) In silico predictions of solubility, Comprehen Med
Chem II, edited by Taylor, J. B., Triggle, D.
J. 5, 627648.
Congreve, M., Carr, R., Murray, C., Jhoti,
H. (2003) A rule of three for fragmentbased lead discovery. Drug Discov Today 8,
876877.
Mayer, M., Meyer, B. (1999) Characterization of ligand binding by saturation transfer
difference NMR spectroscopy. Angew Chem
Int Ed 38, 17841788.
Wang, Y., Liu, D., Wyss, D. F. (2004) Competition STD NMR for the detection of highaffinity ligands and NMR-based screening.
Magn Reson Chem 42, 485489.
Dalvit, C., Pevarello, P., Tato, M., Vulpetti,
A., Sundstrom, M. (2000) Identification of
compounds with binding affinity to proteins
via magnetization transfer from bulk water. J
Biomol NMR 18, 6568.
Baurin, N., Aboul-Ela, F., Barril, X., Davis,
B., Drysdale, M., Dymock, B., Finch, H.,
Fromont, C., Richardson, C., Simmonite,
H., Hubbard, R. E. (2004) Design and characterization of libraries of molecular fragments for use in NMR screening against
protein targets. J Chem Inf Comput Sci 44,
21572166.
Hajduk, P. J., Huth, J., Tse, C. (2005)
Predicting protein druggability. Drug Discov
Today 10, 16751682.
Lau, W. F., Hepworth, D., Magee, T. V.,
Du, J., Bakken, G. A., Miller, M. D., Hendsch, Z. S., Thanabal, V., Kolodziej, S. A.,
Xing, L., Hu, Q., Narasimhan, L. S., Love,
R., Charlton, M. E., Hughes, S., Van Hoorn,

21.

22.
23.

24.
25.

26.
27.

28.

29.
30.
31.
32.

33.

239

W., Mills, J. E., Withka, J. M. (2010) Design

of a multi-purpose fragment screening library
using molecular complexity and orthogonal
diversity metrics. J Comput-Aided Mol Des.
Manuscript in preparation.
Hu, Q., Yan, J., Withka, J. M., Sahasrabudhe,
P., Moore, C., Na, J., Narasimhan, L. S.
(2009) Computational analysis on NMR
screenings of the Pfizer Fragment Initiative collection. 238th ACS National Meeting,
Washington, DC, United States.
Hopkins, A. L., Groom, C. R., Alex, A.
(2004) A useful metric for lead selection.
Drug Disc Today 9, 430431.
Huuskonen, J., Rantanen, J., Livingstone, D.
(2000) Prediction of aqueous solubility for
a diverse set of organic compounds based
on atom-type electrotopological state indices.
Eur J Med Chem 35, 10811088.
Chemical Computing Group Inc., Montreal,
H3A 2R7 Canada.
Baurin, N., Baker, R., Richardson, C.,
Chen, I., Foloppe, N., Potter, A., Jordan,
A., Roughley, S., Parratt, M., Greaney, P.,
Morley, D., Hubbard, R. E. (2004) Druglike annotation and duplicate analysis of a 23supplier chemical database totalling 2.7 million compounds. J Chem Inf Comput Sci 44,
643651.
Siegal, G., Ab, E., Schultz, J. (2007) Integration of fragment screening and library design.
Drug Discov Today 12, 10321039
Chen, I., Hubbard, R. E. (2009) Lessons for
fragment library design: analysis of output
from multiple screening campaigns. J Comput Aided Mol Des 23, 603620.
Hubbard, R. E., Davis, B., Chen, I., Drysdale, M. (2007) The SeeDs approach: integrating fragments into drug discovery. Curr
Top Med Chem 7, 15681581.
Meiboom, S., Gill, D. (1958) Modified spinecho method for measuring nuclear relaxation times. Rev Sci Instrum 29, 688691.
Hajduk, P. J., Huth, J. R., Fesik, S. (2005)
Druggability indices for protein targets. J
Med Chem 48, 25182525.
Halgren, T. A. (2009) Identifying and characterizing binding sites and assessing druggability. J Chem Inf Model 49, 377389.
Ruppert, J., Welch, W., Jain, A. N. (1997)
Automatic identification and representation
of protein binding sites for molecular docking. Prot Sci 6, 524533.
Brewer, M., Ichihara, O., Kirchhoff, C.,
Schade, M., Whittaker, M. (2008) Assembling a fragment library. Fragment-Based
Drug Discovery: A Practical Approach, in
(Zartler, E., Shapiro, M. J. eds.), pp.
3962.

240

Na and Hu

34. Hartshorn, M. J., Murray, C. W., Cleasby,

A., Frederickson, M., Tickle, I. J., Jhoti,
H. (2005) Fragment-based lead discovery
using X-ray crystallography. J Med Chem 48,
403413.
35. Card, G.L., Blasdel, L., England, B. P.,
Zhang, C., Suzuki, Y., Gillette, S., Fong,
D., Ibrahim, P. N., Artis, D. R., Bollag, G.,
Milburn, M. V., Kim, S., Schlessinger, J.,
Zhang, K. Y. J. (2005) A family of phosphodiesterase inhibitors discovered by cocrystallography and scaffold-based drug design.
Nat Biotechnol 23, 201207.

36. Blaney, J., Nienaber, V., Burley, S. K. (2006)

Fragment-based lead discovery and optimization using X-ray crystallography, computational chemistry, and highthroughput organic synthesis, Fragment-Based
Approaches in Drug Discovery, in (Jahnke,
W., Erlanson, D. A., Mannhold, R., Kubinyi,
H., and Folkers, G., eds.), pp. 215248.
37. Erlanson, D.A., Ballinger, M. D., Wells,
J. A. (2006) Tethering, Fragment-Based
Approaches in Drug Discovery, in (Jahnke, W.,
Erlanson, D. A., Mannhold, R., Kubinyi, H.,
and Folkers, G., eds.), pp. 285312.

Chapter 12
Fragment-Based Drug Design
Eric Feyfant, Jason B. Cross, Kevin Paris, and Dsire H.H. Tsao
Abstract
Fragment-based drug design (FBDD), which is comprised of both fragment screening and the use of
fragment hits to design leads, began more than 15 years ago and has been steadily gaining in popularity
and utility. Its origin lies on the fact that the coverage of chemical space and the binding efficiency of
hits are directly related to the size of the compounds screened. Nevertheless, FBDD still faces challenges,
among them developing fragment screening libraries that ensure optimal coverage of chemical space,
physical properties and chemical tractability. Fragment screening also requires sensitive assays, often biophysical in nature, to detect weak binders. In this chapter we will introduce the technologies used to
address these challenges and outline the experimental advantages that make FBDD one of the most
popular new hit-to-lead process.
Key words: Fragment-based drug design, fragment screening, ligand efficiency, NMR, X-ray
crystallography.

1. Introduction
1.1. General Views

In recent decades, high-throughput screening (HTS) has become

the most established method in the pharmaceutical industry for
identifying potential lead compounds. Despite extensive effort
in designing better and larger libraries for screening, the attrition rate of compounds entering clinical trials has continued to
increase. The industry has attempted to address this issue by
focusing on the improvement of compound properties, using
schemes such as the Lipinski Rule of 5 for oral drugs (1). A
recent study (2) has shown that the decisive factor in designing successful clinical candidates is the quality of the initial HTS
hit. Since the primary criterion for HTS hit selection has been
potency, often at the expense of other important physicochemical

J.Z. Zhou (ed.), Chemical Library Design, Methods in Molecular Biology 685,
DOI 10.1007/978-1-60761-931-4_12, Springer Science+Business Media, LLC 2011

241

242

Feyfant et al.

properties, these initial leads are less likely to evolve to successful

candidates.
To improve the lead identification process, the pharmaceutical industry has invested in new approaches. Fragment-based drug
design (FBDD) is a relatively new technology that has shown
remarkable potential in a short period of time. The foundation
of FBDD was described by Jencks (3) and supported by Nakamura and Abeles (4), who showed that drug-like molecules can be
regarded as the combination of several binding epitopes or fragments. FBDD presents two main advantages over HTS screening
methods. The first is that fragment libraries can cover more of
chemical space than HTS screening libraries. Even very large HTS
screening decks, with over a million compounds, can explore only
a vanishingly small fraction of the estimated 1060 compounds (5)
with up to 30 heavy atoms. Since the number of possible compounds increases exponentially with molecular size, a library of
only a few thousand fragments with fewer than 17 heavy atoms
is capable of covering a larger fraction of chemical space. The
second advantage lies with the concept of ligand efficiency, i.e.,
the average contribution of each atom of the molecule to the
binding affinity. This concept was introduced by Kuntz et al. in
1999 (6) and later proposed as a criterion for hit selection by
Hopkins (7). Interestingly, this model suggests that the level of
energy contribution of each ligand atoms to the binding energy is
inversely proportional to the molecular weight. Even for fragment
binders whose affinity ranges from high micromolar to millimolar,
whereas HTS hits range from high nanomolar to low micromolar, fragments are more efficient binders. On the other hand, weak
binders are more difficult to detect by common high-throughput
techniques like displacement assays. This chapter will illuminate
the principles used to design fragment libraries for drug discovery and also describe two different screening methods, NMR and
X-ray crystallography.
1.2. Designing
Fragment Screening
Libraries

Fragment libraries differ from drug-like and lead-like libraries primarily by having members with a significantly lower molecular
weight (MW), typically in the 140300 Da range. However, as
fragment screening programs have matured over recent years,
other key factors that help improve the success rates of these
projects have been identified.
There has been much effort in identifying physical properties of ideal fragment libraries. Since fragments tend to be smaller
than most drugs, clinical candidates, leads, or high-throughput
screening (HTS) hits, they are able to make fewer interactions
on average and tend to have lower affinities for their protein
targets. Affinities are often in the high micromolar or low millimolar range, necessitating solubility at least to that degree and
potentially higher depending on the assay protocol. Congreve and

Fragment-Based Drug Design

243

coworkers (8) introduced the Rule of 3, which showed that

diverse hits from fragment libraries tend to have the following
physical properties: MW < 300 Da, clogP 3, H-donors 3,
H-acceptors 3, polar surface area (PSA) 60 2 , and rotatable
bonds 3. These properties not only partially address fragment
solubility, but also help to ensure that compounds resulting from
the elaboration of fragment hits have a higher likelihood obeying
the Rule of 5. (1)
In addition to physical properties, the chemical structure of
fragments can play a role in their success as screening hits. Hajduk
and coworkers showed that certain privileged scaffolds tend to
show up repeatedly in successful fragment screening campaigns
(9). Bemis and Murcko (10) analyzed known drugs in an effort
to identify common features and scaffolds, which could be used
to bias fragment libraries toward drug-like structures. An optimal
molecular complexity was also discussed by Hann and coworkers
(11), which would ensure that fragments have sufficient chemical
features to keep from being overly promiscuous while at the same
time not making them overly specific by introducing too many
features. Having a good balance of chemical features also ensures
that fragment hits will have a sufficient number of chemical handles to allow for synthetic chemistry follow-up and elaboration.
Schuffenhauer and coworkers (12) took this idea one step further
by suggesting that fragments should have chemical features that
mask reactive functional groups, thereby simplifying synthesis of
analogs. Having a diverse set of chemical structures in a fragment
library is not only ideal for improving the odds of finding interesting hits that bind to the target, but can also assist in deconvolution of fragment structures if they are screened as mixtures.
The Novartis approach for synthetic accessibility is very attractive
and can be managed successfully for a smaller library. At Wyeth,
we created a larger library of 10,000 compounds, since surface
plasmon resonance would be the primary screening technique,
which is able to support a higher throughput than NMR or Xray crystallography. Due to the elaborate work in synthesizing a
parent library and a screening library of such size, a more practical but less exact approach was taken, by choosing fragments
from our corporate library and predicting synthetic accessibility
as a function of number and diversity of substituent of the fragment core. The fragment core is always a ring system and was
considered synthetically tractable if at least two distinct analogs
existed in our compound catalogs (internal and external chemical
collections).
Although many groups that use fragment screening develop
their own internal libraries, many commercial vendors now offer
fragment screening collections that are Rule of 3 compliant
and optimized for chemical diversity (13). Some of these libraries
are even targeted at specific screening methodologies, such as

244

Feyfant et al.

brominated fragment libraries for use in X-ray crystallographic

screening and fluorinated fragment libraries for use in NMR
screening.
1.3. Screening
Methods

The main challenge for any fragment screening method is the

detection of weak binders. The methods commonly used to detect
fragments can be broadly broken down into biophysical and functional assays.
Several biophysical methods are commonly employed in
fragment-based screening, including NMR, X-ray crystallography,
mass spectrometry (MS), and surface plasmon resonance (SPR).
NMR is a popular screening method since it can detect weak
binders and is also a flexible technique. Large fragment libraries
can be screened using ligand-observe NMR experiments, such as
saturation transfer difference (14) (STD) and WaterLOGSY (15).
Structural information detailing the fragments binding mode can
be obtained through protein-observe experiments such as heteronuclear single quantum coherence (HSQC) (42). X-ray crystallography is used as a primary screening tool less often than
NMR because it often requires higher protein quantity, has a
slower throughput, and needs robust crystallization conditions.
Crystallography is commonly used for the follow-up of fragments
identified using other methods (16), since the structural information gained can be used to direct fragment elaboration. MS can
also be used as a detection method for fragment binding (17), as
it forms the basis of the tethering strategy used by Sunesis (18).
In this case, a native or mutated cysteine residue adjacent to the
binding site in the protein target is exposed to a library of disulfide fragments. A covalent bond between a fragment and the protein will then be formed if the fragment has sufficient affinity for
the binding site. SPR is gaining popularity not only as a primary
screening tool but also as an orthogonal confirmation for fragments identified by other assays (19). This method requires that
either the protein or the fragment is immobilized onto the chip.
Immobilization of the protein provides an excellent confirmation
of fragment binding in cases where the protein is poorly behaved
in the primary screening assay (i.e., aggregation). Immobilization
of the fragment library leads to high-throughput fragment screening with a robust signal, but the protein must be soluble in the
assay conditions.
High-concentration screening (HCS) assays tend to be used
less frequently than biophysical methods, but can be employed to
good effect due to their high-throughput capacity. These assays
require biochemical function of the protein, up front knowledge
of biochemical activity, and it can be difficult to reliably detect
weak binders in the millimolar range. Fluorescence correlation

Fragment-Based Drug Design

245

techniques have also been employed successfully (20) and have

the advantage that protein modulation (i.e., stabilization) can
be detected, but this method requires large quantities of soluble
protein.
Computational chemistry techniques can also be employed in
fragment screening to great effect. Virtual screening using methods such as pharmacophore matching and molecular docking or
determination of hot spots within the protein active site using
computational tools such as HSITE (21), HIPPO (22), or MCSS
(23, 24) can be used to bias fragment libraries for a particular
protein target, as long as information detailing the protein structure or chemical structures of several ligands are known. Some
fragment identification methods are based on the use of molecular dynamics, in particular grand canonical ensemble calculations
(25). All computational methods require the use of an experimental screening assay to validate binding.
In this overview of fragment-based drug design, we will
expand on the application of NMR and X-ray crystallography
as the tools for biophysical screening. Common protocols will
be presented, as well as the strengths and limitations of each
approach.
1.4. From Fragment
to Lead

Although fragments can be considered efficient binders, given

their size, their binding potencies are still order(s) of magnitude
too weak for them to be considered true leads. There are two
ways of optimizing these ligands: either linking them when two
fragments are bound in distinct, but not too distant, parts of
the binding pocket or growing them using combinatorial chemistry with (or without) the support of computational methods. A
thorough review of the use of computational approaches to the
fragment-based de novo design problem can be found in reference
(20).
In the growing approach an initial fragment is grown in an
attempt to add interactions between the receptor and the ligand.
There are numerous examples of software applying the growing approach, e.g., SPROUT (26, 27), LEGEND (28), LUDI
(29), GROWMOL (30), SkelGen (31, 32), SMoG (33), and LigBuilder (34). In the linking approach, fragment building blocks
are positioned in the target binding site as described above and
then computationally connected to each other by linkers to yield
a complete molecule that satisfies all of the key interactions. The
tacit assumption here is that the binding affinity of fragments is
additive, the loss in rigid body entropy on binding of all components of the molecule is small (35), and, moreover, that the
affinity contribution from the linker is negligible or favorable.
CONFIRM (36), LUDI (29), HOOK (37), PRO_LIGAND

246

Feyfant et al.

(38), CAVEAT (39, 40) are examples of software using the

linking approach. ReCore (41) and MOE Scaffold Replacement are capable of performing both fragment linking and
building.

2. Materials
2.1. STD NMR
Screening

1. Fragment stock solutions in DMSO, usually 80 mM or

higher. Fragments should have good solubility in biological
buffer, at least 200 M.
2. Biological buffer where protein remains stable for a few
hours. Typically neutral buffer such as HEPES, Tris with salt
(10200 mM). Protein target needs to be stable in the presence of DMSO (110%), usually 5%.
3. NMR spectrometer operating at 500 MHz or higher in frequency, use of cryoprobe preferable for higher sensitivity.
4. 5, 3, or 1.7 mm id NMR tubes (Wilmad).
5. High purity (>95%) protein target, at concentrations
between 0.2 and 10 M.
6. Known inhibitor to check binding specificity (ex: ATP or
staurosporine for kinases).

2.2. X-Ray Screening

1. Protein crystals suitable for soaking. Preferable if known

inhibitor(s) have already verified the suitability of the crystals
for soaking. If co-crystallization with the fragments is to be
used instead of soaking, sufficient protein to grid around a
robust crystallization condition for all fragments of interest
is required.
2. Fragment stock solutions in DMSO, usually 100 mM or
higher (see Note 7).
3. (Optional) Liquid handling robot to dispense the solutions
for co-crystallization.
4. (Optional) Crystallization robot to dispense protein and
fragments for co-crystallization. Ability to do small volume
drops is very desirable.
5. Trays and cover slips appropriate to the technique used.
6. An X-ray generator with high flux (home source) or access
to a synchrotron beam line.
7. (Optional) Robotic sample changer to facilitate exchange of
samples for diffraction analysis (see Note 8).
8. Computers and software to analyze the X-ray diffraction
data.

Fragment-Based Drug Design

247

3. Methods
3.1. STD NMR
Screening

1. Check protein activity and integrity in the working buffer

and the final DMSO amounts (110% v/v), at 0.210 M
concentration.
2. Prepare NMR samples of the fragments at the desired concentration for solubility, integrity, and reference spectrum,
usually at 200500 M.
3. Mixtures of six to eight fragments are prepared at 200
500 M each, with the final % DMSO calculated to be 5%.
Protein for a final working concentration 5 M and an
internal inert standard, such as TSP, are added. Final volume
will be 500 L (if working with a 5 mm NMR tube), 180
L (3 mm NMR tube), or 30 L (1.7 mm tube).
4. Sample is loaded in the NMR spectrometer, parameters are
optimized for the STD experiment, and data are acquired.
5. If binding is observed with positive STD signal, the binder is
identified by comparison of the STD signal and the reference
data from step 2.
6. Confirmation of binding is performed by preparing the fragment binder with the protein and the STD acquired again.
7. Add known competitor and acquire the STD experiment
again, to confirm the fragment binds in the site of interest. If there is good competition, the STD signal for the
fragment should decrease with the addition of a tight
inhibitor.
8. Monitor target integrity in the NMR sample by comparing
the protein background signal with time.

3.2. X-Ray Screening

1. The fastest way to obtain co-structures with a protein and

fragments is to soak the fragments into existing crystals.
Since each protein is unique, trial and error will be necessary to deduce the conditions where your protein crystals
are stable and the fragments are suitably soluble (referred to
as the protein stabilization buffer) (see Note 9).
2. The organic solvent of choice is DMSO. Most fragments
are soluble in DMSO and DMSO also has cryo-preservation
qualities that assist in vitrifying crystals.
3. A suitable starting concentration for DMSO in a soaking
experiment is 5%. Combine 9.5 L of protein stabilization
buffer with 0.5 L fragment solution on a cover slip and
mix thoroughly. Transfer protein crystal(s) to this solution
and invert over a crystallization well (see Note 10).

248

Feyfant et al.

4. If co-crystallization is indicated by properties of the protein or the fragment library, then prepare a solution of the
protein with suitable concentration of fragment(s), (suggested 100 mM). Screen this solution around known crystallization conditions for the protein. If no crystals are
observed a full screen using numerous conditions may be
indicated.
5. Prepare a cryopreservation solution compatible with your
crystals, starting with the protein stabilization buffer. Upon
testing, the amount of cryo agent (DMSO, low molecular weight PEG, glycerol, ethylene glycol, etc.) used should
produce a clear glass effect with no water rings when analyzed with X-rays.
6. Treat the crystal exposed to fragment(s) with the cryopreservation buffer and vitrify the sample with liquid nitrogen (see
Note 11).
7. During data collection from crystals exposed to fragment(s),
collect a data set that is complete in the low-resolution shells
and has high redundancy. Also beneficial will be the highest resolution data possible, so examination of multiple crystals to select the one with suitable qualities is crucial (see
Note 12).

4. Notes
4.1. NMR Screening

1. The NMR screening samples can be prepared in an automated fashion with a programmable platform such as Tecan
(by Tecan) and samples can be automatically loaded into the
spectrometer by using a Sample Rail (by Bruker Biospin).
This allows for maximum spectrometer time and the sample
is always freshly prepared prior to data acquisition.
2. The protein stock concentration should be in the NMR running buffer at concentrations slightly higher than what is
used in the NMR samples. Alternatively, if the protein stability is better in a different buffer, the protein could be stored
at high concentration (80 M or higher) and a small aliquot
diluted into the NMR running buffer for sample preparation.
3. Fragments in mixtures can sometimes precipitate due to
the high total fragment concentration in solution, which
could be up to 5 mM. In most cases where precipitation
is observed, we have noted that the other fragments in the
mixture are still soluble and give good NMR signals. Thus
the mixture is still usable.

Fragment-Based Drug Design

249

4. The higher the DMSO percentage used, the higher the fragment mixture solubility will be. The protein needs to be
stable and active for at least a couple of hours under these
conditions for data collection.
5. Competition experiments can be performed within the same
NMR sample mixture used for screening if protein amounts
are limited. The competitor is just added to the NMR solution in the tube and mixed well.
6. Fragments in the mixtures that bind to the protein target can easily be identified by comparing the NMR
spectrum of the hit with the spectra of the individual
fragments.
4.2. X-Ray Screening

7. The more concentrated the fragment sample the less a dilution effect is observed when added to the protein. For
those proteins or crystals highly sensitive to DMSO concentration we have found that soaking is problematic and
co-crystallization is indicated. Fragments are generally lowaffinity compounds and in order for weakly binding compounds to be observed with X-ray crystallography they
need to possess excellent solubility. A rule of thumb used is
ten times the binding constant. Applying this to fragment
screening, it is desirable to have the compound at 100 mM
during the experiments. While this level of solubility is easily obtained in DMSO, the addition of an aqueous component will be an issue as precipitation of the small molecule
often occurs. During soaking experiments it is not uncommon to have a successful experiment despite heavy precipitate or even crystallization of the small molecule upon addition of the fragment to the protein stabilization solution.
When co-crystallizing protein with fragments, centrifugation prior to screening will be required in cases where precipitation is observed.
8. For automation in our lab, we use a Hamilton STAR
for creating/dispensing crystallization solutions, a TTP
LabTech mosquito liquid handling robot for setting
up crystallization drops, a Formulatrix robotic storage/retrieval/imaging system for crystallization trays, and
a Rigaku ACTOR robot for automatic crystal handling for
testing of diffraction properties.
9. Prior to initiating the FBS soaking experiments a substantial amount of investigation needs to be completed on the
methodology that will be used. The parameters that should
be considered and optimized include
The length of time to soak the compound into the
crystal

250

Feyfant et al.

The amount of DMSO (or other organic liquids) needed

to maintain compound solubility as well as crystal
integrity
The protocol necessary to freeze the crystal for data collection
The default values for the data collection software used
The automation of structure solution
10. The Hampton Research VDXm Plate with sealant (Part
HR3-306) is recommended. A 10 L drop inverted on this
tray will last for at least 7 days at 18 C with no additional
solution added to the well. For longer soaks or for security add 100 L of the protein stabilization buffer to the
well prior to inverting the cover slip containing the soaking experiment. The length of time to soak a crystal with
compound is one of the, if not the most, critical steps. Too
short and you may not find binding. Too long and you risk
damaging the crystals. In addition, each protein and each
crystal form for that protein are different. The successfully
designed experiment will allow sufficient time for binding
to occur as well as any remodeling of the protein required
to accommodate the fragment(s).
11. To prevent backsoaking of the fragment(s) from the crystals swipe the crystal quickly through the cryopreservation
buffer. If longer soaks in the cryopreservation buffer are
required add fragment(s) to it so that equilibrium does not
remove the weak binding fragments from the crystal.
12. It is not uncommon for data at low resolution to be inconclusive. Data collected from similarly treated crystals can
often show inhibitor when the resolution is extended where
it was not visible at low resolution.
References
1. Lipinski, C. A., Lombardo, F., Dominy, B.
W., Feeney, P. J. (2001) Experimental and
computational approaches to estimate solubility and permeability in drug discovery and
development settings. Adv Drug Deliv Rev
46, 326.
2. Keseru, G. M., Makara, G. M. (2009) The
influence of lead discovery strategies on the
properties of drug candidates. Nat Rev Drug
Discov 8, 203212.
3. Jencks, W. P. (1981) On the attribution
and additivity of binding energies. Proc Natl
Acad Sci U S A 78, 40464050.
4. Nakamura, C. E., Abeles, R. H. (1985)
Mode of interaction of beta-hydroxybeta-methylglutaryl coenzyme A reductase
with strong binding inhibitors: compactin

5.

6.

7.
8.

and related compounds. Biochemistry 24,

13641376.
Bohacek, R. S., McMartin, C., Guida, W.
C. (1996) The art and practice of structurebased drug design: a molecular modeling perspective. Med Res Rev 16, 350.
Kuntz, I. D., Chen, K., Sharp, K. A., Kollman, P. A. (1999) The maximal affinity of
ligands. Proc Natl Acad Sci U S A 96,
999710002.
Hopkins, A. L., Groom, C. R. (2002) The
druggable genome. Nat Rev Drug Discov 1,
727730.
Congreve, M., Carr, R., Murray, C., Jhoti,
H. (2003) A rule of three for fragmentbased lead discovery? Drug Discov Today 8,
876877.

Fragment-Based Drug Design

9. Hajduk, P. J. (2006) Fragment-based drug
design: how big is too big? J Med Chem 49,
69726976.
10. Bemis, G. W., Murcko, M. A. (1996) The
properties of known drugs. 1. Molecular
frameworks. J Med Chem 39, 28872893.
11. Hann, M. M., Leach, A. R., Harper, G.
(2001) Molecular complexity and its impact
on the probability of finding leads for
drug discovery. J Chem Inf Comput Sci 41,
856864.
12. Schuffenhauer, A., Ruedisser, S., Marzinzik,
A. L., Jahnke, W., Blommers, M., Selzer, P.,
Jacoby, E. (2005) Library design for fragment based screening. Curr Top Med Chem
5, 751762.
13. Chessari, G., Woodhead, A. J. (2009) From
fragment to clinical candidatea historical
perspective. Drug Discov Today 14, 668675.
14. Moriz, M., Bernd, M. (1999) Characterization of ligand binding by saturation transfer
difference NMR spectroscopy. Angew Chem
Intl Ed 38, 17841788.
15. Dalvit, C., Pevarello, P., Tato, M., Veronesi,
M., Vulpetti, A., Sundstrom, M. (2000)
Identification of compounds with binding
affinity to proteins via magnetization transfer
from bulk water. J Biomol NMR 18, 6568.
16. Jhoti, H., Cleasby, A., Verdonk, M.,
Williams, G. (2007) Fragment-based screening using X-ray crystallography and NMR
spectroscopy. Curr Opin Chem Biol 11,
485493.
17. Annis, D. A., Nickbarg, E., Yang, X., Ziebell,
M. R., Whitehurst, C. E. (2007) Affinity
selection-mass spectrometry screening techniques for small molecule drug discovery.
Curr Opin Chem Biol 11, 518526.
18. Erlanson, D. A., Braisted, A. C., Raphael,
D. R., Randal, M., Stroud, R. M., Gordon,
E. M., Wells, J. A. (2000) Site-directed ligand discovery. Proc Natl Acad Sci U S A 97,
93679372.
19. Neumann, T., Junker, H. D., Schmidt, K.,
Sekul, R. (2007) SPR-based fragment screening: advantages and applications. Curr Top
Med Chem 7, 16301642.
20. Hesterkamp, T., Barker, J., Davenport,
A., Whittaker, M. (2007) Fragment based
drug discovery using fluorescence correlation: spectroscopy techniques: challenges
and solutions. Curr Top Med Chem 7,
15821591.
21. Danziger, D. J., Dean, P. M. (1989) Automated site-directed drug design: the prediction and observation of ligand point positions at hydrogen-bonding regions on protein surfaces. Proc R Soc Lond B Biol Sci 236,
115124.

251

22. Gillet, V., Myatt, G., Zsoldos, Z., Johnson,

A. (1995) SPROUT, HIPPO and CAESA:
Tools for de novo structure generation and
estimation of synthetic accessibility. Perspect
Drug Discov Design 3, 3450.
23. Evensen, E., Joseph-McCarthy, D., Weiss,
G. A., Schreiber, S. L., Karplus, M. (2007)
Ligand design by a combinatorial approach
based on modeling and experiment: application to HLA-DR4. J Comput Aided Mol Des
21, 395418.
24. Miranker, A., Karplus, M. (1991) Functionality maps of binding sites: a multiple copy
simultaneous search method. Proteins 11,
2934.
25. Clark, M., Meshkat, S., Talbot, G. T.,
Carnevali, P., Wiseman, J. S. (2009)
Fragment-based computation of binding free
energies by systematic sampling. J Chem Inf
Model 49, 19011913.
26. Gillet, V., Johnson, A. P., Mata, P., Sike, S.,
Williams, P. (1993) SPROUT: a program for
structure generation. J Comput Aided Mol
Des 7, 127153.
27. Gillet, V. J., Newell, W., Mata, P., Myatt, G.,
Sike, S., Zsoldos, Z., Johnson, A. P. (1994)
SPROUT: recent developments in the de
novo design of molecules. J Chem Inf Comput Sci 34, 207217.
28. Nishibata, Y., Itai, A. (1991) Automatic creation of drug candidate structures based
on receptor structure. Starting point for
artificial lead generation. Tetrahedron 47,
89858990.
29. Bohm, H. J. (1993) A novel computational tool for automated structurebased drug design. J Mol Recognit 6,
131137.
30. Bohacek, R. S., McMartin, C. (1994) Multiple highly diverse structures complementary to enzyme binding sites: results of extensive application of a de novo design method
incorporating combinatorial growth. J Am
Chem Soc 116, 55605571.
31. Todorov, N. P., Dean, P. M. (1997) Evaluation of a method for controlling molecular scaffold diversity in de novo ligand design. J Comput Aided Mol Des 11,
175192.
32. Todorov, N. P., Dean, P. M. (1998)
A branch-and-bound method for optimal
atom-type assignment in de novo ligand
design. J Comput Aided Mol Des 12,
335349.
33. Ishchenko, A. V., Shakhnovich, E. I. (2002)
SMall Molecule Growth 2001 (SMoG2001):
an improved knowledge-based scoring function for protein-ligand interactions. J Med
Chem 45, 27702780.

252

Feyfant et al.

34. Wang, R., Gao, Y., Lai, L. (2000) LigBuilder:

a multi-purpose program for structure-based
drug design. J Mol Model 6, 498516.
35. Murray, C. W., Verdonk, M. L. (2002) The
consequences of translational and rotational
entropy lost by small molecules on binding
to proteins. J Comput Aided Mol Des 16,
741753.
36. Thompson, D. C., Denny, R. A., Nilakantan,
R., Humblet, C., Joseph-McCarthy, D., Feyfant, E. (2008) CONFIRM: connecting fragments found in receptor molecules. J Comput
Aided Mol Des 22, 761772.
37. Eisen, M. B., Wiley, D. C., Karplus, M.,
Hubbard, R. E. (1994) HOOK: a program
for finding novel molecular architectures that
satisfy the chemical and steric requirements
of a macromolecule binding site. Proteins 19,
199221.
38. Clark, D. E., Frenkel, D., Levy, S. A., Li,
J., Murray, C. W., Robson, B., Waszkowycz,
B., Westhead, D. R. (1995) PRO-LIGAND:
an approach to de novo molecular design.
1. Application to the design of organic

39.

40.

41.

42.

molecules. J Comput Aided Mol Des 9,

1332.
Lauri, G., Bartlett, P. A. (1994) CAVEAT:
a program to facilitate the design of organic
molecules. J Comput Aided Mol Des 8,
5166.
Yang, Y., Nesterenko, D. V., Trump, R.
P., Yamaguchi, K., Bartlett, P. A., Drueckhammer, D. G. (2005) Virtual hydrocarbon and combinatorial databases for use
with CAVEAT. J Chem Inf Model 45,
18201823.
Maass, P., Schulz-Gasch, T., Stahl, M., Rarey,
M. (2007) ReCore: a fast and versatile
method for scaffold hopping based on small
molecule crystal structure conformations.
J Chem Inf Model 47(2), 3909.
Mori, S., Abeygunawardana, C., Johnson, M.
O., van Zijl P. C. (1995) Improved sensitivity of HSQC spectra of exchanging protons at short interscan delays using a new
fast HSQC (FHSQC) detection scheme that
avoids water saturation. J Magn Reson B
108(1), 948.

Chapter 13
LEAP into the Pfizer Global Virtual Library (PGVL) Space:
Creation of Readily Synthesizable Design Ideas
Automatically
Qiyue Hu, Zhengwei Peng, Jaroslav Kostrowicki, and Atsuo Kuki
Abstract
Pfizer Global Virtual Library (PGVL) of 1013 readily synthesizable molecules offers a tremendous opportunity for lead optimization and scaffold hopping in drug discovery projects. However, mining into a
chemical space of this size presents a challenge for the concomitant design informatics due to the fact
that standard molecular similarity searches against a collection of explicit molecules cannot be utilized,
since no chemical information system could create and manage more than 108 explicit molecules. Nevertheless, by accepting a tolerable level of false negatives in search results, we were able to bypass the
need for full 1013 enumeration and enabled the efficient similarity search and retrieval into this huge
chemical space for practical usage by medicinal chemists. In this report, two search methods (LEAP1 and
LEAP2) are presented. The first method uses PGVL reaction knowledge to disassemble the incoming
search query molecule into a set of reactants and then uses reactant-level similarities into actual available
starting materials to focus on a much smaller sub-region of the full virtual library compound space. This
sub-region is then explicitly enumerated and searched via a standard similarity method using the original query molecule. The second method uses a fuzzy mapping onto candidate reactions and does not
require exact disassembly of the incoming query molecule. Instead Basis Products (or capped reactants)
are mapped into the query molecule and the resultant asymmetric similarity scores are used to prioritize
the corresponding reactions and reactant sets. All sets of Basis Products are inherently indexed to specific
reactions and specific starting materials. This again allows focusing on a much smaller sub-region for
explicit enumeration and subsequent standard product-level similarity search. A set of validation studies were conducted. The results have shown that the level of false negatives for the disassembly-based
method is acceptable when the query molecule can be recognized for exact disassembly, and the fuzzy
reaction mapping method based on Basis Products has an even better performance in terms of lower
false-negative rate because it is not limited by the requirement that the query molecule needs to be
recognized by any disassembly algorithm. Both search methods have been implemented and accessed
through a powerful desktop molecular design tool (see ref. (33) for details). The chapter will end with a
comparison of published search methods against large virtual chemical space.
Key words: LEAP, PGVL, combinatorial chemistry, library design, similarity search, disassembly,
Basis Product, symmetric similarity score, asymmetric similarity score, lead hopping.

J.Z. Zhou (ed.), Chemical Library Design, Methods in Molecular Biology 685,
DOI 10.1007/978-1-60761-931-4_13, Springer Science+Business Media, LLC 2011

253

254

Hu et al.

1. Introduction
The high attrition rate across multiple stages of the modern drug
discovery process has significantly hampered the productivity of
the pharmaceutical industry as a whole (1). One of the countermeasures implemented by pharmaceutical companies against this
challenge is to build a large and diverse library of combinatorially
enabled molecules to boost productivity in hit identification and
lead optimization (2). Through a multi-year multi-million dollar
investment in collaborations with ChemBridge, Tripos, ChemRx,
and Arqule, Pfizer has developed validated reactions for parallel
synthesis, and implemented those protocols, to expand its corporate compound collection for biological screening to 3 million
(35).
As an integral part of the parallel synthesis of these arrays
of compounds, these collaborations and internal effort produced
and validated 2500 parallel synthetic protocols spanning across
757 diverse chemical reactions. These combinatorial reactions,
their synthetic procedures, and their reactant scope and limitations are well defined and have been captured electronically for
future library production (6). Those experimentally validated synthetic protocols and their corresponding reactant sets compatible
with their reaction conditions implicitly lead to a huge chemical compound space (PGVL, or Pfizer Global Virtual Library)
with more than 1013 virtual yet synthetically feasible compounds.
All starting materials are known, specified, and available. Only
a very small fraction of PGVL has ever been synthesized (106
out of 1013 ). Conceptually a medicinal chemist can use a query
(or seed) molecule as input to search for similar molecules inside
PGVL and thereby retrieve new analogs for lead optimization and
scaffold hopping. Previous work has demonstrated that there are
many lead- and drug-like molecules in this type of large virtual
compound space spanned by combinatorial reactions selected by
medicinal chemists and existing reactant sets (7). Yet there are significant challenges inherent in making the desired similarity search
practical against such a huge chemical space. The standard similarity search methods require the construction of a file or database
containing explicit molecules. However, as of today, no chemical
information technology is known to enumerate and store more
than 108 molecules, for example, both CAS (Chemical Abstract
Service) (8) and Pubchem (9) have collections of substances in
the 107 scale.
In the publications from the Tripos group, the authors had
demonstrated that they could bypass the need for full enumeration of a huge virtual space and enable similarity search by extensively leveraging the reactant-level information (10). Even though

LEAP into the Pfizer Global Virtual Library (PGVL) Space

255

the same authors went on to demonstrate that many drug-like

molecules were found in their validation studies, their method did
not gain wide usage within the community of medicinal chemists
who are engaging in drug discovery. One could speculate that the
long turnaround time (days instead of hours or even minutes) of
a typical search session is the leading factor that has prevented this
method from being widely adopted.
As combinatorial chemistry has been fully integrated into
the modern drug discovery process, more computational search
methodologies against large virtual combinatorial compound
spaces have been steadily developed in recent years (1116). A
detailed summary and comparison of those published methods
are reported in the Section 5 and in Table 13.6. A good review
on this subject could also be found in the publication by Boehm
and coworkers (16).
In this report, two methods (LEAP1 and LEAP2, LEad-based
Analog hoPping) for performing similarity search into PGVL are
presented. The results of validation studies under controlled conditions are given to characterize their search performance profiles
in terms of false-negative rate and search speed. Finally results
from recent applications to advance two drug discovery projects
are also included in this report to highlight the fact that LEAP1
and LEAP2 are fully integrated into chemists molecular design
workflow and have been in use for more than 5 years.

2. Methods
The standard similarity search problem is commonly defined as
the following: Given an input query molecule, find the molecules
within a collection of compounds that are most similar (either top N
or within a predefined similarity threshold) to the query molecule. Of
course a molecular similarity measure has to be given between a
pair of molecules. The Tanimoto distance calculated on the basis
of molecular fingerprints is the most commonly used similarity
measure (17). Medicinal chemists routinely perform this type of
similarity searches against molecular databases containing 106
108 explicit molecules. The set of molecules returned by a similarity search is expected to be well defined by the search parameters
such as the query molecule, the search domain, and the similarity measure in combination with the underlying molecular fingerprints. In this report, we refer to this set of molecules returned
from such a search into a standard explicit database as the reference set to be used in comparison with search results obtained by
new similarity search methodologies.

256

Hu et al.

As stated before, PGVL is too large to be fully enumerated

practically. Therefore our strategy is to find a way to focus in
a just-in-time manner on much smaller sub-regions (104 ) of
PGVL for subsequent on-the-fly enumeration followed by standard similarity search against the same query molecule. It is intuitively evident that a virtual compound space built from parallel
synthesis reaction protocols has inherent array structures in the
form of implicit arrays of related just-in-time enumerated compounds, even if those compounds do not have their molecular
structures yet enumerated at the time this inherent array structure is exploited.
Hypothetically, if we could compare the set of molecules
returned by such a high-speed approach with the reference
set derived from the hypothetical search into fully enumerated
PGVL, we would expect both false positives and false negatives.
It is easy to understand the source of false negatives since only a
much smaller sub-region of PGVL is searched and some true positives outside this sub-region of PGVL would be missed. If top
N hits are returned, there would be false positives among them
since some true top N positives are missed by the limited search
and replaced by lower ranked false positives. On the other hand,
there would not be any false positives if we ask for enumerated
molecules with a similarity threshold with respect to the query
molecule. The molecules, post-enumeration, are either similar or
not, by that threshold. By accepting a tolerable level of false positives and false negatives, a similarity search strategy can be implemented to return interesting search results for practical usage by
medicinal chemists.
Even with the same general just-in-time enumeration and
search strategy, different methods for identifying and retrieving
the required smaller sub-regions of PGVL can be developed.
Their performance can be characterized by the rate of false positives and false negatives as well as search speed and ease of use.
We have implemented two search methods (LEAP1 and LEAP2)
which will be discussed in subsequent paragraphs. Yet this is an
active research area open for further innovations. A summary
comparison of LEAP1/2 with other published search methods
is given in the Appendix.
2.1. LEAP1

If a query molecule can be disassembled into combinations of

virtual reactants by in silico disconnection using one or more
reaction schemes within PGVL, then those virtual reactants can
be used to identify the most similar genuine reactants out of
all suitable genuine starting materials for those known reactions.
For a given precise reaction, similar reactant combinations always
lead to similar product molecules. This is the basic principle
used by the LEAP1 method to focus on smaller sub-region(s) of
PGVL for explicit just-in-time enumeration and similarity search.

LEAP into the Pfizer Global Virtual Library (PGVL) Space

257

More explicitly, the four key steps of LEAP1 are depicted in

Fig. 13.1:
(1) Automatic scan over all PGVL reactions for retro-synthetic
feasibility of the incoming query molecule and disassemble the
query molecule into combinations of virtual reactants. This
step is carried out automatically using the known reaction
cores from each reaction scheme within the PGVL reaction knowledge system. In the case where no reaction can
be found by the in silico disconnection engine to break
up the query molecule, then LEAP1 will fail to return any
search result back to the user. This is the major limitation of
LEAP1. For other cases, multiple reactions are identified to
disassemble the same query molecule in different ways. This
suggests that more than one sub-region of PGVL should be
explicitly enumerated and searched. This is not a problem
and is in fact a benefit. The output of this step is a list, each
entry containing an explicit parallel reaction scheme and
a specified combination of virtual reactants. By definition,
one could use each reaction scheme and the corresponding virtual reactants to form the very same original query
molecule.

O
O
N
O

Step1: Automatic scan over all PGVL

reactions for retro-synthetic feasibility of
the incoming query molecule; and
Disassemble the query molecule into
combinations of virtual reactants.

N
H

Input:
Query

Output:
Search
result

Step2: Identify suitable reactants most

similar to the corresponding virtual
reactants obtained from Step 1. in order to
focus on the most relevant sub-regions.

Step4: Perform
standard similarity
searches against
those explicit virtual
molecules using the
query molecule.

Step3: Enumerate
On-the-fly of those
identified subregions (~102 to 106)

LEAP

Fig. 13.1. Internal flowchart for the LEAP1 fully automated process. The diagram illustrates that there are three reactions
identified whose chemical spaces are colored as pink, green, and yellow, which LEAP1 automatically identified as disconnection routes. LEAP1 then retrieves the most relevant sub-region within the chemistry space, followed by on-the-fly
enumeration of those identified sub-regions. The final step can be any 2D/3D virtual screening algorithms. LEAP1 was
implemented using Scitegic fingerprint technologies.

258

Hu et al.

(2) Identify suitable reactants most similar to the corresponding

virtual reactants obtained from step 1 in order to focus on the
most relevant sub-regions. But the disconnection does not
necessarily result in bona fide known and available starting materials, after just step 1. Consider as an example a
two-component reaction which in the PGVL has M suitable bona fide reactants for the first reaction component
and N suitable bona fide reactants for the second reaction
component. Two similarity searches are used in the step to
select m (out of M) and n (out of N) reactants based on
two virtual reactants as seeds, which arose from the exact
disconnection of the query molecule. In most cases, M and
N are 103 , and m and n are 102 . Here extra search
parameters need to be specified and/or optimized for each
reaction component.
(3) Enumerate on-the-fly the sub-region(s) using the optimized
sets of bona fide reactants identified in step 2. One can see
that reduction in reactant set sizes makes explicit enumeration of product structures practical (for the same example
used, m n = 104 vs. M N = 106 and both are << the
total size of PGVL).
(4) Perform standard similarity search using the original query
molecule against the enumerated sub-region(s) obtained in
step 3. Looking at these four steps in the above discussion,
one can see that steps 1, 3, and 4 are rather straightforward,
whereas step 2 requires more tuning/optimization to get a
balanced sampling of bona fide reactants for each reaction
component to enable the precise and optimal enumeration
sub-space to achieve the best overall search results. As mentioned above, either top-N or certain similarity threshold
can be used to sample the reactant space. To balance the
performance in terms of the adequate sampling and within
reasonable runtime, top 20 reactants are used as the default
value for each component list; users do have the flexibility
to tune this number.
Since LEAP1 was built based on Pipeline Pilot technology,
multiple molecular fingerprints and similarity methods can be
applied at disposal, which currently include MDL Public Keys and
different levels of FCFPs and ECFPs (18).
2.2. LEAP2

LEAP2 was developed to overcome the major limitation of

LEAP1, the need to successfully disconnect the query molecule
into combinations of virtual reactants using reaction schemes
inside PGVL. Even though PGVL contains 757 combinatorial
reaction schemes, still experience has shown that there are many
interesting hits and lead molecules whose structures could not
be precisely disassembled. A fuzzy reaction mapping and reaction

LEAP into the Pfizer Global Virtual Library (PGVL) Space

259

retrieval step is instead required. So in LEAP2, the identification

of suitable candidate reactions and the subsequent focusing to
their optimal corresponding bona fide starting materials is done
with the help of Basis Products (BP) (19) as well as an asymmetric
similarity measure of BPs using the query molecule. For any given
query molecule, LEAP2 always returns search results to user.
Before proceeding further, a short discussion on Basis Products
and the asymmetric similarity measure between two molecules is
given in the following paragraphs.
2.2.1. Basis Products

For a given combinatorial reaction and its associated fully enumerated product space spanned by all suitable reactants, Basis
Products (BP) form a much smaller subset within the full product space and at the same time provide a systematic and efficient
sampling of all reactants suitable for that reaction. Figure 13.2
depicts an example for a two-component reaction. A Basis Product contains information about the R-groups as well as the reaction core, which can be expressed in the following statement (see
Fig. 13.2):
BP = R-groups of one reaction component
+ Reaction Core + CAP(s)from other components(s)
where CAP is the R-group of the smallest reactant from each reactant list.
In Fig. 13.2, the first row and the first column of products
are defined as the Basis Products for that reaction. Basis Products
have an one-to-one relationship with their corresponding reactants. It can be seen also that in a two-component reaction, there
are two sets of Basis Products; in a three-component reaction,
there are three sets of Basis Products; always one set per reaction
component. M plus N reactants lead to M plus N Basis Products,
while the fully combinatorial product space is M N in size. Currently there are 106 Basis Products in PGVL, far smaller than
the full PGVL space about 1013 1014 in size.
Importantly, all Basis Products are real products, members within the product space, and, like the simple truncated
R-groups, retain no transient reactant-only functional groups
(reactive halides, aldehydes, etc.); in R-group methods these disappear by clipping, whereas in Basis Products these are transmuted in the reaction transformation preparing the Basis Products. Yet unlike truncated R-groups, Basis Products also incorporate the full reaction core (all of the newly formed bonds) as
part of the BP structure. Furthermore the collection of available
starting materials, e.g., aliphatic amines, aldehydes, acyl chlorides,
benzyl halides, collapses to a fewer number of unique R-groups
when clipped, whereas the same set of starting materials expands

260

Hu et al.

a)

VRXN-2-00051
O

R1

H +
R2
N
H

N
A

R2

H
O

A_CAP

R2

Basis products for all

Br Alpha-halo ketones
(plus atom level annotations)

A_CAP + Core + R2

N
N

N
N

Basis products
for all 2-amino
heterocylces
(plus atom level
annotations)

A: aminoheterocycles

B: Alpha-halo ketones

R1

B_CAP

R1

b)

Br

Basis Product of B:
VRXN-2-00051_B_1

R1 + Core +
R1+ Core + B_CAP

N
N

N
N

Basis Product of A:
VRXN-2-00051_A_1

Full Products

Fig. 13.2. Illustration of the basic concept of Basis Products. (a) The PGVL reaction scheme of VRXN-2-00051 (formation
of the H-imidazo[1,2-a]pyridine ring system using aminoheterocycles and alpha-halo ketones) is used for the illustration;
(b) The Basis Products of A are formed by all A reactants with one constant B reactant (B_CAP, 1-bromopropan-2-one).
The Basis products of B are formed by all B reactants with a constant A reactant (A_CAP, 2-amino pyridine). The blue
triangle and yellow hexagon represent two such basis products. The red star represents a product molecule which is
related to those two corresponding basis products.

to many more unique BPs since each of these starting materials

can typically participate in many different reactions yielding different reaction product cores, hence multiple BPs arise from the
same starting material. Simply put, the structure of each Basis
Product encodes within it, and through associated database fields,
the precise combination of one reaction and the one starting
material.
All Basis Products in the PGVL have been explicitly enumerated to support numerous molecular design, fragment-based
design, and 2D and 3D methods; they also provide here a rigorous basis for the fuzzy reaction retrieval in the LEAP2 method.
In our previous publication, we have shown that knowledge
of a useful set of physicochemical molecular properties of (M+N)

LEAP into the Pfizer Global Virtual Library (PGVL) Space

261

Basis Products can be used to provide a remarkably accurate and

efficient estimation of the same molecular properties for any product molecule within a fully combinatorial product space without
enumeration (19). Of course, with the just-in-time enumeration
provided by LEAP2, such important ADMET molecular properties can also be explicitly and efficiently calculated now on product analogs rapidly mined by LEAP2. Additionally, Basis Products
have been used to anchor structure-based library design methods
when the 3D structure of a binding site of a target protein is
known (20). There is a deep connection between Basis Products,
fragment-based structure-based design (21), and parallel synthesis chemistry. In this report, we show that Basis Products are again
instrumental for the implementation of LEAP2.
2.2.2. Asymmetric
Similarity Between Two
Molecules

Asymmetric similarity measure has been first described by Tversky (22) to provide a general mathematical framework for the
perception of similarity and later adapted to molecular similarity
by Bradshaw (23). The mathematical formula for both similarity
measurements against BPs are shown below:
Symmetric similarity (SS) favors maximum common features
and penalizes non-common features:
SS =

Number of features in both Query and Basis Product

[1]
Number of features in either Query or Basis Product

Asymmetric similarity (AS) favors retrieval of basis products

with the most features embedded within the query.
AS =

Number of feature in both Query and Basis Product

[2]
Number of features in Basis Product

The well-known symmetric similarity measure rewards common features shared by two molecules and penalizes unique features present in either molecule which are not found in the other.
Its value reaches 1 only when both molecules are identical. The
asymmetric similarity measure focuses on the degree to which a
test molecule (BPs in our case) can map into the original query
molecule. When a BP molecule, which is typically smaller, is
mapped into the query molecule, the asymmetric similarity measure can still reach 1.0 when the BP can be fully mapped into the
query molecule, in another words the BP is a substructure of the
query molecule. Figure 13.3 uses a query molecule within PGVL
and its corresponding Basis Products to highlight the difference
between symmetric and asymmetric similarity measures. From the
differences of the AS and SS scores of the same BP, it is seen that
indeed the standard symmetric similarity measure penalizes any
differences between two molecules, while the asymmetric similarity measure used in LEAP2 focuses on mapping the Basis Product into the query molecule, while ignoring the unique features

262

Hu et al.

Query molecule
O

N
N
N

Symmetric Similarity (SS)

Asymmetric Similarity (AS)

Basis Products

Basis Products

SS=82%

AS=98%

VRXN-2-00051_A_1

VRXN-2-00051_A_1

SS=84%

AS=100%

N
N

VRXN-2-00051_B_1

N
N

VRXN-2-00051_B_1

Fig. 13.3. Comparison of symmetric and asymmetric similarity scores. A virtual product
from VRXN-2-00051 is used as a query molecule. The two corresponding Basis Products
are VRXN-2-00051_A_1 and VRXN-2-00051_B_1. In reference to the query molecule,
their corresponding similarity scores are listed under SS and AS (see equations [1] and
[2] for details), respectively, depending on the similarity methods used.

in the query molecule which extend beyond the Basis Product

structure and those are analyzed using AS with the other BP sets
from the other reaction components of the same reaction, which
serve to scan these other R-group positions. Since AS is a similarity measure, high AS can be achieved without the need for precise
substructure embeddability, hence this is still a fuzzy mapping.
We have hypothesized that when a Basis Product has a high
asymmetric similarity value to a query molecule, there should
be a higher probability that the candidate reaction and the specific available reactant encoded by the Basis Product will be associated with sub-regions of PGVL space where full-size product
molecules most similar to the query molecule will be found. This
is the principle based on which LEAP2 focuses the reaction search
and retrieval. Sub-regions of PGVL spanned by reactions and
reactants encoded by those Basis Products that map most favorably into the query molecule are detected by AS and these subregions advance to the next step.
2.2.3. Search Steps in
the LEAP2 Algorithm

(1) Search a database of Basis Products using Asymmetric Similarity measure. Here this search is done using the query
molecule against a database of 106 explicit enumerated
Basis Products. The asymmetry similarity search in the BP
database is implemented using MDL Keys finger print (24)
with ISIS host technology (25).

LEAP into the Pfizer Global Virtual Library (PGVL) Space

263

The output is a set of Basis Products with high asymmetric similarity (AS) values (the default cutoff value is set
to 90%) when they are mapped into the query molecule.
The reaction schemes and reactants encoded by those Basis
Products are then extracted, ranked, and used to form subregions of PGVL for subsequent just-in-time enumeration
and symmetric similarity search against the query molecule.
Similarly to LEAP1, the top 20 similar molecules per
reaction component list are used, as a default setting, to
ensure balanced sampling of reactants for each reaction
component and the reasonable performance. This is user
adjustable.
(2) Enumerate sub-region(s) using the smaller sets of reactants
identified in step 1. This on-the-fly enumeration step is identical to step 3 of LEAP1.
(3) Perform standard similarity search using the original query
molecule against the enumerated products from the subregion(s) obtained in step 2. This is identical to step 4 of
LEAP1.
Since LEAP2 was also built based on Pipeline Pilot technology, multiple molecular fingerprints and similarity methods can
be applied at disposal, which currently include MDL Public Keys
and different levels of FCFPs and ECFPs (18).

3. Results and
Discussion
3.1. Method
Validation and
Performance
Profiling

As mentioned before, LEAP1 and LEAP2 are the results of conscious choices between accuracy and practical execution performance. Therefore it is important to conduct a set of controlled
validation studies to assess the accuracy in terms of rates of false
positives and false negatives in their search results and performance in terms of end-to-end search turnaround time.
To reach those objectives, we conducted validations to answer
the following questions:
(1) Given a set of molecules known to be inside PGVL as
query molecules, what is the success rate for returning the
expected molecules identical to the query molecules (100%
similarity threshold)? This is by definition a baseline test
that a validated search strategy must pass.
(2) Given a sub-region of PGVL that can be enumerated
explicitly and a query molecule, compare search results
obtained by a LEAP search with the reference sets obtained
by the exhaustive search against the fully enumerated

264

Hu et al.

sub-region. What are the rates of false positives and false

negatives?
The false negative is referred to those true positives outside the sub-regions of PGVL found by LEAP methods
which would be missed as hits.
The false positives are due to the top N similar approach
which resulted in some true top N positives missed by the
limited search and replaced by lower ranked false positives.
(3) Given a set of drug-like molecules as query molecules, what
is the success rate of a search method in returning interesting search results not only similar to the input queries but
also pertinent to lead optimization and/or lead hopping?
Test One: Thirteen product molecules from 13 PGVL reactions (five 2-component VRXNs, three 3-component VRXNs,
and five 4-component VRXNs) were randomly selected as query
molecules for the first validation test. LEAP1 identified all 13
PGVL reactions and returned all 13 expected identical molecules.
LEAP2 correctly located all 13 expected PGVL reactions and 12
expected molecules in its search results using the default setting
(90% AS score and top 20 similar molecules per component list
per VRXN). Due to the relative larger molecular size of CAP
molecule vs. the reactant, the BP corresponding to the missing
molecule shows only a modest 62% AS score, so it was not found
at the BP level (LEAP2 step 1). If we consider lowering the AS
score cutoff, then many more molecules per reaction component
will have to be included in the interior of the processing (LEAP2
steps 2 and 3) which will result in impact on performance. This
case also highlighted the nature of the balancing act between
speed and accuracy.
Test Two: A much small sub-region of PGVL with only
224,700 product molecules was constructed explicitly. Table
13.1 gives the details of the sub-region, which spans seven PGVL
reactions (four of them are two-component reactions and three
of them are three-component parallel synthesis reactions). At the
time of the test, the full virtual space spanned by those seven reactions in combination with their suitable reactants was about 389
million in size. We randomly selected smaller sub-sets of those
reactants to form this sub-region as a controlled environment for
this validation study. The query molecule is also given in Table
13.1, which was chosen so that it can be disassembled by all seven
VRXNs. The validation results are given in Table 13.2 for LEAP1
with different similarity thresholds used. As the similarity threshold used in the searches dropped from 1.0 (for exact match) to
0.48, the number of returned molecules went from 1 to 1807
for the exhaustive search against the enumerated set of 224,700
explicit molecules. It is reassuring to see that 94% or more of
expected molecules were correctly identified by LEAP1 while the

LEAP into the Pfizer Global Virtual Library (PGVL) Space

265

Table 13.1
Construction of a fully enumerated virtual library
space (VL) for the second validation study
Mapped

Seed Structure

VRXN-2-00004
VRXN-2-00006
VRXN-2-00010
VRXN-2-00011
VRXN-3-00063
VRXN-3-00064
VRXN-3-00065

O
N
S

Real VL size

VRXNs

Validation VL size

438 x 1171
544 x 264
3371 x 6635
449 x 7044
19 x 721 x 5697
77 x 389 x 5175
44 x 632 x 444

60 X 60
50 X 50
60 X 60
50 X 50
18 X 50 X 50
50 X 50 X 50
17 X 50 X 50

Total: 388,798,585

Total:
224,700

Table 13.2
True-positive and false-negative rates of the LEAP1 method as a function of search
threshold for molecular similarity

Similarity
threshold

No of cpds
retrieved by
LEAP1

No of cpds
retrieved by
exhaustive
search

No of true
positives in % true positive
LEAP1
in LEAP1

No of false
negatives
in LEAP1

% false
negative
in LEAP1

100

0.9

100

0.8

11

11

11

100

0.7

51

52

51

98

0.6

249

257

249

97

0.55

530

557

530

95

27

0.52

915

968

915

95

53

0.5

1331

1410

1331

94

79

0.48

1699

1807

1699

94

108

rate of false negatives remained less or equal to 6%. Figure 13.4

graphically depicts the true-positive rate and false-negative rate
of this validation test. Using a more common similarity threshold of 0.8, LEAP1 gave identical search results as those from the
exhaustive search method. The false-positive rate is zero.
Table 13.3 shows the performance comparison of LEAP1
method vs. the exhaustive search. The speedup factor is the
ratio between search times required by the exhaustive search and
LEAP1. It is seen that in exchange for a 6% false-negative rate
we can get more than a 27,000-fold speedup. If we assume that

Hu et al.

100
90
80
70
% of cpds

266

60

%True positive in LEAP1

50

%false negative in LEAP1

40
30
20
10
0
0.4

0.5

0.6

0.7
0.8
Similarity threshold

0.9

Fig. 13.4. Performance of LEAP1 when compared against the exhaustive search in the
second validation study. See main text for details.

Table 13.3
Comparison of performance of LEAP1 vs. exhaustive search
Method
LEAP1
Exhaustiv_search
Speedup factor

Validation VL (s)
446
9700
22

Real VL (s)
602
16,783,917a
27880

a Estimated based on the reasonable assumption that standard search time is propor-

tional to the size of the VL to be searched. The exhaustive search time against a
smaller VL of 224,700 is 9700 s. Therefore we have estimated to the first approximation that an exhaustive search against the real VL of 388,798,585 would take
16,793,917 s (or about 194 days) to complete. See Table 13.1 for VL sizes. Time is
in the unit of second and based on single 3 GHz Pentium CPU.

search time required by the exhaustive search is proportional to

the size of the VL to be searched, it is obvious that 194 days
of exhaustive search is not practical; however, the 602 s LEAP1
search can be performed routinely.
Test Three: For the third validation test, we selected 24 known
drugs on the market as query molecules (see Fig. 13.5). This is
a very realistic and challenging set in terms of diversity in their
molecular structures and complexities required for their synthesis. The top 10 most similar virtual compounds to each query
molecule were identified and plotted as color dots in Fig. 13.6
for both LEAP1 and LEAP2.
For every query molecule, LEAP2 is able to return top 10
molecules with best similarity scores ranging from 0.4 for sertraline to 0.9 for celecoxib. Five of 24 searches return PGVL
hits 80% or more similar to the query molecules for meaningful follow-up. If the threshold is relaxed to 0.7, then 11 of 24

LEAP into the Pfizer Global Virtual Library (PGVL) Space

267

AZITHROMYCIN

CAFFEINE

CELECOXIB

VALSARTAN

EFAVIRENZ

VENLAFAXINE

FLUCONAZOLE

FLUOXETINE

ALENDRONATE

GABAPENTIN

IBUPROFEN

ATORVASTATIN

OLANZAPINE

NELFINAVIR

ESOMEPRAZOLE

AMOLDIPINE

PAROXETINE

CLOPIDOGREL

LANSOPRAZOLE

RANITIDINE

RISPERIDONE

SILDENAFIL

SIMVASTATIN

SERTRALINE

Fig. 13.5. A diverse set of 24 drug molecules on the market is compiled for the third validation study.

searches lead to PGVL hits for meaningful follow-up. The PGVL

reactions identified at the same time can also be utilized by medicinal chemists to propose and evaluate new (not yet available) reactants for closer-in lead optimization and/or scaffold hopping, as
indicated by the unique needs of the target active site. This point
becomes even more significant given the observation that many
(17/24) LEAP2 searches beneficially yielded top 10 hits originating from more than two PGVL reactions, while celecoxib, atorvastatin, amlodipine, lansoprazole, efavirenz, sildenafil, and sertraline
are the exceptions to the trend.
As expected, due to the intrinsic requirement for precise
disassembly of query molecules using PGVL reactions, LEAP1
failed to give any search results for fluconazole, alendronate,
gabapentin, esomeprazole, and efavirenz. For the remaining 19
cases with hits, only 3 (3/19) LEAP1 searches lead to top 10
compounds originating from more than one PGVL reaction (fluoxetine/2, nelfinavir/2, and simvastatin/2). For the remaining
16 cases, precisely one PGVL reaction is identified. Three of 24
LEAP1 searches return PGVL hits 80% or more similar to the
query molecules for meaningful follow-up. If the threshold is

268

Hu et al.
LEAP1

LEAP2

0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2

drug

Fig. 13.6. Results from the third validation study. The y-axis represents the Tanimoto similarity score of returned hits
with respect to their corresponding query molecule, calculated based on the FCFP4 molecular fingerprints (31). The xaxis are drug molecules in Fig. 13.5. Search hits are color coded by the PGVL reactions (VRXN) where they are originated
from.

relaxed to 0.7, then 5 of 24 searches lead to PGVL hits for meaningful follow-up.
Based on the validation study it seems that the typical search
time required for both LEAP1 (15 min) and LEAP2 (45 min)
are short enough for routine and practical usage by medicinal
chemists. On average LEAP2 is about three times slower than
LEAP1, due to its larger VRXN coverage. LEAP2 in essence uses
a fuzzy reaction retrieval strategy which returns more candidate PGVL regions of interest in the intermediate steps of the
algorithm.
3.2. Application of
LEAP1 and LEAP2

Two examples are included here to highlight how both LEAP1

and LEAP2 have been used, routinely, by chemists for idea generation and lead hopping.

3.2.1. LEAP1

In the first example, LEAP1 was used to help generate novel

lead series against an anti-obesity target, MCH (melanin concentrating hormone) (26, 27). Fourteen query molecules consisting of both in-house and literature leads were used as search
input, and the product similarity threshold was set as the top 50
final output molecules per VRXN per query, with default settings for remainder of the parameters. The LEAP1 search led
to 7200 hits, all synthesizable based on PGVL chemistries and
parallel synthesis protocols. An additional design step based on

LEAP into the Pfizer Global Virtual Library (PGVL) Space

269

Table 13.4
Two example virtual hits from among hundreds in the 14 LEAP1 searches
Structure

Name

monomerID

VRXN

simScore

Est_IC50 (nM)*

Br

N
H

MCH-1

MFCD01443686:
VRXN-2-00001
MFCD00238752

0.37

0.64

MCH-2

MN-011201:MN017087

0.33

0.69

F
FF N

Cl
N
Cl
N

VRXN-2-00001

a IC50 estimated by 3D Pharmacophore model of MCH, their corresponding mappings are shown

in Fig. 13.7.

MCH-1

MCH-2

Fig. 13.7. Novel synthesizable compounds (see 2D structures in Table 13.4) produced
by LEAP1 searches with high score judged by a project-specific 3D pharmacophore
model (red blob: basic feature, light blue blob: hydrophobe, green vector blob: hydrogen
bond acceptor).

a 3D pharmacophore model using Catalyst TM (28) was used

to further filter those virtual hits by MCH activity, thus leading
to 21 final compounds. Two example virtual hits are shown in
Table 13.4 together with their similarity score to the respective
query molecules and estimated IC50 based on their corresponding 3D pharmacophore mappings shown in Fig. 13.7. A targeted
library using the corresponding chemistry, VRXN-2-00001, was
launched, which resulted in 61 compounds prepared with an average 2D similarity around 30% to the original 14 query molecules.
Thirteen hits from the pharmacophore-directed LEAP1 targeted
library have shown greater than 60% inhibition at 1 M in the
MCH enzymatic assay. Due to the transfer of project and therapeutic area elimination, the lead development for those hits was
stopped.
3.2.2. LEAP2

In the second example, LEAP2 was utilized to help generate novel

leads against an anti-angiogenesis target, caspase-3 (29, 30). A literature compound (PAC-1 with reported 3.1 M IC50 ) and five
of its analogs (31) were used as query molecules, and the product similarity threshold was set as top 50 final output molecules

270

Hu et al.

per VRXN per query molecule, with default settings for remainder of the parameters. The LEAP2 search resulted in 900 hits
originating from 18 PGVL chemical reactions. Three targeted
libraries were subsequently designed and synthesized based on
those LEAP2 hits. The efforts resulted in 281 compounds synthesized, of which 13 yielded IC50 ranging from 1 to 20 M
(see Table 13.5). The result demonstrated that LEAP2 method
is capable of generating multiple different design ideas which can
be implemented quickly and fruitfully by the project team.

Table 13.5
Hits from the caspase-3 targeted libraries
Compound_Number

IC50 (M)

VRXN_IDa

Cpd-1

1.01

VRXN-2-00086

Cpd-2

1.85

VRXN-2-00086

Cpd-3

3.36

VRXN-2-00086

Cpd-4

3.56

VRXN-2-00086

Cpd-5

5.82

VRXN-2-00086

Cpd-6

6.03

VRXN-2-00086

Cpd-7

7.46

VRXN-2-00086

Cpd-8

7.69

VRXN-2-00086

Cpd-9

12.5

VRXN-2-00010

Cpd-10

14.2

VRXN-2-00086

Cpd-11

16.7

VRXN-2-00086

Cpd-12

17.5

VRXN-2-00086

Cpd-13

19.5

VRXN-2-00010

a VRXN-2-00086 (hydrazone formation) and VRXN-2-00010 (amide formation)

4. Conclusion
It is very useful to emphasize systematic data capture within an
organization as large as Pfizer. It has been beneficial to derive
knowledge from those data in projects and sites different from
the original settings which led to the original development of a
given reaction protocol and most valuable if this knowledge can
be reused in the essential operations on a regular basis.
The PGVL system is a large-scale knowledge system derived
from rigorous multi-year systematic reaction knowledge capture,
including the registration of large numbers of bona fide starting materials and validated parallel synthesis reaction protocols.
LEAP chemistry space mining methodologies are ways to enable

LEAP into the Pfizer Global Virtual Library (PGVL) Space

271

the efficient reuse of this knowledge in a practical manner, and

this capacity is unleashed simply by entering the structure of the
new lead at hand. In this sense, the usage of LEAP1 or LEAP2 is
a lead-centric mining capability, as far as the user is concerned.
The validation studies show that the LEAP methods produce
results reasonably comparable to exhaustive search and enable
medicinal chemists with a practical method for the automated
suggestion of synthesizable analogs for lead optimization and lead
hopping.
In order to retrieve readily synthesizable virtual compounds
from PGVL that are useful for virtual screening and for formulating the next synthesis plan, the LEAP-based methods can be
used by itself or coupled together with other fundamental targetspecific design methods, such as 3D pharmacophore modeling,
docking, and SBDD, by simply replacing the final product similarity step with those well-known 3D design methods.

5. Notes on
Comparison with
Other Published
Search Methods
Against Large
Virtual Chemical
Space

Table 13.6 provides a comparison among several leading search

methods in terms of their origin, search time, scope and nature
of chemical space, format of input query ligand, and molecular
similarity measure used.
1. Origin. Molecular similarity search into very large VLs is of
great interest for drug discovery and is becoming a thriving
research field. Methods 12 are from commercial software
companies. The rest are developed by major pharmaceutical
companies to facilitate their internal drug discovery. Four of
the six are from Pfizer alone, all based on the PGVL chemical
space (6).
2. Turnaround time of a search. The performances of most
methods are within minutes for any single run with one
query molecule, except for Method 1 and Method 2. The
relatively long run time for Methods 1 and 2 is mainly due to
the associated 3D searching technologies. For normal drug
discovery process, run time within minutes or even hours
are acceptable. For a computational technology to make a
real impact, run time in month or even week scale is hard to
justify the investment, at least in routine manner. With the
current hardware and software advancement, one can imagine a coarse-grained parallelization for those methods with
3D searching need to significantly speed up those processes.
In summary, minutes or even hours are acceptable but days

NA

AllChem

FTree-FS

LEAP1

LEAP2

MoBSS

CoLibri/
FTrees-FS

CoLibri/
FTrees-FS

15

16

14

11, this
report

11, this
report

13

12

Ref#

Boehringer
Ingelheim (BI)/
BioSolvIT

Pfizer/BioSolvIT

Pfizer

Pfizer

Pfizer

Roche

Tripos

Algodign

Origin

Min

Min

Min

Min

Min

Min

Hour

Month

PGVL
PGVL
PGVL

534a
441a
358a

BI CLAIM
(Comprehensive
Library of
Accessible
Innovative
Molecules)

PGVL (Pfizer
Global Virtual
Library)

157a

NA

RECAP/TOPAS

Tripos Discovery
Research (TDR)

Literature

Source of chemical reactions

11

100

400

# of
chemical
reactions

1.00E+11

1.00E+13

1.00E+13

1.00E+13

1.00E+13

1.00E+18

1.00E+20

1.00E+13

Size of
virtual
library
space

2D ligand

2D ligand

2D ligand

2D ligand

2D ligand

2D ligand

3D/2D
ligand

3D target

Query
input

2D FeatureTree

2D FeatureTree

2D Atom Pair
(AP)

2D

2D

2D FeatureTree

3D Topomer

3D docking

Similarity
measure

Although there are 700+ VRXNs in the PGVL system, not all of them are registered to the full extent to enable the LEAP1 and LEAP2 search. For MoBSS and FTree-based
methods, due to the assumptions made in the finger print additivity, some VRXNs, such as variable ring formation which depends on the reactant combinations used, were
excluded from the implementation. For CoLibri/FTrees-FS method, the final enumeration step was implemented using CoLibri technology which is different from the PGVL
foundation system, so certain VRXNs are excluded as well.

a Those methods based on PGVL are implemented at different times, LEAP1 is the first among all four methods. The rest of the three are second-generation methods.

Method
name

No.

Search
turnaround
time

Table 13.6
Summary and comparison of representative methods to search into large virtual chemical space indexed by combinatorial
libraries

272
Hu et al.

LEAP into the Pfizer Global Virtual Library (PGVL) Space

273

and beyond are not for practical application to impact drug

discovery.
3. Scope and nature of chemical space. For Algodign, the entire
chemical space is constructed based on chemistry from literature (7). For Tripos, Tripos Discovery Research (TDR),
the former contract research division, provided most of
the chemistry foundation for the virtual chemistry space
(12). For Method 3, 11 simple reaction schemes implemented in RECAP (32) are used for both fragmentation
and building block assembly. All four methods from Pfizer
(LEAP1/2 in this report and two others in references (14)
and (16)) are built based on PGVL (6) and enriched with
library chemistry from File Enrichment (3, 4). Method from
Boehringer Ingelheim is also built based on a collection of
in-house library chemistries (15). The key differentiation
factor here is synthetic feasibility of the result molecules. If
the virtual space are constructed based entirely on a large
pool of validated chemistry with step-by-step procedures for
every library protocol and available starting materials, then
it ensured that every hit found can be rapidly made and
expanded synthetically. Size matters, but synthetic feasibility is even more significant. Methods with a large pool of
validated chemistries, protocols, and starting materials, such
as PGVL and BI CLAIM, have this advantage.
4. Similarity measure. Method 1 uses a structure-based de
novo-like scoring function. The input query is not a ligand structure but a 3D structure of the active site for a
target (7). For Method 2, it seems that search input can
be either complete query structure or individual synthon;
and the search result can be evaluated using any combination of filters such as (topomer) shape similarity (automatically generated topomer CoMFA), potency predictions,
size, hydrophobicity, chemical reactivity, and synthetic accessibility (12). FeatureTree is used by Methods 3, 7, and 8
to compute molecular similarity. Two of them are implemented within the CoLibri library tool from BioSolvIT
(15, 16). Method 6 employed atom pair (AP) descriptors
derived from inter-atomic topological distances to compute
molecular similarity (14). LEAP1 uses retro-synthetic analysis to break the input query molecule and applies similarity
search at the fragment-level and product-level consecutively.
LEAP2 uses the asymmetric similarity score between query
molecule and Basis Products to focus on a subset of reactants. Then the standard symmetric similarity score between
the query and the explicit product molecules is used to select
final hits by LEAP2. Both LEAP methods are built based on
Pipeline Pilot technology, so multiple molecular fingerprints

274

Hu et al.

and similarity methods can be applied at disposal, which

currently include MDL Public Keys and different levels of
FCFPs and ECFPs (18). It is also expected that other similarity algorithms can be applied in the similar manner, as
long as they can be integrated into the Pipeline Pilot (18)
framework.
LEAP1/2 are easy to understand, implement, and have been
in service since 2005 for idea generation and lead hopping inside
a powerful and user-friendly molecular design tool called PGVL
Hub (33). It also offers the general framework that encompasses
all search methods against large combinatorial virtual spaces published so far in the literature (two main steps: 1. reactant/Rgroup focusing to reduce a large combinatorial virtual space into
a much smaller and manageable one; 2. product-level similarity
search within the reduced space). This framework suggests that
the reactant/R-group focusing step is the major time saver while
the extra saving harvested by estimation of product FPs based
on additive nature of certain types of FPs (atom pair in MoBSS
(14) and FTrees-FS in CoLibri (15, 16)) is only secondary with
considerable cost introduced in terms of additional complexity
to encode and ensure that combination rule is working properly
for many combinatorial reactions and the restriction in choices
of fingerprint and similarity measure. This framework also suggests that by working with explicitly enumerated products within
that much reduced virtual space, one can apply any molecular
finger print and any similarity measure without any approximation to the steps of reactant/R-group focusing as well as the final
produce-level similarity search. This would allow users the flexibility to choose the more familiar similarity measures based on
Tanimoto coefficient on top of FPs from MDL MACSS, Daylight, and SciTegic for close analogs or use atom pair FPs or FTree
with higher abstraction for more aggressive and non-obvious lead
hopping.
Finally we hope to see that more validation studies are conducted to compare any new search method with the reference
exhaustive search (of course on a smaller validation virtual space of
104 106 ). Only through this type of rigorous validation studies,
one can truly probe the rates of false positives and false negatives
as well as the fold increase in search speed. This in turn allows end
users to make informed decisions on which search method will be
a best match for their specific tasks.

Acknowledgments
The authors would like to thank the following Pfizer colleagues
for their generous help and support: Bo Yang, Thom Shulok,

LEAP into the Pfizer Global Virtual Library (PGVL) Space

275

Sarathy Mattaparti, Bo Chao, Tom Thacher, and Joe Zhou (for

their work on the PGVL software and its reaction and starting
material data foundation which enabled the development and
deployment of LEAP1/2); Bob McDonough, Zi Yang, and Da
Tse (for informatics support); and Gigi Paderes, Klaus Dress,
Dilip Bhumralkar, and Michele Ramirez-Weinhouse (for being
the early adopters of LEAP1/2 and applying them vigorously in
their drug discovery projects); Ben Burke and Zhongxiang (Joe)
Zhou for their valuable comments, suggestions, and proof reading the draft. We also appreciate the technical support from Derek
Stonich and Anne Li-Zhong of SciTegic/Accelrys.
References
1. Kola, I., Landis J. (2004) Can the pharmaceutical industry reduce attrition rates? Nat
Rev Drug Discov 3, 711715.
2. Milne, G. M. (2003) Pharmaceutical productivity: the imperative for new paradigms.
Annu Rep Med Chem 38, 383396.
3. Estep, K. (2004) File Enrichment and Hit
Follow Up: Evolution and Examples. Poster
Presentations at the ALA LabFusion.
4. Smith, G. F. (2006) Enabling HTS
Hit follow-up via Chemo informatics, File Enrichment, and Outsourcing.
High Throughput Medicinal Chemistry II;
MMS Conferencing & Events Ltd., Institute of Physics; London. This article is
also available on-line via this web link
(http://www.mmsconferencing.com/pdf/
htmc/g.smith.pdf).
5. Borman, S. (2006) Improving efficiency. To
eliminate R&D bottlenecks, drug companies
are evaluating all phases of discovery and
development and are using novel approaches
to speed them up. Chem Eng News 84,
5678.
6. Peng, Z., Yang, B., Mattaparti, S., Shulok, T.,
Thacher, T., Kong, J., Kostrowicki, J., Hu,
Q., Na, J., Zhou, J. Z., Klatte, K., Chao, B.,
Ito, S., Clark, J., Coner, C., Waller, C., Kuki,
A. (2010) PGVL Hub: an integrated desktop tool for medicinal chemists to streamline design and synthesis of chemical libraries
and singleton compounds. Chemical Library
Design, in (Zhou, J. Z., ed.), Humana Press,
New York, NY.
7. Nikitin, S., Zaitseva, N., Demina, O.,
Solovieva, V., Mazin, E., Mikhalev, S.,
Smolov, M., Rubinov, A., Vlasov, P., Lepikhin, D., Khachko, D., Fokin, V., Queen,
C., Zosimov, V. (2005) A very large diversity
space of synthetically accessible compounds
for use with drug design programs. J Comput
Aided Mol Design 19, 4763.

8. Chemical Abstract Service: http://www.cas.

org/, under substances count
9. Pubchem: http://www.ncbi.nlm.nih.gov/
sites/entrez?cmd=search&db=
pccompound&term=all[filt].
10. Andrews, K. M., Cramer, R. D. (2000)
Toward general methods of targeted library
design: topomer shape similarity searching
with diverse structures as queries. J Med
Chem 43, 17231740.
11. Hu, Q., Kostrowicki, J., Peng, Z., Kuki, A.
(2008) LEAP into the Pfizer Global Virtual
Library (PGVL) space creation of the readily synthesizable design ideas automatically,
Scitegic Pipeline Pilot User Group Meeting,
San Diego, CA.
12. Cramer, R.D., Soltanshahi, F., Jilek, R.,
Campbell, B. (2007) AllChem: generating
and searching 1020 synthetically accessible
structures. J Comput Aided Mol Des 21,
341350.
13. Rarey, M., Stahl, M. (2001) Similarity
searching in large combinatorial chemistry spaces. J Comput Aided Mol Des 15,
497520.
14. Yu, N., Bakken, G. A. (2009) Efficient
exploration of large combinatorial chemistry spaces by monomer-based similarity searching. J Chem Inf Model 49,
745755.
15. Lessel, U., Wellenzohn, B., Lilienthal, M.,
Claussen, H. (2009) Searching fragment
spaces with feature trees. J Chem Inf Model
49, 270279.
16. Boehm, M. Wu, T., Claussen, H., Lemmen,
C. (2008) Similarity searching and scaffold
hopping in synthetically accessible combinatorial chemistry spaces. J Med Chem 51,
24682480.
17. Chen, X., Reynolds, C. H. (2002) Performance of similarity measures in 2D fragmentbased similarity searching: comparison of

276

18.
19.

20.

21.

22.
23.

24.

25.
26.

27.

Hu et al.
structural descriptors and similarity coefficients. J Chem Inf Comput Sci 42,
14071414.
Pipeline Pilot from SciTegic: http://www.
scitegic.com/
Shi, S., Peng, Z., Kostrowicki, J., Paderes,
G., Kuki A. (2000) Efficient combinatorial
filtering for desired molecular properties of
reaction products. J Mol Graph Model 18,
478496.
Zhou, Z., Shi, S., Na, J., Peng, Z.,
Thacher, T. (2009) Combinatorial librarybased design with basis products. J Comput
Aided Mol Des 23, 725736.
Lau, W., Hepworth, D., Magee, T., Du,
J., Bakken, G., Miller, M., Hendsch, Z.,
Thanabal, V., Kolodziej, S., Xing, L., Hu,
Q., Narasimhan, L., Love, R., Charlton,
M., Hughes, S., Van Hoorn, W., Mills, J.,
Withka, J. (2010) Design of a multi-purpose
fragment screening library using molecular
complexity and orthogonal diversity metrics.
J Comput-Aided Mol Des.
Tversky, A. (1977) Features of similarity. Psycholog Rev 84, 327352.
Bradshaw, J. (1997) Introduction to the
Tversky Similarity Measure. Presented at
Daylight MUG Meeting, Laguna Beach, CA,
URL http://www.daylight.com/meetings/
mug97/agenda97/Bradshaw/MUG97/
tvtversky.html.
Durant, J. L., Leland, B. A., Henry, D.
R., Nourse, J. G. (2002) Reoptimization of
MDL keys for use in drug discovery. J Chem
Inf Comput Sci 42, 12731280.
ISIS host from Symyx: http://www.symyx.
com/products/software/cheminformatics/
isis-host/index.jsp
Qu, D., Ludwig, D.S., Gammeltoft, S. et al.
(1996) A role for melanin-concentrating hormone in the central regulation of feeding
behavior. Nature 380, 243247.
Saito, Y., Nothacker, H., Wang, Z., et al.
(1999) Molecular characterization of the

28.

29.

30.

31.

32.

33.

melanin-concentrating hormone receptor.

Nature 400, 265269.
Li, H., Sutter, J., Hoffmann, R. (2000)
HypoGen: an automated system for generating predictive 3D Pharmacophore Models.
Pharmacophore Perception, Development, and
use in Drug Design,in (Gner, O. F., ed.),
International University Line, La Jolla, CA.
Nachmias, B., Ashhab, Y., Ben-Yehuda, D.
(2004) The inhibitor of apoptosis protein family (IAPs): an emerging therapeutic target in cancer. Semin Cancer Biol 14,
231243.
Schimmer, A. D., Dalili, S., Riedl, S. J.
(2006) Targeting XIAP for the treatment
of malignancy. Cell Death Different 13,
179188.
Putt, K. S., Chen, G. W., Pearson, J. M.,
Sandhorst, J. S., Hoagland, M. S., Kwon,
J. T., Hwang, S. K., Jin, H., Churchwell, M. I., Cho, M. H., Doerge, D. R.,
Helferich, W. G., Hergenrother, P. J. (2006)
Small molecule activation of procaspase-3 to
Caspase-3 as a personalized anti-cancer strategy. Nat Chem Biol 2, 543550.
Lewell, X. Q., Judd, D. B., Watson,
S. P., Hann, M. M. (1998) RECAP
retrosynthetic combinatorial analysis procedure: a powerful new technique for
identifying privileged molecular fragments
with useful applications in combinatorial
chemistry. J Chem Inf Comput Sci 38,
511522.
Peng, Z., Yang, B., Mattaparti, S., Shulok,
T., Thacher, T., Kong, J., Kostrowicki, J.,
Hu, Q., Na, J., Zhou, J. Z., Klatte, K.,
Chao, B., Ito, S., Clark, J., Coner, C., Waller,
C., Kuki, A. (2011) PGVL Hub: an integrated desktop tool for medicinal chemists to
streamline design and synthesis of chemical
libraries and singleton compounds, in (Zhou,
J. Z. ed.) Chemical Library Design. Humana
Press, New York, Chapter 15.

Section IV
Library Design for Kinase Family

Chapter 14
The Design, Annotation, and Application
of a Kinase-Targeted Library
Hualin Xi and Elizabeth A. Lunney
Abstract
We present here a workflow for designing a kinase-targeted library (KTL) with the goal of capturing
known kinase inhibitor chemical space. We validated our design retrospectively using recent, highthroughput screening data and found significant enrichment of kinase inhibitor hits while retaining
majority of the active kinase inhibitor series. To further assist kinase projects in triaging KTL screen
hits, we also developed a methodology to systematically annotate known kinase inhibitors in the KTL
with regard to their binding modes.
Key words: Protein kinase, kinase-targeted library, library design, kinase chemical cores,
substructure search, SMARTS Query, subsetting, binding mode annotation.

1. Introduction
The protein kinase family is one of the largest gene families
encompassing almost 2% of the human genome. The enzymes
phosphorylate proteins through the catalytic transfer of phospho
groups from ATP to the protein substrates. Protein kinases play
key roles in numerous cellular pathways that impact multiple cellular events such as growth, division, differentiation, and apoptosis. From a pharmaceutical perspective, kinases have been targeted
in drug design across multiple therapeutic areas. The most prominent of these is oncology, for which eight small molecule kinase
inhibitors have currently been approved in the USA.
This family of proteins exhibits a common fold that results in
a two-lobe structure: a smaller N-terminal region connected by
J.Z. Zhou (ed.), Chemical Library Design, Methods in Molecular Biology 685,
DOI 10.1007/978-1-60761-931-4_14, Springer Science+Business Media, LLC 2011

279

280

Xi and Lunney

a hinge segment to the larger C-terminal portion. ATP binds in

a well-defined pocket located between the two lobes and forms
hydrogen bond interactions with the hinge. In addition, most
protein kinases exist in both an active and an unactivated state,
the latter of which can be very flexible. Targeting the ATP binding site in the active form plus the various unactivated conformations in drug design has led to discovery of numerous inhibitor
cores or templates that can bind to members of the kinase family.
Compiling a kinase-targeted library (KTL) of compounds representing this chemical landscape can greatly assist in jump-starting
an early-stage project by affording a very efficient means of discovering lead matter and tool compounds. The collection can be
readily screened and the analysis of assay results provides insights
into structureactivity relationships and selectivity profiles. Annotation of inhibitors in the KTL in terms of chemical cores and
potential binding modes further assists scientists in identifying
compound series for hit-to-lead optimization.

2. Materials
Kinase protein/ligand crystal complexes were retrieved from the
Pfizer Crystal Structure Database, an in-house X-ray structure
repository that contains internally solved structures and selected
ones imported from the Protein Data Bank (1). Kinase assay data
were obtained by querying against Pfizer screening database for
screens associated with any kinase target and tagged with IC50 or
Ki as the endpoint type.

3. Methods
3.1. Kinase Domain
and ATP Binding Site

The majority of protein kinases have a common fold (2) that

includes a small N-terminal lobe, which is mainly beta sheet but
contains a conserved -helix C, connected by a hinge region
to a larger more -helical C-terminal lobe (Fig. 14.1). In the
active conformation of the catalytic domain the large activation
loop, A-loop, orients away from the ATP binding site to allow
access to this pocket as well as the substrate docking region. This
conformation is often stabilized through phosphorylation of one
or more residues in the A-loop. Key residues are conserved in
the binding region to align the ATP for catalytic transfer. These
include a Glu in -helix C that forms an ionic bond with a con-

Design, Annotation, and Application of a Kinase-Targeted Library

281

Fig. 14.1. Ribbon structure (magenta) of the phosphorylase kinase crystal structure
2PHK (20) bound with ATP (green carbons, colored by atom type) and substrate peptide
(light blue ribbon). The N- and C-terminal lobes are highlighted; the hinge region is
shown in cyan, the -C helix in gray, and the A-loop in orange.

served Lys in -strand 3, which in turn coordinates with the ATP

- and -phosphate groups.
The ATP subunits (adenine, ribose, and tri-phosphate) bind
in a cleft formed between the N- and C-lobes. Interactions are
formed with the hinge backbone through hydrogen bonds: adenine N1 is an acceptor with a backbone NH and the 6-amino
group is a donor to a backbone CO (Fig. 14.2, D1, A). Analogous polar interactions can be targeted in inhibitor design. Furthermore, a proximal hinge carbonyl group that does not interact
with ATP (Fig. 14.2, D2) is positioned for hydrogen bond formation with a ligand. While residues in the ATP binding region
that participate in the catalytic process are well conserved across
the protein kinases, others vary and can be targeted to gain specificity. In addition, pockets and potential interaction points that
exist beyond those utilized by ATP can be targeted in inhibitor
design to enhance potency as well as selectivity (Fig. 14.2). These
regions have been designated as NE (Northeast or Selectivity
Pocket) near the Gatekeeper residue, SE (Southeast) delimited
from the Phosphate pocket by the Asp of the DFG (Asp -PheGly) segment at the beginning of the A-loop, W (West) extending toward solvent, and N (North), above the hinge segment.
Therefore, although protein kinases bind a common endogenous
ligand, the topology and the electrostatics of the ATP sites and
adjacent regions vary and provide the opportunity for specificity.

282

Xi and Lunney

Fig. 14.2. Protein kinase active site of JNK3 bound with an ATP analogue extracted
from an X-ray structure (1JNK) (21). Hinge region and DFG segment of the A-loop are
shown; the sugar and phosphate binding areas are circled. Highlighted schematically:
substrate binding region and the adjacent sites: NE, SE, W, and N. ATP hydrogen-binding
interactions with the hinge region are labeled D1 (Donor1) and A (Acceptor). Another
potential donor interaction (D2) with a hinge carbonyl is shown.

Unactivated states of the protein kinases have also been targeted in inhibitor design. In fact, five of the eight approved
small molecule drugs have been reported to target the unactivated form of the enzyme (3). Two characterized, unactivated
forms are the DFG-out (4, 5) and the C-Glu-out conformations
(6) (Fig. 14.3). In the DFG-out conformation, the DFG Phe, at
the beginning of the A-loop, reorients from a buried pocket near
the -helix C and can extend to the ATP binding region. This
in turn opens a pocket that can be accessed by inhibitor ligands.
The multikinase inhibitors imatinib, which targets Abl in chronic
myelogenous leukemia (CML) patients, and sorafenib, which targets angiogenesis in renal cell carcinoma, are approved drugs that
probe this pocket (5, 7).
In the C-Glu-out conformation the -helix C moves away
from the ATP site, such that the conserved Glu in the helix
does not form an ionic interaction with the conserved Lys in the
-strand 3. Here again a pocket is formed for inhibitor binding
(Fig. 14.3). The EGFR-targeted drug, lapatinib, takes advantage
of this pocket in binding the tyrosine kinase (8), as do the MEK
inhibitors identified by the Parke-Davis (PD) group (9). However, the latter compounds differ from lapatinib in that they do

Design, Annotation, and Application of a Kinase-Targeted Library

283

Fig. 14.3. An extension of Fig. 14.2 illustrating the DFG-out and C-Glu-out binding
regions. (Acknowledgment to J. F. Ohren for original mapping of the scheme).

not extend to the hinge region of the ATP binding site and are
non-competitive with ATP.
Another category of kinase inhibitors include compounds targeting the substrate binding region (Figs. 14.1 and 14.2), which
are also non-competitive with ATP (10). In addition, inhibitors
that have been designed to target the related phosphoinositol
kinases (PIK) can be characterized accordingly.
3.2. Compilation of
the Kinase-Targeted
Library

Our initial goal of the KTL is to compile a subset of compounds

from Pfizer corporate compound collections that can comprehensively represent the existing kinase chemotypes. Several targeted library design methods including docking, use of privileged
fragments and various ligand-based statistical models have been
reported in literature in recent years (1114). Here we applied a
substructure query-based method to identify kinase inhibitor-like
compounds in our corporate compound collection and implemented a series-based subsetting method to reduce the number down to a manageable collection of 70K compounds. The
substructure query-based approach has the advantage of being
intuitive to medicinal chemists, therefore, enables computational
scientists to effectively engage in design discussions with experimental scientists.
The overall design workflow for the KTL is schematically
depicted in Fig. 14.4. In order to capture the majority of the
known kinase chemotypes that represent the various binding
modes, we first compiled a set of substructure queries based on

284

Xi and Lunney

Fig. 14.4. Overall design workflow of the KTL.

the co-crystallized kinase ligands present in the corporate crystal

structure database (CSDB). Our CSDB contains internally solved
kinase structures and selected structures imported from the Protein Data Bank (1), Greater than 1000 kinase ligands from the
CSDB representing a variety of binding modes were first clustered
into 150 major chemical series (as defined by compounds sharing common core structures) using Wards clustering in combination with Daylight fingerprint and Tanimoto similarity metrics
(15). Approximately 130 substructure queries (labeled as CSDB
substructure queries) were then manually derived to capture the
core structures of these chemical series. Some series such as
staurosphorin-like structures were not included in the queries due
to their known promiscuity or lack of interest from project teams.
Then atoms in these substructure queries were replaced with
query atoms while preserving the aromaticity of rings and hydrogen bonding potentials of heteroatoms. The use of query atoms
allows the search to pick up additional series that resemble the
known kinase chemotypes but are potentially novel. Some examples of the CSDB substructure queries are shown in Table 14.1.
A test of these queries against the CSDB ligands correctly identified all known series of interests, while only 15% of compounds
were found when the queries were run versus a set of randomly
selected compounds. In order to capture additional kinase series

Design, Annotation, and Application of a Kinase-Targeted Library

285

Table 14.1
Examples of substructure queries
H

[C,N] N

N N
N

[C,N]
O

[C,N]

[C,N] [C,N]
[C,N]
A

[C,N]

[C,N,O,S]

N
2

[C,N,O,S]

N
N

that did not yet have a solved structure in the CSDB, we mined
our corporate screening database for any compounds with an
IC50 less than 1 M in either functional or enzymatic kinase
assays. A total of 34K compounds from 200 kinase assays were
found. For these kinase active compounds, we filtered out compounds already represented by the CSDB substructure queries
and then clustered the rest into structural series. From these, an
additional 100 substructure queries were derived from the maximal common substructure of each series.
In addition to these substructure queries, a small number
of SMARTS (16) queries were derived to capture the more
general hydrogen bond Donor-Acceptor-Donor (D-A-D) motif
that is frequently observed in core structures interacting with
kinases at the hinge regions. To make the search more specific, these SMARTS queries capture the presence of at least
two out of three hydrogen bond features at the D-A-D motif
(Table 14.2). Although single acceptor cores (for example, the
pyridine moiety of Gleevec binding to Abl (17)) are missed by
these SMARTS queries, the ones existing in the CSDB would
be captured by the CSDB substructure queries. We tested the
sensitivity and specificity of these SMARTS queries on CSDB
ligands and the randomly selected compound set. While the
SMARTS queries matched 75% of kinase ligands in CSDB, only
24% of the compounds in the random set were matched. Only
40% of the hits from the random set are also found in the hits
from CSDB substructure queries indicating that the SMARTS
query could potentially identify additional kinase inhibitor-like
compounds.
With these sets of substructures and SMARTS queries, we
searched our corporate compound collections and identified

286

Xi and Lunney

Table 14.2
Example of SMARTS queries
Motifs

Examples

SMARTS

AD

Pyrazole

[N, n;!H0][nX2;H0; R1]

AD

Amide in a ring

[N, n;!H0;R1] [$(O=[C, S])]

AD

Azaindole, adenosine,
amino-pyrimidine, etc

[$([N, n;!H0] [$([nX2;

H0; R1])]);!$(n1 n 1);!$(NC=N)]

AD

Pyrrolepyrmidine

[N, n;!H0] !: [nX2; H0]

AD

amine, carbonyl attached

to aromatic ring

[N; !H0]a[ ; R1][$(O=[C, S])]

[n; !H0] [$(O=[C, S])]

AD
Other cases for amides

5-member-aryl-amide

[N; H2]C(=O)-a 1aaaa1

Other cases for amides

6-member-aryl-amide

[N; H2]C(=O)-a 1aaaaa1

Other cases for amides

Biaryl urea

a 1aaaaa1-[N;!H0]C(=O)[N;!H0]a

840K hits from a total 2.8 M compounds. Then a set of druglikeness filters were applied to these compounds to reduce down
the total number of compounds to 720K. We then split these
720K compounds into two collections the library set (270K
compounds) amenable to combinatorial synthesis with library
synthesis protocols available and the medchem set (450K compounds) mostly made through traditional medicinal chemistry
synthesis. To further prioritize these hits, compounds in the
library sets were grouped by library protocol id and compounds
in the medchem set were clustered into structural series using
Wards clustering with Daylight fingerprints. Then four representative structures from each library or series were selected. A
panel of experienced kinase chemists were then asked to review
and prioritize the represented compound library or series based
on the physical properties, synthetic doability, as well as structural novelty. The review process focused on the chemical series as
opposed to individual compounds. Although each kinase expert
might unintentionally be biased toward a subset of chemical series
that he or she worked on in the past, by having multiple experts
in the review process, we were aiming to have a more unbiased
representation of the kinase chemical space collectively. In the
end, 310K compounds were retained after pooling the chosen
series together. We validated this selection retrospectively using
data from two recent kinase HTS projects (HGK and JNK1) that
screened the full compound collection in Pfizer and data from
Pfizer kinase selectivity panel screens (Table 14.3). For HGK,
82% of the Rule of 5 (Ro5) (18) compliant confirmed hits were
recovered in the 310K collection representing an 8-fold hit rate

Design, Annotation, and Application of a Kinase-Targeted Library

287

Table 14.3
Enrichment of kinase inhibitors in the initial compilation of
the KTL (310K compound collection) using substructures and
SMARTS queries prior to applying subsetting
HGK

JNK1

1,600,000

1,600,000

# confirmed actives

945

1455

# of actives passed filter

(MW<=550, clogP<=7,
RotB<12)

833

1376

# of actives passed filter and present

in 310k collection

685

920

% recovered

82

67

# Compounds Screened in HTS

enrichments (defined as hit rate in the 301K collection divided by

the overall hit rate in the HTS). Similarly for JNK1, 67% of confirmed hits were recovered representing a 6.7-fold hit rate enrichment. For the kinase selectivity panel hits (compounds with 50%
inhibition against any kinase on the panel at 10 M concentration), 80% of Ro5 compliant compounds were recovered. Overall these validations indicate reasonable combination of sensitivity
and specificity for this 310K compound collection.
To further reduce the total number of selected compounds
to a manageable subset for screening, a final subsetting step was
applied. We use a series-based subsetting method where compounds in each series were randomly sampled. The percentage
of compounds selected from each series depends on the size of
the series, ranging from 16% (1/6) for the largest clusters (clusters with >1000 compounds) to 100% (i.e., selecting all compounds) for the series with one or two compounds. This subsetting approach enabled us to remove overrepresentation of some
large chemical series without significantly affecting representation of small series. This final step reduced the total number of
compounds to 73K, an overall 4-fold reduction of the collection. To evaluate the effect of the subsetting step, we analyzed
the coverage of active series from the two actual kinase screens,
HGK and JNK1. As shown in Table 14.4, 688 active compounds
(representing 40 series) from the HGK screen and 730 active
compounds (representing 44 series) from the JNK1 screen were
found in the initial set of 310k compounds before the subsetting.
After the subsetting, as expected only 2530% of unique, active
compounds were retained. In contrast, 8590% of the series are
retained indicating the subsetting step has minimal impact at the
series level.

288

Xi and Lunney

Table 14.4
Validation of the subsetting algorithm retrospectively using
data from two HTS screening projects. Majority of the active
series were retained after the subsetting

3.3. Kinase-Targeted
Library Annotations

HGK

JNK1

# compounds (series) before subsetting

688 (40)

730 (44)

# compounds (series) found in after

subsetting

263 (37)

208 (37)

Individual inhibitors in the KTL with known kinase activity and

related structural information can be annotated based on the
types of binding interactions made with the protein. This process can greatly assist the project team in triaging the screening results and in identifying chemical series to prosecute. At the
highest level, inhibitors can be defined by the site of binding or
whether they are a known phosphoinositol kinase series: ATP site,
DFG-out, PD-MEK-type inhibitor site, substrate site, and PIK
inhibitors. The compounds can be further categorized by the key
binding template or series core.
The majority of inhibitors in the KTL bind in the ATP site
and the cores are defined by the rings or groups that form
hydrogen bonds with the hinge segment. For example, the P38
inhibitor, SB203580 (19), shown in Fig. 14.5 interacts with the
hinge region through the pyridine ring nitrogen as the acceptor. This template or core would be designated Pyri(mi)dine5-MemberHeterocycle_4 with one interaction with the hinge
region: A. A compound with an amino group at the 2-position
of the pyrimidine would have a second hydrogen bonding contact with the hinge and would be defined with a unique core,
Pyri(mi)dine-5-MemberHeterocycle_3, with two hinge interactions: D2, A. By analyzing bound cores in X-ray structures, the
substitution sites can be annotated according to which pocket(s)
they would probe and thus a specific compound could be so
labeled. The inhibitor example in Fig. 14.5 would be fully
annotated as ATP site; A; Pyri(mi)dine-5-MemberHeterocycle_4;
Phos, NE, indicating that the compound targets the ATP site, has
one interaction with the hinge (A), which was made by the pyridine core Pyri(mi)dine-5-MemberHeterocycle_4 and extends to
the phosphate and NE regions (Phos, NE). Examples of annotations for inhibitors that target the non-activated state of the protein are shown in Fig. 14.6: sorafenib (DFG-out binder) and the
MEK inhibitor, PD318088 (9). In these cases, subsite binding
would not be annotated.

Design, Annotation, and Application of a Kinase-Targeted Library

289

Fig. 14.5. Annotation for the ATP site inhibitor, SB203580. The core is Pyri(mi)dine-5-MemberHeterocycle_4, which is a
hydrogen bond acceptor with the hinge region. The compound probes the Phosphate (Phos) and NE sites.

COMPOUND

CORE
F
F

O
O

N
H

Cl

A
A

N
H

N
H

Sorafinib
OH
O

H
N

A
A

A
A

O
H
N

Br

Bayer_like_dfg

OH

H
N
I

PD318088

MEK_like_3

Fig. 14.6. Annotations for inhibitors that bind to unactivated kinase conformations.
Sorafenib binds to the DGF-out conformation and its core is defined as Bayer_like_dfg.
PD318088 binds to the C-Glu-out conformation and its template is MEK_like_3.

3.4. Performance of
KTL and Future Plans

Since the establishment of the KTL in Pfizer, many kinase projects

have screened the KTL collection. Among the first seven screens
completed, the hit rate (defined as retest confirmed hits divided
by total number of compounds screened in the KTL) ranges from
0.5 to 3%, several fold higher than a typical full HTS campaign
with confirmed hit rate in the range of 0.10.3%. For many of the
projects, KTL screens have led to interesting chemical series for
project teams to pursue hit-to-lead optimization.

290

Xi and Lunney

While this version of the KTL successfully captured known

kinase inhibitor series, the goal for our next-generation KTL
would be to apply de novo design methods to incorporate novel
chemotypes and to incorporate more nonclassical chemotypes
that bind to a protein kinase beyond the typical ATP pocket.

4. Notes
4.1. Advantage of
Using Substructure
Query-Based Method
for KTL Design

We presented here a workflow to design kinase-targeted library

using a substructure query-based method. Compared to various
de novo kinase-targeted library design methods, our approach has
the advantage of ensuring a comprehensive coverage of known
kinase inhibitor chemotypes.
In the past 10 years, there has been a large number of kinase
projects conducted in Pfizer for a range of therapeutic areas. By
deriving substructure queries from kinase ligands in our corporate crystal structure database as well as from active compounds
identified in in-house kinase assays, we were able to capture the
institutional knowledge on kinase inhibitors in the KTL design. It
is a common observation that inhibitors against different kinases
often share the same core structure. This is supported by the overall high similarity of active sites among kinases. The use of substructure searches in our design workflow guarantees coverage
of all compounds containing any of these common kinase cores
from our corporate compound collection . The subsequent seriesbased subsetting provides a sampling of each core series. Such
diverse sampling is critical as the KTL will be screened against
novel kinase targets.

4.2. Use of KTL Core

Annotation for HTS
Triage

The KTL core annotations have been integrated into several inhouse desktop applications for compound design and HTS hit
triage. The annotations provide key binding information for the
inhibitors and can be used to cluster compounds or to search for
inhibitors with a particular binding feature. Overall these insights
can help accelerate the HTS triage process and allow project teams
to advance chemical matter in a timely manner.

References
1. Berman, H. M., Westbrook, J., Feng, Z.,
Gilliland, G., Bhat, T. N., Weissig, H.,
Shindyalov, I. N., Bourne, P. E. (2000) The
Protein Data Bank. Nucleic Acids Res 28,
235242.
2. Johnson, L. N., Lowe, E. D., Noble, M. E.
M., Owen, D. J. (1998) The structural basis

for substrate recognition and control by protein kinases. FEBS Lett 430, 111.
3. Alton, G. R., Lunney, E. A. (2009)
Targeting the unactivated conformations
of protein kinases for small molecule
drug discovery. Expert Opin Drug
Discov 3, 595605.

Design, Annotation, and Application of a Kinase-Targeted Library

4. Pargellis, C., Tong, L., Churchill, L., Cirillo,
P. F., Gilmore, T., Graham, A. G., Grob, P.
M., Hickey, E. R., Moss, N., Pav, S., Regan,
J. (2002) Inhibition of p38 MAP kinase by
utilizing a novel allosteric binding site. Nat
Struct Biol 9, 268272.
5. Schindler, T., Bornmann, W., Pellicena, P.,
Miller, W. T., Clarckson, B., Kuriyan, J.
(2000) Structural mechanism for STI-571
inhibition of Abelson tyrosine kinase. Science
289, 19381942.
6. Levinson, N. M., Kuchment, O., Shen, K.,
Young, M. A., Koldobskiy, M., Karplus, M.,
Cole, P. A., Kuriyan, J. (2006) A Src-like
inactive conformation in the Abl tyrosine
kinase domain. PLoS Biol 4, 753767.
7. Reeves, D. J., Liu, C. Y. (2009) Treatment
of metastatic renal cell carcinoma. Cancer
Chemother Pharmacol 64, 1115.
8. Wood, E. R., Truesdale, A. T., McDonald,
O. B., Yuan, D., Hassell, A., Dickerson, S.
H., Ellis, B., Pennisi, C., Horne, E., Lackey,
K., Alligood, K. J., Rusnak, D. W., Gilmer, T.
M., Shewchuk, L. . (2004) A unique structure for epidermal growth factor receptor
bound to GW572016 (lapatinib): relationships among protein conformation, inhibitor
off-rate, and receptor activity in tumor cells.
Cancer Res. 64, 66526659.
9. Ohren, J. F., Chen, H., Pavlovsky, A., Whitehead, C., Zhang, E., Kuffa, P., Yan, C.,
McConnell, P., Spessard, C., Banotai, C.,
Mueller, W. T., Delaney, A., Omer, C.,
Sebolt-Leopold, J., Dudley, D. T., Leung,
I.K., Flamme, C., Warmus, J., Kaufman,
M., Barrett, S., Tecle, H., Hasemann, C.A.
(2004) Structures of human MAP kinase
kinase 1 (MEK1) and MEK2 describe novel
noncompetitive kinase inhibition. Nat Struct
Mol Biol 11, 11921197.
10. Vanderpool, D., Johnson, T. O., Chen, P.,
Bergqvist, S., Alton, G., Phonephaly, S., Rui,
E., Luo, C., Deng, Y. -L., Grant, S., Quenzer, T., Margosiak, S., Register, J., Brown, E.,
Ermolieff, J. (2009) Characterization of the
CHK1 allosteric inhibitor binding site. Biochemistry 48, 98239830.
11. Bradley, E. K., Miller, J. L., Saiah, E.,
Grootenhuis, P. D. (2003) Informative
library design as an efficient strategy to identify and optimize leads: application to cyclindependent kinase 2 antagonists. J Med Chem
46, 43604364.

291

12. Lowrie, J. F., Delisle, R. K., Hobbs, D. W.,

Diller, D. J. (2004) The different strategies for designing GPCR and kinase targeted libraries. Comb Chem High Throughput
Screen 7, 495510.
13. Prien, O. (2005) Target-family-oriented
focused libraries for kinasesconceptual
design aspects and commercial availability.
Chembiochem 6, 500505.
14. Stahura, F. L., Xue, L., Godden, J. W.,
Bajorath, J. (1999) Molecular scaffold-based
design and comparison of combinatorial
libraries focused on the ATP-binding site of
protein kinases. J Mol Graph Model 17, 19,
5152.
15. Daylight. Daylight Cheminformatics Toolkits. http://www.daylight.com.
16. Daylight. Daylight SMARTS Theory.
http://www.daylight.com/dayhtml/doc/
theory/theory.smarts.html.
17. Nagar, B., Bornmann, W. G., Pellicena, P.,
Schindler, T., Veach, D. R., Miller, W. T.,
Clarkson, B., Kuriyan, J. (2002) Crystal
structures of the kinase domain of c-Abl in
complex with the small molecule inhibitors
PD173955 and imatinib (STI-571). Cancer
Res 62, 42364243.
18. Lipinski, C. A., Lombardo, F., Dominy, B.
W., Feeney, P. J. (2001) Experimental and
computational approaches to estimate solubility and permeability in drug discovery and
development settings. Adv Drug Deliv Rev
46, 326.
19. Regan, J., Breitfelder, S., Cirillo, P., Gilmore,
T., Graham, A.G., Hickey, E., Klaus, B.,
Madwed, J., Moriak, M., Moss, N., Pargellis,
C., Pav, S., Proto, A., Swinamer, A., Tong,
L., Torcellini, C. (2002) Pyrazole urea-based
inhibitors of p38 MAP kinase: from lead
compound to clinical candidate. J Med Chem
45, 29943008.
20. Lowe, E. D., Noble, M. E., Skamnaki, V.
T., Oikonomakos, N. G., Owen, D. J., Johnson, L. N. (1997) The crystal structure of a
phosphorylase kinase peptide substrate complex: kinase substrate recognition. EMBO J
16, 66466658.
21. Xie, X., Gu, Y., Fox, T., Coll, J.T., Fleming,
M.A., Markland, W., Caron, P.R., Wilson,
K.P., Su, M.S. (1998) Crystal structure of
JNK3: a kinase implicated in neuronal apoptosis. Structure 6, 983991.

Section V
Library Design Tools

Chapter 15
PGVL Hub: An Integrated Desktop Tool for Medicinal
Chemists to Streamline Design and Synthesis of
Chemical Libraries and Singleton Compounds
Zhengwei Peng, Bo Yang, Sarathy Mattaparti, Thom Shulok,
Thomas Thacher, James Kong, Jaroslav Kostrowicki, Qiyue Hu,
James Na, Joe Zhongxiang Zhou, David Klatte, Bo Chao, Shogo
Ito, John Clark, Nunzio Sciammetta, Bob Coner, Chris Waller,
and Atsuo Kuki
Abstract
PGVL Hub is an integrated molecular design desktop tool that has been developed and globally deployed
throughout Pfizer discovery research units to streamline the design and synthesis of combinatorial
libraries and singleton compounds. This tool supports various workflows for design of singletons, combinatorial libraries, and Markush exemplification. It also leverages the proprietary PGVL virtual space
(which contains 1014 molecules spanned by experimentally derived synthesis protocols and suitable reactants) for lead idea generation, lead hopping, and library design. There had been an intense focus on ease
of use, good performance and robustness, and synergy with existing desktop tools such as ISIS/Draw and
SpotFire. In this chapter we describe the three-tier enterprise software architecture, key data structures
that enable a wide variety of design scenarios and workflows, major technical challenges encountered and
solved, and lessons learned during its development and deployment throughout its production cycles.
In addition, PGVL Hub represents an extendable and enabling platform to support future innovations
in library and singleton compound design while being a proven channel to deliver those innovations to
medicinal chemists on a global scale.
Key words: Drug discovery, chem-informatics, molecular design, combinatorial chemistry, combinatorial library, synthesis protocol, PGVL, reactant, product, enumeration, filtering integration,
workflow, streamline, desktop tool, software deployment.

J.Z. Zhou (ed.), Chemical Library Design, Methods in Molecular Biology 685,
DOI 10.1007/978-1-60761-931-4_15, Springer Science+Business Media, LLC 2011

295

296

Peng et al.

1. Introduction
Like many new technologies introduced to drug discovery, combinatorial library design and synthesis have matured during the
last 15 years (15). This is reflected by the fact that their practice
has been shifted significantly from experts, such as computational
and combinatorial chemists, to general medicinal chemists working in the pharmaceutical industry (627). In terms of library
design technology, we have seen a similar shift of focus from
methodology development to integration and deployment. With
the maturing of this technology, three new needs have emerged.
First, a significant amount of chemistry knowledge has accumulated in the form of detailed synthetic protocols. These protocols not only contain step-by-step synthesis instructions but
also specify what is considered to be suitable chemical reactants
compatible with the reaction conditions explored and validated
experimentally (termed the scope and limitation of the given
synthetic protocol). Systematic capturing, mining, and reuse of
such knowledge would bring tremendous value and competitive advantage to a pharmaceutical company in hit generation,
hit follow-up, and lead optimization. In a separate publication,
we have described Pfizers effort for the past 10 years in this
type of systematic knowledge capture and reuse which led to the
PGVL (Pfizer Global Virtual Library) chemistry knowledge base
(28). Second, unlike the expert community which is more interested in the latest design algorithms or flexibility in customizable
solutions, medicinal chemists working on drug discovery projects
place more emphasis on ease-of-use, start-to-finish workflow
management, and robust deployment as their top needs (2227).
For example, the ADEPT system was developed by scientists from
Galaxo and Daylight and deployed to the Glaxo discovery chemistry community as an integrated suite of Web tools on its corporate intranet for reactant selection, library enumeration, and
molecular property profiling and library design (25). The REALISIS system from Johnson & Johnson essentially accomplished
the same major goals of combinatorial library design, with more
focus on medicinal chemists and utilization of more advanced
software architecture and components (26). Finally, a modern
medicinal chemist uses many software packages to increase productivity. To gain acceptance by the chemist, a new software package needs to work with existing software packages synergistically
to reduce training effort and to enable richer and more powerful
workflows. Over the years, both commercial vendor offers and inhouse chem-informatics solutions have evolved to address these
three aforementioned needs and have achieved varying degrees of
success (627). Within Pfizer, we have witnessed a similar path

PGVL Hub

297

starting from an expert-only tool running on the UNIX platform

(23, 24), moving to a Web-based two-tier solution for a single
department or a selective group of users, and eventually reaching
a three-tier enterprise solution with a very capable and interactive
Java client on the scientists desktop computer.
In this chapter, we shall discuss the major requirements for
PGVL Hub to meet the three needs listed above, the software
architecture and key enabling data structures we have employed,
and the major technical challenges encountered and solved. Some
simple library design examples will be used throughout this report
to showcase the main features of PGVL Hub and the enabled
molecular design workflows. Finally, we will conclude by discussing the impact of PGVL Hub in terms of adoption and usage
by medicinal chemists over the past several years.

2. Major
Requirements
There are a few key design scenarios commonly requested by
medicinal chemists. They are described in the following sections.
2.1. Singleton Design

In this workflow, chemists want to draw or import a set of

molecules, profile their molecular properties, such as computed
ADME&T (absorption, distribution, metabolism, excretion, and
toxicity) properties, estimated activities against specific protein
targets based on existing SAR models, and make selections based
on the analysis of structural features and computed molecular
properties of those singleton molecules.

2.2. Markush
Exemplification

Similar to the singleton design scenario, but the input set

of molecules is created automatically based on a user-supplied
Markush drawing with R-groups attached to a molecular core
structure along with sets of explicit examples of those R-groups.
This design scenario is very popular with medicinal chemists in
terms of expressing their chemistry ideas or in the analysis of compounds commonly represented in patent literature.

2.3. Library Design

Using User Reaction
and Reactant Sets

This is the standard library design scenario which is supported

by most in-house and commercial vendor tools. The user-defined
reaction is usually a Markush reaction drawing commonly in the
format of MDL ISIS sketch or .rxn file (35). Chemists may
also supply reactant sets for each reaction component either by
loading pre-defined sets of molecules or by retrieving them via
searches into chemical reactant databases. Molecular property calculations and analysis on the reactants are performed, and selections are made based on these results. This is commonly known

298

Peng et al.

as the Reactant-Based Library Design, and the outcome is a

fully combinatorial library. Conversely, chemists can also generate explicit products via a product enumeration tool and perform
property calculations and analysis, then make decisions based on
product properties. If the final product selections are done at
the level of individual product molecules, then a cherry-picked
(sparse matrix) library will be the outcome. On the other hand,
if the final selections are still done at the reactant level, then the
outcome remains a fully combinatorial library, even when explicit
product properties are aggregated and used to guide and shape
the reactant selection. Both cherry-picking and fully combinatorial approaches are used by medicinal chemists. The full combinatorial design has an advantage in terms of library production efficiency in maximizing the number of products synthesized with
a given number of reactants to be handled during library production. However, the cherry-picked library design is more flexible and allows the designer to ensure that all products satisfy
user-imposed design criteria. With an increased level of synthesis
automation and popularity of small yet highly targeted libraries,
the cherry-picked library design is becoming the dominant mode
in pharmaceutical industry.
2.4. Library Design
Using PGVL
Reactions and
Pre-mined Suitable
Reactants

In this design scenario, users can take advantage of the captured chemistry knowledge inside the PGVL chemistry knowledge base by using various readily available components, ranging from high-quality product enumeration instructions to premined reactants lists suitable to pre-validated synthetic protocols,
in order to streamline library design and production (28). Since
a significant portion of modern corporate screening compound
collections originates from combinatorial chemistry, HTS hits
resulting from this subset synthesized via combinatorial chemistry
can be followed-up quickly and effectively with targeted libraries
using the same pre-validated synthetic protocols and pre-mined
compatible reactant sets. This is one of the unique objectives for
PGVL Hub.

2.5. Initiating Library

Design via Lead
Centric Mining
(LEAP)

Conceptually, the PGVL virtual product space is a collection of

molecules that can be made via one or more of the registered
synthetic protocols. While medicinal chemists routinely perform
similarity searches into vendor and corporate molecular databases
(approximately 107 in size) based on the given lead molecules
to gather information on synthetic routes, formulate SAR relationships, or generate lead-hopping hypotheses, it is also desirable to perform similarity searches into the PGVL product space
(approximately 1014 in size) (28). The only challenge is that the
technology managing the current corporate compound collections cannot be directly used for this purpose due to the enormous size of the PGVL product space. One innovative solution

PGVL Hub

299

has been developed via the lead centric mining tool, and its design
concepts and application will be described in detail by a separate
publication (29). The output of a LEAP search is a collection of
PGVL virtual compounds, each linked to an available combinatorial synthesis protocol and a combination of explicit reactants that
fully describes how this compound can be synthesized. Chemists
can then use these results to further evaluate each hit and launch
one or several targeted library designs to follow up on the LEAPderived hits.
2.6. Additional
Considerations

In addition to the above design scenarios which provide the

core requirements for the design and implementation of PGVL
Hub, there are several other requirements to be considered. As a
desktop tool for medicinal chemists, PGVL Hub should be very
graphical with good performance and robustness, easy to learn
and use, and easy to deploy and update with minimum administrative effort. More specifically, it needs to manage a fairly rich set
of hierarchical data structures (molecule, set of molecules, reaction, reactant, product, library, design, work session, etc.) and
ensure their capture either as a saved file on a desktop computer or
as a registration entry into downstream chemical information systems for library registration and synthesis. It should also have a set
of design features and a pluggable framework to easily integrate
new features to be developed in the future. Finally, PGVL Hub
has to integrate closely and seamlessly with existing tools to realize
synergies for a more streamlined and more powerful design experience. Among the software packages PGVL Hub interfaces with
are ISIS/Draw (30) for structure and reaction drawing; Microsoft
Excel (31) for list and table management; SpotFire (32) for data
visualization, analysis, and decision making; and various other 2D
and 3D molecular design tools.

3. PGVL Hub
3.1. Three-Tier
Enterprise
Architecture

We have chosen the J2EE (http://www.sun.com/java/) threetier enterprise software architecture for PGVL Hub (Fig. 15.1).
The client side is a J2SE GUI built as a Java Web Start (33)
deployable application which enables easy and automated deployment and update of both prototype and production versions globally from a single central server machine. Chemists can easily
install the client-side software component via a Web link available on an internal Web page. The Java Web Start technology
also provides automatic version check and upgrade of the clientside component each time PGVL Hub is launched by the user.
This mechanism was used to deploy PGVL Hub during all stages

300

Peng et al.

PGVL Hub
client-side GUI
component

Corporate Compound
Structure Service

PGVL Services

PGVL
Data

Product
Enumeration
Service

Other PGVL
Computing
Service

PGVL specific services

Corporate
Molecular
Property
Computing
Service

Library
Planning &
Production
systems

Corporate
Compound DBs

Services not specific to PGVL

Fig. 15.1. Three-tier J2EE architecture used by PGVL Hub. The client is a Java Swing GUI deployed via the Java Web
Start technology (33) to chemists desktop. The middle-tier is WebLogic J2EE server (51). PGVL data are hosted by
Oracle (52). And the product enumeration service and the PGVL computing service are hosted on SciTegic Pipeline Pilot
server (34). The corporate compound structure service provides support to PGVL Hub client for compound ID to structure
look-up, query searches into corporate databases, inventory checking, and compound duplicate checking. The corporate
molecular property computing service (41) returns computed properties chosen by user for a set of submitted molecules.

of the software life cycle with maximum ease and flexibility while
minimizing administrative cost.
The J2EE middle tier manages the interaction between clientside and server-side backend resources (e.g., various databases and
molecular property computing services) and enables the clientside and server-side software components to be updated independently. The server-side backend has access to various corporate
compound databases and the PGVL chemistry knowledge base
which delivers captured combinatorial reactions, synthetic protocols, and pre-mined and indexed suitable reactants for these protocols (28). It also contains computational services that perform
product enumeration and compute various molecular properties
at the request of the client-side GUI component. Since several
of the server-side services are not unique to PGVL Hub, we were
able to leverage emerging general purpose services via the serviceoriented architecture (SOA). Reciprocally, the PGVL server side
is also designed and deployed as a service, so software packages
other than the client-side component of PGVL Hub can tap into
the PGVL services as a valuable resource. The underlying database
engine for PGVL data is Oracle, and the underlying engine for
product enumeration is SciTegic Pipeline Pilot (34). More discussion on the PGVL services and captured chemistry knowledge
can be found in a separate publication (28).
3.2. Data Structures

A hierarchical data structure was designed to capture all data elements created during a design session, as shown in Fig. 15.2.
At the very basic level of Molecular Structure, a CTAB string

PGVL Hub

301

Collection
Design

Molecule

Reaction
CTAB
1

n Property

Name
Value

Reaction
component

* Reactant
Collection

Lib i
Libraries
1

Session

*
Library

Design

Generic Mol.
Collection

LCM result

Reaction

Reaction
component

P d
Product
Collection

Reactant
Collection

Fig. 15.2. Key data structures within PGVL hub for library design. The three data structures (Collection, Design, and
Session) within PGVL Hub are hierarchical; each is built on top of previous ones sequentially. A Collection contains
a set of molecules and their molecular properties. By default, one-to-one relationship between two data entities is
assumed, unless marked as either 1n which means for each parent entity, there are n (1, 2, 3, . . .) instances of child
entities associated. 1 means that for one parent entity, there could be any number (0, 1, 2, 3, . . .) of child entities
associated. MDL CTAB string (35) is used to represent the 2D structure of a molecule. Design encapsulates everything
about a library design work using one chemical reaction scheme. The Reaction is either an Rxn ID pointing to a preregistered PGVL reaction (28) or a MDL RDF string (35) for user-drawn reaction scheme for product enumeration, or even
a Markush drawing with R-groups hanging off the core (also in MDL CTAB string format, see Ref. (35)) for the Markush
exemplification workflow. The Rxn Component is created for each reaction component in reaction to be a place holder
for various reactant sets (or R-group sets for the Markush exemplification design workflow) for that reaction component
to enable reactant-based library design. The Libraries folder is a place to hold many explicit designed libraries. This
folder also enables chemists to compare and combine individual libraries (designed based on different design objectives
and protocols) within this folder into new ones. The Library concept contains all elements of a single explicit library
designed, and is self-contained with the Reaction information used to form the library. The single set of reactant sets
(one for each reaction component) now is fully synchronized with the product collection automatically during a library
design to ensure self-consistency. Finally Session contains everything a user has worked on during a single working
session after launching PGVL Hub and can be easily saved as an XML file for later use. Inside a Session folder, one
can have multiple Design folders, each with a unique reaction. The Generic Molecule Collection is intended for
simple singleton design workflow. It is also a convenience place to share sets of molecules between different Design
folders via Drag-and-Drop or Copy-and-Paste operation. The LEAP Result folder is a specialized Collection which
holds the Lead-Centric-Mining (LEAP) results as a collection of PGVL molecules with their combinatorial origin (reaction,
protocol, and reactant combination to make each individual molecule).

(a molecular format published by MDL) with 2D atomic coordinates is used for molecular representation (35). PGVL Hub
not only renders molecular structures graphically based on the
2D atomic coordinates inside the CTAB string but also allows
chemists to update those 2D atomic coordinates via ISIS/Draw
to improve the 2D layout of a molecule if desired. At the
Molecule level, users can add any number of properties (textual or numerical) to a molecule, with the format being a combination of name and value. We have found this data structure
to be adequate, although it could be extended to handle more

302

Peng et al.

complex requirements. Continuing with Fig. 15.2, we move

from Molecule to a Collection of molecules and then to the
concept of Design which is central to PGVL Hub. A Design
contains one chemical reaction, one or more reaction components, and one Libraries folder. Each reaction component can
contain any number of reactant collections to enable reactantbased library design. The Libraries folder contains individual
Library objects all sharing the same chemical reaction scheme,
and each Library represents an explicit library design which
contains one and only one set of reactant collections and one
product collection. Multiple library objects can be used to
explore different hypotheses or molecular design strategies. This
data structure enables designs of both fully combinatorial libraries
and cherry-picked libraries. Moreover, it enables list-logic operations (AND, OR, MINUS) between two reactant lists or two
libraries within a single Design. Finally a Session object can
contain any number of Designs plus other data elements of a
design session. A single Library, a Design, and a full design
Session can all be exported or imported back into PGVL Hub
as data objects in three XML file formats for storage, sharing, and
refinement.
We have taken several steps in working with the XML data format to ensure integrity of the data objects. To save a CTAB string
as a data element into an XML file, it is encoded via the BASE64
scheme and compressed into an XML-safe string using gzip
(3638). A user-drawn or imported reaction scheme is represented as an MDL RDF (35) string, which is likewise encoded
and compressed before saving into an XML file. The encoding and compression steps ensure robustness and compactness
of the XLM files written by PGVL Hub. One special circumstance involves the use of special characters in molecular properties. These characters may have special meanings to the XML
format and have the potential to corrupt the XML files. We have
solved this problem by identifying all instances of these characters
and encoding them into XML-safe strings before they are saved
into XML files. All encoding and compression steps are reversed
when importing an XML file back into PGVL Hub.
3.3. Performance

Library design, unlike singleton design, usually involves collections of many molecules. It is fairly common to encounter design
cases starting with hundreds to thousands of reactants for each
reaction component which could potentially lead to huge enumerated libraries. Such a large collection of molecules poses three
performance challenges to PGVL Hub. First, how do we move
them quickly between the client and server components of PGVL
Hub? Second, how do we store and manage them once they arrive
at a client machine? Lastly, how do we enhance the user experience when performing molecular property calculation on such

PGVL Hub

303

large collections? Some molecular properties such as Rule-of-Five

(39) can be computed quickly, but others like LogD (40) and
docking scores can be computationally expensive and time consuming.
We initially encountered the network throughput bottleneck while moving a large number of molecules between the
PGVL server and the client components. This problem was
addressed by deploying multiple PGVL servers, one for each
Pfizer research site. By having a PGVL server within close network
proximity of the intended users, the throughput bottleneck was
greatly reduced. Furthermore, having multiple PGVL servers provided redundancy which greatly improved the availability of the
PGVL service. As network throughput improved, we eventually
returned to the original deployment model with a single global
server cluster running multiple instances of WebLogic Application Server (51). Nevertheless, the multiple-servers-at-multiplesites approach has provided great value throughout the development, deployment, and user support of PGVL Hub.
At the early stage of the PGVL project, molecules were stored
in the RAM of the client machine so that users could browse
20100 molecular structures per viewing page quickly and efficiently. This speedy structure browsing capability of PGVL Hub
was warmly received by medicinal chemists. However, when
PGVL Hub encountered large molecule collections numbering
in the tens of thousands or more (10 K molecules required
approximately 1 GB of RAM for storage), its performance would
drop significantly. This situation can be further exacerbated when
other memory-intensive applications are running concurrently on
the same client machine. To overcome this problem, we have
implemented a hard disk-based cache system which operates with
great efficiency and performance. All molecules either loaded
into PGVL Hub or generated during a design session (e.g., an
enumerated virtual library) are indexed and saved onto the disk
file system which is much larger than the physical RAM space
on a typical desktop computer. PGVL Hub would then use
a fast lookup scheme to load only those molecules to be displayed on the screen. This solution enables chemists to work on
extremely large library designs without experiencing any performance degradation. With a much reduced RAM footprint, PGVL
Hub also becomes a good desktop citizen among the many
software packages concurrently running on a typical machine.
The Pfizer molecular property computing service contains
many properties related to ADME&T endpoints as well as
project-specific SAR models (41). These computed molecular
properties are often critical for medicinal chemists to ensure that
the designed singleton and/or library compounds have good
drug-like properties and to explore new SAR hypotheses against
specific protein targets. While some molecular properties can be

304

Peng et al.

calculated quickly, others can be much more time consuming.

Adding more computing power to speed up molecular property calculations was utilized as part of the solution to reduce
turn-around time. More importantly, users should be allowed to
continue working with PGVL Hub after submitting jobs for
molecular property calculations. This requires an asynchronous
computing model which is available from the Pfizer molecular
property calculation service. To further improve performance,
PGVL Hub also utilizes a divide and conquer approach which
breaks up a large calculation job into smaller ones, each with
500 molecules, and submit each smaller job to the molecular property calculation service. The service then sends a job ID
for each job submission back to PGVL Hub, which it uses to
check with the central computing service periodically for computed results. Whenever the job with a given job ID is finished,
PGVL Hub then downloads the computed results and merges
them with their corresponding molecules automatically. During
this time, chemists can continue working with a design session
without having to wait for the results to return. If a design session is ended before the computed results are returned, PGVL
Hub would save the job IDs so that the computed results can
be retrieved next time the design session is restored. This asynchronous job submission mode was proven to be absolutely essential for optimal user experience while working on designs with
large molecule collections.
3.4. Workflow

One of the strategic goals of PGVL Hub was to streamline the

most common workflows in combinatorial library design. In general, the design workflow begins with a chemical reaction that is
either downloaded from the PGVL chemistry knowledge base,
imported from a pre-drawn reaction file (in the MDL ISIS/Draw
.rxn format), or the chemist can create a new reaction on the fly
using the imbedded reaction drawing page (Fig. 15.3). Once the
reaction is defined, PGVL Hub then creates a Design with folders for reactant lists of each reaction component. The reactant lists
are defined by the chemists, either by importing pre-defined lists
or by using the pre-registered reactant lists available within PGVL
Hub. For list importation, PGVL supports the SDF file format as
well as compound IDs from in-house corporate IDs or MFCD
numbers. If the reaction selected comes from one of the several
hundred pre-registered reactions in PGVL, users can choose the
reactant lists associated with each synthetic protocol for that given
reaction, with the advantage that each reactant list has been prefiltered against the scope and limitations of the associated protocol to ensure the best synthetic success of the designed library.
For virtual compounds, chemists also have the option to sketch
them on-the-fly using MDL ISIS/Draw and use them in PGVL
Hub.

PGVL Hub

305

a)

d)

b)
O

z3
z1
N

z1

O
N

z1

S
O

z3

O
O

N
R1

R3

c)

N
R2

z2
z2
z2

Fig. 15.3. Library design workflow: initiate design by defining reaction and gathering reactant sets: (a) Common workflows supported by PGVL Hub. Here the Monomers inside the figure means reactant sets. Clicking on a bottom that
represents a step of the workflow leads to more detailed GUI panels for users to further specify what need to be done.
(b) One can define a reaction either by search, browse, and select a pre-defined PGVL reaction or draw a new reaction
on the fly using ISIS/Draw. (c) A Markush core capped by R-groups plus actual R-group fragment sets are used for the
Markush exemplification workflow in place of reaction and reactant sets. (d) There are many ways to get reactant lists
into PGVL Hub (see main text). Here one GUI component of ChemSelect (AQB) (50) was shown. Names of chemistry
functional groups are listed in the menu for user to specify what functional group should and should not appear in the
desired reactants. The query will be searched against in-house inventory systems and search results will be loaded into
PGVL Hub as reactant lists.

Once the reactant lists are loaded, chemists can filter them
through various molecular property calculations, substructure
search/mapping, and similarity score against one or a set of lead
structures. The substructure mapping and similarity scores against
a collection of molecules are performed by a server-side SciTegic
Pipeline Pilot component on the fly, eliminating the need for preregistration. For library synthesis considerations, PGVL Hub also
enables calculation for the amount of each reactant required for
library synthesis and determines its availability from the corporate
reagent inventory system. These calculations and filtering are in
place to enable the designer to derive lists of desirable reactant for
library synthesis.
Having the desired reactant lists, the chemist can now create a virtual library by enumerating the product structures in a
fully combinatorial manner. Enumeration instructions are prevalidated for all PGVL registered reactions; for a user-specified
reaction, PGVL Hub enables enumeration via the Markush
representation of the reaction scheme. Once the products are

306

Peng et al.

enumerated, further filtering can be done via molecular property

calculations, substructure mapping, similarity scoring, and structure matching with the Pfizer corporate databases for possible
duplicates (product molecules already exist in the corporate collection). The enumerated products can be exported out of PGVL
Hub and processed by additional 3D molecular design tools such
as docking and scoring based on 3D pharmacophore models or
target protein structures for further design refinement. Users can
also prioritize and select desired product subsets for further consideration (cherry-picked library), perform list logic operations
(AND, OR, MINUS) between two reactant sets or two libraries,
which allow comparisons between and combinations of different
design hypotheses.
When the chemist is satisfied with the library designs, various output methods can be used to export the results. Individual
reactant lists can be exported as ID lists or as SDF files, while
the final products can be exported as SDF files with corresponding combinations of reactant IDs. Or, the entire design or session file containing the library information can be exported in the
XML format. PGVL Hub even provides the option to upload a
design directly into a downstream chemistry informatics system
to initiate library registration and synthesis. The XML format for
library design has also been used as a preferred data exchange format between Pfizer and external chemistry outsource partners for
library production. The Grid View in Fig. 15.4 shows a screen
shot of this basic workflow.
Using the same set of software capabilities, PGVL Hub also
supports other design scenarios such as Singleton Design that is
commonly practiced by medicinal chemists, Markush Exemplification where the reaction scheme becomes a Markush core structure, and reactant sets become R-group sets, and Lead Centric
Mining (29).
3.5. Desktop Synergy

The desktop computer of a modern medicinal chemist usually

contains software packages designed to increase productivity. Two
such examples are Microsoft Excel (31) for data analysis and SpotFire (32) for data visualization, along with other 2D or 3D molecular design tools. In designing PGVL Hub, we recognized the
importance of integration with some essential software packages
to realize synergies as well as making PGVL Hub a more powerful design tool. Toward this goal, we have achieved seamless
integration between PGVL Hub and MDL ISIS/Draw (30) and
SpotFire (32). These two software packages are critical for structure drawing and SAR viewing, two of the most common practices by a medicinal chemist. From within PGVL Hub, chemists
can launch ISIS/Draw and sketch a new molecule or reaction and
then return the newly drawn molecule or reaction back to PGVL
Hub via a single click (Fig. 15.5). This is the same synergistic

PGVL Hub

307

Fig. 15.4. Library design workflow: analysis and filtering of molecular collections: (a) A grid view and a table view panels
are showed here displaying a collections of product and reactant molecules. Both views can be sorted by properties,
and visibility and column location for each molecular property can be customized by user. Both views allow user to
browse through large collections of molecules very efficiently and mark molecules of interest with various color pens.
Also notice that the top workflow bar keeps track of design progress and highlights workflow steps already initiated.
The first example in Grid View also offers a good view of all content within a single Design. (b) User finds, selects,
and submits computation jobs to the Pfizer molecular property calculation service (41). This service contains many in
silico models for ADME&T and even project-specific SAR activity models. (c) User makes all filtering decisions within this
Decision Maker panel. All molecular properties available can be used by Decision Maker. The user hand-marking
using color pens or selection returned from a SpotFire session are used as binary input. Numerical data are filtering
using range slider bars. The color histograms display the property distributions of the starting set (in green color) and the
current set (in blue color) before a filtering action is fully committed. This gives user an immediate and dynamic feedback
on possible consequences of the current filtering setting. Textual properties can also be used for decision making (not
shown here) via string searches such as Exact Match, contains, Starts with, and Ends with in comparison with
user-specific string. Also one can use SpotFire for data visualization and selection.

behavior between ISIS/Base and ISIS/Draw, already familiar to

most medicinal chemists. The integration between PGVL Hub
and SpotFire offers an even richer set of behaviors (Fig. 15.6).
To use the SpotFire viewer within PGVL Hub, simply select a
set of molecules (either reactants or products), then launch SpotFire directly within PGVL Hub. SpotFire would then display the
molecular properties of the molecule collection, such as compound ID and various imported and/or computed ADME&T
properties. Mousing over a data point displayed within the SpotFire window will display the associated molecular structure inside
a structure viewing panel in PGVL Hub. Selections made within
the SpotFire window will be automatically passed back into the

308

Peng et al.

Fig. 15.5. Integration between PGVL Hub and ISIS/Draw. By clicking on any structural box in PGVL Hub window, ISIS/Draw
will appear with any molecular structure inside the PGVL Hub structural box if exists. Once the user is done creating or
polishing a structure drawing inside ISIS/Draw, a single click on ISIS/Draw will transfer the new structure drawing back
to the PGVL Hub structural box, and the ISIS/Draw window will disappear automatically.

a)

b)
Fig. 15.6. Seamless integration between PGVL Hub and SpotFire. User can launch SpotFire within PGVL Hub to visualize
the molecular properties associated with a molecule collection. Any selection done within SpotFire is dynamically passed
back to PGVL Hub as marking on individual molecules. And user then can use the Decision Maker within PGVL Hub to
make selections based on the SpotFire marking.

PGVL Hub

309

PGVL Hub window. Such integration allows seamless behavior

between the two tools so that the user experience feels like a single piece of software. Additionally, PGVL Hub also has a one-way
connection with Microsoft Excel and several in-house molecular design tools which can be launched within PGVL Hub while
passing appropriate data to those applications (e.g., textual and
numerical data for Excel, and SDF files for other 2D and 3D tools
capable of reading such file format. Details on the MDL SDF file
format can be found in reference (35)). By realizing these synergies, chemists can easily access and combine the best features
from several applications to realize an even more powerful design
experience. Furthermore, by not having to manually shuttle data
between several applications, chemists can enjoy a more streamlined workflow. From a software development point of view, these
synergies also allow PGVL Hub to leverage the best features from
other software packages without re-inventing the wheel.
3.6. Design Features

Over a period of several years, we had worked closely with user

communities and implemented many singleton and library design
features. The most basic but heavily utilized is the Browse and
Mark capability (see Fig. 15.4), where a chemist can display many
molecular structures and their properties in grid or table view,
page through them quickly and efficiently, and mark molecules
of interest using different color makers for later decision making. Both the grid and the table views can be sorted based on
molecular IDs or properties. We also implemented a SpotFirelike Decision Maker component (Fig. 15.4), where filtering can
be performed on user-selected molecules as well as on numerical and textural properties. For numerical properties, a range
filter is implemented where user can enter numerical values as
well as using a sliding bar to perform the filtering process. For
textual properties, we have implemented Exact Match, Starts
with, Ends with, and Contains to allow very flexible filtering operations. Color marking created by manual Browser and
Mark steps as well as selection marking created from a SpotFire session can also be used by the decision maker component
for molecule filtering. The histograms inside the decision maker
provide chemists with immediate feedback on the consequence
of the filtering action before committing to a filtering operation.
For product collections, filtering on the products will result in
a cherry-picked library (sparse matrix, not fully combinatorial).
This design approach maximizes flexibility to ensure all product
molecules are compliant with desired property ranges, although
the consequence is lower production efficiency since not all products within the combinatorial matrix are made. An alternative
approach is to use aggregated product properties to shape and
optimize the reactant selection so that the library outcome is still
fully combinatorial while maintaining a desired profile of product

310

Peng et al.

molecular properties. Previous publications (27) have provided

several possible solutions to the design objective described above.
We have incorporated within PGVL Hub a combinatorial shaping
design feature that is simple, graphical, and intuitive for medicinal
chemists. For each product molecule in a given library, we calculate a user-customizable Pass/Fail score. Then for each reactant
molecule, the number of Failed products associated with this
reactant is calculated and used as an aggregated property called
Failed Score, which can then be used to sort the reactant list
and reorder the product matrix according to this score. Removing the reactants with the highest Failed Scores would make
the most impact to improve the quality of the remaining library in
terms of molecular property compliance. As shown in Fig. 15.7,
users can simply use the slider bars within the combi-shaping
panel to remove the reactants with the highest failed score, and

a)
Reactants are sorted based
on # of failed products
they are associated with.
Drag the slider(s) to
remove reactants with
highest # of failed
products

Status on
library

Effect of removing
reactants on the remaining
library are shown by the
display

b)
Fig. 15.7. Interactive combinatorial library shaping (a) A user-defined Pass/Fail score (such as Rule-of-Five) can be
constructed and computed for product molecules based on existing molecular properties. This type of Pass/Fail scores
can be used for combinatorial library shaping. (b) After user selects a library for combinatorial shaping, PGVL Hub allows
user to pick the appropriate Pass/Fail scores as input, then sort reactants for each reaction component from low to high
based on number of Failed product molecules a reactant molecule is associated with and plot the library status visually.
User then uses the slider bars (one for each reaction component) to remove the worst reactant(s) and get an immediate
feedback from the status report. The green curves are static based on the input reactant sets; the blue curves are
updated dynamically to indicate the possible outcome. User can explore various strategies in reducing worst offenders
in reactant sets to reduce the number of Failed products while still maintaining a fully combinatorial library of good
size and production efficiency (number of products to be synthesized vs. number of reactants to be handled). User then
commits the shaping by creating a new library.

PGVL Hub

311

PGVL Hub provides immediate feedback of the action in terms

of updated property ranges and library size. This instant graphical
feedback enables chemists to review the results before committing
to the library shaping steps, and have the option to either updating the existing library or creating a new library with modified
product properties and updated reactant sets.
One other useful feature of PGVL Hub is substructure
searching and mapping enabled by SciTegic Pipeline Pilot (34).
This feature allows chemists to determine what substructures are
mapped to a target molecule and where as well as how many
times each substructure is mapped. The substructure query can
be entered by a user via a pop-up panel or run as a set of
pre-built substructure queries as a part of a molecular property calculation service. An example is shown in Fig. 15.8,
where substructure fragments are compiled at the corporate level
to flag undesirable structural elements (41b) so they can be
flagged and avoided in molecular designs. Such practice provides
another example where PGVL Hub enables the reuse of valuable
knowledge collectively captured within the Pfizer drug discovery
community.

Fig. 15.8. Substructure mapping, highlighting, and drill-down. Based on on-the-fly substructure query and mapping
capability within SciTegic Pipeline Pilot, PGVL Hub allows user to perform substructure queries into a set of target
molecules. In the example shown, a set of substructure queries globally collected and validated as undesirable substructure features to be avoided are mapped into target molecules (41b).

312

Peng et al.

4. Remarks
4.1. Deployment,
Usage, and Impact

Since PGVL Hub client side is Java Web Start deployable and
Pfizer-supported desktop computers all have the Java Web Start
utility installed, the installation of PGVL Hub is easily done by
users themselves directly via a Pfizer internal web site. This web
site also contains other resources to facilitate the distribution,
training, and support of PGVL Hub. Training materials include
the users guide, presentation slides, animated tutorials, and scientific literatures pertaining to singleton and library designs. Support is provided by a global support team, a network of local
power users at various Pfizer research sites, and a global steering
committee comprising PGVL champions.
The initial deployment began around 2003 targeting a small
yet highly motivated community of approximately 100 beta
testers comprising mainly medicinal chemists and computational
chemists. Based on feedback from these users and results of two
formal software usability studies conducted at various Pfizer sites
involving medicinal chemists with some or no prior exposure to
PGVL Hub, we were able to add further enhancements to make
the software even better and more user-friendly. A full deployment was initiated around 2005 to the global discovery chemistry
community of over 2000 potential users at that time. The adoption and usage of PGVL Hub are tracked, and reports of usage
statistics are provided through an internal Web page; one graph
within a typical report is shown in Fig. 15.9. For the past few
years, PGVL Hub usage has been steady at 60100 launches per
day. Of the approximately 1000 registered users, 30% are considered to be experienced users based on their numbers of logins
during the last 12 months. The usage tracking tool also identified expert users as well as novices which helped the support
team recognize opportunities for training and allocation of support efforts. The feedback from PGVL local champions and the
usage data collected so far illuminate some aspects of PGVL Hub
success, including penetration and adoption by the intended user
community (>50%), frequency of usage, and level of expertise
reached by expert users (about 1 in 6 is a frequent user). However,
the true impact of PGVL Hub should be measured by increased
quality of singletons and library compounds chemists designed to
move drug discovery projects forward and the productivity gained
due to PGVL Hub usage (5356). Unfortunately we do not have
a systematic way of tracking these factors directly other than feedback from medicinal chemists and their research leadership. Ultimately, the success of PGVL Hub to enable smarter designs and
higher productivity should be better assessed by successful drug
candidates coming out of drug discovery projects.

PGVL Hub

313

Number of user login daily

07/01/04

02/16/09

From 07/01/2004 to 02/16/2009

Total # of user login: 104,866
Total unique users : 1,586

Fig. 15.9. Tracking usage of PGVL Hub. Each time a user logins into PGVL Hub, information about user ID and time stamp is recorded into a tracking database. A usage report
can be generated via a Web reporting tool.

4.2. Comparison with

Published Integrated
Library Design Tools
Used in
Pharmaceutical
Industry

It is likely that every pharmaceutical company would have an integrated library design tool in place as part of the strategy to incorporate the combinatory library approach into its drug discovery
process. Since most of them were not published, we could only
make comparisons among ADEPT, REALISIS, and PGVL Hub
(see Table 15.1). Here we shall leave the details for readers to
explore further while making a general statement that all three are
designed to address essentially the same set of major questions in
library design. They only differ in level of software engineering,
GUI capability and intuitiveness, and scope of feature coverage.

4.3. A Proven
Platform for Future
Enhancement and
Innovation

PGVL Hub has been well entrenched as one of the key desktop
molecular design tools used by Pfizer medicinal chemists. Its solid
three-tier enterprise architecture and powerful client-side component easily deployed by Java Web Start provide a very attractive
platform with a proven track record for future enhancement and
innovations in singleton and library design. There are many possibilities for further enhancement based on user requests as well
as attractive methodologies and algorithms already published in
the literature (627). Here we would like to list a few, with some
already being prototyped.

4.3.1. Multiple Property

Optimization

This is a well-known area of research with significant practical

implications for molecule design (11, 12, 14, 18, 19, 4246).

On-the-fly searching using

SMARTS either pre-defined or
user provided

SMIRKS either pre-defined or

user entered/chemical transformation based

Limited set mainly available in

daylight

Yes

Yes

No

Reaction encoding/
enumeration method

Product property prediction

Support for fully combinatorial shaping using product

properties

Cherry-picking library design

based on product properties

Support for Markush exemplification

ADEPT (25) (Glaxo & Daylight,

1999)

Source for reactant

Feature or capability

No

Yes

No

Limited set

ISIS reaction scheme

either pre-defined or
user entered/chemical
transformation based

On-the-fly searching using

ISIS query

REALISIS (26) (J & J, 2004)

Table 15.1
Comparison of three integrated library design tools from the pharmaceutical industry

Yes

Yes

Very large collection, including many vendorsupplied and internally developed in silico
models for ADME&T end points and target SAR
Yes

ISIS reaction scheme entered by user or predefined reaction object for 500 PGVL
registered combinatorial reactions/reactant
clipping and assembly of a Markush core
and R-groups

One-the-fly searching using ISIS queries

either pre-defined or user provided; Load
lists of compound IDs; Load pre-mined lists
of reactants suitable for registered PGVL
reaction reactions; Load SDF files; User
drawn

PGVL Hub (Pfizer, fully deployed to 1200

users in 2005)

314
Peng et al.

Java GUI, two-tier, access to various DBs and SciTegic Pipeline

Pilot via SOAP and ODBC

Limited. SDF file of various reactants and products are created

and stored in user directory
during design

Mainly for medicinal chemists

(70%)

Web GUI with CGI scripts

on the server side for integration. two-tier

Software architecture

Computational and medicinal chemists

No

Session file for persistence

and design sharing

Numerical property filters and

histograms

Targeted users

Numerical property filters

and histograms

Integrated decision
making

Limited

Multi-thread during reactant

mining against multiple
reactant databases

Limited

Structure and property

viewing, sorting, and
exporting

N/A since there is no predefined virtual library space

Software performance
enhancements

N/A since there is no

pre-defined virtual library
space

Similarity search into predefined virtual library

space for idea generation
and lead hopping

Table 15.1
(continued)

Mainly for the medicinal chemists, but it has also

been used extensively by computational chemists

Multi-thread, batch data fetching or batch job

submission, asynchronous mode for ADME/Tox
property prediction. Client-side disk-based fast
cache for molecular structures

Java GUI deployed via Web Start for easy deployment and update, J2EE three-tier. Access to
various DBs, SciTegic Pipeline Pilot, and other
molecular property predicting services

Yes, intermediate results of a design session can all

be saved into a single XML-based session file for
later use or share with collaborators

Powerful SpotFire-like filter for both numerical and

textural properties; also integrated with SpotFire
directly for data exploration and decision making

Very general and powerful. Both table- and gridviewers are fully configurable by user. User can
also select and re-order property columns for both
display and export

Yes, virtual library space of 1014 molecules (PGVL)

spanned by 500 combinatorial reactions are
searchable directly through PGVL Hub (see Ref.
(29) for details)

PGVL Hub
315

316

Peng et al.

The key challenge is to make it intuitive so that it is easy to use

and easy to interpret the results. Significant progress in this area
has been made and reported in the literature (47).
Genetic-Algorithm (GA)-Driven Library Design and Lead
Centric Mining: An abundance of literature on singleton and
library design utilizing genetic algorithms already exists (6, 8,
1619). The usage of GA is based on a Goodness score function, for example, a combination of similarity to a known set
of lead compounds, various ADME&T molecular properties, or
computed activity scores based on specific project SAR models
against a protein target. The GA methodology will act as an agent,
explore automatically the vast compound space either defined by
the user or by PGVL, and return a set of molecules with enhanced
or optimized Goodness scores. In essence it is a virtual screening methodology against a virtual chemical space not fully enumerated. In the context of similarity search, it is a generalized
form of our current Lead-Centric-Mining methodology (29).
4.3.2. Structure-Based
Library Design (SBLD)

For discovery projects where target protein structures are available, a structure-based library design strategy would be highly
desirable. In the current version of PGVL Hub, 2D library
molecules are exported out of PGVL Hub and imported to
SBDD-enabled molecular design tools; it would be desirable to
have a more integrated workflow. The challenge is how to deal
with the one-to-many relationship between a 2D molecule inside
PGVL Hub and its many potential 3D conformers in complex
with the target protein binding sites. By utilizing an aggregation
step to reduce the one-to-many relationship into a one-to-one
relationship (e.g., keeping just the best docking score of the best
3D conformation), the SBDD aspect is reduced to the best docking score, and a very simple molecular property is returned to
PGVL Hub for decision making. This approach is very simple and
intuitive, but is best for smaller libraries due to the computationintensive nature of docking and scoring. One way to extend the
range of SBLD coverage to a much larger virtual space is through
the usage of Basis Products (21). The detail of this SBLD effort
using Basis Products has been described in a separate publication
(48).
All of the more advanced library design capabilities described
above can be integrated into the PGVL Hub platform in an intuitive way and utilized by medicinal chemists routinely to impact
progression of drug discovery projects.

5. Conclusions
PGVL Hub is an integrated desktop tool which has been developed and globally deployed throughout Pfizer discovery research

PGVL Hub

317

units for singleton and library design and synthesis. It has a highly
intuitive and interactive GUI, an excellent performance profile,
and is easy to install and update. For the past several years it
has been routinely accessed by hundreds of medicinal chemists
and other scientists for compound and library design work. This
tool provides direct access to Pfizers proprietary PGVL chemistry knowledge base to enable fast HTS hit follow-up and lead
optimization. It offers a very rich and intuitive set of design capabilities and covers a wide range of workflows commonly used
by medicinal chemists. PGVL Hub also has the advantage of
being integrated with other desktop tools, such as ISIS/Draw,
Microsoft Excel, SpotFire, and other 2D and/or 3D molecular
design tools, and it leverages the best features of those tools to
provide synergies and an integrated workflow. Its three-tier J2EE
enterprise architecture and a powerful GUI provide a proven platform and delivery mechanism for future enhancements. Beyond
its usage statistics, the true measure of PGVL Hubs positive
impact in design quality and productivity should be an increase of
attractive chemical leads emerging from drug discovery projects.

Acknowledgments
Over the years, the PGVL development team has received strong
support and help from Pfizer research management, PGVL site
champions and steering committee, user communities in medicinal chemistry and computational chemistry, research informatics, and sister software development projects. We would like to
express our deepest gratitude and apologize for not being able to
list all their names explicitly here.

References
1. Hogan, J. C. Jr. (1997) Combinatorial chemistry in drug discovery. Nat Biotechnol 15,
328330.
2. Hall, S. E. (1997) The future of combinatorial chemistry as drug discovery paradigm.
Pharm Res 14(9), 11041105.
3. Salemme, F. R., Spurlino, J., Bone, R. (1997)
serendipity meets precision: the integration
of structure-based drug design and combinatorial chemistry for efficient drug discovery.
Structure 5, 319324.
4. Floyd, C. D., Leblanc, C., Whittaker, M.
(1999) Combinatorial chemistry as a tool
for drug discovery. Prog Med Chem 36,
91163.

5. Beeley, N., Berger, A. (2000) A revolution in

drug discovery. Combinatorial chemistry still
needs logic to drive science forward. BMJ [Br
Med J] 321(7261), 581582.
6. Singh, J., Ator, M. A., Jaeger, E. P., Allen, M.
P., Whipple, D. A., Soloweiij, J. E., Chowdhary, S., Treasurywala, A. M. (1996) Application of genetic algorithm to combinatorial
synthesis: a computational approach to lead
identification and lead optimization. J Am
Chem Soc 118, 16691676.
7. Blaney, J. M., Martin, E. J. (1997) Computational approaches for combinatorial library
design and molecular diversity analysis. Curr
Opin Chem Biol 1, 5459.

318

Peng et al.

8. Brown, R. D., Martin, Y. C. (1997)

Design combinatorial library mixtures using
a genetic algorithm. J Med Chem 40,
23042313.
9. Agrafiotis, D. K., Myslik, J. C., Salemme
F. R. (1998) Advances in diversity profiling
and combinatorial series design. Mol Diversity 4, 122.
10. Bures, M. G., Martin, Y. C. (1998) computational methods in molecular diversity and
combinatorial chemistry. Curr Opin Chem
Biol 2, 376380.
11. Zheng, W., Cho, S. J., Waller, C. L, Tropsha, A. (1999) Rational combinatorial library
design. 3. Simulated annealing guided evaluation (SAGE) of molecular diversity: a
novel computational tool for universal library
design and database mining. J Chem Inf
Comput Sci 39, 738746.
12. Gillet, V. J., Willett, P., Bradshaw, J., Green,
D. V. S. (1999) Selecting combinatorial
libraries to optimize diversity and physical properties. J Chem Inf Comput Sci 39,
169177.
13. Spellmeyer, D. C., Grootenhuis, P. D. J.
(1999) Computational approaches to combinatorial chemistry. Annu Rep Med Chem 34,
287296.
14. Brown, R. D., Hassan, M., Waldman,
M. (2000) Combinatorial library design
for diversity, cost efficiency, and druglike characters. J Mol Graph Model 18,
427437.
15. Jamois, E. A., Hassan, M., Waldman, M.
(2000) Evaluation of reagent-based and
product-based strategies in the design of
combinatorial library subsets. J Chem Inf
Comput Sci 40, 6370.
16. Douguet, D., Thoreau, E., Grassy, G. (2000)
A genetic algorithm for the automated generation of small organic molecules: drug design
using an evolutionary algorithm. J Comput
Aided Mol Design 14, 449466.
17. Sheridan, R. P., SanFeliciano, S. G., Kearsley, S. K. (2000) Designing targeted libraries
with genetic algorithm. J Mol Graph Model
18, 320334.
18. Gillet, V. J., Khatib, W., Willett, P., Fleming,
P. J., Green, D. V. S. (2002) Combinatorial
library design using a multiobjective genetic
algorithm. J Chem Inf Comput Sci 42,
375385.
19. Chen, G., Zheng, S., Luo, X., Shen, J., Zhu,
W., Liu, H., Gui, C., Zhang, J., Zheng,
M., Puah, C. M., Chen, K., and Jiang, H.
(2005) Focused combinatorial library design
based on structural diversity, drug likeness
and binding affinity score. J Comb Chem 7
(3), 398406.

20. Mason, J. S., Beno, B. R. (2000) Library

design using BCUT chemistry-space descriptors and multiple four-point pharmacophore
fingerprints: simultaneous optimization and
structure-based diversity. J Mol Graph Model
18, 438451.
21. Shi, S., Peng, Z., Kostrowicki, J., Paderes,
G., Kuki, A. (2000) Efficient combinatorial
filtering for desired molecular properties of
reaction products. J Mol Graph Model 18,
478496.
22. Gobbi, A., Poppinger, D., Rohde, B. (1997)
Developing an in-house system to support
combinatorial chemistry. Perspect Drug Discov Des 7/8 (Combinatorial Methods for the
Analysis of Molecular Diversity), 131158.
23. Polinsky, P., Feinstein, R. D., Shi, S., Kuki,
A. (1996) LiBrain: software for automated
design of exploratory and targeted combinatorial libraries, in (Chaiken, I. M., Handa, K.
D., eds.) Molecular Diversity and Combinatorial Chemistry. American Chemical Society,
Washington, DC, pp. 219232.
24. Shi, S., Kuki, K., Zhou, Z., Na, J., Thacher,
T., Yanovsky, A., Polinsky, P. LiBrainTM , An
Intelligent System for the High-Throughput
Design of Combinatorial libraries in Drug
Discovery. Poster Presentations at the Fifth
International Conference on Chemical Structures (1999) and the Second European Conference on Strategies and Technologies for Identification of NOVEL BIOACTIVE COMPOUNDS (1998).
25. Leach, A. R., Bradshaw, J., Green, D. V.
S., Hann, M. M., Delany, J. J., III (1999)
Implementation of a system for reagent selection and library enumeration, profiling, and
design. J Chem Inf Comput Sci 39, 1161
1172; also see a review article from Leach, A.
R., Hann, M. M. (2000) The in silico world
of virtual libraries. Drug Discovery Today 5,
326336.
26. Yasri, A., Berthelot, D., Gijsen, H.,
Thielemans, T., Marichal, P., Engles, M.,
Hoflack, J. (2004) REALISIS: a medicinal
chemistry-oriented reagent selection, library
design, and profiling platform. J Chem Inf
Comput Sci 44, 21992206.
27. (a) Truchon, J., Bayly, C. I. (2006) GLARE:
A new approach for filtering large reagent
lists in combinatorial library design using
product properties. J Chem Inf Model 46,
15361548. (b) Stanton, R. V., Mount,
J., Miller, J. L. (2000) Combinatorial
library design: maximizing model-fitting
compounds within matrix synthesis constraints. J Chem Inf Comput Sci 40, 701705.
28. Peng, Z., et al. PGVL: a vast virtual space
of synthetic feasible compounds based on

PGVL Hub

29.

30.
31.
32.
33.
34.
35.

36.
37.

38.
39.

40.

41.

captured knowledge of combinatorial chemistry synthesis protocols at the enterprise

level. Manuscript in preparation.
Hu, Q., Peng, Z., Kostrowicki, Kuki, A.
(2011) LEAP into the Pfizer Global Virtual
Library (PGVL) space creation of readily
synthesizable design ideas automatically, in
(Zhou, J. Z. ed.) , Chemical Library Design.
Humana Press, New York, Chapter 13.
ISIS-Draw, MDL Information Systems, Inc.
http://www.mdli.com.
Microsoft Excel: http://office.microsoft.com/
en-us/excel/FX100487621033.aspx
SpotFire for data visualization and decision
making: http://spotfire.tibco.com/
Java web Start from Sun Microsystems, Inc:
http://java.sun.com/products/javawebstart/
Pipeline Pilot from SciTegic: http://www.
scitegic.com/
Dalby, A., Nourse, J. G., Hounshell, W .D.,
Gushurst, A. K. I., Grier, D. L., Leland,
B. A., Laufer, J. (1992) Description of several chemical structure file formats used by
computer programs developed at molecular
design limited. J Chem Inf Comput Sci 32,
244255.
A good resource for BASE64 encoding could
be found in: http://en.wikipedia.org/wiki/
Base64
Resource on data compression in Java could
be found in: http://java.sun.com/developer/
technicalArticles/Programming/
compression/
Resource about XML (http://en.wikipedia.
org/wiki/XML ) and its special characters
(http://www.devx.com/tips/Tip/14068)
Lipinski, C. A., Lombardo, F., Dominy, B.
W., Feeney, P. J. (1997) Experimental and
computational approaches to estimate solubility and permeability in drug discovery and
development settings. Adv Drug Deliv Rev
23 (13), 325.
More info on the LogD prediction tool
from of the ACD Lab could be found:
http://www.acdlabs.com/products/phys_
chem_lab/logd/
a) A Pfizer in-house service and framework
for development, validation, and deployment
of in silico models for ADME-Tox and target specific SAR prediction. New ADMETox models published into this service will
be automatically visible to all client software
packages (such as PGVL Hub) subscribed
into this service. b) Kalgutkar, A. S., Gardner, I., Obach, R. S., Shaffer, C. L., Callegari, E., Henne, K. R., Mutlib, A. E., Dalvie,
D. K., Lee, J. S., Nakai, Y., ODonnell, J. P.,
Boer, J., Harriman, S. P. (2005) A comprehensive listing of bioactivation pathways of

42.

43.

44.

45.

46.

47.

48.

49.

50.

319

organic functional groups. Curr Drug Metab

6, 161225.
Steuer, R. E. (1986) Multiple Criteria Optimization: Theory, Computations, and Application, John Wiley & Sons, Inc., New York,
ISBN 047188846X.
Sawaragi, Y., Nakayama, H., Tanino, T.
(1985) Theory of Multiobjective Optimization
(vol. 176 of Mathematics in Science and Engineering). Academic Press Inc., Orlando, FL,
ISBN 0126203709.
Messac, A., Ismail-Yahaya, A., Mattson, C.
A. (2003) The normalized normal constraint
method for generating the pareto frontier.
Struct Multidis Optim 25(2), 8698.
Das, I., Dennis, J. E. (1998) Normalboundary intersection: a new method for
generating the pareto surface in nonlinear
multicriteria optimization problems. SIAM J
Optim 8, 631657.
Deb, K., Pratap, A., Agarwal, S., Meyarivan, T. (2002) A fast and elitist multiobjective genetic algorithm. NSGA-II. IEEE Trans
Evol Comput 6(2), 182197.
Two examples: Mobius from Coalesix, Inc.: A GA driven, R-group
based compound/library design tools
(http://www.coalesix.com/FAQ.html); and
C2-LibX from Accelrys, Inc. contains a GA
based library design module (more from
http://accelrys.com/ ).
Zhou, Z., Shi, S., Na, J., Peng, Z.,
Thacher, T. (2009) Combinatorial librarybased design with basis products. J Comput
Aided Mol Des 23, 725736.
SciTegics re-implementation of Canonical SMILES string based on the original
publication: Weininger, D., Weininger, A.,
Weininger, J. L. (1989) SMILES. 2. Algorithm for Generation of Unique SMILES
Notation. J Chem Inf Comput Sci 29(2),
97101.
ChemSelect AQB (Advanced Query
Builder): A Pfizer in-house developed reusable Java component that allows users to
query various molecular structure databases
within Pfizer based on molecular structural information and/or other properties,
retrieve, manage, and export retrieved hits.
It has a functional role similar to MDL
ISIS/BASE, but with many enhanced capabilities. PGVL Hub has embedded this reusable Java component within itself so that
users can search for suitable reactants from
various corporate reactants databases and
inventory houses and return the hits seamlessly back into PGVL Hub design session.
From users point of view, the ChemSelect
AQB component is just part of PGVL Hub.

320

51.

52.
53.

54.

Peng et al.
This is one of the successful examples that
re-usable components are developed and
shared among multiple software development projects within Pfizer.
WebLogic is a J2EE middle-tier software
suit from the former BEA Systems, now
part of Oracle (http://www.oracle.com/
appserver/weblogic/weblogic-suite.html)
Oracle is a database application software from
Oracle (http://www.oracle.com/index.html)
Smith, G. F. (2006) Enabling HTS
Hit follow-up via Chemo informatics,
File Enrichment, and Outsourcing. High
Throughput Medicinal Chemistry II; MMS
Conferencing & Events Ltd., Institute of
Physics; London, 2006. This article is also
available on-line via this web link (http://
www.mmsconferencing.com/pdf/htmc/g.
smith.pdf).
Clark, J. D., Hu, Q., Kuki, A., Peng, Z.,
Sciammetta, N., Smith, G. F., Ramirez-

Weinhouse, M., Van Hoorn, W. (2006)

Pfizer global virtual library: one-stop-shop
for design on the desktop. A poster given
by Nunzio Sciammetta during the 2006 Gordon Conference on Combinatorial Chemistry,
August 2025, 2006 in The Queens College
Oxford, United Kingdom.
55. Teng, M., Zhu, J., Johnson, M. D., Chen,
P., Kornmann, J., Chen, E., Blasina, A., Register, J., Anderes, K., Rogers, C., Deng, Y.,
Ninkovic, S., Grant, S., Hu, Q., Lundgren,
K., Peng, Z., Kania, R. S. (2007) Structurebased design of (5-Arylamino-2H-pyrazol-3yl)-biphenyl-2 ,4 -diols as Novel and Potent
Human CHK1 Inhibitors. J Med Chem 50
(22), 52535256.
56. Peng, Z., Hu Q. (2011) Design of targeted
libraries against the human Chk1 kinase using
PGVL Hub, in (Zhou, J. Z. ed.) Chemical
Library Design. Humana Press, New York,
Chapter 16.

Chapter 16
Design of Targeted Libraries Against the Human Chk1
Kinase Using PGVL Hub
Zhengwei Peng and Qiyue Hu
Abstract
PGVL Hub is a Pfizer internal desktop tool for chemical library and singleton design. In this chapter, we
give a short introduction to PGVL Hub, the core workflow it supports, and the rich design capabilities
it provides. By re-creating two legacy targeted libraries against the human checkpoint kinase 1 (Chk1)
as a showcase, we illustrate how PGVL Hub could be used to help library designers carry out the steps
in library design and realize design objectives such as SAR expansion and improvement in both kinase
selectivity and compound aqueous solubility. Finally we share several tips about library design and usage
of PGVL Hub.
Key words: PGVL Hub, combinatorial chemistry, library design, reaction, synthesis protocol,
reactant, product, enumeration, filtering, Chk1, kinase, inhibitor, SAR, ADME&T (Adsorption,
Distribution, Metabolism, Excretion, and Toxicity), selectivity, solubility, proteinligand complex.

1. Introduction
1.1. PGVL Hub

PGVL Hub was developed for and deployed within Pfizer global
chemistry communities (1). The main goal of PGVL Hub is to
offer bench chemists a very capable desktop tool to (a) access
Pfizers proprietary chemistry knowledge database containing
information about many experimentally validated combinatorial
chemistry synthesis protocols; (b) support and streamline the full
cycle of library design, synthesis, and registration; and (c) harness
the power of synergy with many other desktop software packages
(ISIS/Draw, MS-Excel, SpotFire (http://spotfire.tibco.com/),
and additional 2D/3D molecular design tools). PGVL Hub has

J.Z. Zhou (ed.), Chemical Library Design, Methods in Molecular Biology 685,
DOI 10.1007/978-1-60761-931-4_16, Springer Science+Business Media, LLC 2011

321

322

Peng and Hu

been used by more than 1000 users within Pfizer with more than
110,000 user logins accumulated since 2004. Its application to
targeted library design is the focus of this chapter.
In 2007, Teng and coworkers published a potent and selective
lead series of human Chk1 inhibitors (2). Their work was initiated by two advanced lead matters 1808-1 and 1819-1 obtained
through two rounds of targeted library design and synthesis based
on a high-throughput screening (HTS) hit Cpd-1 (see Fig. 16.1
for the progress of this hit through the library-based lead optimization). This report distills the essence of those two legacy targeted libraries to showcase the design process using PGVL Hub.
The PGVL Hub screen shots used in this report are recreations
of our legacy design efforts in the past.

1.2. A Sample Case

of Targeted Library
Design Against the
Chk1 Kinase Domain

OH

HO

OH
H
N
N
HN

HN
O

H
N

Series1808
N

2x44, VRXN-2-00010

HN

OH

HN
O

H
N

Series1819
N
2x88, VRXN-2-00010

2x44

HN

2 x8 8

Cpd-1
CHK1_Ki = 0.005uM

1808-1
CHK1_Ki = 0.3uM

1819-1
CHK1_Ki = 0.0005uM

ClogP = 4.6
ACD_logD = 3.2 (pH=7.4)

ClogP = 3.9
ACD_logD = 2.3 (pH=7.4)

ClogP = 3.5
ACD_logD = 2.3 (pH=7.4)

CDK1_Ki = 0.009uM
CDK2_Ki = 0.015uM
VEGF_Ki = 0.0066uM
LCK = 37%@1uM

CDK1_Ki = 24uM
CDK2_Ki = 5uM
VEGF_Ki = 53.9%@1uM
LCK = 46%@1uM

CDK1_Ki =4%@1uM
CDK2_Ki = 0.242uM
VEGF_Ki = 29%@1uM
LCK = 14%@1uM

Fig. 16.1. Progress of two rounds of Chk1 targeted libraries. Cpd-1 is the original HTS hit with a broad kinase inhibition
profile and based on which the first round library was designed and synthesized. 1808-1 is the best hit from the first
round targeted library with improved kinase selectivity profile, based on which the second round library was designed
and synthesized. 1819-1 is the best lead with improved potency, kinase selectivity, and solubility. Co-crystal structures
of Chk1 kinase domain and corresponding lead compounds were solved and extensively utilized in structure-based
singleton and library designs. For details of the X-ray co-Crystal structures, please refer to the publications from Ming
and et al (4a) and Foloppe and et al for details (4b).

Design of Targeted Libraries

323

2. Materials
More information about the biological roles of the Chk1 gene
and its protein product can be found in Ref. (3). In short, DNA
damages are sensed and passed to the Chk1 kinase to activate the
checkpoint between the G2 (secondary gap) and the M (mitosis) phases of the cell cycle. The cell cycle rest at this checkpoint
allows the cell to repair its DNA damages before proceeding to
the M phase (see Fig. 16.2). Cells entering into the M phase with
un-repaired DNA damages tend to suffer from mitotic catastrophe, which ultimately leads to cell death via the apoptosis pathway. Since the anti-cancer drugs used in standard chemotherapies
are mostly DNA-damaging agents intended to induce cancer cell
death in the M phase, it has been hypothesized that a Chk1 kinase
inhibitor could synergistically enhance the anti-cancer effect of
those DNA-damaging agents. What makes this approach even
more attractive is the observation that normal cells tend to arrest
at various checkpoints in both G1 and G2 phases after DNA damage, while many cancer cells tend to heavily rely on the G2/M
checkpoint to repair DNA damages where activity of Chk1 is critical. This implies selective kill of cancer cells vs. normal cells (3).
The first X-ray crystal structure for the Chk1 kinase domain was
solved by Pfizer scientists at the Pfizer La Jolla site (4a), and the
key structural features around its hinge region in the ATP binding
site are also depicted in Fig. 16.2. Due to its above-mentioned

G2

Chk1
M

G1

G0

resting

Fig. 16.2. Cell cycle, Chk1 at the G2/M checkpoint, and key structural features of the
Chk1 kinase domain (4a). The highlighted location is the hinge region of the Chk1 kinase
domain, which also corresponds to the same regions of proteinligand structures shown
in Fig. 16.1.

324

Peng and Hu

connections with cancers, Chk1 has been identified as an attractive oncology target (5) for inhibition by small organic molecules
(6). For details on the biological assays and solutions of protein
ligand X-ray crystal structures cited in this report, please refer to
Refs.(2) and (4a) directly.

Structure view panel

Property view panel

Two viewer panels

Decision Maker

Integration
with SpotFire

Fig. 16.3. Screen shots of PGVL Hub. It has two ways to display molecules and their properties (Structural Viewer panel
and Table viewer panel). It has been integrated with SpotFire for data visualization. It also has a decision maker capable
of handling numerical and textual data as well as user selections by hand.

The library design tool used for this report is PGVL Hub (1).
Figure 16.3 contains several screen shots of PGVL Hub, highlighting its capabilities in viewing molecules and their properties,
decision making, and integration with SpotFire for data visualization. Figure 16.4 describes the main workflow of library design
seen through the workflow manager of PGVL Hub and the key
questions a library designer would ask and address with the help
of a library design tool such as PGVL Hub. Those key questions
in library design can be summarized into the following list: What
chemical reaction should be used? What reactants are available to
start the library design? Which reactants (also called monomers)
or products should be chosen for testing design hypotheses as
well as satisfying various constraints such as ADME&T compliance? How can the library design be communicated to collaborators as well as a downstream synthesis process? Figure 16.5
describes more about the design strategies and PGVL Hubs
capabilities that allow a library designer to analyze the possible

Design of Targeted Libraries

(1) Initiate a library

design with a userdrawn reaction or a
pre-registered PGVL
VRXN reaction to
enable product
structure enumeration.
(What chemistry do I
want to use for my
library?)

(3) Analyses and design

decisions can be done at
monomer level. (Which
monomers are good for
realizing my design
objectives, testing my
hypotheses, or satisfying
giving design constrains?)

(2) Input monomers into a

design session (e.g., user
drawn templates , downloading pre-mined
monomers suitable for a
specific registered
synthesis protocol) (Which
monomers can I start
with for my design?)

325

(6) Many ways

to export design
results to
various forms

(4) Sooner or
later, one wants
to enumerate
product
structures
explicitly

(5) Even more analyses and

design decisions can be done
at the product level. Hub
allows one to design via
cherry-picking products or
shape a fully-combinatorial
library based on product
properties. (Which products
are good for my design
objectives, testing my
hypothesis, or satisfying
design constrains?)

Fig. 16.4. The basic library design workflow enabled in PGVL Hub.

At the monomer level:

At the product level:

Monomer availability (filter out those

low in stock room(s));

Overlap with known compound collection. (Avoid re-making same

molecules when possible);

Rough MW cutoff at monomer level;

Many more molecular properties can
be calculated and used for monomer
prioritization and selection;

Many molecular properties can be calculated and used for product

prioritization and selection (RO5, structure-Alert, more ADME&T
estimations, project specific activity model, etc);

Similarity to known template(s);

substructure mapping and filtering;
Identify and remove duplicates;
Cluster analysis and sampling;
random sampling;

Similarity to known template(s);

substructure mapping and filtering;
Identify and remove duplicates;
Cluster analysis and sampling;
random sampling;
Make decision directly on individual product which yields a cherrypicking library (best product property profile but lower library production
efficiency);
Make decision on monomers based on property profile of their
corresponding products (Combinatorial-shaping). This leads to a fully
combinatorial library design to retain efficiency in library production;

Fig. 16.5. Library design strategies and features enabled in PGVL Hub.

reactants/products one can use/synthesize and select a subset

from which to realize the intended objectives of a library. For
more detailed information, please refer to our report dedicated to
PGVL Hub (1).

326

Peng and Hu

Other computational models and software packages were also

used for design of the targeted libraries showcased in this report,
such as Rule-Of-Five (8), ClogP (9), LogD (10), and a protein
structure-based docking and scoring tool called AGDOCK (12)
and its associated proteinligand score function called HT-Score
(13). For the library design, we have also used the Tanimoto coefficient (11) computed based on the molecular fingerprints from
SciTegic Pipeline Pilot (14) as the measure of molecular similarity.

3. Methods
3.1. Information
Known Before the
Design of the First
Targeted Library

The initial HTS hit Cpd-1 was originally synthesized as an

inhibitor against another human kinase called VEGF (7) with a
measured Ki value of 6.6 nM. Nevertheless, it is not very selective and possesses a broad inhibition profile against many kinases
such as Chk1 (5 nM), CDK1 (9 nM), CDK2 (15 nM), and LCK
(37% at 1 M) (see Fig. 16.1). Cpd-1 is also known to have low
aqueous solubility, as indicated by the high values of calculated
ClogP (4.6) and LogD (3.2 at pH 7.4) values (9, 10). The Xray crystal structure of Cpd-1 with the Chk1 kinase domain was
solved through an in-house effort (see Fig. 16.1). Binding site
comparison of this proteinligand complex with others containing different kinase domains and different ligands in Pfizers crystal structure knowledge database (those Pfizer proprietary complex structures are not shown in this report) was conducted. This
analysis suggested that the Chk1 region accessed by the righthand side of Cpd-1 might provide opportunities to gain specificity toward Chk1. Therefore the objectives for the first targeted
library focused on improving selectivity and aqueous solubility
while expanding the body of SAR information around the selectivity pocket with 244 = 88 product molecules (see Fig. 16.6).

3.2. Design Steps of

the Targeted Library
3.2.1. Selection of
Reactions

As shown in Fig. 16.6, we planned to use a pre-registered combinatorial chemistry protocol (LJ0194) to synthesize the targeted library. PGVL Hub allowed us to easily search and load
this pre-registered reaction scheme into a design session without
the need to draw a reaction scheme required for product enumeration (see Fig. 16.7). Even for this simple reaction, a simple
reaction scheme drawn by users may not be sufficient to ensure
proper formation of product structures in the case where bonds
associated with chiral centers are near the reactive sites on the
reactants.

Design of Targeted Libraries

327

Design hypothesis: Explore the right hand side of Cpd-1 for specificity
(selectivity) and new ways to achieve affinity.
VRXN ID: VRXN-2-00010; Synthetic protocol: LJ0194
selectivity Pocket
Hinge Region

OH

H
N

N
HN

H
N

H
N

O
OH

H
N

H
N

H
N

R1
N
R2

2 acids

H
N

O
R1
N
R2

H
N

OH
N

H
N

R1
N
R2
O

44 amines

Fig. 16.6. Objectives of the first round library. The main goal is to explore the protein pocket probed by the right-hand
side of Cpd-1 to improve kinase selectivity and further build SAR knowledge. The single arylaryl bond was replaced by
three bonds containing an amide group with more flexibility. A registered combinatorial synthesis protocol (LJ0194) was
used for this library and a 244 plate format was planned before the design of the library.

3.2.2. Selection of
Reactants

In addition to the pre-registered reaction scheme, the pre-mined

reactants (acids and amines) suitable for the reaction conditions
used in the library synthesis protocol LJ0194 were also available
for download directly through PGVL Hub (also see Fig. 16.7).
Since the two special acids of the library are already chosen,
the task of library design now is to select 44 amines from the
8449 amines that are compatible with the synthesis protocol (see
Fig. 16.8).

3.2.3. Reactant Analysis

and Selection

Due to the small size of the binding pocket accessed by the righthand side of Cpd-1, we first hypothesized that smaller amines
would have a higher chance of fitting to that binding pocket. By
applying the MW and ClogP calculations and filtering, we significantly reduced the possible choices of amines from the original
set of 8449. Using molecular similarity score (11, 14) with respect
to 4-amino-2-methoxyphenol, we ensured that the neighborhood
of the right-hand side of Cpd-1 was well sampled (see Fig. 16.1).
Finally we looked up the reactant amount available in our chemical inventory for each reactant and only used those with sufficient amount for library production so that the designed library
could be synthesized without further delays. With those two considerations, we were able to focus the remaining choices further

328

Peng and Hu

Fig. 16.7. Accessing the pre-registered reactions and pre-mined suitable reactants for synthesis protocol LJ0194 to
initiate library design. PGVL Hub makes it simple to load the pre-registered reaction scheme and suitable reactants for
the reaction conditions specified in synthesis protocol LJ0194. The product structure enumeration and the synthetic
feasibility of those product molecules are taken care of by the PGVL Hub through extensive knowledge capturing and
reusing, so that the library designers can focus their efforts on design issues such as target binding, selectivity, and
ADME&T.

to a subset of a few hundred amines. Then we used the Structural Viewer panel of PGVL Hub (see Fig. 16.3) to display many
amines in a single page and browsed through them visually. Each
molecule was examined in terms of possible hypothesis it could
help to form and validate. Molecular diversity of the final library
is also a consideration. Desirable ones were marked with color
markers provided by PGVL Hub and included in the final set of
44 (plus a few backups). Even though the actual legacy design also
contained input from the project medicinal chemists and library
production chemists, the first target library was designed mainly
based on the reactant-level considerations described so far. The
first targeted library yielded several hits with weaker potency than
Cpd-1 (see Fig. 16.9); however, assay data on kinase selectivity
suggested that the top hit 1808-1 had a much improved kinase
selectivity profile and some improvement in solubility (see Fig.
16.1). This prompted the project team to solve the co-crystal

Design of Targeted Libraries

329

Fig. 16.8. Reactant-level (pre-enumeration) design steps. This is a screen shot of PGVL Hub during the design of the
two libraries. The reactant sets for this two-component reaction and the generated explicit libraries and products are
all captured during the design session (see the left-hand side). The A-component is for acids and the B-component
for amines. The molecular structures of the two special acids are shown in Fig. 16.6. Many annotations can be added
to reactants to aid their analysis and selection. Here ClogP, molecular weight (MW), similarity (SIMI) with respect to a
user-defined reactant, and reactant amount available in the inventory house are just a few examples.

structure of 1808-1 bound to the kinase domain of Chk1 (see

Fig. 16.1) and plan another round of targeted library using a
288=176 format (see Fig. 16.10) to further refine 1808-1.
3.2.4. Product Analysis
and Selection

For the second targeted library, we focused our attention on the

enumerated products directly. After certain initial selection steps
for reactants similar to what were used for the first targeted library,
a few thousand product molecules were enumerated. Then we
subjected them to various property calculations for additional
analysis and filtering. MW, ClogP (9), and LogD (10) values were
computed to shape molecular size and solubility. Molecular similarity score (SIMI, see (11) and (14)) with respect to 1808-1 was
computed to ensure that the chemical neighborhood of 1808-1
was well sampled. A protein binding pocket was created based
on the experimental X-ray structure of 1808-1 bound to the
Chk1 kinase domain, and all product molecules were exported
out of PGVL Hub and docked and scored using the in-house
proteinligand docking software (12, 13). The numeric docking
scores (HT_Score, (13)) were then imported back into PGVL
Hub and displayed along with the product structures (see Fig.
16.11). We also leveraged activity models built by project teams
working on other kinase targets to provide some early read about
kinase selectivity (e.g., Predicted_CDK2 activity, see Fig. 16.11),
even though no such model existed at the time when we designed

330

Peng and Hu

Fig. 16.9. Top hits from the first targeted library. One can see that a fairly diverse set of small amines are all tolerated
by the binding pocket but with a significant reduction in potency when compared with the initial HTS hit Cpd-1 (5 nM).
On the other hand, the top hit 1808-1 shows significant improvement in kinase selectivity and improved solubility (see
data given in Fig. 16.1).

Design hypothesis: Explore the right hand side of 1808-1 for specificity
(selectivity) and further improve potency.
VRXN ID: VRXN-2-00010; Synthetic protocol: LJ0194
selectivity Pocket

Hinge Region
H
N

OH
HN

N
O
HN
N

H
N

H
N

O
H
N

OH

H
N

H
N

R1
N
R2
H

2 acids

N
N

H
N

O
R1
N
R2

H
N

OH
N

H
N

R1
N

R2

88 amines

Fig. 16.10. Objectives of the second round targeted library. The main goal is to further expand the SAR knowledge around
the right-hand side of 1808-1 with the aim to improve potency while retaining kinase selectivity. A same combinatorial
synthesis protocol (LJ0194) was used and a 288 plate format was planned before the design of the library.

Design of Targeted Libraries

331

Fig. 16.11. Product-level design steps. In this screen shot, product structures and their calculated properties are listed in
the table. Those annotations are key to implementing various design considerations such as ADME&T profile (e.g., Rule of
Five (8, 9), LogD (10)), molecular similarity with respect to lead compound 1808-1 (SIMI), proteinligand complementation
(docking score again the binding pocket in the Chk1 kinase domain initially occupied by 1808-1, such as HT score (13)),
kinase selective (predicted CDK2 activity based on an in silico model), and finally checking for duplicates against the
corporate compound database (PRGL_Lookup).

our legacy libraries in the past. Finally we wanted to know if any of

our designed molecules had already been registered into Pfizers
corporate compound database (PGRL_lookup, see Fig. 16.11).
If duplicates were found, we could either remove them from the
design or strategically include a few of them in our design as internal controls for the library production and biological assays. With
all those desired molecular properties calculated and combined,
we conducted further focusing and filtering to reduce our choices
to about a few hundreds, then finalized our 88 amines (plus some
backups) through visual inspection of product structures using
PGVL Hubs Structural Viewer panel. Again, input from project
medicinal chemists and library production chemists was included
in the actual legacy library design. The result of the second round
library is given in Fig. 16.12. One compound 1819-1 has a much
higher potency than 1808-1. Additional data also suggested a
very sharp SAR among 1819-1, 1808-1, 1819-2, 1819-5, and
1819-6. The X-ray structure of 1819-1 bound to the kinase
domain of Chk1 was solved subsequently to provide significant
molecular insights for the observed sharp SAR. It showed that
the OH group at the ortho position was able to replace two
tightly bound waters and induce a significant rearrangement in
that region of the protein binding pocket (see Fig. 16.1). This

332

Peng and Hu

Fig. 16.12. Top hits from the second round library. One compound 1819-1 (0.5 nM) shows significant improvement over
the lead 1808-1 (300 nM) in terms of Chk1 inhibition results. More data show (see Fig. 16.1) that it is even more potent
than the initial hit Cpd-1 (5 nM) while having a much improved kinase selectivity profile and better solubility.

in-depth structural knowledge generated by 1819-1 was further

used in the additional lead optimization effort through one-onone synthesis (2).
3.2.5. Integration of
Final Design (List
Operation at the
Reactant/Product Level)

The library workflow depicted in Fig. 16.2 seems to imply that

library design is a straightforward step-by-step linear process. In
reality, it involves multiple iterations and revisions through collaborations with other stakeholders. It is quite common for a
designer to simultaneously pursue several design hypotheses and
strategies that yield multiple intermediate library designs. PGVL
Hub offers a list of logic operations (AND, OR, and MINUS)
at the reactant as well as the library/product levels so that the
designer can easily compare and/or combine two reactant lists or
two libraries. The final design submitted for library production
is usually a combination of several individual designs intended to
test multiple hypotheses.

3.3. Result Summary

and Project Impact

Two rounds of targeted libraries (244 and 288) were

designed, synthesized, and assayed within 6 months. This effort
led to a new lead matter 1819-1 with improved potency (10
folds), kinase selectivity (>100 folds), and better solubility (1
log unit or 10 folds) when compared with the original HTS hit
Cpd-1. The extensive SAR information spanned by those 264

Design of Targeted Libraries

333

library compounds, supplemented by the two co-crystal structures associated with 1801-1 and 1819-1 showing significant
protein flexibility around the ligand binding site, provided a
solid foundation for additional lead optimization effort on this
project (2).
There are many design requirements, constraints, and
hypotheses for a given library design case. In our example, we
have touched upon several reactant as well as product-level design
considerations. These considerations include but not limited to
ADME&T properties, similarity with respect to one or more lead
molecules, docking and scoring against a given protein binding
site, activity models for selectivity profiling, and even practical
issues such as reactant availability in chemical inventory systems
and duplication check against corporate compound collections.
Historical usage tracking strongly suggests that PGVL Hub is a
proven, streamlined, and highly effective design environment to
fulfill many of those diverse library design objectives in the hands
of Pfizer medicinal and computational chemists.

4. Notes
1. Shorten the cycle time: The designsynthesistest cycle of lead
optimization is the dominant workflow in drug discovery
projects. The library designsynthesistest cycle should be
short enough to be compatible with the progression of the
project so that relevant project questions can be proposed
and answered by targeted libraries in a timely manner. Effective communication and coordination among the library
designer, his/her project collaborators, and the library production team is essential to reduce the cycle time. PGVL
Hub allows one library designer to save a full design session
into a file and share it with another collaborator to enable
effective communication and coordination. Selecting only
reactants that are readily available from chemical inventory
systems is another way to bypass the wait required to restock missing reactants, and the inventory check feature of
PGVL Hub makes this check straightforward.
2. Complementary to singleton synthesis: In terms of lead optimization, library synthesis works as a shot-gun approach,
with multiple shots on the goal while its resolution is limited by availability of types of combinatorial chemistries and
suitable reactants. On the other hand, the one-on-one singleton synthesis practiced by standard medicinal chemistry
offers the highest resolution, yet at a lower throughput. So
for a well-explored SAR region, the singleton approach is
the best way to further refine project leads effectively, while

334

Peng and Hu

the library approach is best for exploring a new SAR region,

as in the example showcased in this report. In addition, the
singleton approach could be used to synthesize a few special template reactants (like those two special acids shown
in Fig. 16.6), which are then subsequently amplified extensively by targeted libraries.
3. Control the size of the enumerated library: As the name
implies, combinatorial libraries can explode in size very
quickly. Therefore one must perform reactant-level selections before product enumeration in most design cases. As
shown in the example library, molecular weight (MW) is an
effective filter to cut down number of reactants, so is reactant
availability inside the reactant inventory system. As a matter
of principle, more expensive computational approaches (e.g.,
proteinligand docking and scoring) should be applied only
to smaller subsets of reactants or products.
4. The importance of visual inspection: Visual inspection of reactants and product molecules offers a library designer tremendous value in terms of what product molecules can or cannot
be synthesized to help formulate and address SAR hypotheses. Project medicinal chemists have fondly called this popular approach cerebral processing. PGVL Hub has provided a capable environment to enable this approach (e.g.,
Structural Viewer panel with many molecules per page for
fast browsing, sorting of displayed molecules by molecular
properties, and multiple color markers to label molecules for
further processing).
5. Leverage externally computed molecular properties: Many
molecular properties pertinent to ADME&T and projectspecific activity and/or selectivity models are available within
PGVL Hub for use by library designers. If a computational model is not available within PGVL Hub, one could
export pertinent molecules out from PGVL Hub, compute
the desired molecule properties using the external software
package, and then import the computed results back into
PGVL Hub for further decision making in an integrated
manner. The docking and scoring calculation used in this
report (HT_Score, see Fig. 16.11) exemplifies this type of
use case.

Acknowledgments
We would like to express our gratitude toward other members of
the Chk1 project team for their design input and their efforts in
library production (Haresh Vazir and Dr. Ming Teng), bio-assays

Design of Targeted Libraries

335

(James Register), X-ray structures of proteinligand complex (Dr.

Ping Chen, Dr. Chun Luo, and Yali Deng), and in-depth medicinal chemistry follow-ups (Dr. Ming Teng and her team). We also
appreciate the strong Chk1 project leadership provided by Dr.
Karen Lundgren. Finally we would like to thank Drs. Joe Zhongxiang Zhou and Ben Burke for their valuable comments and suggestions and Dr. David Simon for proof reading the draft.
References
1. Peng, Z., Yang, B., Mattaparti, S., Shulok,
T., Thacher, T., Kong, J., Kostrowicki, J.,
Hu, Q., Na, J., Zhou, J. Z., Klatte, K.,
Chao, B., Ito, S., Clark, J., Coner, C., Waller,
C., Kuki, A. (2011) PGVL Hub: an integrated desktop tool for medicinal chemists to
streamline design and synthesis of chemical
libraries and singleton compounds, in (Zhou,
J. Z., ed.) Chemical Library Design. Humana
Press, New York, Chapter 15.
2. Teng, M., Zhu, J., Johnson, M. D., Chen,
P., Kornmann, J., Chen, E., Blasina, A., Register, J., Anderes, K., Rogers, C., Deng, Y.,
Ninkovic, S., Grant, S., Hu, Q., Lundgren,
K., Peng, Z., Kania, R. S. (2007) StructureBased Design of (5-Arylamino-2H-pyrazol3-yl)-biphenyl-2,4-diols as Novel and
Potent Human Chk1 Inhibitors. J Med Chem
50(22), 52535256.
3. (a) Zhou, B., Elledge, S. J. (2000) The DNA
Damage Response: Putting Checkpoints in
Perspective. Nature 408, 433439.
3. (b) Melo, J., Toczyski, D. (2002) A unified
view of the DNA-damage checkpoint. Curr
Opin Cell Biol 14, 237245.
4. (a) Chen, P., Luo, C., Deng, Y., Ryan,
K., Register, J., Margosiak, S., TempczykRussell, A., Nguyen, B., Myers, P., Lundgren, K., Kan, C. C., OConnor, P. M.
(2000) The 1.7 Crystal structure of
human cell cycle checkpoint kinase Chk1:
implications for Chk1 regulation. Cell 100,
681692.
4. (b) Foloppe, N., Fisher, L. M., Francis, G.,
Howes, R., Kierstan, P., Potter, A. (2006)
Identification of a buried pocket for potent
and selective inhibition of Chk1: Prediction and verification. Bioorganic & Medicinal
Chemistry 14, 17921804.
5. Li, Q., Zhu, G.D. (2002) Targeting serine/threonine protein kinase B/Akt and cellcycle checkpoint kinases for treating cancer.
Curr Top Med Chem.2, 939971.
6. (a) Jackson, J. R., Gilmartin, A., Imburgia,
C., Winkler, J. D., Marshall, L. A., Roshak,
A. (2002) An indolocarbazole inhibitor of
human checkpoint kinase (Chk1) abrogates

6.
6.
6.

6.

6.

7.

8.

9.

10.

cell cycle arrest caused by DNA damage.

Cancer Res 60, 566572.
(b) Prudhomme, M. (2006) Novel checkpoint 1 inhibitors. Rec Pat Anti-Cancer
Drug Discov 1, 5568.
(c) Tao, Z. F., Lin, N. H. (2006) Chk1
inhibitors for novel cancer treatment. AntiCancer Agents Med Chem 6, 377388.
(d) Foloppe, N., Fisher, L. M., Francis, G.,
Howes, R., Kierstan, P., Potter, A. (2006)
Identification of a buried pocket for potent
and selective inhibition of Chk1: prediction
and verification. Bioorg Med Chem 14, 1792
1804, and references cited therein.
(e) Tao, Z. F., et al. (2007) Structure-based
design, synthesis, and biological evaluation
of potent and selective macrocyclic checkpoint kinase 1 inhibitors. J Med Chem 50 (7),
15141527.
(f) Tong, Y., et al. (2007) Discovery of 1,4dihydroindeno[1,2-c]pyrazoles as a novel
class of potent and selective checkpoint
kinase 1 inhibitors. Bioorg Med Chem 15,
27592767.
Kania, R. S., Bender, S. L., Borchardt, A.,
Braganza, J. F., Cripps, S. J., Hua, Y., Johnson, M. D., Johnson, T.O.J., Luu, H.T.,
Palmer, C. L., Reich, S. H., TempczykRussell, A. M., Teng, M., Thomas, C., Varney, M. D. (2001) Wallace, M. B. Patent WO
0102369.
Lipinski, C.A., Lombardo, F., Dominy, B.
W., Feeney, P. J. (1997) Experimental and
computational approaches to estimate solubility and permeability in drug discovery and
development settings. Adv Drug Deliv Rev
23 (13), 325.
CLOGP v 4.0 for Unix, BioByte Corp.,
201 W. Fourth Street, Suite 204, Claremont,
CA 91711; see Leo, A. J. (1993) Calculating LogPoct from structures. Chem Rev 93,
12811306 and references cited within for
details.
More info on the LogD prediction tool
from of the ACD Lab could be found:
http://www.acdlabs.com/products/phys_
chem_lab/logd/

336

Peng and Hu

11. (a) Tanimoto, T. T. (1957) IBM Internal Report 17th Nov. 1957; and (b)
Wikipedia entry on the Tanimoto coefficient:
http://en.wikipedia.org/wiki/Tanimoto_
coefficient#Tanimoto_Coefficient_.28
Extended_Jaccard_Coefficient.29
12. (a) Gehlhaar, D. K., Verkhivker, G. M.,
Rejto, P. A., Sherman, C. J., Fogel, D. B.,
Fogel, L. J., Freer, S.T. (1995) Molecular recognition of the inhibitor AG-1343
by HIV-1 protease: conformationally flexible docking by evolutionary programming.
Chem Biol 2(5), 317324.
12. (b) Gehlhaar, D., Bouzida, D., Rejto,
P. A. (1999) Reduced dimensionality in

ligand-protein structure prediction: covalent

inhibitors of serine proteases and design of
site-directed combinatorial libraries, in (Parrill, A. L., Reddy, M. R. eds.) ACS Symposium
Series 719: Rational Drug Design. ACS Press,
New York, pp. 292311.
13. Marrone, T. J., Luty, B. A., Rose, P. W.
(2000) Discovering high-affinity ligands
from the computationally predicted structures and affinities of small molecules bound
to a target: a virtual screening approach.
Perspect Drug Discov Design 20, 209230.
14. Pipeline Pilot from SciTegic and its molecular
fingerprint system: http://www.scitegic.com/

Chapter 17
GLARE: A Tool for Product-Oriented Design of Combinatorial
Libraries
Jean-Franois Truchon
Abstract
Combinatorial chemistry with two or more diversity points often leads to an immense number of theoretical products. It is sensible to select the reagents based on the desired properties of the products in the
hope of maximizing the usefulness of the synthesized molecules. The presented tool enables the filtering
of reagents such that any further reagent selection will form products matching the desired properties.
Virtual combinatorial library leading to thousands of billions of products can be rapidly assessed. The
publicly available software (http://glare.sourceforge.net) and key algorithmic elements are discussed.
Key words: Combinatorial
multi-objective optimization.

library

design,

computer

algorithms,

product

properties,

1. Introduction
The design of chemical libraries often requires the selection of
only a small number of reagents compared to what is available
commercially or from proprietary repositories. The combinatorial
nature of multi-reagent synthesis can lead to far more products
than can be synthesized or screened. It is thus of great interest to
apply filtering schemes that are adapted to the goal of the chemical library (1, 2). In order to help the chemist focus on only the
reagents that would form a library with desired computed properties, it is attractive to use a computational algorithm to eliminate the unproductive reagents. In practice, it is often difficult
to map the desired product properties into reagent-based filtering rules. For example, if the products need to fit within a log
J.Z. Zhou (ed.), Chemical Library Design, Methods in Molecular Biology 685,
DOI 10.1007/978-1-60761-931-4_17, Springer Science+Business Media, LLC 2011

337

338

Truchon

P range, filtering rules applied to the reagents are rather difficult to find. Although one can guess a reagent-based threshold, it
would be difficult to account for the fact that some reagent classes
are fundamentally greasier than others. This sort of strategy has
been found to be misleading (3). The daunting task of generating chemical structures for millions or billions of virtual products
and assessing their properties in order to find the best reagents to
form the combinatorial matrix is clearly challenging. This is worsened by the multi-objective thresholds needed when more than
one property is monitored simultaneously.
A practical solution to this problem, called GLARE (Global
Library Assessment of REagents) (4), has been developed and
validated in our laboratories and is explained in this chapter. We
will focus on a specific chemical combinatorial library to illustrate
the workflow and the use of the software.

2. Materials
2.1. Computer
Program GLARE

The computer program used in this chapter has been made publicly available under an Open Source Initiative BSD License at
http://glare.sourceforge.net. This program is written in C++ and
has been successfully compiled on diverse platforms such as Mac
OS 10.X, Linux, IBM AIX, and Windows XP. Under its current form, GLARE is mainly a command line application that is
invoked identically on any of the platforms. Details of the parameters and options can be found at the aforementioned web site.

2.2. Chemical
Databases

There exist a multitude of chemical reagent sources. The Available Chemical DirectoryTM (ACD) collection, from Symyx Technologies Inc., lists as many as 1,160,000 unique chemicals with
chemical structure, pricing, supplier, purity, forms, etc.

2.3. The Oxazolidine

Library

The different steps of the protocol to filter chemical reagents

based on the product properties will be illustrated with an oxazolidine library (5) for which reagents adding diversity (dimension) are shown in Fig. 17.1. Even though only few hundreds of
reagents are picked from the ACD in each dimension, the total
number of theoretically accessible products is over 59 million.
The objective of GLARE is to filter the reagents such that any
further reagent selection would lead to a good (see Note 1) product. Figure 17.2 shows the property distributions of the products
before (black) and after (grey) the use of GLARE that considers a product good if its properties fall within the goodness range
(dashed vertical lines) for each property.

GLARE

339
R1

R1
H2N

R2

O
R3

O
OH
(651)

O
R4

(637)

O
S

R2

O2 S N O
R4 R
3

Cl

(143)

(59,300,241)

Fig. 17.1. The oxazolidine library used to illustrate how the GLARE tool works. The
number of reagents considered and the total number of products they can potentially
form are given in parenthesis under each chemical class.

30

20

Frequency (%)

Frequency (%)

Initial
Final

40

Initial
Final

25

15
10

30

20

10

>7
0

0
2

5
6
7
8 9 10 11 12 13 14
Number Hydroge-Bond Acceptors

50

25

40

Initial
Final

20

Initial
Final

30

Frequency (%)

Frequency (%)

1
2
3
4
5
6
7
Number of Hydrogen-Bond Donors

20
10

15
10
5

> 65
0

10

15

20 25 30 35 40 45 50 55
Number of Non-Hydrogen Atoms

60

65

4 3 2 1 0

1 2 3 4 5
Calculated logP

9 10

Fig. 17.2. Properties profile of the products in the oxazolidine library formed with all available reagents (black) and after
filtering the reagents (grey) based on the product properties with GLARE. The multi-objective thresholds are illustrated
by the dashed vertical lines. The initial library is formed by 651 637 143 products and the filtered library by 144
143 92 products (aminoalcohols aldehydes sulfonyl chlorides).

3. Methods
The different steps, files, and parameters necessary to optimize
the oxazolidine library are discussed in this section.
3.1. Selection of the
Reagents

It is standard practice to remove chemical functionalities susceptible to interfere with subsequent synthetic steps in the library. We

340

Truchon

used a Merck & Co proprietary web tool to this end called Virtual
Library Toolkit (VLTK) described elsewhere (6).
3.2. Product
Properties and Offset
Calculations

One of the strategies behind GLARE is to take advantage of

the additivity of many of the computed properties in a chemical library. In other words, one can calculate the property of a
product by summing the properties of its diversity contributing
reagents corrected by an offset kept constant for the entire library.
Although this may seem relatively obvious for a property like the
number of non-hydrogen atoms, a real-value property such as the
calculated logarithm of the octanol/water partition (log P) or the
polar surface area (PSA) are also well approximated by this scheme
(4, 7). We write the property P of any product of the chemical
library as
N





P reagenti
P product = Poffset +

[1]

where Poffset is the constant offset correction for the entire

library of property P, P(reagenti ) the property of the ith diversity
reagent. In practice, the offset is calculated from a single example:


Poffset = P product

N




P reagenti

[2]

This has been shown to work well for a diverse set of libraries
(see Note 2) (4). In Fig. 17.3, the offset is calculated for the
oxazolidine library for properties related to Lipinskis rule of five
(8): the number of hydrogen bond acceptors (HBA), the number
of hydrogen bond donors (HBD), the number of non-hydrogen
atoms (NHA), and the calculated log P (9).

NH3+

+
OH

Cl

R1

P
R1
R2
R3
Offset

O2
S

R2

HBA
3
1
1
2
1

R3

HBD
0
4
0
0
4

NHA
18
6
4
10
2

N SO
2

logP
1.3
0.3
0.3
1.5
0.2

Fig. 17.3. Calculation of the oxazolidine offsets from a specific example. From the product (P ) property, each of the reagent (Ri ) property is subtracted. Four properties are
considered: the number of hydrogen bond acceptors (HBA), the number of hydrogen
bond donors (HBD), the number of non-hydrogen atoms (NHA), and the logarithm of the
octanol/water partition constant (log P ).

GLARE

341

The use of an additive scheme has the obvious advantage of

avoiding the explicit generation of each product structure, which
would be impractical whenever a large number of products are
possible. Indeed, the only requirement is the calculation of the
properties of each unmodified reagent, avoiding even the complication of forming the synthons. There are many commercial
and non-commercial software suitable for this task providing a
2D structure of a reagent. Just to name a few: the OEChem Tk
(10), the Molecular Operating Environment (MOE) (11), and
JOELib (12). In this specific work, we have used a Merck & Co
proprietary cheminformatics platform to calculate the properties
of each reagent list, each of which sprang from a substructure
query to the ACD.
3.3. Preparation of
the Input Files

With GLARE, there are only two file types that need to be prepared. First, the reagent property files contain one reagent per
line in a text file starting with a reagent ID followed by a list
of numbers corresponding to the reagent properties. The offset
information is given in a separate file with the same format and
contains only one line.
Second, the virtual library is combined according to the
instructions outlined in the library definition file. An example
for the oxazolidine library is given in Fig. 17.4. The keyword
DIMDEF associates a list of reagent property files to one combinatorial dimension identified by a user-defined alias (e.g., ALDEHYDES). The listed reagent files are simply appended in the program. The LIBDEF keyword is followed by a user-defined library

# Defines the combinatorial dimensions and list the reagent property files that are
combined.
DIMDEF AMINOALCOHOLS amino_alcohols_acd.gli amino_alcohols_inhouse.gli
DIMDEF ALDEHYDES aldehydes_acd.gli
DIMDEF SULFONYLS sulfonyl_chlorides_acd.gli
DIMDEF OFFSET oxazolidine_offset.gli
# Defines a combinatorial library called oxazolidines formed by the matrix of products
from the combination of the listed dimensions
LIBDEF OXAZOLIDINES AMINOALCOHOLS ALDEHYDES SULFONYLS
OFFSET
# Defines a property name with the expected minimum and maximum value of the
products.
PROPDEF HBD 0 5
PROPDEF HBA 0 10
PROPDEF NHA 0 35
PROPDEF LOGP -2.4 5.0
# Gives the order of the properties found in the property input files.
INPUTDEF HBA HBD NHA LOGP

Fig. 17.4. Example of the library definition file for GLARE. Bold text indicates a dedicated keyword, text in italic a user-defined alias, lines starting with a hash mark are
comments, and normal text gives the keyword-associated parameters.

342

Truchon

name (e.g., OXAZOLIDINES) and the list of the dimensions

that form the combinatorial matrix (e.g., AMINOALCOHOLS
ALDEHYDES SULFONYLS OFFSET ). When two or more
libraries share common intermediates the filtering of the common reagents can be achieved by specifying more than one LIBDEF keyword. This tells GLARE to simultaneously consider all
libraries in the filtering. The PROPDEF keyword associates the
minimum and maximum values for a well-behaved product to
a user-defined property name. Finally, the INPUTDEF keyword
lists the order of the properties read in the reagent property input
files.
3.4. Recommended
Optimization
Parameters

GLARE uses an iterative filtering that stops when a user-defined

fraction of the products formed by the remaining reagents comply
with the desired product property ranges. We call this fraction the
goodness. It is not sufficient to identify a set of reagents leading
to good products, but one wants to find the largest set of such
reagents to provide enough choice to a chemist who also needs
to account for other properties. We discuss here the impact of
the different optimization parameters on the resulting number
of reagents. We measure the efficiency of the filtering with an
effectiveness metric, which corresponds to the average fraction of
reagents left after optimization compared to what was available
initially. Quantitatively, the effectiveness E is defined by
$
#D
1  Ni, final
E=
D
Ni, initial

[3]

i=1

where D corresponds to the number of dimension (three for the

oxazolidine library), Ni, final to the number of reagents in the
dimension i after GLARE has been applied, and Ni, initial the number of reagents input before the filtering. When a high compliance
to the desired product properties is requested, more reagents are
pruned. Figure 17.5 shows the effectiveness of the oxazolidine
library as a function of the goodness threshold used. The obvious
drawback of a lower goodness threshold is a potential deterioration of the properties of the final library when the reagents are
selected for synthesis. We have found, more generally, that whenever a high compliance to the product property rules is needed, a
goodness threshold of 95% is the most appropriate as the last 5%
unduly reduces the effectiveness.
The scaled pruning strategy is an optional feature that is useful when one of the reagent sets is significantly smaller than the
others like the sulfonylchloride reagents of the oxazolidine library.
It is often difficult to retain enough diversity in these less populated reagent sets while maintaining high library goodness. The
principles of the scaled pruning is to eliminate reagents in the

GLARE

343

dimensions with more reagents faster. The iterative procedure

initially eliminates only reagents from the larger list and progressively starts to prune reagents from the smaller list as the lists
become of comparable size. The switching function that turns on
the pruning of smaller dimensions depends on a single parameter () (4). The final number of reagents in the three dimensions
of the oxazolidine library after applying GLARE with different
values of is shown in Fig. 17.6. A small has no effect and
as its value increases, proportionally more sulfonylchlorides are
kept and less of the other two more populated dimensions. As we
found more generally, a value of between 1 and 10 (a value of 6
is our default) leads to a more evenly distributed diversity across
the dimensions.
The third user-defined parameter discussed is related to the
partitioning scheme implemented in GLARE. To avoid the combinatorial explosion that makes a product-based filtering algorithm impractical, the reagent sets can be optionally partitioned
such that each reagents ability to form a good product is
evaluated in a sub-library formed by combining the individual
partitions in a systematic way. This is in contrast with the examination of all combinatorial products. The partitioning approximation systematically leads to libraries matching the desired goodness when verified with all the products (4). However, for a given
targeted goodness, the partitioning scheme reduces the effectiveness of the resulting library. Figure 17.7 shows the effectiveness
of the oxazolidine library as a function of the minimum number
of reagents in the created partitions. A partition size of 16 (corresponding to a value of 4 on the x-axis of Fig. 17.7) is generally
optimal.

Effectiveness of filtered library (%)

100
90
80
70
60
50
40
30
20
0

10 20 30 40 50 60 70 80 90 100
Goodness filtering threshold (%)

Fig. 17.5. This figure shows how requiring a higher fraction of the products to comply
(goodness) with the desired product properties reduces the fraction of retained reagent
(effectiveness). The initial goodness of the oxazolidine library used here is 18%.

Truchon

Number of reagents left at 95% goodness

200
180
160
140

aminoalcohols
aldehydes
sulfonylchlorides

120
100
80
60
40
20
0,001

0,01

0,1
1
10
Scaling parameter ()

100

1000

Fig. 17.6. This figure shows the final number of reagents left in each dimension once
a 95% goodness threshold is obtained as a function of the scaling parameter () displayed on a log scale. The larger is , the more reagents from the initially less populated
dimension are left. We found that = 6 generally leads to useful results.
40%
Optimized oxazolidine library effectiveness

344

111 s

50.6 s

18.7
18.7s

38%

3.94 s
1.03 s

36%

no partitioning

34%
0.24 s

32%
0.07 s

30%
28%

0.04 s

26%
24%
1

log2 (number reagent per partition)

Fig. 17.7. This figure illustrates the advantages and disadvantages of using partitioning. On the one hand, the timings (shown next to the individual points in seconds) are
tremendously reduced, on the other hand the effectiveness of the optimized oxazolidine library is sub-optimal with smaller partitions. A partition of 16 reagents seems best
overall.

In summary, the two main parameters related to the compliance to the product property rules (goodness) and the number
of reagents left for further selection (effectiveness) can be controlled by adjusting the algorithmic goodness threshold, the scaling parameter, and the size of the reagent partitions. Each library
being different, it may sometimes be useful to deviate from the
proposed defaults. The oxazolidine library is a good surrogate
for the relationships normally involved and Figs. 17.5, 17.6, and
17.7 can be used to assess sensitivity and expected effects of modifying these parameters.

GLARE

345

4. Notes
1. Here the word good and goodness are strictly related
to the binary classification that a product is good only if it
fits all the multi-objective criteria. GLARE could easily be
adapted to work with a scalar fitness score.
2. Most spectacular exceptions to the property additivity
scheme come from nitrogen atoms that can change their
basicity, their polar surface area, their number of donors,
etc. If this becomes an issue for a library, the reagents can
be initially split according to each case and a different offset
used.
3. When the partitioning scheme is used, only a small subset of
the products is examined and the goodness is then defined
as the fraction of the examined products with the desired
product properties.

Acknowledgments
The author thanks Dr. Christopher Bayly from Merck Frosst
Canada for his initial important contribution to GLARE and for
a careful proofreading of this chapter.
References
1. Gillet, V. J. (2008) New directions in library
design and analysis. Curr Opin Chem Biol 12,
372378.
2. Song, C. M., Bernardo, P. H., Chai, C. U.,
Tong, J. C. (2009) CLEVER: Pipeline for
designing in silico chemical libraries. J Mol
Graph Model 27, 578583.
3. Truchon, J. -F., Bayly, C. I. (2006) Is there
a single Best Pool of commercial reagents
to use in combinatorial library design to conform to a desired product-property profile?
Aust J Chem 59, 879882.
4. Truchon, J. -F., Bayly, C. I. (2006)
GLARE: a new approach for filtering
large reagent lists in combinatorial library
design using product properties. J Chem
Inf Model 46, 15361548. http://glare.
sourceforge.net
5. Conde-Frieboes, K., Schjeltved, R. K., Breinholt, J. (2002) Diastereoselective synthesis of

2-aminoalkyl-3-sulfonyl-1,3-oxazolidines on
solid support. J Org Chem 67, 89528957.
6. Feuston, B. P., Chakravorty, S. J., Conway, J. F., Culberson, J. C., Forbes, J.,
Kraker, B., Lennon, P. A., Lindsley, C.,
McGaughey, G. B., Mosley, R., Sheridan, R.
P., Valenciano, M., Kearsley, S. K. (2005)
Web enabling technology for the design,
enumeration, optimization and tracking of
compound libraries. Curr Top Med Chem 5,
773783.
7. Shi, S. G., Peng, Z. W., Kostrowicki, J.,
Paderes, G., Kuki, A. (2000) Efficient combinatorial filtering for desired molecular properties of reaction products. J Mol Graph
Model 18, 478496.
8. Lipinski, C. A., Lombardo, F., Dominy, B.
W., Feeney, P. J. (1997) Experimental and
computational approaches to estimate solubility and permeability in drug discovery and

346

Truchon

development settings. Adv Drug Deliv Rev

23, 325.
9. Klopman, G., Li, J. Y., Wang, S. M.,
Dimayuga, M. (1994) Computer automated
log P calculations based on an extended
group-contribution approach. J Chem Inf
Comput Sci 34, 752781.
10. OpenEye Scientific Software Inc OEChem
Toolkit, Santa Fe, NM, USA, 2009.
www.eyesopen.com

11. The Molecular Operating Environment

(MOE), Chemical Computing Group
Inc., Montreal, QC, Canada, 2008. www.
chemcomp.com
12. JOELib a Java based cheminformatics
library, version 2; University of Tuebingen, Tuebingen, Germany, 2009. http://
sourceforge.net/projects/joelib

Chapter 18
CLEVER: A General Design Tool for Combinatorial Libraries
Tze Hau Lam, Paul H. Bernardo, Christina L.L. Chai,
and Joo Chuan Tong
Abstract
CLEVER is a computational tool designed to support the creation, manipulation, enumeration, and
visualization of combinatorial libraries. The system also provides a summary of the diversity, coverage,
and distribution of selected compound collections. When deployed in conjunction with large-scale virtual
screening campaigns, CLEVER can offer insights into what chemical compounds to synthesize, and,
more importantly, what not to synthesize. In this chapter, we describe how CLEVER is used and offer
advice in interpreting the results.
Key words: Virtual combinatorial library, Markush technique, compound analysis, chemoinformatics, chemistry.

1. Introduction
Combinatorial chemistry has become increasingly essential in the
modern drug discovery pipeline (1, 2). Through the discovery of
new chemical reactions and commercially available reagents, the
size of these libraries has amplified exponentially over the past few
years (3). Often, such libraries are far too large to be synthesized
and screen in their entirety. Moreover, the output frequently produces high level of redundancy in terms of the similarity in the
physiochemical properties of the derived compounds. Therefore,
a rational approach for combinatorial library design is desirable
in order to maximize the outcome of an expensive synthesis and
screening campaign (4).
Here we introduce CLEVER (Chemical Library Editing,
Visualizing, and Enumerating Resource), a platform-independent
J.Z. Zhou (ed.), Chemical Library Design, Methods in Molecular Biology 685,
DOI 10.1007/978-1-60761-931-4_18, Springer Science+Business Media, LLC 2011

347

348

Lam et al.

tool that allows not only the enumeration of chemical libraries

using customized fragments but also the computation of the
physicochemical properties of the generated compounds along
with filtering functionalities for evaluating their drug likeness.
CLEVER may also be used for visualizing the generated chemical
compounds in 3D space, as well as charting various graphs based
on the innate properties of the chemical libraries. The system is
available at http://datam.i2r.a-star.edu.sg/clever/.

2. Materials
1. Java version 1.6 and above.
2. SmiLib v2.0 (5) for rapid combinatorial library generation in Simplified Molecular Input Line Entry Specification
(SMILES) (6).
3. The Chemistry Development Kit (CDK) Application Programming Interface (API) (7), OpenBabel (8), or CORINA
(9) for generating 3D coordinates (SDF format) from
SMILES strings.
4. Jmol (10) for interactive display of molecular structures in
3D space.
5. JFreeChart (http://www.jfree.org/jfreechart/) for generating histograms and 2D scatter plots for chemical compound
analysis.

3. Methods
CLEVER is implemented using the Java 3D API (see Section
4.1). The main framework is made up of five key modules for
chemical library editing, enumeration, conversion, visualization,
and analysis. The operations of these functionalities are accomplished by the various applications at the resource layer. For the
purpose of illustration, the compound calothrixin B, a secondary
metabolite isolated from the Calothrix cyanobacteria (1113), is
used as the scaffold molecule with the variable functional groups
[Rn] attached (Fig. 18.1). The calothrixins are redox-active natural products which display potent antimalarial and anticancer
properties and thus there is interest in probing the physical as
well as biological profiles of their derivatives (14). In this exercise,
six functional groups have been selected as the building blocks
(Table 18.1).

CLEVER

349

Fig. 18.1. Compound CID: 9817721 and its corresponding scaffold structure for
enumerating novel library.

Table 18.1
SMILES string configuration for scaffold and building blocks

3.1. Data Preparation

Scaffold

SMILES

S1

O=C(C(C(C=C([R3])C=C1)=C1N=C2)=
C2C3=O)C4=C3C5=CC=C([R2])C([R1])
=C5N4

Attachment blocks

SMILES

B1

C[A]

B2

C(C)(C)([A])

B3

F[A]

B4

CC[A]

B5

C=C[A]

B6

C1=CC=CC=C1[A]

1. Use the library editor to create a library file for the compounds under study (Fig. 18.2). Library files are essentially plain text files that contain a record on each line,
with an entry identifier and a SMILES string for the

350

Lam et al.

Fig. 18.2. Illustration of the library editor.

scaffold or building blocks (delimited by a tab character) (see

Section 4.2).
2. Define the chemical scaffolds, attachment blocks, linkers,
and reaction schemes for the compounds under study.
Attachment points on the blocks are represented by [A],
while functional groups to be permutated on the scaffolds
are depicted by [Rn], where n is a numerical value unique
to each functional group to be varied (Fig. 18.1). Linker
is the intersection between the scaffolds and the attachment
blocks (see Sections 4.34.6).
3. Click on the Convert SMILES button to perform the conversion of the linear SMILES strings into 3D coordinates
(SDF format). To browse automatically, click the Start
Visualizer button for the systematic viewing of the 3D
molecular structures from the chemical library (Fig. 18.3).
3.2. Chemical Library
Enumeration
3.2.1. Full Library
Enumeration

1. Click on the Enumerator tab to proceed to the library enumeration workspace.

2. Enter the library name.
3. Open both the scaffold and the building blocks text files
(Fig. 18.4a).
4. Select the appropriate scaffold and building blocks from the
scaffold and block lists (Fig. 18.4b).
5. Ensure the full combination and the empty linker options
are selected.

CLEVER

351

Fig. 18.3. CLEVER SMILES conversion and 3D structure visualization.

Fig. 18.4. Chemical library enumeration. (a) Initiation for the scaffold and block lists. (b)
Illustration on the usage of the enumerator.

352

Lam et al.

6. Click on the Enumerate Library button to start enumeration. A full enumeration will generate a new library consisting of 216 compounds derived from the systematic permutation of the variable sites with the six attachment blocks on
the core scaffold.
3.2.2. Flexible Library
Enumeration

1. Click on the Enumerator tab to proceed to the library enumeration workspace.

2. Enter the library name.
3. Open both the scaffold and the building blocks text files.
4. Unclick the full combination option to enable access to userdefined reaction schemes.
5. Within the Reaction Scheme text box, define the scaffold
for each reaction scheme in the first column, followed by
pairs of linkers and blocks to be used for each attachment
site Rn, where n is a numerical value unique to each functional group to be varied (see Sections 4.34.6). For example, columns two and three denote the linker and the blocks
for the first attachment site, while columns four and five for
the second attachment site (Fig. 18.5).
6. Users can also prepare pre-defined reaction schemes for
batch upload.

3.2.3. Library
Enumeration Using
Linkers

1. Enter the library name.

2. Open both the scaffold and the building blocks text files.

Fig. 18.5. Reaction schemes definition.

CLEVER

353

Fig. 18.6. Enumeration using different linkers.

3. Unclick empty linker option to allow addition and modification of the linkers (Fig. 18.6). In this exercise, we only
demonstrate enumeration using two linkers. More linkers
could be included for chemical library construction.
3.3. Chemical Library
Analysis
3.3.1. Computation of
Physiochemical
Properties

1. Click on the Properties tab to proceed to the workspace.

2. Load and select the library for analysis.
3. Click on the Compute button to calculate physiochemical
properties including the number of hydrogen bond acceptors and donors, XlogP (partition coefficient) values, molecular weights, number of rotatable bond, and the Topological
Polar Surface Area (TPSA) of compounds.
4. User may also save the results for future reference.

3.3.2. Filtering of
Chemical Library-Based
Predefined Schemes

3.3.3. Evaluation of
Chemical Libraries

1. To initiate the filtering function, click on the Filter button,

a Filter Library window will appear.
2. User can select one of the six predefined filtering schemes
for drug likeness, lead likeness, or fragment likeness from
the Filter Scheme dropdown list (Fig. 18.7). Users may
also define their own criteria for filtering.
To analyse the distribution of chemical compounds of a certain
physiochemical property,

354

Lam et al.

Fig. 18.7. Physiochemical properties computation and the filtration of chemical libraries
based on predefined scheme.

Fig. 18.8. Distribution of compounds of a selection collection(s).

CLEVER

355

Fig. 18.9. Scatter plot for one or more libraries.

1. Select chemical collection(s) from the Available Chemical

List display space.
2. Select the Property combo list to choose a property for the
distribution graph.
3. Click on the Display Chart button to display histograms on
the distribution of chemical compounds (Fig. 18.8).
To analyse the diversity and coverage of the selected chemical
library
1. Select chemical collection(s) from the Available Chemical
List display space.
2. Select the physicochemical properties for the X and Y axes.
3. Click on the Display Chart button to show the 2D scatter
plot (Fig. 18.9).

4. Notes
1. Install a Java Virtual Machine (a runtime version of Java,
or JRE 1.6 and above). JVM is compatible to all the major
operating systems including Windows, MacOS, and Linux.
2. Ensure the input scaffold and the building block plain text
lists are saved in the .smi extension format. Any other extension formats are unrecognizable by the CLEVER enumerator and will generate an error.
3. CLEVER only allows up to a maximum of 90 [Rn] functional groups to be defined. However, there is no restriction
on the number of scaffolds, linkers, and building blocks.

356

Lam et al.

4. The [Rn] functional groups defined on the scaffolds and the

attachment points [A] groups defined on the building blocks
should not be linked to more than one atom. Examples such
as C[R1]C, C1CC[R1]1, C[A]C, and C1CC[A]1
are invalid.
5. The [Rn] and the [A] groups have to be attached to its
neighbouring atom by a single bond. Instances such as
[R1]#C, C(=[R1])C, and [A]=C are invalid.
6. SMILES format inputs with [Rn] groups attached to atoms
with E/Z isomerism specification are not allowed.
Examples such as [R1]/C=C(F)/I and Br/C(Cl)=
C(O/C=C/F)/[R1] are invalid.
References
1. Martin, E. J., Critchlow, R. E. (1999)
Beyond mere diversity: tailoring combinatorial libraries for drug discovery. J Comb Chem
1, 3245.
2. Valler, M. J., Green, D. (2000) Diversity
screening versus focussed screening in drug
discovery. Drug Discov Today 5, 286293.
3. Jamois, E. A. (2003) Reagent-based and
product-based computational approaches in
library design. Curr Opin Chem Biol 7,
326330.
4. Leach, A. R., Hann, M. M. (2000) The in
silico world of virtual libraries. Drug Discov
Today 5, 326336.
5. Schller, A., Hhnke, V., Schneider, G.
(2007) SmiLib v2.0: a Java-based tool
for rapid combinatorial library enumeration.
QSAR Comb Sci 26, 407410.
6. Weininger, D. (1988) SMILES, a chemical
language and information system. 1. Introduction to methodology and encoding rules.
J Chem Inf Comput Sci 28, 3136.
7. Steinbeck, C., Hoppe, C., Kuhn, S., Floris,
M., Guha, R., Willighagen, E. L. (2006)
Recent developments of the chemistry development kit (CDK)an open-source java
library for chemo- and bioinformatics. Curr
Pharm Des 12, 21112120.
8. Guha, R., Howard, M. T., Hutchison, G.
R., Murray-Rust, P., Rzepa, H., Steinbeck,
C., Wegner, J., Willighagen, E. L. (2006)
The Blue Obeliskinteroperability in chem-

9.

10.
11.

12.

13.

14.

ical informatics. J Chem Inf Model 46,

991998.
Sadowski, J. (1997) A hybrid approach for
addressing ring flexibility in 3D database
searching. J Comput Aided Mol Des 11,
5360.
Angel, H. (2006) Biomolecules in the computer: Jmol to the rescue. Biochem Educ 34,
255261.
Rickards, R. W., Rothschild, J. M., Willis, A.
C., de Chazal, N. M., Kirk, J., Kirk, K., Saliba, K. J., Smith, G. D. (1999) Calothrixins
A and B, novel pentacyclic metabolites from
Calothrix Cyanobacteria with potent activity
against malaria parasites and human cancer
cells. Tetrahedron Lett 55, 1351313520.
Bernardo, P. H., Chai, C. L. L., Heath, G.
A., Mahon, P. J., Smith, G. D., Waring,
P., Wilkes, B. A. (2004) Synthesis, electrochemistry, and bioactivity of the Cyanobacterial Calothrixins and related quinones. J Med
Chem 47, 49584963.
Bernardo, P. H., Chai, C. L. L., Le Guen,
M.,Geoffrey D., Smith, G. D., Waring,
P. (2006) Structureactivity delineation of
quinones related to the biologically active
Calothrixin B. Bioorg Med Chem Lett 17,
8285.
Khan, Q. A., Lu, J., Hecht, S. M. (2009)
Calothrixins, A new class of human DNA
topoisomerase I poisons. J Nat Prod 72,
438442.

SUBJECT INDEX

Adamantyl amide. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .191212

ADME&T (Adsorption, Distribution, Metabolism,
Excretion, and Toxicity) . . . . . . . . 297, 303, 307,
314, 316, 324325, 328, 331, 333334
ADMET Predictor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
Affymaxs thiolacyl library . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
AGDOCK . . . . . . . . . . . . . . . . . . . . . . 193, 195196, 202, 326
Agglomerative clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
Algorithm
computer algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
deterministic annealing . . . . . . . . . . . . . . . . . . . . . . . 7577
evolutionary algorithm . . . . . . . . . . . . . . . . . . . . 5556
genetic algorithm . . . . . . . . . . . . . . . . . . 56, 120, 137, 140,
144, 316
multi-objective evolutionary algorithm . . . . . . . . . . . . . 58
Alignment-based . . . . . . . . . . . . . . . . . . . . . . . . . 122123, 125
Alignment-free . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122125
Analysis tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40, 122
Angiotensin converting enzyme (ACE) . . . . . . . . . . . . 1214
Antagonist . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1921, 117,
126, 128
Applications . . . . . . . . . . . . . . . . . 91107, 111129, 140147,
268270, 279290
Aqueous solubility . . . 33, 140, 144, 196197, 229, 236, 326
Aromaticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284
2-Aryl indole . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1516
Arylamine N -acetyltransferase . . . . . . . . . . . . . . . . . . . . . . 128
Asymmetric similarity score . . . . . . . . . . . . . . . . . . . . 262, 273
Available Chemicals Directory (ACD) . . . . . . . . . . . . . . . 114,
117, 138, 142, 159, 164, 168, 177178, 183,
197, 207208, 227, 322, 338, 341

Caco-2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
Calculations . . . . . . . . . . 29, 60, 64, 105, 122123, 164, 177,
193194, 196, 210, 225, 227, 231, 245,
297298, 302, 304307, 311, 327, 329, 334,
340341
Carbo index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
Catalyst . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
CDK2 . . . . . . . . . . . . . . . . . . 18, 232233, 322, 326, 329, 331
Cell-based . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
Cell-based partitioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
Centroid . . . . . . . . . . . . . . . . . . . . . . . . . . . 7778, 82, 228229
Chem-Diverse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
Chemical library . . . . . . . . . . . . . . 324, 2728, 48, 111130,
156, 165, 167168, 180, 231, 295317, 337,
340, 347348, 350355
Chemical reactions . . . . . . . . . . . . 29, 31, 165, 167, 179, 188,
254, 270, 272, 301302, 304, 324, 347
Chemical representation . . . . . . . . . . . . . . . . . . . . . . . . . . 2832
Chemical space . . . . . . . . . . . . . 28, 3340, 4345, 48, 54, 62,
102, 106, 115, 136, 156, 167, 170, 220, 236,
242, 254, 257, 271274, 286, 316
Cheminformatics . . . . . . . . . . . . . . . . 112113, 129, 296, 341
Chemistry
combinatorial . . . . . . . . . . 5, 45, 54, 71, 77, 91, 106, 112,
156, 163, 167, 170, 175176, 245, 255, 298,
321, 326, 347
high throughput . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
medicinal . . . . . . . . . . 4, 16, 45, 115, 135136, 286, 333
Chemoinformatics . . . . . . . . . . . . . . . . . . . . . . 2749, 57, 176
Cherry-picking . . . . . . . . . . . . . 94, 112, 225, 298, 302, 306,
309, 314, 325
Chk1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321335
Chromosome . . . . . . . . . . . . . . . . . . . . . . . . 5960, 65, 68, 140
Chronic myelogenous leukemia (CML) . . . . . . . . . . 92, 282
cLipE (calculated lipophilic efficiency) . . . . . . . . . . 205206
cLogD (calculated LogD) . . . . . . . . . . . . . . . . . . . . . 198, 206,
208209
cLogP . . . . . . . . . . . 7, 11, 32, 197, 205207, 225, 236237,
243, 287, 322, 326327, 329
Cluster . . . . . . . . . 3839, 6061, 66, 68, 7374, 7782, 84,
88, 144145, 149, 229, 287, 290, 303, 325
Clustering . . . . . . . . . . . 39, 4344, 60, 66, 68, 88, 137, 160,
229, 232, 284, 286
Collaborations . . . . . . . . . . . . . . . . . . . . . . . . . . . 57, 254, 332
COMBIBUILD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166167
CombiDock . . . . . . . . . . . . . . . . . . . . . . . . . 161, 163, 167168
CombiGlide . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166167, 180
CombiLibMaker . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159, 180
Combinatorial . . . . . . . . . . . 45, 79, 16, 2223, 3940, 43,
4547, 54, 7188, 91107, 112, 114, 117, 128,

B
Basis Product . . . . . . . . . . . . . . . 166167, 259263, 273, 316
BCUT descriptors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
Binary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4, 35, 114, 136, 140,
307, 345
Binding mode annotation . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
Bioactivity data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
Bioinformatics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
Biological activity . . . . . . . 4, 8, 1718, 40, 9798, 101, 111,
114115, 121, 124
Biologically active compounds . . . . . . . . . 3, 8, 113114, 128
Bond order . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
Builder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162, 165, 167, 245
Building . . . . . . . . . . . . . 4, 1011, 28, 3841, 44, 47, 57, 59,
6162, 64, 6667, 97101, 106, 112, 137, 156,
179180, 220222, 224230, 245246, 273,
348350, 352, 355356

J.Z. Zhou (ed.), Chemical Library Design, Methods in Molecular Biology 685,
c Springer Science+Business Media, LLC 2011
DOI 10.1007/978-1-60761-931-4, 

357

CHEMICAL LIBRARY DESIGN

358 Subject Index

135150, 156, 159, 161, 163164, 166169,

175176, 180, 193, 224225, 245, 254255,
258259, 261, 272, 274, 286, 296, 298302,
304, 305, 309310, 314315, 321, 325327,
330, 333334, 337345, 347356
Combinatorial chemistry . . . . . . . . . . . 5, 45, 54, 71, 77, 91,
106, 112, 156, 163, 167, 170, 175176, 245,
255, 298, 321, 326, 347
Combinatorial explosion . . . . . . . . . . . . . . . . . . . . . . . 161, 343
Combinatorial library . . . . . . . . . . . . 89, 3940, 43, 4547,
7188, 91107, 128, 135150, 159, 161, 163,
166167, 175176, 180, 224225, 272, 296,
298, 302, 304, 310, 325, 334, 337345,
347356
Combinatorial library design . . . . . . . . . . . . . . . . . 3940, 45,
7188, 91107, 135150, 159, 175176, 296,
304, 325, 347
Combinatorial optimization . . . . . . . . . . . . . . . . . . 22, 72, 77
CombiSMoG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166167
Complexity . . . . . . . . . 11, 34, 37, 4041, 54, 57, 63, 72, 77,
138, 226, 228, 230, 233235, 243, 266, 274
Components . . . . . . . 12, 3738, 43, 4546, 55, 80, 86, 101,
115116, 129, 166, 225, 245, 249, 258259,
262264, 296305, 309310, 313, 329
Compound analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348
Computational filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
Computational model . . . . . . . . . . . . . . 32, 44, 207, 326, 334
Computational tool . . . . . . . . . . 28, 177, 194, 196197, 245
CONFIRM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
Conformation . . . . . . . . . 126127, 136, 158, 181, 183, 188,
195196, 200201, 230, 280, 281, 289, 316
Conformational search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
Connection table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2930
Consensus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118119, 236
Constitutional . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3436, 86
Conversion . . . . . . . . . . . . . . . . . . . . . . 42, 192, 348, 350351
CORINA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348
Correlation . . . . . . . . . . 5, 7, 94, 96, 99101, 115116, 118,
129, 181, 229, 244
Crossover . . . . . . . . . . . . . . . . . . . . . . . . . 5556, 59, 6162, 64
Cross-validation/ed . . . . . . 99101, 107, 118119, 178, 227
Cytochrome P . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

D
Database . . . . . . . . . . . . . . . . . 5, 32, 112, 118119, 124126,
128129, 137138, 141144, 149, 157,
168170, 177178, 180, 183, 226227, 229,
254255, 260, 262, 280, 284285, 290, 300,
313, 321, 326, 331
Data mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28, 3234, 117
tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Daylight . . . . . . . . . . . . 31, 34, 125, 137, 274, 284, 286, 314
fingerprints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34, 286
Degrees of freedom . . . . . . . . . . . . . . . . . . . . . . . . . . . 195, 210
De novo design . . . . . . . . . . . . . . . . . . . . . . . 58, 177, 245, 290
Dependent variable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9798
Descriptors . . . . . . . . . . . . 28, 3338, 4041, 60, 65, 86, 93,
9798, 101, 103, 106, 113114, 118120,
122125, 129, 139, 207, 227, 273
Design
approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
based library . . . . . . . 137, 155170, 175187, 261, 298,
301302, 316
chemical library . . . . . . . . . . . . . . . 323, 28, 48, 111129
Desktop tool . . . . . . . . . . . . . . . . . . . . . . . . 192, 295317, 321

Determinant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
Deterministic annealing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7577
2D fingerprints . . . . . . . . . . . . . . . . . . . . . . . . . 34, 3839, 128
3D fingerprints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
Diaminopyrimidine . . . . . . . . . . . . 95, 97, 99100, 102105
Dimensionality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3738
Dimension reduction . . . . . . . . . . . . . . . . . . . . . 28, 3438, 40
Directory . . . . . . . . . . . . . . . . . . . . . . . . . . . 159, 227, 315, 338
Disassembly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
Discriminant analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
Dissimilarity . . . . . . . . . . . . . . 3740, 65, 142, 145, 147, 149
Distance range . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
Diverse libraries . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44, 66, 145
Diversity
analysis . . . . . . . . . . . . . . 39, 60, 136, 162, 177178, 226
library . . . . . . . . . . . . . . . . . . 142, 145146, 148149, 176
Diversity oriented synthesis (DOS) . . . . . . . . . . . 5, 7, 1112
DOCK . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163, 165, 167, 178
Docking . . . . . . . . . 4344, 63, 67, 104105, 125, 155170,
176178, 180183, 187, 193196, 198,
200210, 245, 271272, 280, 283, 303, 306,
316, 326, 329, 331, 333334
3D pharmacophore . . . . . . . . . . . . . . . . . . 231, 269, 271, 306
DRAGON . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
DREAM++ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165, 167
Drug discovery . . . . . . . . . . . . 45, 8, 28, 3233, 4144, 48,
5354, 58, 7172, 86, 113, 121, 126128,
135136, 155159, 170, 181, 219, 227, 236,
242, 255, 271, 273, 275, 296, 312313,
316317, 333, 347
Drug-likeness . . . . . . . . . . . . . . 42, 45, 56, 66, 111, 348, 353

E
EC50 . . . . . . . . . . . . . . . . . . . . . . . 18, 118, 129, 192, 205206
EGFR inhibitors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
Eigenvalue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35, 78
Electron density map . . . . . . . . . . . . . . . . . . . . . . . . . . 158, 187
Empirical . . . . . . . . . . . . . . . . . . . . 33, 35, 114, 135, 163, 208
Encoding . . . . . . . . . . . . . . . . . . . . 5, 59, 6263, 66, 124, 138,
200, 302, 314
Enrichment factor . . . . . . . . . . . . . . . . . . . . . . . . 125, 127, 168
Enumeration . . . . . . . . . . . . . 4647, 72, 102, 162, 164, 167,
177179, 187188, 196, 254, 256258, 261,
263, 272, 296, 298, 300, 314, 325326,
328329, 334, 348, 350353
Enzyme
inhibition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
selectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23, 94
Erlotinib . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
Euclidean distance . . . . . . . . . . . . . . . . . . . . . . . . . . . 37, 65, 78
Evaluation
algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5556
programming . . . . . . . . . . . . . . . . . . . . 193, 195, 200, 210
Excretion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54, 156, 297

F
Features . . . . . . . . . . . . . . . . . . . . 8, 30, 63, 77, 119, 129, 136,
146, 163, 179, 194, 228229, 243, 261, 285,
297, 299, 309311
Filtering . . . . . . . . . . . . . . . . . . . . . 43, 47, 59, 63, 65, 67, 139,
158160, 162, 176177, 181, 187, 197, 208,
224, 228229, 305307, 309, 325, 327, 329,
331, 337339, 342343, 348, 353
integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307

COMPUTATIONAL LIBRARY DESIGN

359
Subject Index
Filters . . . . . . . . . . . 4243, 56, 5960, 62, 64, 67, 169, 177,
179181, 225227, 229, 286, 315
Fingerprints . . . . . . . . . . . . . . . 3436, 3839, 135149, 228,
255, 258, 263, 268, 273, 286, 326
FIRM/organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5, 270
FlexX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166, 178, 180
FlexXc . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
Focused
libraries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44, 156, 167
library design . . . . . . . . . . . . . . . . . . . . . . . . . 178180, 182
Focusing . . . . . . . . . . . . . . . . . . . . . . . 241, 253, 259, 274, 331
Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
Fragment based drug design . . . . . . . . . . . . . . . . . . . 241250
Fragment based lead discovery . . . . . . . . . . . . . . . . . 219238
Fragment screening . . . . . . . . . . . . . . . . . . 219221, 224227,
230231, 236, 238, 242245, 249
FRED . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
Free-Wilson . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41, 91107
Functions . . . . . . . . . . . . . . . 14, 33, 40, 44, 56, 76, 122123,
144, 163, 170, 181, 187, 199, 221
Fuzzee . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57, 6364

G
Gastrointestinal stromal tumor (GIST) . . . . . . . . . . . . . . . 92
Gaussian functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122123
Gefitinib . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
Genetic Algorithm . . . . . . . . . . . 56, 120, 137, 140, 144, 316
Gleevec . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92, 285
Glide . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125, 170, 178, 180
GOLD . . . . . . . . . . . . . . . . . . . . 177178, 180, 182184, 186
GPCR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1517, 45
Graph theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30, 34
GROWMOL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245

H
Hamming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
HCV NS5B . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181187
hERG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32, 140, 144
High throughput chemistry . . . . . . . . . . . . . . . . . . . . . . . . 37
High-throughput screening (HTS) . . . . . . . . . . . . . . . 33, 38,
45, 58, 91, 127, 155156, 170, 175, 194,
219221, 231, 241242, 286290, 298, 317,
322, 326, 330, 332
HIPPO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
Histone deacetylases (HDACs) . . . . . . . . . . . . . . . . .117119
Hit rate . . . . . . . . . . . . . . . . . . . 118, 128, 205, 226, 231232,
235, 286287, 289
Homology model . . . . . . . . . . . . . . . . . . . . 157158, 169, 177
HOOK . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .245
HSITE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
HSP70 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231232, 234235
Human rhinovirus 3C protease . . . . . . . . . . . . . . . . . 5, 1920
Hydrogen bond acceptor (HA) . . . . . . . . . . . . . . . . . . . . . 138
Hydrogen-bond donor (HD) . . . . . . . . . 138, 146, 197, 339
11 Hydroxysteroid dehydrogenase type 1
(11-HSD1) . . . . . . . . . . . . . . . . . . . . . . . 191212

I
IC50 . . . . . . . . . . . . 14, 18, 21, 23, 32, 57, 95, 97, 119120,
182, 184, 186, 269270, 280, 285
Imatinab-resistant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
Imatinib . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92, 282
Independent variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4, 123, 145, 231

Indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
Informatics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296, 306
Inhibitor . . . . . . . . 5, 12, 1922, 58, 92, 107, 127, 168170,
182186, 192, 246247, 250, 280283, 285,
288290, 323, 326
Iressa . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
ISIS . . . . . . . . . 262, 297, 299, 301, 304308, 314, 317, 321

J
JNK3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232233, 282

K
Kappa opioid receptor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Kinase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279290, 321334
Kinase chemical cores . . . . . . . . . . . . . . . . . . . . . . . . . 280, 290
Kinase targeted library (KTL) . . . . . . . . . . . . . 280, 283284,
287290
K-means . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
k nearest neighbor (kNN) . . . . . . . . . . . . . . . . . . 118120, 160
Knowledge-based . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21, 167

L
LCK . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322, 326
Lead hopping . . . . . . . . . . 128, 264, 268, 271, 274, 298, 315
Lead-likeness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
LEAP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253274, 298299
LEAP1 . . . . . . . . . . . . . . . . . . . . 255258, 263269, 271275
LEAP2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255264, 266273
Leave-one-out (LOO) . . . . . . . . . . . 100101, 107, 118119
LEGEND . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
Lennard-Jones . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
Library design . . . . . . 324, 2749, 5368, 7188, 91107,
111130, 135150, 155170, 175188,
191212, 219238, 241250, 253275,
279290, 295317, 321335, 337345,
347356
Library design strategies . . . . . . . . . . . . . . . . . . 137, 140, 325
Ligand efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .238, 242
LigBuilder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
LIGSITE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
Linear regression . . . . . . . . . . . . . . . . . . . . . . . . . . . 9394, 207
LipE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
Lipophilic groups (LIP) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
LogD . . . . . . . . . . . . 197198, 205209, 303, 326, 329, 331
LogP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
LUDI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245

M
MACCS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
Markush . . . . . . . . . . . . 4647, 297298, 301, 305306, 314
Markush exemplification . . . . . 46, 297, 301, 305306, 314
Markush technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
MCSS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
MDL . . . . . . . . . . . . . 30, 258, 262263, 274, 297, 301302,
304, 306, 309
MDL ISIS/Draw . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304, 306
MedChem . . . . . . . . . . . . . . . . . . . . . . . . . . 138, 141, 226, 286
Median . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
Medicinal chemistry . . . . . . . . . . . . . . . . . . . . . 4, 16, 45, 115,
135136, 286, 333
MEGALib . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .59, 6168
Melanin-concentrating hormone receptor 1
(MCHR1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .128

CHEMICAL LIBRARY DESIGN

360 Subject Index

Mercaptoacyl pharmacophore library . . . . . . . . . . . . . . 1213

Methods . . . . . . . . . . 28, 37, 4044, 4748, 5368, 92106,
113114, 120, 122125, 128129, 135136,
140, 155170, 176187, 197207, 219221,
224225, 230, 236238, 242, 244245,
247248, 254264, 271274, 280290, 306,
326333, 339344, 348355
Metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73, 88, 342
MLR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94, 9899
Model validation . . . . . . . . . . . . . . . . . . . . . . 41, 113115, 119
Modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114, 348
MOE . . . . . . . . . 36, 117, 119, 179180, 183, 228, 246, 341
Molar refractivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
MolConnZ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36, 118119
Molecular complexity . . . . . . . . . . . . . . . . . . . . . . . . . 228, 243
Molecular conformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
Molecular descriptors . . . . . . . . . . 28, 3335, 3738, 4041,
86, 114, 139
Molecular design . . . . . . . . . . . 7, 33, 36, 232235, 260, 274,
297, 299, 302, 306, 309, 311, 313, 316317, 321
Molecular diversity . . . . . . . . . . . . . . . . . . . . . . . . . . .5, 28, 328
Molecular dynamics . . . . . . . . . . . . . . . . . . . . 29, 44, 187, 245
Molecular graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Molecular library design . . . . . . . . . . . . . . . . . . . . . . . . . 5368
Molecular mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
Molecular similarity . . . . . . . . . . . . .28, 38, 57, 63, 255, 261,
265, 271, 273, 326327, 329, 331
Monte Carlo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44, 137, 167
MoSELECT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
MoSELECT II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
Multi-objective . . . . . . . . . . . . . . . . . . . . 5368, 338339, 345
Multi-objective evolutionary . . . . . . . . . . . . . . . . . . . . . . . . . 58
Multi-objective genetic algorithm (MOGA) . . . . . . . . . . 56
Multi-objective library design . . . . . . . . . . . . . . . . . . . . 5962
Multiobjective optimization . . . . . . . . . . . 28, 4143, 4748,
5368, 73, 76, 88
Multiple linear regression analysis (MLR) . . . . . . 94, 9899
Multi-property lead optimization . . . . . . . . . . . . . . . 191, 208
Mutation . . . . . . . . . . . . . . . . . . . 5556, 59, 61, 64, 201, 352

N
National Cancer Institute (NCI) . . . . . . . . . . . . . . . 119, 126
Negative charge centre (NEG) . . . . . . . . . . . . . . . . . . . . . . 138
Neural networks . . . . . . . . . . . . . . . . . . . . . . . . 40, 44, 75, 160
NMR . . . . . . . . . 5, 157, 176177, 187, 219220, 224238,
242243, 245249
NMR screening . . . 224225, 227230, 238, 244, 246249
Node . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
Non-dominated solution . . . . . . . . . . . . . . . . . . . . . 5455, 61
Nonoligomeric library . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
Non-small cell lung cancer (NSCLC) . . . . . . . . . . . . . . . . 92
Normal distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
NSisFragment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57, 64

O
OEChem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57, 341
OptiDock . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166167
Optimization library . . . . . . . . . . . . . . . . . . . . . . . . 1921, 333
ORIENT++ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

P
PAMPA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
Pareto . . . . . . . . . . . . . . . . . . . . . 42, 5456, 5961, 6465, 67
Pareto-optimal solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

Partial atomic charges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

Partition coefficient . . . . . . . . . . . . . . . . . . . . . 32, 4041, 353
Patents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8, 46, 297
PDE . . . . . . . . . . . . . . . . . . . . . 9397, 99100, 102104, 107
PDPK1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232233
Peptide library . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8, 10
Peptoid library . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
Pfizer global virtual library (PGVL) . . . . . . . 192198, 207,
253274, 295317, 321334
PGVL Hub . . . . . . . . . . . . . . . 192194, 196198, 207, 274,
295317, 321334
Pharmacophores
fingerprints . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34, 135149
mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
modeling . . . . . . . . . . . . . . . . 44, 111129, 269, 271, 306
Phase . . . . . . . . . . . 46, 9, 20, 59, 77, 78, 81, 156, 203, 323
PICCOLO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
Piecewise linear . . . . . . . . . . . . . . . . . 193, 195, 199200, 212
Piecewise linear potential . . . . . . . . . . . . . 195, 199200, 212
Pipeline Pilot . 258, 263, 273274, 300, 305, 311, 315, 326
Platform . . . . . . . . . . . . . . . . 63, 93, 227, 236, 248, 297, 313,
316317, 338, 341, 347
POCKET . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
4-Point pharmacophores . . . . . . . . . . . . . . . . . . . . . . . 136, 138
Polar surface area (PSA) . . . . . . . . . . . . 32, 34, 48, 197198,
207, 221, 237, 243, 340, 345, 353
Positive charge centre (POS) . . . . . . . . . . . . . . . . . . . . . . . 138
Potential . . . . . . . . 4, 14, 16, 32, 44, 64, 71, 81, 91, 93, 102,
112113, 120, 124, 140, 160, 164, 168, 175,
195, 199200, 207, 212, 229, 241242,
280282, 302, 312, 316, 342
Prediction . . . . . . . . . . . . . . . 33, 97, 100, 113, 115116, 119,
121, 137, 195, 208, 229, 314315
Predictors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116, 196, 207
Principal component analysis (PCA) . . . . . . . . . . . . . . 37, 86
Pro Ligand . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164, 245
Pro Select . . . . . . . . . . . . . . . . . . . . . . . . . . . 161, 163, 167168
Probabilities . . . . . 47, 54, 59, 61, 64, 7779, 139, 168, 262
Product
basis . . . . . . . . . . . . . . . . . . . 166167, 259263, 273, 316
properties . . . . . . . . . 137, 140, 298, 309, 311, 314, 325,
337343, 345
Profile
activity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
biological . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348
perfect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
property . . . . . . . . . . . . . 33, 45, 137138, 140, 144145,
197, 207208, 325
selectivity . . . . . . . . 94, 96, 102107, 280, 322, 328, 332
Property-encoded shape distributions (PESD) . . . 123124
Property-encoded surface translator (PEST) . . . . . 123124
ProSAR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137149
Protein kinase . . . . . . . . . . . . . . . . . . . . . . . . . . . 92, 9596, 99,
102105, 229, 279282, 290
Protein-ligand complex . . . . . . . . . . 177, 181, 195, 198, 326
Protein-ligand docking . . . . . . . . . . . . . . . . . . . . . . . . 329, 334
PubChem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64, 254
Purine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13, 16, 18
Pyrazolopyrimidine . . . . . . . . . . . . . . . . . 9697, 99, 101106
Pyrrolopyrazole . . . . . . . . . . . . . . . . . . . . . . . . . 95, 97, 99103
Pyrrolopyrimidine . . . . . . . . . . . . . . . . . . . . . . . 95, 97, 99103

Q
Quantitative structure activity relationship (QSAR)
methods . . . . . . . . . . . . . . . . . . . . . . . . . . 28, 113114, 120

COMPUTATIONAL LIBRARY DESIGN

361
Subject Index
modeling . . . . . . . . . . . . . . 3334, 4041, 4344, 94, 98,
101102, 107, 112121, 129
Quantitative structure property relationship
(QSPR) . . . . . . . . . . . .28, 3334, 36, 4041, 115
Quinazoline . . . . . . . . . . . . . . . . 92, 95, 97, 99100, 102103

R
Raf . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2124
Random library . . . . . . . . . . . . . . . . . . . . . . 8, 10, 12, 144145
Rapid Overlay of Compound Structures (ROCS),
122123, 125127
REACT++ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
Reactant . . . . . . 46, 196, 257260, 272, 297302, 304311,
314315, 327329, 332334
Reaction transform . . . . . . . . . . . . . . . . . . . . . . . . . . 5, 31, 259
Reagent selections . . 56, 137, 139142, 144, 176178, 338
REALISIS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296, 313314
RECAP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57, 61, 272273
Regression . . . . . . . . . . . . . . . . 9395, 98101, 116, 207, 227
Renal cell carcinoma (RCC) . . . . . . . . . . . . . . . . . . . . . 92, 282
Research and development . . . . . . . . . . . . . . . . . . . . . . . . . 155
Review . . . . . . . . . . . . . . . . 28, 112, 115, 117, 126, 136, 157,
159160, 219, 221, 225226, 229, 236, 238,
245, 255, 286, 311
Rings . . . . . . . . . . . . . . . . . . . . . 1213, 1516, 22, 3132, 34,
105106, 147, 165, 182, 227228, 233234,
243, 248, 260, 272, 284, 286, 288
Root node . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
Rule-of-five . . . . . . . 42, 48, 6364, 303, 310, 326, 331, 340
Rxn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297, 301, 304

S
Scaffold . . . . . . . . . . 13, 1516, 66, 120, 167, 178, 182186,
349352, 355356
Scaffold hopping . . . . . . . . . . . 125, 127129, 136137, 177,
254, 267
Scalable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7188
Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37, 80, 199, 344
SciTegic . . . . . . . . . . 257, 274275, 300, 305, 311, 315, 326
Scoring . . . . . 44, 59, 63, 112, 128, 159161, 163, 170, 176,
180181, 184, 187188, 201, 212, 273, 306,
316, 326, 333, 334
Screen . . . . . . . . 89, 91, 113, 219220, 225226, 228231,
233, 235236, 248, 280, 286287, 289, 303,
306, 322, 324, 329, 331, 347
Screening collection . . . . . . . . . . . . . . . . . . . . . . 219238, 243
SEARCH++ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
Search . . . . . . . 144, 200201, 257, 262263, 265266, 268,
271274, 305, 315
Searching . . . . . . . . . . . . . . . . 42, 45, 126, 194, 271, 311, 314
SeeDs . . . . . . . . . . . . . . . . . . . . . . . . . . 227, 229230, 232, 258
Selection . . . . . 3940, 4748, 5962, 64, 68, 72, 112, 118,
139141, 307308, 325329, 339340, 354
Selective library design . . . . . . . . . . . . . . . . . . . . . . . . . . . 6366
Selectivity . . . . .89, 15, 1924, 54, 5758, 6364, 91107,
280281, 286287, 322, 326330, 332334
Shannon entropy . . . . . . . . . . . . . . . . . . . . 139, 144145, 147
SHAPE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126127
Shape complementarity . . . . . . . . . . . . . . . . . . . . . . . . 121122
Similarity . . . . . . . . . . . . . . 28, 34, 3740, 4245, 4748, 54,
5657, 6365, 67, 103, 112, 115, 121129, 137,
142145, 147, 149, 158, 169170, 183, 194,
254259, 261266, 268269, 271274, 284,
290, 298, 305306, 315316, 325327, 329,
331, 333, 347

Similarity coefficient . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3839

Similarity search . . . . . . . 121, 126, 137, 142, 144, 254258,
262263, 271, 273274, 298, 315316
Simulated annealing . . . . . . . . . . . . . . . . . . . . 56, 75, 137, 195
Singleton . . . . . . . . . . . . . . . . . . . . . . . . . . 47, 68, 72, 295317,
322, 333334
SkelGen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
SLogP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233234
SMARTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31, 138, 180,
285287, 314
SMARTS query . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
SMILES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2932, 139, 228,
348351, 356
SMoG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167, 245
Software . . . . . . . . . . . 3436, 46, 5657, 65, 136, 156, 159,
167, 170, 177, 179180, 208, 225, 245246,
250, 271, 275, 296297, 299300, 303, 306,
309, 312313, 315, 321, 326, 329, 334, 338, 341
Software deployment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
Software tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159, 167
Solubility . . . . . . . . 3234, 40, 48, 140, 144, 156, 192194,
196198, 206211, 221, 225, 227, 229, 236,
242243, 246247, 249250, 322, 326,
328330, 332
Spotfire . . . . . . 198, 208, 299, 306309, 315, 317, 321, 324
SPROUT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
Statine pharmacophore library . . . . . . . . . . . . . . . . . . . . 13, 15
Statistical partitioning methods . . . . . . . . . . . . . . . . . . . . . . 40
Streamline . . . . . . . . . . . . . . . . . . . . . . . . . . 295317, 321, 333
Structure-activity relationship (SAR) . . . . . . . . . . . 5, 21, 23,
2728, 33, 40, 42, 93, 97, 112, 115, 129,
135137, 139, 146147, 149, 224, 226227,
237, 280, 297298, 303, 306307, 314, 316,
326327, 330334
Structural alerts (STA) . . . . . . . . . . . . . . . . . . . . . . . . 197, 207
Structural keys . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
Structure-based
design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185, 261
drug design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122, 175
library design . . 155170, 175188, 191212, 261, 316
Structures
3D . . . . . . . . . . . .121, 129, 157159, 177, 261, 273, 351
array . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
core . . . . . . . . . . . 180, 196, 202, 284285, 290, 297, 306
crystal . . . .103104, 136, 157158, 168170, 176177,
181, 193194, 196, 200201, 211, 280281,
284, 290, 322324, 326, 333
data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58, 297, 299302
molecular . . . . 2932, 40, 178, 256, 266, 300301, 303,
307309, 315, 329, 348, 350
protein . . . . . . . . 157, 177, 201202, 245, 306, 316, 326
searching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
X-ray . . . . . . . . . . 94, 104, 106, 126127, 177178, 184,
256, 280, 282, 288, 329, 331, 335
Subgraph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57, 68
Subsetting . . . . . . . . . . . . . . . . . . . . . . . 33, 283, 287288, 290
Substituent constant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Substructure searching . . . . . . . . . . . 183, 197, 290, 305, 311
Summary . . . . 121, 147, 255256, 271272, 332333, 344
Sunitinib . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
Support vector machines (SVM) . . . . . . . 44, 118119, 160
Surflex-Dock . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178, 180
SURFNET . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
Sutent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
Symmetric similarity score . . . . . . . . . . . . . . . . . . . . . 262, 273
Symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261263, 273

CHEMICAL LIBRARY DESIGN

362 Subject Index

Symyx . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30, 159, 338

Synthesis protocol . . . . . . . . . . 268269, 286, 299, 321, 325,
327328, 330
Systematic Elaboration of Libraries Enhanced by
Computational Techniques
(SELECT). . . . . . . . . . . . .56, 161, 163, 167168

T
Tanimoto . . . . . . . . . . . . 3839, 63, 103, 128129, 137, 142,
144145, 147, 149, 232, 255, 268, 274, 284, 326
Tanimoto coefficient . . . . . . . . . . . . . . 38, 129, 232, 274, 326
Tarceva . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
Targeted . . . . . . . . . . . . . . . 8, 1119, 92, 102, 111130, 192,
226, 243, 269270, 279290, 298299, 315,
321335, 343
Targeted library . . . . . . . . 8, 1214, 19, 112, 129, 192, 226,
269270, 279290, 298299, 321335
Tautomers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .158159
Techniques . . . . . 4, 8, 16, 21, 54, 5758, 6061, 71, 92, 97,
99101, 106, 114115, 117, 126, 141, 156157,
160, 163, 195, 200, 224,
226227, 230, 236, 242246
Thiazolone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182184, 186
Tools . . . . . . . . . . . . 2728, 33, 40, 4648, 62, 71, 113122,
125126, 137, 156, 159, 167, 176177, 192,
194, 196197, 202, 245, 295317, 321335,
337345, 347356
Topological pharmacophore . . . . . . . . . . . . . . . . . . . . 136139
Toxicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5354, 205, 297
Training and test sets . . . . . . . . . . . . . . . . . . . . . 115, 117118
Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40, 44, 60
Tripos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159, 254, 272273
Tversky . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
Tyrosine kinase . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92, 170, 282

U
Undesirable functional group . . . . . . . . . . . . . . . . . . . . . . . 227

V
Validation . . . . . . . . 41, 4344, 98101, 107, 113119, 123,
125126, 129, 138, 141, 176, 178, 187, 255,
263268, 271, 274, 287288
Virtual combinatorial library . . . . . . . . . . . . . . . . . . . . 45, 128
Virtual libraries . . . . . . . . . . . . . . . 3940, 4344, 46, 56, 94,
101102, 106, 112113, 121, 128129,
166167, 178179, 181, 183, 192193,
196197, 201202, 208, 210212, 225,
253275, 296, 303, 305, 315, 340341
Virtual screening . . . . . . . . . . . . . . . 28, 34, 4344, 112122,
124129, 157, 160, 176177, 180181,
183184, 187, 195, 245, 257, 271, 316
VSA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

W
Workflow . . . . . . . . . . . . . . . 4041, 104, 114, 116119, 127,
137, 255, 283284, 290, 296297, 301,
304307, 309, 316

X
X-ray . . . . . . . . . . . . . . . . . . 94, 104, 106107, 126127, 136,
176178, 181185, 187, 194, 219, 230, 237,
242, 244249, 280, 282, 288, 322324, 329,
331, 335

Z
Zinc metalloprotease . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12