m 2 .
.5 a
AL. . .

.
...

.3.
“an“! tiiuzcfi
. V BMW“. .5433 1.4.." $17.2.
. -IrT. 1 . 1,55
. I37 a. 7.243} . A 9
£971.; is .1V 3,. 3.1... 333‘")?
; “.muiasium.» ‘ Ifi

 

 

.1

Egg em

 

 

LIBRARY
Michigan State
University

This is to certify that the
dissertation entitled
COMPUTATIONAL TECHNIQUES FOR MODELING
PROTEIN-LIGAND INTERACTIONS AND THEIR

APPLICATION TO SERINE PROTEASES AND
ASPARAGINYL-tRNA SYNTHETASE

presented by
Paul C. Sanschagrin

has been accepted towards fulfillment
of the requirements for

DOCTOR 0F PHILOSOPHYdegl-eein BIOCHEMISTRY

WA/M

Major professor

mag/z 5/0 I

MS U is an Affirmative Action/Equal Opportunity Institution 0-12771

 

PLACE IN RETURN Box to remove this checkout from your record.
To AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DATE DUE DATE DUE DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

6/01 cJCIFIC/DateDuepes-p. 15

COMPUTATIONAL TECHNIQUES FOR MODELING
PROTEIN-LIGAND INTERACTIONS AND THEIR
APPLICATION TO SERINE PROTEASES AND
ASPARAGINYL-tRNA SYNTHETASE

By

Paul C. Sanschagrin

A DISSERTATION

Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of

DOCTOR OF PHILOSOPHY

Department of Biochemistry and Molecular Biology

2001

 

ABSTRACT
COMPUTATIONAL TECHNIQUES FOR MODELING PROTEIN-LIGAND
INTERACTIONS AND THEIR APPLICATION To SERINE PROTEASES AND
ASPARAGINYL-tRNA SYNTHETASE
By

Paul C. Sanschagrz'n

This thesis describes several techniques for modeling protein-ligand interactions, in-
cluding interactions between water molecules and proteins, as well as application of these
techniques to thrombin, trypsin, and asparaginyl-tRNA synthetase. A complete-linkage hi-
erarchical cluster analysis technique for determining the degree of conservation of water
sites in a series of related protein structures is presented. This technique was applied to the
study of conserved water binding sites in the serine proteases thrombin and trypsin, with
analysis of the implications for conserved water sites in the active site and nearby sodium
ion site in thrombin. Cluster analysis was also used to identify conserved water binding
sites in bovine pancreatic trypsin inhibitor (BPTI) and in the trypsin-BPTI complex. The
conserved sites in the trypsin-BPTI complex were compared to those in trypsin and BPTI,
showing that only about half of the interfacial water sites in the complex exist in the free
form of either protein, while the remaining half are recruited or shuffled Upon complex for-
mation. The results of cluster analysis also allow inclusion of highly conserved water sites

in protein and drug design.

In addition to examination of water molecules as protein ligands, modeling sites of
favorable potential interactions with proteins in the SLIDE technique for computational lig-
and screening was improved. SLIDE is a multi-step algorithm which eliminates potential
ligand molecules from a screening database via increasingly more stringent and compu-
tationally expensive steps to yield a ranked list of potential ligands for the protein target.
Improvements made to the modeling of Sites of favorable hydrophobic interaction for both
the protein template and the potential ligand molecules are described and evaluated. Addi-
tional improvements made by my colleague Maria Zavodszky to the description of hydro-
gen bonding points in the protein template are also briefly described and evaluated. These
improvements were tested using 42 thrombin and 16 glutathione S-transferase complexes.
Both the protein template and database molecule representation improvements yield ligand
dockings that are closer to those seen in the crystallographic complex. An enrichment of
known ligands selected from a set of molecules from the Cambridge Structural Database

(CSD) is also shown.

Application of SLIDE to asparaginyl-tRNA synthetase from Brugia malayi, a human
pathogenic nematode, was performed. Screening against the CSD identified three potential
ligands of particular interest: variolin B, with possible antitumor and antiviral properties;
cercosporamide, which has known phytotoxic and fungitoxic properties; and phlorizin, a
sodium/glucose transport inhibitor. Suggested binding modes for two of 16 in vitro high-

throughput screening hits are also described.

For my wife, Suzanne. Without your support, encouragement,
understanding, and love over the years I could not have dreamed to make it

this far.

iv

ACKNOWLEDGMENTS

I would like to first thank my advisor, Dr. Leslie Kuhn for her support, encouragement,
and guidance throughout my graduate career. I consider myself lucky to have met such
an intelligent and energetic person at the right time in my education. She kindled my
interest in research and persuaded me to continue into graduate school. The opportunity to
work with such a dynamic person at this stage is one I will always treasure. I would also
like to thank the other members of my committee, Dr. Doug Gage, Dr. Michael Garavito,
Dr. Bill Punch, and Dr. John \Vilson for their guidance during my graduate tenure and for
keeping my disparate research interests on a single track. Additional appreciation goes to
Dr. Michael Kron, who piqued my interested in working with asparaginyl-tRNA synthetase
and provided valuable input on the synthetase portion of my thesis work, and to Dr. Stephen
Cusack, who provided the structure of the Brugia AsnRS. Also, thanks go to Dr. Laurie
Kaguni and Dr. Li Fan for their great assistance in the phage library project I undertook,

which went infinitely further with their help than it would have otherwise.

I would also like to thank the American Heart Association and their Michigan affiliate

for funding under grants to Dr. Kuhn.

 

I also am thankful for current and past members of the MSU Protein Structural Analy-
sis and Design group, including Carrie Barkham, Sridhar Venkataraman, Vishal Thakkar,
Trevor Barkham, Rajesh Korde, and Brandon Hespenheide for providing an exciting and
stimulating research environment and for their many ideas. I am especially grateful to
Volker Schnecke for helping me to start on the main portion of my thesis research, to Maria
Zavodszky for working with SLIDE alongside me and ensuring that I had a good grasp of
the theory and methods behind the work, and to Michael Raymer for his immeasurable help
in getting started in computational research, his innumerable ideas and careful criticisms,

and his important friendship.

Finally, I would like to thank my family, who made me realize both the importance and
the joy of learning so many years ago. I would not have come this far if it wasn’t for you.
Of special importance is the support in so many ways my wife, Suzanne, has given me over
the last 9 years. Her constant encouragement and understanding have helped me get past
the times of slow progress and difl‘icult work and have meant more to me than I could ever

express.

vi

TABLE OF CONTENTS

LIST OF TABLES x
LIST OF FIGURES xi
LIST OF ABBREVIATIONS xiii

1 Computational Docking and Screening Methods — A Review of Algorithms,

Scoring Functions, and Applications 1
1.1 Protein-Ligand Binding Sites ........................... 1
1.2 Computational Docking Methods ......................... 4
1.2.1 Ligand Manipulation Docking Methods .................... 4
1.2.2 Recombination Docking Methods ....................... 10
1.2.3 Incremental Construction Docking Methods .................. 12
1.2.4 Water in Computational Docking ........................ 14
1 .3 Computational Screening ............................. 16
1.4 Docking and Screening Scoring Functions .................... 18
1.5 Evaluation of Docking and Screening Methods ................. 26
1.6 Creation of Targeted Computational Screening Databases ............ 29
1.7 Molecular Clustering ............................... 31
1.8 Database Comparisons .............................. 33
1.9 Conforrner Generation .............................. 34
1.10 Successful Application of Docking and Screening Methods ........... 35
1.11 Motivation for this Thesis Work ......................... 36

2 Identification of Conserved Water Binding Sites in Proteins 38
2. 1 Introduction .................................... 3 8
2.1.1 The Role of Water Molecules in Proteins ................... 38

2.2 Conservation of Water Molecules among Several Crystal Structures of a Protein 41

2.3 Water Site Clustering Methods .......................... 44
2.3.] Structure Selection ............................... 44
2.3.2 Hierarchical Clustering ............................. 46
2.3.3 Crystal Contact Calculation .......................... 51
2.3.4 Evaluation of Bound Water Environments ................... 52

2.3.5 Calculation of Overlapping Microclusters between Thrombin and Trypsin . . 53

2.4 Results ....................................... 55
2.4.1 Clustering Statistics ............................... 55
2.4.2 Environmental Analysis ............................ 57
2.4.3 Effects of Crystal Contacts on Bound Water Conservation .......... 59
2.4.4 Spatial Analysis of the Conserved Microclusters ............... 59
2.4.5 Overlapping Water Sites between Thrombin and Trypsin ........... 62

2.4.6 Contribution of Conserved Water Molecules to the TrypsinzBPTI Complex . 65

2.5 Discussion ..................................... 70
2.5.1 Conservation of Water Sites in Thrombin and Trypsin ............. 70
2.5.2 Conserved Water Sites and Ligand Specificity ................. 71
2.6 Conclusions .................................... 72

3 Computational Ligand Screening -— An Improved Model of Protein-Ligand In-

teractions 73
3. 1 Introduction .................................... 73
3.2 Methods ...................................... 75
3.2.1 A General Overview of the SLIDE Method .................. 75

viii

3.2.2 Assignment of Interaction Points to Molecules in the Screening Database . . 75

3.2.3 Identification and Assignment of Protein Template Points ........... 78
3.2.4 Matching Molecular Interactions to the Template: The Screening Step . . . . 82
3.2.5 Testing Databases ................................ 96
3.3 Results ....................................... 98
3.3.1 Visual Examination of the New Template and Interaction Point Methods . . . 98
3.3.2 SLIDE Docking of Known Ligands ....................... 102
3.3.3 Improved Enrichment .............................. 115
3.3.4 Results Summary ................................ 1 19
3.4 Discussion ..................................... 120
4 Computational Screening of Asparaginyl tRNA Synthetase 125
4. 1 Introduction .................................... 1 25
4.2 Methods ...................................... 127
4.2.1 Available Asparaginyl-tRNA Synthetase Structures and Ligands for Virtual
Screening Studies ........................... 127
4.2.2 High-throughput Screening for Asparaginyl-tRNA Inhibitors and Conformer
Generation of Selected Ligands ................... 131
4.3 Results ....................................... 132
4.3.1 High-Throughput Screening .......................... 132
4.3.2 Computational Screening using a Ligand Based Template ........... 134
4.3.3 Computational Screening using an Unbiased Template ............ 140
4.4 Discussion - Analysis of Potential Ligands selected by SLIDE ......... 149
APPENDICES 156

A Summary of Publications Outside of the Scope of the Work Presented in this
Dissertation 156

BIBLIOGRAPHY 162

ix

2.1

2.2

2.3

2.4

2.5

3.1
3.2
3.3

3.4

3.5

3.6

3.7

3.8

3.9

LIST OF TABLES

Database of thrombin, trypsin, BPTI, and trypsinzBPTI structures for analysis

of conserved water sites ............................

Clustering statistics ................................

Linear correlation coefl‘icients between degree of conservation and environ-

mental features ................................

Overlapping conserved water sites between thrombin and trypsin ........

Functionally relevant conserved water sites unique to thrombin or trypsin

42 Thrombin protein-ligand complexes used for testing SLIDE modifications . .

16 GST protein-ligand complexes used for testing SLIDE modifications
Known ligand dockings for constant template point method experiments . . .

RMSDS for known ligands docked with both the original and modified interac-

tion point methods for constant template point method experiments .....

Known ligand dockings for constant interaction point assignment method ex-

periments ...................................

RMSDs for known ligands docked with both the original and modified template

method for constant interaction point method experiments .........

Known ligand dockings for combined new template and interaction point meth-

ods compared to the original methods ....................

Scores for the known ligands docked using the new template and new interac-

tion point methods compared to using the original methods .........

Enrichment factors for thrombin and GST test screening runs ..........

48

56

63

67

97

99

. 105

107

108

109

110

112

119

2.1

2.2

2.3

2.4
2.5
2.6

3.1

3.2
3.3
3.4
3.5

3.6
3.7
3.8

3.9

LIST OF FIGURES

Mobility distributions of two BPTI structures, used as a quantitative tool to
screen for structures with uncertain water positions .............. 47

Example of complete linkage clustering applied to water sites from several

BPTI structures ................................. 50
Correlation between water site conservation and environmental features for (A)

thrombin, (B) trypsin, and (C) BPTI ...................... 58
Conserved water sites in thrombin ......................... 6O
Overlapping conserved water sites between thrombin and trypsin. ....... 66
Conservation of water sites in the trypsinzBPTI interface. ............ 69

Summary of rules used to assign hydrophobic interaction points to molecules

in the screening database ............................ 77
Two example anchor fragments for an example molecule ............. 84
Hashing scheme implemented in SLIDE ...................... 85
Screening algorithm implemented in SLIDE .................... 86

Percentage of buried carbon ligand atoms in the 89 complexes used to tune the
SLIDE scoring function ............................. 92

Unbiased template for the estrogen receptor generated using the original method. 100
Unbiased template for the estrogen receptor generated using the new lks *method. 101

A comparison of hydrophobic interaction point assignment methods for (A)
estradiol from CSD code BEQJ IQ (Parrish and Pinkerton, 1999) and (B)
S-nonyl-glutathione fi'om PDB code 12gs (Oakley et a1., 1997) ........ 103

Overview of experiments t 0 test interaction modeling modifications made to
SLIDE ...................................... I 04

3.10 Dockings of estradiol to the estrogen receptor ................... 1 13

xi

3.11

3.12

3.13

4.1

4.2

4.3

4.4
4.5
4.6

4.7

4.8
4.9

4.10

4.11

4.12

4.13

Dockings of ligand BMS-182282 to thrombin ................... 114

Enrichment of known thrombin ligands in a set of random CSD molecules
using (A) the SLIDE scoring function and (B) the DrugScore scoring function. 1 16

Enrichment of known glutathione S-transferase ligands in a set of random CSD
molecules using (A) the SLIDE scoring function and (B) the DrugScore
scoring function. ............................... 118

Structure of Brugia malayi asparaginyl-tRNA synthetase complexed with S-

adenosyl-asparagine. ............................. 128
Asparaginyl-tRNA synthetase active-site pocket. ................ 130
Potential asparaginyl-tRNA synthetase inhibitors identified by high-throughput

screening. ................................... 133
Conformers generated for HTS hit 9 (A and B) and HTS hit 15 (C and D). . . . 135
'I‘wo-dimensional structure of variolin B, CSD code LEPWIM. ......... 137
Variolin B docking as determined by SLIDE .................... 138

Variolin B docking determined by superimposing the ring structure highlighted
in Figure 4.5 onto the purine ring of the AMP moiety of the S-AMP-Asn
ligand ...................................... 139
Best scoring SLIDE docking for high-throughput screening hit number seven. . 141
Best scoring SLIDE docking for high-throughput screening hit number 13. . . . 142

Molecules from the CSD selected as potential ligands by SLIDE screening with
an unbiased template .............................. 143

Docking of cercosporamide, CSD code SIVXIE, with Brugia asparaginyl-
tRNA synthetase. ............................... 145

Docking of phlorizin, CSD code CEWWACZO, with Brugia asparaginyl-tRNA
synthetase .................................... 146

Docking of CSD code MSFURY with Brugia asparaginyl-tRNA synthetase. . . 148

xii

 

ACD
ADN
AHP
AsnRS

AngrBVAL

BPTI
BVAL

CASP2

CMC
COX-2
CSD
DES
DFT

DHFR

LIST OF ABBREVIATIONS

Available Chemicals Database

atomic density

atomic hydrophilicity

asparaginyl t-RNA synthetase

average protein B-value, average B-value of the protein en-
vironment surrounding a water molecule

bovine pancreatic trypsin inhibitor

B—value, Debye-Waller factor

Second Meeting on the Critical Assessment of Techniques
for Protein Structure Prediction

Comprehensive Medicinal Chemistry (database)
cyclooxygenase-2

Cambridge Structural Database

diethylstilbestrol

discrete Fourier transform

dihydrofolate reductase

xiii

 

DME

ER

GA

GST

HTS

IleRS

LGA

MHCI

MMFF

MMP3

MOB

PDB

PLP

PMF

PPACK

PrBHBD

RMSD

SAS

S-AMP-Asn

SG-DOCK

distance matrix error

estrogen receptor

genetic algorithm

glutathione S-transferase

hi gh-throughput screening
isoleucyl-tRNA synthetase

Larnarckian genetic algorithm

major histocompatibility complex class I
Merck molecular forcefield

stromelysin-l (matrix metalloproteinase 3)
mobility, normalized protein mobility measure
Protein Data Bank

piecewise linear potential

potential of mean force
phenyl-prolyl-arginiyl-chloroketone
(Phe-Pro-Arg-chloroketone)

protein hydrogen bonds, number of hydrogen bonds formed
between a water molecule and a protein
root mean square deviation

solvent accessible surface
S—adenosyl-asparagine

similarity-guided docking

xiv

 

 

SLIDE

SP-DOCK

TPrBVAL

WatHBD

WDI

Screening for Ligands with Induced-fit Docking Efficiently
similarity-penalized docking

total protein B-value, sum of the B-values for protein atoms
in a water molecule’s environment

water hydrogen bonds, number of hydrogen bonds formed
between a water molecule and other water molecules

World Drug Index

XV

 

Chapter 1

Computational Docking and Screening Methods
- A Review of Algorithms, Scoring Functions,

and Applications

1.1 Protein-Ligand Binding Sites

Proteins perform many of the processes that are responsible for giving life to the cell.
While proteins play important structural roles, much of the interest in proteins involves the
study of their roles in catalytic and signaling events, both of which involve binding to other
molecules. The molecular partners of proteins range in size from single water molecules
and metal ions to large, multimeric protein complexes. Understanding the mechanisms of
protein binding is a key to understanding the function of proteins and the particular niche

each protein inhabits in the large, extraordinarily complex system of a living organism.

The classic view of protein-ligand binding is the “lock-and-key” concept (Fischer,

1

1894), in which the ligand “key” acts as the complement to the binding site “lock”. Part
of the interaction between protein and ligand is simply the steric fit of the two pieces, i.e.,
similar to a key fitting a lock. Of course, chemistry also plays an important in establishing
complementarity between the proteins and their ligands. Examination of protein surfaces
showed that, on average, 57% of the surface involves non-polar residues (Miller et a1.,
1987). Interaction between proteins in regions of non-polar character are predominantly
steric in nature. Given that the remaining 43% of the surface of a protein consists of polar
and charged residues, it is not surprising that interactions between these types of residues
also play a significant role in ligand binding. Since the non-polar interactions are generally
non-specific in nature, the polar and charged residue interactions provide the specificity of
interaction. A study of 15 protease-inhibitor complexes and 4 antigen-antibody complexes
determined these complexes generally form a large number of protein-ligand hydrogen

bonds (J anin and Chothia, 1990).

Since various residue types can contribute similar chemical properties to a ligand
binding site, it is interesting to examine if there is a preference for any specific residue
types in ligand-binding sites. In a study of 46 monomeric proteins, Miller and colleagues
(1987) showed that some residue types are significantly overrepresented (lysine, serine,
and glycine) compared to the overall amino acid composition of proteins, while others,
(methionine) are significantly underrepresented. A similar analysis focused on protein-
ligand binding sites in 50 crystallographic complexes was performed (Villar and Kauvar,
1994). They showed an overabundance of tryptophan, histidine, and tyrosine located close

to the ligand, indicating that these may play an important role in ligand binding. Proline

 

is found to be significantly underrepresented in the binding site relative to its abundance in
proteins overall, perhaps due to its general structural role, which may be important internal
to the protein and away from the binding site. Lysine is also significantly underrepresented
in ligand-binding sites, while it constitutes a higher than expected proportion of the protein
surface. In addition to examination of the relative abundance of each amino acid type in
ligand-binding Sites relative to the over-Al protein surface, they studied the percentage of
the binding site, defined as those amino acids within 4.0 A of the ligand, made up of each
of the amino acids. This examination showed that glycine, serine, arginine, and tyrosine
predominate in the binding Site. The distributions for the binding sites are Significantly

different from those for the protein surface and for the protein overall.

The key may not be quite the perfect match to the lock as a simple lock-and-key model
would suggest, even with the inclusion of chemistry as part of the lock and the key. Many
proteins exist which bind a variety of ligands. The major histocompatibility complex class
I (MI-1C I) binds a large variety of peptides with high aflinities (\Vrlson and Fremont, 1993).
One way to extend the “lock-and-key” concept is that of a flexible lock and/or a flexible key,
i.e., the protein and/or ligand changes conformation during ligand binding. A recent study
of 39 complexes showed that proteins generally undergo some conformational change upon
ligand binding (Betts and Stemberg, 1999), including moderate backbone conformational
changes in addition to side-chain movements. Inclusion of protein and ligand flexibility
greatly increases the complexity of computational models for protein-ligand binding. Sev-
eral docking and screening methods, including their approaches to handling flexibility, are

discussed below.

 

1.2 Computational Docking Methods

Computational docking can be described as the process of modeling the binding orientation
of a specific ligand to a specific protein of interest, i.e., a “receptor”. Docking and screening
methods (screening is described below in Section 1.3) provide for a further understanding
of the mechanisms involved in protein-ligand binding in general, as well as helping to
understand the details of the interactions in a specific protein or protein-ligand complex
of interest. In addition to gathering such knowledge for understanding the processes, this
knowledge can be used to improve the design of ligands with high specificity and high

afiinity toward a specific target.

1.2.1 Ligand Manipulation Docking Methods

The classical algorithm implemented for computational docking is that of DOCK (Kuntz
et a1., 1982; DesJarlais et a1., 1988; Shoichet et a1., 1992; Ewing and Kuntz, 1997). DOCK
operates by generating a set of spheres to describe the volume, or negative image, of the
binding site and uses the centers of these spheres as sites for matching to ligand atoms.
Sets of receptor spheres are matched to sets of ligand atoms to generate a ligand orienta-
tion, which can then be scored according to their complementarity with the protein. The
early internal DOCK scoring function, GRID (Meng et a1., 1992), is a grid-based scoring
function in the method of Goodford and colleagues (1985). Later implementations have
used more robust scoring functions which are also grid-based. It is possible to use the lig-

and docking method of DOCK with an externally supplied scoring function. The initial

 

implementation of DOCK used only steric fit and electrostatics as a determinant for ligand
docking, but later versions implemented chemical type matching to better model chemical
complementarity between ligand and receptor groups, including hydrogen bonding inter-
actions. It should be noted that DOCK uses only rigid-body translations and rotations,

including no flexibility within the docking algorithm.

Several improvements have been made to the DOCK algorithm since its inception, in-
cluding extension to include information about known ligand docking orientations through
the development of similarity-penalized docking (SP-DOCK) and similarity-guided dock-
ing (SG-DOCK; Fradera et a1. 2000). During docking with the SG-DOCK algorithm, the
docking score is weighted according to the similarity, defined by MIMIC (Mestres et a1.,
1997) to a known ligand structure and/or pharrnacophore structure, the spatial arrangement
of key ligand functional groups identified by analysis of a set of known ligands, for each
scoring during the docking. This causes the dockings to be biased towards that of the
known ligand/pharmacophore. When using SP-DOCK, the final scores of dockings per-
formed without a pharmacophore bias are weighted by the similarity, which results in a
resorting of the final docking orientations. The similarity measure plays a much more im-
portant role in the SG-DOCK procedure as it can drive the orientational and conformational

search.

Another adaptation of DOCK involves the observation that protein-protein interfaces
generally contain more hydrophobic contacts than other surface regions. Vakser and Aflalo
(1994) implemented an algorithm which reduces the model of the ligand protein surface

to include only points attributed to hydrophobic atoms. This algorithm resulted in only a

 

small improvement in docking predictions for three of the four cases tested by the authors
compared to standard DOCK, but this method uses a reduced surface representation, an
important consideration for computationally intensive docking and screening methods, and

is more tolerant to conformational changes.

Another rigid body docking procedure, developed for protein-protein docking, is the
PUZZLE algorithm (Helmer-Citterich and Tramontano, 1994), which maps protein sur-
faces into two-dimensional matrices, consisting of distances between adjacent points on
the surface along the edge of a surface slice of fixed height, and then identifies matching
submatrices. The PUZZLE algorithm has been modified to include a more comprehen-
sive scoring function and improved mapping of small protein surfaces, dubbed ESCHER
(Ausiello et a1., 1997). This algorithm operates by slicing each of the proteins, describ-
ing a polygon for each slice, and then finding complementary shape matchings for these
polygons by way of translations and rotations of one relative to the other. Alternate sets of
polygons are generated by three-dimensional translation and rotation in fixed increments.
Matching polygons are then scored based on steric and electrostatic parameters. A short-
coming of this procedure is the relatively coarse set of three-dimensional rotations em-
ployed, only taken at 10° increments. The ESCHER program is generally able to achieve
correct dockings for well-buried ligands, but fails to do so for ligands which are bound to a

shallow binding site.

Additional rigid-body docking algorithms have been developed. One is that devel-
oped by Fischer and colleagues (1995). This algorithm is a geometric-based approach,

in which the protein and the ligand are represented by “critical” points and sets of points

 

are matched using geometric hashing, whereby the geometry for the spatial arrangement
of points of potential interaction are precomputed and entered in a lookup table for later
reference. Once possible matchings are identified, they are scored based on the contact
area between molecules and their electrostatic interactions. This procedure was able to
dock test ligands to within 1.5 A root-mean-square deviation (RMSD) relative to the posi-
tion of the ligand in the crystallographic complex in 18 of 19 test cases, though often the
best docking in terms of RMSD was not ranked near the top based on scoring. Scoring of
computational dockings, equivalent to predicting binding aflinity, is a key component of
identifying the “best” docking and remains a Significant challenge. A discussion of scoring
functions is presented below in Section 1.4. Another rigid-body docking algorithm is the
LIGIN program developed by Sobolev et a1. (1996). In LIGIN, a complementarity func-
tion including terms for favorable contacts, unfavorable contacts, and the contact surface is
defined. Minimal receptor flexibility is modeled by allowing the user to define one or more
residues whose side chains, from Cg to the side-chain termini, are ignored in calculation of
the complementarity function. LIGIN was tested on a set of 14 complexes and was able to

dock all ligands with reasonable RMSDS relative to the crystal structure.

More complex approaches to handle conformational flexibility in ligand docking have
been developed. In the latest version of DOCK, limited ligand flexibility is modeled by the
use of rigid-body dockings of ligand conformers, with later versions of DOCK including
an internal Conformation generator using a genetic algorithm (GA; Oshiro et a1. 1995). An
alternative to stochastic sampling of the rotational degrees of freedom is to use previously

observed low energy states. The algorithm of Leach (1994) implements such a technique

 

to explore both protein side chain flexibility, via the use of rotarner libraries, and ligand
flexibility, via the use of conformational analysis. Instead of relying on rigid-body dock-
ings of ligand conformations, flexibility is addressed more directly in the latest version of
DOCK using an incremental construction algorithm based on that developed by Leach and
Kuntz (1992). Incremental construction docking begins by placing a fragment of the ligand
into the binding site and then adding functional groups to build up the ligand. Incremental

discussion methods are further discussed below in Section 1.2.3.

A further derivative of DOCK is F LOG (Miller et a1., 1994). FLOG uses the same
matching approach of DOCK, but expands the types of atoms assigned based on their
chemistries. It also includes some ligand flexibility by including generated ligand confor-
mations in the database, but the docking procedure remains a rigid-body one. The authors
of FLOG developed an enhanced grid-based scoring frmction which includes electrostatic,
hydrogen bonding, hydrophobic, and van der Waals potentials. FLOG is able to select

known inhibitors from a large database of drug-like molecules.

In all docking algorithms, there is a tradeoff between how extensively and accurately
the orientation and conformational space is explored and the computational requirements to
perform the exploration. Several methods use stochastic approaches to the docking problem
to achieve a higher degree of accuracy at the expense of exploration. A popular algorithm is
AutoDock (Morris et a1., 1996), which employs a Monte-Carlo simulated annealing method
to sample binding orientations and ligand conformations, by randomized rotation of rotat-
able torsions in the ligand. AutoDock is inexpensive, easy to use, achieves reasonable

dockings, and has been applied to several cases (Goodsell et a1., 1996). An extension to

 

AutoDock replaces the Monte-Carlo search with a Lamarckian genetic algorithm (LGA),
a hybrid GA and local search method (Morris et a1., 1998). The LGA differs from a tradi-
tional GA by its performing a finer local search on the orientation of the ligand relative to

the protein and of the ligand ’3 conformation.

A second stochastic approach based on the use of a GA was developed by Jones and
colleagues (1995). Their approach uses a simple GA operating on rotational angles in the
protein, rotational angles in the ligand, and on hydrogen bonds from ligand to protein and
protein to ligand. The fitness or scoring function encompasses terms for the hydrogen bond
energy between protein and ligand, for the van der Waals energy between protein and lig-
and, and for the internal van der Waals energy of contacts within the ligand. Improvements
to this algorithm resulted in the development of the GOLD algorithm (Jones et a1., 1997),
with changes in the representation of angles and hydrogen bonds and inclusion of a more
robust scoring firnction. The GOLD algorithm achieved “acceptable” dockings for 71 of
100 test cases. Similar to GA approaches is the evolutionary programming approach AG-
DOCK (Gehlhaar et a1., 1995b; Verkhivker et a1., 1999), which is able to generally correctly
dock, defined here as docking within 1.5 A of the complex crystal structure, methotrexate
into dihydrofolate reductase (DHFR) and a proprietary ligand, AG-1343, into HIV pro-

tease.

Another approach is the ICM algorithm (Abagyan et a1., 1994; Totrov and Abagyan,
1997), which uses a complex procedure to finely dock a ligand. In this approach the
molecules are represented by a set of internal variables (rotatable angles, e.g., rotatable

side-chain single bonds) for their relative positions and conformations, which are randomly

 

changed, energy minimized, and selected via the Metropolis algorithm for each step. This
approach successfully docked lysozyme to its antibody (Totrov and Abagyan, 1994) and
fl-lactamase and its inhibitor (Strynadka et a1., 1996), but the study presented here docked
only three of eight ligands used in the Second Meeting on the Critical Assessment of Tech-
niques for Protein Structure Prediction (CASP2; Moult et a1. 1997) with an RMSD relative
to the crystal structure of less than 5.0 A. Generally, dockings with RMSDS lower than 2.0

to 2.5 A are considered as correct.

The SLIDE algorithm, which constitutes a major focus of this thesis work, belongs to
this first, ligand manipulation, class of docking algorithm types and is described in Chap-

ter 3.

1.2.2 Recombination Docking Methods

An alternative to the approaches described above, in which the docking is performed via
manipulation of a ligand molecule, is the approach of placing each piece of the ligand
independently and then linking the docked fragments to form a docking of the complete
ligand. One implementation of this approach is the empirically based GEMINI algorithm
(Singh et a1., 1991), which docks peptidyl ligands. In this method, a database of side-
chain packing arrangements generated from 52 protein structures (Singh and Thornton,
1990) is used to map potential orientations of each peptide ligand Side chain relative to
the protein binding-site side chains. The three-dimensional distribution for each ligand
residue/protein residue type pair is orientated based on the protein side chain’s positions.

These distributions can then be superimposed to identify potential orientations of the ligand

10

side chains which simultaneously occur in regions of high frequency in the distributions.
Some simple constraints on covalent bond formation to other ligand side-chains, i.e., along
the peptidyl ligand chain, can be included to yield a small set of possible ligand binding
orientations and conformations. The technique has the limitation that it is applicable to
only peptidyl and, perhaps, a few other limited types of ligands as an empirical database of
side-chain or firnctional group orientations must be known. This requires that a statistically
significant number of example interactions exist in the structural database, which is the case
for peptidyl packings as one can use the proteins in the Protein Data Bank (PDB; Berman
et a1. 2000; Bernstein et a1. 1977). However, for many ligands of interest, there may be

only a few or no structures with similar groups bound.

The algorithm developed by Sandak and colleagues (1998b) is another recombination-
based approach. In this method, the ligand is represented by a set of predefined hinges and
set of interaction points, representing Sites of favorable interaction. Each triplet within a
part of the ligand, i.e., portions of the ligand on one side of the hinge, are matched by triplets
of interaction points in the protein. The best dockings of each piece and hinge orientation
are tabulated, recombined, and then scored using an interatomic contact scoring function.
This algorithm can be applied to hinges in either the ligand or the receptor, but not both,
during a docking run. Acceptable dockings were achieved for synthetic peptides binding to
calrnodulin and to HIV protease (Sandak et al., 19983), which undergo a clamping motion

upon ligand binding.

11

1.2.3 Incremental Construction Docking Methods

A third docking approach is that of incremental construction, which initially docks only
a portion of the ligand and then builds up the remaining ligand groups based on this ini-
tial placement. The classical algorithms for this docking approach are GROW (Moon and
Howe, 1991), for peptidyl ligands, and FLEXX (Rarey et al., 1996a), for generalized lig-
ands. In FLEXX, the flexibility of the ligand is described by torsion angle rotations at dis-
crete steps as in the MIMUMBA algorithm (Klebe and Mietzner, 1994). In the first stage
of docking, the base, or anchor, fragment is chosen, which is then placed into the bind-
ing site. This placement in done based on local interaction surfaces as described in Rarey
et al. (1996b), involving a pose, or orientation, clustering step. Once the anchor fragment
is placed, in one or more orientations, a tree representing the possible conformations for
adding the remaining pieces of the ligand is constructed. Each node of the tree represents a
set of currently docked fragments and their orientations and conformations, with the base
node(s) representing the base fragment docking(s). Branches from a node represent the
next ligand fragment to add, with each branch including the identity of the fragment and
the rotation about the newly formed bond, set to one value in a set of fixed increments of the
rotation angle. Ligand groups connected by a rigid, i.e., nonrotatable, bond are considered
a single fiagment, so the bond formed by all additions is rotatable. For example, branches
from a current node may include a methyl group and a rotation angle of 10°, a methyl group
and a rotation angle of 20°, etc. to 360° (36 branches), and a second set of branches from
the same node for a carboxyl group and rotation angles of 10—360° (36 branches), for a

total of 72 branches. Due to the computational complexity of exploring the complete tree,

12

only the top It, with I: set to 500 in the work presented by the authors, scoring branches
from each node are explored. The solution set, consisting of all dockings, is then clustered
to reduce dockings that are very similar which arose from independent traversals of the
docking tree. Scoring is done with a energy fimction derived fi'om that of Béhm (1992a,b).
This algorithm was tested on a set of 19 protein-ligand complexes which had between 1
and 8 x 1010 theoretical ligand conformations, i.e., combinations of incremental torsional
angle rotations in the ligand, including ligands which may contain internal steric collisions.
In 15 of 19 cases, a docking with RMSD from the crystallographic ligand of less than 1.0 A
was achieved, but in some cases the lowest energy docking was significantly different from
the crystallographic complex, with RMSDS up to 4.5 A. Also, the energy scoring function

used in this work generally did not correlate well with the experimental binding energy.

A later study on the performance of FLEXX using a test set of 200 protein-ligand com-
plexes was done (Kramer et al., 1999). For this test set, the top ranked docking was within
2.0 A RMSD of the ligand in the crystallographic structure in 47% of the cases, and a dock-
ing within 2.0 A of the crystallographic ligand position was found in 70% of the cases if all
dockings are examined, not just those with the best score. In general, the FLEXX algorithm
achieves better dockings with simpler ligands. More complex ligands, those with more than
15 components, yielded a correct docking in only 25% of the cases. This study also showed
that, in general, the algorithm is able to cross dock most ligands, i.e., dock a ligand from
one crystallographic structure into the conformation of the protein extracted from another

crystallographic structure that contains the protein bound to a different ligand.
To address side-chain and main-chain flexibility of the receptor structure in docking, an

13

extension to FLEXX, FLEXE (Claussen et al., 2001), was created. FLEXE docks ligands
into an ensemble of structures of the receptor instead of a single receptor binding site. The
binding site ensembles can be from different crystallographic structures, as in the study
described here, from a homology model with uncertain side-chain positions, fi'om a series
of molecular dynamics time steps, or from another source. The key component of this
algorithm is that multiple conformations of the protein can be used as a docking target si-
multaneously. In this algorithm, the receptor structures are merged, with regions of similar
conformations reduced to a single structure and regions with dissimilar conformation con-
stituting alternate positions. While this algorithm may handle some backbone movement
in addition to side-chain rotations, the authors claim it is not able to work with large do-
main movements and limit their test set to protein ensembles with similar backbone traces.
FLEXE was able to generate a docking ranked in the top ten potential dockings which was
within 2.0 A of the crystallographic structure in 67% of the 105 structure test set, compared
to 63% with the FLEXX algorithm. In addition, FLEXE was able to effectively cross dock

two potent inhibitors which FLEXX was not.

1.2.4 Water in Computational Docking

Water molecules are known to play a key role in many protein-ligand complexes, reviewed
by Ladbury (1996), but are often ignored in computational docking approaches as they
are difficult to model. Many of the scoring functions do consider the energy of desolvation,
generally as a function of buried and exposed hydrophobic surface area of the ligand and/or

protein, but there are two problems with using this as the only water modeling technique:

14

(1) it assumes that the best protein-ligand interface will be completely desolvated, while in
many instances water molecules are retained in the interface and contribute key hydrogen
bonding interactions, and (2) the shape of the binding pocket surface to which the ligand
binds in reality, which includes the water molecule(s), is different from the surface used in
the docking, which does not include the water molecule(s). The second point is especially
important as steric fit between the ligand and the protein is Often a key component of the

docking and scoring procedures.

Since there are no covalent constraints on their positions relative to the ligand or the
protein, treatment of water molecules in docking procedures must be different from the
procedures used to model either the protein receptor or the ligand molecule. One approach
is to retain bound water molecules fi'om the free protein structure as a fixed part of the
protein target in the docking process. However, this can lead to overrepresentation of inter-
facial watcr as roughly two-thirds of the water molecules in the binding site of the protein
are lost upon ligand binding (Raymer et al., 1997). A method to model protein-bound wa-
ter sites conserved upon ligand binding is to predict which bound water molecules will be
conserved and which will be displaced. This approach is used with the lc-nearest-neighbor
genetic algorithm application Consolv (Raymer et al., 1997), allowing those water sites
predicted to be conserved to be included in docking, while removing those predicted to be
displaced. Because this prediction is not 100% correct and will depend, to some extent,
on the ligand shape and chemistry, water molecules wrongly predicted as conserved may
incorrectly bias the docking. A better method may be to allow the docking algorithm to

displace any water molecule, but penalize the displacement of water molecules according

15

-.- gar-nu my

 

to their predicted likelihood of being conserved, as is done with SLIDE (Section 3.2.4).
F LEXX uses a different idea to address the problem; it places water molecules at favorable
sites during the docking process by placing “particle phantoms” at favorable positions in
the binding site prior to docking. These phantoms can be turned on during docking when
they can make favorable interactions to the protein and to the growing ligand and turned
off if they are involved in steric collisions or fail to meet angular hydrogen bond constraints
(Rarey et al., 1999). Another method is that developed by Jackson et a1. (1998), which uses
a finely spaced grid of potential hydration sites around the ligand as it is being docked. The
energy of the docked ligand can be computed including hydration at subsets of the grid

points, yielding favorable hydration sites for the docking.

1.3 Computational Screening

Computational screening is the process of identifying molecules that bind to a protein of
interest from a database using computational methods, as reviewed in Walters et al. (1998).
It can be considered as a computational equivalent to traditional hi gh-throughput screening.
In theory, any of the docking methods described above can be used in a screening mode by
attempting to dock all molecules in a database, such as the Cambridge Structural Database
(CSD; Allen and Kennard 1993). In practice, however, this method is not practical due to
the time needed to dock each molecule. Most of the docking algorithms described above
are reported to take several minutes to hours to dock a single ligand. Most of the molecular
databases of interest contain 100,000-plus molecules. Even if the fastest method is used,

at one minute per ligand, a screening of 100,000 molecules would take over two months

16

of computational time. When factoring in the desirability of docking 10—20 different con-
formers of each ligand to better represent their flexibility, the computational time expands
to years for a single database screening. One approach is to develop a docking algorithm
designed for screening, such as the SLIDE algorithm originally presented by Dr. Volker

Schnecke and extended as described in this thesis (Chapter 3).

An alternative screening approach to database docking is closely related to the incre-
mental construction docking algorithms. In such an approach, the ligand backbone is ini-
tially docked, followed by the addition of various functional groups, akin to combinatorial
chemistry methods. This approach is used in LUDI (BOhm, l992a,b), Grow (Moon and
Howe, 1991), SPROUT (Gillet et al., 1993, 1994), GroupBuild (Rotstein and Murcko,
1993b), BUILDER (Roe and Kuntz, 1995), HOOK (Eisen et al., 1994), and SMOG (De-
Witte and Schaknovich, 1996). A technique in which the ligand is built up atom by atom
instead of fragment by fragment is employed in the Genstar (Rotstein and Murcko, 1993a),
Legend (Nishibata and Itai, 1993), MCDNLG (Gehlhaar et al., 1995a), and CONCEPTS

(Pearlman and Murcko, 1995) programs.

Closely related to these incremental construction methods for screening is the F LEXS
algorithm (Lemmen et al., 1998). This method requires knowledge of the binding orien-
tation for at least one ligand, which is used as the rigid component in flexible alignment.
The structures in the screening database are broken into fi'agments, as in the previous incre-
mental construction screening methods; however, the base fiagments, and later the flexible
groups, are placed based on matching chemical properties, such as hydrophobicity, hydro-

gen bonding character, and partial charge, with the known ligand instead of the protein

17

 

binding site. F LEXS was able to extract and highly rank ligands with known function
against the fibrinogen receptor from a database of 984 drug-like compounds. In terms of
docking ability, FLEXS was able to reasonably reproduce dockings resulting fi'om super-

position of crystal structures for test set consisting of 14 protein targets and a total of 284

superimposed ligands.

1.4 Docking and Screening Scoring Functions

One of the difliculties in computational docking and scoring methods is that of accurate
scoring, measuring the affinity between protein and ligand. The scoring function is key
to being able to identify the most realistic ligand dockings, in docking experiments, or
the most promising potential ligands for further study, in screening experiments (Greer
et al., 1994). Scoring functions can be classified into two general categories: molecular

mechanics-based and empirical.

Molecular mechanics-based scoring functions are those that describe the energy of
ligand binding in terms of a physical chemistry function summing the component bind-
ing energies, such as van der Waals contacts, electrostatic interactions, and covalent bond
stretching, bending, and torsional energies. Such scoring functions includes the forcefield-
based measures such as AMBER (Weiner et a1., 1984, 1986), which is not commonly used
in docking and screening methods as it has been tuned to reproduce protein and nucleic
acid energies, rather than arbitrary small organic ligands. A commonly used molecu-

lar mechanics-based scoring function for docking and screening is the Merck Molecular

18

”In”. I‘ L

Forcefield (MMF F; Halgren 1996), which was developed as a combined “organic/protein”
forcefield for molecular dynamics of such systems. One problem with this scoring function
is that it is highly detailed. While this feature enables it to be relatively accurate, it also
makes the function computationally expensive, which causes significant problems when
scoring a large number of docking orientations for many potential ligands. One method
to reduce the computational cost is to precompute the atomic potential for the protein on
a grid, so that only the ligand potential changes with each new cycle. This is done with
GRID (Meng et al., 1992), implemented in DOCK, and FLOG (Miller et al., 1994) scoring
functions. Other grid-based scoring algorithms developed by Blom and Sygusch (1997)
and Mandell et al. (2001) in DOT uses a discrete Fourier transform (DFT) correlation ap-
proach to solve a moderately complex molecular mechanics-based scoring function with
reasonable computational cost. Docking experiments done with this scoring function on
four complexes were able to dock the ligand molecule close to the observed orientations in
the crystal structure. The docking time of the DFT correlation approach is generally low,
but much too slow, on the order of 2 to 30 hours per ligand depending on grid resolution

and ligand size, to be of use in screening experiments.

Given their extensive use in docking and screening algorithms, a brief note about the
general methodology behind grid-based scoring procedures is warranted. In general, a
three-dimensional grid is established in the binding site to calculate energies. The initial
step is to precompute the energy of interaction between the target and each type of atom
residing at each grid point. After this step, the scoring function constitutes a lookup table of

the energy for each type of atom at each grid point. When a ligand is placed into the binding

19

site in an arbitrary orientation during the docking procedure, each atom of the ligand is
assigned to the nearest grid point. The energy for the docking orientation is then calculated
as the sum of the energies for those grid points to which ligand atoms have been assigned,
using the assigned ligand atom type energy calculated for the grid point. Since grid-based
scoring functions are precomputed prior to docking and screening and simply looked up
during the run, they are very fast to compute during the actual docking or screening run.
The drawback of grid-based scoring methods is that they use only an approximate ligand
atom position. Problems resulting fi'om this approximation can be limited through the
use of a finely-spaced grid; however, this can greatly increase the precomputation time,
especially with complex scoring functions, and somewhat increase the run time, due to the

ligand atom to grid point assignment step.

Most of the commonly used docking and screening tools implemented use empirically
tuned scoring functions. One of the earliest such scoring functions is SCOREl , developed
by BOhm for the LUDI program (Béhm, l992a, 1994). SCORE] seeks to calculate the
binding aflinity, AGbinding’ as the sum of the energy from polar interactions, including
hydrogen bonding and ionic contact; nonpolar interactions, constructed from the surface
area buried in the complex; and flexibility, defined as an “energy” for loss of the ability
to rotate a rotatable bond in the ligand. This scoring function was generally able to pre-
dict binding energies close to those experimentally observed. Examination of complexes
with large deviations between the predicted and observed binding energies led to the de-
velopment of SCORE2 (BOhm, 1998), which includes additional parameters for penalizing

cavities in the binding site, an improved electrostatics model, an term for aromatic group

20

interactions, and a term for desolvation effects. This improved scoring function improved
the correlation between predicted and observed K ,- values and reduced the standard devia-

tion in predicted K.— from 1.7 orders of magnitude to 1.3 orders of magnitude.

Other similar energy-based scoring functions have been developed. One example is the
VALIDATE scoring function developed by Head et al. (1996), which combines molecular
mechanics approaches with empirical descriptors. The enthalpy of binding in the protein-
ligand complex is computed from a molecular mechanics forcefield, while additional prop-
erties, including fixing of rotatable side chains, buried surface area complementarity, and
steric compatibility, are used to estimate the entropy of binding. Overall the VALIDATE
scoring functions consists of 12 terms, and while accuracy is quite good, the estimated
cross-validation error for a set of 51 complexes in predicted K.- was 1.1 order of magnitude,
the time to compute the score is over 1000 times slower than the original BOhm SCOREI
function. Another similar scoring function is that developed by Jain (1996), which includes
terms for hydrophobic complementarity, for hydrophilic complementarity (hydrogen bond-
ing and salt bridges), for electrostatic repulsion, for desolvation, and an entropic term. This
function was tested on a set of 34 protein-ligand complexes and resulted in a mean pre-
dicted K4 error of 0.7 orders of magnitude. Yet a third is the piecewise linear potential
(PLP; Gehlhaar et al. l995b,Verkhivker et al. 2000), which simplifies the energy firnction
to only four terms to yield a piece-wise linear approximation of the hydrogen-bond and
lipophilic interaction wells. A fourth approach is the ChemScore algorithm developed by
Murray and colleagues (Eldridge et al., 1997; Murray et al., 1998), which includes terms

for hydrogen bonding interactions, metal interactions, hydrophobic interactions, and a term

21

for loss of entropy, i.e., rotational freedom. This function predicts the observed affinities

with a cross-validated error of 1.3 orders of magnitude.

Several additional methods have been developed which take advantage of the structural
information available in the PDB. In all of these methods, the key is the creation of a statis-
tical distribution for various contacts between atom types in the ligand and in the protein.
One such example is that developed by Miigge and Martin (1999), which uses the statistical
information to derive a set of fi'ee energies of protein-ligand atom pair interactions, or po-
tentials of mean force (PMFS), which are summed to provide the final score. The method of
deriving the PMFs implicitly includes entropic and desolvation effects. Tests performed on
a database of eight protein-ligand complex sets, which each contained between 11 and 77
complex structures, showed reasonable correlation between the score and observed binding
aflinities. The authors note that one set gave poor correlation, likely due to the large and

variate Size of the inhibitors.

Other methods which use statistical information derived from structures reduce the rep-
resentation further away from one of energetics to include only the distribution informa-
tion. An early implementation of such a method is that developed by Klebe (1994), which
extracts information on interaction angles from structures in the Cambridge Structural
Database. This method was able to reasonably predict the key binding sites for methotrex-
ate in dihydrofolate reductase (DHFR), for a peptidic inhibitor in endothiapepsin, and for
tyrosinyladenylate in tyrosyl-tRNA synthetase. This method has been extended by Nissink
et al. (2000) through the use of Isostar (Bruno et a1., 1997), which tabulates the orientations

of specific chemical groups with respect to another specific chemical group of interest

22

from PDB and CSD structural data. In this method, the orientation data are used to derive
propensity plots, or smoothed spatial distribution firnctions, which can be used as a basis
for scoring the spatial relationships in computationally docked ligands. A second extension
of Klebe’s work is the development of DrugScore (Gohlke et al., 2000a,b). In DrugScore,
the model of Spatial relationships is further reduced to a set of one-dimensional radial dis-
tribution functions for each set of atom-type interactions. DrugScore was shown to be an
improvement when examining the RMSD between the top-ranked ligands and the ligands
in the crystallographic structures during docking experiments with FLEXX (Kramer et al.,
1999). The top-ranked docking was within 2.0 A RMSD from the crystallographic struc-
ture in 73% of 91 test protein-ligand complexes using DrugScore, versus only 54% for the
F LEXX scoring function. DrugScore resulted in standard deviations in predicted pK; val-
ues of 0.7 to 2.2 on a test group of 9 test sets, consisting of between 16 and 71 complexes,
though linear correlations with observed binding aflinities were poor for some of the test

cases.

Stahl and Rarey (2001) compared several scoring functions in terms of the docking
algorithm of FLEXX for a computational screening procedure. The test database consisted
of seven proteins, with 36 to 128 known inhibitory compounds per protein. The F LExX,
PLP, DrugScore, and PMF scoring functions were tested. In general, the F LEXX scoring
function performs best with compounds for which the binding is dominated by hydrogen
bonds, such as for thrombin and neuramirridase, but poorly with compounds whose binding
is predominated by hydrophobic interactions, such as for cyclooxygenase-2 (COX-2). In

contrast, the DrugScore algorithm performed well with the COX-2 screening, but poorly

23

for neuraminidase. The PLP algorithm tended to perform well with shallow sites that still
had significant hydrogen bonding interactions, while the PMF algorithm performed poorly

with very narrow and/or restricted binding sites.

Given the preponderance of scoring firnctions available, one may consider a method
of combining a set of functions to overcome shortcomings in any one particular scoring
function. This approach was implemented by Charifson et al. (1999) and by Stahl and
Rarey (2001), who both showed improvements in screening efficiency, in terms of rank-
ing active compounds highly compared to inactive compounds. Both tested against a set
of diverse targets, with consensus scoring providing an overall increase in performance
across the set. In some cases, the consensus method performed worse for a specific target,
but without a priori knowledge of which scoring function may be best suited for the tar-
get of interest, it is not possible to improve on the overall performance. When examining
combinations of only two scoring functions, as is done in the thesis work presented here
(Section 4.3.3), there is a very limited number of ways to combine the scoring function re-
sults. If more scoring functions are included, the possible ways to combine them increases.
Wang and Wang (2001) present a computational experiment to explore the effects of difl‘er-
ent combination methods of the resulting overall selection of ligands fiom a computational
screening run. They constructed a set of virtual ligands by assigning a “true energy”, i.e.,
equivalent to the experimentally observed energy, based on a Gaussian distribution. This
experimental dataset was then used to create ten predicted datasets by adding an random
error to each experimental score, representing the error introduced by each often indepen-

dent scoring functions. They showed that use of increasing numbers of scoring functions

24

 

results in an increase in the ability to include an “active” ligand in the top 100 ligands from
the screening database using two combination methods, but the increase in performance
becomes significantly less when including more than four to five scoring functions. The
use of additional scoring functions is undesirable for docking and screening applications
since each additional scoring function evaluated will require additional computational time,
which is at a premium during docking and screening runs. Another concern arises out of
statistical pattern recognition and that is the fact that, with a finite amount of data, overall
scoring accuracy is likely reach a peak and then decline as additional scoring functions,
i.e., dimensions, are added to the system. This phenomenon is often termed the curse of
dimensionality (Jain and Chandrasekaran, 1982). A classic construction to illustrate such a

case was presented by Trunk (1979).

One concern with all empirically tuned scoring functions is that of the measurement
of the binding affinities. Generally, the observed aflinity values are taken from literature
sources; however, it is very uncommon to have affinity values measured at standardized
conditions, even for a set of ligands to a single protein. By tuning the scoring function to
these observed values, one makes the assumption that differences in the observed values
due only to differences in affinity for the ligands. One must question the validity of tuning
the scoring functions to data which may not be self-consistent, and one can ask how much
of the deviation seen between predicted values and observed values is likely due to inaccu-
racies in the scoring function and how much is due to differences in experimental methods

and conditions used to measure the binding affinity.

One interesting idea related to scoring is the use of a post-processing filtering step

25

(Stahl and BOhm, 1998). A limitation of most scoring firnctions is the absence of strong
penalties for unfavorable interactions, especially in terms of leaving cavities in between the
buried surfaces of the protein, which are not generally seen in protein-ligand complexes.
The filtering algorithm presented in this study includes four terms: the size of cavities in
the internal protein-1i gand interface, the portion of solvent accessible surface (SAS) of the
ligand which is hydrophobic, the fraction of ligand volume buried in the binding cavity,
and the presence of pairs of polar atoms in close contact which do not participate in hydro-
gen bonds. Testing against a set of 32 complexes docked with FLEXX showed a general
decrease in the RMSD of the best ranked ligand and a dramatic increase in the number
of complexes for which the the docking closest to the crystal structure was ranked within
the top 20 ligand dockings. The authors also note than the docking with the best RMSD
relative to the crystal structure was not lost after filtering for any of the 32 complexes. For
ligand manipulation docking and screening methods, e. g., SLIDE, a filtering step could be

directly incorporated into the scoring function.

1.5 Evaluation of Docking and Screening Methods

As there are several docking and scoring methods available, it is interesting to examine
their performance on identical problems. Performance of various scoring functions im-
plemented in FLEXX is discussed above. One of the key studies examining each of the
docking algorithms independently is the docking section of CASP2, the results of which
were summarized by Dixon (1997). In this study, seven small molecule-protein complexes

and one protein-protein complex were used as targets. Target protein structures were pro-

26

SC

vided to several docking algorithm research groups along with the two-dimensional ligand
structures, but no ligand conformational, i.e., three-dimensional, information was given.
Of the approaches discussed above, the ICM method of Abagyan et al. (1994), FLEXX
(Rarey et al., 1996b), the DFT method of Blom and Sygusch (1997), and LIGIN (Sobolev
et al., 1996) were examined. The mean RMSD-based score between the docked and crys-
tallographic ligand orientation for all docking methods ranged from 2.5 to 7.2, indicating
that while some targets are easier to dock overall, none regularly dock correctly for every
method. The DFT docking method performed poorly, having RMSDS between 15 and 28 A
for the three targets for which predictions were performed. Of the other three approaches,
the overall performance were roughly equivalent. Some of the targets were clearly easier to
dock, with most of the presented methods docking ligands within 3 A, while other targets
had no dockings within 4 A. For the protein-protein trial, none of the algorithms achieved
a close docking when examining detailed geometry. However, some methods were able to

correctly predict some of the interactions which occur in the binding site.

Bissantz et al. (2000) also examined a series of docking algorithms and scoring func-
tions for screening against thymidine kinase and the estrogen receptor. They found that
most algorithms were able to extract roughly 70% of known ligands from a database of
990 random molecules and that consensus scoring generally enhanced hit rates. However,
they saw no relationship between the accuracy of producing the correct docking orientation
and correctly ranking the orientations. It was also not possible to accurately predict bind-
ing energies. Of note is the authors’ suggestion to use a limited size database and several

screening methods to determine the best tool for the protein in question and then to screen

27

a large database.

Other studies have examined the performance of docking and screening algorithms in
more limited ways. Knegtel and Wagener (1999) used the DOCK algorithm to explore the
efficacy of screening methods for thrombin inhibitors and for selectivity for progesterone
receptor ligands versus estrogen receptor ligands. This study applied two scoring fimc-
tions: an energy scoring function consisting of the AMBER forcefield (Weiner et a1., 1984,
1986) and a chemical scoring function consisting of the AMBER forcefield with the attrac-
tive van der Waals interaction energy applied only to interaction between complementary
atom types. For thrombin, a set of 32 active inhibitors and a set of ten chemically similar,
but inactive, compounds was compared. Rigid-body docking yielded a slight bias towards
active compounds in the ligands ranked in the top 100 ligands, but allowing flexibility ef-
fectively eliminated this bias, allowed more highly ranked inactive compounds. Neither
scoring function had a strong ability to differentiate between active inhibitors and inactive
compounds, though the chemical scoring methods had a slight advantage over the AMBER
scoring function. Docking against the progesterone receptor was performed on a set of 28
known agonists and 20 chemically similar estrogen receptor ligands. A similar analysis to
the thrombin case Showed that the both scoring functions had a somewhat more pronounced
discriminant ability, but that the energy scoring function performed somewhat better. This
is likely due to the highly hydrophobic nature of the progesterone receptor site whose inter-
actions are predominantly based on van der Waals forces. The AMBER forcefield is likely

to more accurately model such forces.

A second such study was performed on stromelysin-l (matrix metalloproteinase 3;

28

MMP3) by Ha et al. (2000). The authors of this study co-crystallized MMP3 with 6
biphenyl-based inhibitors, all of which dock in highly conserved orientations and are used
to construct “correct” dockings for a set of 61 biphenyl ligands with IC50 values in the low
micromolar to low nanomolar range. DOCK with the PMF scoring function performed
best, with a mean RMSD of 1.8 A between the docked and crystallographic binding modes
and oriented nearly all ligands with the biphenyl moiety in the correct binding pocket.
FLEXX and DOCK with the AMBER forcefield performed worse and generally produced
ligand orientations with the biphenyl group in other, unoccupied pockets around the bind-
ing site. However, orientations calculated by F LEXX which did place the biphenyl into the
correct pocket generally had a lower RMSD relative to the crystal structure, indicating that

FLEXX may be better able to fine tune the ligand orientation.

1.6 Creation of Targeted Computational Screening

Databases

Computational screening’s requirement for very fast handling of individual ligands puts
severe limits on the detail one can use in the algorithm. A possible method to reduce the
effective time per molecule and add complexity is to limit the screening to ligands that
are of particular interest, for example by removing molecules from the database that do not
resemble drug molecules. Bemis and Murcko (1996) showed that 50% of orally deliverable
human drug molecules in the Comprehensive Medicinal Chemistry (CMC) database are

characterized by only 31 of l 179 graph frameworks, Le, a connectivity graph ignoring atom

29

 

31]

1&1

3P

types and bond orders, and 24% are characterized by only 41 of 2506 atom frameworks, i.e.,
graph fi’ameworks which consider atom types and bond orders. In fact, 8.5% of these drug
molecules have a benzene fiamework. Databases could also be pruned to include only those
molecules which chemically resemble known drugs. The well-known Lipinski rule of fives
(Lipinski et al., 2001), based on examination of a 2245 orally active subset of the World
Drug Index (WDI; Derwent Information, London, UK), gives the following guidelines for
molecules that are unlikely to be adequately soluble and permeable to function as oral

pharmaceuticals:

0 contains more than five hydrogen bond donors,

0 contains more than ten hydrogen bond acceptors,

c has a molecular weight greater than 500 Da, or

c has a calculated Log P (CLogP; the calculated octanol/water partition coeflicient)

greater than 5 (or has a Morigucchi Log P, MLog P >4.15).

Limiting the database to only those molecules which contain a common framework and
which meet the Lipinski rule would greatly decrease the number of molecules in most
chemical databases, allowing a greater exploration of each of the molecules. Other re-
strictions on the screening database contents, such as limiting it to molecules which con-
tain a specific functional group, could also be implemented. One disadvantage to these
approaches is that any new ligands identified will resemble drug-like molecules and any

unusual novel compounds will be overlooked.

30

 

1.7 Molecular Clustering

Another method to reduce the effective size of the database is to eliminate molecules which
bear close resemblance to other molecules in the database, e.g., by clustering. However,
when clustering is used, one must ensure the molecules contained in the clustered database
represent a sufliciently broad scope of chemical space. The breadth of this scope is not
necessarily in relation to the whole of chemical space, but may be restricted to regions
of space with particular interest. Matter (1997) examined the diversity in a set of 1283
biologically active compounds using various similarity measures. Compounds were clas-
sified with simple 2D fingerprints or with 2D fingerprints combined with more complex
descriptors. Clustering was then performed, taking the structure closest to the center of the
cluster as the representative. Chemical space coverage was measured by the percentage
of biological classes which were included in the final set of database molecules. The best
coverage was achieved using only 2D fingerprints. While using this approach to prune the
database achieves a reduction in the number of molecules which must be screened, there
is a potential that the representative molecule in a cluster may not be selected during a
screening run or may be inactive upon experimental investigation while another molecule
contained in the cluster would be a very good ligand. Other clustering methods include the
Jarvis-Patrick nonhierarchical and Ward hierarchical methods, their application to chemi-
cal structures discussed in Brown and Martin (1996), the clustering of Markush structures
using a lc-means clustering algorithm (Barnard et al., 2000), and clustering using molecular

field matching algorithms (Mestres et al., 1997).
A related approach is to restrict the number of molecules which must be analyzed after

31

a screening run is performed as effective analysis can often be difficult given the large
numbers of selected potential ligands. This is an especially serious problem when the
screening database contains many compounds which closely resemble known ligands for
the targets as any novel potential ligands will be difficult to extract from these known ligand
hits. Su et al. (2001) propose a method by which the database is initially grouped into
families based on common frameworks. Each family consists of a base fragment along
with an ensemble of attached functional groups, each transformed into the same reference
flame based on the base fi'agment. The base fragment is then rigidly docked into the binding
site, the ensemble functional groups transformed into the binding site based on this original
docking, and then each family molecule is scored independently. The transfonnation of the
ensemble as an entity means that only a single transformation needs to be done instead
of a transformation for each database molecule, greatly reducing the amount of necessary
calculation. In the final score list, only the best scoring member of each family is identified,
meaning that potential ligands previously ranked below a large molecular family are pulled
to a higher ranking. The authors show that families which contain known ligands are pulled
into ranks which could be considered to be reasonably examined fi'om ranks below what
would normally be analyzed. One concern with this approach are that if the best scoring
member of a family shows no experimental inhibitory activity, any inhibitors in this family
could be missed. These concerns can be reduced, though not eliminated, by analyzing the
top It potential ligands in the family. Another concern is that molecules which contain a
common base fragment but which are otherwise unrelated are placed in the same family.

The use of other modes of clustering could be explored to alleviate this concern.

32

 

1.8 Database Comparisons

Aside fi'om the notion of the best method to group the molecular structures in a Single
database is the question of which database to use. Several studies have compared some of
the available databases (Shemetulskis et al., 1995; Cummins et a1., 1996; Bernard et al.,
1998). The most recent and exhaustive comparison is that done by Voigt et al. (2001).
They compared the open NCI database (Milne and Miller, 1986), the publicly available
portion of the National Cancer Institute anti-cancer and anti-AIDS screening database;
the Available Chemicals Database (ACD; MDL Information Systems, Inc., San Leandro,
CA), a database of commercially available compounds; the ChemACX database (CamSoft,
Cambridge, MA), a second database of commercially available chemicals; the Maybridge
Catalog (Maybridge, Plc, Cornwall, England), a third database of commercially available
chemicals; the Ansinex database (Asinex, Ltd., Moscow, Russia), a database of commer-
cially available chemicals with emphasis on compounds fi'om combinatorial chemistry; the
Sigma-Aldrich Catalog (Sigma-Aldrich, St. Louis, M0); the World Drug Index (WDI);
and the Cambridge Structural Database (CSD). These databases contain between 55,000
and 249,000 available threedimensional structures. All of the databases have some dupli-
cation of entries, ranging fi'om 0.02% for the Asinex database to 13% for the ChemACX
database. Diversity analysis showed that the CSD is significantly more diverse than the
remainder of the databases, which is not surprising given its origin as a repository for in-
formation about all types of organic compounds and not only those commercially available
or likely or known pharmaceutical compounds. Combination of all eight databases yields

a database of 681,000 unique structures. Given the large number of compounds in these

33

publicly available databases combined with what may be proprietary databases of several
hundred thousand molecules makes it clear the efficiency needed to computational screen-

ing techniques.

1.9 Conformer Generation

The final issue to arise in computational docking and screening is that of molecular con-
formers. Confonner generation is of in importance for two reasons. For many molecular
databases, many of the structures do not have three-dimensional coordinates attached to
them and would therefore be inappropriate for most docking and screening algorithms.
Secondly, most of the structures in the molecular databases are taken from crystal struc-
tures of the free ligand and/or fi‘om crystal structures of the ligand with other proteins. It is
quite likely that the conformation of the ligand as bound to the target of interest will differ
from the conformations in the database (Betts and Stemberg, 1999), especially in the case
for databases of small molecule crystal structures as the crystal packing forces are large
compared to the size of the molecule. Many of the docking and screening algorithms al-
low for minor conformational changes, but the bound conformation could be significantly
different and beyond the range for which the program can compensate for. Common con-
formation generators generally follow an empirical method, where rotations are based on
observed structures, such as protein Side-chain rotamer libraries (Maeyer et al., 1997; Dun-
brack and Cohen, 1997; Lovell et al., 2000) and the MIMU MBA program (Klebe and Mi-
etzner, 1994); a systematic method, where each rotatable bond is altered by a fixed angle,

generating a tree of conformers, as implemented by the systematic conformational search

34

 

function in MOE (Chemical Computing Group, Montreal, Quebec); or follow a stochastic
method, as implemented in the stochastic conformational search function in MOE. In the
stochastic method, each rotatable bond in the molecule is rotated by a random amount,
the structure is then energy minimized, and then it is compared to previously generated
conformers. If it is close in energy and/or conformation, it is not saved. This process is
repeated until a set number of conformers is generated or a set number of failures, i.e.,
generation of a conformer which is similar to a previously generated conformer, occurs.
Instead of random torsion rotations, some algorithms also employ random atom displace-
ments and subsequent energy minimization. Bostrt'im et a1. (1998) compared energy rrrin-
imized uncomplexed ligand structures with those bound to the protein and found that, for
most, protein-ligand complexes, the energy difference between the bound and flee struc-
ture is small, 3 3.0 kcal/mol. This gives a guideline for effective conformer generation for
screening and docking methods. On the receptor side, rotations between flee and liganded
structures are generally small and lie within the range of the SLIDE screening algorithm

(Maria Zavodszky, unpublished results).

1.10 Successful Application of Docking and Screening

Methods

Structure-based drug design has become a common addition to drug discovery projects, re-
viewed by Klebe (2000). The first successful application dates back to 197 3 (Beddell et a1.,

1976; Goodford, 1984) when a hemoglobin efiector mimic of diphosphoglycerate was de-

35

veloped. Later, researchers at Dupont-Merck used a pharmacophore model to screen the
CSD and identified the DMP-323 cyclic urea inhibitor (Lam et al., 1994), which reached
phase I clinical trials. Other successes include identification of an inhibitor with a low
micromolar IC50 using LUDI (Klebe, 2000), identification of four inhibitors of farnesyl-
transferase with moderate to high micromolar IC50’s (Perola et al., 2000), identification
of isoforrn specific inhibitors of adenylyl cyclase through pharmacophore screening (Onda
et al., 2001), identification of retinoic acid receptor antagonists using ICM docking methods
to screen the ACD (Schapira et al., 2000), and identification of ligands which bind specif-
ically to the RNA hairpin HIV-1 TAR RNA by screening a subset of the ACD using the
ICM docking method (Filikov et al., 2000). Many other successes of computational dock-
ing and screening are likely to reside within pharmaceutical companies. Given the array of
available techniques for computational docking and screening and the dramatic growth in
computational power and speed, it is likely that the use and importance of computational

docking and screening methods will continue to grow.

1.11 Motivation for this Thesis Work

The thesis work presented in this dissertation seeks to improve techniques for modeling
protein-water and protein-small molecule interactions and to apply these techniques to gain
knowledge about such interactions in systems of interest. Previous examinations of water
molecule binding have been limited to experimental studies, which are often arduous and
time consuming, or have relied on using a single crystallographic structure as a reference.

Chapter 2 describes a technique applying hierarchical clustering to computationally analyze

36

water molecule conservation in a series of crystallographic structures without the necessity
for assigning a structure as a reference. Application of this technique to the serine proteases
thrombin and trypsin shows that it can be effective at assisting in explaining specificity
differences between related enzymes and in examining the water content of protein-protein

interfaces.

Extension flom examination of water molecules as ligands has led to the development
of a computational screening technique. Previous ligand docking and screening tools lim-
ited the modeling of the conformational changes which occur upon a ligand binding to pro-
tein to the ligand molecule or contained very limited protein receptor flexibility, through
rotamer libraries. Chapter 3 describes a screening algorithm, SLIDE, which allows for both
ligand flexibility and protein side-chain flexibility, without resorting to rotamer libraries.
This thesis work focuses on improvements made to the description of the hydrophobic
character of the interaction, including results flom testing on thrombin and glutathione S-
transferase (GST). Chapter 4 describes the application of SLIDE to analyze potential dock-
ing orientations of molecules selected by in vitro high-throughout screening and to identify

a limited set of potential new ligands for asparaginyl-tRNA synthetase.

37

 

Chapter 2

Identification of Conserved Water Binding Sites

in Proteins

This research has been previously published as M. L. Raymer, P. C. Sanschagrin, W. F.
Punch, S. Venkataraman, E. D. Goodman, and L. A. Kuhn. Predicting conserved water-
mediated and polar ligand interactions in proteins using a k-nearest-neighbors genetic al-

gorithm. J. Mol. Biol., 265:445-464, 1997.

2.1 Introduction

2.1.] The Role of Water Molecules in Proteins

Water molecules play an important role in protein structure and function. In addition to
providing the driving force behind protein folding via the hydrophobic effect (Kuntz and
Kauzmann, 1974; Eisenberg and McLachlan, 1986), they play a significant role in medi-

38

ating protein-ligand interactions. In some protein complexes, water molecules play key
roles in establishing the specificity of ligand binding, such as the interface water molecules
in the Trp repressor which allow the protein to make base specific interactions with the
repressor element DNA (Otwinowski et al., 1988; Joachimiak et al., 1994). In thymidy-
late synthase, a water molecule conserved in all crystal structures has been shown to allow
the protein to distinguish between substrate and product nucleotides (Fauman et al., 1994).
HIV-1 protease shows a similar highly conserved water molecule (Wlodawer et al., 1989).
Inclusion of a carbonyl oxygen group to displace the water molecule and satisfy the water
molecule ’s position in the protein’s hydrogen bonding network enabled the construction of

a high-affinity inhibitor (Lam et al., 1994).

In contrast to the above function, water molecules have also been found to contribute
to the plasticity of ligand binding in some protein complexes, such as in the class I ma-
jor histocompatibility complex (MI-1C I) where bound water molecules rearrange to allow
the protein to bind to several peptidyl ligands (VVrlson and Fremont, 1993). The different
peptides have varying side chains, with water molecules bridging the gaps which would
otherwise occur between the bound peptide and the protein. Water molecules can also be
directly involved in the protein’s catalytic function, as in the case of the hydrolytic mecha-
nism for peptide bond breakage by serine proteases (Blow et al., 1969; Perona et al., 1993;
Singer et al., 1993). In fact, proteins which are stripped of their primary hydration level are

observed to lose catalytic firnction (Rupley and Careri, 1991).

In addition to playing a such a direct role, water molecules have been shown to sta-

bilize protein structures via formation of extensive hydrogen bond networks (Baker and

39

 

Hubbard, 1984) and by filling grooves on the protein surface (Kuhn et al., 1992a). It has
been shown crystallographically that bound water molecules remain an integral part of the
protein structures, even after repeated rinsing of protein crystals with anhydrous organic

solvent (Fitzpatrick et al., 1993; Travis, 1993).

In general, several techniques exist for the identification of bound water molecule sites
in a single protein, reviewed by Levitt and Park (1993) and Karplus and Faerrnan (1994),
which measure somewhat different aspects of water binding. Protein structures solved by
X-ray and neutron crystallography ofien assign water molecules bound to very favorable
binding sites as the water molecules bound to minimally favored sites and water molecules
in the bulk solvent are too mobile to appear as electron or neutron density peaks. A con-
cern with using crystallographic structures for identifying favored water sites is the influ-
ence of crystal packing contacts, which can either exclude water, leading to undiscovered
sites, or trap water, leading to sites that are not biologically relevant. A second method
of identifying favored water binding sites is through NMR, which can measure the time
during which a water molecule occupies a given site, i.e., the residence time. Aside flom
limitation of NMR to small to moderate-sized proteins, water site identification remains
a significant challenge. Water sites that are too far flom a proton group, water sites that
are close to rapidly exchanging protons, and water sites that exchange extremely rapidly
cannot be identified. Several NMR studies of protein structure have identified long-lived
buried water sites that coincide with crystal structure sites (Otting and Wiithrich, 1989;
Clore et al., 1990; Forman-Kay et al., 1991; Xu et al., 1993). Otting et al. (1991) were able

to observe some rapidly exchanging surface water molecules with NMR and found that

40

 

water molecules that corresponded to surface water sites in the crystal structure had similar

residence times to each other.

2.2 Conservation of Water Molecules among Several

Crystal Structures of a Protein

Given the multiple and important roles water molecules play in protein structure and func-
tion, the ability to quantitatively define conserved water sites flom crystallographic protein
structures has a number of practical applications, including drug design, allowing the de-
sign of ligands that displace conserved bound water molecules (Ladbury, 1996; Wang and
Ben-Naim, 1996), and analysis of protein ligand interfaces to identify such sites (Raymer
et al., 1997; Sanschagrin and Kuhn, 1998). A typical method for analysis of crystallo-
graphic bound water molecules is to use molecular graphics to visualize the water bound
in a single protein structure, or small number of closely related structures which have been
superimposed, and their proximity to catalytic or ligand-binding residues. As the num-
ber of superimposed structures increases, the ability to effectively analyze the conservation
of water molecules decreases dramatically. In general, the water molecules located at the
same binding site will be somewhat shifted in position due to minor changes in neighboring
protein atom positions and due to minor variations in both the actual location of the water
molecule in the crystal and variations in its placement by the crystallographer. Visualiza-
tion of more than a few structures simultaneously will cause the waters to become nearly a

continuous shell of hydration, losing all definition of preferred sites.

41

A second, quantitative approach is to use a single, chosen structure as a reference to
judge conservation in a series of related protein structures. This approach has been used to
study the solvation of FKBP12 complexes with the irnmunosuppressant FK506 (Faerrnan
and Karplus, 1995) and the solvation of T4 lysozyme (Zhang and Matthews, 1994). Water
site conservation is defined based on sites which occur in the reference also occurring
in the remainder of analyzed structures, i.e., are the water sites located in the reference
structure also observed in the other structures? This causes the results to be dependent on
which structure of a homologous set is chosen as the reference. The AquariusZ algorithm
(Pitt et al., 1993) uses a knowledge base of protein-water molecule interactions, with each
interaction tabulated and referenced to a set of common functional groups and side chains,
in a series of unrelated protein structures to derive a three-dimensional probability map for
locating bound water sites in protein structures in general. Analyses of water molecule
binding sites remain subject to limitations in crystallographic fitting and refinement (Levitt
and Park, 1993; Karplus and Faerman, 1994), but limitations due to assignment in any
given single structure can be minimized through the use of multiple, independently solved

structures as a knowledge base for analysis and design.

This section presents work employing the statistical method of complete linkage hierar-
chical clustering to define consensus water sites in thrombin, trypsin, and bovine pancreatic
trypsin inhibitor (BPTI), with the goal of determining the extent to which water sites are
conserved for each protein and between the two serine proteases and their relationship to
ligand binding. This technique circumvents the problem of using the bound water sites

flom a single structure as a reference set, because all sites flom each of the different pro-

42

tein structures are equally weighted in cluster analysis. Thrombin was chosen as a protein
of focus for several reasons: there are a number of structures solved at good resolution
with different ligands bound, thrombin is an important pharmaceutical target for regulating
blood coagulation, and highly conserved water molecules are known to surround the bind-
ing site of its allosteric regulator, Na” (Di Cera et al., 1995; Zhang and Tulinsky, 1997).
Thrombin is a serine protease at the junction between blood coagulation and anticoagu-
lation pathways and can initiate both processes, reviewed by Furie and Furie (1988) and
Esmon (1992). In addition to binding its receptor, proteolytic substrates, and several phys-
iological inhibitors, thrombin also binds exogenous inhibitors such as hirudin (produced
as an anticoagulant agent by leeches) and D-Phe—Pro-Arg chloromethylketone (PPACK), a
substrate transition-state analog. Thrombin contains two major ligand binding sites: the
active site, which binds fibrinogen at the cleavage site, and an exosite, which provides ad-
ditional substrate binding surface, enhancing the aflinity for fibrinogen and hirudin and its
analogs (Vijayalakshmi et al., 1994). The variety of crystallographic proteinzligand com-
plexes available for thrombin provides the ability to study water sites that are conserved

regardless of ligand, as well as those water sites that are ligand specific.

Another goal was to identify water sites that are shared by thrombin and trypsin, a serine
protease not involved in blood coagulation, in order to identify water sites that are essential
in serine proteases and also point to water molecules that are specific to thrombin or trypsin
ligand-binding sites. Trypsin is a serine protease which proteolytically activates other di-
gestive proteases. The loop which binds Na+ in thrombin cannot bind Na+ in trypsin due

to a change in conformation and chemistry associated with the T‘yr 255 to Pro sequence

43

 

change (Dang and Di Cera, 1996). This results in the absence of allosteric regulation
by Na+ in trypsin. Several high-resolution trypsin structures are available in the Protein
Data Bank (PDB), and its water structure has been studied via several techniques includ-
ing room-temperature and low-temperature X-ray crystallography (Earnest et al., 1991),
neutron-diffraction (Finer-Moore et al., 1992), and D2 O-HgO difference neutron diffraction
(Kossiakofl‘ et al., 1992). BPT I is a natural inhibitor of trypsin and its water interactions
have been studied using NMR and molecular dynamics (van Gunsteren et al., 1983; Brunne
et al., 1993; Denisov et al., 1996) and simultaneous NMR and X-ray difflaction refinement
(Schiffer et al., 1994). Several high-resolution structures of BPTI are available in the PDB,
along with X-ray diffraction structures of the trypsinzBPTI complex, providing the ability
to examine the fate of water molecules bound to free trypsin and flee BPTI upon formation
of the trypsin:BPTI complex. Given the wealth of structural information available for these
serine proteases, they provide an ideal system for testing the technique presented here for

determination and analysis of conserved water binding sites.

2.3 Water Site Clustering Methods

2.3.1 Structure Selection

Thrombin, trypsin, and BPTI structures were selected flom the Protein Data Bank based
upon the absence of unusual crystallization conditions (e.g., low pH), sequence insertions,
deletions, or point mutations, and a resolution of 52.0 A for trypsin and BPTI and 32.4 A

for thrombin. Ligand-flee structures were selected; no ligand-flee structure thrombin struc-

44

 

tures were available, but 6 of the 10 structures analyzed here have no ligand in one of the
two sites, the active site or the exosite. The availability of 10 thrombin structures for anal-
ysis helped compensate for their somewhat lower resolution as compared to the trypsin
structures. Visual inspection of the superimposed molecules for each protein eliminated
those with regions of large structural deviation likely to affect water site conservation.
Only structures with refined water molecule positions were included. The quality of water
refinement was assessed using a mobility measure designed to normalize and combine the
crystallographic temperature factor (Debye-Waller factor; B-value) and the occupancy as

defined below (Craig et al., 1998):

MObilitywater molecule =

B'Valuewater molecule/ Average B-valuegn waters in structure (2.1)
OCCUPancywater molecule/ Average Occupancyan waters in structure

This facilitates comparison of atomic mobility between protein structures refined with dif-

ferent protocols, in particular, those structures in which occupancy as well as B-value were

allowed to vary during refinement of the water molecules.

Using this normalization, a water molecule (or other atom) with a high degree of rigidity
has a mobility value near 0, a water molecule with average mobility relative to other atoms
in the protein has a mobility value of 1, and a highly mobile water molecule has a mobility
value greater than 1. In general, if a water molecule’s mobility is 2:, then it is 2: times
as mobile as the average water molecule. In practice, the mobility of a water molecule
is determined by its oxygen atom, since hydrogen atom positions are not assigned in the

majority of structures. Histograms of the water mobility values for each structure showed

45

 

whether there were a number of water sites with high mobility (>2); a preponderance
of such sites was found, by analysis of inter-water molecule distances, to indicate water
molecules placed too close to each other (<2.6 A). Such structures were excluded flom
this analysis. AS an example, Figure 2.1 compares the mobility distributions of water sites
in two BPTI structures. Structures selected using all the above criteria and selected for

this work are presented in Table 2.1.

2.3.2 Hierarchical Clustering

The following steps were‘performed independently for the thrombin, trypsin, and BPTI
structural sets (Table 2.1). The chosen structures were superimposed onto a reference struc-
ture using main-chain least-squares superposition in Insight” (Accelrys, San Diego, CA)
to transform the protein structure and water molecules into the same reference flame. The
2:, y, z coordinates for these transformed water molecules were then extracted and used
for clustering. Clustering is in an iterative process. The first step is to generate a ma-
trix of all inter-element distances. Here, the simple Euclidean distance between points is
used, though, in general, any distance metric can be used. The first cluster is then formed
flom the two closest elements and the distance between this initial cluster and the remain-
ing elements is calculated. Once again, the two closest elements, one of which could be
the previously formed cluster, are joined into a new cluster. The process repeats until all
elements are joined in a single cluster or, as is the case here, until a distance threshold,
representing the maximum distance between any of the elements assigned to a single clus-

ter, is reached. There are several methods of calculating the distance between a cluster of

46

 

 

 

30%

 

- 1pr
D 4pti

 

 

 

25% .2

 

 

 

 

 

20% .2

 

‘1 5%

 

10% ~

 

Percentage of Water Molecules

 

 

5%

 

 

 

 

 

 

 

 

. I . I 4“—
3.5 4

 

0%
. 1 1 .5 2 2 .5 3
Average to Law Mobility
Mobtlty Water: I I High Moblty Waters

Figure 2.1: Mobility distributions of two BPT I structures, used as a quantitative tool to
screen for structures with uncertain water positions. The 4pti distribution is narrow and
shows that most water molecules have nearly average mobility and none are highly mobile.
The lbpi distribution is broad and has an extended right tail, indicating the presence of a
number of water molecules with high mobility (22; at least twice as mobile as average).
Further analysis showed one-half of these high-mobility sites could be explained by the
occupancies of overlapping sites summing to 51, suggesting that they represent alternate
locations of a single water molecule. However, including multiple copies of single wa-
ter molecules corresponding to their different, partially occupied sites would introduce a
statistical bias into the cluster analysis and has been avoided in this work.

 

47

 

Table 2.1: Database of thrombin, trypsin, BPTI, and trypsin:BPTI structures for analysis of
conserved water sites

 

 

 

 

Ligand Binding Site
PDB Active F ibrinogen Resolution Main-Chain Number of
Code Site Binding Site (A) RMSD Crystallographic

(Exosite) (A) Bound Waters

Thrombin StructuresI
lhai PPACK 2.4 0.000 194
labj PPACK 2.4 0.694 196
1 ppb PPACK 1.9 0.802 409
ltrnb Cyclotheonamide A Hirugen 2.3 0.560 239
lhah Hirugen 2.3 0.345 205
ltmt —— CGP50,856 —— 2.2 0.458 1 l l
labi —— Hirulog-3 —— 2.3 0.409 246
lthr Hirullin 2.3 0.350 190
lths MDL-28050 2.2 0.439 140
lihs —-———— Hirutonin-Z —— 2.0 0.481 146
Trypsin Structures”
ltpo 1.7 1.395 84
2pm 1.6 0.103 82
3pm 1.7 0.266 82
BPTI structures3
4pti 1.5 0.000 60
5pti" 1 .0/ 1.8 0.403 63
6pti 1.7 0.436 73
9pti 1.2 0.418 67
Trypsin/BPTI complex structures5
2ptc 1.9 0343/0479 157
ltpa 1.9 0.336/0.638 159

 

 

lSuperpositions and RMSDS are relative to lhai.

’Superpositions and RMSDS are relative to ltpo, except for ltpo which is relative to lhai.

’Superpositions and RMSDS are relative residues 1—46 of 4pti.

“Resolution is for X-ray difflaction/neutron diffraction data.

5RMSDS are reported for the trypsin chain of the complex superimposed onto ltpo and for
the BPTI chain of the complex superimposed onto 4pti.

 

48

multiple elements and other clusters or remaining unclustered elements: (1) the shortest
distance between any pair of elements in each of the clusters (single linkage), (2) the dis-
tance between the cluster centroids, which are the mean 2:, y, z coordinates of the elements
of each cluster (average linkage), and (3) the maximum distance between any pair of ele-
ments in each of the clusters (complete linkage). Complete linkage clustering was chosen
for this work as it yields compact, globular clusters and allows the specification of a maxi-
mum diameter for any cluster by defining the maximum distance between cluster elements.
This ability is useful when defining water sites as it can ensure the water molecules flom
different structures which contribute to a cluster can form the same approximate hydrogen

bonds and are within hydrogen bond forming distance (2.4 A).

An example cluster analysis for a subset of the water molecules flom the BPT I structure
set is shown in Figure 2.2. Complete linkage clustering begins by placing the two closest
elements together in a cluster; 6pti 108 and 4pti 108 are less than the maximal distance of
2.4 A apart, the basis for this threshold is given below, and are grouped into a cluster (ar-
bitrarily numbered 109). Next, the distance between this cluster and each of the remaining
data elements is computed; for complete linkage clustering, this is defined as the maximum
distance between that element and all the elements in the clusters. In Figure 2.2, the dis-
tance between 4pti 139, 9pti 103, and 6pti 238 (which are not yet clustered) and cluster
109 is calculated as the distance to 4pti 108, since it is the furthest element of cluster 109.
This process is repeated until no further elements can be clustered without exceeding the
selected maximum distance, 2.4 A in this work. Any elements not included in clusters

at this point are considered to define single-element clusters; for example, cluster 134 in

49

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

4pti 139
3 .\ }Cluster 108
2.4 A Q
______________ ..______-_(Z”I°I'--_ 9pti103

i

8 2

C

5 .

.2 Clurster 134
D A
a .

’6 6pti 238.
3

.I. 1 A

9 6pti 108.19”

5 O 3?

a»
..‘.-'Q?‘"
o 0‘) ID 0
s 2 2 2 a was
i 6* IS 3:. I? u...-
Cluster 109

Cluster 108 Cluster 109 Cluster 134

Figure 2.2: Example of complete linkage clustering applied to water sites flom several
BPTI structures. A portion of the BPTI clustering tree is shown at left, based on the inter-
water distances flom the superimposed BPTI structures, shown at right. Note that 6pti
water 108 and 6pti water 238 are not clustered together even though they are closer than
the cutoff distance of 2.4 A, since 6pti 108 belongs to a cluster in which one water (4pti
108) is too far flom 6pti 238 to meet the 2.4 A threshold. This feature of complete linkage
clustering guarantees that no cluster contains water sites separated by more than 2.4 A.
At this distance, all water molecules in a microcluster are overlapping, and it is unlikely
that more than one water molecule will be included flom any given protein structure (they
would be too close). Cluster numbers are arbitrary sequential indices, whereas individual
water molecules are labeled by the residue number flom the corresponding PDB file.

 

50

.57”

Figure 2.2 consists only of water molecule 6pti 238.

A maximum diameter of 2.4 A was chosen, resulting in clusters with a maximum
inter-water distance of 2.4 A, as measured flom oxygen center to oxygen center. This
value was chosen because water molecules have an approximate effective radius of
1.6 A, which includes the radius of the oxygen and a correction for the contribution
of the hydrogen atoms, whose positions are typically unknown. Thus, if two water
molecules are placed with their oxygen atoms at a center-to-center distance of 2.4 A,
their radii will overlap by 50%. This almost always prevents water sites flom the same
structure flom being included in the same cluster, since at <2.4 A apart, they would
be positioned too closely. Complete linkage clustering results in the set of maximally
dense clusters (in terms of average number of water molecules per cluster). These will be
referred to as “microclusters” to emphasize that all water molecules within a single cluster
physically overlap. The WatCH (Waters Clustered Hierarchically) software package
developed in this work is implemented in C and has been made available via the intemet at

http: / /www. bch . msu . edu/ labs /kuhn/web/software/WatCH/doc . html.

2.3.3 Crystal Contact Calculation

To observe the possible effects of crystal contacts on water site conservation, crystal con-
tacts in the seven thrombin structures in space group C2 were calculated using Chain (Sack,
1988), where interactions were included for protein crystal lattice symmetry mate atoms
within 4.0 A. Crystal contact residue and atom lists were generated for each of the crys-

tallographic structures, with water sites represented by the microclusters observed in that

51

 

protein. The number of times each microcluster appeared in a crystal contact was calcu-
lated for the seven thrombin structures, and software was developed to convert Chain’s
crystal contact lists and the microcluster lists into InsightII subsets for visualization of the

spatial relationship between crystal contacts and microcluster conservation.

2.3.4 Evaluation of Bound Water Environments

The degree of conservation of the water microclusters, each representing a favored site
for water binding, was calculated as the number of individual water molecules contained
in the microcluster divided by the number of structures used for clustering. To assess
the influence of the shape and chemistry of the water binding site on its conservation in

different structures, measures of eight environmental features were calculated:

0 atomic density (ADN), measured as the number of protein atoms within van der
Waals packing distance, 3.6 A, of the water molecule, which correlates with whether
the site is in a groove (high density of protein neighbors) or a protrusion (low density

of neighbors) (Kuhn et al., 1992b);

0 local atomic hydrophilicity (AHP), measured by the sum of the atomic hydrophilicity

of all protein and water atoms within 3.6 A of the water site (Kuhn et al., 1995);

c crystallographic temperature factor (Debye-Waller factor”, B-value; BVAL), a mea-
sure of the atom’s thermal mobility and spread in the in the electron density, read

flom the protein’s PDB file;

the number of hydrogen bonds to neighboring protein atoms (PrHBD);

52

o the number of hydrogen bonds to neighboring water molecules (WatHBD), using a

distance of 33.5 A between donor and acceptor atoms;
0 the water site mobility (MOB), a normalized measure of mobility (see Section 2.3.1);

o the summed B-values for all protein atoms within 3.6 A of the water site (TPrBVAL);

and

o the average B-value for these neighboring protein atoms (AngrBVAL).

Several of these features are related, and the goal here was to see determine which features

best correlate with degree of water site conservation.

For each microcluster, the value for each of the eight features was averaged over the
individual environments of its water molecules. To assess the correlation between conser-
vation of the nricroclusters and their environments, feature values were also averaged over
all microclusters with a given degree of conservation (e.g., those containing waters flom 6

of 10 structures).

2.3.5 Calculation of Overlapping Microclusters between Thrombin

and Trypsin

Distances were calculated between the centroids of microclusters in the superimposed
structures of thrombin and trypsin, and overlapping clusters were defined as those with
a centroid-to-centroid distance of 31.8 A. “fith a maximum diameter of 2.4 A for each mi-

crocluster, the microclusters’ radii overlap by 50% when their centroids are within 1.8 A.

53

 

To analyze the effect of using different overlap criteria and provide a list of nricrocluster
overlaps between thrombin and trypsin using less stringent criteria, overlapping micro-
clusters were also tabulated using thresholds up to 2.4 A (where two microclusters would
just touch). To determine the significance of the observed number of overlapping sites
between thrombin and trypsin, in a separate experiment, microcluster centroids were ran-
domly placed in the thrombin structure at the density of rrricroclusters experimentally ob-
served for thrombin, and the same was done for trypsin. Then, the number of overlaps
between thrombin and trypsin was calculated using these random distributions. Because
many of the water sites in thrombin and trypsin are buried in the proteins, the microcluster
density was calculated based on the number of microclusters per A3 of protein volume,
which was calculated for each protein using the PQMS routine of the Molecular Surface
Package, version 2.6 (Connolly, 1983). Random placement of microcluster centroids and
subsequent counting of overlaps was repeated 100 times to obtain statistical means and
standard deviations for the number of overlaps as a function of overlap criterion (1.8 to
2.4 A). For analyzing conserved water site proximity to functionally important sites (e.g.,
residues in the catalytic triad), a distance threshold of 3.6 A flom the microcluster to the
functionally important atom(s) was used. Interaction with an active-site or exosite ligand

was determined by measuring the distance to all ligands bound in the structure.

Images in this dissertation are presented in color.

54

 

2.4 Results

2.4.1 Clustering Statistics

To identify shared versus unique conserved water sites for thrombin and trypsin, complete
linkage clustering was performed on their respective water sites (Table 2.2). Clustering of
2,075 water sites flom the ten thrombin structures yielded 708 microclusters with an aver-
age of 2.93 waters each, indicating that the average water site was observed in 29.3% of the
structures. Of the 708 microclusters, 18.5% were found in at least half of the 10 structures.
Clustering of 248 water sites flom the three trypsin structures yielded 106 microclusters,
conserved on average in 78.0% of the structures. Of these microclusters, 56.6% were ob-
served in all three structures. This high degree of conservation was surprising, but two of
the structures (PDB codes ltpo and 2pm) were solved by the same crystallographers and
have very similar water sites; however, mobility plots (data not shown) indicated that the
water assignments in both structures were reasonable. (Consideration was given to ana-
lyzing additional trypsin structures, but there were only three ligand-flee, wild-type bovine
structures solved under typical crystallization conditions.) Given the similarity in water
assignments for ltpo and 2ptn, trypsin water sites were considered to be highly conserved
only if they appeared in all three structures. A similar analysis of BPTI clustered 263 water
sites flom four structures into 134 microclusters, with an average conservation of 49.0%.

Of these microclusters, 54.5% were found in at least half of the BPTI structures.

55

 

Table 2.2: Clustering statistics

 

 

Thrombin (10 superimposed structure)
Number of water molecules
Number of water clusters
Average conservation (waters/cluster)
Number of clusters with 250% conservation
Number of clusters with 100% conservation
Mean protein volume (A3)
Cluster density (clusters/A3)
Conserved cluster1 density (clusters/A3)

Trypsin (3 superimposed structure)
Number of water molecules
Number of water clusters
Average conservation (waters/cluster)
Number of clusters with 250% conservation
Number of clusters with 100% conservation
Mean protein volume (A’)
Cluster density (clusters/A3)
Conserved cluster1 density (clusters/A3)

BPTI (4 superimposed structure)
Number of water molecules
Number of water clusters
Average conservation (waters/cluster)
Number of clusters with 250% conservation
Number of clusters with 100% conservation
Mean protein volume (A’)
Cluster density (clusters/A3)
Conserved cluster1 density (clusters/A3)

 

2075
708

2.93 (29.3%)
131 (18.5%)
28 (4.0%)
38.27
0.0185
0.0034

248
106
2.34 (78.0%)
82 (77.3%)
60 (56.6%)
27.21
0.0039
0.0030

263
134
1.96 (49.0%)
73 (54.5%)
18 (13.4%)
7.31
0.0180
0.0100

 

lConserved clusters are those with conservation 250%.

 

56

 

Table 2.3: Linear correlation coefficients between degree of conservation and environmen-
tal features

 

 

 

Environmental Correlation Coeffrcients

Feature Thrombin Trypsin BPTI Combined
ADN 0.412 0.451 0.418 0.408
AHP 0.483 0.467 0.304 0.475
BVAL -0.377 0.443 -0.355 -0.474
PrHBD 0.457 0.463 0.439 0.485
WatHBD 0.218 0.075 -0.040 0.177
MOB -0.525 0.450 -0.458 -0.478
TPrBVAL 0.068 0.371 0.146 -0.061
AngrBVAL -0.387 0.028 -0.328 -0.514

 

 

 

2.4.2 Environmental Analysis

Analysis of water site environments provided insights into the determinants of conserved
water binding. All protein-bound microclusters, i.e., those containing at least one water
molecule making direct contacts (33.6 A) with the protein, were analyzed. There were
521 protein-bound microclusters for thrombin, 98 for trypsin, and 117 for BPTI. Highly
conserved water molecules occupied somewhat different environments than less conserved
environments (Figure 2.3). Linear correlation coeflicients for each feature are given in
Table 2.3. Conserved microclusters had more neighboring protein atoms (atomic density;
ADN), made more hydrogen bonds to the protein (PrHBD), and were in a more polar envi-
ronment, indicated by more hydrophilic neighboring atoms (atomic hydrophilicity; AHP).
Most mobility measures, the water site’s B-value (BVAL), its mobility (MOB), and the av-
erage B-value of the neighboring protein atoms (AngrBVAL), were negatively correlated

with conservation, indicating that water sites with high conservation tend to reside in less

57

 

 

 

 

 

 

 

A — ADN

10 _. f —I . _ I I I n r n /l\ . —— AHP
g . w- \ r « \ \ , BVAL
'6 8 _ x ~ PrHBD
> . ‘ q ’ , X i. — — WatHBD
g \ ///’ \2 - 7 MOB
'5 ~/_’\ “ - » TPrBVAL
o - ' -. A P VAL
LL. \ \ N\\ j vg r8
'0 \ "\.
2 \ \
8 \ \ 2 \\7
0) ‘Pl

 

 

 

0 l ‘ l I r l I
10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Water Site Conservation

 

 

 

 

 

 

B

- - J - 7 I I I I fii
d) T T ‘ ~— _ T 7 2 y
2 T T \ \ /'/
2 \ \ E /’/
3 E s1 "
8 / I y > ;.
LL / , ’ ’ \ > \ —
E 2,..—-//// IKK
8 A / / \ \T
m - ,

1 1 1 ‘ ’ l
43% 53% 63% 73% 83% 93%

C Water Site Conservation
0:
3
(6
>
9
2
8
LL
8
(I)

 

 

 

o l I A L I I I
25% 35% 45% 55% 65% 75% 85% 95%
Water Site Conservation

Figure 2.3: Correlation between water site conservation and environmental features for (A)
thrombin, (B) trypsin, and (C) BPTI. Shown are the average values of eight environmental
features for the water microclusters as a function of their degree of conservation. Features
are those described in Section 2.3.4. The feature values have been averaged within mi-
croclusters as described in Section 2.3.4, averaged over the microclusters with the given
degree of conservation, and normalized to range between 1 and 10 to allow visualization
on the same plot. The curve of AHP for trypsin superimposes with that of PrHBD, and is
therefore not apparent on the plot. Approximately linear correlation with conservation is
seen for many of the features, as described in Results.

 

58

mobile portions of the protein. The number of hydrogen bonds to other water molecules
(WatHBD) did not correlate strongly with the conservation, suggesting that consensus wa-

ter sites are not strongly stabilized by hydrogen-bonded water networks.

2.4.3 Effects of Crystal Contacts on Bound Water Conservation

The effects of crystal contacts upon water binding were examined by spatially correlating
water site conservation with contacts in the protein lattice. To address whether water sites
were preferentially excluded flom or trapped in these contacts, the location of conserved
water sites along with the crystal contact residues for the seven thrombin structures in the
C2 space group were visualized. Crystal contacts had fewer conserved water sites than
surrounding areas, consistent with the observed expulsion of interfacial bound water upon

dirnerization of chyrnotrypsin (Blevins and Tulinsky, 1985).

2.4.4 Spatial Analysis of the Conserved Microclusters

To explore how microclusters of different conservation levels are distributed spatially
around the protein, molecular graphics visualization was used. For thrombin, a concen-
tration of highly conserved microclusters (in 250% of the structures; yellow spheres in
Figure 2.4) was found near the sodium site but not observed in the active site, perhaps
due to water displacement by the presence of active-site ligands in 7 of the 10 structures.

Other conserved microclusters were observed in deep grooves or cavities within the pro-

tein, as expected flom the known correlation between water site conservation and groove

59

 

 

Figure 2.4: Conserved water sites in thrombin. Thrombin microclusters containing water
sites from three of the ten structures are colored blue, sites found in four are green, and
sites found in at least five are yellow. The backbone ribbon of lhai is shown colored by
B-value (dark blue equals a B-value of 0, white ~30, red ~50, and yellow >50 A“). The
catalytic triad Asp, Ser, and His side chains are rendered as pink tubes at center. PPACK,
an active-site ligand lhai. is shown in blue tubes, and hirugen, an exosite ligand from lhah,
is shown in green (structurally conserved region) and orange (structurally divergent region)
at right. The sodium ion (labeled as water 410 in lhai) is rendered as a large blue sphere
at lower left. A concentration of conserved water sites exist near the sodium site and its
channel, at lower left; many other sites are buried.

 

60

topography (Kuhn et al., l992a) and previous studies on the conservation of buried water
molecules in serine proteases (F iner-Moore et al., 1992; Rashin et al., 1986; Meyer, 1992;
Sreenivasan and Axelsen, 1992). When the exosite ligands were superimposed, a struc-
turally conserved region, comprising the six N-terminal ligand residues (green tubes at the
bottom right of Figure 2.4, and a structural variable region, extending flom the seventh
residue to the C-terminus of the ligand (orange tubes are rightmost edge of Figure 2.4),
were found; water sites associated with the structurally conserved region in the exosite lig-
ands were also generally conserved. Similar patterns of buried water site conserved were

observed for trypsin.

Given the functional importance of the Na+ binding site for switching between the co-
agulant (Na+ bound) and anticoagulant (water bound) forms of thrombin (Di Cera et al.,
1995), this region of the structure was analyzed in detail. The Na”r sites in structures lhai
and lhah assigned by Zhang and Tulinsky (1997), which were originally labeled as water
molecules in the PDB structures and later confirmed by rubidium replacement to represent
a Na+ site (Di Cera et al., 1995), occur in two overlapping microclusters (centroids 1.2 A
apart) containing the Na+/water molecules flom all 10 structures. The 38 water micro-
clusters in the channel coupling the Na+ site with the active site are >50% conserved on
average, consistent with the recent discovery of this conserved solvent channel (Zhang and

Tulinsky, 1997).

61

2.4.5 Overlapping Water Sites between Thrombin and Trypsin

To define water sites shared between these serine proteases involved in distinct biochem-
ical pathways, overlaps between conserved (2 50%) Sites in thrombin and trypsin were
evaluated (Table 2.4 and Figure 2.5). The number of overlapping water microclusters in
thrombin and trypsin, 37, is statistically significant, since only 7.9 overlaps would be ex-
pected if the conserved water sites in thrombin and trypsin were distributed randomly (see
2.3.5). Seven of the conserved microclusters were in the active-site region, four being
near at least one of the catalytic triad residues. Three overlapping clusters were near the
Na” binding site of thrombin, with two additional ones in the surrounding solvent channel.
Conservation of solvent in this region (Figure 2.5, lower left), which regulates the coag-
ulant/anticoagulant function of thrombin via Na+ binding displacement, suggests it may
also be important in trypsin. To assess whether water Site conservation between thrombin
and trypsin is associated with conservation of nearby side chains and their conformations,
the 37 shared waters sites were evaluated in the context of PDB structures lhai (thrombin)
and ltpo (trypsin). Ninety-two percent of the shared water sites had chemically and con-
formationally similar environments, based on no more than one side-chain substitution and
no more than one residue with a significant (1.5—2 A) shift. Larger shifts were considered
structurally dissimilar, yet even substituted side chains tended to be similar through the
7-carbon. Of the 37 shared sites, 38% were structurally very similar, with no side-chain
substitutions and no positional shifts exceeding 1.5 A. Thus, conserved protein structure
between thrombin and trypsin largely accounted for their water site conservation, which

can be considered a shared feature of their structure and function as serine proteases.

62

 

Table 2.4: Overlapping conserved water sites between thrombin and trypsin

 

 

 

 

 

 

 

Thrombin Trypsin Distance
Cluster Percent Representative Cluster Percent Representative between
Number Conserved1 Water Number Conserved1 Water Thrombin
Residue Residue and
Numberz Numbera Trypsin
Centroids"
(A)
1021 70 570 25 100 470 0.16
899 LSC5 100 417 45 100 415 0.30
996 100 423 22 100 717 0.41
954 100 461 5 100 430 0.43
1 197 70 515 54 67 806 0.48
1017 LC 100 407 33 L5 100 416 0.52
935 A5 100 430 6 A5 100 703 0.54
857 A 100 445 20 100 701 0.56
951 A 100 468 19 100 408 0.56
874 100 401 28 100 708 0.56
970 50 5516 72 100 752 0.57
885 100 404 18 100 721 0.62
888 LC 100 403 31 100 704 0.65
926 100 414 30 100 429 0.67
1214 80 480 21 67 751 0.72
1075 80 489 59 100 736 0.75
948 50 5547 62 67 754 0.75
852 90 441 13 100 473 0.76
1016A 100 436 10L 100 410 0.77
963 100 439 17 100 722 0.78
1051 70 455 66 100 728 0.80
878 100 405 9 100 406 0.82
972 SC 90 448 35 100 705 0.86
1032 80 469 55 67 803 0.94
981 100 467 38 100 709 0.96
955 100 412 29 100 716 0.99
1139 70 537 42 100 530 1.01
1150 E5 90 507 8 67 738 1.06
832 50 458 65 67 801 1.10
Continued on next page.

 

63

 

Table 2.4 (cont’d)

 

 

 

 

 

 

Thrombin Trypsin Distance

Cluster Percent Representative Cluster Percent Representative between

Number Conserved1 Water Number Conservedl Water Thrombin
Residue Residue and

Number2 Number3 Trypsin

Centroids"
(A)
1111 90 546 60 100 744 1.11
1086 SC 70 4096 44 100 562 1.14
890 90 406 24 100 5 16 l .29
870 90 443 4 100 726 1.32
916 80 452 34 100 604 1.40
921 100 413 1 100 746 1.58
1038 90 451 2 100 741 1.61
1 108 80 539 16 100 725 1.66
l 150 90 507 71 100 733 1.86
1259 70 426 52 67 750 1.93
1053 80 457 56 67 735 2.10
981 100 467 24 100 516 2.13
1170 60 494 66 100 728 2.15
964 SC 90 450 35 100 705 2.15
948 50 5547 65 67 801 2.19
1032 80 469 37 100 720 2.25
827 100 446 18 100 721 2.29
995 90 43 1 27 67 743 2.30
1119 50 505 2 100 741 2.34
857 100 445 10 100 410 2.36

 

 

1Only clusters with at least 50% conservation are tabulated.
2Representative thrombin waters are flom lhai unless there is no member water flom lhai,

in which case the source structure is noted.
aRepresentatrve trypsrn waters are flom ltpo.

‘A line divides the table into highly overlapping water microclusters with centroids g 1.8 A
apart (cluster radii overlap by 250%), shown in the top section of the table, flom some-
what overlapping microclusters with centroids 1.8-2.4 A apart.

5Labels indicate overlapping clusters that interact with (are 33.6 A flom) active site lig-
ands (L), active-site catalytic triad residues (A), Na+ channel waters (C), exosite ligands
(E), or Na+ site (S).

°Representative thrombin water is flom lhah.

7Representative thrombin water is flom labj.

 

64

Several water sites were highly conserved in firnctionally important regions of throm-
bin or trypsin, but were not shared between them (Table 2.5). These may contribute
to their specificity differences. Four more microclusters were specifically associated with
active-site ligands in trypsin than were seen in thrombin, in reflecting the larger inhibitor
in trypsin; 13 residues of BPTI interact with trypsin, whereas the thrombin active-site lig-
ands are only three to seven residues long. Five Na" binding site and channel clusters were
shared between thrombin and trypsin (Table 2.4); however, 15 conserved sites in this region
were found only in thrombin (Table 2.5). Combined with the five conserved exosite water
positions found uniquely in thrombin and eight active-site water positions found uniquely
in trypsin (Table 2.5), it is apparent that bound water can make a significant contribution to

ligand specificity.

2.4.6 Contribution of Conserved Water Molecules to the

TrypsinzBPTI Complex

Trypsin provides an ideal system to test the applicability of a lock-and-key mechanism for
the contributions of protein-bound and 1i gand-bound water molecules to serine protease
complex formation because several high-resolution ligand-flee structures are available for
trypsin, BPTI, and the trypsin:BPTI complex. Using water nricroclusters identified for
trypsin and BPTI, the conserved water sites flom each protein were compared with water
sites conserved in the complex structures (2ptc and lpta). The flee trypsin structures were
superimposed onto the trypsin chain of the 2ptc complex, and the flee BPTI structures

were superimposed onto the BPTI chain of 2ptc. Three conserved microclusters flom the

65

 

 

Figure 2.5: Overlapping conserved water sites between thrombin and trypsin. Water sites
conserved in at least half of the structures of thrombin (water sites shown as blue spheres)
or trypsin (red spheres) are shown. It is important to note that each cluster shown cor-
responds to one cluster in an overlapping pair; nonoverlapping clusters are omitted from
the figure. The backbone of thrombin (represented by lhai) is shown as a magenta rib-
bon and the backbone of trypsin (represented by ltpo) is shown as a red ribbon. Catalytic
triad residues are shown as white tubes (center of figure), PPACK (a thrombin active-site
inhibitor) is shown in blue, and hirugen (a thrombin exosite inhibitor) is shown in green
and orange (structurally conserved and divergent regions, respectively). The region of the
trypsin inhibitor BPTI which contacts trypsin is shown in yellow and superimposed from
2ptc; note the conformational similarity between PPACK and BPTI, extending downward
from the Pro residue of PPACK. The Na+ from lhai is rendered as a large blue sphere at
lower left. A concentration of overlaps between conserved thrombin and trypsin water sites
is observed near the Na+ site, despite trypsin having no known functional similarity here;
these conserved water molecules form a network which extends towards the active site.
There is also a number of overlapping sites located in the exosite, though these are more
spatially spread.

 

66

 

Table 2.5: Functionally relevant conserved water Sites unique to thrombin or trypsin

 

 

Representative
Cluster Percent Water Residue
Number Conservation‘ Number2

 

Thrombin
Active Site Catalytic Triad Residues
No Nonoverlapping Conserved Water Sites

Active Site Ligands
1153 100 408
1196 90 428
Exosite Ligands
821 60 560
949 50 576
1179 70 415
1241 50 490 (lhah)
1278 70 496 (lhah)
Na+ Binding Site
1195 80 418
976 100 424
1121 90 482
838 100 514
Na+ Channel Waters
1001 100 409
1195 80 418
976 100 424
1196 90 428
788 70 463
914 100 464
944 50 474
1121 90 482
915 90 497
838 100 514
1229 60 457 (lhah)

 

 

Continued on next page.

 

67

 

Table 2.5 (cont’d)

 

 

 

Representative
Cluster Percent Water Residue
Number Conservation‘ Number2
Trypsin
Active Site Catalytic Triad Residues
48 100 747
80 100 702
Active Site Ligands
80 100 702
61 100 710
48 100 747
64 67 807
23 67 808
77 100 805

 

 

lOnly waters with 250% conservation are tabulated
1"Representative thrombin waters are flom lhai unless noted. Representative trypsin waters
are flom ltpo.

 

flee structures overlapped with the conserved water sites in the complex (large spheres
in Figure 2.6), two being contributed by trypsin and one by BPTI. Thus, three of the
seven trypsin:BPTI interfacial water molecules were donated by the flee proteins, while
four were newly recruited or shuffled upon complex formation. This contrasts with the
contributions of water molecules bound to the flee structures of lysozyme and the D1 .3
antibody, which contribute 20 of the 25 water molecules observed in the antibody-lysozyme
interface (Braden et al., 1995). Thus, the hydration structure of the flee protein and ligand
and the creation of new environments favorable for water binding upon docking of the
protein and ligand should both be considered in inhibitor design. One approach to doing so

in the pattern recognition application Consolv (Raymer et al., 1997).

68

 

 

Figure 2.6: Conservation of water sites in the uypsinzBPTI interface. This close-up of the
interface between trypsin (red ribbon, from PDB ltpo) and BPTI (blue ribbon, PDB 4pti),
superimposed into the analogous chain of the trypsin:BPTI complex (PDB 2ptc), shows the
conservation of water sites between the free structures and their complex. The orientation
is approximately a 90° rotation about the horizontal axis relative to Figure 2.5. Conserved
(250%) water sites from the free trypsin structure are shown as red spheres, those from
BPTI shown in blue, and interfacial waters found in both structures of the trypsinzBPTI
complex (PDB 2ptc and ltpo) are shown in green. Water sites overlapping between the
complexes and free structures are rendered as large spheres, while non-overlapping sites
are rendered as small spheres. The catalytic triad of trypsin is shown in white, and the side
chain of the inhibitory Lys 15 from BPTI is shown in cyan. Of the seven trypsinzBPTI
interfacial water sites, three are contributed by either trypsin or BPTI.

 

69

2.5 Discussion

2.5.1 Conservation of Water Sites in Thrombin and Trypsin

A number of water sites were conserved in at least half of the thrombin and trypsin struc-
tures, and several sites were found in all of the structures examined (Table 2.2). An earlier
detailed study of the solvent structure of trypsin (Finer-Moore et al., 1992) defined 211
consensus water sites via high-resolution X-ray difflaction data for the waters’ oxygen
atoms, verified by D2 O-H20 difference neutron scattering density for the waters’ hydrogen
atoms. In this work, significantly fewer consensus water sites were identified, 60, perhaps
due to comparing three structures. A key goal was to distinguish conserved water sites
characteristic of serine proteases in general flom those contributing to ligand specificity.
Thirty-seven overlapping conserved water sites were found between thrombin and trypsin,
four and a half times the number expected for a random distribution of water sites. Finer-
Moore et al. (1992) evaluated similarity in solvent structures between pairs of eight trypsin
and trypsinogen structures and also found significant similarity between them. Ten of the
37 shared sites observed here were in contact with ligands or associated with the solvent
channel proximal to the Na” site (Table 2.4). This is consistent with the observation of
Krem and Di Cera (1998) that one-third of the conserved internal water sites in serine pro-
teases (Sreenivasan and Axelsen, 1992) are located near the Na+ site; they proposed that
the water structure stabilizes this pocket associated with the substrate specificity (Krem and
Di Cera, 1998). Two water sites conserved between thrombin and trypsin in a channel lead-

ing from non-catalytic triad Ser 214, which interacts with Asp 102 of the catalytic triad,

70

were also found. This solvent channel has been proposed as an exit path for the protons

produced during catalysis (Meyer, 1992).

2.5.2 Conserved Water Sites and Ligand Specificity

To identify water sites that can contribute to substrate specificity, water sites which were
conserved in functionally important regions of thrombin and trypsin but not conserved
between the two enzymes were analyzed. The 22 water sites conserved in the active site,
Na+ binding region, and exosite of thrombin but not in trypsin, and the eight active-
site water molecules conserved in trypsin but not in thrombin (Table 2.5) are likely to
contribute to their different substrate specificities. Design of thrombin inhibitors may be
optimized by mimicking these water interaction, as has been achieved for HIV protease
(Lam et al., 1994) and cyclophilin-A (Mikol et al., 1995). Results presented in Raymer
et al. (1997) on a study of 20 nonhomologous proteins bound to diverse ligands showed
that water molecules in ligand-binding sites can be displaced by similarly polar ligand
atoms, but also that water-mediated bridges between protein and ligand are ubiquitous,
with an average of 19 water-mediated hydrogen-bond interaction between proteins and
small ligands. Thus, the positions of conserved interfacial water molecules can used to
specify a template of favorable hydrogen bonds for ligands to satisfy, providing another

strategy for optimizing ligand design.

71

2.1

The

defi

llllI.
C01"

Me

811

an;

2.6 Conclusions

The work presented here demonstrates hierarchical clustering as a useful tool for unbiased
definition and analysis of consensus water sites when several independent structures of a
protein are available. This approach is particularly useful for resolving the continuum of
water site overlaps that occurs when a number of structures are superimposed. Analysis
of colocalization between thrombin and trypsin water sites showed a small, but significant
number of overlaps, predominantly surrounding the sodium ion site in thrombin and the
corresponding region in trypsin. Cluster analysis of water sites and their environments also

identified the features associated with highly conserved water sites:

1. a high density of protein atom neighbors, indicating the water site is in a protein

groove or cavity instead of being associated with a surface protrusion,

2. several hydrogen bonds being formed to the protein,

3. a hydrophilic environment, and

4. low thermal mobility of the site.

Since cluster analysis is a general statistical method, it is also expected to be useful for

analyzing side-chain and ligand atom positions and their chemistries.

72

Chapter 3

Computational Ligand Screening - An Improved

Model of Protein-Ligand Interactions

3.1 Introduction

In addition to examining water molecules as a set of special ligands, it is desirable to expand
analysis to small, organic molecules which act as ligands. One approach to this is the devel-
opment of computational screening techniques which can be used to screen large databases
of molecules efficiently using computers versus experimental approaches. One method of
computational screening is to approach the problem as an extension to computational dock-
ing, which seeks to find the position of a known protein ligand in the protein’s binding site.
Using such a method for screening would involve docking each of the molecules in the
database into a defined binding site and ranking them based on the quality of the docking.
While such an approach would work in theory, using the best docking algorithms available,
which include firll ligand flexibility (Welch et al., 1996; Rarey et al., l996b,a), would cause

73

the problem to be computationally intractable as these take at least on the order of a few
minutes to dock a single molecule. Such a screen of several hundred thousand molecules
would take an infeasible amount of time. Even allowing for just one minute per ligand,
screening a database of 100,000 molecules would take 10 weeks. However, one could
consider using a simplified docking algorithm which reduces the docking time to a second
or less per ligand and enables the screening to completed in a day. Such a technique is

presented here.

The classical view of protein-ligand binding, introduced by Fischer (Fischer, 1894), is
that of the “lock-and-key”, where the ligand fits as a key into the protein lock. However, a
more recent study of 39 complexes showed that the “lock” and the “key” are often flexible
(Betts and Stemberg, 1999), meaning that a simple steric fit docking is not sufficient to
function as a screening method. In an ideal case, both the protein and the ligand would
be fully flexible, but this retums to the problem of computational tractability. Instead, a
method where sufficient, but not additional, flexibility is included in the model would be
optimal. An additional step to further increase the efliciency of the docking is to represent
the binding site of the protein by a series of points which reflect the possible interactions
which can be made to a potential ligand. In the docking tool DOCK (Shoichet et al., 1992;
Shoichet and Kuntz, 1993), this is generally a set of around 100 spheres which constitute a
negative image of the binding site. When a docking search is performed, the set of points
representing the protein binding site is matched with the set of points representing the

possible interactions a potential ligand could make.

74

3.2 Methods

3.2.1 A General Overview of the SLIDE Method

The overall method of the SLIDE (Screening for Ligands with Induced-fit Docking Efli-

ciently) algorithm involves three stages, all of which will be discussed in detail:

1. assignment of a set of points which represent the types of interactions a molecule

could make when bound to a protein (performed once per database of molecules),

2. identification of the sites of favorable interaction in the protein and reduction to a set

of favorable template points (performed once per protein of interest), and

3. matching the database molecules, via their interaction points, to the protein, via its

template points, i.e., the actual screening process.

Both molecule interaction point and protein template points can be one of four types: (1) a
hydrogen bond acceptor, (2) a hydrogen bond donor, (3) a hydrogen bond doneptor (donor

and acceptor), or (4) a hydrophobic or non-polar point.

3.2.2 Assignment of Interaction Points to Molecules in the Screening

Database

Assignment of interaction points to database molecules is based upon an atom by atom

examination of the molecule in question. Hydrogen bonding points are placed at atom

75

positions which can make a hydrogen bond. A hydrogen bond donor point is assigned to
each nitrogen which is bonded to a hydrogen or which is positively charged. A hydrogen

bond acceptor point is assigned to each of the following atoms:

0 carboxylate oxygen atoms,

0 SP2 oxygen atoms,

0 .pra oxygen atoms which are not in hydroxyl groups and which are not bound to a

nitrogen atom,

o fluorine atoms in a C—F bond, and

o chlorine atoms in a C—Cl bond.

A hydrogen bond doneptor point is assigned to hydroxyl groups as the lone-pair electrons
on the oxygen can act as hydrogen bond acceptors while the oxygen can donate the hydro-

gen to another hydrogen bond acceptor.

Assignment of hydrophobic interaction points is done using a set of rules summarized
in Figure 3.1. These rules strive to place a hydrophobic interaction point every 1.5 to 2
carbon atoms along hydrophobic chains and around the edges of hydrophobic rings. The
method originally implemented in SLIDE assigned a hydrophobic interaction point to every
hydrophobic carbon, i.e., those bonded only to other carbons, hydrogens, or sulfurs, and to
the center of hydrophobic rings. This caused a significant overassigrrment in long aliphatic
carbon chains, such as those contained in fatty acid molecules, due to the assignment of

points at every carbon position. The previous method also resulted in underassignment

76

 

Methyl Group lsopropyl Group Tetrahedral Group

   

Hydrophobic Ring with Hydrophiiic Hydrophobic Ring with Single Hydrophobic Ring with Multiple
Substituent Hydrophobic Substituent Hydrophobic Substihrents

 

lntemal Hydrophobic Atom Triplet

 

Figure 3.1: Summary of rules used to assign hydrophobic interaction points to molecules
in the screening database. The overall goal is to assign a point every 1.5 to 2 carbon atoms.

 

77

for hydrophobic rings due to the assignment of a single point in the center of the ring.
The new method presented in this work sought to eliminate both the overassignment and

underassigrrment to better represent the hydrophobic character of the molecules.

3.2.3 Identification and Assignment of Protein Template Points

Once the chemistry of molecules in the screening database has been described by the as-
signment of interaction points, the chemistry of the protein binding site, the template, must
be described. This is done by the assignment of a set of template points in the binding
site. These points represent favorable interaction positions of potential ligands and rep-
resent the negative image, in terms of both shape and chemistry, of the binding site. AS
for the database molecules’ interaction points, the template points are assigned one of four
types: hydrogen bond acceptor, hydrogen bond donor, hydrogen bond doneptor (donor and
acceptor), or hydrophobic point. There are two methods of placing template points: based

on known ligands or in an unbiased approach.

Creation of a Protein Template Based on Known Ligand Binding Modes

If the structure of at least one protein-ligand complex is known, a template can be created
which is based on the binding orientation of this known ligand or ligands. This technique is
useful when it is desirable to identify potential ligands which represent the chemistry of the
known ligand(s) and can be usefirl to screen a subset of a larger molecular database and/or
when a particular set of protein-ligand interactions want to be exploited. The first step of

template creation is to assign interaction points to each of the docked ligands, as above in

78

Section 3.2.2, with each ligand in the reference flame of the protein binding site. These
points are then clustered using complete linkage clustering (see Section 2.3.2 for a detailed
explanation of complete-linkage clustering) to reduce the set to a representative sample of
the ligands’ chemistries. In the case where a single ligand is used as a basis for the creation
of the template, the resulting template is simply the set of interaction points of the single

ligand.

Creation of an Unbiased Template

The unbiased approach to template creation is the preferred method when searching a
molecular database for novel potential ligands, i.e., potential ligands which do not resemble
ligands in the available protein-ligand complex structures. It is also the only method avail-
able when the only available protein structures do not contain a ligand. Hydrogen bond
forming points are placed based on geometry of residues residing in the binding site in a

technique developed by my colleague Maria Zavodszky.

To identify potential hydrophobic interaction centers, a set of points are initially placed
at the vertices of a three-dimensional grid, generally with spacing of 0.5 A, in the binding
site, defined by a box surrounding it. This generally results 10,000-40,000 points, depend-
ing on the size of the binding site. An earlier method placed points randomly in this box,
but this resulted in uneven sampling which often caused areas of possible interaction to
be unrepresented in the resulting set of template points. The set of points is then reduced
to include only those within a shell between 3.0 and 5.2 A flom the protein surface. This

step generally reduces the set of potential template points to 2,000-10,000. Each of these

79

remaining points are then checked for hydrophobic character by calculating a hydrophobic

enhancement score as follows:

Enhancement Score = (3.1)
Number of Hydrophobic Atoms in Protein Environment —

Number of Hydrophilic Atoms in Protein Environment

The protein environment is defined as a sphere of radius 5.2 A centered on the potential
hydrophobic interaction point. This measure encompasses the idea of having a more hy-
drophobic environment when the point neighborhood contains more hydrophobic atoms.
This is in contrast to a measure which involves the average hydrophobic character of the
point’s protein environment, giving equal weight to an environment with a single hydropho-
bic atom and one with many hydrophobic atoms. By adjusting the enhancement score used
as a cutoff to assign a potential template point as being a hydrophobic template point, the
number of hydrophilic atoms allowed within this environment can be adjusted. By exam-
ination of the predominantly hydrophobic diethylsilbestrol (DES) ligand of the estrogen
receptor (PDB code 3erd; Tanenbaum et al. 1998), a cutoff of 3 was chosen. After deter-
mining which points reside in hydrophobic environments, they are clustered using com-
plete linkage clustering, generally with a clustering threshold of 3.0 A, which provides for
an approximate inter-cluster nearest-neighbor distance of 1.5 A. It is important to note at
this point that while the hydrophobic template points are initially placed at grid vertices,
the clustering results in them being assigned to arbitrary, non-grid positions. Hydrophobic

interaction points which overlap with hydrogen bond interaction points, defined as having

80

a center-to-center distance of less then 1.5 A, are eliminated in favor of the geometrically
placed hydrogen bond points. The remaining hydrophobic points and the hydrogen bond

points constitute the final template.

Earlier Method of Unbiased Template Design

As a major focus of this work is the change in the model of the protein binding site, a
brief note on the original method by which the the binding site is modeled is warranted.
The original method described the protein binding site as a set of template points which
reflect the favored sites of potential interactions with ligands, as the new method does. The
original method also assigned points as hydrogen bond acceptor, hydrogen bond donor,
hydrogen bond doneptor, or hydrophobic, as does the new method, but these were assigned
differently. Instead of being placed at geometrically preferred positions, hydrogen bond
points were selected flom the shell of all points as those which can form hydrogen bonds
to protein atoms. Each type of hydrogen bond point, i.e., acceptor, donor, or doneptor,
was then clustered and rechecked for the ability to still participate in a hydrogen bond.
The same set of points constituting the shell around the protein was then probed for points
which reside in a hydrophobic environment based on the average hydrophobicity of the
atoms in the potential point ’s protein environment. The hydrophobicity for a protein atom
in the template point’s environment is defined as the average number of instances when
a water was bound to the atom per 1000 occurrences of the atom in a study of 53 non-
homologous protein structures (Kuhn et al., 1995). As in the new method, points classified

as hydrophobic were then clustered and combined with the hydrogen bond points to form

81

the final template set.

3.2.4 Matching Molecular Interactions to the Template: The Screen-

ing Step

Once the protein binding site has been described as a set of template points and each of the
database molecules have been described as a set of potential interaction centers, identifica-
tion of compatible matches can be achieved. As stated in the introduction, the use of a fill]
scale docking approach for each ligand would prove computationally intractable. The ap-
proach described here implements several techniques to reduce the overall screening time
to enable the screening of databases on the order of a 100,000 molecules in approximately

one day.

Use of flashing Techniques to Rapidly Eliminate Infeasible Dockings

The first step in the screening process is to define a set of four hash tables to describe
the triangles present in the set of template triangles. These four hash tables include the

following parameters of the template point triangles:

1. the chemical type, i.e., hydrogen bond acceptor, hydrogen bond donor, hydrogen
bond doneptor, or hydrophobic, of the three template points which define the triangle

(20 hash entries),

2. the perimeter of the template point triangle, generally over a range of 3—25 A in bins

of 0.25 A (88 hash entries),

82

3. the length of longest side of the triangle, generally over a range of 1—10 A in bins of

0.125 A (72 hash entries), and

4. the length of the shortest Side of the triangle, generally over a range of 1—5 A in bins

of 0.125 A (32 hash entries).

The lengths described above are used in most screening cases, but can be altered for spe-
cialized runs. This tabulation, while somewhat computationally costly, is performed only a

single time for a screening run.

Identification and Docking of the Anchor Fragment

Each set of three interaction points in a molecule in the database describes an anchor flag-
ment for that ligand (Figure 3.2). An anchor flagment is the substructure of the molecule
which is rigid when allowing only torsion angle rotations, i.e., if any of the bonds in the
anchor flagment were to be rotated, the triangle defined by the interaction centers would
be distorted. The screening process examines all of the triangle mappings in each of the
database molecules, leading to an exhaustive approach. Each anchor flagment of a database
molecule is used as a potential basis for docking the molecule into the protein binding site.
The previously calculated hash tables are used to very quickly eliminate template point
triangles which cannot feasibly match the anchor flagment currently being explored, as
shown in Figure 3.3. In order to eliminate edge effects that may occur when the measure
of a particular geometric property for an anchor flagment triangle lies near the boundary of

bins, the template triangles in bins on either side of the matched on are also included.

83

 

Side Chain Anchor Fragment

    

 

 

 

Anchor Fragment

Figure 3.2: Two example anchor flagrnents for an example molecule. An anchor flagment
is defined for each set of interaction point triplets for each molecule in the database. Rota-
tion of any rotatable bond within the anchor flagment would cause a distortion in the anchor
flagment triangle, disturbing the initial triangle matching. Portions of the molecule outside
of the anchor fiagrnent are ligand side chains and can be rotated to alleviate collisions with
the protein.

 

Once a set of feasible template triangle matches to the anchor flagment is identified,
the remaining screening process, summarized in Figure 3.4, is performed. The overall
idea is to perform the least computationally expensive steps early on, discarding molecules
which fail to meet particular thresholds at each step. In this way, the most costly steps are
only performed on molecules most likely to dock. After the set of feasible matching tem-
plate triangles have been extracted flom the hash tables, each template triangle is examined
individually. The six possible triangle one-to-one triangle mappings are investigated. Ini-
tially, the chemical complementarity of the mappings is checked, e.g., to ensure acceptor
database molecule interaction points are mapped onto only template acceptor or doneptor
points. Database molecule interaction points are mapped onto the same type of template
points since the template represents the negative image of the protein binding site. For all

complementary triangle mappings, the distance matrix error (DME) for the side lengths of

84

 

SetofFeasible

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

D Template Triangle
5 A D 6.2 A A 3.6 A D Matches tothe
. - A . .
- . A UgandTnande
DLIQMU T080919 6.5 A 6.7 A 3.5A 6 3 ¢
.3 A A A 6.6
A 3.7 A A A
6.6 3.7 A
A Tables Indexing Protein Template Triplets
E
000 Triangle 10.0105 5.0052571
ADD Pam“ 10511.0 Triangle 5255.50 A
$31 i Longest :
e:— I
1.1—‘65 11 am
-,—" 17.0-17.5 6.6A 6.75-7.00
D = 1180110 Donor - . o
A I m M : 0 c
H = Hydrophobic Point . ' '
- 24. 11. 11.7 5,5375 ,
HHH “ 11.751200 4.75500
1st Index Table 2nd Index Table 3rd Index Table 4th Index Table

Figure 3.3: Hashing scheme implemented in SLIDE. The template hash tables are calcu-
lated a single time at the beginning of the screening run and are used to quickly eliminate
template triangles which cannot feasibly match to the ligand anchor flagrnent triangle un-
der current examination. This initial step reduces the number of template triangles which
have to undergo more computationally expensive triangle fitting steps. To eliminate efl‘ects
which may occur when the measured property is near the boundary between bins, the tem-
plate triangles in bins adjacent to the matched one are also included as potential further
matches.

 

85

 

For All Poeelble Anchor Fragments Defined by All Triplets of
Interaction Centers In Each of the Screened Molecules

Identity Chemically and

Flexible Side Chain Feaelble
Superposition of Ligand Triangle

A onto Template Triangle

  
 
 

——>
Identify Mawhing Template
Triangles by Mufti-Level Chemistry 1
m s“. Chain and Geometry Based Hashing

    
 

Figure 3.4: Screening algorithm implemented in SLIDE. SLIDE’s docking of potential
ligands into the binding site is based on mapping triplets of ligand interaction centers (H-
bond donors, acceptors doneptors, or hydrophobic) onto triangles of template points located
above the protein surface. Feasible template triangles for each possible triplet in a screened
molecule are directly accessed via a multi-level hash table, and the corresponding mapping
is used to dock the rigid anchor flagment of the potential ligand. Single bonds in the flexible
parts of both molecules are rotated to generate a shape-complementary interface, before
the complex is scored by the number of intermolecular hydrogen bonds and hydrophobic
complementarity of the contact surfaces. In all steps the ligand triplets or dockings that do
not meet a particular threshold are discarded.

 

86

each molecule side, m;, and the corresponding template site, t,- is computed as follows:

 

3

DME = \l; 2 (mi—1,)2 (3.2)

i=1

The DME’provides an approximation of the root-mean-square deviation, (RMSD) of the
superimposed triangles and is simpler to calculate. Therefore, it can be use as a first ap-
proximation to eliminate infeasible matchings and to find the best superposition between
the molecule and template triangles. Both the DME and RMSD must be below a defined
threshold for the anchor triangle to pass. In both cases, a looser fit is required for the hy-
drophobic template/molecule interaction point match to allow for the fact that hydrophobic

interactions are less specific.

Modeling of Induced Complementarity

Until this point in the algorithm, SLIDE has been working with the database molecule
anchor flagment and protein template in a reduced form as a simple triangle of interaction
points, but now the algorithm introduces a more realistic model by introducing the atoms
included in the database molecule’s anchor fragment and the atoms in the protein’s main
chain and C3 atoms. A check for intermolecular collisions between the anchor flagrnent
atoms and the protein main-chain atoms is performed. If atoms are found to overlap, the
anchor flagrnent is translated away flom the protein in the direction which alleviates the
collision(s) the minimum distance necessary to remove the overlap(s). This direction and
distances can be calculated as the sum of the vectors which lie along the collision axes.

This translation is limited, generally to 0.2 A, to maintain the original triangle matching

87

and is repeated, up to 100 times, effectively shaking the molecule in the binding site, but in

a directed method.

If a docking with no overlaps between the protein main-chain atoms and molecule an-
chor flagment atoms is found, induced flexibility is modeled by rotation of protein and
molecule side chains. In this context, a database molecule’s side chains are those flag-
ment connected to the anchor flagment by rotatable bonds (Figure 3.2). If intermolecular
overlaps are found between the database molecule and the protein, using a full atom rep-
resentation, they may be resolved by rotation of a bond that will move either the database
molecule atom or the protein side-chain atom involved in the collision. Often times, there
are multiple bonds which can be rotated to resolve the collision, each displacing a different

set of atoms a different amount.

The approach presented here for modeling induced complementarity and deciding the
best bonds to rotate to resolve a set of database molecule/protein collisions is based on
mean-field theory (Jackson et al., 1998; Koehl and Delarue, 1994, 1996). This method al-
lows the rotation of the best of any rotatable bond to resolve one or more of the collisions.
A key part of this method is the creation of a probability matrix, P(i, j), which describes
the probability that a collision i will be resolved by rotation of bond j. Initially, all in-
termolecular collisions are identified. These form one dimension of the matrix. If more
than 20 collisions are identified, the docking is discarded as it is unlikely that this many
collisions will be resolvable. All rotatable bonds which can be used to resolve at least one
collision and do not cause a new intramolecular collision in the current configuration form

the other matrix dimension. It is important to note that there is no differentiation between

88

database molecule and protein side chains. All rotations which can resolve a particular
collision, i.e., all entries P(1', j) such that rotation j resolves collision i, are assigned equal
initial probabilities. For each probability entry, P(i, j), a force, F(i, j), is computed that
reflects the cost of rotating bond j to resolve collision 1‘. The force assigned in this work is
simply the product of the absolute value of the angle of rotation of the bond and the number
of atoms which will be displaced by rotating the bond. Such a force penalizes rotating a
larger number of atoms a larger number of degrees, as this is more likely to cause additional

collisions elsewhere.

After initialization, several iterations of mean-field optimization are performed by up-
dating the probability matrix P to converge to high probabilities for those rotations which
provide the lowest cost conformational change for both the database molecule and the pro-
tein and which resolve the largest number of collisions. In each iteration, a mean force,

E(i, j), is computed for each rotation, as follows:

3033') = F011) + h a; kdePKI'IJ'). (11.10] -P(h, II) ' F(h, k) (33)

The value of dep[(i, j), (h, It)] is a measure of the dependency between probability entries
P(i, j) and P(h, Ir). It is set to — 1.0 if both entries refer to the same bond and both rotations
are in the same direction, i.e., j = 1:. If this is the case, two collisions can be resolved at
once by a rotation of this single bond. Assignment of a dependency of —1.0 results in a
lower mean force, E(i, j), thereby favoring this rotation. If probability entries P(1', j) and
P(h, k) refer to the same bond (j = 1:), but the rotations are in opposite directions, the

dependency is set to +1.0, penalizing this rotation. If bond 3' lies on the path to bond 1:,

89

e.g., bond j is between C )9 and C, and bond I: is between C, and C5 of the same side chain,
the dependency is also set to +1.0. This is penalized, since if rotation (1', j) were applied,
bond I: would be displaced, invalidating the assumptions about the current conformation of

the both the database molecule and the protein in the current optimization iteration.

At the end of each iteration, the entries of the probability matrix are updated based on

the mean force using the Boltzmann principle:

e_E(ir j)/"
2“. e_E(ir k)/I‘

 

Pour) = (3.4)
p is the average value of all computed mean forces. Convergence of the values in the
probability matrix is generally seen in fewer than ten iterations, and those rotations with the
highest probability are chosen to resolve the collisions. It is necessary to check for negative
correlations between bonds again at this time. Although checked for during the mean field
optimization, two correlated bonds can receive high probabilities if they are the only bonds
which will resolve a particular set of collisions or if alternative rotations are much more
expensive. Also, it is not possible to anticipate complex dependencies, e.g., which ligand
rotations influence protein bonds related to other collisions, during mean field optimization.
Since it is unlikely that all intermolecular collisions can be resolved by a single application
of optimization, up to 10 cycles are executed. Database molecule dockings are discarded
if they have more than 20 collisions at any time during the optimization or have remaining

collisions after 10 cycles of mean field optimization.

90

Scoring

Once a collision flee complex is identified, the final step in determining if the current
complex is a valid potential ligand is to calculate the chemical complementarity between
the database molecule and the protein. As a first pass, all complexes with poor shape
complementarity are eliminated. In the 89 complexes (Eldridge et al., 1997) used to tune
SLIDE’S scoring function, an average of 88% of the ligand carbon atoms where located
within 4.0 A of a protein atom, and all ligands buried at least 55% of their carbon atoms
against the protein surface (Figure 3.5). Based on this observation, all dockings in SLIDE
with fewer than 50% of carbon atoms buried, i.e., within 4.0 A of a protein atom, are

discarded.

The complexes are then assessed for chemical complementarity by a scoring function,
SCORE(P, M), which consists of a term for the number of intermolecular hydrogen bonds
formed between the protein, P, and the database molecule, M, HBOND(P, M), and a
term for hydrophobic complementarity between the protein and the database molecule,
HPHOB( P, M). For calculation of the number of intermolecular hydrogen bonds, hydro-
gen bonding is considered for cases where the distance between donor and acceptor is less
than 3.5 A. For proteins and database molecules with no hydrogen atoms provided in the
crystallographic structure, the positions of the hydrogens are computed based on known
bond angle and length constraints and optimal placement for hydrogen bonding when sev-
eral positions are possible (l-looft et al., 1996), such as for the hydrogen in a hydroxyl
group. For cases when the hydrogen atoms are given in the structure, their positions are

taken as given. Rotatable hydrogens are rotated to optimize hydrogen bonding when appli-

91

“l

 

 

 

35 f f . I I T fl I

30

l

N N
C 01
I I

Number of Complexes
at

1

l
I
I

 

 

 

 

o 1
50% 60% 70% 80% 90% 100%
Buried Ligand Carbon Atoms

Figure 3.5: Percentage of buried carbon ligand atoms in the 89 complexes used to tune
the SLIDE scoring function. Complexes were derived flom Eldridge et al. (1997). All
complexes had at least 55% of their carbons buried against the protein surface, defined
as being within 4.0 A of any protein atom. Based on this observation, SLIDE rejects any
ligand docking in which less than 50% of the carbon atoms are buried as an initial screen

before scoring.

 

92

cable. Donation to multiple acceptors is allowed if the angular constraints are fulfilled. The
following constraints are used to qualify a hydrogen bond: (1) a donor- - acceptor distance
of 3.5 A, (2) a donor—hydrogen distance of 1.0 A, and (3) a donor—hydrogen- - acceptor

angle of 120° to 180° (Haberrnann and Murphy, 1996).

Hydrophobic complementarity is calculated based on the residue and atoms type in the
protein and the atom type of the database molecule. The values are calculated as the average
number of hydrations per 1000 occurrences of the atom and are taken flom a statistical
study of 56 protein structures (Kuhn et al., 1995). (The values for protein atoms are flom
Table II and the values for ligand atoms are flom Table III). The hydrophilicity values
range florn 0, maximally hydrophobic, Co, to 635, maximally hydrophilic, the tyrosyl
hydroxyl oxygen atom. The hydrophobic complementarity between database molecule M

and protein P is calculated as follows:

avg{h’(Me), I3(0)}

 

HPHOB(P, M) = "REM max{abs(h’(M1)- 73(2)), 32} (3'5)
#R>0
where
h'(M,') = max{317 — h(M,'), 0} (3.6)

considers only the hydrophobic contribution of the database molecule atoms, M5, since
values larger than 317 refer hydrophilic atoms. If the atom is hydrophilic, i.e., the atom’s
hydrophilicity value is > 317, 317 - h(M,-) is < 0 and the hydrophobic character of the
database molecule atom becomes zero. The hydrophobicity, II(P,-), of the protein neigh-

borhood P,- for a single database molecule atom, M ,-, is the average hydrophobic character

93

‘1\

of all protein atoms, pj, within 4.0 A of the database molecule atom:

 

I.“ 2 ..,.),.}

PjEPI

The denominator in each element of the hydrophobic score, HPHOB(P, M) is set to be
greater than or equal to 32, which is 10% of the maximum score for a single database
molecule atom. This prevents a very few contacts with a large difference flom dominating
the hydrophobic score. The overall score for a complex is simply a linear combination of

the hydrophobic complementarity term and the number of intermolecular hydrogen bonds:
SCORE(P, M) = A - HPHOB(P, M) + B - HBONDS(P, M) (3.8)

The weights A and B have been chosen to optimize the fit between the scoring firnction
and the aflinities of 89 high—resolution complexes (Eldridge et a1., 1997). These com-
plexes had an average hydrophobic complementary term of 28.7 and made an average of
7.8 hydrogen bonds. Values of 0.59 and 2.76 for A and B, respectively, give a reasonable
approximation to the series of measured afl‘inity values (linear correlation coeflicient of
0.615), which yields a relative contribution of 1.321 .0 of the hydrogen bond term over the
hydrophobicity term. The overall goal of SLIDE’S scoring function is to provide a relative
rank for the potential ligands. At this point, a minimum score cutoff can be used to include
only favorable complexes, resulting in a set of 100-500 potential ligands. Optimization of
the binding mode and prediction of the binding aflinity can be done using a more detailed

conformational search and/or docking algorithm on the top-ranked potential ligands.

94

Another aspect of the scoring function is the modeling of salvation in SLIDE, though
not used in the work presented here. In other screening and docking algorithms, water
molecules are generally either considered to be a fixed part of the binding site or are ig-
nored. Some approaches seek to identify favorable water molecule positions in the protein
binding site prior to docking (Rarey et al., 1999) or seek to solvate the ligand (Shoichet
et al., 1999), but the water molecules in question are still fixed in position during the screen-
ing. It is well known that while some water molecules are key to providing high aflinity
binding (Ladbury, 1996), but it has also been shown that generally there are many water
molecules which are displaced from the binding site upon binding of the ligand (Raymer
et al., 1997). SLIDE uses the approach developed by Raymer et al. (1997) in Consolv, which
is a Ic-nearest-neighbor/genetic algorithm application to predict which water molecules are
conserved and which are displaced prior to the screening run. Consolv uses only informa-
tion about the protein binding site and uses no information about the bound ligand. The
information about Consolv’s predictions is used in SLIDE via inclusion of the prediction
confidence, the proportion of votes for conservation in the Ic-nearest-neighbor classifier, for
the water molecules predicted to be conserved. SLIDE can displace water molecules which
are predicted to be conserved, but a penalty proportional to the confidence of the conserved
prediction is assessed in the score. Water molecules which are predicted to be displaced

are removed fiom the binding site before screening.

One post-screening method that can be used is to apply a more costly, but more so-
phisticated, scoring function to fiirther refine the ranking of potential ligands. The empiri-

cally based scoring function DrugScore (Gohlke et al., 2000a,b) is such a scoring function.

95

DrugScore was created by examination of crystallographic structures for a series of protein-
ligand complexes to generate a database of radial distribution functions for each type of
possible ligand atom-protein atom pair. The favorability/unfavorability of a SLIDE gen-
erated protein-ligand complex can be computed by comparing the distance between each
ligand atom-protein atom pair in the docking to the functions derived from the empirical
study. Distances which match common distances are considered favorable, while distances

which match those only rarely seen are unfavorable.

3.2.5 Testing Databases

To examine the effects of changes to SLIDE, human a-thrombin and glutathione S-
transferase (GST) were selected. Thrombin and GST are good cases to test for several
reasons: there is a high resolution crystal structure of the ligand-flee protein available in
the Protein Data Bank (PDB), there are several high resolution protein-ligand complex
structures available, and there is a moderate diversity of ligands in the complex structure
of each protein. For thrombin, no truly ligand-free structure is available, but this work fo-
cuses on active-site ligands so structures with only non-active site ligands can be used as
a ligand-flee structure. The protein structures, shown in Tables 3.1 and 3.2, were chosen
to provide a set of unique ligands so as to prevent biasing the results towards one type of
ligand. In cases when more than one PDB entry contained the same ligand, the structure

with the best resolution was used.

Images in this dissertation are presented in color.

96

 

Table 3.1: 42 Thrombin protein-ligand complexes used for testing SLIDE modifications

 

 

 

PDB Code Ligand Resolution (A)
la2¢ Aeruginosin298-A 2.1
la3b Borologl 1.8
la3e Borolog2 1.9
la46 fl-strand mimetic inhibitor 2.1
la4w Ans-Arg-2ep-Kth l .8
135g Bic-Arg-Boa 2. 1
1a61 Mol-Arg-Lom 2.2
lad8 MDL103752 2.0
lae8 Eoc-D-Phe-Pro-azaLys-Onp 2.0
l afe Cbz-pro-azaLys-Onp 2.0
laht p-Amidino-phenyl-pyruvate l .6
lai8 PhCHgOCO-D-Dpa-Pro-boroMpg 1.9
laix PhCH20CO-D-Dpa-Pro-boroVal 2.1
lawf GR133487 2.2
lawh GR133686 3.0
lay6 Hmf-Pro-Arg-Hho l .8
leg Bcc-Arg-Thz 2.1
lba8 Pms-Ron-Gly-Arg 1 .8
lbe Pms-Ron-Gly-3 ga 2. l
lbcu Proflavin 2.0
lbhx SDZ 229-357 2.3
lbmm EMS-186282 2.6
lbmn EMS-189090 2.8
ldwb Benzamidine 3.2
ldwc MD-805 (Argatroban) 3.0
ldwd NAPAP 3.0
lfpc Ans-Arg-Epi (DAPA) 2.3
lhdt Alg-Phe-Alo-Phe-CH, (EMS-183507) 2.6
llhc Ac-D-Phe-Pro-boroArg-OH 2.0
llhd Ac-D-Phe-Pro-boroLys-OH 2.3
llhe Ac-D—Phe-Pro-boro-N-butyl-amidino-glycine-OH 2.2
llhg Ac-D-Phe-Pro-borohomoomithine-OH 2.2
lnrs Leu-Asp-Pro-Arg 2.4
lppb PPACK 1.9
1th Dpn-Pro-Prg-Bot 2.3

 

 

Continued on next page.

 

97

 

Table 3.1 (cont’d)

 

 

 

PDB Code Ligand Resolution (A)
ltmb Cyclotheonamide A 2.3
ltmt Phe-Pro-Arg 2.2
Item Methyl-Phe-Pro-amino-cyclohexyl glycine l .8
luma N,N-dimethylcarbamoyl-a-azalysine 2.0
3hat F ibrinopeptide A mimic 2.5
7kme SEL271 1. 2.1
8kme SEL2770. 2.1

 

 

 

3.3 Results

Presented below are the results of a series of tests on the changes made to SLIDE during
the course of this research work. Previously published reports have shown that SLIDE can
identify known ligands from a large database, can rank known ligands as better potential
ligands, and can correctly dock known ligands (Schnecke et al., 1998; Schnecke and Kuhn,

1999, 2000a,b).

3.3.1 Visual Examination of the New Template and Interaction Point

Methods

A first step was a visual examination of the template changes and the hydrophobic interac-
tion point assignment changes. Example unbiased templates for the estrogen receptor (ER)
in comparison to the diethylstilbestrol (DES) ligand using the original and the modified
methods are shown in Figure 3.6 (original method) and Figure 3.7 (new method). The

figures show a much better representation in the hydrophobic space occupied by the two

98

 

Table 3.2: 16 GST protein-ligand complexes used for testing SLIDE modifications

 

 

 

PDB Code Ligand Resolution (A)
10gs Benzylcysteine phenylglycine 2.2
12gs S-nonyl-cysteine 2.1
13 gs Sulfasalazine (SAS) 1.9
l 8 gs l-(S-glutathionyl)-2,4-dinitrobenzene l .9
l9gs Phenol-l ,2,3,4-tetrabromophtha1ein-3’,3”-

disulfonic acid ion 1.9
laqv p-Bromobenzylglutathione 1.9
laqw Glutathione 1.8
laqx S-(2,3,6-trinitrophenyl)cysteine 2.0
lgss S-hexylcystine 2.8
lpgt S-hexylglutathione 1 .8
203s Cibacron blue 2.5
2 l gs Chlorambucil l .9
2gss Ethacrynic acid (EAA) 1.9
2pgt (9R,10R)-9-(S-glutathionyl)-l0-hydroxy-9,10 1.9

dihydrophenanthrene
3 gss Ethacrynic acid-Glutathione conjugate 1.9
3pgt (+)-Anti-BPDE 2.1

 

 

 

99

 

 

Figure 3.6: Unbiased template for the estrogen receptor generated using the original
method. The template points are represented by spheres colored according to type: green,
hydrophobic; red, hydrogen bond acceptor; blue, hydrogen bond donor; and white, hydro-
gen bond doneptor (donor/acceptor). The solvent-accessible ER surface and the diethyl-
stilbestrol (DES) ligand are colored by atom (green, carbon; blue, nitrogen; red, oxygen;
yellow, sulfur). The DES ligand is shown for comparison only and is not used in the tem-
plate generation. Compared to new method in Figure 3.7, one can see the the very poor
representation of the hydrophobic area corresponding to the left benzyl ring of DES and
the moderately poor representation of the right DES benzyl ring. (planar in the figure).

 

 

 

Figure 3.7: Unbiased template for the estrogen receptor generated using the new method.
The template points are represented by spheres colored according to type: green, hydropho-
bic; red, hydrogen bond acceptor; blue, hydrogen bond donor; and white, hydrogen bond
doneptor (donor/acceptor). The solvent-accessible ER surface and the diethylstilbestrol
(DES) ligand are colored by atom (green, carbon; blue, nitrogen; red, oxygen; yellow,
sulfur). The DES ligand is shown for comparison only and is not used in the template gen-
eration. Comparison to the original method in Figure 3.6 shows a better representation of
the hydrophobic DES rings, especially in the area corresponding to the DES benzyl ring at
left.

 

101

benzyl rings of the DES ligand.

Additional visual analysis was done for the changed hydrophobic interaction point as-
signment method. A comparison between assignments using both the original and new
methods is shown in Figure 3.8. The new method shows a more balanced hydropho-
bic representation, especially between the hydrophobic rings and the aliphatic tails. The
overrepresentation seen in such tails and the underrepresentation seen in rings has been

eliminated.

3.3.2 SLIDE Docking of Known Ligands

While visual examination shows that the new methods for template point assignment and
hydrophobic interaction point assignment are likely to be an improvement, further work
needed to be done with screening and docking runs to confirm this result. A series of ex-
periments were performed with the original and new template methods and with the original
and new hydrophobic interaction point assignment methods. In one set, the template model
was held constant while the hydrOphobic interaction point method was changed, isolating
the effects of using the new interaction point method. In another set, the interaction point
model was held constant while the template method was changed from the original one to
the new one, isolating the effects of changing the template. In a third set, both the template
and interaction point models were changed to judge the combined effects. A graphical

summary of these experiments is shown in Figure 3.9.

Initial tests examined dockings of the known ligands with SLIDE compared to the crys-

tallographic structure dockings. Dockings of the known ligands from the complex struc-

102

 

 

 

00
00

Original Method New Method
B k
Original Method New Method

Figure 3.8: A comparison of hydrophobic interaction point assignment methods for
(A) estradiol from CSD code BEQJIQ (Parrish and Pinkerton, 1999) and (B) S-nonyl-
glutathione from PDB code 12gs (Oakley et al., 1997). Interaction points are represented
as spheres, colored by type: hydrogen bond acceptor, red; hydrogen bond donor, blue;
hydrogen bond doneptor, white; and hydrophobic, green. The new assignment method
provides a more balanced hydrophobic representation of the molecules, especially in the
hydrophobic tail of the S—nonyl-glutathione.

 

103

 

A

 

 

 

 

 

1’
B Original Tenplate Original Template
Original Interaction Points New Interaction Points
C
X‘
New Template New Template
Original Interaction Points New Interaction Points

 

 

 

 

Figure 3.9: Overview of experiments to test interaction modeling modifications made to
SLIDE. (A) represents experiments in which the template was maintained as a control and
the method for interaction point assignment was altered (results in Tables 3.3 and 3.4). (B)
represents experiments in which the interaction points were maintained as a control and the
template generation method was altered (results in Tables 3.5 and 3 .6). (C) represents tables
in which both the template generation method and interaction point assignment method
were changed (results in Tables 3.7 and 3.8).

 

tures were docked into the ligand-flee structure via superposition of the protein’s active-
site residues. Table 3.3 shows the number of known ligands which were docked for the
experiments in which the template was held constant but the hydrophobic interaction point
assignment method was changed from the original one to the new one. The table is in-
terpreted such that changing the interaction point method to assign interaction points to
the known thrombin ligands fi'om the original method to the new method while using the
template derived by the original method for the thrombin binding site enables SLIDE to
dock two known ligands not docked using the original interaction point method, but one
known ligand is no longer docked, yielding a net of +1 (without using an RMSD cutofl).
There were 34 known ligands that were docked using both the original and new interac-

tion point methods along with the original template point method. The key lines to note

104

 

Table 3.3: Known ligand dockings for constant template point method experiments

 

 

Template RMSD Dockings Dockings Dockings Net
Used Cutoff (A)"‘ Gained Maintained Lost Docked

 

Thrombin
Original None 2 34 1 +1
2.5 2 26 3 -1
l .0 6 4 5 +1
New None 1 33 0 +1
2.5 1 31 0 +1
1 .0 1 l9 2 -1
GST
Original None 0 15 O O
2.5 2 6 1 +1
1.0 1 1 1 0
New None 0 l6 0 0
2.5 0 12 0 0
l .0 0 12 0 0

 

 

‘The RMSD cutoff is the maximum RMSD relative to the crystallographic docking that a
SLIDE docking must have to be considered successfirl. None indicates that all dockings
were allowed. An RMSD of 2.5 A means the SLIDE docking must be moderately close
to the crystallographic docking, while an RMSD limit of 1.0 A means only very close

dockings are considered successful.

 

105

are the RMSD cutoff 2.5 A lines (shown in boldface). These show data for the dockings
which were reasonably close to the crystallographic complex, and that while a docking
is sometimes lost, generally additional dockings are successfully achieved using the new
interaction point method. One can now ask if there is a quantitative improvement in the
dockings of ligands that were docked using both interaction point methods. The data,
shown in Table 3.4, clearly indicate that these dockings are better. More dockings had
a lower RMSD using the modified interaction point assignment method and the average
change was generally near or below zero. Dockings of known ligands using the new inter-
action point method also generally had improved scores, using either the scoring function
implemented in SLIDE or DrugScore, compared to dockings using the original interaction

point method.

A set of experiments similar to the above was performed by maintaining the interaction
point method constant and altering the template creation method used. Table 3.5 shows
the number of known ligands docked in these experiments. This table is interpreted
the same as Table 3.3. In these experiments, there are a few cases where changing the
template causes a loss of a few ligands, but when examining only the ligands docked at
least reasonably well, RMSD cutoff of 2.5 A, an average of 3.3 additional ligands are
docked that were not docked using the original method. It is especially key in the GST
experiments where only a few (7 to 8) known ligands are reasonably well docked using
the original interaction point method, but 11 to 12 of the 16 (the number maintained +
the number gained) are reasonably docked using the new template method. The ability to

correctly dock known ligands is an important feature of the SLIDE algorithm and has an

106

 

Table 3.4: RMSDS for known ligands docked with both the original and modified interac-
tion point methods for constant template point method experiments

 

 

 

Template RMSD Dockings with Mean Standard
Used Cutoff (A)1 Better Equal Worse RMSD Deviation
RMSD RMSD RMSD Change (A)2
Thrombin
Original None 17 2 15 -0.02 0.05
2.5 14 0 12 +0.05 0.05
1.0 3 0 l -0.16 0.02
New None 13 10 10 +0.01 0.05
2.5 13 10 8 +0.002 0.05
1.0 9 5 5 -0.04 0.06
GST
Original None 5 0 10 -l .36 0.79
2.5 2 0 4 +0.02 0.75
1.0 0 0 1 0.14 —
New None 6 6 4 -2.01 0.06
2.5 3 6 3 +0.03 0.01
1.0 3 6 3 +0.03 0.01

 

 

For an overview of these experiments, refer to Figure 3.9. Please also note that it is not

possible to directly compare values within a column as they relate to varying numbers of

docked ligands. The changes reported in each line of this table are equivalent to traversing

lefi to right in Figure 3.9 (A).

1The RMSD cutoff is the maximum RMSD relative to the crystallographic docking that a
SLIDE docking must have to be considered successfirl. None indicates that all dockings
were allowed. An RMSD of 2.5 A means the SLIDE docking must be moderately close
to the crystallographic docking, while an RMSD limit of 1.0 A means only very close

dockings are considered successful.
1'RMSD change is scaled such that values < 0.0 are better while values > 0.0 are worse.

 

107

 

Table 3.5: Known ligand dockings for constant interaction point assignment method exper-
iments

 

 

Interaction RMSD Dockings Dockings Dockings Net
Point Method Cutoff (A) Gained Maintained Lost Docked

 

Used
Thrombin
Original None 1 32 3 -2
2.5 4 27 2 +2
1 .0 14 7 2 +12
New None 1 33 3 -2
2.5 6 26 2 +4
1.0 12 8 2 +10
GST
Original None 1 15 0 +1
2.5 4 7 0 +4
1 .0 10 2 0 +10
New None 1 15 1 0
2.5 4 8 0 +4
1 .0 10 2 0 +10

 

 

 

108

 

Table 3.6: RMSDS for known ligands docked with both the original and modified template
method for constant interaction point method experiments

 

 

 

Interaction RMSD Dockings with Mean Standard
Point Method Cutoff (A) Better Equal Worse RMSD Deviation
Used RMSD RMSD RMSD Change (A)*
Thrombin
Original None 27 0 5 -0.67 0.07
2.5 23 0 4 -0.38 0.02
1.0 5 0 2 -0.20 0.19
New None 25 0 8 -0.77 0.03
2.5 20 0 6 -0.37 0.12
1.0 6 0 1 -0.17 0.001
GST
Original None 13 0 2 -l.36 0.31
2.5 7 0 0 -0.84 0.03
1.0 2 0 0 -0.22 0.14
New None 15 0 0 -2.01 1.04
2.5 8 0 0 -0.80 0.11
1.0 2 0 0 -0.44 0.06

 

 

For an overview of these experiments, refer to Figure 3.9. Please also note that it is not
possible to directly compare values within a column as they relate to varying numbers of
docked ligands. The changes reported in each line of this table are equivalent to traversing
top to bottom in Figure 3.9 (B).

’RMSD change is scaled such that values < 0.0 are better while values > 0.0 are worse.

 

important role in ensuring that ligands are not missed during screening runs. Once again,
one can question if these dockings are better, and the answer is yes. Table 3.6 shows that
in all cases, there are significantly more ligands are docked with better RMSDS using the
new template method compared to the original method. Also, in all of the experiments
performed, the mean RMSD decreased, in some cases by a large amount, the most dramatic
change being for the GST using the new template method for all docked ligands. Given the

relatively small size of these ligands, a drop of mean RMSD of 2.01 A is very significant.

109

 

Table 3.7: Known ligand dockings for combined new template and interaction point meth-
ods compared to the original methods

 

 

RMSD Dockings Dockings Dockings Net
Cutoff (A) Gained Maintained Lost Docked

 

Thrombin
None 1 33 2 -1
2.5 4 28 1 +3
1.0 14 6 3 +1 1
GST
None 1 15 0 +1
2.5 5 7 0 +5
1.0 10 2 0 +8

 

 

 

Changes in scores also show a similar improvement in known ligand docking.

Now that it has been shown that the new interaction point assignment method and the
new template point method are improvements independently, the next experiment com-
pared dockings done with both new methods to dockings done with both original methods.
The combination of new methods also shows a significant increase in the ability to dock
known ligands, as seen in Table 3.7. Once again, there is generally an increase in the
number of known ligands docked, especially when focusing on only those ligands which
were reasonably well docked in comparison to the crystallographic docking. This indi-
cates that the combination of new template method and new interaction point method is
an improvement over using the original methods for both template generation and interac-
tion point assignment. Also, the combination of new methods generally does better than
the introduction of either the new template method or the new interaction point method
by itself. Table 3.8 shows the improvements in scores for each of the scoring functions,

SLIDE’s built-in scoring function, DrugScore, and RMSD, for the combined new methods

llO

versus the combined original methods for the thrombin and GST known ligands. From
this, it can be seen that the dockings using the combination of new template and interaction
point methods are quantitatively better in all measures with all successfirl docking RMSD
cutoffs. The improvement is more pronounced for the GST ligands, but is still significant
for the thrombin ligands. In fact, for GST, 13 of the 16 docked ligands using the combined
new methods have RMSDS relative to the crystallographic dockings lower than 1.0 A, while
only two of the dockings using the original methods have RMSDS lower than 1.0 A. The
scores and RMSDS are as good or better than seen using either single change, indicating
that using both the new method for template creation and for interaction point identification

is better than using either change on its own.

The question may arise as to the nature of the known ligands which cannot dock to
the protein. Most of the known thrombin ligand docking failures, 63% of the new method
docking attempts and 86% of the original method docking attempts, occur at the side-chain
collision resolution stage. This result is consistent with many dockings runs seen during
the course of SLIDE development, indicating that this stage is the most critical for dock-
ing, which is not surprising given the complexity needed to effectively model induced fit.
Other failures commonly occur at the ligand anchor fi'agment/protein main chain collision
resolution stage, which is likely related to the lack of modeling of protein main-chain con-

formational changes in the available docking and screening tools, including SLIDE.

In addition to quantitative measure of the quality of dockings of known ligands, one
can examine the dockings visually. Figures 3.10 and 3.11 show two such comparisons.

In both these cases, the docking determined using the new template and interaction point

111

 

Table 3.8: Scores for the known ligands docked using the new template and new interaction
point methods compared to using the original methods

 

 

 

RMSD Scoring Dockings with Mean Standard
Cutofl‘ (A) Function Better Equal Worse Score Deviation
Used Score Score Score Change1
Thrombin

None SLIDE 2 27 O 6 +6.8 2.8

None DrugScorea 27 0 6 -59300 6200

None RMSD 26 0 7 -0.63 0.01
2.5 SLIDE 22 0 6 +6.1 3.2
2.5 DrugScore 23 0 5 -51800 8200
2.5 RMSD 22 0 6 -0.37 0.04
1.0 SLIDE 5 0 1 +6.0 1.1
l .0 DrugScore 6 0 0 -6 1400 12400
1.0 RMSD 4 0 2 -0.233 0.18

GST

None SLIDE 4 15 O 0 +135 2.45

None DrugScore5 l4 0 1 -127700 5100

None RMSD l4 0 1 -l .59 0.25
2.5 SLIDE 7 0 0 +9.4 1.53
2.5 DrugScore 6 0 1 -92200 30000
2.5 RMSD 6 0 1 -0.79 0.03
1.0 SLIDE 2 0 0 +13.l 0.17
1.0 DrugScore l 0 l -810 47000
1.0 RMSD l 0 l -0. 15 0.21

 

 

 

For an overview of these experiments, refer to Figure 3.9. Please also note tlraTIt is not

possible to directly compare values within a column as they relate to varying numbers of

docked ligands. The changes reported in each line of this table are equivalent to traversing

diagonally upper-left to lower-right in Figure 3.9 (C).

1SLIDE scores are scaled such that values > 0.0 are better while values < 0.0 are worse.
DrugScore and RMSD changes are scaled such that values < 0.0 are better while values

> 0.0 are worse. _ . . . _
2SLIDE score rs umtless. For comparison, the mean score for known thrombin lrgands In

the crystallographic dockings is 36.3.

aDrugScore score is unitless. The mean score for known thrombin ligands in the crystallo-
graphic dockings is -468000.

4The mean known GST ligand SLIDE score for the crystallographic dockings is 40.2.

5The mean known GST ligand DrugScore score for the crystallographic dockings is
-378000.

 

112

 

 

Figure 3.10: Dockings of estradiol to the estrogen receptor. Shown are the dockings of
the estradiol ligand computed using the original template and interaction point methods,
shown in grey, and computed using the new template and interaction point methods, shown
in blue, compared to the position on the crystal structure, colored by atom (taken from PDB
1a52; Tanenbaum et a1. 1998). The solvent-accessible surface of the estrogen receptor (ER)
binding site is shown colored by atom. The new method docking is significantly improved
compared to the original method docking, which is rotated roughly 90° about the long axis
of the ligand and can no longer form a hydrogen bond with the protein at the left of the
binding site. The new method docking is in a similar position to the known ligand and can
make both hydrogen bonds (left and right of the binding site) as seen in the crystal structure
complex.

 

113

 

 

Figure 3.11: Dockings of ligand BMS-182282 to thrombin. Shown are the dockings of a
thrombin ligand to the ligand-free thrombin structure using the new template and interac-
tion point methods, shown in blue, and using the original template and interaction point
methods, shown in grey, compared to the crystallographically docked ligand, shown col-
ored by atom. The solvent-accessible surface of the thrombin active site is shown colored
by atom. The docking computed using the new methods is similar to the crystallographic
docking while the docking computed using the old methods is much less similar, especially
in the middle region of the ligand where the new method docking and the crystallographic
docking track quite closely, but the original method docking follows a significantly differ-
ent path.

 

114

A _:t

methods is much more similar to the known ligand, from the crystallographic complex
for the estrogen receptor or docked into the ligand-flee crystal structure via active site
superposition for the thrombin, than the docking determined using the original template

and interaction point methods.

3.3.3 Improved Enrichment

Now that it has been shown that the new methods to identify points of favorable interaction
in the template and in the database molecules improves the docking of known ligands,
it remains to show that the new methods result in improved screening results, i.e., better
selection of potential ligands from a database of molecules. The way to measure this is to
examine if known ligands are generally docked with higher scores than non-ligands, and,
therefore, would reside at the top of the screening hit list. To explore this, the first 14691
molecules of the 87326 molecule Cambridge Structural Database (CSD) screening subset
were selected. The vast majority of these molecules are unlikely to be true ligands, so
comparing the rankings of any hits resulting fi'om screening against these molecules to the
rankings of known ligands can yield a measure of how likely one is to select true ligands as
the top ranking screening hits. Shown in Figure 3.12 is a plot of enrichment for selection of
known ligands over random molecules for thrombin. If no enrichment was seen, the curves
would have a slope of l, which is clearly not the case; therefore significant enrichment
was seen for both the original and new template point and interaction point identification
methods. The plots do not reach 100% on the vertical axis as they are scaled to set 100% to

the full dataset of known thrombin ligands, 42 molecules, but only 35 (original method) and

115

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

A B
100% / 100% /
/ /
90% r / 90% + , /
» _-__/_/____ » /__._.

30% ,,,,, ‘ , 80% ~ ....... v ‘ '7
‘8 . ————————— / f , , ’ /
Ԥ 70% --- Original Method 70% ~ ---- Original Method
3 —— New Method ~ — New Method
2 60% x 60% » x
a / /
5" ’ / ” / /
g 50% / 50% f," / /
5 40% 40% a" / /
.— . /
37 30% 30% l /
0 ‘ /
a. ,' /

20% - 20% ,- / /

l
10% 10% .
’ /
00/° . l 1 a 1 A 1 a m/o a 1 4 1 1 L l 4
0% 20% 40% 60% 80% 100% 0% 20% 40% 60% 30% 100%
Percent Molecules Docked

Figure 3.12: Enrichment of known thrombin ligands in a set of random CSD molecules
using (A) the SLIDE scoring function and (B) the DrugScore scoring function. The plots are
scaled such that if the known ligands were randomly mixed in with CSD molecules in terms
of score, i.e., no enrichment was achieved, the plots would have a slope of l (demarcated
by the thin dashed line). This is clearly not the case for either the original method or the
new method for template point and interaction point identification, indicating significant
enrichment was achieved. The new method curve is shifted to the lefi, indicating the new
method yields an increase in the enrichment over the original method, especially when the
dockings of scored using the DrugScore scoring function (B). The y-axes are scaled such
that docking of all 42 thrombin ligands in the database would yield 100%.

 

116

34 (new method) ligands are docked. It can also be seen that the enrichment was greater for
the new method compared to the original method as the enrichment curve is shifted to the
left, i.e., a higher percentage of known thrombin ligands rank above the same number of
random molecules for the new method, indicating that the new method is an improvement
when measured using the enrichment. It can also be said that the scoring functions are
able to distinguish reasonable dockings from questionable ones based on these plots. A

similar analysis done for glutathione S-transferase showed a similar increase in enrichment.

(Figure 3.13).

A quantitative measure of enrichment, or enrichment factor, can also be calculated. One

such measure is derived from that developed by Knegtel and Wagener (1999):

F : Nact(p)/p

3.9
Nact/N ( )

where N a ) is the number of active compounds/known ligands in the top p ranked lig-

ct(p
ands and N act is the number of active compounds/known ligands in the complete docking
set, i.e., all molecules that were chosen as potential ligands, of size N. This factor is a
measure of the proportion of known ligands in the top p selected molecules of the database
relative to the proportion of known ligands in the set of all docked molecules. In general, p
is selected as a fixed proportion of the database, e.g., the top 1% of the docked ligands. A

second metric was developed for this work which calculates the proportion of total ligands

docked with higher scores than the top I: percent of known ligands:

12%

Eh = 27% (3.10)

117

 

Percent Known Ligands Docked

A

 

100%
90°/o
80% A
70% .
60% r
50%
40%
30%
20%

1 0%

 

L

 

 

 

— - - Original Method
— New Method

 

 

 

J 4. 4L

90%

70%

40%

30%

20%

10%

 

 

0%
0%

20°/o

40°/o

60%

 

1 00%
80%

60%~

50% .

Y ‘1 v V-v.-L -7-

 

    

 

 

- - - - Original Method
’ — New Method

 

 

 

 

4L 4‘ A_

 

1 ‘ 0%»
80% 100% 0%

20%

Percent Molecules Docked

40% 60% 80% 100%

Figure 3.13: Enrichment of known glutathione S-transferase ligands in a set of random
CSD molecules using (A) the SLIDE scoring function and (B) the DrugScore scoring furrc-
tion. The plots are scaled such that if the known ligands were randomly mixed in with CSD
molecules in terms of score, i.e., no enrichment was achieved, the plots would have a slope
of l (demarcated by the thin, dashed line). As can be seen, this is clearly not the case for
either the original or the new methods for template point and interaction point identifica-
tion, indicating significant enrichment was achieved. The new method curve (solid line) is
shifted to the left, indicating the new method yields an increase in the enrichment over the
original method (dotted line).

 

118

 

Table 3.9: Enrichment factors for thrombin and GST test screening runs

 

 

Protein Scoring
Function

Original
Method Method Increase

New

Fold

 

Knegtel and Wagener Measure

Thrombin SLIDE score
DrugScore

GST SLIDE score
DrugScore

New Measure

Thrombin SLIDE score
DrugScore

GST SLIDE score
DrugScore

28.6
25.7
46.7
20.0

12.1
7.7
6.5
3.0

44.1
76.5
62.9
71.9

26.0
74.4
93.6
18.3

1.5
3.0
1.3
3.6

2.1

9.6
14.3
6.0

 

 

 

where E]: is the enrichment factor for docking of 1: percent of the known ligands, generally

70%, and T% is the percent of total ligand dockings which were ranked higher than the top

I: percent of docked known ligands. Enrichment factors calculated using this measure are

shown in Table 3.9 Both measures of enrichment show quantitatively that the new method

yields improved enrichment, sometimes dramatically. It is not informative to compare the

enrichment factors calculated here to those determined by Knegtel and Wagener using their

method as their database was significantly smaller, only 1000 molecules compared to the

roughly 15,000 used here, resulting in their enrichment factors being significantly lower,

generally five to ten.

3.3.4 Results Summary

The method used to model the interactions a ligand molecule can make with a protein,

i.e., modeling of the ligand’s interaction points, has been modified as described in Sec-

119

tion 3.2.2. This was shown to be an improvement using both visual analysis of example
database molecules (Figure 3.8) and quantitative comparison between dockings achieved
with SLIDE and dockings achieved via superposition of the binding sites of the ligand-free
and complex crystallographic structures. There was generally an increase in the number
of known ligands docked (Table 3.3) and the quality of those dockings (Table 3.4). Addi-
tional changes were made to the model of the interactions made by the protein with a 1i g-
and molecule, i.e., the protein template, as described in Section 3.2.3. These changes were
shown to be improvements via both visual analysis of generated templates (Figures 3.6
and 3.7) and by quantitative measures comparing the dockings achieved by SLIDE to those
achieved by superposition of the binding sites. As with the new interaction point method,
the RMSD between the SLIDE dockings and the crystallographic dockings generally im-
proved using the new method compared to the original method (Tables 3.5 and 3.6). Com-
bination of the new methods also showed an improvement, generally greater than either
new method individually, as seen by comparing known ligand dockings (Tables 3.7 and
3.8) and by visual analysis of the dockings (Figures 3.10 and 3.11). Combination of the
new methods also improved enrichment of the screening results (Figures 3.12 and 3.13;

Table 3.9).

3.4 Discussion

SLIDE is an efficient tool for virtual ligand screening which includes ligand and protein
side chain flexibility. Databases on the order of 100,000 molecules can be screened in a few

hours to a day, depending on template size and screening parameters. Full ligand docking

120

is computationally too expensive to be performed on such a large database of molecules,
so initial matchings are calculated based on point representations of the protein binding
site and of the database molecules’ potential interaction centers. This approach enables
the elimination of infeasible matchings very quickly, thereby reducing the need to perform
the expensive computational operations on these infeasible matchings. The work presented
here describes changes made to the model used to identify both the sites of interaction in

the protein binding site and on the database molecules.

The method used to identify sites of potential interaction in the protein binding site
identifies sites where ligands could make favorable interactions to the protein and mirror
the binding site in both shape and chemistry. Each site is represented by a point with an
associated chemistry type, hydrogen bond acceptor, hydrogen bond donor, hydrogen bond
doneptor, or hydrophobic, and reflects the favorable ligand atom type to place at that site.
Altering the definition of a hydrophobic interaction site from one that is in an environment
which is hydrophobic on average to one which is in an environment which contains several
hydrophobic atoms proved to produce an improved model of the protein template site as ev-
idenced by the improved docking of known thrombin and glutathione S-transferase ligands.
While it may seem more logical to base an assignment of hydrophobic character on the av-
erage environment of the point in question, there is no sense of the size of the hydrophobic
area in question since having an environment consisting of only a single carbon atom is
treated as equally hydrophobic as an environment consisting of many carbon atoms. How-
ever, it is clear these environments are not equal in terms of their hydrophobic character.

Instead, it may be more important to have several hydrophobic atoms in the environment,

121

allowing the presence some hydrophilic atoms while necessitating an overabundance of
hydrophobic atoms. This is partially reflected in the fact that bound water molecules are
more often located in depressions in the protein surface than adjacent to protrusions, even
when neglecting the hydrophobic character of the water molecules’ environments (Kuhn
et a1., l992b). One could reasonably expect this observation to be stronger for hydrophobic
ligand atoms as there will be a stronger push for isolation from the surrounding solvent.
Use of the concept of the number of hydrophobic atoms in the environment could easily
be included in docking and screening scoring functions to further differentiate the more
favorable potential ligands and/or docking orientations selected. Often, scoring potential
ligands can be a key factor in extracting the best of the potential ligands selected, so an im-
provement such as this in the scoring function could prove to be an important enhancement

of the overall screening process.

A change in the model used for sites of potential hydrophobic interaction in the database
molecules was also implemented. This change yields a much better balanced representation
of the molecule’s hydrophobic character than was seen previously. One key feature of
the new model is the assignment of hydrophobic points to the edges of ring structures in
the molecules. Previously, a single point was placed at the center of each hydrophobic
ring. This leads to difficulties with determining the proper orientation in which to dock
the ring, i.e., stacked against the protein surface or edge-on relative to the protein surface,
and with limited placement of the ring. Since only a single point was assigned to the
ring, there was a severe limitation on the way that ring could match a site of favorable

hydrophobic interaction in the protein binding site, but it has been shown that while there

122

are some preferences observed for ring-ring packing interactions, there is also a fair amount
of variability (Burley and Petsko, 1985; Mitchell et al., 1997). Given that a 6-membered
aromatic ring is approximately 3 A wide and hydrophobic template points are generally
placed at an approximate nearest-neighbor distance of 1.5 A, there would be only one to
two matching sites for a ligand ring structure. By placing points around the edge of the
ring, more possibilities are allowed for ring placement as there are more ring points to
match to template points. The initial docking of the ring can also shift substantially more
by mapping points from different sides of the ring onto the same template point, firrther
increasing the docking space explored. The exploration of this additional docking space is
important when searching for potential ligands so as to decrease the probability that a true

ligand is missed due to a misalignment of a key feature.

In addition, to change the model for hydrophobic ligand rings, a new model of hy-
drophobic ligand atoms outside of rings was introduced. This model averages the hy-
drophobic character of the molecule throughout carbon chains compared to the previous
model, which assigned a hydrophobic interaction point to every hydrophobic atom bonded
only to other hydrophobic atoms. While one can argue that the reduction in the number of
points in this case reduces the docking space sampled, which would be unfavorable, reduc-
ing the number of points without significantly reducing the sampled space can be achieved,
as happens here. The new model of points, while reduced, still allows for a significant
amount of sampling as the points are placed approximately every 1.5 carbon atoms along
a chain. This is an approximate spacing to match the approximate 1.5 A template point

spacing, and adequate sampling of the space is still achieved.

123

The idea of docking space sampling is always a concern when performing computa-
tional screening. As the sampling becomes finer, such as with a finer template, the screen-
ing time can increase dramatically. In fact, docking tools, such as DOCK (Shoichet et al.,
1992; Shoichet and Kuntz, 1993), AutoDock (Morris et al., 1996), and GOLD (Jones et al.,
1995), which seek to find the optimal binding orientation of a single ligand, can be thought
of as finely sampling the docking space of a particular protein-ligand complex. The key
to effective screening methods will be to reduce the time needed to explore an adequate
amount of this docking space to provide accurate enough docking orientations to be able to
effectively analyze the resulting selected ligands. The ultimate goal of all virtual screening
techniques is to provide a set of potential ligands which show binding to the protein of
interest and can be used for further ligand optimization or as a set of probes for functional

studies, e. g., via differential inhibition. The work presented here comes closer to this goal.

124

Chapter 4

Computational Screening of Asparaginyl tRNA

Synthetase

4.1 Introduction

In addition to development of improvements to the computational screening algorithm
SLIDE, work was undertaken to identify potential novel inhibitors of asparaginyl-tRNA
synthetase (AsnRS) using SLIDE. Aminoacyl-tRNA synthetases are responsible for cat-
alyzing the addition of an amino acid onto the 3’ ribose of its cognate tRNA via a two-step
ATP dependent reaction. Initially, the amino acid is activated by the addition of ATP to form
an enzyme-bound arninoacyl adenylate intermediate. The arninoacylation reaction and its
specificity are vital to protein synthesis. The 20 aminoacyl-tRNA synthetases can be di-
vided into two general classes, I and 11, based on sequence motifs and putative structural
domains, as reviewed by Cusack (1995) and Amez and Moras (1997). In general, class I
synthetases contain a Rossmann Fold in their active sites and are active as monomers,

125

whereas the class II synthetases contain an anti-parallel fl-fold and are active as (12 homod-
imers or azflg heterotetramers. Class II aminoacyl-tRNA synthetases (AARS) are defined
by the presence of three class specific motifs of 10 (motif 1), 18-26 (motif 2), and 16 (motif
3) residues in their catalytic domains. It is possible to further divide class 11 into subclasses
Ila, 11b, and 11c based on the presence of specific domains, which generally play a role in
anticodon recognition. In addition to these subclass specific domains, a variable insertion
of 60 to 280 residues occurs between motifs 2 and 3 in the catalytic domains of class II
AARSs. AsnRS is a member of subclass IIb, which also includes aspartyl- and lysyl-tRNA

synthetases.

Lymphatic filariasis caused by Brugia malayi infection affects an estimated 100 mil-
lion people worldwide, and more than 1 billion people live in areas where the disease is
actively transmitted (Awadzi, 1997). There are currently no effective preventive medicines
against filariasis, as none are effective against the larvae, which are transmitted to humans
by mosquitos. Brugia AsnRS is an excellent target for filarial drug development as it is
highly expressed in the worms, has been well characterized biochemically and structurally
in several species, and can be recombinantly expressed to facilitate in vitro studies. Also,
the sequence and structure of the Brugia enzyme is difl‘erent fi'om the human AsnRS, pro-

viding for the possibility of identification of inhibitors specific for the Brugia synthetase.

Pieces of the protein translation apparatus have long been targets for antibacterial
agents, reviewed by Schimmel et al. (1998), including streptomycin and tetracycline, which
target the 30 S ribosomal subunit, and erythromycin and chloramphenicol, which target the

50 S ribosomal subunit. Aminoacyl-tRNA synthetases are currently a promising target for

126

anti-infective agents as an increased number of pathogens become resistant to traditional
antibiotics. Targeting new anti-infectives to aminoacyl-tRNA synthetases is promising as
the enzymes are specific for the transacylation of tRNAs and are somewhat species spe-
cific, reducing the possibility of cross-reactivity with host enzymes and the resulting side-
effects. Pseudomonic acid (mupirocin) is a natural product synthesized by Pseudomonic
fluorescens (Fuller et a1., 1971) that has been shown to be an inhibitor of isoleucyl-tRNA
(IleRS) synthetase from Gram-positive infectious bacteria, including antibacterial-resistant
S. aureus (Casewell and Hill, 1985). It has been shown to have an approximately 8000-fold
selectivity for pathogen IleRS over mammalian IleRS (Hughes and Mellows, 1980). Sev-
eral other natural products have been shown to inhibit aminoacyl-tRNA synthetases (N ass
et al., 1969; Paetz and Nass, 1973; Tanaka et al., 1969; Ogilvie et al., 1975; Werner et al.,
1976; Konrad and Roschenthaler, 1977; Konishi et al., 1989), indicating the potential for

use of natural and synthetic products to be effective against aminoacyl-tRNA synthetases.

4.2 Methods

4.2.1 Available Asparaginyl-tRNA Synthetase Structures and Ligands
for Virtual Screening Studies

A 1.9 A structure of Brugia malayi asparaginyl-tRNA synthetase complexed with an S-

adenosyl-asparagine (S-AMP-Asn) substrate analog was provided by Dr. Stephen Cusack

(EMBL, Grenoble, France), shown in Figure 4.1. The active-site pocket of AsnRS, shown

in Figure 4.2, is relatively deep and consists of two lobes, one of which binds the adenosyl

127

 

 

Figure 4.1: Structure of Brugia malayi asparaginyl-tRNA synthetase complexed with S-
adenosyl-asparagine. Shown is the backbone structure of AsnRS, with one chain of the
dimer colored green and the other chain colored blue. The S-adenosyl-asparagine ligand is
rendered as ball-and-stick and colored by atom type and the associated Mg ion is rendered
as a large white sphere. While there is a substantial amount of contact between the chains
of the dimer in the interface, the active site is isolated within each chain and, therefore,
only a single chain need be considered during the computational screening procedure.

 

128

moiety (right side of Figure 4.2) with the asparaginyl group extending into the other lobe
(left side of Figure 4.2). This crystal structure with the ligands removed was used to
construct two templates for SLIDE screening. The ligand positions fi'om this structure,
along with the positions of the asparagine ligand from E. coli AsnRS (PDB code llas;
Nakatsu et al. 1998) and the AMP and asparagine ligands fiom a second E. coli AsnRS
crystal structure (PDB code 12as; Nakatsu et al. 1998) were used to generate a ligand-
based template. The ligands from the E. coli structures were transformed into the same
reference fi‘arne as the Brugia structure via superposition of their active site atoms onto
the Brugia active site. A second, unbiased SLIDE template (i.e., a template based only
on the chemistry of the binding site, incorporating no information about the structures or
positions of known ligands bound in the site) was generated using the Brugia structure.
This template was modified to eliminate points outside of the pocket and to label points in
the deep lobes as key points. By labeling points as key points, SLIDE requires any ligand
docking to match at least one of these points, ensuring that the ligands are well situated in
these deep portions of the pocket. These ligand-based and unbiased templates were used to

screen a set of three databases:

1. a database of six known ligands consisting of three asparagine molecules (two taken
from PDB files llas and 12as (E. coli AsnRS structures) and a third generated
from the asparaginyl portion of the Brugia S-adenosyl-asparagine ligand), two AMP
molecules (one fi'om 12as and one generated from the AMP portion of the Brugia

complex ligand), and the S-adenosyl-asparagine ligand fi'om the Brugia structure;

2. a database of the 481 conformers generated for a set 16 high-throughput screening

129

 

 

Figure 4.2: Asparaginyl-tRNA synthetase active-site pocket. The pocket’s solvent-
accessible surface is colored by the type of atom which contributes to it: carbon surface,
green; oxygen, red; and nitrogen, blue. The S-adenosyl-asparaginyl ligand is also colored
by atom type, with sulfur in magenta. The two lobes of the pocket which hold the adeno-
syl and asparaginyl moieties are oriented in the upper-left and lower right, respectively.
Though difficult to visualize in this two-dimensional figure, these lobes extend deeper than
the central portion of the pocket, which forms a ridge between the two deeper parts of the
binding site.

 

130

ligands (described below); and

3. an 85,000 molecule subset of the CSD of crystal structures of small organic
molecules, consisting of those molecules with three-dimensional structures, nonpoly-

meric nature, and the absence heavy metal atoms, e. g., Fe, Mg, Zn, U.

In order for ligand candidates identified by screening to be guaranteed to have the ability
to span the entire binding site, encompassing both lobes of the AsnRS pocket (Figure 4.2),
SLIDE was set to require that at least two of the three ligand interaction points matching

the template be a minimum of 9.0 A apart.

4.2.2 High-throughput Screening for Asparaginyl-tRNA Inhibitors

and Conformer Generation of Selected Ligands

High throughput screening (HTS) for asparaginyl-tRNA synthetase inhibitors was
performed by Discovery Technologies (Allschwil, Switzerland) for our collaborator,
Dr. Michael Kron (Department of Medicine, College of Human Medicine, Michigan State
University). Screening was performed using a library of 11,700 compounds selected from
Bionet, MayBridge, SPECS, Aldrich, Analyticon, and several university compound li-
braries. Activity assays were performed using unfractionated yeast RNA as a tRNA sub-
strate along with l“C-asparagine under the protocol developed by Dr. Michael Hartlein
(EMBL, Grenoble, France). For each potential inhibitor, (“HTS hit”) the IC50 and relative
specificity for the Brugia versus human enzyme were determined. Since the HTS ligands

were provided as two-dimensional structures, three-dimensional structures must be gener-

l3l

ated to perform structural analysis of their mode of interaction with AsnRS. A search of
the Cambridge Structural Database (CSD) based on the structure of the HTS identified in-
hibitors yielded no entries, indicating that the crystal structures are not available for any of
these compounds. Initial three-dimensional structures were generated fi‘om SMILES codes
(a one-dimensional representation of the molecule’s chemical structure) and energy min-
imized using the Molecular Operating Environment (MOE; Chemical Computing Group,
Inc, Montreal, Quebec). Each of these minimized structures was used as input for the
stochastic conformer generation firnction in MOE to generate a set of three-dimensional
conformers for each of the HTS ligands. This conformer generation was run using a fixed

set of parameters that yielded a fine sampling of conformational space.

Images in this section of this dissertation are presented in color.

4.3 Results

4.3.1 High-Throughput Screening

High-throughput screening for asparaginyl-tRNA synthetase inhibitors by Discovery Tech-
nologies resulted in a set of 16 potential inhibitors, shown in Figure 4.3. The compounds
exhibited a range of structures, though they are generally small, with molecular weights be-
tween 178 and 542 Da, and contain some aromatic structure. Inhibitory activity, measured
as the concentration of ligand which reduces enzyme activity to 50% of native activity
(IC50), ranged fiom 6.7 to 171 pM. Specificities for the Brugia enzyme relative to the hu-

man enzyme, measured as the ratio of IC50 for human over the Brugia 1C5 0, ranging fiom

132

 

 

 

 

 

 

 

 

 

 

HTS 01 HTS 02 HTS 03 HTSO4
o l )N _ o
- N
o-n o H 0&0 0 l
o
o
O Nae W .8. O
o
. \o oo
. ~ I: s'
O 0 c.
N.
o ~o
HTSOS HTSOG HTSO7 HT508
SJ 3 O
A s "O
no
>=~ .
s o
H
HNH
HTSOQ HTSIO HTS" HTSTZ
H H
N N O‘N‘O
sc __ —
arts ~ .0. .21.
N /\©/ 5:
: ,N~ —o o——
o' o
HTSIS HTSl4 HT515
H s O N O F F
.N HN F
o
\~
'/ ° 0
N
\
o

 

 

 

 

Figure 4.3: Potential asparaginyl-tRN A synthetase inhibitors identified by high-throughput
screening. These HTS ligands contain at least some aromatic character while also con-
taining hydrogen bond forming groups. Of particular interest are ligands numbered 2, 7,
l3, and 15 which are suggested by Basilea scientists working on AARS inhibitor develop-
ment as promising compounds (Malcolm Page and Frank Daniels, private communication).
Compounds 2, 7, and 15 have been shown to have some toxicity towards either adult worms

or larvae.

 

133

 

1.7 to 63. Compounds 2 and 15 were shown to be toxic to adult Brugia worms, while com-
pound 7 has been shown to have only moderate toxicity against adult worms and strong

toxicity to worm larvae via in vivo assays (Michael Kron, unpublished results).

Conformer generation for these 16 HTS ligands resulted in an average of 30 conformers
(standard deviation 35.6) per ligand. This large standard deviation reflects the fact that
the flexibility of the HTS ligands varies considerably. Ligand 8 is completely rigid with
respect to its non-hydrogen atoms and yields only a single conformer, while ligand 7 is
quite flexible and yields 100 conformers under the same conformer generation parameters.

Generated conformers for two example HTS ligands are presented in Figure 4.4.

4.3.2 Computational Screening using a Ligand Based Template

To identify ligand candidates which are related to the known ligands, a ligand-based
screening template was generated from the six available AsnRS ligands: three asparagine
molecules, two AMP molecules, and one S-adenosyl-asparagine (S-AMP-Asn) molecule.
This template consisted of 13 points and was able to successfully dock the two AMP
molecules and the S-AMP-Asn ligand, but was not able to dock any of the three Asn ligands
from the crystallographic structures. This is quite likely due to the increased variability of
placement of the asparagine moieties, such that when the template is generated by aver-
aging the atom positions, the template points are placed outside of the favorable binding
positions. Screening with this template against the database of conformers generated from
the 16 HTS ligands yielded no dockings for any of the HTS ligands. One possible expla-

nation for this result is that none of the HTS ligands bind in the active site, which was

134

 

 

 

 

 

 

 

 

Figure 4.4: Conformers generated for HTS ligand 9 (A and B) and HTS ligand 15 (C and
D). Shown in A and C are the 2-dimensional structures of the the HTS ligands; shown
in B and D are the 3-dimensional structures of the generated conformers, each of which is
superimposed onto the first conformer using the phenyl ring in the upper left as a reference.
Ligand 15 is much less flexible and yields a small number of conformations (6) that are
somewhat similar while ligand 9 yields a broader range of conformations (20) due to its
increased flexibility.

 

135

screened against, which could be tested with additional activity assays to determine the
nature of the inhibition. Another explanation based on the screening algorithm is that the
HTS ligands do not strongly resemble the ligands from the crystallographic structure. One
would not expect a template generated from these crystallographic ligands to be able to
select for these different HTS ligands as they do not resemble the known ligands and are

likely to take advantage of different sets of interaction sites within the binding site.

In addition to the screening against the databases of known ligands, either fi'om the
crystallographic structures or from the high-throughput screening tests, the ligand-based
template was used to screen against the CSD for new ligands. A adenosyl compound,
adenosine-5’-methylphosphonate (CSD code ADMPOT 10; Barnes and Hawkinson 1979),
a substrate analog, was the top scoring ligand using either the internal SLIDE scoring func-
tion or the DrugScore scoring function (Gohlke et al. 2000a; discussed briefly in Sec-
tion 3.2.4). The docking of this CSD ligand closely resembles the docking orientation of the
AMP portion of the S-AMP-Asn ligand from the Brugia AsnRS crystallographic structure.
This confirms that the ligand-based template is able to select ligands that mimic the nat-
ural ligands’ binding modes. Other adenosine analogs were also docked with high scores
during the screening. Another ligand that was highly ranked by both the internal SLIDE
scoring firnction and the DrugScore scoring function was variolin B (CSD code LEPWIM;
Perry et al. 1994). Variolin B is a natural product derived from an Antarctic sponge and has
been shown to have possible antiviral and antitumor activity. Other marine sponge products
have been shown to have antihelrninthic and antibiotic effects (Alvi et al., 1991). The two-

dimensional structure of variolin B is presented in Figure 4.5 and the docking orientation

136

 

 

’--------

 

 

 

 

Figure 4.5: Two-dimensional structure of variolin B, CSD code LEPWIM. Circled is the
ring structure mapped onto the purine ring of the AMP portion of the S-AMP-Asn ligand
in the crystal structure of Brugia AsnRS to achieve a second docking of the variolin ligand.

 

as calculated by SLIDE is presented in Figure 4.6. An additional, manual docking was
generated by mapping the variolin B ring circled in Figure 4.5 onto the purine ring of the
adenosyl group, with subsequent rotation of rotatable side-chain and ligand bonds to alle-
viate intermolecular and intramolecular collisions in Insight]! (Accelrys, San Diego, CA).
This docking assessed whether is was possible for variolin B to bind such that it matched
the binding of the adenosyl moiety in S-AMP-Asn. This docking, found to be feasible, also
places a hydroxyl group of variolin B onto the adenosyl N6 group, the hydroxyl oxygen can
act as either a hydrogen bond acceptor, the “normal” hydrogen bond forming role for oxy-
gen atoms, or a donor, thereby mimicking the adenosyl nitrogen’s firnction. This docking,
shown in Figure 4.7, scored higher than the docking in Figure 4.6 when using the internal

SLIDE scoring function, but somewhat lower when using the DrugScore scoring function.

137

 

 

Figure 4.6: Variolin B docking as determined by SLIDE. The S-AMP-Asn ligand from the
Brugia AsnRS crystallographic structure is shown in yellow (AMP portion) and magenta
(Asn portion). Variolin B is colored by atom type: carbon, grey; oxygen, red; and nitrogen,
blue. There is a clear match between the 6-membered ring of the purine group and the
isolated ring of the variolin (upper right). The amine nitrogen of this ring coincides with
N5 of the adenosine, making the same hydrogen bond to the AsnRS protein.

 

138

 

 

Figure 4.7: Variolin B docking assessed manually by superimposing the ring structure high-
lighted in Figure 4.5 onto the purine ring of the AMP moiety of the S-AMP-Asn ligand.
The S-AMP-Asn ligand from the crystallographic structure is shown in yellow (AMP por-
tion) and magenta (Asn portion). Variolin B is colored by atom type: carbon, grey; oxygen,
red; and nitrogen, blue. The S-AMP-Asn ligand is in the same orientation as in Figure 4.6.
This docking orientation yields a more filled AMP lobe of the binding pocket compared to
the SLIDE orientation and scored somewhat higher with the internal SLIDE scoring func-
tion. Thus, the binding modes predicted in this figure and Figure 4.6 are both feasible and
have favorable complementarity with AsnRS.

 

139

 

Given these two favorable binding modes and the known biological activities of variolin B,

it was suggested as a potential ligand for further testing.

4.3.3 Computational Screening using an Unbiased Template

To identify ligand candidates that do not necessarily resemble the known ligands, an unbi-
ased template of 140 points was generated fiom the Brugia AsnRS crystallographic struc-
ture. This template had 50 points which reside at the extreme ends of the S-AMP-Asn
ligand marked as key points, meaning that all matches in SLIDE must include at least one
of these points. Screening with this template against the database of six ligands derived
from the crystallographic structures resulted in dockings of all six of these ligands, indi-
cating the unbiased template can also correctly dock known ligands into the binding site.
SLIDE was also able to dock all 16 of the HTS ligands in at least one, and generally more
than one, orientation. The best scoring dockings, according to the DrugScore scoring firnc-
tion, for two of the HTS compounds are shown in Figures 4.8 and 4.9. This results in
prediction of the mode of binding of the HTS inhibitors to AsnRS, which is not known

experimentally.

In addition to predicting docking orientations of potential AsnRS inhibitors identified
by high-throughput screening, SLIDE was used to identify additional novel potential in-
hibitors, without bias towards known ligands. This was done by screening with the un-
biased template against the CSD as described above. Three of the CSD compounds were
docked and ranked in the top 10 ligands using both the internal SLIDE scoring function and

the DrugScore scoring function (Figure 4.10), indicating these potential ligands form very

140

 

 

 

 

Figure 4.8: Best scoring SLIDE docking for high—throughput screening ligand number
seven. The S-AMP-Asn ligand from the Brugia crystallographic structure is shown in
yellow (AMP portion) and magenta (Asn portion) and the ligand is shown colored by atom
(C, grey; 0, red; N, blue). Side chains in the Brugia AsnRS that were rotated by SLIDE
during the screening process are shown in green, with the crystallographic conformations
shown in grey.

 

141

 

 

Figure 4.9: Best scoring SLIDE docking for high-throughput screening ligand number 13.
The S-AMP-Asn ligand from the Brugia crystallographic structure is shown in yellow
(AMP portion) and magenta (Asn portion) and the ligand is shown colored by atom (C,
grey; 0, red; N, blue). Side chains in the Brugia AsnRS that were rotated by SLIDE during
the screening process are shown in green, with the crystallographic conformations shown
In grey.

 

142

favorable interactions with AsnRS. One of these candidates, MSFURY, has a second crys-
tal form entered in the CSD as entry code MSFURYOI. This second form was also selected

in the screening with rankings of 15 and 2 (SLIDE score and DrugScore respectively).

Cercosporamide, CSD code SIVXIE (Sugawara et al., 1991), was the top scoring can-
didate based on the SLIDE scoring function and was the 10th ranked ligand based on the
DrugScore scoring firnction; it is shown in its docked orientation in Figure 4.11. This
potential ligand fills the AMP lobe of the AsnRS binding pocket quite well, which is likely
to be important for both binding affinity and specificity. Cercosporamide, a phytotoxin
produced by a cassava pathogen, has been shown to have biological activity as a toxin to
plant protoplasts as well as to a variety of fungi (Sugawara et al., 1991). Given the high
scores and the known activities, this would be a good candidate to test for activity against
both Brugia AsnRS and against the Brugia nematodes. It will also be important to test this
and other potential ligands for inhibition of human AsnRS, which would be undesirable
in therapeutical applications. The ultimate goal is to obtain compounds that specifically

inhibit Brugia AsnRS, but not human AsnRS.

A second potential ligand of interest selected during computational screening was phlo-
rizin, CSD code CEWWAC20 (Auf’mkolk et al., 1986). Phlorizin, a dihydrochalcone gly-
coside produced by apple trees, has been shown to inhibit Na+ and glucose transport with
high nanomolar concentration via direct interaction with the Na+/ glucose cotransporter
(Hirayarna et al., 2001). It has also been suggested that it may interact with the NADPH
binding site in some mammalian catalases (Kitlar et al., 1994) and has been shown to be

toxic to malarial pathogens (Loyevsky and Cabantchik, 1994). The docking orientation

143

 

IT ‘

 

 

CEWWAC20 MSFURY

O

 

Figure 4.10: Molecules from the CSD selected as potential ligands by SLIDE screening
with an unbiased template. These molecules (SIVXIE, cercosporarnide; CEWWAC20,
phlorizin; and MSFURY, (E)-4,4’-dimesityl-but-3-enolidylidene-but-3’-enolide) were
ranked in the top 10 potential ligands using both the SLIDE scoring firnction and the
DrugScore scoring function, indicating that they are highly favorable ligands to pursue
further through in vitro and in vivo experimental testing.

 

144

 

 

 

 

Figure 4.11: Docking of cercosporarriide, CSD code SIVXIE, with Brugia asparaginyl—
tRNA synthetase. The S-AMP-Asn ligand from the Brugia crystallographic structure is
shown in yellow (AMP portion) and magenta (Asn portion) for comparison, and the ligand
is shown colored by atom (C, grey; 0, red; N, blue). Rotated side chains are shown in green
with the native positions shown in grey. The docking orientation shown was achieved by
screening with the AsnRS unbiased template against the CSD and was ranked as the best
potential ligand using the internal SLIDE scoring function and the 10th best potential ligand
using the DrugScore scoring function. The AMP pocket of AsnRS is well filled by the
potential ligand, lower center of the figure.

 

145

determined by SLIDE is shown in Figure 4.12. These results suggest the compound is
bioavailable, but also, particularly in the case of binding to mammalian catalases, suggest
the possibility of toxicity or other side effects due to non-specific interactions of this po-
tential inhibitor. Phlorizin was ranked as the 10th best potential ligand using the internal
SLIDE scoring function and the ninth best potential ligand based on the DrugScore scor-
ing function, indicating that this docking to B. malayi AsnRS is still favorable, but not as
favorable as the interactions of the cercosporarnide previously analyzed. The docking de-
termined by SLIDE also fills the AMP lobe of the AsnRS binding pocket, but to a lesser
extent. Also, more of the ligand extends away from the deepest portion of the binding site
(the portion extending down on the right side of Figure 4.12). There are several side chain
rotations in this docking, though most are still quite small, except for the movement of
tyrosine 223, behind the ligands on the left side of the figure. This degree of movement
is still not large and is comparable to what is seen between ligand-flee and ligand-bound
crystal complexes (Maria Zavodszky, unpublished results). In addition to the known activ-
ities, phlorizin is commercially available (Sigma-Aldrich) and would be a good candidate

for in vitro testing against AsnRS and in viva testing against Brugia larvae and nematodes.

The third molecule that was ranked as the 10th or better ligand with both the internal
SLIDE scoring function and the DrugScore scoring function (ranked eighth with both scor-
ing functions) was (E)-4,4’-dimesityl-but-3-enolidylidene-but-3'-enolide, CSD code MS-
FURY (Begley et al., 1981). Figure 4.13 shows the docking orientation determined by
SLIDE for this potential ligand. Unfortunately, little biological work has been done with

this molecule, so no information about potential biological activity exists. It is interesting to

146

 

 

 

‘

Figure 4.12: Docking of phlorizin, CSD code CEWWAC20, with Brugia asparaginyl-tRNA
synthetase. The S-AMP-Asn ligand from the Brugia crystallographic structure is shown
in yellow (AMP portion) and magenta (Asn portion) for comparison, and the ligand is
shown colored by atom (C, grey; 0, red; N, blue). Rotated side chains are shown in green,
with the native positions shown in grey. The docking orientation shown was achieved by
screening with the AsnRS unbiased template against the CSD and was ranked as the 10th
best potential ligand using the internal SLIDE scoring function and the ninth best potential
ligand using the DrugScore scoring function. Phlorizin fills a fair amount of the AMP
binding lobe, lower left. This potential ligand extends further away from the Asn binding
lobe compared to cercosporamide.

 

147

 

 

Figure 4.13: Docking of (E)—4,4’-dimcsityl-but-S-enolidylidene-but-3’—enolide, CSD code
MSFURY, with Brugia asparaginyl-tRNA synthetase. The S-AMP—Asn ligand from the
Brugia crystallographic structure is shown in yellow (AMP portion) and magenta (Asn
portion) for comparison, and the ligand is shown colored by atom (C, grey; 0, red; N,
blue). Rotated side chains are shown in green with the native positions shown in grey. The
docking orientation shown was achieved by screening with the AsnRS unbiased template
against the CSD database and was ranked as the 8th best potential ligand using both the
internal SLIDE scoring function and the DrugScore scoring function. MSFURY also binds
mostly in the AMP lobe of the binding pocket, but extends away from the binding site of
AsnRS (towards the upper right of the figure).

 

148

 

 

note that a second crystal form of MSFURY exists in the CSD, code MSFURYOI, in which
the distal substituated phenyl rings are rotated approximately 90° from planar with respect
to the central rings. (In the MSFURY entry, all four of the rings are roughly coplanar.) This
conformation was also selected by SLIDE as a potential ligand, with SLIDE making con-
formational changes to the ligand to rotate the distal phenyl rings closer to being coplanar
with the central rings. The resulting dockings are very similar, indicating the SLIDE has
the ability to select different conformations of the same molecule as potential ligands. This
potential ligand is a possibility for testing given its high scores and multiply docked con-
formations, however given its significant hydrophobic character, suggesting low solubility,

and the absence of commercial availability, it could prove diflicult to test.

4.4 Discussion — Analysis of Potential Ligands selected by

SLIDE

As a summary of the above results, a computation screening study was performed using
asparaginyl-tRNA synthetase (AsnRS) fi'om Brugia malayi as a target. A template gener-
ated from six ligands taken fiom the crystallographic structures was generated and used to
screen against a subset of the Cambridge Structural Database (CSD) and identified variolin
B as a potential ligand. An unbiased template was generated and used to suggest binding
orientations of the 16 inhibitors selected fiom in vitro high-throughput screening against
AsnRS. This unbiased template was also used to screen the CSD subset and identified three

potential ligands: cercosporarnide, phlorizin, and (E)-4,4’-dimesityl-but-3-enolidylidene-

149

but-3’-enolide.

One of the difl‘iculties encountered in virtual screening experiments is the selection of
the best potential ligands. Generally, a large number of molecules pass through to the scor-
ing step. The number of molecules reaching this step can be controlled by changing the
stringency of the screening parameters, such the allowed interatomic overlap parameters
in SLIDE. However, increasing the stringency of these steric aspects of, which decreases
the number of potential ligands, also increases the chance of rrrissing worthwhile potential
ligands. The ideal case would be to use loose parameters, allowing many molecules to
pass, and then have a good scoring function for protein:ligand complementarity with which
to choose the most promising candidates for further work, such as in vitro inhibition as-
says followed by drug lead optimization for successful inhibitors. While there are several
scoring firnctions and docking forcefields available (Bdhm 1994; Jain 1996; Eldridge et al.
1997; Béhm 1998; Murray et al. 1998; Miigge and Martin 1999; reviewed in Section 1.4),
all have shortcomings. One problem with many of the currently available forcefields is
their sensitivity to minor changes in relative protein-1i gand orientation. During the screen-
ing process, it is not possible to finely tune each orientation due to computational time
costs, but the untuned orientation may have a significantly lower score, thereby seeming
like a poor potential ligand when minor changes could make it rise much higher in the score
rankings. Such problems could potentially be handled by doing subsequent refinement of
the orientation, using such a forcefield coupled with a molecular dynamics algorithm, for

the best scoring candidates.

The approach of consensus scoring, suggested by Charifson et al. (1999) and applied in

150

 

several cases (Bissantz et al., 2000; Stahl and Rarey, 2001; Tripos Associates, 2001), is to
score each docking with several scoring functions and then choose the dockings that score
well with several scoring firnctions. The work presented here used a simple consensus
function in which potential ligands which score well with both the SLIDE scoring function
and the DrugScore scoring function are examined in detail. Various methods have been
suggested and evaluated with respect to weighting and combination of the scoring functions

(Wang and Wang, 2001).

One advantage of the internal SLIDE scoring function over DrugScore is the clear sep-
aration of the score into the hydrophobic contribution and the hydrogen bonding contribu-
tion. This separation of score into components can be a very usefiIl to increase the aflinity
of a particular ligand as it can point towards deficiencies in the current ligand, e.g., that
it is too hydrophobic or has too high a positive charge. Several other scoring functions
employ this separation, but many of the empirically based scoring functions do not make
such distinctions. Unfortunately, since DrugScore does not include a separation of scores in
component parts, it is not possible to examine the reasons why some potential ligands score
quite highly with one scoring function, but quite poorly with another. In fact, in general,
the scoring firnctions do not correlate well with each other; over seven various screening
runs, the mean correlation was only 059 (standard deviation of 0. l 7), indicating why con-
sensus scoring is important. The ability to disentangle the score into its component parts
is an advantage to forcefield based approaches, but often empirical approaches yield better
correlation with binding affinities. The advantages and disadvantages with each scoring

function point directly to the problem that none of the currently available scoring func-

151

tions are extremely accurate and that additional research must be done to improve both the
accuracy and computational complexity of current functions to yield functions which can
accurately compute the binding affinity of a docked protein-ligand complex in a reasonable
amount of time. However, development of an improved scoring function is beyond the

scope of this work.

Another currently difficulty of analysis of potential ligands selected by computational
screening techniques is that of multiplicity, i.e., the identification of multiple binding ori-
entations of a single ligand. In cases where the desire is to simply obtain a list of potential
ligands, this may not be a concern, but in cases when one wants to use the docking results
to obtain insight into how a protein may fimction, it becomes a significant problem. This
situation arises in SLIDE since each database molecule interaction point triplet is matched
to each template point template. Two sets of pairings may orient the database molecule
in virtually the same orientation with respect to the target protein. The simplest approach
to resolve this would be to employ a clustering algorithm on the ligand positions, giving
a set of most similar orientations. This issue becomes more prevalent when dealing with
multiple conformers of a ligand. Each of the multiple conformers may dock in multiple
orientations, leading to a large increase in the number of dockings to be analyzed. While
the docking orientations of very different conformations cannot be the same, molecular
conformers which start in reasonably close conformations can merge into a very similar
docking when SLIDE rotates ligand bonds. However, even with varying final, docked con-
formations, it may be instructional to identify groups in the ligands which tend to bind in

similar locations in the target binding site. Direct visual analysis for a few dockings is pos-

152

sible, but when many favorable orientations are identified, as in this study, where 3 to 5300
orientations were identified for each HTS ligand, visual analysis becomes impossible. In
this case, one could generate distribution maps for functional groups of particular interest,
such as those generated in Isostar (Bruno et al., 1997) for molecules in the CSD and Protein

Databank (PDB) databases.

Yet a third question arises from analysis of a set of potential ligands: are there
trends in the types of molecules selected? Techniques for clustering molecules using
a one-dimensional chemical representation of the molecules (Barnard et al., 2000), i.e.,
the SMILES string (Weininger, 1988), using one-dimensional bitstring representations
of the molecules (Matter, 1997), and using a compatibility based Tanimoto coefficient
(Verkhivker et al., 2000), have been previously implemented and could be applied in SLIDE
as a post-screening step. The identification of groupings of potential ligands is useful to
determine general chemical characteristics of potential ligands (i.e., are there specific func-
tional groups which all selected molecules share?) and to reduce the ligands to a set which
is analyzable. Also, having such clusters would yield information about the diversity of
molecules which can bind to the target by analysis of the molecule clusters produced, i.e,
are there many clusters containing only a small number of ligands produced, indicating di-
verse binding, or are there only a few, well occupied clusters produced, indicating binding
over only a narrow range of molecules. A practical result of clustering would be to provide
information about chemically similar neighbors of potential ligands that are commercially
available for unavailable selected molecules or may be easier to work with experimentally

than selected molecules. A second application of such a technique would be to estimate

153

 

the promiscuity of a protein, i.e., the ability of a protein to bind to diverse types of ligands,

and may help identify difficult targets for drug design work.

The use of computational screening tools is to provide a set of molecules with likely
binding to a protein of interest for further testing. The ability to dock known ligands
strongly suggests that ligands which reside in the database used for screening will be se-
lected in the list of potential ligands. The scoring function can rank these potential ligands
to give a set of more probable actual ligands, but knowledge of the researcher can play
a significant role in what potential ligands are most likely to dock, or, in pharmaceutical
companies, are promising candidates for drug development. One approach would be to
limit the database to molecules that are “drug-like” molecules, such as those that follow
Lipinski’s “rule of fives” (Lipinski et a1., 2001). Other options would be simply limit the
molecular weight of the molecules or exclude molecules which have a certain functional
group. Another problem that sometimes arises is the availability of selected compounds
for fmther work. While this may be less of a concern to a large, pharmaceutical company
with the ability to readily synthesize many compounds, small research entities may wish to
limit the screening database to only available compounds, such as those on the Available
Chemicals Database (ACD). All of this gets down to limiting the screening database to the

most promising candidates, which is the main goal of computational screening techniques.

The use of computational screening is only likely to increase. One idea not yet widely
considered is the idea of reverse screening, i.e., beginning with a ligand molecule and
identifying which proteins in a database, e.g., the PDB, it could dock to. This would

require a automated technique to identify potential binding sites on proteins. Such a method

154

 

could provide leads as to what targets compounds of unknown function may bind too,
leading to elucidation of their role in the cell, as well as provide an estimate of cross-
reactivity potential for a drug candidate. While impossible to do now due to inadequate
algorithmic techniques and computational resources, perhaps a distant future screening
application would be to do computational hybrid screening, i.e., screening a protein of
interest for binding to a set of database proteins, which could prove especially applicable as
the proteomics movement comes into its own. One can be assured that new and innovative

screening techniques and applications will continue to arrive on the scene.

155

Appendix A

Summary of Publications Outside of the Scope of

the Work Presented in this Dissertation

o M. L. Raymer, P. C. Sanschagrin, W. F. Punch, S. Venkataraman, E. D. Goodman,
and L. A. Kuhn. Predicting conserved water-mediated and polar ligand interactions
in proteins using a k-nearest-neighbors genetic algorithm. J. Mol. Biol., 265:445—

464, 1997.

Water-mediated ligand interactions are essential to biological processes,
from product displacement in tlrymidylate synthase to DNA recognition
by Trp repressor, yet the structural chemistry influencing whether bound
water is displaced or participates in ligand binding is not well character-
ized. Consolv, employing a hybrid k-nearest-neighbors classifier/genetic
algorithm, predicts bound water molecules conserved between free and
1i gand-bound protein structures by examining the environment of each wa-
ter molecule in the free structure. Four environmental features are used:

156

 

the water molecule’s crystallographic temperature factor, the number of
hydrogen bonds between the water molecule and protein, and the den-
sity and hydrophilicity of neighboring protein atoms. After training on 13
non-homologous proteins, Consolv predicted the conservation of active-
site water molecules upon ligand binding with 75% accuracy (Matthews
coefficient Cm = 0.41) for seven new proteins. Mispredictions typically
involved water molecules predicted to be conserved that were displaced
by a polar ligand atom, indicating that Consolv correctly assesses polar
binding sites; 90% accuracy (Cm = 0.78) was achieved for predicting con-
served active-site water or polar ligand atom binding. Consolv thus pro-
vides an accurate means for optimizing ligand design by identifying sites
favored to be occupied by either a mediating water molecule or a polar lig-
and atom, as well as water molecules likely to be displaced by the ligand.
Accuracy for predicting first-shell water conservation between indepen-
dently determined structures was 61% (Cm=0.23). The ability to predict
water-mediated and polar interactions from the fi'ee protein structure in-
dicates the surprising extent to which the conservation or displacement of
active-site bound water is independent of the ligand, and shows that the
protein micro-environment of each water molecule is the dominant influ-

ence.

e M. L. Raymer, W. F. Punch, E. D. Goodman, P. C. Sanschagrin, and L. A. Kuhn.

Simultaneous feature scaling and selection using a genetic algorithm. In T. Back,

157

 

editor, Proceedings of the Seventh International Conference on Genetic Algorithms,

pages 561—567. Morgan Kaufrnann Publishers, 1997.

Statistical pattern recognition techniques classify objects in terms of a
representative set of features. The selection of features to measure and
include can have a significant effect on the cost and accuracy of an auto-
mated classifier. Our previous research has shown that a hybrid between
a k-nearest-neighbors (knn) classifier and a genetic algorithm (GA) can
achieve greater classification accuracy than a knn alone by weighting fea-
tures during knn classification. Here we describe an extension to this ap-
proach which further enhances feature selection through the simultaneous
optimization of feature weights and selection of key features by includ-
ing a masking vector on the GA chromosome. We present the results of
our GA/knn feature selection method on two important problems from
biochemistry and medicine: identification of conserved water molecules
bound to protein surfaces, and diagnosis of thyroid deficiency. By allow-
ing the GA to explore the effect of eliminating a feature from the classi-
fication ‘without losing the weight knowledge already learned, the feature
masking technique allows the GA/knn to efliciently examine noisy, com-
plex, and high-dimensionality datasets to find combinations of features
which classify the data more accurately. In both biomedical applications,
use of the feature masking technique resulted in equivalent or better ac-

curacy than feature weighting alone, while using fewer features for the

158

classification.

0 L. Craig, P. C. Sanschagrin, A. Rozek, S. Lackie, L. A. Kuhn, and J. K. Scott. The
role of structure in antibody cross-reactivity between peptides and folded proteins. J.

Mol. Biol., 281:183—201, 1998.

Peptides have the potential for targeting vaccines against pre-specified
epitopes on folded proteins. When polyclonal antibodies against native
proteins are used to screen peptide libraries, most of the peptides isolated
align to linear epitopes on the proteins. The mechanism of cross-reactivity
is unclear; both structural mimicry by the peptide and induced fit of the
epitope may occur. The most effective peptide mimics of protein epitopes
are likely to be those that best mimic both the chemistry and the structure
of epitopes. Our goal in this work has been to establish a strategy for char-
acterizing epitopes on a folded protein that are candidates for structural
mimicry by peptides. We investigated the chemical and structural bases of
peptide-protein cross-reactivity using phage-displayed peptide libraries in
combination with computational structural analysis. Polyclonal antibodies
against the well-characterized antigens, hen eggwhite lysozyme and worm
myohemerythrin, were used to screen a panel of phage-displayed peptide
libraries. Most of the selected peptide sequences aligned to linear epi-
topes on the corresponding protein; the critical binding sequence of each
epitope was revealed from these alignments. The structures of the critical

sequences as they occur in other non-homologous proteins were analyzed

159

 

using the Sequery and Superpositional Structural Assignment computer
programs. These allowed us to evaluate the extent of conformational pref-
erence inherent in each sequence independent of its protein context, and
thus to predict the peptides most likely to have structural preferences that
match their protein epitopes. Evidence for sequences having a clear struc-
tural bias emerged for several epitopes, and synthetic peptides represent-
ing three of these epitopes bound antibody with sub-nricromolar aflinities.
The strong preference for a type II beta-tum predicted for one peptide
was confirmed by NMR and circular dichroism analyses. Our strategy
for identifying conforrnationally biased epitope sequences provides a new

approach to the design of epitope-targeted, peptide-based vaccines.

o L. Fan, P. C. Sanschagrin, L. S. Kaguni, and L. A. Kuhn. The accessory subunit of
mtDNA polymerase shares structural homology with arninoacyl-tRNA synthetases:
Implications for a dual role as a primer recognition factor and processivity clamp.

Proc. Natl. Acad. Sci. USA, 96(17):9527—32, 1999.

The accessory subunit of the heterodirneric mtDNA polymerase (p017)
from Drosophila embryos is required to maintain the structural integrity
or catalytic efficiency of the holoenzyrne. cDNAs for the accessory sub-
unit from Drosophila, man, mouse, and rat have been identified, and com-
parative sequence alignment reveals that the C-terrninal region of about
120 aa is the most conserved. Furthermore, we demonstrate that the ac-

cessory subunit of animal p017 has both sequence and structural similarity

160

 

with class Ila aminoacyl-tRNA synthetases. Based on sequence similar-
ity and fold recognition followed by homology modeling, we have devel-
oped a model of the three-dimensional structure of the C-terminal region
of the accessory subunit of p017. The model reveals a rare five-stranded
beta-sheet surrounded by four alpha-helices with structural homology to
the anticodon-binding domain of class Ila aminoacyl-tRNA synthetases.
We postulate that the accessory subunit plays a role in the recognition
of RNA primers in mtDNA replication, to recruit poly to the template-
primer junction. A similar role is served by the 7-complex in Escherichia
coli DNA polymerase III, and indeed our accessory subunit model shows
structural similarity with the N-terrrrinal domain of the 6’ subunit of the 7—
complex. Structural similarity is also found with E. coli thioredoxin, the
accessory subunit and processivity factor in bacteriophage T7 DNA poly-
merase. Thus, we propose that the accessory subunit of p017 is involved

both in primer recognition and in processive DNA strand elongation.

161

Bibliography

R. Abagyan, M.. Totrov, and D. Kuznetsov. ICM-a new method for protein modeling
and design: Applications to docking and structure prediction from the distorted native
conformation. J. Comput. Chem, 15(5):488—506, 1994.

F. H. Allen and O. Kennard. 3D search and research using the Cambridge Structural
Database. Chemical Design Automation News, 8:1 & 31—37, 1993.

K. A. Alvi, L. Tenenbaum, and P. Crews. Antlrelmintic polyfunctional nitrogen-containing
terpenoids from marine sponges. J. Nat. Prod, 54(1):71—78, 1991.

J. G. Amez and D. Moras. Structural and functional considerations of the aminoacylation
reaction. Trends Biochem. Sci, 22(6):211—216, 1997.

M. Auf’mkolk, J. Koehrle, R. D. Hesch, S. H. Ingbar, and V. Cody. Crystal structure of
phlorizin and the iodothyronine deiodinase inhibitory activity of phloretin analogues.
Biochem. Pharmacol, 35(13):2221—2227, 1986.

G. Ausiello, G. Cesarani, and M. Helmer-Citterich. ESCHER: A new docking procedure
applied to the reconstruction of protein tertiary structure. Prots, 28:556—567, 1997.

K. Awadzi. Research notes from the Onchocerciasis Chemotherapy Research Cen-
tre,Ghana. Ann. Trop. Med. Parasitol, 91:703—71 1, 1997.

E. N. Baker and R. E. Hubbard. Hydrogen bonding in globular proteins. Prog. Biophys.
Mol. Biol., 44(2):97-179, 1984.

J. M. Barnard, G. M. Downs, A. von Scholley-Pfab, and R. D. Brown. Use of Markush
structure analysis techniques for descriptor generation and clustering of large combina-
torial libraries. J. Mol. Graph. Model, 18(4-5):452—463, 2000.

C. L. Barnes and S. W. Hawkinson. Structure of adenosine 5’-methylphosphonate hemihy-
drate. Acta Crystallogn B, 35:1724—1727, 1979.

C. R. Beddell, P. J. Goodford, F. E. Norrington, S. VVIlkinson, and R. Wootton. Compounds
designed to fit a site of known structure in human haemoglobin. Br: J. Pharmacol, 57
(2):201—209, 1976.

M. J. Begley, L. Crombie, G. L. Griffiths, R. C. F. Jones, and M. Rahmani. Charge trans-
fer and noncharge transfer forms of (e)-5,5’-dimesity1bifuranylidenediones: An X-ray
structural investigation. J. Chem. Soc.,Chem. Commun, 16:823—825, 1981.

162

 

G. W. Bemis and M. A. Murcko. The properties of known drugs. 1. Molecular frameworks.
J. Med. Chem, 39(15):2887—2893, 1996.

H. M. Berrnan, J. Westbrook, Z. Feng, G. Gilliland, T. N. Bhat, H. Weissig, I. N.
Shindyalov, and P. E. Boume. The Protein Data Bank. Nucleic Acids Res., 28:235—
242, 2000. http: / /www. rscb . org/pdb/.

P. Bernard, A. Golbraikh, D. Kireev, J. R. Chretien, and N. Rozhkova. Comparison of
chemical databases: Analysis of molecular diversity with self-organising maps (SOM).
Analysis, 26:333—341 , 1998.

F. C. Bernstein, T. F. Koetzle, G. J. B. Williams, E. F. Meyer, Jr., M. D. Brice, J. R. Rodgers,
0. Kennard, T. Shimanouchi, and M. Tasumi. The Protein Data Bank: A computer-based
archival file for macromolecular structures. J. Mol. Biol., 112:535-542, 197 7 .

M. J. Betts and M. J. E. Stemberg. An analysis of conformational changes on protein-
protein association: Implications for predictive docking. Protein Eng, 12:271—283,
1999.

C. Bissantz, G. Folkers, and D. Rognan. Protein-based virtual screening of chemical
databases. 1. Evaluation of different docking/scoring combinations. J. Med. Chem, 43
(25):4759—4767, 2000.

R. A. Blevins and A. Tulinsky. Comparison of the independent solvent structures of dirneric
alpha-chymotrypsin with themselves and with gamma-chymotrypsin. J. Biol. Chem, 260
(15):8865—8872, 1985.

N. S. Blom and J. Sygusch. High resolution fast quantitative docking using Fourier domain
correlation techniques. Proteins, 27(4):493—506, 1997.

D. M. Blow, J. J. Birktofi, and B. S. Hartley. Role of a buried acid group in the mechanism
of action of chymotrypsin. Nature, 221:337—340, 1969.

H. J. Biihm. The computer program LUDI: a new method for the de novo design of enzyme
inhibitors. J. Comput. Aided. Mol. Des., 6(1):61—78, l992a.

H. J. Biihm. LUDI: Rule-based automatic design of new substituents for enzyme inhibitor
leads. J. Comput. Aided. Mol. Des., 6(6):593—606, l992b.

H. J. Br’ihm. 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(3):243-256, 1994.

H. J. Biihrn. Prediction of binding constants of protein ligands: a fast method for the
prioritization of hits obtained fiom de novo design or 3D database search programs. J.
Comput. Aided. Mol. Des., 12(4):309—323, 1998.

J. Bostriim, P. O. Norrby, and T. Liljefors. Conformational energy penalties of protein-
bound ligands. J. Comput. Aided. M01. Des., 12(4):383—396, 1998.

163

 

B. C. Braden, B. A. Fields, and R. J. Poljak. Conservation of water molecules in an
antibody-antigen interaction. J. Mol. Recognit., 8(5):3l7-—325, 1995.

R. D. Brown and Y. C. Martin. Use of structure-activity data to compare structure-based
clustering methods and descriptors for use in compound selection. J. Chem. Inf Comput.
Sci. , 36(3):572—584, 1996.

R. M. Brunne, E. Liepinsh, G. Otting, K. Wiithrich, and W. F. van Gunsteren. Hydration
of proteins. A comparison of experimental residence times of water molecules solvating
the bovine pancreatic trypsin inhibitor with theoretical model calculations. J. Mol. Biol. ,
231:1040—1048, 1993.

I. J. Bruno, J. C. Cole, J. P. Lommerse, R. S. Rowland, R. Taylor, and M. L. Verdonk.
IsoStar: a library of information about nonbonded interactions. J. Comput. Aided. Mol.
Des., 11(6):525—537, 1997.

S. K. Burley and G. A. Petsko. Aromatic-aromatic interaction: a mechanism of protein
structure stabilization. Science, 229(4708):23—28, 1985.

M. W. Casewell and R. L. Hill. In-vitro activity of mupirocin (’pseudomonic acid’) against
clinical isolates of Staphylococcus aureus. J. Antimicrob. Chemother, 15(5):523—531,
1985.

P. S. Charifson, J. J. Corkery, M. A. Murcko, and W. P. Walters. Consensus scoring: A
method for obtaining improved hit rates fi'om docking databases of three-dimensional
structures into proteins. J. Med. Chem, 42(25):5100—5109, 1999.

H. Claussen, C. Buning, M. Rarey, and T. Lengauer. FLEXE: Efficient molecular docking
considering protein structure variations. J. Mol. Biol., 308(2):377—395, 2001.

G. M. Clore, A. Bax, P. T. VVmgfield, and A. M. Gronenbom. Identification and localization
of bound internal water in the solution structure of interleukin 1 beta by heteronuclear
three-dimensional 1H rotating-frame Overhauser 15N-1H multiple quantum coherence
NMR spectroscopy. Biochemistry, 29(24):5671-5676, 1990.

M. L. Connolly. Solvent-accessible surfaces of proteins and nucleic acids. Science, 221:
709—713, 1983. http: / /www. biohedron . com.

L. Craig, P. C. Sanschagrin, A. Rozek, S. Lackie, L. A. Kuhn, and J. K. Scott. The role
of structure in antibody cross-reactivity between peptides and folded proteins. J. Mol.
Biol., 281:183—201, 1998.

D. J. Cummins, C. W. Andrews, J. A. Bentley, and M. Cory. Molecular diversity in chem-
ical databases: Comparison of medicinal chemistry knowledge bases and databases of
commercially available compounds. J. Chem. Inf Comput. Sci, 36(4):750—763, 1996.

S. Cusack. Eleven down and nine to go. Nat. Struct. Biol., 10:824—831, 1995.

164

 

O. D. Dang and E. Di Cera. Residue 225 determines the Na+-induced allosteric regulation
of catalytic activity in serine proteases. Proc. Natl. Acad. Sci. USA, 93(20): 10653—10656,
1996.

V. P. Denisov, J. Peters, H. D. H6rlein, and B. Halle. Using buried water molecules to
explore the energy landscape of proteins. Nat. Struct. Biol., 3:505—509, 1996.

R. L. DesJarlais, R. P. Sheridan, G. L. Seibel, J. S. Dixon, 1. D. Kuntz, and R. Venkataragha-
van. Using shape complementarity as an initial screen in designing ligands for a receptor
binding site of known three-dimensional structure. J. Med. Chem, 31(4):722—729, 1988.

R. S. DeWitte and E. I. Schaknovich. SMoG: de novo design method based on simple,
fast, and accurate free energy estimates. 1. methodology and supporting evidence. J.
Am. Chem. Soc., 118:1651—1663, 1996.

E. Di Cera, E. R. Guinto, A. Vindigni, O. D. Dang, Y. M. Ayala, M. Wuyi, and A. Tulinsky.
The Na+ binding site of thrombin. J. Biol. Chem, 270(38):22089—22092, 1995.

J. S. Dixon. Evaluation of the CASP2 docking section. Proteins, Suppl: 198-204, 1997 .

R. L. Dunbrack, Jr. and F. E. Cohen. Bayesian statistical analysis of protein side-chain
rotamer preferences. Protein Sci, 6(8): 1661—1681, 1997.

T. Earnest, E. Fauman, C. S. Craik, and R. Stroud. 1.59 A structure of trypsin at 120
K: Comparison of low temperature and room temperature structures. Proteins, 10(3):
171—187, 1991.

M. B. Eisen, D. C. Wiley, M. Karplus, and R. E. Hubbard. HOOK: a program for find-
ing novel molecular architectures that satisfy the chemical and steric requirements of a
macromolecule binding site. Proteins, 19(3):]99—221, 1994.

D. Eisenberg and A. D. McLachlan. Solvation energy in protein folding and binding.
Nature, 319(6050):199—203, 1986.

M. D. Eldridge, C. W. Murray, T. R. Auton, G. V. Paolini, and R. P. Mee. Empirical scoring
functions: I. The development of a fast empirical scoring function to estimate the binding
aflinity of ligands in receptor complexes. J. Comput. Aided. Mol. Des., 11(5):425—445,
1997 .

C. T. Esmon. The protein C anticoagulant pathway. Arterioscler: Thromb., 12(2):135—145,
1992.

T. Ewing and I. D. Kuntz. Critical evaluation of search algortihms for automated molecular
docking and database screening. J. Comput. Chem, 18:1175-1189, 1997.

C. H. Faerman and P. Andrew Karplus. Consensus preferred hydrtation sites in six
FKBP12-drug complexes. Proteins Struct. F unct. Genet, 23: 1—1 1, 1995.

165

E. B. Fauman, E. E. Rutenber, G. F. Maley, F. Maley, and R. M. Stroud. Water-
mediated substrate/product discrimination: The product complex of thyrnidylate syn-
thase at 1.83 A. Biochemsitry, 33(6):]502—1511, 1994.

A. V. Filikov, V. Mohan, T. A. Vickers, R. H. Griffey, P. D. Cook, R. A. Abagyan, and T. L.
James. Identification of ligands for RNA targets via structure-based virtual screening:
HIV-l TAR. J. Comput. Aided. Mol. Des., l4(6):593—610, 2000.

J. S. Finer-Moore, A. A. KossiakofiZ‘J. H. Hurley, T. Earnest, and R. M. Stroud. Solvent
structure in crystals of trypsin determined by X-ray and neutron diffraction. Proteins
Struct. Funct. Genet, 12:203—222, 1992.

D. Fischer, S. L. Lin, H. L. Wolfson, and R. Nussinov. A geometry-based suite of molecular
docking processes. J. Mol. Biol., 248(2):459—477, 1995.

E. Fischer. Einfluss der configuration auf die wirkung der enzyme. Berichte Deutsche
Chemische Gesellschaft, 27:2985—2993, 1894.

P. A. Fitzpatrick, A. C. Steinmetz, D. Ringe, and A. M Klibanov. Enzyme crystal structure
in a neat organic solvent. Proc. Natl. Acad. Sci. USA, 90(18):8653—8657, 1993.

J. D. Forman-Kay, A. M. Gronenbom, P. T. \Vrngfield, and G. M. Clore. Determination of
the positions of bound water molecules in the solution structure of reduced human tlriore-

doxin by heteronuclear three-dimensional nuclear magnetic resonance spectroscopy. J.
Mol. Biol., 220(2):209—2 16, 1991.

X. Fradera, R. M. Knegtel, and J. Mestres. Similarity-driven flexible ligand docking. Pro-
teins, 40(4):623—636, 2000.

A. T. Fuller, G. Mellows, M. Woolford, G. T. Banks, K. D. Barrow, and E. B. Chain.
Pseudomonic acid: an antibiotic produced by Pseudomonas fluoreScens. Nature, 234
(5329):416—417, 1971.

B. Furie and B. C. Furie. The molecular basis of blood coagulation. Cell, 53(4):505—518,
1988.

D. K. Gehlhaar, K. E. Moerder, D. Zichi, C. J. Sherman, R. C. Ogden, and S. T. Freer. De
nova design of enzyme inhibitors by Monte Carlo ligand generation. J. Med. Chem, 38
(3):466—472, 1995a.

D. K. Gehlhaar, G. M. Verkhivker, P. A. Rejto, C. J. Sherman, D. B. Fogel, L. J. Fogel,
and S. T. Freer. Molecular recognition of the inhibitor AG-l343 by HIV-l protease:

Conforrnationally flexible docking by evolutionary programming. Chem. Biol., 2(5):
317—324, l995b.

V. Gillet, A. P. Johnson, P. Mata, S. Sike, and P. Williams. SPROUT: a program for structure
generation. J. Comput. Aided. Mol. Des., 7(2): 127—153, 1993.

166

V. J. Gillet, W. Newell, P. Mata, G. Myatt, S. Sike, Z. Zsoldos, and A. P. Johnson. SPROUT:
Recent developments in the de novo design of molecules. J. Chem. Inf Comput. Sci, 34
(l):207—217, 1994.

H. Gohlke, M. Hendlich, and G. Klebe. Knowledge-based scoring function to predictedge-
based scoring function to predict protein-ligand interactions. J. Mol. Biol., 295:337—356,
2000a.

H. Gohlke, M. Hendlich, and G. Klebe. Predicting binding modes, binding affinities and
‘hot spots’ for protein-1i gand complexes using a knowlege-based scoring function. Per-
spectives in Drug Discovery and Design, 20:1 15—144, 2000b.

P. J. Goodford. Drug design by the method of receptor fit. J. Med. Chem, 27(5):558—564,
1984.

P. J. Goodford. A computational procedure for determining energetically favorable binding
sites on biologically important macromolecules. J. Med. Chem, 28(7):849-857, 1985.

D. S. Goodsell, G. M. Morris, and A. J. Olson. Automated docking of flexible ligands:
Applications of AutoDock. J. Mol. Recognit., 9(1): 1—5, 1996.

J. Greer, J. W. Erickson, J. J. Baldwin, and M. D. Vamey. Application of the three-
dirnensional structures of protein target molecules in structure-based drug design. J.
Med. Chem, 37(8):]035—1054, 1994.

S. Ha, R. Andreani, A. Robbins, and I. Muegge. Evaluation of docking/scoring approaches:
a comparative study based on MMP3 inhibitors. J. Comput. Aided. Mol. Des., 14(5):
43 5—448, 2000.

S. M. Habermann and K. P. Murphy. Energetics of hydrogen bonding in proteins: a model
compound study. Protein Sci, 5(7): 1229—1239, 1996.

T. A. Halgren. Merck Molecular Forcefield. '1. Basis, form, score, parametrization, and
performance of MMFF94. J. Comput. Chem, 17:490—519, 1996.

R. D. Head, M. L. Smythe, T. I. Oprea, C. L. Waller, S. M. Green, and G. R. Marshall.
VALIDATE: A new method for the receptor-based prediction of binding affinities of
novel ligands. J. Am. Chem. Soc, 118(16):3959—3969, 1996.

M. Helmer-Citterich and A. Tramontano. PUZZLE: a new method for automated pro-
tein docking based on surface shape complementarity. J. Mol. Biol., 235(3): 1 021—103 1 ,
1994.

B. A. Hirayarna, A. Diez-Sampedro, and E. M. Wright. Common mechanisms of inhibition
for the Na(+)/ glucose (hSGLT‘l) and Na(+)/Cl(-)/GABA (hGATl) cotransporters. Br. J.
Pharmacol., l34(3):484—495, 2001.

R. W. Hoofi, C. Sander, and G. Vriend. Positioning hydrogen atoms by optimizing
hydrogen-bond networks in protein structures. Proteins, 26(4):363—37 6, 1996.

167

J. Hughes and G. Mellows. Interaction of pseudomonic acid A with Escherichia coli B
isoleucyl-tRNA synthetase. Biochem. J., 19l(l):209—219, 1980.

R. M. Jackson, H. A. Gabb, and M. J. E. Stemberg. Rapid refinement of protein interfaces
incorporating solvation: Application to the docking problem. J. Mol. Biol. , 27 6:265—285,
1998.

A. K. Jain and B. Chandrasekaran. Dirnensionality and sample size considerations in pat-
tern recognition in practice. In P. R. Krishnaiah and L. N. Kanal, editors, Handbook of
Statistics, volume 2, pages 835-855. North-Holland, 1982.

A. N. Jain. Scoring noncovalent protein-ligand interactions: A continuous differentiable
function tuned to compute binding aflinities. J. Comput. Aided. Mol. Des., 10:427-440,
1996.

J. Janin and C. Chothia. The structure of protein-protein recognition sites. J. Biol. Chem,
265(27):]6027—16030, 1990.

A. Joachimiak, T. E. Haran, and P. B. Sigler. Mutagenesis supports water mediated recog-
nition in the Trp repressor-operator system. EMBO J., 245 :43—53, 1994.

G. Jones, P. VVrllett, and R. C. Glen. Molecular recognition of receptor sites using a genetic
algorithm with a description of desolvation. J. Mol. Biol., 245 :43—53, 1995.

G. Jones, P. \Vrllett, R. C. Glen, A. R. Leach, and R. Taylor. Development and validation
of a genetic algorithm for flexible docking. J. Mol. Biol., 267(3):727—748, 1997.

P. A. Karplus and C. Faerman. Ordered water in macromolecular structure. Curr: Opin.
Struct. Biol., 4:770—776, 1994.

T. Kitlar, F. Doring, D. F. Diedrich, R. Frank, H. Wallrneier, R. K. Kinne, and J. Deutscher.
Interaction of phlorizin, a potent inhibitor of the Na+/D-glucose cotransporter, with the
NADPH-binding site of mammalian catalases. Protein Sci, 3(4):696—700, 1994.

G. Klebe. The use of composite crystal-field environments in molecular recognition and
the de novo design of protein ligands. J. Mol. Biol., 237(2):212—235, 1994.

G. Klebe. Recent developments in structure-based drug design. J. Mol. Med, 78(5):269—
281, 2000.

G. Klebe and T. Mietzner. A fast and efficient method to generate biologically relevant
conformations. J. Comput. Aided. Mol. Des., 8(5):583-606, 1994.

R. M. Knegtel and M. Wagener. Efficacy and selectivity in flexible database docking.
Proteins, 37(3):334—345, 1999.

P. Koch] and M. Delarue. Application of a self-consistent mean field theory to predict
protein side-chains conformation and estimate their conformational entropy. J. Mol.
Biol., 239:249—275, 1994.

I68

 

P. Koehl and M. Delarue. Mean-field minimization methods for biological macromolecules.
Curr: Opin. Struct. Biol., 6:222—226, 1996.

M. Konishi, M. Nishio, K. Saitoh, T. Miyaki, T. Oki, and H. Kawaguchi. Cispentacin, a new
antifimgal antibiotic. 1. Production, isolation, physico-chernical properties and structure.
J. Antibiot. (Tokyo)., 42( 12):1749—1755, 1989.

I. Konrad and R. Roschenthaler. Inhibition of phenylalanine tRNA synthetase fi'om Bacillus
subtilis by ochratoxin A. FEBS Lett., 83(2):341—347, 1977.

A. A. Kossiakoff, M. D. Sintchak, J. Shpungin, and L. G. Presta. Analysis of solvent struc-
ture in proteins using neutron D20-H20 solvent maps: Pattern of primary and secondary
hydration in trypsin. Proteins Struct. F unct. Genet, 12:223—236, 1992.

B. Kramer, M. Rarey, and T. Lengauer. Evaluation of the FLEXX incremental construction
algorithm for protein-ligand docking. Proteins, 37(2):228-241, 1999.

M. M. Krem and E. Di Cera. Conserved water molecules in the specificity pocket of serine
proteases and the molecular mechanism of Na+ binding. Proteins, 30(1):34—42, 1998.

L. A. Kuhn, M. A. Siani, M. E. Pique, C. L. Fisher, E. D. Getzoff, and J. A. Tainer. The
interdependence of protein surface topography and bound water molecules revealed by
surface accessibility and fractal density measures. J. Mol. Biol., 228:13—22, 1992a.

L. A. Kuhn, M. A. Siani, M. E. Pique, C. L. Fisher, E. D. Getzoff, and J. A. Tainer. The
interdependence of protein surface topography and bound water molecules revealed by
surface accessibility and fractal density measures. J. Mol. Biol., 228:13—22, l992b.

L. A. Kuhn, C. A. Swanson, M. E. Pique, J. A. Tainer, and E. D. Getzoff. Atomic and
residue hydrophilicity in the context of folded protein structures. Proteins Struct. F unct.
Genet, 23:536—547, 1995.

I. D. Kuntz, J. M. Blaney, S. J. Oatley, R. Langridge, and T. E. Ferrin. A geometric approach
to macromolecule-ligand interactions. J. Mol. Biol., 161(2):269—288, 1982.

I. D. Kuntz and W. Kauzmann. Hydration of proteins and polypeptides. Adv. Protein
Chem, 28:239—345, 1974.

J. E. Ladbury. Just add water! The effect of water on the specificity of protein-ligand
binding sites and its potential application to drug design. Chem. Biol., 3:973—980, 1996.

P. Y. Lam, P. K. Jadhav, C. J. Eyermann, C. N. Hodge, Y. Ru, L. T. Bacheler, J. L. Meek,
M. J. Otto, M. M. Rayner, Y. N. Wong, and et al. Rational design of potent, bioavailable,
nonpeptide cyclic ureas as HIV protease inhibitors. Science, 263(5 145):380—3 84, 1994.

A. R. Leach. Ligand docking to proteins with discrete side-chain flexibility. J. Mol. Biol.,
235(1):345—356, 1994.

A. R. Leach and I. D. Kuntz. Conformational analysis of flexible ligands in macromolecular
receptor. J. Comput. Chem, 13 :7 30—748, 1992.

169

 

C. Lemmen, T. Lengauer, and G. Klebe. FLEXS: A method for fast flexible ligand super-
position. J. Med. Chem, 41(23):4502—4520, 1998.

M. Levitt and B. H. Park. Water: Now you see it, now you don’t. Structure, 1:221—226,
1993.

C. A. Lipinski, F. Lombardo, B. W. Dominy, and P. J. F eeney. Experimental and computa-
tional approaches to estimate solubility and permeability in drug discovery and develop-
ment settings. Adv. Drug Deliv. Rev, 46(1-3):3—26, 2001.

S. C. Lovell, J. M. Word, J. S. Richardson, and D. C. Richardson. The penultimate rotamer
library. Proteins, 40(3): 389—408, 2000.

M. Loyevsky and Z. I. Cabantchik. Antimalarial action of hydrophilic drugs: Involvement
of aqueous access routes to intracellular parasites. Mol. Pharmacol., 45(3):446—452,
1994.

M. De Maeyer, J. Desmet, and I. Lasters. All in one: a highly detailed rotamer library irn-
proves both accuracy and speed in the modelling of sidechains by dead-end elimination.
Fold Des., 2(1):53—66, 1997.

J. G. Mandell, V. A. Roberts, M. E. Pique, V. Kotlovyi, J. C. Mitchell, E. Nelson,
1. Tsigelny, and L. F. Ten Eyck. Protein docking using continuum electrostatics and
geometric fit. Protein Eng, l4(2):105—113, 2001.

H. Matter. Selecting optimally diverse compounds from structure databases: a validation
study of two-dimensional and three-dimensional molecular descriptors. J. Med. Chem,
40(8):1219—1229, 1997.

E. C. Meng, B. K. Shoichet, and I. D. Kuntz. Automated docking with grid-based energy
evaluation. J. Comput. Chem, 13(6):505—524, 1992.

J. Mestres, D. C. Rohrer, and G. M. Maggiora. MIMIC: A molecular-field matching pro-
gram. Exploiting appicability of molecular similarity approaches. J. Comput. Chem, 18:
934—954, 1997.

E. Meyer. lntemal water molecules and H-bonding in biological macromolecules: a review
of structural features with functional implications. Protein Sci, 1(12):1543—1562, 1992.

V. Mikol, C. Papageorgiou, and X. Borer. The role of water molecules in the structure-
based design of (5-hydroxynorvaline)-2-cyclosporin: Synthesis, biological activity, and
crystallographic analysis with cyclophilin A. J. Med. Chem, 38(17):3361—3367, 1995.

M. D. Miller, S. K. Kearsley, D. J. Underwood, and R. P. Sheridan. FLOG: A system to
select ’quasi-flexible’ ligands complementary to a receptor of known three—dimensional
structure. J. Comput. Aided. Mal. Des., 8(2):153—l74, 1994.

S. Miller, J. J anin, A. M. Lesk, and C. Chothia. Interior and surface of monomeric proteins.
J. Mol. Biol., l96(3):64l—656, 1987.

170

 

G. W. A. Milne and J. A. Miller. The NCI Drug Information System. 1. system overview.
J. Chem. Inf. Comput. Sci, 26:154—159, 1986.

J. B. Mitchell, R. A. Laskowski, and J. M. Thornton. Non-randomness in side—chain pack-
ing: the distribution of interplanar angles. Proteins, 29(3):370—380, 1997.

J. B. Moon and W. J. Howe. Computer design of bioactive molecules: a method for
receptor-based de novo ligand design. Proteins, 1 l(4):314—328, 1991.

G. M. Morris, D. S. Goodsell, R. S. Halliday, R. Huey, W. E. Hart, R. K. Below, and
A. J. Olson. Automated docking using a Lamarckian genetic algorithm and an empirical
binding fiee energy function. J. Comput. Chem, 19:1639—1662, 1998.

G. M. Morris, D. S. Goodsell, R. Huey, and A. J. Olson. Distributed automated docking
of flexible ligands to proteins: Parallel applications of AutoDock 2.4. J. Comput. Aided.
Mal. Des., 10:293—304, 1996.

J. Moult, T. Hubbard, S. H. Bryant, K. Fidelis, and J. T. Pedersen. Critical assessment of
methods of protein structure prediction (CASP): Round 11. Proteins, Supplement 1:2—6,
1997. http: / /predictioncenter . llnl . gov/ caspZ /Casp2 . html.

1. Miigge and Y. C. Martin. A general and fast scoring fimction for protein-ligand interac-
tions: a simplified potential approach. J. Med. Chem, 42(5):79l—804, 1999.

C. W. Murray, T. R. Auton, and M. D. Eldridge. Empirical scoring functions. 11. The testing
of an empirical scoring function for the prediction of ligand-receptor binding aflinities
and the use of Bayesian regression to improve the quality of the model. J. Comput.

Aided. Mal. Des., 12(5):503—519, 1998.

T. Nakatsu, H. Kato, and J. Oda. Crystal structure of asparagine synthetase reveals a close
evolutionary relationship to class II aminoacyl-tRNA synthetase. Nat. Struct. Biol. , 5(1):
15-19, 1998.

G. Nass, K. Poralla, and H. Zahner. Efl‘ect of the antibiotic Borrelidin on the regulation
of threonine biosynthetic enzymes in E. coli. Biochem. Biophys. Res. Commun, 34(1):
84—91, 1969.

Y. Nishibata and A. Itai. Confirmation of usefulness of a structure construction program
based on three-dimensional receptor structure for rational lead generation. J. Med
Chem, 36(20):2921—2928, 1993.

J. W. M. Nissink, M. L. Verdonk, and G. Klebe. Simple knowledge-based descriptors to
predict protein-ligand interactions. methodology and validation. J. Comput. Aided Mal.
Des., 14(8):787-803, 2000.

A. J. Oakley, M. L. Bello, A. Battistoni, G. Ricci, J. Rossjohn, H. O. Villar, and M. W.
Parker. The structures of human glutathione transferase P l -l in complex with glutathione
and various inhibitors at high resolution. J. Mal. Biol., 274(1):84—8100, 1997. PDB 12gs.

171

 

A. Ogilvie, K. Wiebauer, and W. Kersten. Inhibition of leucyl-transfer ribonucleic acid
synthetasyrnol. Biochem. J., 152(3):511—515, 1975.

T. Onda, Y. Hashirnoto, M. Nagai, H. Kuramochi, S. Saito, H. Yarnazaki, Y. Toya, I. Sakai,
C. J. Homcy, K. Nishikawa, and Y. Ishikawa. Type-specific regulation of adenylyl cy-

clase; Selective pharmacological stimulation and inhibition of adenylyl cyclase isoforms.
J. Biol. Chem, 2001. in press.

C. M. Oshiro, I. D. Kuntz, and J. S. Dixon. Flexible ligand docking using a genetic algo-
rithm. J. Comput. Aided. Mal. Des., 9(2):1 13—130, 1995.

G. Otting, E. Liepinsh, and K. Wiithrich. Protein hydration in aqueous solution. Science,
254:974—980, 1991.

G. Otting and K. Wiithrich. Studies of protein hydration in aqueous solution by direct
NMR observation of individual protein-bound water molecules. J. Am. Chem. Soc., 1 l 1:
1871—1875, 1989.

Z. Otwinowski, R. W. Schevitz, R.-G. Zhang, C. L. Lawson, A. Joachimiak, R. Q. Mar-
morstein, B. F. Luisi, and P. B. Si gler. Crystal structure of trp repressor/operator complex
at atomic resolution. Nature, 335:321—329, 1988.

W. Paetz and G. Nass. Biochemical and immunological characterization of threonyl-tRNA
synthetase of two borrelidirr-resistant mutants of Escherichia coli K12. Eur. J. Biochem. ,
35(2):331—337, 1973.

D. A. Parrish and A. A. Pinkerton. Estradiol methanol hemihydrate. Acta Crystallogr: C,
55:1UC9900100, 1999.

D. A. Pearlrnan and M. A. Murcko. Concepts: New dynamic algorihrn for de nova drug
design. J. Med. Chem, 38(3):466—472, 1995.

E. Perola, K. Xu, T. M. Kollrneyer, S. H. Kaufinann, F. G. Prendergast, and Y. P. Pang. Suc-
cessful virtual screening of a chemical database for famesyltransferase inhibitor leads.
J. Med. Chem, 43(3):401—408, 2000.

J. J. Perona, C. S. Craik, and R. J. Fletterick. Locating the catalytic water molecule in
serine proteases. Science, 26l(5121):620—622, 1993.

N. B. Perry, L. Ettouati, M. Litaudon, J. W. Blunt, M. H. G. Munro, S. Parkin, and H. Hope.
Alkaloids from the antarctic sponge Kirkpatrickia varialasa. Part 1: Variolin B, a new
antitumour and antiviral compound. Tetrahedron, 50:3987—3992, 1994.

W. R. Pitt, J. Murray-Rust, and J. M. Goodfellow. AQUARIUSZ: Knowledge-based mod-
eling of solvent sites around proteins. J. Comput. Chem, 14:1007—1018, 1993.

M. Rarey, B. Kramer, and T. Lengauer. The particle concept: Placing discrete water
molecules during protein-ligand docking predictions. Proteins, 34(1): 17—28, 1999.

172

 

M. Rarey, B. Kramer, T. Lengauer, and G. Klebe. A fast flexible docking method using an
incremental construction algorithm. J. Mol. Biol., 261(3):470—489, 1996a.

M. Rarey, S. Wefing, and T. Lengauer. Placement of medium-sized molecular fiagments
into active sites of proteins. J. Comput. Aided. Mal. Des., 10(1):41-54, 1996b.

A. A. Rashin, M. Iofin, and B. Honig. lntemal cavities and buried waters in globular
proteins. Biochemistry, 25:3619—3625, 1986.

M. L. Raymer, P. C. Sanschagrin, W. F. Punch, S. Venkataraman, E. D. Goodman, and L. A.
Kuhn. Predicting conserved water-mediated and polar ligand interactions in proteins
using a k-nearest-neighbors genetic algorithm. J. Mol. Biol., 265:445—464, 1997.

D. C. Roe and I. D. Kuntz. BUILDER V.2: Improving the chemistry of a de novo design
strategy. J. Comput. Aided. Mal. Des., 9(3):269-282, 1995.

S. H. Rotstein and M. A. Murcko. GenStar: A method for de nova drug design. J. Comput.
Aided. Mal. Des., 7(1):23—43, 1993a.

S. H. Rotstein and M. A. Murcko. GroupBuild: a fiagrnent-based method for de nova drug
design. J. Med. Chem, 36(12):]700—1710, 1993b.

J. A. Rupley and G. Careri. Protein hydration and firnction. Adv. Protein Chem, 41 :37—172,
1991.

J. S. Sack. CHAIN: A crystallographic modeling program. J. Mol. Graph, 2:224—225,
1988.

B. Sandak, R. Nussinov, and H. J. Wolfson. A method for biomolecular structural recog-
nition and docking allowing conformational flexibility. J. Comput. Biol., 5(4):631—654,
1998a.

B. Sandak, H. J. Wolfson, and R. Nussinov. Flexible docking allowing induced fit in pro-
teins: Insights from an open to closed conformational isomers. Proteins Struct. Funct.
Genet, 32:159—174, 1998b.

P. C. Sanschagrin and L. A. Kuhn. Cluster analysis of consensus water sites in thrombin
and trypsin shows conservation between serine proteases and contributions to ligand
specificity. Protein Sci, 7:2054—2064, 1998.

M. Schapira, B. M. Raaka, H. H. Samuels, and R. Abagyan. Rational discovery of novel
nuclear hormone receptor antagonists. Proc. Natl. Acad. Sci. USA, 97(3): 1008—1013,
2000.

C. A. Schiffer, R. Huber, K. Wuthrich, and W. F. van Gunsteren. Simultaneous refinement
of the structure of BPTI against NMR data measured in solution and X-ray diffraction
data measured in single crystals. J. Mol. Biol., 241(4):588-599, 1994.

P. Schimmel, J. Tao, and J. Hill. Aminoacyl tRNA synthetases as targets for new anti-
infectives. FASEB J., 12:1599-1609, 1998.

173

V. Schnecke and L. A. Kuhn. Database screening for HIV protease ligands: the influence
of binding-site conformation and representation on ligand selectivity. Proc. Int. Conf
Intell. Syst. Mol. Biol., pages 242—251, 1999.

V. Schnecke and L. A. Kuhn. Modeling induced fit and controlling molecular diversity
during database screening for ligands. Proteins Struct. F unct. Genet, 2000a. In press.

V. Schnecke and L. A. Kuhn. Virtual screening with salvation and ligand-induced comple-
mentarity. Perspectives in Drug Discovery and Design, 20: 171—190, 2000b.

V. Schnecke, C. A. Swanson, E. D. Getzoff, J. A. Tainer, and L. A. Kuhn. Screening a
peptidyl database for potential ligands to proteins with side-chain flexibility. Proteins
Struct. Funct. Genet, 33:74—87, 1998.

N. E. Shemetulskis, J. r. Dunbar JB, B. W. Dunbar, D. W. Moreland, and C. Humblet.
Enhancing the diversity of a corporate database using chemical database clustering and
analysis. J. Comput. Aided. Mal. Des., 9(5):407—416, 1995.

B. K. Shoichet, D. L. Bodian, and I. D. Kuntz. Molecular docking using shape descriptors.
J. Comput. Chem, 13:380—397, 1992.

B. K. Shoichet and I. D. Kuntz. Matching chemistry and shape in molecular docking.
Protein Eng, 6(7):723—732, 1993.

B. K. Shoichet, A. R. Leach, and I. D. Kuntz. Ligand solvation in molecular docking.
Proteins, 34(1):4—16, 1999.

P. T. Singer, A. Smalas, R. P. Carty, W. F. Mangel, and R. M. Sweet. The hydrolytic
water molecule in trypsin, revealed by time-resolved Laue crystallography. Science, 259
(5095):669—673, 1993.

J. Singh, J. Saldanha, and J. M. Thornton. A novel method for the modelling of peptide
ligands to their receptors. Protein Eng, 4:251—261, 1991.

J. Singh and J. M. Thornton. SIRIUS. An automated method for the analysis of the pre-
ferred packing arrangements between protein groups. J. Mol. Biol., 211(3):595—615,
1990.

V. Sobolev, R. C. Wade, G. Vriend, and M. Edelrnan. Molecular docking using surface
complementarity. Proteins, 25(1):120—129, 1996.

U. Sreenivasan and P. H. Axelsen. Buried water in homologous serine proteases. Biochem-
istry, 31(51):]2785—12791, 1992.

M. Stahl and H. J. Biihm. Development of filter functions for protein-ligand docking. J
Mol. Graph. Model, 16(3):121—132, 1998.

M. Stahl and M. Rarey. Detailed analysis of scoring functions for virtual screening. J. Med.
Chem, 44(7):1035—1042, 2001.

174

 

N. C. Strynadka, M. Eisenstein, E. Katchalski-Katzir, B. K. Shoichet, I. D. Kuntz,
R. Abagyan, M. Totrov, J. Janin, J. Cherfils, F. Zimmerman, A. Olson, B. Duncan,
M. Rao, R. Jackson, M. Stemberg, and M. N. James. Molecular docking programs
successfully predict the binding of a beta-lactamase inhibitory protein to TEM-l beta-
lactamase. Nat. Struct. Biol., 3(3):233—239, 1996.

A. 1. Su, D. M. Lorber, G. S. Weston, W. A. Baase, B. W. Matthews, and B. K. Shoichet.
Docking molecules by families to increase the diversity of hits in database screens: Com-
putational strategy and experimental evaluation. Proteins, 42(2):279—293, 2001.

F. Sugawara, S. Strobel, G. Strobel, R. D. Larsen, D. L. Berglund, G. Gray, N. Takahashi,
S. J. Coval, T. J. Stout, and J. Clardy. The structure and biological activity of cercospo-
rarnide fi'om Cercosparidium henningsii. J. Org. Chem, 56:909—910, 1991.

K. Tanaka, M. Tamaki, and S. Watanabe. Effect of furanomycin on the synthesis of
isoleucyl-tRNA. Biochim. Biophys. Acta, l95(l):244—245, 1969.

D. M. Tanenbaum, Y. Wang, S. P. “Williams, and P. B. Sigler. Crystallographic comparison
of the estrogen and progesterone receptor’s ligand binding domains. Proc. Natl. Acad.
Sci. USA, 95(11):5998—6003, 1998.

M. Totrov and R. Abagyan. Detailed ab irritio prediction of lysozyrne-antibody complex
with 1.6 A accuracy. Nat. Struct. Biol., 1(4):259—263, 1994.

M. Totrov and R. Abagyan. Flexible protein-ligand docking by global energy optimization
in internal coordinates. Proteins Struct. F unct. Genet, Supplement 1:215—220, 1997.

J. Travis. Proteins and organic solvents make an eye—opening mix. Science, 262(5138):
1374, 1993.

Tripos Associates. The SYBYL Software. St. Louis, MO, 2001.
http: / /www. tripos . com.

G. V. Trunk. A problem of dimensionality: A simple example. IEEE Trans. on Pattern
Anal. and Machine Intel., 1:306—307, 1979.

I. A. Vakser and C. Aflalo. Hydrophobic docking: a proposed enhancement to molecular
recognition techniques. Proteins, 20(4):320—329, 1994.

W. F. van Gunsteren, H. J. Berendsen, J. Hermans, W. G. H01, and J. P. Postrna. Computer
simulation of the dynamics of hydrated protein crystals and its comparison with x-ray
data. Proc. Natl. Acad. Sci. USA, 80(14):4315—4319, 1983.

G. M. Verkhivker, D. Bouzida, D. K. Gehlhaar, P. A. Rejto, S. Arthurs, A. B. Colson, S. T.
Freer, V. Larson, B. A. Luty, T. Marrone, and P. W. Rose. Deciphering common failures
in molecular docking of ligand-protein complexes. J. Comput. Aided. Mal. Des., 14(8):
73 1-751, 2000.

175

 

G. M. Verkhivker, P. A. Rejto, D. Bouzida, S. Arthurs, A. B. Colson, S. T. Freer, D. K.
Gehlhaar, V. Larson, B. A. Luty, T. Marrone, and P. W. Rose. Towards understand-
ing the mechanisms of molecular recognition by computer simulations of ligand-protein
interactions. J. Mol. Recognit., 12(6):371—389, 1999.

J. Vrjayalakshmi, K. P. Padmanabhan, K. G. Mann, and A. Tulinsky. The isomorphous
structures of prethrombin2, hirugen-, and PPACK-thrombin: Changes accompanying
activation and exosite binding to thrombin. Protein Sci, 3(12):2254—2271, 1994.

H. O. Villar and L. M. Kauvar. Amino acid preferences at protein binding sites. FEBS
Lett., 349(1):125—130, 1994.

J. H. Voigt, B. Bienfait, S. Wang, and M. C. Nicklaus. Comparison of the NCI open
database with seven large chemical structural databases. J. Chem. Inf Comput. Sci, 41
(3):702—7 12, 2001.

W. P. Walters, M. T. Stahl, and M. A. Murcko. Virtual screening — an overview. Drug
Discov. Today, 3(4):l60—l78, 1998.

H. Wang and A. Ben-Naim. A possible involvement of solvent-induced interactions in drug
design. J. Med. Chem, 39:1531—1539, 1996.

R. Wang and S. Wang. How does consensus scoring work for virtual library screening? an
idealized computer experiment. J. Chem. Inf Comput. Sci, 41(5):]422—1426, 2001.

S. J. Weiner, P. A. Kollmann, D. A. Case, U. C. Singh, C. Ghio, G. Alagona, S. Profeta Jr.,
and P. Weiner. A new forcefield for molecular mechanical simulation of nucleic acids
and proteins. J. Am. Chem. Soc., 106:765—784, 1984.

S. J. Weiner, P. A. Kollmann, D. T. Nguyen, and D. A. Case. An all atom forcefield for
simulations of proteins and nucleic acid. J. Comput. Chem, 7:230—252, 1986.

D. Weininger. SMILES 1. Introduction and encoding rules. J. Chem. Inf Comput. Sci, 28:
31, 1988.

W. Welch, J. Ruppert, and A. N. Jain. Hammerhead: Fast, firlly automated docking of
flexible ligands to protein binding sites. Chem. Biol., 3:449—462, 1996.

R. G. Werner, L. F. Thorpe, W. Reuter, and K. H. Nierhaus. Indolmycin inhibits prokaryotic
tryptophanyl-tRNA ligase. Eur: J. Biochem., 68(1):1—3, 1976.

I. A. VVrlson and D. H. Fremont. Structural analysis of MHC Class I molecules wirh bound
peptide antigen. Semin. Immunol., 5:75-80, 1993.

A. Wlodawer, M. Miller, M. Jaskolski, B. K. Sathyanarayana, E. Baldwin, 1. T. Weber,
L. M. Selk, L. Clawson, J. Schneider, and S. B. Kent. Conserved folding in retroviral
proteases: Crystal structure of a synthetic HIV-1 protease. Science, 245(491 8):616—621,
1989.

176

R. X. Xu, R. P. Meadows, and S. W. F esik. Heteronuclear 3D NMR studies of water bound
to an F K506 binding protein/immunosuppressant complex. Biochemistry, 32(10):2473-—
2480, 1993.

E. Zhang and A. Tulinsky. The molecular environment of the Na+ binding site of thrombin.
Biophys. Chem, 63(2-3):185—200, 1997.

X.-J. Zhang and B. W. Matthews. Conservation of solvent-binding site in 10 crystal forms
of T4 lysozyme. Protein Sci, 3: 103 1—1039, 1994.

177

IIIIIIIIIIIIIIIIIIIII

iiilililliliiilljllill