BOSTON UNIVERSITY

GRADUATE SCHOOL OF ARTS AND SCIENCES

AND

COLLEGE OF ENGINEERING

Dissertation

DEVELOPMENT OF STRUCTURE-BASED COMPUTATIONAL METHODS

FOR PREDICTION AND DESIGN OF PROTEIN-PROTEIN INTERACTIONS

by

BRIAN GREGORY PIERCE

B.S., Duke University, 2000

Submitted in partial fulfillment of the

requirements for the degree of

Doctor of Philosophy

2008 UMI Number: 3298669

INFORMATION TO USERS

The quality of this reproduction is dependent upon the quality of the copy submitted. Broken or indistinct print, colored or poor quality illustrations and photographs, print bleed-through, substandard margins, and improper alignment can adversely affect reproduction. In the unlikely event that the author did not send a complete manuscript and there are missing pages, these will be noted. Also, if unauthorized copyright material had to be removed, a note will indicate the deletion.

® UMI UMI Microform 3298669 Copyright 2008 by ProQuest LLC. All rights reserved. This microform edition is protected against unauthorized copying under Title 17, United States Code.

ProQuest LLC 789 E. Eisenhower Parkway PO Box 1346 Ann Arbor, Ml 48106-1346 Approved by

First Reader "23 Zhiping Weng, Ph.D. Associate Professor of Biomedical Engineering

Second Readei Charles DeLisi,Ph.D. Arthur G. B. Metcalf Professor of Science and Engineering ACKNOWLEDGEMENTS

There are so many people that have provided help and inspiration to me as I've performed this

work, far too many to recount in this text. That being said, I will now acknowledge some of the

most prominent.

First, I would like to thank my wife, Laura, for inspiration, moral support, listening, and

understanding as I worked so diligently on this research. As I grew and became involved in

science, my parents and sister have also provided invaluable support.

Without question, this work could not have been performed without the guidance and expertise of

my advisor, Professor Zhiping Weng. She gave me the opportunity to learn and pursue a variety

of projects, providing insightful comments and criticism when necessary. I would also like to

thank my thesis committee, Professor Charles DeLisi, Professor John Straub, Dr. Enoch Huang,

and Professor Scott Mohr for their time, and for their useful comments and suggestions.

All of the members of Zlab and the BU Bioinformatics Program have provided invaluable

discussions, ideas, and feedback; in particular, Jaafar Haidar, Yong Yu, Kevin Wiehe, Weiwei

Tong, Julian Mintseris, Rong Chen, and Howook Hwang. Yong Yu, Jaafar Haidar, and Bruce

Miller deserve special mention for their hard work on the experiments related to the TCR, CD4,

and LANA proteins.

For computing support and expertise, I would like to thank Mary Ellen Fitzpatrick, and for

administrative support, Caroline Lyman, Jessica Barros, and Robert Henry.

iii I would also like to acknowledge the late Professor Michael Laskowski for gift of the dataset of

mutation energies for the OMTKY system to my lab. Also, Marcia Osburne and Cassidy Dobson for the gift of the pelB-CD4 constructs, the late Prof. Don Wiley for the constructs for HLA-A2,

and Prof. Brian Baker for the A6 TCR constructs.

This work was supported by NSF grants DBI-0078194, DBI-0133834 and DBI-0116574.

iv DEVELOPMENT OF STRUCTURE-BASED COMPUTATIONAL METHODS

FOR PREDICTION AND DESIGN OF PROTEIN-PROTEIN INTERACTIONS

(Order No. )

BRIAN GREGORY PIERCE

Boston University Graduate School of Arts and Sciences and College of Engineering, 2008

Major Professor: Zhiping Weng, Associate Professor of Biomedical Engineering

ABSTRACT

Protein-protein interactions play a key role in the functioning of cells and pathways, and

understanding these interactions on a physical and structural level can help greatly in developing therapeutics for diseases. The large amount of protein structures available presents an immense

opportunity to model and predict protein interactions using computational techniques. Here we describe the development of algorithms to predict protein complex structures (referred to as

protein docking) and to design proteins to improve their interaction affinities. We also present experimental results validating our protein design approach.

The protein docking work we present includes the symmetric multimer docking program M-

ZDOCK as well as ZRANK which rescores docking predictions using a weighted potential. Both programs have been successful when applied to docking benchmarks and in the CAPRI experiment. In addition, we have used the M-ZDOCK program to produce a tetrameric model for a disease-associated protein, the latent nuclear antigen of the Kaposi's sarcoma-associated herpesvirus.

v We have also developed a protein design algorithm to improve the binding between two proteins, given their complex structure. This was applied to a T cell receptor (TCR) to enhance its binding to the Major Histocompatibility Complex and peptide. Several of the point mutations predicted by our algorithm were verified experimentally to bind several times stronger than wild type; we then combined these mutations to produce a TCR with approximately 100-fold affinity improvement.

Further testing of combinations of TCR point mutations has led to striking results regarding the kinetics and cooperativity of the mutations. Finally, we have used our protein design algorithm to predict designability of protein complexes from the Protein Data Bank, and identified the complex between CD4 and HIV gpl20 as a target for future structure-based design efforts.

Preliminary results for this project are given.

vi TABLE OF CONTENTS

List of Tables viii

List of Figures x

List of Abbreviations xii

Chapter 1 Introduction 1

Chapter 2 M-ZDOCK: Incorporating Symmetric Searching 4 into Rigid-Body Docking

Chapter 3 ZRANK: Optimal Reranking of Initial-Stage Protein 27 Docking Predictions

Chapter 4 Refining Protein Docking Predictions Using ZRANK 51 and RosettaDock

Chapter 5 Modeling Protein Interaction Affinity Enhancement, 81 and Application to TCR/peptide/MHC complex

Chapter 6 Studying Cooperativity of TCR Mutations Through 110 Modeling and Experiments

Chapter 7 Exploring the In Silico Designability of Protein 126 Complexes, and Affinity Enhancement of CD4/gpl20

Chapter 8 Summary and Future Directions 143

List of Journal Abbreviations 146

Bibliography 148

Curriculum Vitae 161

vii LIST OF TABLES

2.1 The Unbound Multimer Test Cases 14

2.2 Residues Removed from Multimeric Structures Before Determining Interface 15 Ca Atoms

2.3 M-ZDOCK Results for Quasi-Bound and Bound Test Cases 17

2.4 M-ZDOCK Results for Unbound Test Cases 18

3.1 Results from ZDOCK 2.1 and ZDOCK 2.3 Before and After ZRANK 36

3.2 Charge Terms Used for the ZRANK Candidate Electrostatics Functions 41

4.1 Results for Near-Hit Cases of the ZD3.OZR Set, Before and After Refinement 64

4.2 Number of Cases with Top-Ranked Hits Before and After Refinement 67

4.3 CAPRI Scoring Results Using ZRANK and RosettaDock 69

5.1 Binding Kinetics and Prediction Method for A6 TCR Point Mutants 90

5.2 TCR Point Mutations that Exhibited No Measurable Binding to pepMHC 91

5.3 Weighted ZAFFI Terms, Total Energy Scores, and Measured AAGs for TCR 93 Point mutations

5.4 Contribution of ZAFFI Scoring Terms to Correlation with Measured TCR 95 Point Mutations

5.5 ZAFFI Scores, Filter Scores, and Measured Binding for All TCR Point 96 Mutants

5.6 Binding Kinetics of Combinations of Point Mutants, and Cooperativity 98 of the Energetics

5.7 Binding Kinetics and Specificity of WFGMT TCR Mutant for HLA-A2 with 99 Tax Peptide and V7R Point Mutant of Tax Peptide

6.1 Binding Energy Changes of Combinations of A6 TCR Point Mutants, and 113 Cooperativity

6.2 Inter-Residue Energy Terms for Combinations of Point Mutations on the 115 TCR a Chain

viii 6.3 Association Rates, Dissociation Rates, and Binding Energy Changes for 116 Combinations of Mutations from Different Chains 7.1 In Silico Designability of Nonredundant Transient Enzyme/Inhibitor and 132 Other Complexes

7.2 In Silico Designability of Nonredundant Transient Antibody/Antigen 132 Complexes LIST OF FIGURES

2.1 Diagram of Successive Rotations Through Euler Angles

2.2 The Relative Positions of the Subunits of a C3 Multimer 24

2.3 The Highest-Ranked Hits of M-ZDOCK Superposed onto the Structures 25 of the Complexes

2.4 Structural Model of the Latent Nuclear Antigen (LANA) Carboxy Terminal 26 Domain Tetramer Bound to the LBS-1 and LBS-2 DNA Binding Sites

3.1 Success Rate Versus the Number of Predictions for Reranked Predictions 47

3.2 Score Versus Interface RMSD Plots Using ZDOCK and ZRANK for 48 Several Test Cases

3.3 Success Rate for Reranking Various Numbers of ZDOCK Predictions, and 49 Success Rate comparison for Various Short-Range Electrostatics Formulations

3.4 Success Rate as a Function of the Quality of the Predictions for ZDOCK 2.3 50 and Reranked ZDOCK 2.3

4.1 Hit Success Rate and Hit and Near-hit success rate for ZDOCK 2.3 and 73 ZDOCK 3.0 with and without ZRANK

4.2 Protocol Employed for Docking and Refinement 74

4.3 Histogram of Interface RMSD Change for All Hit and Near-Hit Models After 75 Refinement

4.4 Percent of Models with RMSD Improvement for Several Search/Scoring 76 Strategies

4.5 Percent of Models with Hits After Refinement for Several Search/Scoring 77 Strategies

4.6 Refinement of Three Test Cases, with Rosetta Scores and ZRANK Scores 78 Versus Interface RMSD

4.7 Success Rates of Refinement for Predictions for Hit and Near-Hit 79 Cases for Various Numbers of RosettaDock Refinement Models

4.8 Refined Structure for Test Case 1IQD 80

x 5.1 A AG k^ and A AG k^ versus A AG for 18 Measured TCR Point Mutations 104

5.2 Models of Mutant Complexes for Point Mutations aD26W and aG28T 105

5.3 ZAFFI Scores Versus Experimentally Measured Binding Energy Changes 106 for TCR Point Mutations

5.4 Rosetta Scores Versus Experimentally Measured Binding Energy Changes 107 for TCR Point Mutations

5.5 Biacore SPR Sensorgrams for Binding of Tax/HLA-A2 to Designed 108 and Wild Type TCRs

5.6 Global Fits and Residuals for Wild Type and WFGMT TCRs 109

6.1 Structural Models of WRTMT and WRMMT TCR mutants 121

6.2 SPR Sensorgram for Combination Mutant G28T-phage 122

6.3 Additive AAG Versus Measured AAG for Combinations of TCR Point 123

Mutations

6.4 Association Rate Additivity for Combination TCR Mutations 124

6.5 Dissociation Rate Additivity for Combination TCR Mutations 125 7.1 Residue Transition Frequencies for In Silico Improvement of Binding 139 Affinity

7.2 Predicted Versus Experimentally Measured Binding Energy Changes 140 for CD4 Mutants

7.3 Crystal Structure of the HIV gpl20 Protein Bound to the CD4 Protein 141

7.4 Binding Sensorgrams for CD4 Wild Type and CD4 Mutant H27Y 142 Proteins to YU2 gpl20

xi LIST OF ABBREVIATIONS

3D 3-Dimensional

6D 6-Dimensional

AA Antibody/Anti gen

ACE Atomic Contact Energy

Ca alpha Carbon atom

CAPRI Criticial Assessment of PRedicted Interactions

CHARMM Chemistry at HARvard Molecular Mechanics

CONGEN CONformation GENerator Program

Default Pert RosettaDock with Default Perturbation Size

DFIRE Distance-scaled Finite Ideal Gas REference

DNA Deoxyribonucleic Acid

EBNA-1 Epstein-Barr Nuclear Antigen 1

EI Enzyme/Inhibitor

FFT Fast Fourier Transform

gpl20 HIV Glycoprotein 120

HIV Human Immunodeficiency Virus

HLA Human Leukocyte Antigen

HTLV-1 Human T-cell Lymphotropic Virus 1

HPLC High Pressure Liquid Chromatography

IFACE Interface Atomic Contact Energy

KSHV Kaposi's Sarcoma-Associated Herpesvirus

LANA Latent Nuclear Antigen

xii Large Pert RosettaDock with Large Perturbation Size

LBS Latent Nuclear Antigen Binding Site

MHC Major Histocompatibility Complex

MPI Message Passing Interface

NP Number of Predictions pepMHC Peptide/MHC Complex

PDB Protein Data Bank

RMSD Root Mean Square Distance (or Deviation)

RU Response Units

SCOP Structural Classification of Proteins

SPR Surface Plasmon Resonance

TCR T Cell Receptor vdW van der Waals

ZD2.3ZR ZDOCK 2.3 followed by ZRANK

ZD3.0 ZDOCK3.0

ZD3.0ZR ZDOCK 3.0 followed by ZRANK

Xlll 1

Chapter 1

Introduction

This thesis describes the research that I conducted under the guidance of Professor Zhiping Weng for my doctorate research over a period of 4 1/2 years, and falls into the broad category of computational protein modeling and design. The organization of this thesis is chronological, in that the chapters are arranged in the order of when that work was performed (although there were some overlaps in the times of conducting these projects). In addition, it happens that the research

(and thus the organization of the chapters) is in the order of the steps that a researcher could theoretically utilize in a project regarding a structurally uncharacterized protein-protein interaction: modeling the interaction with initial-stage docking (Chapter 2), refining this interaction model (Chapters 3 and 4), and finally designing the protein-protein interaction based on the structural model (Chapters 5,6, and 7). Though we have not in fact performed all of these steps for one particular protein-protein interaction system (we performed these to some extent for the Kaposi's sarcoma-associated herpesvirus LANA protein, described in Chapter 2), we feel that this is indeed possible and hope that the research community benefits from some, or all, of the methods developed and optimized during this thesis work.

In Chapter 2, the development, testing, and application of an initial-stage docking program for

Cn symmetric multimers, M-ZDOCK, is discussed. This is a modified version of the hM-based docking program ZDOCK [4], so that it performs a full rigid-body search in symmetric space.

The study includes testing on a benchmark of four test cases of proteins that have been crystallized both as monomers and symmetric multimers. We also discuss the application of this 2 program to model a disease-related protein, the latent nuclear antigen from Kaposi's sarcoma- associated herpesvirus.

In Chapter 3, we describe the development of a program to rerank initial-stage docking predictions, ZRANK. Motivated by the success of ZDOCK to produce hits within a large set of predictions but not to rank them well, ZRANK rescores the ZDOCK predictions using an optimized potential to improve the rank of the hits. This is verified using ZDOCK predictions from protein-protein docking Benchmark 2.0 [6].

Having developed and tested ZRANK to rerank predictions, we then applied this program to refine the structures from initial-stage docking in Chapter 4. By combining the scoring of

ZRANK with the protein modeling program RosettaDock [7] for moving the side chains and repositioning the structures, we achieved significant structural improvement when refining protein-protein docking benchmark cases, as well as in the double-blind protein docking experiment, CAPRI [8]. The success rate was further improved after optimizing the ZRANK function for scoring refined structures rather than rigid-body predictions.

In Chapter 5, we discuss the design of a T cell receptor to improve its binding to the peptide-

MHC complex using a computational design algorithm based on ZRANK. The algorithm we developed used the mutagenesis module of Rosetta [9] to model the structures and a weighted energy function (based on ZRANK) to score the structures. This algorithm, named ZAFFI, predicted several TCR mutations that were found to improve binding, and when combining these point mutations we had nearly a 100-fold peptide-MHC binding affinity improvement. 3

In Chapter 6 the TCR mutations are studied in more detail, with more thorough testing of combinations of point mutations, and also modeling of some of these combinations of mutations.

This analysis reveals some striking trends regarding the additivity of multiple mutations, and a theory as to why certain combinations of mutations did not have the degree of energetic additivity that was expected.

Chapter 7 details the application of the ZAFFI protein design function to model a large set of protein-protein interface structures. This study yields some interesting results regarding the in silico mutational tendencies for amino acids and interfaces, and identifies targets for experimental protein interface design. We then selected one of these targets, the CD4 protein of the CD4-HIV gpl20 complex, to attempt to improve its gpl20 affinity via point mutagenesis. Experimental work is underway; we present the initial experimental results from this exciting study.

The dynamics and structures of protein-protein interactions are extraordinarily complex, yet essential to understand if we are to utilize protein structural information to understand and cure diseases. The following thesis presents computational models that have predictive capabilities to determine how structures interact and how to improve their interaction, and these studies, if not the programs themselves, should be useful in furthering our research in computational structural biology. 4

Chapter 2

M-ZDOCK: Incorporating Symmetric Searching into Rigid-Body Docking

ABSTRACT

Computational protein docking is a useful technique for gaining insights into protein interactions.

We have developed an algorithm for predicting the structure of cyclically symmetric (Cn) multimers based on the structure of an unbound (or partially bound) monomer. Using a grid- based Fast Fourier Transform approach, a space of exclusively symmetric multimers is searched for the best structure. This leads to improvements both in accuracy and running time over the alternative, which is to run a binary docking program and filter the results for near-symmetry.

The accuracy is improved because fewer false positives are considered in the search, thus hits are not as easily overlooked. By searching four instead of six degrees of freedom, the required amount of computations is reduced. This program has been tested on several known multimer complexes from the Protein Data Bank, including four unbound multimers: three trimers and a pentamer. For all of these cases, M-ZDOCK was able to find at least one hit, whereas only two of the four test cases had hits when using ZDOCK and a symmetry filter. We additionally describe the usage of M-ZDOCK to predict the tetrameric structure of a structurally uncharacterized protein, the latent nuclear antigen of the Kaposi's sarcoma-associated herpesvirus. M-ZDOCK is freely available to academic users at: http://zlab.bu.edu/M-ZDOCK/.

INTRODUCTION

An important subclass of interactions between proteins is the case where two or more identical proteins interact to form a homomultimer. A common form of symmetry found in homomultimers 5

is Cn symmetry or cyclic symmetry, which is a ring-shaped complex. For symmetric dimers, trimers, pentamers, and heptamers, this symmetry is necessarily the case, while this symmetry is also found for other numbers of protein subunits. For instance, membrane channels and chaperonins often have oligomers with Cn symmetry.

To efficiently and accurately predict Cn multimer complexes, we have implemented a program called Multimer ZDOCK, or M-ZDOCK. This program takes advantage of the properties of Cn symmetry to perform a simplified search for the correct complex.

There are many instances where this program can be applied. A number of proteins have been solved as monomers or in a complex with another protein but exist in a homomultimeric state under different conditions in vivo (e.g. heat shock, pH changes, viral fusion).

The recently solved crystal structure of adeno-associated virus Rep40 provides evidence that it oligomerizes for nucleotide binding, possibly as a Cn hexamer [10]. Using M-ZDOCK with this monomelic crystal structure as input, the structure of the hexamer can be modeled. Another example is the protein Chaperonin-60 (Cpn60), which is expressed under heat shock and other forms of stress, is a homologue of E. coli GroEL and is typically found in a double ring structure composed of 14 protomers. However it has been found that Mycobacterium tuberculosis has lower order oligomers of this protein [11]; a ring-shaped model for this structure can be obtained by multimer docking. Also, Korkhov et al. have devised a model for the dimeric structure of

GAB A transporter-1 (GAT1) [12]. By computationally predicting possible structures of dimeric

GAT1, multimer docking would help to support this model or provide new ones regarding the structure. 6

Since the interface between two adjacent subunits is the same for all interfaces of the complex, only one out of the n interfaces needs to be considered, reducing the problem to two monomers for any degree of Cn symmetry. In addition, since all Cn multimers can be aligned in a plane (as they are rotated around a single axis), one spatial degree of freedom can be ignored. Finally, since there is redundancy when rotating a Cn complex around its rotational axis (the resultant complex will be the same), this rotational degree of freedom is eliminated. Thus the problem becomes 4 dimensional instead of 6 dimensional; this reduces the amount of searching and the computational time.

Another type of symmetry seen in proteins is D2 (dihedral) symmetry, which is composed of two homodimers interacting symmetrically, or a dimer of dimers (four asymmetric units). From an interaction standpoint, this case differs from Cn symmetry in that there are two interfaces to predict rather than n identical interfaces as is seen in Cn symmetry. Recently Berchanski and

Eisenstein filtered and combined the pairwise complexes between monomers generated with a

FFT-based generic docking algorithm [13] to predict the structure of D2 multimers [14]. They tested the subunits taken directly from the complex structure, as well as homology modeled monomers, and reported promising results. A similar approach was used earlier to construct the helically symmetric protein coat of the tobacco mosaic virus [15]. However, due to the discrete nature of the FFT algorithm, the vast majority of these binary predictions are not symmetric, and many of the ones that pass the filter are not truly symmetric.

Here we have developed a new docking algorithm M-ZDOCK that only explores the search space conforming to the Cn symmetry. We observe a significant improvement in accuracy, lower redundancy, and fewer false positives, as shown in a direct comparison with docking and 7 filtering. In addition, since only perfectly symmetric multimers are explored in the search space, less computational time is required.

In the process of developing M-ZDOCK, we have carefully curated a set of test cases that exist in both monomelic and multimeric forms in physiological conditions. Although small, this set represents an exhaustive search of such test cases in the Protein Data Bank (PDB) [16]. It should prove useful for future docking studies on multimers.

We have also applied the M-ZDOCK program to predict the quaternary structure of a viral protein, the latent nuclear antigen (LANA) of the Kaposi's sarcoma-associated herpesvirus

(KSHV). As the name indicates, KSHV is known to cause cancer, in particular in individuals that have suppressed immune function, such as those with HIV [17]. The LANA protein performs a wide variety of functions, including tethering the viral DNA to the host chromosome by forming an oligomer and binding the terminal repeat sequence of the viral DNA [18]. Using M-ZDOCK, we predicted the two putative dimer interfaces for the homology-modeled monomer, and assembled a tetrameric model. The resultant structural prediction is presented in light of known information regarding this complex. 8

METHODS

Scoring Function

The scoring function used by this program is based on the scoring used in the latest version of

ZDOCK [4]. ZDOCK is an initial-stage docking algorithm designed to predict the structure of the complex of two proteins, referred to as the receptor and the ligand. It takes into account surface complementarity, electrostatics, and desolvation to find the optimal fit between two proteins.

Surface complementarity is calculated using pairwise shape complementarity (PSC), which consists of a favorable term determined by the number of atom pairs within a distance cutoff, and a penalty term determined by the number of clashes. ACE, or Atomic Contact Energy [19], is used to score desolvation, and the electrostatic term is calculated by applying Coulomb's equation to the partial charges of the ligand in the electrostatic field of the receptor.

The search strategy of ZDOCK is to discretize both ligand and receptor onto a grid, and use an

FFT to determine the best position of the ligand relative to the receptor. This discretization and

FFT is performed for a complete set of angular orientations of the ligand (relative to a fixed receptor).

Results have demonstrated that this approach performs well against a docking benchmark and in the international double-blind docking experiment CAPRI [20], Critical Assessment of PRedicted

Interactions. 9

Euler Angles

The Euler angle conventions used here refer to these successive rotations from the initial configuration, used in Goldstein's Classical Mechanics [21]:

1. Rotation by $ around the z-axis.

2. Rotation by 8 around the original x-axis.

See Figure 2.1 for a diagram of these rotations. Typically Euler angles are sets of three angles; in this case, the third angle, op, is not necessary as it would be redundant in the symmetric search

(see the Search Space section for an explanation).

Rotational/Translational Search

M-ZDOCK uses the convention that the rotational axis will be parallel to the z-axis, and searches in the x-y plane for the optimal position of this axis. To perform the search for the best conformation of a multimer based on the structure of a monomer, it has been necessary to make modifications to the search methodology that is used for ZDOCK 2.3. The new search algorithm is outlined below:

1. Center the receptor (the input monomer) at the origin.

2. Rotate the receptor by an angle <|> around the z-axis, and then 0 around the x-axis.

3. Copy the receptor, and rotate it by 360/n degrees around the z-axis to create the ligand.

4. Discretize both the ligand and receptor, with a grid spacing of 1.2 Angstroms (the same as

ZDOCK 2.3). 10

5. Perform the 3D FFT and correlation, and search in the x-y plane for the best scoring multimer position for that rotational orientation.

6. Repeat steps 2-5 for various other sets of § and 8.

Search Space

In order to fully explore the space of multimers, it is necessary to vary $ from 0 to 360 degrees, and 6 from 0 to 90 degrees. 8 does not need to sample a full 360 degrees because for a given

<|) there are redundancies at 180 - 8,180 + 8, and -8 due to the symmetric nature of these angles around the z and x axes.

It is not necessary to sample ip angles (a third rotation around the z-axis), as these are symmetric around the z-axis and therefore would be redundant for the same values of <|> and 8. This corresponds to the loss of a rotational degree of freedom that is referred to in the Introduction section.

M-ZDOCK uses a set of 1500 angles; this was found to be a good balance between computational time and predictive performance. In addition, given that ZDOCK 2.3 uses 54000 angles for 3 degrees of angular freedom (6 degree sampling), the number of angles that M-ZDOCK covers is mathematically reasonable as it is approximately 5400020. 11

Reconstructing the Multimer

Based on the optimal relative position of two adjacent monomers in the x-y plane (output from the FFT), it is possible to reconstruct the full multimer. The only constraint is that the monomers need to be rotated by 360/n degrees with respect to one another around the z-axis. Referring to

Figure 2.2, the vector representing the displacement between the two adjacent monomers is L and the vector from the monomer to the symmetry axis (in the x-y plane) is d. b is the angle around the Cn symmetry axis between two multimer centers of mass, 360/n. The angle between the vectors L and d is a, computed by (180 - P)/2. The magnitude of d can be computed as

L/(2*cos(a)).

Once the rotational axis is found, the monomer needs to be rotated around this axis n times by beta degrees to form the multimer. Thus, given the vector between two adjacent monomers in the

Cn multimer (and the symmetry number), it is possible to reconstruct the entire multimer. To illustrate this concept, a Java applet has been written and is publicly available at http://zlab.bu.edu/M-ZDOCK.

Symmetry Filter

In order to compare the results of M-ZDOCK with results from an existing method of docking, we implemented a symmetry filter that will choose only near-symmetric complexes. It is designed to process the results from a docking tool such as ZDOCK which produces many predictions

(54000 in the case of ZDOCK with dense sampling). 12

The filter determines the angle and axis between the monomers of the prediction, as well as the center of mass translation between the monomers. For perfect symmetry, the angle between the center of mass translation and the axis is 90 degrees, and the angle of rotation around the axis is

360/n, but a certain range must be allowed as the predictions are not perfectly symmetric. In the case of Berchanski et al. [14] the angular range for the rotation around the axis was ±6°, and the between the axis and translation the angular range was ±3°. To allow for a comparison with the

M-ZDOCK results so there would be approximately 1500 predictions per test case this range was increased to ±18° and ±9°, respectively.

Multimer Test Cases

We tested M-ZDOCK with two categories of test cases, bound/quasi-bound and unbound.

Quasi-bound and Bound Test Cases

To ensure that the search space is covered entirely and that the algorithm is valid for various types of Cn symmetry, both bound and quasi-bound docking test cases were used. The bound test cases were generated by extracting the monomer from the multimeric structure so that the docking algorithm can attempt to reassemble the multimer. These test cases should be relatively simple to dock as there is no conformational change to account for. If the correct structure is not found with these cases, then there is perhaps some problem with the searching algorithm. Though found in the PDB as both monomers and multimers, quasi-bound test cases are most likely biologically multimers. Therefore the conformational change involved is of little or no significance, making these cases similar (but slightly more difficult than) the bound test cases. 13

The monomer structure that is found in the PDB is used as input to the docking algorithm, while the multimer structure in the PDB is used to evaluate the docking results.

Unbound Test Cases

The second type of test cases is unbound structures. These test cases are significantly more difficult to predict, both due to the conformational change of the proteins inherent in unbound docking, and because of the low affinity of the complexes, as these cases must coexist in both monomer and multimer forms to be found and verified experimentally.

Four proteins were found in the Protein Data Bank [16] (PDB) for which different symmetric forms exist, according to Protein Quaternary Structure server classification. Here is a brief summary of these proteins:

1. RNase A. Bovine pancreatic RNase A was crystallized in monomeric [22] and trimeric [23] forms. The trimer in this case is one of two trimeric forms of thought to exist in mildly acidic solutions Rnase A. Notable about this structure is a domain-swapped C-terminal beta strand.

2. Phospholipase A2. The Naja naja naja (Indian cobra) phospholipase A2 (PLA2) was obtained from the venom and crystallized in trimeric [24] form using random crystallization screening. The monomeric version [25] is the Naja naja atra (Chinese cobra) PLA2, which was crystallized with a

2+ 2+ lower concentration of PLA2 and higher concentration of Ca (the Ca is seen in the structure of the monomer but not the trimer). In Segelke et al. it is discussed that the trimeric form may be a means of shielding the active site and thus "protecting the snake from its own venom". 14

3. Havivirus Envelope Protein. This is the fusion envelope protein of the tick borne encephalitis virus (TBEV E protein). The input structure is taken from the homodimer structure [26]. The trimeric form, which occurs at low pH during membrane fusion, was recently solved [27].

4. Bovine Trypsin Inhibitor. This test case is the bovine pancreatic trypsin inhibitor (BPTI), which occurs as a monomer [28] at basic pH and a decamer [29] at acidic pH levels. As the decamer is comprised of two C5 symmetric pentamers, the target for this case is one half of the decamer.

Table 2.1 summarizes these test cases. To provide a measure of the difficulty of docking each complex, interface Ca atoms from unbound monomers were fitted to two adjacent subunits of the complex. As M-ZDOCK is a rigid-body docking algorithm, the RMSD in this case can be seen as the lower limit for the RMSD of the predictions.

Table 2.1. The unbound multimer test cases.

Test Case PDB Ids" Symmetry RMSD" RNase A 9RAT/1JS0 Trimer 033

Phospholipase A2 (PLA2) 1POA/1A3F Trimer 0.79 Havivirus Envelope Protein (TBEV E) 1SVB/10ML Trimer 2.08 Bovine Trypsin Inhibitor (BPTI) 3PTI/1B0C Pentamer 0.41 The first PDB code is for the structure used as input for docking, while the second one is the bound multimer. b Interface Ca RMSD change between unbound/bound structures. 15

RMSD Calculations and Hits

To evaluate bound and unbound predictions, the RMSDs of interface alpha Carbon (Ca) atoms were used. The interface Ca atoms were determined from the crystal structure of the multimer. If any atom of a residue is within 10 Angstroms of any atom of another chain, the Ca atom from that residue is determined to be an interface Ca. In addition, to avoid false negatives due to large domain movements, regions of residues with large movement from unbound to bound (> 4 A) were removed before determining interface Ca atoms. These residues are given in Table 2.2.

Table 2.2. Residues removed from multimeric structures before determining interface Ca atoms.

Test case Residues Removed RNase A 113-124 Phospholipase A2 none Flavivirus Envelope Protein 2-18,189-193,294-401 Bovine Trypsin Inhibitor none

Once the Ca residues are known, two adjacent subunits of the predicted structure are fitted to two adjacent subunits of the complex using the interface Ca's, and the root mean square deviation

(RMSD) between the interface Ca's is computed. Hits are defined as predictions that have an interface Ca RMSD s 2.5 Angstroms.

Structure Modeling of LANA Tetramer Bound to DNA

The tertiary structure of latent nuclear antigen (LANA) residues 868 through 960 was modeled with the 3D-Jigsaw modeling tool [30], using the EBNA-1 crystal structure as template [31].

LANA residues 929-939 did not have defined coordinates after 3D-Jigsaw and were modeled 16 using Modloop [32]. LANA residue numbering follows the numbering in Garber et al. [5]

(GenBankAAK50002).

To predict the quaternary structure of LANA, M-ZDOCK was run using the LANA homology model to perform a full search of possible homodimeric interfaces. The output models from M-

ZDOCK were then filtered based on similarity to the EBNA-1 dimer (for the dimer interface), ability fit to double-stranded DNA (for both the dimer and dimer-dimer interfaces), and score of the model from ZRANK [33]. Two M-ZDOCK models were selected using these criteria, and were joined to construct the model of the tetramer.

The DNA structure was taken from the structure of EBNA-1 bound to DNA [34] and fit to the two dimers in the LANA tetramer model. The linking DNA between the two dimer binding sites was produced by extending the existing DNA strands. Rosetta [35] was then used to mutate the

DNA sequence to the LBS-1 and LBS-2 sequences and repack the LANA side chains.

RESULTS

Structure Prediction: Quasi-bound and Bound

Eight quasi-bound and bound test cases were used to ensure the coverage and basic functionality of M-ZDOCK for a variety of symmetries. The results in Table 2.3 clearly demonstrate that M-

ZDOCK is capable of predicting structures with Cn symmetry. For all of the structures the number one ranked prediction was a hit, and in addition there were a number of hits in the top 20 for every test case. 17

Table 23. M-ZDOCK results for quasi-bound and bound test cases.

Test Case" Symmetry Hits" RMSDd References Quasi-bound Test Cases 1NSP/1B996 Trimer 11 1.06 [36,37] 1KKU/1GZU6 Trimer 7 0.88 [38,39] 1AUS/1AA16 Tetramer 8 0.93 [40] 1EXB/1QRQ Tetramer 7 1.15 [41,42] Bound Test Cases 1AF6 Trimer 6 0.73 [43] lA8Re Pentamer 10 0.78 [44] 1QNU Pentamer 16 0.75 [45] 1G31 Heptamer 17 1.81 [46] a PDB IDs of the test cases, with the PDB ID of the input structure for M-ZDOCK listed first for the quasi-bound test cases. b Number of hits in the top 20 (out of 1500) predictions, as ranked by M-ZDOCK. cRank of the first hit. d RMSD (in A) of the first hit. e The bound structures in these cases are in fact dimers of the Cn multimer; just the Cn contacts are predicted so the other interface is ignored.

Structure Prediction: Unbound

The structure prediction capabilities of M-ZDOCK are shown to be superior to filtering normal docking predictions, across the unbound multimer benchmark (Table 2.4). For M-ZDOCK, all of the first hits are in the top third of the predictions, whereas for filtering there were two cases where no hit was found, and the two other cases were in the bottom third of the predictions. 18

Table 2.4. M-ZDOCK results for unbound test cases.

M-ZDOCK ZDOCK +filtering Test case a 0 d e c d ND Hits" Rank RMSD ND Hits" Rank RMSD Rnase A 1500 1 476 2.44 1875 0 PLA2 1500 6 33 1.11 1595 1 1417 2.05 TBEVE 1500 2 62 2.31 1476 0 BPTI 1500 20 384 2.25 1164 1 1064 1.83 "Number o ' predictions produced by M-ZDOCK (the number is always 1500). b Number of hits among the predictions. c Rank of the first hit. d RMSD (in A) of the first hit. e Number of predictions remaining after running ZDOCK and filtering the 54000 predictions for symmetry.

M-ZDOCK successfully predicted a hit for RNase A (Figure 2.3a), while the near-symmetric predictions of ZDOCK failed to produce a hit. This is despite the fact that 375 more predictions were produced by ZDOCK plus filtering. This complex was difficult to predict due to the strand swapping that takes place upon multimerization, which explains the relatively high rank of 476 for the first M-ZDOCK hit. Although the swapped strands are not included in the RMSD calculation, they are clearly part of the interface making the prediction non-trivial. The swapped strands are highlighted in Figure 2.3a.

The symmetric trimer PLA2 was successfully predicted by M-ZDOCK (Figure 2.3b). In this case

M-ZDOCK predicted 6 hits, one of them with the particularly high rank of 33. While ZDOCK followed by filtering obtained a hit, the rank of the hit is 1417 and the RMSD of this hit is higher.

Perhaps the most striking results are for the TBEV E protein, where two hits were found by M-

ZDOCK, while no hits were found with ZDOCK. This protein is somewhat difficult to dock due to the large C-terminal conformational change upon trimerization that helps to stabilize the interaction. The difficulty is also reflected in the lower bound for the RMSD of 2.08 A (Table 19

2.1), which leaves little room for error for rigid-body docking to obtain a hit (under 2.5 A).

However M-ZDOCK is able to predict this structure, giving the first-ranked hit an impressive rank of 62 (see Figure 2.3c for the structure).

The BPTI pentamer (Figure 2.3d) had a large number of predictions produced by M-ZDOCK. As with the other test cases, M-ZDOCK performed better with regard to hits and the rank of the first hit. In this case the RMSD of the first hit was slightly better for the ZDOCK prediction. But of the 20 M-ZDOCK hits, 6 of them had a better rank and RMSD than the top ZDOCK prediction, so clearly M-ZDOCK is superior in this case as well.

Prediction of LANA Complex Structure

Having tested M-ZDOCK against a benchmark set of multimers, we used M-ZDOCK to predict a currently unknown structure, that of the Kaposi's sarcoma-associated herpesvirus latent nuclear antigen (LANA) protein. As described in more detail in the Methods, we modeled the monomer based on the EBNA-1 protein structure, and produced possible dimer interfaces for this model using M-ZDOCK. By joining predictions for two dimer interfaces, we produced a model for tetrameric LANA protein. This model is shown bound to DNA in Figure 2.4.

The LANA model in Figure 2.4 indicates some interesting features of this protein complex, some of which can be verified against known experimental results. The teterameric structure features two dimer interfaces, and a dimer-dimer interface. The dimer interface resembles the EBNA-1 structure unbound and bound to DNA [31, 34], which is not surprising as this was the source of the monomer for the homology model. However, the protein docking helped to confirm that with 20 the LANA residues would still permit this mode of dimerization to be favorable. Studies have shown that interactions are not necessarily similar when sequence identities fall below 30-40%

[47], and this LANA domain has only a 20% sequence similarity with EBNA-1. In addition to the dimer interface, the docking predicted a dimer-dimer interface, which allows a curved DNA fragment, based on the curved DNA in the bound EBNA-1 structure, to fit under the LANA protein tetramer.

Such a DNA interface is likely for several reasons. First, footprinting analysis of the LANA binding site identified two adjacent near-repeats (referred to as LANA Binding Site, or LBS, 1 and 2) with spacing of 22 base pairs between the centers [48]. 22 base pairs are between the binding site centers in the model, showing that the DNA sequence from footprinting fits into this

DNA-bound model. Furthermore, another study measured the bend in the DNA bound by LANA and found it to be 57° for LBS-1, and 110° for both LBS-1 and LBS-2, similar to the bending found for the EBNA-1 binding sites [49]. We have indicated the three helices of the LANA model in Figure 2.4, along with a selected portion of the DNA binding site. The three DNA bases highlighted in Figure 2.4, which contact Helix 2 extensively, were found to be crucial for LANA binding in a mutagenesis study of LBS-1 [3]. Another study that performed alanine scanning of the LANA protein found that several of the residues in Helix 2 are critical for DNA binding [50].

Computational Performance

The benchmark docking predictions reported in this study were performed on an IBM p690 workstation with 32 1.3 Ghz Power4 processors, using MPI for parallelization. Due to the increased efficiency of the M-ZDOCK search, a significant savings in running time can be seen 21 using this approach. The discretization of the receptor at every angle set (as described in the

Methods section), which costs more than regular ZDOCK, is more than compensated by the faster search. On average, M-ZDOCK runs 30-40% faster than ZDOCK.

M-ZDOCK was also compiled and run on Linux in serial and parallel, and on Mac OS X in serial.

Versions of M-ZDOCK for all of these platforms are available at: http://zlab.bu.edu/M-ZDOCK.

DISCUSSION

A possible future modification to the M-ZDOCK algorithm would be to incorporate the degree of packing into the algorithm. Since the algorithm currently used considers only the interface between adjacent subunits, interactions between non-adjacent subunits is ignored. This would possibly be an issue in the case of a multimer that has a structure similar to the spokes of a wheel, i.e. tightly packed (versus a "daisy chain" that is shown in Figure 2.2). However these non- neighbor interactions would clearly be less significant overall than the interactions taking place in the interface between adjacent subunits.

In summary, there are several advantages to using M-ZDOCK versus filtering docking predictions from a normal binary docking program:

1. Greater accuracy, with improvements in both hit count and the rank of the first hit.

2. Increased efficiency, due to a reduced search space based on the knowledge of Cn

symmetry.

3. Perfectly symmetric multimers are automatically output; there is no need for

approximation/fitting to generate the other subunits. 22

We have shown that it is possible to perform an intelligent search of the space of exclusively symmetric Cn multimers, and have incorporated this into the M-ZDOCK program. Based on its performance, with regards to both accuracy and speed, M-ZDOCK is an effective program for predicting complexes of Cn symmetry based on the structure of its subunit, showing clear superiority over traditional docking and subsequent filtering.

In addition to testing against a multimer benchmark, we have applied this program to predict the structure of a complex for which no structure currently exists, the LANA carboxy-terminal domain tetramer. We then compared the result with various experimental studies in the literature regarding this protein, and found that the model corroborates several characteristics of the complex obtained from these studies. This indicates the utility of this algorithm in modeling the structures of multimers that are of interest to the research community. 23

Figure 2.1. Diagram of successive rotations through Euler angles ty and 9, the angles used to describe the rotational configuration of the ligand and receptor. In this case, 4> = 90° and 8 = 45°. 24

Figure 2.2. The relative positions of the subunits of a C3 multimer. The vector L is the relative position between the receptor and the ligand (which is the receptor rotated by p degrees; in this case p = 120°). The magnitude of vector d to the axis of symmetry and the angle a between vectors L and d can be determined algebraically. Thus, the interface between the ligand and receptor is evaluated by M-ZDOCK, and then the rest of the multimer (in this case the subunit represented by the dashed lines) can be generated. 25

Figure 23. The highest ranked hits of M-ZDOCK superposed onto the structures of the complexes, with each predicted and actual chain colored separately. Images generated with Pymol [1]. (a) The interface region of RNase A, with arrows indicating the b- strands that swap in the trimeric structure. Predicted: magenta, green, blue; Actual: purple, pink, yellow (b) Phospholipase A2. Predicted: purple, pink, yellow; Actual: magenta, green, blue (c) TBEV E Protein, truncated at the highly mobile C-terminal domain. Predicted: purple, pink, yellow; Actual: magenta, green, blue (d) Bovine Pancreatic Trypsin Inhibitor. Predicted: magenta, yellow, blue, red, grey; Actual: slate, green, aqua, peach, orange. 26

Figure 2.4. Structural model of the latent nuclear antigen (LANA) carboxy terminal domain tetramer bound to the LBS-1 and LBS-2 DNA binding sites. DNA is colored green and cyan, and three LBS-1 bases found to ablate LANA binding when mutated [3] are colored yellow. LANA Helix 1 (residues 870-882) is blue; Helix 2 (residues 907-917) is red, and Helix 3 (residues 950-966) is magenta. Protein residue numbering follows the numbering in Garber et al. [5] (GenBank AAK50002). Figure generated using Pymol [1]. 27

Chapter 3

ZRANK: Optimal Reranking of Initial-Stage Protein Docking Predictions

ABSTRACT

Protein-protein docking requires fast and effective methods to quickly discriminate correct from incorrect predictions generated by initial-stage docking. We have developed and tested a scoring function that utilizes detailed electrostatics, van der Waals and desolvation to rescore initial-stage docking predictions. Weights for the scoring terms were optimized for a set of test cases, and this optimized function was then tested on an independent set of nonredundant cases. This program, named ZRANK, is shown to significantly improve the success rate over the initial ZDOCK rankings across a large benchmark. The amount of test cases with No. 1 ranked hits increased from 2 to 11 and from 6 to 12 when predictions from two ZDOCK versions were considered.

ZRANK can be applied either as a refinement protocol in itself or as a preprocessing stage to enrich the well-ranked hits prior to further refinement.

INTRODUCTION

Protein-protein interactions are an essential part of many biological processes. Understanding the mode of interaction between two proteins is important for identifying drug targets and optimizing or eliminating protein-protein interactions via site-directed mutagenesis. While there are many crystallized structures of protein complexes available in the Protein Data Bank [16], it has become increasingly useful to complement these known structures with in silico predictions of protein interactions using protein-protein docking. 28

Several reviews have been published on protein-protein docking, including those by Smith and

Sternberg [51], and Halperin et al. [52]. An overview of protein-protein interactions can be found in Jones and Thornton [53]. There have been many exciting developments in protein-protein docking algorithms, but it is still generally performed in two sequential stages, due to the complexity of the problem. The initial stage, which treats proteins as rigid bodies and generates many predictions (10,000 or more) is followed by the refinement stage, which performs any combination of detailed scoring, energy minimization, side chain searches and clustering on these predictions.

A variety of approaches have been used in initial-stage docking, with respect to both searching and scoring. The programs FTDOCK [54], GRAMM [55], and ZDOCK [4] all use grid-based spatial searches that are sped up with a Fast Fourier Transform (FFT), a method first applied by

Katchalski-Katzir et al. [13] The program HEX [56] also utilizes the FFT, but in this case it is to speed up a rotational search using spherical harmonics. Other approaches for initial-stage docking searches include Monte Carlo based searching [7, 57] and geometric hashing [58]. The scoring functions for initial-stage docking all employ some measure of shape complementarity, and they generally include electrostatics and desolvation as well; more about initial-stage docking scoring can be found in the reviews mentioned above.

Docking refinement has also made progress with the recent development of several algorithms.

Examples include the web server ClusPro [59] that performs clustering of rigid body docking predictions from DOT [55, 60], GRAMM [55], and ZDOCK [4]. The program RosettaDock [7] performs docking refinement on its predictions using a Monte Carlo based approach, optimizing 29 side chain positions and rigid body position. It employs an energy function that includes terms for van der Waals, hydrogen bonding, electrostatics, pair potential, and desolvation. The refinement program MultiDock uses side chain rotamers and rigid body minimization to relax the interfaces of rigid body docking predictions [61].

The initial-stage docking program ZDOCK has been proven effective both against a docking benchmark [4] and in several rounds of the CAPRI experiment [20, 62]. ZDOCK is a grid-based docking algorithm that performs a systematic search in 6D and typically can output 3600 or

54,000 predictions, depending on the sampling density in the rotational space (15° or 6° sampling, respectively). It has been shown that 6° sampling yields more near-native predictions in the top

2000, and the refinement of these predictions can improve their ranking [4].

The refinement program RDOCK was shown to be successful in refining ZDOCK predictions

[63]. It uses the CHARMM program [64] to perform energy minimization on the top ZDOCK predictions (the top 2000 is recommended), and reranks these predictions using desolvation and electrostatics. While it is successful, there are some limitations to RDOCK. The energy minimization step takes roughly one minute per test case and therefore RDOCK is only feasible for a limited subset of ZDOCK predictions. Also, following the recommended usage, RDOCK success is limited by the number of near-native structures produced by ZDOCK in the top 2000.

To accommodate this, we have developed the program ZRANK (Zlab Rerank) that quickly and accurately reranks the rigid body docking results from ZDOCK. It uses a more detailed potential than ZDOCK but is fast enough to quickly process and rerank the 54,000 predictions that are produced by the ZDOCK 6° sampling search. It significantly improves the success rate of 30

ZDOCK when tested against the newly released protein-protein docking Benchmark 2.0 [6].

Thus, it can be used to rerank predictions on its own, or else can be used as a preprocessing step for refinement programs such as RDOCK.

METHODS

Scoring Function

ZRANK utilizes a scoring function that can be quickly computed and effectively employed to discriminate hits from non-hits. The scoring function is a linear weighted sum of van der Waals attractive and repulsive energies, electrostatics short and long-range attractive and repulsive energies, and desolvation:

jcore = wvdw _at,vdW _a + wvdw_rhvdW_r + welec _sraEelec _sra +w E + w E + w F elec _ srr elec _ srr elec _ Ira elec _lra elec _ Irr elec _ Irr + WdsEds

Here is a more detailed explanation of the energy components: van der Waals: This is calculated using the Lennard-Jones 6-12 potential, which is calculated for all atoms i and j for inter-atom distance r^ < 8.0 A:

I \12 y Evdw(iJ) = £ij -2 r r.. ^\ v I \ y / 31

The coefficients for the well depth en and width Oy are from the CHARMM 19 polar hydrogen potential [65]. For distances ry < 0.6o-y, a repulsive linearization is used, as in Gray et al.[7]

Electrostatics: For electrostatics interactions, the Coulomb equation is used, with a 1/r distance dependent dielectric:

Eekc (i,j) = 332^-

For short-range electrostatics (ry < 5.0 A), partial charges from the CHARMM 19 polar hydrogen potential [65] are used. The minimum distance used in the calculation is the width of the van der

Waals energy well between the atom pair, Ey, to avoid extraordinarily large values due to clash.

For electrostatics interaction at distances greater than 5.0 A, only fully charged side chain atoms are used, with charges assigned as in Gray et al. [7]. The long-range electrostatics (> 5.0 A) is represented by a separate term, so that short-range interactions (namely hydrogen bonding) can be weighted separately.

Desolvation: Pairwise Atomic Contact Energy [66] (ACE) is used to calculate the desolvation energy:

Eds(iJ) = atj

The term ay is the ACE term for the atom pair of types i an j; this is zero if the atoms are greater than 6.0 A apart and for hydrogens. 32

There are several major differences that distinguish this scoring function from that of

RosettaDock. One is that the short-range electrostatics is determined using polar hydrogen partial charges. This allows for the hydrogen bonding and polar forces to be calculated within the electrostatics. In RosettaDock, the hydrogen bonding potential is in a separate term (only full charges are in the electrostatics terms). Another notable difference is the use of pairwise ACE in the desolvation calculation. Also our simplified scoring function has many fewer terms and thus can be rapidly evaluated for a large number of structures.

This scoring function also differs from that of RDOCK. While they both use CHARMM 19 and pairwise ACE, ZRANK breaks down the electrostatics terms into components and applies weights to them separately. Additionally, the van der Waals energy is included in the scoring function rather than just being a filter. Finally, RDOCK computes its final scores only after several rounds of CHARMM minimization, whereas for ZRANK only the input structure is scored, allowing for faster evaluation.

ZDOCK Predictions

To generate initial rigid-body docking predictions, ZDOCK was used with 6° rotational sampling and randomized start positions. At this sampling density, ZDOCK generates 54,000 predictions by keeping the best translational solution for each rotational angle set. For the Antibody/Antigen test cases, the search was restricted to the CDR portions of the antibodies.

For the test set used in this study, two different versions of ZDOCK were used: ZDOCK 2.1 and

ZDOCK 2.3. ZDOCK 2.1 uses Pairwise Shape Complementarity [67] (PSC) to score predictions. 33

ZDOCK 2.3 [4] uses a combination of PSC, electrostatics and desolvation in its scoring function; this has been shown to improve performance against a docking benchmark.

Evaluation of Docking Predictions

The predictions from ZDOCK were evaluated by calculating the Root Mean Square Deviation

(RMSD) between the bound and unbound interface alpha carbon (Ca) atoms, as previously described [4]. Hits are predictions with interface Ca RMSD less than 2.5 A.

Addition of Hydrogen Atoms

The Benchmark 2.0 PDB files do not have hydrogens, so it was necessary to add polar hydrogens to the structures prior to rescoring. To accomplish this, the Rosetta program [7] was used to add hydrogens to the receptor and ligand PDBs separately. These structures were then superposed onto the ZDOCK predictions for these cases. We used CHARMM to add the hydrogens to two cases that were too large to be imported into Rosetta.

Training Weights Using Benchmark 1.0 Test Cases

An important aspect of this scoring function is a set of weights that highlight the relative values of the terms of the scoring function. This was particularly crucial due to the fact that the predictions being reranked were from rigid-bodydockin g for which a certain degree of softness in the scoring function was essential. 34

Optimal weights for the terms were obtained by training the function with a subset of Benchmark

1.0 cases [68]. Benchmark 2.0 cases [6] were used for testing. To ensure independent training and testing sets, all Benchmark 1.0 cases that were homologous with Benchmark 2.0 cases, through

SCOP superfamily (non-antibody antigen cases) and manual inspection (antibody antigen cases), were filtered out. After this filtering, 15 cases remained for training; all of these were then docked using ZDOCK 2.3, generating 54,000 predictions per test case. All predictions for these cases were then classified as hits or non-hits, and scored using the seven terms described above.

After obtaining and scoring the ZDOCK predictions, a downhill simplex minimization algorithm

[69] was used to determine the optimal set of weights. The downhill simplex algorithm allows for quick minimization in a multidimensional space. Briefly described, for N dimensional minimization it selects N points around a starting point (by taking a given offset from the point in each dimension) and successively moves in the direction of the lowest valued point. In addition to its speed and ease of use, the simplex was an attractive option due to its lack of need for taking derivatives, and in the case of our training criterion (rank of first hit) the function is not necessarily smooth.

The downhill simplex performed minimization in 7 dimensions (corresponding the each of the 7 energy terms), and simultaneously optimized the ranks of the first hits of the 14 cases (one of the

15 nonredundant cases did not have any hits from ZDOCK). The target function was the sum of the top ranking hit for each case, and all ranks over 50 were set to 51, to avoid over-optimizing poor rankings (changing this cutoff did not significantly affect the resultant weights). To avoid local minima, five random restarts were used when the minimum was found by the simplex (as outlined in Press et al. [69]), by reinitializing the simplex at each minimum (taking N points 35 around the minimum, as described above). Finally, the simplex was run using 12 separate times, using different starting positions. The weights corresponding to the top set of rankings from all of the simplex runs were kept. Performing the minimization using different random seeds led to little difference in the resultant weights.

The weights that were obtained and then applied to reranking the Benchmark 2.0 complexes are as follows:

van der Waals attractive: 1.0 van der Waals repulsive: 0.009

Electrostatics short-range attractive: 0.31

Electrostatics short-range repulsive: 0.34

Electrostatics long-range attractive: 0.44

Electrostatics long-range repulsive: 0.50

Desolvation: 1.02

Using Weighted Function to Rescore Predictions

The scoring function with weights optimized using ZDOCK predictions for Benchmark 1.0 was applied for reranking of ZDOCK predictions for Benchmark 2.0. For the test set, all cases that had ZDOCK hits for Benchmark 2.0 were used, amounting to 62 cases (out of 76 rigid and medium difficulty cases). The scoring function was applied to all 54,000 predictions obtained from the 6° sampling for each case, and the reranked predictions were then analyzed for the ranks of the hits. 36

RESULTS

Reranking Benchmark 2.0 Predictions

The performance results of ZRANK on Benchmark 2.0 test cases are presented in Table 3.1.

There is a clear improvement in the number of highly ranked cases for both ZDOCK versions.

Although the weights were trained on ZDOCK 2.3 predictions (which uses a different scoring function than ZDOCK 2.1), it is encouraging to see improvements in performance for predictions from the latter as well.

Table 3.1. Results from ZDOCK 2.1 and ZDOCK 2.3 before and after reranking. ZDOCK 2.1 ZDOCK 2.1 ZR ZDOCK 2.3 ZDOCK 2.3 ZR Test case" Hitsb Rankc Hits" Rankc Hits" Rankc Hits" Rank0 RMSD" Rigid Body Enzyme-Inhibitor 1AVX 0 2863 44 7 7 449 40 9 2.17 1AY7- 0 5584 24 111 0 11358 12 134 2.20 1BVN 13 502 33 14 52 23 48 14 1.79 1CGI 9 145 11 23 0 2423 3 34 2.33 1D6R 0 2951 2 984 0 3538 2 1312 2.30 1DFJ 40 9 16 1 73 1 18 1 1.89 1E6E 0 22643 6 3 18 103 67 4 1.51 1EAW 62 3 37 62 108 13 22 77 1.65 1EWY 2 259 34 65 11 113 28 129 2.38 1EZU 3 1100 0 4323 0 5267 0 20347 1.44 1F34 13 5 8 69 16 26 4 160 1.42 1MAH 9 92 73 1 71 1 85 1 0.89 1PPE 218 1 217 1 324 1 193 1 1.13 1TMQ 11 314 3 342 21 126 5 389 1.90 1UDI 4 258 2 673 7 13 0 4259 2.19 2MTA 0 — 0 — 0 35227 1 1722 2.37 2PCC 0 — 0 — 0 12916 1 1037 1.45 2SIC 24 173 69 1 45 39 55 1 1.15 2SNI 0 17906 0 2868 0 5079 6 300 1.99 7CEI 24 106 115 1 186 1 157 1 1.09 Other 37

1AKJ 0 4872 5 425 28 96 6 40 2.37 1B6C 2 1717 12 1 6 168 18 1 2.26 1BUH 0 14556 7 475 0 22962 4 903 1.65 1E96 0 3094 7 16 2 399 16 16 2.02 1F51 4 230 10 2 17 11 25 3 1.76 1FQJ 0 9889 0 5551 0 17028 0 6057 2.15 1GCQ 0 24339 1 429 0 23148 2 762 2.29 1HE1 0 4672 2 349 1 1146 2 258 1.82 1KAC 0 2896 3 72 1 1523 2 94 2.42 1KTZ 0 53599 0 4251 0 10395 7 804 1.28 1KXP 1 1734 14 2 33 7 20 1 1.99 1ML0 21 36 50 1 60 1 56 1 1.43 1QA9 0 5672 2 850 0 13606 1 1502 1.59 1RLB 0 — 0 — 2 302 38 4 2.37 1SBB 0 — 0 — 0 17342 0 6089 1.32 2BTF 0 — 0 — 0 13881 7 96 2.17 Antibody -Antigen 1AHW 21 268 24 8 46 56 24 10 0.98 1BVK 0 3970 0 9314 0 5390 0 10315 2.41 1DQJ 0 2287 0 7904 0 10257 0 34026 2.26 1E6J 34 15 121 1 77 19 115 2 1.91 UPS 9 171 18 53 19 181 20 71 1.07 1MLC 12 110 39 5 33 7 38 5 1.19 1VFB 0 2734 1 60 0 4969 1 104 1.79 1WEJ 8 465 40 2 39 102 55 1 1.01 2VIS 0 2747 22 37 1 369 17 60 1.83 Antibody -Antigen: UBB 1BJ1 49 129 61 2 67 18 52 4 1.05 1FSK 105 1 89 1 163 1 78 1 1.14 1I9R 41 50 16 20 37 90 14 24 1.44 1IQD 5 612 29 73 21 99 33 98 1.16 1K4C 0 20806 3 359 1 1575 34 345 1.47 1KXQ 13 212 28 1 13 301 27 1 1.51 1NCA 47 14 28 5 50 4 17 166 0.86 1NSN 5 185 1 923 1 445 1 1775 1.81 1QFW 7 257 10 94 23 74 11 61 1.20 2JEL 33 45 6 632 23 86 2 975 1.81 2QFW 3 832 41 1 18 179 63 1 1.49 Medium Other 1GRN 2 1704 3 1661 6 807 1 1523 2.26 1HE8 0 — 0 — 0 47386 0 11558 2.36 1I2M 0 — 0 — 0 34162 0 34492 2.43 38

1IJK 0 52731 0 8688 0 6357 9 205 1.52 1K5D 0 - 0 - 0 6012 8 134 2.29 1WQ1 2 1101 10 6 5 27 0 2899 1.52

"Only Benchmark 2.0 test cases with ZDOCK hits in the top 54,000 predictions are listed. Test cases in bold had a hit in the top 20 predictions for the reranked ZDOCK 2.3 predictions. 'The number of hits in the top 2000 predictions. cRank of the first hit."—" denotes that no hits were produced. "Interface Ca RMSD of the first hit.

The count of No. 1 hits for ZDOCK 2.3 reranking increases from 6 to 12 test cases, and for

ZDOCK 2.1 it increases from 2 to 11. In addition for ZDOCK 2.3 reranking there are 22 cases that have hits in the top 20 (in bold in Table 3.1).

ZRANK made notable improvements for many cases. The test case 1RLB (Transthyretin/Retinol binding protein) had a significant improvement in both the number of hits and the rank of the first hit over ZDOCK 2.3 (ZDOCK 2.1 had no hits for this test case). Interestingly, for the medium difficulty test case 1WQ1 (Ras GTPase/Ras GAP), ZRANK improved the ranking of the ZDOCK

2.1 predictions (with a reranked hit in the top 10 predictions), however such improvement was not seen for ZDOCK 2.3. This is most likely due to the conformational change between the unbound and bound structures leading to unfavorable energies due to the particular arrangements between the two proteins of the hits among the ZDOCK 2.3 predictions.

Success Rate

To provide a clearer overall view of the performance of the original and reranked ZDOCK results, success rate plots are shown in Figure 3.1. The plot for All Test Cases reiterates the points mentioned above. Both success rates for ZRANK show significant improvements, with the 39 greatest improvement seen for ZDOCK 2.1 reranking, and the greatest overall success rate for the

ZDOCK 2.3 reranking.

In addition, the success rates are shown for each individual category of cases. While improvements can be seen for every category, the best improvements can be seen for the Other category, which contains the most heterogeneous, the most difficult, and arguably the most biologically interesting test cases of the three categories.

Score versus RMSD plots

While ranking of the top hit and the number of hits are very important statistics, it is also useful to examine the score versus RMSD plots. If the scoring function is accurate enough, one would expect to see an "energy funnel". The score/RMSD plots for three cases (one Enzyme Inhibitor, one Antibody Antigen and one Other) that improved with ZRANK are shown in Figure 3.2. For all three cases, there is an energy funnel after ZRANK, which did not exist or was not as prominent as for ZDOCK. This supports the improved accuracy of the more detailed scoring function of ZRANK over ZDOCK alone.

Reranking Different Numbers of Predictions

While ZRANK has been discussed in terms of reranking 54,000 predictions, it was also examined to see if the performance changes when reranking the top 5000, 10,000, and 20,000 predictions from ZDOCK 2.3 (Figure 3.3a). It can be seen that the success rate improves moderately when reranking more predictions; this is encouraging, as the increased number of false positives subject 40

to reranking does not negatively affect the success. In addition, when NP > 100, the success rate curves diverge and it is clear that the more predictions subject to reranking, the better the success rate becomes.

Examination of the Scoring Potential

As electrostatics is a sensitive term, we tested four formulations of the short-range electrostatics; these are summarized in Table 3.2. All of the scoring functions shown (apart from ZDOCK and

PolH vdW) employed a minimum short-range electrostatics distance of 3.0 A. The ZDOCK

Charges function used the same (non-hydrogen) partial charges as in ZDOCK 2.3. The Full

Charges function used only fully charged side chain residues, mimicking the treatment of the long-range electrostatics. The PolH 3.0 used the CHARMM 19 polar hydrogen partial charges with a 3.0 distance cutoff. The PolH vdW function is the one selected in this paper, and is described in detail in the Methods.

Table 3.2. Charge terms used for the electrostatics functions.

Potential Name Partial Charge Source Minimum Distance Cutoff PolH vdW CHARMM 19 vdW Minimum PolH 3.0 CHARMM 19 3.0 A Full Charges RosettaDock Electrostatics 3.0 A ZDOCK Charges ZDOCK 2.3 (no hydrogens) 3.0 A 41

For each formulation, all seven terms were optimized using Benchmark 1.0 predictions as described in the Methods section. The success rate results are shown in Figure 3.3b. The Full

Charges reranking performed approximately the same as ZDOCK 2.3, while the ZDOCK Charges and PolH 3.0 each performed successively better. In fact the PolH 3.0 function behaved approximately the same in terms of success after the top 30 predictions. As the polar hydrogens used by this scoring function represent an inclusion of more detailed electrostatics (namely hydrogen bonding), this shows that such a term is helpful in discriminating hits. However the

PolH 3.0 function did not have as many No. 1 hits as the PolH vdW function given in this paper, thus showing that the electrostatics cutoff based on the vdW minima may help to reduce noise that is introduced by having a constant distance for a minimum cutoff.

Success Rate versus the Quality of Input Predictions

While there is much improvement in the success rate given by the reranking potential, a few cases were reranked less effectively. We hypothesized that the quality of the input docking predictions would affect the outcome of the reranking. To examine this, we binned the test cases according to the lowest RMSD of the ZDOCK predictions in the top 2000. The success rate for each set of cases was then evaluated for ZDOCK 2.3 and the reranking, as shown in Figure 3.4. Success rate is computed as the percentage of test cases in the given bin with a hit in the top 20 predictions. As expected, the success rate declines as the quality of the predictions decreases, for both the reranking and ZDOCK. The reranking has significantly greater success than ZDOCK in all the bins, and the improvement is even greater for lower-quality predictions. 42

Computational Time

One important aspect of ZRANK is that it is possible to rerank all 54,000 predictions from

ZDOCK, given its relatively low computational cost. It performs efficiently because the hydrogens only need to be generated once for the initial structure and simple energy calculations are performed on each rigid-body prediction. Reranking all 54,000 predictions takes an average of 180 minutes on an Intel Pentium III 2.0 GHz machine, or about 5 predictions per second. This process can be easily parallelized using the Message Passing Interface (MPI).

DISCUSSION

We have implemented and tested an algorithm, ZRANK, for quickly and effectively reranking rigid-body docking predictions. It has been shown to increase the success rate across a benchmark set of cases from the initial ZDOCK rankings, and dramatic improvement in the success rate and hit count was seen for several cases. We have also explored variants of this scoring scheme and amount of docking predictions to be reranked (Figures 3.3a and 3.3b), to determine how the ranking success on a benchmark set of cases changes according to different protocols.

One notable feature of ZRANK is that the success rates after reranking ZDOCK 2.3 and ZDOCK

2.1 predictions are quite similar (Figure 3.1) when considering all test cases, although ZDOCK

2.3 alone significantly outperforms ZDOCK 2.1. This demonstrates that the improvement for reranking ZDOCK 2.1 is much greater. Nonetheless, there is a slight edge for the reranked

ZDOCK 2.3 predictions, most likely due to the higher initial hit counts from ZDOCK 2.3. This is 43 in contrast to the performance from RDOCK, where the overall success was higher for reranking the ZDOCK 2.1 predictions [63].

Furthermore, the Other category of cases seems to have improved most using ZRANK from the initial ZDOCK ranking. This may be because the PSC function (used in ZDOCK 2.1 and 2.3), which is effective in discriminating the large pockets of binding sites of enzyme/inhibitor complexes [67], has slightly less success in scoring the non enzyme/inhibitor interfaces. Thus the van der Waals based shape complementarity of the reranking function could provide an advantage in evaluating these cases, which represent a larger percentage of the test cases in

Benchmark 2.0.

While the basic energy function used by ZRANK is simple, the weights allow for the high degree of discrimination between hits and non-hits within a set of rigid-body docking predictions. The van der Waals repulsive term is significantly smaller than the van der Waals attractive term; this allows for the softness required when evaluating the rigid body predictions (in addition to the short range linearization used for the repulsive term). Thus, hits with some degree of clash (but compensated by favorable energies in the other terms) can be ranked well.

Another notable part of the weights is that the repulsive electrostatics terms are greater than the attractive terms. This may help to filter out predictions with unfavorable (repulsive) electrostatics; electrostatic repulsion could preclude the formation of the encounter complex between the proteins in that configuration. In addition, a recent study has noted that there is an "asymmetric screening" of charged spheres in the presence of a dielectric [70], wherein the repulsive force is seen to be stronger than the attractive force. This supports the greater weight of the repulsive 44 electrostatics terms obtained for ZRANK, and the effectiveness of the weighted scoring function when reranking. The long-range electrostatic terms are larger than their short-range counterparts, this may be due to the necessity of long-range electrostatics to form an encounter complex, and it may also be due to more noise in the short-range terms (for example, due to side chain movements upon binding).

The ZRANK algorithm was developed to explore the limits of scoring rigid-body predictions, to see if the energy terms from initial-stage docking predictions can allow hits to be discriminated and well-ranked. Recent studies have indicated that certain portions of interfaces in protein- protein interactions do not change significantly between bound and unbound conformations.

Camacho et al. have found that there are "anchor residues" and corresponding recognition motifs on the binding partner that are very important in the binding process [71]. Another study performed molecular dynamics simulations on a set of protein-protein interfaces, and found that many interfaces have a core region that is less mobile than the periphery of the interface [72]. It is therefore possible that through careful examination of the rigid-body docking predictions these important characteristics can be discriminated, without using side chain and/or backbone movement of the predictions.

As with any scoring program, the success of ZRANK is limited by the quality of the input. This is illustrated to some extent in Figure 3.4, with regard to the RMSDs of the predictions from

ZDOCK. In addition, conformational change upon binding (which is essentially the lower bound of the rigid-body prediction RMSD) of the docking cases limits the success of the initial-stage searching; only about half of the medium difficulty cases in Benchmark 2.0 had any hits from

ZDOCK. However, it can be seen both in the benchmark and in the CAPRI docking experiment 45 that many proteins exhibit small conformational change upon binding and are thus within the reach of rigid-body docking. As for the remaining cases, work is being performed to modify

ZDOCK to allow for better predictions of cases with conformational change. This includes addition of a statistical potential recently developed using a non-redundant set of transient protein interfaces [73].

While the scoring scheme presented in this paper has been developed and tested for reranking protein docking predictions, it is possible that other uses can be employed. For instance, the scores of a set of predictions can be analyzed in terms of whether there is any near-native prediction at all. Preliminary results indicate that this would require additional information such as binding energy of the proteins and a classification tool (e.g. a Support Vector Machine), but such a method, if effective, would certainly be useful for analyzing protein docking output. In addition to this, the ZRANK scoring can also be combined with clustering to help to eliminate redundancy and false positive predictions. One group has recently used ZDOCK scores to help determine the AAG's from mutations at a protein-DNA interface [74]; it is possible that ZRANK scores can be used in this context as well.

The performance of ZRANK indicates that it is possible to greatly increase the ranking results of a set of rigid-body docking predictions without structural refinement. Thus the number of well- ranked hits can be enriched after initial-stage docking, prior to further refinement and/or analysis of the predictions. This is particularly useful before running detailed refinement programs, which can take hours to refine single predictions. Recent research in protein-protein docking includes explicitly modeling flexible protein backbones [75]; as refinement using these methods is computationally intensive and would be limited in the amount of predictions that can be 46 accommodated, this underscores the need for retaining near-native predictions from initial-stage docking among the top ranks.

Future work includes combining ZRANK with side chain repacking and/or energy minimization of predictions. It can be applied in an iterative manner to rerank the predictions before structural refinement (to increase the number of near-native structures to be refined) and after the structural refinement to perform a final reranking. This will allow for even greater success in producing accurate structural models of protein complexes. 47

AH Test Cases Enzyme/Inhibitor TT 90% T- "1 ZDOCK 2.3 ...... ' ' 80% - ZDOCK 2.3 + ZRANK —•— ZDOCK 2.1 70% ..*... _—«*" "ZDOCK 2.1+ZRANK —•— --•— **£,**• 7* 60% - 7*~"jr -*•" --•*' 50% *j/t - "-4 * / - 40% t ~ Jr ' t 30% -_--*-^*r * 0 1^"*^ M"^ * - 20% 1 -»—«--*•'--•-•-»—•e' .*' .a' _ 10% - .-•*"' - 0% <>-*" , , , ill - 10 100 1000

Number of Predictions (Np) Antibody/Antigen Other 90% 1 1—I- I I Ml| 1 1 1 I I I I ZDOCK 2.3 --•— 80% ZDOCK 2.3 + ZRANK —•— ZDOCK 2.1 --€>--• 70% ZDOCK 2.1 + ZRANK —•— 60% 50% 40% 30% 20% 10% 0% 10 100 10 100 1000

Number of Predictions (Np) Number of Predictions (Np)

Figure 3.1. Success Rate versus the number of predictions for reranked predictions and the original predictions from ZDOCK 2.1 and 2.3. Only Benchmark 2.0 test cases with ZDOCK hits in the 54,000 predictions are considered in the success rate. All 54,000 predictions were reranked for each test case. 48

1B6C 2QFW 2SIC KM W- •

9 • • 10 15 20 25 30 10 15 20 25 30 5 10 15 20 Interlace RMSD, A Interlace RMSD, A Interface RMSD, A

10 15 20 10 15 20 25 Interlace RMSD, A Interlace RMSD. A

Figure 3.2. Score versus interface RMSD plots for the top 2000 ZDOCK 2.3 (top) and reranked (bottom) predictions for several test cases. The ZDOCK score is negated in order to facilitate comparison between the plots. 49

Reranking Different Numbers of ZDOCK Predictions

- 90% , 1 r • i i i i | -T i i ...... t 54000 Predictions • 20000 Predictions --»-• 80% 10000Predictions ••••• 5000 Predictions •"•»»•• Original ZDOCK Rank -••-• 70% ^^^ mf m

60% h-

B S 50% _ 18 X ,#, # § 40% - «>* to *'

30% - - .# 20% X

10% ^.-.-.-.-.-.«.—•

. . 1 . • 0% 10 100 1000

Number of Predictions (Np) Comparison of Different Scoring Methods

90% •'•'I — '••i PolH vdW "*" '"'... , PolH 3.0 --»-- 80% Full Charges ..«.. —•> ZDOCK Charges ,.,».„ Original ZDOCK Rank »••>• *>.'*x^ 70% - & < ^ *& 60% -• **;*••• " /^ ,#•:$ 50% - V - ^ *••i * •• .« 40%

30%

20% »^— 1.. T>' - i ::::>- ,..*•; 10%

0% . 1 ... § 10 100 1000

Number of Predictions (Np)

Figure 33. a) Success Rate plots for reranking various numbers of ZDOCK predictions: 5000,10,000,20,000, and all 54,000 predictions. b) Success Rate comparison for various short-range electrostatics formulations. PolH vdW is the potential described in this paper, PolH 3.0 uses a constant 3.0 A cutoff for minimum electrostatics distance, Full Charges uses this cutoff but with only fully charged side chains and no partial charges, and ZD Charges uses partial charges with no polar hydrogens (with partial charge terms from ZDOCK 2.3). 50

Success Rate vs. Prediction Quality

80%

70%

,

2 - hj

8 * * of rfk

- it

10% -

no*, . 0.5A-1.0A 1.0A-1.5A 1.5A-2.0A 2.0 A-2.5 A Prediction RMSO

Figure 3.4. Success Rate as a function of the quality of the predictions for ZDOCK 2.3 and reranked ZDOCK 2.3. Success is defined as a hit in the top 20 for that test case. Test cases are binned according to the lowest interface RMSD for of the predictions in the first 2000 ZDOCK predictions. No test cases had a hit below 0.5 A in the top 2000 predictions. 51

Chapter 4

Refining Protein Docking Predictions Using ZRANK and RosettaDock

ABSTRACT

To determine the structures of protein-protein interactions, protein docking is a valuable tool that

complements experimental methods to characterize protein complexes. While protein docking

can often produce a near-native solution within a set of global docking predictions, there are

sometimes predictions that require refinement to elucidate correct contacts and conformation.

Previously, we developed the ZRANK algorithm to rerank initial docking predictions from

ZDOCK, a docking program developed by our lab. In this study, we have applied the ZRANK

algorithm toward refinement of protein docking models, in conjunction with the protein docking

program RosettaDock. This was performed by reranking global docking predictions from

ZDOCK, performing local side chain and rigid-body refinement using RosettaDock, and selecting

the refined model based on ZRANK score. For comparison, we examined using RosettaDock

score instead of ZRANK score, and a larger perturbation size for the RosettaDock search, and

determined that the larger RosettaDock perturbation size with ZRANK scoring was optimal. This

method was validated on a protein-protein docking benchmark. For refining docking benchmark

predictions from the newest ZDOCK version, this led to improved structures of top-ranked hits in

20 of 27 cases, and an increase from 23 to 27 cases with hits in the top 20 predictions. Finally, we

optimized the ZRANK energy function using refined models, which provides a significant

improvement over the original ZRANK energy function. Using this optimized function and the

refinement protocol, the numbers of cases with hits ranked at number one increased from 12 to 19

and from 7 to 15 for two different ZDOCK versions. This shows the effective combination of 52 independently developed docking protocols (ZDOCK/ZRANK, and RosettaDock), indicating that using diverse search and scoring functions can improve protein docking results.

INTRODUCTION

Protein-protein interactions are key to the functioning of all cells and many biological processes.

To understand the mechanism of a protein-protein interaction, the structure of a protein complex is essential. While many high-resolution (x-ray) structures of protein complexes are available in the Protein Data Bank (PDB [16]), a vast number of protein complex structures are not yet determined. Meanwhile, structural genomics projects are underway [76], producing new structures of proteins, many of them monomeric. With the crystal structures (or modeled structures) of the component monomers, protein-protein docking (referred to as protein docking for brevity) can be used to predict the structures of the protein complex when no protein complex structure is available. Recent developments in protein docking allow for atomic-scale protein complex predictions [77], yet work needs to be done to refine these methods so that they can be quickly and reliably applied to unknown protein complexes.

Many protein docking algorithms are divided into several steps: the initial global search and subsequent steps to improve these initial predictions [51]. The global search is a full search of the orientations of the two proteins, typically keeping the larger protein (referred to as the receptor) fixed, while moving the smaller protein (the ligand). This is often a rigid-body search in 6 dimensions, utilizing a Fast Fourier Transform (FFT) for efficiency and softness for small overlaps [13,54,78], but other methods such as Monte Carlo with side chain searching have also been successful [7, 79]. The following steps can include clustering [59, 80] reranking,[81] and 53 structural refinement [61] of the initial set of predictions. Structural refinement is useful in that it can improve the contacts and the accuracy of initial predictions that are close to the correct conformation but also have room for improvement.

Previously we have implemented several algorithms for initial-stage docking and refinement:

ZDOCK, RDOCK, and ZRANK. The program ZDOCK performs a grid-based docking search using Fast Fourier Transform (FFT), and its scoring includes desolvation, electrostatics, and a novel shape complementarity function [4]. It has performed consistently among the top algorithms during the CAPRI docking experiment [62]; using ZDOCK to perform docking led to

5 of 6 recent targets with at least one prediction rated Acceptable or higher [82] (the highest number among all participants). ZDOCK was also found to compare favorably to other FFT- based docking algorithms in a recent study on clustering initial-stage docking predictions [83].

While ZDOCK produces many near-native predictions (hits), they are often not ranked in the top

10. To improve the rank of the hits, RDOCK performs docking refinement by reranking the top

2000 ZDOCK predictions using energy minimization followed by scoring using electrostatics and desolvation [63]. Although RDOCK has been shown to improve the success rate of ZDOCK predictions, it lacks the ability to quickly process all 54,000 predictions from a ZDOCK run.

To account for this, we developed the ZRANK program; it uses a weighted energy function with van der Waals, electrostatics and desolvation terms to quickly and effectively rerank the ZDOCK predictions without energy minimization [33]. It was tested on protein docking Benchmark 2.0

[6], using predictions from two versions of ZDOCK: ZDOCK 2.1 (which employs shape complementarity alone) and ZDOCK 2.3 (which employs shape complementarity, desolvation, and electrostatics). In both cases there was significant improvement in docking performance 54 when using ZRANK to rescore the rigid-body predictions; the number of cases with top-ranked hits increased from 2 to 11 for ZDOCK 2.1 and from 6 to 12 for ZDOCK 2.3.

It was noted that ZRANK could be followed with structural refinement to further improve the docking success rate [33]. To examine this possibility, we have combined the initial-stage docking of ZDOCK and scoring of ZRANK with the structural refinement of RosettaDock [7].

The local refinement of RosettaDock includes side chain repacking and a Monte Carlo search of the local rigid-body space of the ligand. While RosettaDock can be highly successful in obtaining atomically accurate models through its refinement, it is sometimes unsuccessful in locating near- native structures in its initial (Monte Carlo based) global search due to the large size of the search space, particularly for larger proteins [84]. On the other hand, ZDOCK is not as limited by size of the protein structures, as it utilizes the FFT to scan the entire protein translation^ space quickly.

In this study, we tested the effectiveness of refining the initial-stage docking structures from

ZDOCK and ZRANK using RosettaDock, and selecting refined models using either RosettaDock score or ZRANK score. Also we explored using a larger perturbation size in the RosettaDock refinement search, to determine whether this can allow for successful refinement of models that are more distant from native. Finally, we optimized the ZRANK scoring function specifically to evaluate refined structures, which leads to a significant improvement in accuracy. 55

METHODS

Hit and Near-hit Definitions

In this study, hits are defined as predictions with Ca root-mean-square distance (RMSD) of less or equal to than 2.5 A after superposition with the interface atoms in the crystal structure, as described by Chen et al. [4]. Near-hits are defined as having interface Ca RMSD greater than 2.5

A and less than or equal to 4.0 A.

ZDOCK and ZRANK Protocol

The initial-stage docking models were generated by ZDOCK versions 2.3 [4] and 3.0 [85]. For the ZDOCK runs, 6° rotational sampling was used, with different initial rotations for each test case to avoid bias. The 76 rigid-body and medium unbound Benchmark 2.0 cases were used for docking. This was to provide as large a test set as possible, without including the difficult cases which would require explicit modeling of the large interface conformational changes to produce near-native predictions [6]. For the antibody test cases, the search was restricted to the complementarity determining regions for the antibody cases, as described by Chen and Weng

[78].

ZRANK was used to rerank the ZDOCK models as described previously [33], with polar hydrogens added to the unbound proteins using RosettaDock prior to scoring. For the refined structures, hydrogens were already in the structures from RosettaDock. The non-polar hydrogens

(which were also added by RosettaDock) were ignored by ZRANK. 56

RosettaDock Refinement

For the docking refinement protocol, the Monte Carlo refinement method of RosettaDock 2.0 was used [7], with ZDOCK predictions as starting structures. Non-standard amino acids and non­ protein atoms were removed prior to refinement, with exceptions where substitutions were possible (for example modeling MSE as MET). During refinement, extra chil rotamers and chi2 aromatic rotamers were included in the side chain searching. Unbound rotamers were also used, as described by Wang et al. [86], with the exception of the cases with bound antibody structures.

Filtering was turned off, as it was found to lead to no output for many ZDOCK predictions, due to the filter rejecting the models because of small clashes. Three hundred refined models were generated for each starting structure, similar to (but slightly smaller than) the 500-1000 structures generated by Schueler-Furman et al. [87].

The Large Perturbation RosettaDock searching (Large Pert) was achieved through modification of the RosettaDock code and setting Monte Carlo perturbations to 0.4 A and 0.2°, rather than the default perturbation (Default Pert) size of 0.1 A and 0.05° [7].

Optimization of ZRANK Weights

To optimize the weights of the ZRANK terms for scoring refined models, a downhill simplex was used to determine the weights, as was used for the original ZRANK [33]. To generate the docking models for training, all three initial docking protocols used in this study (ZD2.3ZR, ZD3.0,

ZD3.0ZR) were utilized. This provided 37 Benchmark 2.0 cases with near-hits in the top 20 predictions. For all of these cases, the top 20 models for each protocol were refined by 57

RosettaDock to produce 300 refined models. The downhill simplex was then used to maximize the number of hits per test case, selecting the top-scoring prediction (using the candidate weights) from the 300 refined structures for each of the 20 models. The simplex optimized the weights for the seven terms from the original ZRANK, as well as a term for the IFACE potential [85]. To avoid missing the global minimum, 30 different simplex starting points were used as well as five random restarts from each minimum. For the success rate calculation, five-fold cross validation was used. We divided the test cases into five non-overlapping sets, training the weights with four sets and testing on the remaining set. This was performed five times so that each set was tested using weights from the remaining sets.

RESULTS

ZDOCK and ZRANK Success Rates for Hits and Near-hits

To produce initial sets of structures for refinement, ZDOCK versions 2.3 [4] and 3.0 [85] were run on all rigid-body and medium difficulty cases from Benchmark 2.0 [6], and ZRANK [33] was then used to rerank all 54,000 of the initial-stage docking predictions for each ZDOCK run.

ZDOCK 3.0 is a newly developed version of ZDOCK that uses a pairwise interface statistical potential (IFACE) based on improved atom-typing [73], and has been shown to have significantly improved success on a docking benchmark. We did not use ZDOCK 2.1 [67] as its shape complementarity scoring function is contained within ZDOCK 2.3 and ZDOCK 3.0, and its performance is approximately the same or less than that of ZDOCK 2.3 [4]. 58

The success rate for each docking/scoring method for the 63 rigid-body cases is given in Figure

4.1. For each number of predictions (Np) allowed, the success rate denotes the percentage of cases with a hit (or near-hit) ranked within that set of predictions. As defined in the Methods, hits are predictions with interface root-mean-square distance (RMSD) of less than or equal to 2.5 A from structure of the complex, and near-hits are predictions with interface RMSD greater than 2.5 A and less than or equal to 4.0 A from the structure of the complex.

While the success rates of ZDOCK 2.3 and ZRANK have already been investigated [33], Figure

4.1 provides a basis for examining how ZRANK performs when reranking ZDOCK 3.0 models, and also how near-hit success compares with hit success for these protocols. The hit success rate for ZDOCK 2.3 and ZRANK (ZD2.3ZR) versus the original ZDOCK 2.3 (ZD2.3) predictions represents a strong improvement, as has already been noted [33]. For ZDOCK 3.0 followed by

ZRANK (ZD3.0ZR), the success rate is slightly lower than that of ZDOCK 3.0 (ZD3.0) for the top few predictions (Np < 4). After this point, the hit success rate of ZD3.0ZR is better than for

ZD3.0 alone, and surpasses that of ZD2.3ZR at Np = 20.

The near-hit success rates (Figure 4.1, bottom) are shifted up from those of the hits, reflecting the more lenient cutoff. In general, the near-hit success rates follow the same trends as the hit success rates. The top near-hit success rates at Np = 100 are highest for the ZRANK protocols (ZD2.3ZR and ZD3.0ZR), both above 60%. In addition, ZD3.0 gives a relatively high near-hit success rate, particularly for the top predictions. 59

Testing of RosettaDock Sampling and ZRANK Scoring

Based on the success rates for ZRANK and ZDOCK to produce initial hit and near-hit structures, we chose to refine models generated by ZD2.3ZR, ZD3.0, and ZD3.0ZR sets. The ZD2.3ZR,

ZD3.0, and ZD3.0ZR sets have 26,27, and 27 cases, respectively, with hits or near-hits in the top

20 predictions.

The schematic showing the basic steps we employed for docking and refinement is given in

Figure 4.2; the focus of this study is the last two steps. For each test case, the top 20 models from

ZDOCK and ZRANK were refined using RosettaDock to generate 300 models per prediction.

ZRANK was then used to score all 300 models for each prediction, and the best scoring model of the 300 was selected for that prediction. Finally, these 20 refined structures were reranked by

ZRANK score. For comparison, we consider two alternatives: the use of the top ZDOCK models for the input to refinement (rather than ZDOCK and ZRANK) as illustrated by the top dotted line

(which was performed for the ZD3.0 set), and the use of RosettaDock scores to select the refined structures and rerank them (thus skipping the second ZRANK step) as shown by the lower dotted line.

In addition to testing RosettaDock scores instead of ZRANK scores to evaluate the structures, we also explored using a larger rigid-body perturbation size in the RosettaDock structural refinement

(as described in the Methods section), referred to as large perturbation (Large Pert) versus default perturbation (Default Pert). This was performed primarily to determine whether increasing the search space would successfully refine the more distant hits and near-hits. The evaluation of these refinement protocols was performed via several metrics, and is given below. 60

Amount of Structural Improvement

To determine the degree of structural improvement resulting from the refinement and reranking,

we calculated the interface RMSD of the refined structure and compared it with the initial

interface RMSD of the prediction for all models that were initially near-hits (from the three

docking protocols ZD2.3ZR, ZD3.0 and ZD3.0ZR). The histogram of these RMSD changes is

given in Figure 4.3.

For the RosettaDock scoring, the Default Pert searching performed better than the Large Pert. In

particular, the Large Pert had a significant amount of models that were worse than input by > 0.8

A. This can be explained by the fact that the RosettaDock scoring function and search function

were developed together, and the default search size may be optimized for its scoring scheme.

Also in Figure 4.3 the improvement from ZRANK scoring can be seen, resulting in significant

differences in the distributions from RosettaDock scoring. Using the Wilcoxon rank sum test, the

P-values for similarity between the RosettaDock and ZRANK scoring RMSD distributions are

2.7* 108 and < 2.2*1016 for Default Pert and Large Pert, respectively. For all bins representing

structural improvement, the ZRANK scoring had more predictions than for RosettaDock scoring.

Comparing the perturbation sizes for ZRANK scoring, they are approximately equal for the larger improvement bins, while the Large Pert+ZRANK improved more predictions than Default

Pert+ZRANK for the under 0.4 A range. The Default Pert then had more predictions become slightly and moderately worse, and Large Pert had some predictions worsen by 0.8 A or more while default had none. Overall, the large perturbation performed better than default perturbation for the ZRANK scoring. 61

Improved Structures Versus Initial RMSD

To further examine the structural improvement from refinement using these methods, we binned the predictions based on their initial RMSDs and calculated percentage of cases with structural improvement for each bin (Figure 4.4). This indicates which methods perform well for the more distant initial predictions. The dotted line indicates 50% of cases improving; however, it should be noted that random movement of the proteins might not necessarily yield this high a rate of improvement.

It can be seen in Figure 4.4 that the RosettaDock Large Pert+ZRANK gives the greatest overall performance in structural improvement. In four of the six bins it has the highest percentage improved, and in five out of six of the bins it is above 60% improved (all of them are above

50%). The highest percentage improvement is for the bin of 2.5 A to 3.0 A, which represents the most proximal near-hits. Following this method in terms of performance is default perturbation plus ZRANK and default perturbation plus RosettaDock scoring.

Hits After Refinement versus Initial RMSD

In addition to the structural improvement, we also measured the performance for hits after refinement for the same refinement schemes (Figure 4.5). It should be noted that performing no refinement at all would yield 100% hits in the first three bins, and 0% hits in the latter three bins.

In this case, the default perturbation with ZRANK performed slightly better than the large perturbation with ZRANK for the bins with the smallest initial RMSDs. Interestingly, the large 62 perturbation with ZRANK has the most hits for the bin from 3.5 A to 4.0 A starting RMSD, in agreement with that the larger perturbation allowed for more sampling in hit range for those distant predictions than default perturbation. RosettaDock with default perturbation also performed well, but not as high as the ZRANK scoring with either perturbation size.

Score versus RMSD Examples

One means to understand the effectiveness of a scoring function is to plot the scores of the docking models versus the RMSD to see if there is a trend or funnel toward the native structure.

Such funnels are considered to be part of the physical binding process [88-90], thus an accurate energy function should be able to replicate this. Plots of score versus RMSD for three test cases are shown in Figure 4.6, using RosettaDock Large Pert for searching and RosettaDock (top) and

ZRANK (bottom) scoring. For each test case, the top 10 model refinements are shown, to illustrate how the scores and funnels appear for both the near-native structures and those that are far from native for that test case (the top 10 rather than the top 20 were shown to simplify the plots).

In all three cases, there is a hit after refinement when using the ZRANK scoring, and the energy funnels can be seen for the near-hits and hits. This is not as evident when using the RosettaDock scoring for these predictions, as can be expected based on the overall results described above

(Figures 4.3-4.5). Although it is not the top-ranked prediction, the near-hit for 1MLC is refined to

0.98 A using ZRANK scoring to select the top model, close to the minimum rigid-body RMSD for this case (0.6 A). Also using ZRANK scoring, the top-ranked model for 1RLB is a hit with

1.38 A RMSD, and for 1CGI the near-hit model is refined from the initial 3.4 A RMSD to 2.33 A 63

RMSD, thus producing a hit from a near-hit. In the case of 1CGI, the interface RMSD between the superposed unbound and bound structures is 2.02 A, making this one of the more difficult of the rigid-body Benchmark 2.0 cases [6].

Detailed Results: ZD3.0ZR + RosettaDock Large perturbation + ZRANK

Based on the analysis of the four different refinement sampling and scoring schemes (Figures 4.3,

4.4, and 4.5), we chose to utilize the ZRANK scoring and large perturbation of RosettaDock for the remainder of this study.

Numbers of Refined Structures

While we selected to use sets of 300 refined structures for this study, we examined the success rates for using fewer than 300 refined structures from RosettaDock as input to the scoring. In this case, the success rate is out of all hit and near-hit cases from ZD3.0ZR selected as input to refinement. This is provided in Figure 4.7. Random subsets of predictions were selected from the

RosettaDock refined structures to determine the success from sets of fewer than 300 predictions.

The success rates increase upon using more predictions from RosettaDock, with 300 predictions showing the highest overall success rate, in particular for Np > 6. At 20 predictions, using 300 refined structures and ZRANK has a 100% success rate, indicating that all 27 cases that had hits or near-hits in the top 20 prior to refinement had hits after refinement. Based on this analysis, it is possible that greater than 300 refined structures would provide even greater success rate, however this was not tested due to computational limitations. 64

Hit Statistics

To provide an illustration of the specific improvements from this refinement, the detailed results for the refinement of ZD3.0ZR models are given in Table 4.1. As was noted above regarding the success rate (Figure 4.7), all 27 cases had hits after refinement, with four cases becoming hits from near-hits. The number of cases with hits ranked at #1 increased from 7 to 10 after refinement. For 20 of the 27 cases, the RMSD of the top hit was improved indicating the structural improvement resulting from the refinement.

Table 4.1. Results for all near-hit cases of the ZD3.0ZR set, before and after refinement.

ZD3.0ZROrig ZD3.0ZR+Ros+ZR Test Case Hits1 Rank2 RMSD3 Hits1 Rank2 RMS>D 3 1AVX 2 11 1.59 2 1 1.45 1BVN 1 16 1.55 1 3 2.49 1DFJ 3 2 2.06 2 2 2.24 1E6E 7 5 1.96 8 1 1.08 1EAW 0 - - 1 9 1.70 1MAH 6 3 1.10 6 1 0.93 1PPE 19 1 0.76 19 1 0.56 1UDI 0 - - 2 2 2.16 2SIC 9 1 1.38 9 1 0.60 7CEI 11 3 1.34 11 2 1.46 1E6J 9 1 1.58 3 8 1.57 UPS 2 1 1.01 2 5 0.93 1MLC 3 5 1.14 3 9 1.02 1WEJ 4 2 0.75 3 4 0.62 2VIS 1 8 2.02 1 15 2.24 1B6C 4 2 2.38 10 1 2.42 1F51 2 3 1.67 2 2 1.75 1KAC 1 11 2.10 1 2 1.89 1KXP 0 - - 2 9 1.91 1ML0 9 1 1.25 9 1 1.24 1RLB 8 1 2.31 8 1 1.38 1BJ1 1 19 1.18 1 16 1.10 1FSK 17 1 1.05 17 1 1.54 65

1IQD 0 - - 1 18 1.46 1KXQ 2 14 1.29 2 6 0.95 1NCA 1 14 0.90 1 1 0.55 2QFW 3 6 1.80 1 5 1.63

'Number of hits in the top 20 predictions. 2Rank of the first hit; "-" denotes no hit was found in the top 20. interface root-mean-square distance (RMSD) of the first hit, in A; "-" denotes no hit was found in the top 20.

Refined Structure Example

In some cases, there were significant improvements of RMSD, for example 1IQD (Factor

VIII/Fab) for the ZD3.0ZR set, which is shown in Figure 4.8. The original model from ZDOCK, which had an interface RMSD of 4.18 A, was refined by RosettaDock to produce 300 structures, and these models were scored by ZRANK to select the structure shown, with 1.46 A interface

RMSD. Figure 4.8 shows how the ligand in the final structure is both shifted and rotated from the initial prediction to be positioned more correctly on the receptor. Though it is not the typical degree of RMSD improvement (as indicated by Figure 4.3), this demonstrates that it is possible to sample adequately large space in the Rosetta refinement to achieve significantly improved structures from the initial rigid-body prediction, and such structures can be identified by ZRANK.

Interestingly, there were several predictions for this case with initial RMSD less than 4.0 A, and none of these became hits; this is possibly due to the initial positioning having some hindrance preventing the Monte Carlo algorithm from correctly positioning the ligand and its side chains for those predictions. 66

Retraining Weights for Refinement

Based on the success of using the ZRANK scoring function to rescore refined models, we retrained the ZRANK weights to determine whether this would further improve the refinement performance, in particular to rank refined hits at #1. For several instances (such as 1MLC and

1CGI in Figure 4.6) the near-hit structures were refined well using RosettaDock and ZRANK scoring, but the hit predictions were not ranked at #1 among the top 20. This is possibly because the original ZRANK weights were determined using rigid-body models from ZDOCK, and though they are effective they may not be optimal for discriminating refined predictions that should have less clash and better side chain positions. For instance, the van der Waals repulsive weight in ZRANK is significantly smaller than the van der Waals attractive to provide softness for the scoring of the rigid-body predictions; for the refined predictions this softness may not be as necessary.

Weights were retrained as described in the Methods section, using five-fold cross validation with the original ZRANK terms and also incorporating a term for the pairwise IFACE potential [85].

The cross-validation results using these new weights are provided in Table 4.2, along with the initial results for comparison. The number of cases with hits ranked at #1 is significantly higher compared with the original predictions, and also compared with the original ZRANK for refinement scoring. The best performance for the retrained function is seen for the ZD2.3ZR and

ZD3.0ZR sets. Comparing the results using the new weighted refinement to before refinement, the number of cases with hits ranked at #1 increased from 12 initially to 19 (the ZD2.3ZR set) and from 7 initially to 15 (the ZD3.0ZR set). For the ZD2.3ZR set, the 19 cases with hits at #1 comprise over 79% of the 25 cases with hits in the top 20. Both refinement with the ZRANK 67 weights and refinement with the new weights led to significant improvements in the number of cases with hits in the top 20 versus the original unrefined models.

Table 4.2. Number of cases with hits ranked at #1 and in the Top 20, before refinement, after refinement with ZRANK (Ros+ZR) and after refinement with new ZRANK Refine function (Ros+ZR Refine) for three sets of initial predictions. Original Ros+ZR Ros+ZR Refine Input Rank 1 20 1 20 1 20 ZD2.3ZR 12 20 10 24 19 24 ZD3.0 10 19 9 23 13 21 ZD3.0ZR 7 23 10 27 15 26

The weights obtained when training using the entire set of cases are: vdW attractive: 1.0 vdW_repulsive: 0.23 electrostatics short-range attractive: 0.57 electrostatics short-range repulsive: 0.56 electrostatics long-range attractive: 1.09 electrostatics long-range repulsive: 0.29

ACE: 0.7

IFACE: 0.38

As anticipated, the repulsive van der Waals weight is higher than for the original ZRANK weights, which was 0.009, representing less softness in the refinement scoring function. As before, the electrostatics short-range terms are similar to one another. The ACE and IFACE terms both have significant weights and the sum of their weights is approximately the same as the ACE weight for the original ZRANK, where no IFACE term was present. Both IFACE and ACE are 68 contact potentials representing solvent exclusion. ACE was parameterized based on atomic contacts within chains of protein crystal structures [66]. In contrast, the IFACE function was developed using structures of transient protein-protein interfaces, and has 12 atom types rather than the 18 atom types of ACE [73, 85]. Ideally, IFACE should replace ACE entirely when evaluating protein-protein interfaces; however, the amount of available training data is substantially less for IFACE than for ACE, hence some energy terms may be better estimated in

ACE. As the weights and results indicate, these terms complement each other well and help to improve the accuracy of discriminating refined hits from non-hits.

CAPRI Experiment

The Critical Assessment of PRedicted Interactions (CAPRI) is an international experiment for testing protein docking methods where participants make blind predictions of protein complex structures [8]. Recently, the CAPRI experiment has featured a scoring sub-round where a set of initial docking models from several groups (approximately 1000-2000) are rescored and refined by participants, and the top 10 models are submitted for evaluation.

We have used the CAPRI scoring experiment as an opportunity to test the combination of

ZRANK and RosettaDock, with positive results (Table 4.3). The CAPRI evaluation classifies docking predictions as acceptable (*), medium (**) and high (***) accuracy. Our definition of

"hits" is approximately between the criteria for "acceptable" and "medium" for CAPRI. Our general protocol for CAPRI scoring was to rescore input models with ZRANK, filter false- positive models using known biological data (e.g., if a C-term is known not to interact then 69 predictions involving an interface C-term are removed), refine using RosettaDock, and rerank the refined structures using ZRANK.

Table 43. CAPRI scoring results using ZRANK and RosettaDock, with numbers of Acceptable and Medium predictions submitted in the 10 predictions for each target.

Target Protein Acceptable Medium T26 TolB/Pal 1 3 T27.2 E2-25K/Ubc9 7 0 T29 Trm8ATrm82 1 2

For all three targets, we submitted at least one acceptable prediction, and for two targets we submitted medium predictions [82]. For Target 26, where we utilized ZRANK and Rosetta with default perturbation (selecting models based on RosettaDock score) because we had not investigated large perturbation at that time, we achieved 3 medium and 1 acceptable predictions.

In the case of Target 27 (the second interface evaluated), we achieved 7 acceptable predictions, and for Target 29, for which the protocol matches that of the present study with large perturbation and ZRANK, we submitted two medium and one acceptable predictions for the scoring sub- round. For Target 28 (results not shown), no near-hits were provided to the scorers so as a result there were no acceptable predictions from any scorers.

It should be noted that the input predictions for the CAPRI scoring are not necessarily from

ZDOCK; in fact as several groups are involved in producing initial structures some scoring structures are certainly not and may include refined models or more clash than ZDOCK predictions, which was the original intent of ZRANK. However success in the context of the

CAPRI scoring helps to highlight the effectiveness of this algorithm. 70

Computational Performance

The computational time of the Rosetta refinement protocol on a 2.2 GHz Linux machine was on average 9 hours to produce 300 refined structures from the input model. Scoring the 300 refined structures with ZRANK took an average of 4 minutes.

DISCUSSION

Protein docking often requires the effective usage of several steps to produce accurate predictions

[51]. In this study, we have explored an efficient global search with rescoring and refinement, by combining the tools ZDOCK, ZRANK and RosettaDock. The combination of these techniques has led to increased success on a docking benchmark and suggests that this is a promising avenue for further improving protein complex prediction success.

One interesting result from this study is the improvement of using ZRANK scoring over

RosettaDock scoring when selecting refined docking models. It has been shown that RosettaDock scoring, when used in the context of the RosettaDock global search, is effective on a docking benchmark [7]. One major difference in this study is that the models being refined are from rigid- body docking using ZDOCK, rather than from the Rosetta global search. The softness in the scoring function of ZDOCK allows for slight side chain overlaps in the predictions; Rosetta is most likely not as tolerant of these as ZRANK. This also explains the need for removal of the filter when running the RosettaDock refinement, as discussed in the Methods section. On the other hand, the ZRANK scoring function was parameterized to allow it to effectively score rigid- body predictions. 71

The success rates and refinement RMSD changes (Figures 4.3, 4.4, and 4.5) highlight the performance differences between the scoring functions and search strategies explored in this study. It is particularly clear from the success rates in Figure 4.4 that the RosettaDock with large perturbation combined with ZRANK scoring performs well for structural refinement. While

RosettaDock scoring does perform well when rescoring the refined models using RosettaDock default perturbations, it is not as high a success rate as that for either perturbation size with

ZRANK.

The success rates of the refinement procedure described here are further improved by re- optimization of the scoring function for refined docking models. The vast improvement in success of cases with hits ranked at #1 is highly encouraging. Also informative are the weights themselves resulting from the training; indicating that the van der Waals repulsive provides more discrimination after refinement, where models (including hits) no longer have clash that is inherent in rigid-body docking. The IFACE term also helps the scoring function. Though its weight is roughly similar to that obtained for ACE, training the scoring function without the

IFACE term yields lower success rates, though higher than those from the original ZRANK weights (data not shown).

There have been several recent studies that have utilized scoring functions specifically for protein docking refinement. The program FireDock [91] employs two different weighted functions (one for enzyme/inhibitor systems, and one for antibody-antigen systems), each with 11 terms to score refined predictions (after rigid-body and side chain refinement). Compared with this, the scoring function of ZRANK is simpler and does not use separate weights for different types of protein complexes. Another recently developed scoring function, EMPIRE [92], uses an eight term 72 scoring function and a separate side chain energy function, in conjunction with rotamer modeling and CHARMM energy minimization [64]. In that case, the structural improvement of the predictions was more limited than used in this study as it employs CHARMM energy minimization rather than the RosettaDock 6D search.

Future work includes incorporating backbone movements into the refinement search, to overcome limitations imposed by backbone conformational change at the binding interface. Also, the

RosettaDock refinement algorithm can possibly be modified to search more quickly and just a subset of mobile side chains, so that more predictions can be effectively processed. This way the remaining cases from the docking benchmark can conceivably be included (those with hits and near-hits ranked greater than 20) to improve the docking performance on these cases.

In summary, we have shown that it is possible to combine the protein docking tools ZDOCK,

RosettaDock, and ZRANK in a systematic manner to improve the success across a set of cases from a docking benchmark. In this approach, the ZRANK algorithm was found to be effective at rescoring the refined models from RosettaDock, in particular when utilizing a function specifically trained for refined models. 73

Hit Success Rate 70% 1 1—i—i—i—i i 11— -i—i—i—i—i i i ZDOCK2.3 --••- 60% |_ ZDOCK 2.3 + ZRANK —•— ZDOCK3.0 --©-- ZDOCK 3.0 + ZRANK —•— 50% h

0% 10 100

Number of Predictions (Np) Hit and Near-Hit Success Rate 70% 1 1 1—I—I—I I I I— ZDOCK 2.3 --«-- 60% ZDOCK 2.3 +ZRANK —•— ZDOCK 3.0 --©-- ZDOCK 3.0+ZRANK —O— 50% TS CC 40% w w § 30%

Number of Predictions (Np)

Figure 4.1. Hit success rate (top) and hit and near-hit success rate (bottom) for ZDOCK 2.3 and ZDOCK 3.0 with and without ZRANK for the rigid-body cases of Benchmark 2.0, versus number of predictions allowed (Np). Hits are defined as having interface RMSD less than or equal to 2.5 A from the complex structure determined by x-ray crystallography, and for near-hits the RMSD is between 2.5 A and 4.0 A. 74 Receptor, Ligand iUnboun d Structures ZDOCK ( Initial 54,000 Rigid-Body Models | Structural Models ZRANK

20 Rigid-BodT y Models RosettaDock (

Structure I Refinement 20 x 300 Refined Models • and Reranking \

ZRANK

\ 20 Refined Predictions -^— — — Figure 4.2. Protocol employed for docking and refinement (alternative protocols employed in this study are indicated with dashed lines). The initial stage, which produces 20 rigid-body models, includes ZDOCK followed by ZRANK (alternatively the top 20 ZDOCK models are used). The model refinement, which is the focus of this study, employs RosettaDock to refine each model to generate 300 structures per rigid body prediction. These structures are rescored by ZRANK and the top scoring model is selected from each set of 300. The resultant 20 predictions are reranked using ZRANK score (alternatively RosettaDock score is used to select and rerank the structures). 75

Refinement RMSD Change for Near-Hits

<= -0.8 -0.8 to -0.4 -0.4 to 0.0 0.0 to 0.4 0.4 to 0.8 >0.8 Interface RMSD Change, A Closer To Native Farther From Native

Figure 4.3. Histogram of interface RMSD change for all hit and near-hit models after refinement using several search/scoring strategies. Each bin represents the interface RMSD after refinement minus the interface RMSD of the model before refinement. Default Pert = RosettaDock refinement with default perturbation size, Large Pert = RosettaDock refinement with large perturbation size, Rosetta = RosettaDock score used to select the predictions, ZRANK = ZRANK score used to select the predictions. 76

Structural Improvement After Refinement 100%

<= 1.5 1.5 to 2.0 2.0 to 2.5 2.5 to 3.0 3.0 to 3.5 3.5 to 4.0 Initial RMSD, A

Figure 4.4. Percent of models with RMSD improvement for several search/scoring strategies, binned by initial interface RMSD of the models. The dotted line represents 50% success rate. Protocols and abbreviations are the same as Figure 4.3. 77

Hits After Refinement 100% Default Pert + Rosetta [ Large Pert + Rosetta Default Pert + ZRANK Large Pert + ZRANK

<=1.5 1.5 to 2.0 2.0 to 2.5 2.5 to 3.0 3.0 to 3.5 3.5 to 4.0 Initial RMSD, A Figure 4.5. Percent of models with hits after refinement for several search/scoring strategies Figure 4 3 "*"*"* RMS° °f ** m°delS' Pr°t°COlS ^ abbreviations «* the same as 78

1MLC 1RLB 1CGI

o

10 12 14 Interface RMSD, A Interface RMSD Interface RMSD, A

20 0 -20 aml&: -40 Jpiflntv<7S •WW". ^^H ***.mP -60 jt > VHe«KI* 2 -80 F^mw -100 5 10 15 20 25 5 10 15 20 Interface RMSD, A Interface RMSD, A

Figure 4.6. Refinement of three test cases, 1MLC, 1RLB, and 1CGI, with Rosetta scores (top) and ZRANK scores (bottom) versus interface RMSD of the predictions. For each case, 300 refinement models were generated for each of 10 input structures from ZD2.3ZR (1MLC), ZD3.0ZR (1RLB), and 1CGI (ZD3.0), using the large perturbation size for RosettaDock refinement. Each point represents the score for one refinement model, and each point type represents refinement models for one input prediction. For each input model, the top scoring refined model was retained for evaluation. 79

Refinement Success for Various Numbers of Refined Models

20% h

Oo/0 I 1 1 1 1 1 1 1 2 4 6 10 16 20

Number of Predictions (Np)

Figure 4.7. Success rates of refinement for ZD3.0ZR predictions for hit and near-hit cases for various numbers of RosettaDock refinement models. Success is defined as the number of cases (out of 27 hit and near-hit cases from this set) that have a hit in a given number of top-ranked predictions (Np). The large perturbation size was used for RosettaDock, and ZRANK scoring was used to select and rerank the refined model. Random subsets of RosettaDock refined models were selected from a total of 300 for the smaller numbers of predictions. 80

Figure 4.8. Refinement of ZD3.0ZR prediction #12 for test case 1IQD (Factor VIII/Fab), with input ligand (red), refined ligand (green) and bound ligand (blue). The bound receptor is colored gray. The refined structure was chosen by ZRANK score of the 300 refined models from Rosetta. Figure generated with Pymol (www.pymol.org). 81

Chapter 5

Modeling Protein Interaction Affinity Enhancement, and Application to T Cell

Receptor

ABSTRACT

T cell receptors (TCRs) recognize peptides from foreign proteins bound to the self protein called

Major Histocompatibility Complex (MHC) on the surface of an antigen-presenting cell. This

interaction enables the T cells to initiate a cell-mediated immune response to terminate cells displaying the foreign peptide on their MHC. Naturally occurring TCRs have high specificity but low affinity toward the peptide-MHC (pepMHC) complex. This prevents the usage of solubilized

TCRs for diagnosis and treatment of viral infections or cancers. Efforts to enhance the binding affinity of several TCRs have been reported in recent years, through randomized libraries and in vitro selection; however, there have been no reported efforts to enhance the affinity via structure- based design, which allows more control and understanding of the mechanism of improvement.

Here we have applied structure-based design to a human TCR to improve its pepMHC binding.

The design algorithm we developed, named ZAFFI, has a correlation of 0.77 and average error of

0.35 kcal/mol with the binding free energies of 26 point mutations for this system that we measured by surface plasmon resonance. Applying the filter we developed to remove non-binding predictions, this correlation increases to 0.85 and the average error decreases to 0.3 kcal/mol.

Using this algorithm, we predicted and tested several point mutations with significant binding energy improvement, which were then combined and tested, giving a mutant TCR that binds 98- fold more strongly than the wild-type TCR, with the same specificity for the Tax peptide. This 82 energy-based algorithm can be easily applied to improve binding for other systems, requiring only a crystal structure of a protein-protein complex.

INTRODUCTION

T cell receptors (TCRs) are heterodimeric proteins that play a critical role in the cellular immune response. By binding to the complex of a foreign peptide and the Major Histocompatibility

Complex protein (pepMHC), they initiate an immune response against this cell, protecting the host from foreign invasion [93]. The interaction between TCR and pepMHC is highly specific, but of low affinity, making this system an attractive target for protein design. By utilizing protein design, the binding of TCRs can be increased while maintaining their specificity, thus making them useful agents for targeting cancers and pathogens. These designed TCRs should have immense therapeutic potential, for instance to detect and treat virally infected cells. This is complementary to and has distinct advantages over monoclonal antibody-based therapy, which recognizes the surface epitopes of intact virus. Viruses rapidly mutate the surfaces of their coat proteins to evade host immune response. Because TCRs target peptides, some of which are generated from the functional sites of viral proteins and hence can not be easily mutated, TCRs may be more effective than antibodies in treating viral infection. TCRs can also recognize pepMHC complexes on cancerous cells and be applied to cancer therapy.

Efforts in protein design to improve protein-protein interactions have generally fallen into two categories: directed evolution and structure-based design. Directed evolution involves using randomized libraries of mutant proteins (displayed on the surface of phage or yeast, or directly attached to ribosome in a cell-free manner) to select strong-binding clones. In the context of TCR 83 engineering, there have been several directed evolution efforts, improving binding by approximately 100 fold for a mouse TCR [94], and up to 1,000,000 fold for human TCRs [95-97].

However, such technologies can have limitations, such as inadequate library size, and problems with displaying certain mutants due to inefficient protein folding in the expression system [98].

Structure-based methods to enhance protein-protein interactions have become more widely used over recent years, with more crystal structures of protein complexes available, greater computational processing power, and more advanced algorithms to compute binding energy changes. Selzer et al. enhanced the on-rate of the TEM-BLIP enzyme-inhibitor system by over

200-fold, by producing point mutations on the BLIP protein near the TEM interface predicted to enhance electrostatic attraction [99]. While the association rate (k,,,,) and dissociation constant

(KD) were significantly affected, the dissociation-rate (k^) did not change significantly. Another study used several in silico methods to optimize the ICAM-1 protein to bind to LFA-1 protein, resulting in a mutant ICAM-1 with 22-fold binding affinity improvement over the wild-type

[100]. Structure-based design has also been used to alter the specificity of several protein-protein interactions [101, 102]. Structure-based design efforts to enhance the affinities of immunoglobulin proteins (the superfamily that both antibodies and TCRs belong to) have led to four times affinity improvement for a CD8-HLA complex [103] and over seven times affinity improvement for an antibody-antigen system [104]. A recent study performed computational redesign of two antibodies via point mutations; combining five of the tested point mutations led to over 100-fold improvement over wild-type affinity for one of the antibodies to its antigen

[105]. There are currently no studies that have utilized structure-based design to improve the

TCR-pepMHC interaction. 84

We have chosen to engineer the human A6 TCR for enhanced affinity for its in vivo binding partner, the Tax peptide/HLA-A2 MHC complex. HLA-A2 is the most abundant Class I MHC allele, possessed by nearly half of the human population. The peptide is a 9-mer from the Tax protein of the human T cell lymphotrophic virus HTLV-1. In addition to being characterized biophysically [106,107], the wild type structure for this complex has been crystallized [108]. As with other TCR-pepMHC complexes, this system has a relatively low binding affinity (KD ~2 l*M).

In this study, we predicted point mutations of the TCR based on the crystal structure of the complex, and tested their binding affinities for pepMHC using surface plasmon resonance (SPR).

We chose to focus on the A6 TCR alpha chain, as it has a large number of contacts with both peptide and MHC, and the beta chain has been engineered in another study using phage display

[95]. Based on our initial experimental results, we developed a new protein design algorithm named ZAFFI (Zlab AFFInity enhancement) that yielded point mutations with up to 6 times binding improvement. We then combined several point mutations with improved binding to create a single TCR with 98-fold improved affinity for the tax/MHC, which was then tested for specificity with several other peptides bound to the HLA-A2 protein.

METHODS

Initial Modeling of Point Mutations

Initial mutation predictions were produced using molecular modeling of the point mutations using

CONGEN [64,109, 110], followed by scoring using electrostatics and solvation, with a van der 85

Waals filter. Mutations were tested if they improved electrostatics and solvation energies, and passed the van der Waals filter for no severe clashes.

Based on the experimental results from this method, we proceeded to develop the ZAFFI function that ultimately performed the predictions of the mutations. For ZAFFI, the simulation of the point mutations was performed using Rosetta's interface mutagenesis module [9] and the crystal structure of the wild type TCR complex [108]. All 26 interface residues (within a 5 A cutoff of pepMHC) were systematically mutated, generating 494 point mutations. Only the mutant side chain was packed by Rosetta, and the backbone was kept fixed during packing. Standard rotamers were augmented with extra chil, chi2, and chi3 rotamers for all residues with at least 12 neighbors within 10 A. Rosetta output included predicted energies of the mutations, as well as structures of the mutants.

Training of the ZAFFI energy function was performed using a dataset of systematic point mutations at 10 positions on the ovomucoid turkey inhibitor (OMTKY) molecule in four enzyme- inhibitor complexes (from Professor Michael Laskowski, personal communication to Zhiping

Weng), with corresponding PDB IDs: 1CHO, 1PPF, 1YU6, and 3SGB. These 760 mutations were then simulated by Rosetta's interface mutagenesis module [9], using no repacking of neighbor residues, extra chil, 2, and 3 rotamers, and an extra chi cutoff of 12 neighbors (within

10.0 A).

We filtered approximately 100 of the OMTKY mutations prior to training due to excessive clash as measured by the van der Waals repulsive term. These mutant structures most likely required further side chain packing to accommodate the clash and were considered false negatives, which 86 were not of interest as we sought to reduce false positives. Inclusion of these mutations in the training set led to excessively low van der Waals repulsive weights (data not shown).

The scoring terms explored were:

1. van der Waals attractive (vdW_atr)

2. van der Waals repulsive (vdW_rep)

3. explicit solvation (solv)

4. hydrogen bonding (HB)

5. intra-residue clash (intra)

6. short-range electrostatics

7. long-range electrostatics (LRelec)

8. amino acid backbone probability

9. ACE

10. IFACE

11. DFIRE

Terms 1, 2, 6, 7, and 9 are from the ZRANK scoring function [33], except with long-range electrostatics using partial charges rather than full charges. Terms 3,4,5, and 8 are from Rosetta; to obtain them, we modified the Rosetta code to set their weights to 1.0 and output the numbers with more precision than by default. IFACE is a protein interface atomic level statistical potential

[85], and DFIRE is a distance-dependent statistical potential [111], the numbers for which were obtained from the dcomplex program. 87

Using the above terms, we combined all possible sets of sizes 0 to 5 for terms 4 through 11 with terms 1,2, and 3. This led to 219 candidate sets of scoring terms. For each set of scoring terms, multilinear regression was applied, using the GNU Scientific Library

(http://www.gnu.org/software/gsl/) and the filtered mutations for the OMTKY data set described above. This determined a set of weights for the terms that was optimized to fit the energies of the

OMTKY point mutants. Each weighted function from the multilinear regression was then tested using the data set of our measured point mutations for the T cell receptor, and evaluated based on correlation.

Here are the optimal terms and weights obtained from training on the OMTKY data and then applied to our T cell receptor mutations:

van der Waals attractive: 0.24 van der Waals repulsive: 0.017 solvation: 0.24 intra-residue clash: 0.073

ACE: 0.32

ZAFFI Energy Score = 0.24*vdW_atr + 0.017*vdW_rep + 0.24*solv + 0.073*intra + 0.32*ACE 88

ZAFFI Filter

We developed the ZAFFI filter to remove non-binding TCR mutants that were not detected by the

ZAFFI scoring function. This function was developed using all point mutants described in Tables

1 and SI, except for the following, which were produced after filter development:

a chain: D26V, D26W, D26M, R27F, G28M, G28T, G28R, Q30M, Q30H, Q30E,

S51M, K68H, T93Q, T98E, G102D, G102S

P chain: I54R, R95F, G100I

To train the ZAFFI filter, a Monte Carlo algorithm was used to generate candidate weights for pairs of scoring terms. All possible pairs of terms used in ZAFFI energy function development

(described in the previous section) were tested. For each set of weighted terms, the function was judged based on the Area Under the Curve (AUC) of the Receiver Operating Characteristic

(ROC) curve. This was computed using all mutants with ZAFFI energy score less than 0.0, separating the mutants found to bind form those that did not bind, thus separating true positives from false positives from the ZAFFI energy function.

Based on the above training, the ZAFFI filter function was determined, with these terms and weights:

ZAFFI Filter Score = 1.0*HB + 0.01*LRelec 89

This function had an AUC of 0.93 using the test data. A filter score cutoff of 0.05 was used to filter the predictions. Filter scores for all measured data points can be seen in Table 5.5.

Protein Expression and Purification

Wild-type HLA-A2, |32M, TCRa and TCRp\ and mutant TCRa and TCRp proteins were expressed separately as inclusion bodies in E. coli. Mutations were introduced to the constructs via site-directed mutagenesis using standard PCR protocols. The inclusion bodies were refolded using protocols based on those of Garboczi et al. [112, 113]. However, high pressure liquid chromatography (HPLC) (rather than dialysis and ion exchange) was used to purify the refolded proteins from the aggregates; this resulted in much faster purification of the proteins than the published method.

Measurement of Binding Kinetics

Binding of wild type and mutant A6 TCR to the Tax-HLAA2 complex was tested at 25°C using

Biacore 3000 surface plasmon resonance (SPR) biosensor. A6 TCR was immobilized on CM5 chip using standard amine coupling to ~400 Response Units (RU). The Tax-HLAA2 complex was injected over the immobilized TCRs at a flow rate of 100 uJ/min. In order to correct for non­ specific binding of the tax-HLAA2 to the chip surface, tax-HLAA2 was also injected over a surface on which no TCR was bound; this signal was subtracted from those of the TCR-bound cells. After the binding dissociation phase, the baseline was regenerated with a 2 min injection of

0.01 M HEPES (pH 7.4) 1 M NaCl over all channels. HBS-EP (0.01 M HEPES (pH 7.4), 0.15 mM NaCl and 3 mM EDTA, 0.005% v/v Superfactant P20) was used as a running buffer during 90

binding affinity measurements. BIAevaluation (Biacore) was used to determine the kon and k0#of the complex formation by simultaneous global fitting of the data to a 1:1 Langmuir model. When kinetics fitting was not possible, steady-state analysis was used. For all mutants tested, three different Tax-HLAA2 concentration gradients were used to compute the kinetic and equilibrium parameters.

RESULTS

Identification of T Cell Receptor Point Mutations with Enhanced Binding to pepMHC

We used several prediction methods to generate point mutations to test experimentally. These included our original method using the CONGEN program [110], Rosetta's "interface" module

[9], and ZAFFI. ZAFFI uses an optimized energy-based scoring function and Rosetta for side chain packing. These algorithms are described in more detail in the Methods section. In addition, four point mutations comprising the quadruple p chain mutant produced in the study of Li et al.

[95] were included in our study. All mutants were expressed, folded, and measured for binding using the protocol described in the Methods section. Kinetics and binding results for mutations with measurable binding to pepMHC are shown in Table 5.1. The mutations that resulted in loss of binding or immeasurable binding are listed in Table 5.2.

Table 5.1. Binding kinetics and prediction method for measured A6 TCR point mutants. 2 4 1 l 3 Mutation' Pred. k.,n, xlO M V kp^xlQ-' s KD, uM AAG,kcal/mol WT 5AA L08 2AI 0.00 a chain D26M R - - 233 0.19 + 0.11 D26V Z - - 24.50 1.45 ±0.17 D26W Z 2.44 0.08 0.34 -1.08 + 0.13 R27F Z 8.03 1.18 1.47 -0.22 ±0.13 91

G28A I 3.72 2.10 5.65 0.58 ±0.13 G28I I 4.27 0.58 1.35 -0.27 ±0.11 G28L I 6.92 0.68 0.98 -0.46 G28M R 8.59 0.77 0.89 -0.51 ±0.21 G28R R - - 128.60 2.43 ± 0.49 G28T Z 6.76 0.39 0.58 -0.76 ± 0.12 G28V I 2.48 1.34 5.40 0.56 ±0.16 S29A I 3.91 1.45 3.71 0.33 ±0.11 Q30N I 2.88 2.09 7.26 0.73 ±0.17 Q30E Z - - 4.40 0.43 ± 0.14 S51M z 5.38 0.61 1.13 -0.37 ± 0.22 K68H z - - 13.30 1.09 S100A I 4.71 0.97 2.07 -0.01 ±0.14 SIOON I - - 53.79 1.92 ±0.84 SIOOT I 2.58 0.24 0.93 -0.49 ± 0.10 SIOOY I - - 102.00 2.34 ±0.55 P chain I54R R - - 18.52 1.28 ±0.84 A99M L 3.63 0.45 1.25 -0.31 ±0.11 A99K I 2.08 0.52 2.49 0.10 ±0.14 GIOOS L 2.48 0.31 1.24 -0.32 ±0.10 G101A L 1.86 0.06 0.33 -1.09±0.10 R102Q L 2.27 1.03 4.54 0.45 ± 0.23 'The point mutation tested. WT = wild-type A6 TCR. Mutations in bold are those from ZAFFI predictions which improved binding. ^he prediction algorithm responsible for the mutation. I = our initial method; R = Rosetta (using Rosetta score); Z = ZAFFI; L = from Li et al. quadruple mutant [95] 3Binding energy change from the wild type, with uncertainty calculated based on the standard deviation among the three gradients used. For G28L and K68H, the uncertainty could not be calculated because only two gradients were used. "-" kinetics not measurable, KD obtained by steady-state analysis of the sensorgram. Otherwise KD obtained by koff/kon

Table 5.2. TCR point mutations that exhibited no measurable binding to pepMHC. The mutations produced during development of the ZAFFI filter are shown in italics. CDRal CDRa3 CDRfll CDRft3 Q30M T93Q E30A R95F Q30H T93I E30F L98D Q30L T98E E30Q L98F Q30F D99N E30W L98I Q30W S100M E30Y L98M Q30Y S100I A99D S100L A99Y S100F G1001 92

G102S G101M G102D G101V

Table 5.1 indicates improved performance in moving from our initial algorithm to Rosetta to our final ZAFFI algorithm. Our initial predictions ("I") had three mutations with improved binding, seven worse, and one approximately the same (aSlOOA), while using Rosetta scoring to produce mutations ("R"), one improved binding and three became worse. After utilizing these results and developing the ZAFFI function ("Z"), our success improved so that we had four mutations with improved binding (in bold in Table 5.1) and three with worse binding. Among the mutations identified by ZAFFI were ccD26W (the highest affinity point mutation identified by all three methods) and aG28T, which improved binding by 6.2 and 3.6 times, respectively.

For the mutations that improved binding, it can be seen in Table 5.1 that the improvements are largely due to a decrease in the off-rate. This is further shown in Figure 5.1, which divides the

AAGs into k^, and k^ components and compares with the total AAG. Exceptions include aR27F and aG28M, with the former showing significant k^ improvement and the latter showing improvements in both k^ and k^.

To probe the mechanism for binding affinity improvement, we built structural models for the aD26W and aG28T mutants (Figure 5.2). The mutation aD26W (Figure 5.2a and 5.2b) is on the periphery of the interface, near MHC residues E61 and R65. The substitution of a neutral residue

(W) for a negatively charged one (D) in this position of the TCR is electrostatically permitted because the position is closer to E61 than to R65 (3.3 A versus 6.1 A). Furthermore, the mutant

W residue is predicted to make extensive hydrophobic contacts with the MHC, as can be seen in W« & 3 3 H-* tatio n the T .Th e era l : TO e c <; C <•*• I-I o o OS OS S3 a o <-• to GO gerb i iuta t esi d b. T 3 < 5" < Ta b 3 -^ to o s a o CD o> K*• c 63* ti s a t thi in g o f a G o H 5. 3 provi i /- o Waal s e n P o o o o o K> i— N) i— O O a. o owit h th e amode l f o os K> & •*-* '^ >° '£• --J w 4^ bo M W S) O 4^ 0\ o\ o\ w u bo u u it ~ o ^ u> oo ^ 5 w vl si O0 ergetic s a n de s th e t : positio n r th e muta t peptid e re s a 28T , com p o o o o o o i— OOc>OOOOOOOOON)000000 o P O P S Q ^OboOOOQO^Q^^-4^W— o r§ ared ' dde s aile d io n a idu e bou n

o o p o o o wit h th e olvatio n ;G28 T (Figur e 5.2 c a n

t0©OOO O L I an d o oo d th e p P o r66p66ppp o ZAFF I Lj O O *wO » '•_>_*. ^-"~iJ W ^O ft wO\M|O00wWfO U> 00 U\ N) ~J to 4^ to w to VO O OJ .« ~ energ y term s severa l hydr o othe r mutati o epMHC , a G (represente d OOOOQOOOO •- O O © ° i— O o O O O p p o p o p -J © 4^ © O O 8 O 8 vo © ^82823288 to © t-H £l <-ft S 8 82

o o o P o P ooPooopoPoo P o •- o o o o o o > ^ (» b bo U sj Ln L» °-j n 3 a* >-*, to T3 ex '>— 4^ © © !- ^- - - ^o o S3 K- os ^. o LU-l (O 4^ -J C/3 ^ o 00 cr 5.2d ) s 00 -J w io to o »- o >-t H > i c MH C r e s wa s 8- n n en [66] ) a i how s t oin t m positi o 13 | H th e s o o o o o o o o o o o o a — i— O O O i— i— O O to V .Os bo !&>• bo o to -J 4^ to N b ^ A o 8 4^ oo vo £ o to <=> o <-" 00 W VO & OO W M >C to os £ si d 3 sr 3 oi-i n s c Q sw CFQ a> o N>©H-©*-OO©O©©S>©O© ©©;!-••-© 'a a w a o ^ o p o r o w 5? I SUO l sibl e fo r bindi n tsma d nTab l lthou g > o fe o OJ jr u> N> C5 3 t^i vO JO O ,_. 00

n TO cr n 94

Analysis of ZAFFI Performance

As described in Methods, ZAFFI has two components: an energy function and a filter. Figure 5.3 plots ZAFFI energetic scores versus experimentally measured AAGs for the 26 point mutations in

Table 5.1. The correlation coefficient is 0.77 (P-value < 10'5) and the average error from the fit is

0.35 kcal/mol. For comparison, the correlation coefficient between the Rosetta scores and the experimental AAGs is 0.42 (Figure 5.4). Of the points that improved binding, many of them have improvements (negative values) for the van der Waals attractive and the desolvation ACE terms

(Table 5.3). This indicates that these two energy terms are particularly effective for ZAFFI's scoring of true positives.

To more quantitatively examine the contribution of the individual terms to the ZAFFI scoring function, we calculated the correlation with just these terms alone, as well as the correlation drop when removing the terms, for all 26 measured TCR point mutations (Table 5.4). Based on the analysis of the correlations for each term by itself, the two top terms are the explicit solvation and the ACE terms. The van der Waals attractive additionally appears to be benefiting the scoring function; based on the correlation drop when removing the terms, it is second only to the explicit solvation term. Combined with Table 5.3, these results for the energetic terms further indicate that increasing hydrophobic contacts is the primary reason for the affinity improvement for the measured point mutations. 95

Table 5.4. Contribution of ZAFFI scoring terms to correlation with measured TCR point mutations. Term Without Term1 Just Term2 vdW_atr 0.28 0.00 vdW_rep 0.19 0.12 solv 0.39 0.48 intra 0.05 0.08 ACE 0.10 0.42 'Drop in correlation when removing this term from the ZAFFT function. Correlation with TCR mutations when scoring using just the indicated term.

The ZAFFI filter complements the energetic scores by removing non-binding false positive

predictions. The filter uses a weighted sum of two energy terms: hydrogen bonding and long-

range electrostatics, which are not utilized directly in the ZAFFI scoring function. The predictive

improvements due to the ZAFFI filter can be seen by comparison of the binding mutations in

Table 5.1 with the non-binding mutations in Table 5.2. Prior to developing the filter

(corresponding to "I" in Table 5.1), 36 point mutations were tested, of which 26 (72%) ablated

binding to the pepMHC. During and after development of the filter (corresponding to "R" and

"Z" in Table 5.1; in italics in Table 5.2), 19 point mutations were tested, of which 8 (42%) did not

bind. It should be noted that during the development of the filter, several points were produced to

test candidate filter functions. The final filter function has much higher success rate on the tested

point mutations.

The scoring of all experimentally tested point mutations (binding and non-binding) from this filter function can be seen in Table 5.5, which is sorted by ZAFFI score. By ZAFFI scores alone, the top scoring mutations were mutations at residue aQ30 that did not bind; thus the filtering is necessary to remove such false-positives. Using a ZAFFI score cutoff of -0.4 to select predictions, 12 points pass the filter, 9 of which are better binders, out of a total of 11 found in 96 this study (aSlOOA does not bind significantly better than the wild-type and is not counted as a better binder).

Table 5.5. ZAFFI Scores, Filter Scores, and measured binding for all measured TCR point mutants. Chain Mutant ZAFFI Score ZAFFI Filter Pass1 AAG2 a Q30W -3.00 0.46 N NB a Q30F -2.90 0.63 N NB a Q30Y -2.37 0.66 N NB a G28M -1.88 0.03 Y -0.51 a Q30M -1.82 0.63 N NB a Q30H -1.59 0.47 N NB a Q30L -1.50 0.63 N NB a S100L -1.49 0.08 N NB a R27F -1.28 0.15 N -0.22 a S100I -1.07 0.08 N NB a D26W -1.05 0.03 Y -1.08 P G101A -1.00 0.03 Y -1.09 a G28T -0.97 0.01 Y -0.76 a S51M -0.90 0.04 Y -0.37 a G28I -0.89 0.03 Y -0.27 a G100I -0.89 0.00 Y NB a K68H -0.83 0.93 N 1.09 co . A99M -0.72 0.00 Y -0.31 a D26M -0.49 0.03 Y 0.19 a G28A -0.46 0.03 Y 0.58 a R95F -0.46 0.33 N NB a S100T -0.46 0.01 Y -0.49 a G28L -0.43 0.03 Y -0.46 a Q30E -0.42 0.50 N 0.43 P E30Y -0.33 0.25 N NB P E30F -0.33 0.25 N NB P E30W -0.31 0.26 N NB CO . G100S -0.25 -0.02 Y -0.32 a S100A -0.19 0.07 N -0.01 P E30M -0.18 0.25 N NB a T93I -0.17 0.04 Y NB co . A99K -0.14 0.05 N 0.1 a S100M -0.14 0.10 N NB a S100F -0.13 0.08 N NB a G102D -0.12 -0.11 Y NB a D26V -0.11 0.04 Y 1.45 97

p A99D -0.10 -0.01 Y NB a G102S -0.03 -0.61 Y NB P R102Q -0.03 0.14 N 0.45 a T98E -0.01 0.62 N NB a S29A 0.00 0.02 Y 0.33 P E30Q 0.01 0.21 N NB a Q30N 0.10 -0.32 Y 0.73 P E30A 0.10 0.25 N NB P L98F 0.10 0.00 Y NB a T93Q 0.22 -1.49 Y NB a S100Y 0.26 0.22 N 2.29 P L98I 0.49 0.00 Y NB P G101V 0.51 0.03 Y NB a G28V 0.52 0.03 Y 0.56 a SIOOQ 0.52 0.10 N NB a I54R 0.54 -0.74 Y 1.28 a D99N 0.58 0.26 N NB a SIOON 0.58 0.11 N 1.92 P G101M 0.61 0.02 Y NB P L98M 0.96 0.01 Y NB a G28R 1.65 -0.51 Y 2.43 P A99Y 1.92 -0.01 Y NB P L98D 2.48 0.09 N NB

'Whether the mutant passed the ZAFFI filter, defined by having a filter score less than 0.05. 2Measured binding energy change (kcal/mol). NB = no binding detected or unmeasurable binding.

There is also evidence that the filter removes false positive predictions for the binders, in addition to removing non-binders. Using just the points that passed the filter (Figure 5.3, solid points) gives an even greater correlation of 0.85, and a lower average error of 0.30 kcal/mol (versus correlation of 0.77 and average error of 0.35 for all points). 98

Cooperativity of Multiple Point Mutations

Point mutations on the TCR a chain that improved binding to pepMHC were then combined to determine if it was possible to further improve the affinity beyond those of the single point mutations. By comparing with the results for the measured energies with the sums of the component single point mutants, we determined whether the combinations were additive, or exhibited positive or negative cooperativity. Results for binding kinetics and cooperativity are given in Table 5.6. Additionally, comparisons of the SPR sensorgrams of A6 WT, aD26W, and

WFGMT can be found in Figure 5.5, and global fits of the sensorgrams for A6 WT and WFGMT are in Figure 5.6.

Table 5.6. Binding kinetics of combinations of point mutants, and cooperativity of the energetics.

1 1 2 Mutation k^xK^M's' k^,xlO's' KD,mM AAG,kcal/mol Coop WFTMT 1.98 0.0481 0.240 -1.28 ±0.24 1.64 WFGMT 6.32 0.0135 0.0214 -2.72 + 0.10 -0.56

'Residue identities at positions 26,27,28,51, and 100 on the TCRa chain. Mutant residues are in bold. 2 Cooperativity of the mutations, defined by AAGmutant - 2(AAGindividuaI ,,<»,„mutaMs)

The combination of the top binding mutations at five positions on the a chain (WFTMT) had significantly less binding than expected from additivity (negative cooperativity of 1.64 kcal/mol), suggesting that some of the mutations were interfering with one another. Based on analysis of the wild-type crystal structure, we determined that interference between the ctD26W and aG28T mutations was most likely the cause due to the proximity of their side chains, and possibly backbone movements due to the ctG28T substitution. The mutations WFMMT, WFLMT, and

WFIMT did not fold, indicating that mutations of aG28 to larger hydrophobic residues in combination with the other four mutations destabilize the folding of the TCR. We then restored 99 a chain position 28 to glycine to create the WFGMT mutant; this resulted in some positive cooperativity (-0.56 kcal/mol) and a 98-fold KD improvement over the wild type. As with most of the single-point mutations, the kinetics for the WFGMT mutant had a marked decrease in k^, responsible for the majority of the affinity improvement.

Mutant T-cell Receptor Specificity

To determine whether the engineered TCRs in this study have cross-reactivity with peptides other than Tax, we displayed three 9-mer peptides on HLA-A2 and tested the binding of these complexes to the wild type A6 and WFGMT mutant TCR using Biacore. The peptides are known epitopes from the HIV virus, with sequences SLYNTVATL, ILKEPVHGV, and VIYQYMDDL.

None were found to bind either the wild type or WFGMT TCR.

Table 5.7. Binding kinetics and specificity of WFGMT TCR mutant for HLA-A2 with Tax peptide and V7R point mutant of Tax peptide.

peptide k,,, xlO* MV k^xlO'V KD,nM AAG, kcal/mol Specificity' Tax 6.32 1.35 21.4 -2.72 ±0.10 NA V7R 1.14 2.28 200 -1.40 ±0.21 9.3 'Fold change in KD with respect to Tax peptide.

We additionally displayed the V7R point mutant of the Tax peptide (sequence, with mutation in bold, is LLFGYPRYV) on the HLA-A2 and tested its binding to the WFGMT TCR mutant using

SPR (Figure 5.7). The binding affinity of the wild-type A6 TCR for this peptide variant has already been characterized by SPR in another study, and found to be approximately 9.9-fold less than for the Tax peptide [114]. Inspection of the crystal structure of the HLA-A2/TaxV7R/A6

TCR complex [114] indicates that there are no significant changes in the interface of the TCR a 100 chain with the MHC (though the TCR P chain has considerable movement on the CDR3 loop), thus the specificity would not be expected to change significantly for the WFGMT mutant. Our measured specificity of the WFGMT mutant for the Tax peptide with respect to the V7R variant

(Table 5.7) is 9.3-fold, close to that of the wild-type A6 TCR. This indicates that the mutations we introduced into the A6 TCR are able to improve the affinity while maintaining its peptide specificity.

DISCUSSION

This study shows the use of structure-based design to significantly enhance the affinity of a T cell receptor for its peptide-MHC. The final design protocol, named ZAFFI, combines the structural modeling of Rosetta with a novel scoring function that includes shape complementarity and desolvation terms. These terms were found to be optimal after a systematic evaluation of many possible sets of terms, including electrostatics, hydrogen bonding, and several statistical potentials.

To complement the scoring function, ZAFFI also uses a filter that includes hydrogen bond and electrostatics terms to remove false positive non-binders and outlying predictions for binding mutants. As with the scoring function, these terms were found to be optimal after systematic evaluation of possible filter terms on a set of TCR mutations. An earlier protein design study also used satisfaction of hydrogen bonds as a filter, as part of a design function to engineer receptor proteins [115]. 101

We chose a conservative approach toward structural modeling of point mutations. We repacked the mutant side chain using Rosetta, without moving the backbone or neighboring side chains. It is possible that this led to some overlooked substitutions that improve binding (false negatives), of which the conformations require side chain and/or backbone rearrangement. However, the focus of this study was to limit false-positive predictions while reliably predicting some true positives. The low level of noise in our predictions (Figure 5.3) was partly the result of the limited conformational search, and it can be seen in Table 5.1 that we were still able to identify many candidate mutations that improved binding.

The mutations that improved TCR binding affinity in this study were hydrophobic substitutions that increased interface complementarity without losing crucial electrostatic contacts (as seen in

Figure 5.2). The ZAFFI function, trained on the initial TCR mutation results, is optimized for this. It ranks point all mutations with a combined score of desolvation, pair potential and van der

Waals, and is complemented by an electrostatics-based filter. Interestingly, a recent study has found that the hydrophobic residues W, M, and F are the most highly conserved of all residues in protein-protein binding sites [116], thus are key components in naturally evolved binding interfaces. A phage display-derived TCR mutant against another pepMHC also had hydrophobic substitutions, including a CDR2a substitution of QSS to PFW [96]. These highlight the usefulness of hydrophobic substitutions in protein interfaces, in particular for T cell receptors where electrostatics may not the major contributor to the affinity (versus, for instance, the barnase-barstar complex). In contrast, the study of Lippow et al. found that electrostatics was the most useful term in their computational design of antibodies [105]. They additionally found that

"problematic designs were at the binding site periphery", in particular for large residue substitutions with favorable van der Waals. We identified two mutations of large residues, 102 aD26W and aR27F, near the periphery that improved binding significantly (both part of the quadruple mutation WFGMT); the differences between these results may be due to the computational accuracy of the two scoring functions for the terms of interest, as well as the differences between the systems being designed (antibodies and the T cell receptor).

Consistent with the hydrophobic substitutions, the binding improvements seen in this study are largely due to decreases in the off-rate. The correlation between the energy contribution from off- rate and total energy change is 0.87, while this correlation for the on-rate is 0.15 (Supplemental

Figures 5.1a and 5.1b). Decreased off-rate has also been seen as the means of TCR affinity improvement via randomized libraries [95, 96]. The off-rate, which is directly related to the half life of the TCR binding, is known to be related to the functional importance of the TCR signaling

[117].

The results from combining point mutations indicate the complex nature of affinity improvement and protein binding energetics. We observed dramatic negative cooperativity in the case of

WFTMT (possibly due to van der Waals clash), and some positive cooperativity in the case of

WFGMT. Another study noted significant cooperativity when mutating another TCR (based on a phage display mutant), both within and between clusters of residues (referred to as "hot regions")

[118]. We also noted that the p chain point mutations that we tested indicate that the phage display derived mutations for the A6 TCR from Li et al. have significant positive cooperativity. It is unclear whether the cooperativity of -0.56 kcal/mol for the WFGMT mutant arises from within a hot region (residues 26-27) or between regions (involving the relatively distant residues 51 and

100). Further testing of combinations of mutations in these positions should provide more insight. 103

By probing the specificity of our engineered WFGMT TCR mutant, we were able to confirm that our structure-based design did not negatively affect the specificity of the TCR. As the mutants

D26W, R27F, and S51M primarily contact the MHC in the structural models, it was important to confirm that the preference for the Tax peptide would not change. Notably, the specificity of the

WFGMT TCR for the V7R Tax variant was found to be approximately the same as what is measured for the A6 wild-type, and there was no binding to other peptides tested for either wild- type or WFGMT TCRs. TCR mutants that have been optimized by phage display have been shown to maintain peptide specificity as well [97]. This includes the A6 TCR CDR3P variant from Li et al. [95] which was tested for binding two HLA-A2/peptide complexes known to cross- react with the A6 TCR and found to have somewhat greater specificity than the wild-type [119].

Based on the crystal structure of the wild-type A6 TCR complex with HLA-A2/Tax [108], the specificity increase in that case may have been due to contacts between mutant CDR3|3 residues and the peptide at positions where peptide residues varied from Tax.

Future work includes application of the ZAFFI algorithm for specificity enhancement and application to other systems. Crystallization of the multiple mutants described in Table 5.6 would shed more light on detailed structural changes, which could explain the basis of the cooperativity.

There are many other crystal structures of TCR/pepMHC complexes, and any of these can be used for affinity enhancement by ZAFFI. Structure-based design of other protein receptors using approaches similar to our study can yield more effective therapeutics than the wild-type molecules, and can also be used to further improve interfaces that have been optimized using in vitro selection methods. 104

• • •

-1.2 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 AAG, kcal/mol

-i 1

-0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 AAG, kcal/mol

Figure 5.1 AAG k^, (a) and AAG k„ff (b) versus AAG for the 18 measured point mutations with kinetics values in Table 1, with values defined as follows: AAG k„n = RT*ln(kon wt/kon mut), AAG k^ = RT*ln(koff mut/koff wt), and AAG = RT*ln(KD_mut/KD_wt) = RT*(in(koluw/koluI11It) + ln(ksfljD1AflLwt)) = AAG k,,,, + AAG k^. The correlations are: a) 0.15 and b) 0.87. 105

Figure 5.2. Models of mutant complexes for point mutations D26W (a and b) and G28T (c and d) of the TCR a chain, using cartoon (a and c) and spacefill (b and d) to represent the molecules. Wild-type residues and TCR a chain are shown in green, mutant residues are shown in blue, MHC is colored yellow, peptide is colored pink, and TCR p chain is colored slate. In (a) and (c), side chains of peptide and MHC residues in the vicinity of the mutation are labeled. In (a), dotted lines indicate the distance from the TCR D26 OD2 atom to the MHC E58 OE1 atom (3.3 A), and to the MHC R65 NH2 atom (6.1 A). In (c), dotted lines indicate the distances from the (modeled) TCR T28 CG2 atom to the peptide LI CD1 atom (3.4 A), to the MHC W167 CZ2 atom (3.5 A), and to the MHC Y59 CD1 atom (4.3 A). Images generated using Pymol (www.pymol.org). 106

•

•G28V • I54R OS100N

OS100Y .S29A »Q30N- & „ #A99K R102Q • • D26V •GIOO^SIOOA....-^ Q3 W328L .-••• o °gOOA ^100T ..-" «D26M •G28A ...«A99M OK68H ^01>G28T •S51GM8' •D26W OR27F

•G28M i i i_ -1.5 -0.5 0 0.5 1 1.5 2.5 Experimental AAG, kcal/mol

Figure 53. ZAFFI scores versus experimentally measured binding energy changes for 26 measured point mutations for the A6 TCR/tax-MHC complex. The linear regression line is shown as a dotted line. The mutations that do not pass the ZAFFI filter are shown in the plot as empty circles. 107

£. 1— r" 1— i ' i i 'i

1.5 ^ *''

1 -

% 0.5 .,-% -

i i i i i i i -1.5 -0.5 0 0.5 1 1.5 2.5 Experimental AAG, kcal/mol

Figure 5.4. Rosetta scores versus experimentally determined AAGs for the TCR point mutations listed in Table 1. The outlier point S100Y is not shown. Linear fit is given bv the dotted line; the correlation is 0.42. 108

m D26W « - WFGMT — A6WT A6WT * ** 1* \ * to 100 t *

ni t , 3 3 8 30 »k

sspo r --. a: 20 * 50

0 100 150 200 250 30 0H 100 20030040050060 0 700 800 Time, sec Time, sec Figure 5.5. Biacore SPR sensorgrams for binding of Tax/HLA-A2 to designed and wild type TCRs. In (a), the binding for aD26W (dashed line) is compared with binding to wild type (solid line) for Tax/HLA-A2 concentration of 40 ng/ml. In (b), the binding of combined mutant aWFGMT (dashed line) is compared with wild-type (solid line), using Tax/HLA-A2 concentration of 10 (Ag/ml. Approximately 400 response units of TCR were immobilized for each experiment. 109

Figure 5.6. Sensorgrams for the (a) A6 wild type and (b) WFGMT TCRs binding to the tax/HLA-A2 complex. Fit residuals are shown below the sensorgrams. Fits were produced using BIAevaluation software with a Langmuir 1:1 binding model. Approximately 400 RUs of TCR were immobilized for each experiment, and tax/HLA-A2 concentrations of (a) 0.9 ^M, 0.84 \iM, 0.785 |AM, 0.45 nM, 0.42 JAM, 0.3925 \iM, 0.225 ^M, 0.21 ^M, 0.192625 \iM and (b) 0.45 ^M, 0.42 \JLM, 0.225 jxM, 0.21 nM, 0.1125 uM, 0.105 \iM, 0.098125 \xM were injected over the surface. 110

Chapter 6

Studying Cooperativity of TCR Mutations Through Modeling and Experiments

ABSTRACT

Computational protein design requires algorithms that can reliably evaluate energies of single and

multiple mutations. In the previous chapter, we described the development and implementation of

a protein design algorithm, ZAFFI, that accurately predicts the binding energy change of point

mutants of the A6 T cell receptor (TCR) for the tax/HLA-A2 peptide complex (MHC).

Combining several point mutations that were predicted by ZAFFI led to nearly 100-fold binding

energy improvement. However, the limitations of the protein modeling were evident as certain

combinations of mutations had energies that were not additive and some displayed strong

noncooperativity. Here we will explore the basis for the nonadditivity of these mutations through further mutagenesis of the T cell receptor along with investigating the protein models of these

mutations.

INTRODUCTION

Protein design is a valuable tool for understanding proteins, and more importantly for developing

protein-based therapeutics. One avenue of protein design is to modify proteins to make them bind

to their partner proteins more strongly than in vivo to help detect and/or inhibit antigens efficiently. Such designs can be performed using in vitro selection methods, such as phage display and yeast display [94, 95, 97], or using in silico protein design to predict the mutations

[105, 120]. While the latter method can be more attractive to researchers due to greater Ill understanding and control of the mutations, it can become complicated due to possible errors in the energy function, structural modeling, and experimental issues such as folding of the predicted mutant proteins.

One particular difficulty with the modeling of mutant proteins is modeling combinations of mutations. While in Chapter 5 we introduced a method to successfully model point mutations in a

T cell receptor, when combining these point mutations (Table 5.5) the results were quite nonadditive. This is despite the predicted conformations of the mutant residues in the combined mutants having the same conformations of the point mutations, leading to the prediction of additivity in the energies in those combined mutations.

Other studies have been devoted to studying the nonadditivity in the protein-protein interaction energetics of combinted mutations. One early study into this phenomenon determined that the majority of combinations of mutations are additive, however nonadditivity is seen in cases where the interface is not rigid or the mutant residues are in close proximity to one another [121]. Later studies indicated that interfaces are organized into closely packed "hotspots" of residues, and intra-hotspot mutations often lead to nonadditivity [122]. A recent study of combinations of mutations on a T cell receptor in fact found significant cooperativity (nonadditivity where the measured A AG is larger than the sum of individual A AGs) within and between hotspots [118] for mutations based on a TCR mutant obtained from phage display. The inter-hotspot cooperativity was exhibited between clusters separated by over 20 A; the authors attributed this to TCR flexibility and propagation of the energetics through a "dynamic structural network". 112

In this study we explore the nonadditivity of mutations seen in our A6 T cell receptor to determine the basis for these results. Using surface plasmon resonance, we measured the binding of various combinations of a chain and p chain TCR mutations that we have characterized individually in Chapter 5. These comprise mutations both within and between hotspots. We then examine the structural models of certain combinations of mutations to explain the lack of additivity in these cases. Finally, we determine how the association and dissociation rates were affected when combining these mutations, finding a surprising level of additivity of dissociation rates of component mutations.

METHODS

Protein Production and Binding Energy Measurement

All proteins were mutated, expressed, and folded using E. coli and the protocols described in

Chapter 5. Binding energies were measured by Biacore surface plasmon resonance; details of this protocol are also in Chapter 5.

Modeling of TCR Mutations

T cell receptor mutations were modeled using Rosetta as described in Chapter 5. For the testing of intra-chain clash, alpha chain G28 and D26 mutations were modeled separately and scored as separate chains by ZAFFI. 113

RESULTS

Combination of Mutations on the a and P Chain

To expand on the results seen in the previous chapter for multiple mutations, we have measured the binding affinities of 14 combination mutations in the A6 TCRa and TCRp chain. Binding energies and cooperativity for these mutations are given in Table 6.1. The a chain mutations were based on point mutations that improved binding from ZAFFT (ranging from combinations of 2 to

5 mutations), and the p chain mutations were from the phage display mutant from Li et al. [95].

While the greatest cooperativity is seen for the phage display mutation (MSAQ), most of the a chain mutations (9 of 14) show some cooperativity as well, up to 0.77 kcal/mol.

Table 6.1. Binding energy changes of combinations of A6 TCR point mutants, along with the sum of the AAGs of the component point mutations and the cooperativity, defined as the measured A AG minus the sum of the point mutation AAGs. All numbers are in kcal/mol. The a chain mutations are for residue numbers 26,27,28,51, and 100, and the P chain mutations are for residues 99-102. Mutant residues are shown in bold. a Mutations Measured AAG Sum of AAGs Cooperativity DRIST -0.90 -0.76 -0.14 DRISA -0.38 -0.28 -0.10 DRMST -1.77 -1.00 -0.77 DRMSA -0.92 -0.52 -0.40 DRLST -1.03 -0.49 -0.54 DRLSA -0.51 -0.47 -0.04 DRMMT -2.12 -1.37 -0.75 DRTMT -1.62 -1.62 0.00 DFMMT -1.78 -1.59 -0.19 DFTMT -1.61 -1.84 0.23 WRMMT -1.12 -2.45 1.33 WRTMT -1.23 -2.70 1.47 WFGMT -2.72 -2.16 -0.56 WFTMT -1.28 -2.92 1.64

P Mutations MSGR -1.04 -0.63 -0.41 114

AGAQ -1.56 -0.64 -0.92 MSAR -3.59 -1.72 -1.87 MSAQ -3.93 -1.27 -2.66

Certain trends can be seen in Table 6.1, showing dependence of the particular mutations involved.

For instance, the mutation S100T in combination with the mutations at position 28 (G28I, G28M,

G28L) shows more cooperativity than mutations with S100A. Also, introducing the mutation

S51M seems to have very little impact on the cooperativity (i.e. its energy is additive), namely when progressing from DRMST to DRMMT.

Modeling of the Nonadditivity of TCRa Combination Mutants

There are several combination mutants in Table 6.1 for which the binding energies were notably less than additive: WRMMT, WRTMT, and WFTMT. While the mutations D26W, S51M and

S100T are common to these mutants, other mutants containing these mutations did not have this lack of cooperativity, e.g. WFGMT.

We performed a more detailed investigation of the structural modeling of the TCR mutations

WRMMT, WRTMT, and WFTMT. These all have mutations in CDR1, CDR2, and CDR3, and based on the structure of the wild type TCR and the models of the mutants it did not seem likely that the side chains were interacting between these CDRs due to the large distance. However, residues D26W and G28T, and D26W and G28M of the structural models did appear to have side chains in close proximity (Figure 6.1), although as mentioned previously the models of these combination mutations did not require packing of the side chains differently than for the point mutations. 115

To quantify the proximity of the side chains and determine if this could cause unfavorable binding energetics, we used ZAFFI to analyze the interaction between side chains of combinations of D26W and G28 mutants. Results for this analysis are shown in Table 6.2, showing the raw scores for vdW_atr, vdW_rep, and ACE, in addition to the weighted vdW_rep term. In fact, there is some clash as indicated by the large (unfavorable) van der Waals repulsive terms for D26W paired with G28T and G28M. When weighted by the ZAFFI weight (0.017), these values become similar to the measured noncooperativities shown in Table 6.1.

Table 6.2. Inter-residue energy terms for combinations of point mutations on the TCR a chain, with their raw ZAFFI scores, and weighted van der Waals repulsive term. Combination vdW_atr vdW_rep ACE vdw _rep*0.017 D26W-G28T -5.08 55.6 -1.18 0.95 D26W-G28M -5.24 36.3 -1.99 0.62 D26W-G28L -5.72 4.42 -2.05 0.08 D26W-G28I -5.92 4.56 -2.04 0.08

To test the clash predictions in Table 6.2, we attempted to test combinations including D26W and

G28L or D26W and G28I, however these combinations were unable to fold due to the strong hydrophobicity of these mutations.

Combination of Mutations from Different Chains

Having modeled and measured the various TCR a and p chain mutations, we then tested the combination of mutations on both chains. In particular, we sought to improve upon the binding affinity of the phage display-derived p chain mutant that improved binding approximately 1000 times. Results for these combinations are given in Table 6.3. 116

Table 63. Association rates, dissociation rates, and binding energy changes for combinations of mutations from different chains. Mutant1 kgg? kj AAG4 SumAAGres5 SumAAGch6 WFGMT-phage 1.77E+04 1.36E-04 -3.33 -3.43 -6.65 WFGMS-phage 3.31E+04 1.32E-04 -3.71 -2.94 NA G28T-phage 5.22E+04 1.25E-04 -4.02 -2.03 -4.69 D26W-phage 2.22E+04 4.35E-05 -4.13 -2.35 -5.01 G28T-phagel02R 3.26E+04 9.95E-05 -3.87 -2.48 -3.83 'TCR mutant combination tested, alpha chain followed by beta chain, "phage" corresponds to the phage display mutant MSAQ, "phagel02R" corresponds to mutant MSAR Association rate, in units M's-1 dissociation rate, in units s"1 4Binding free energy change compared with wild-type, in units kcal/mol 5Sum of AAG from component point mutations 6Sum of measured AAG from each chain. NA = no data; WFGMS on alpha chain not measured

Table 6.3 indicates several interesting trends in the combination mutants. The combinations with the phage (MSAQ p mutant), notably WFGMT-phage, did not achieve the expected energy when adding the energies from the separate chains (SumAAGch). This indicates some long-range effect or hindrance that does not allow the energy from the a and p chains to combine with additive (or better) energies. In fact, as the mutations become less extensive on the a chain, the binding energy becomes better, in contrast to what was seen when mutating the a chain alone. The

WFGMS mutant, which is expected to bind approximately 0.5 kcal/mol less strongly than

WFGMT, was produced because the S100T mutation is the only mutation near the CDR3P loop in the wild type A6 TCR crystal structure. This in fact improved the binding. Surprisingly, just combining the point mutations G28T and D26W with the phage mutant improved binding further, with D26W-phage binding the best of all measured mutations. This is due to a dramatically low dissociation rate, approximately 2500 times slower than for the wild type A6 TCR. The SPR sensorgram for this mutant is shown in Figure 6.2. 117

Also in Table 6.3 the degree of cooperativity can be seen for the combination mutations. In addition to comparing the energy with the sum of the A AG from the point mutations, we also included the sum of the AAG of the mutations from each separate chain. The latter would indicate the degree of nonadditivity between the two chains, by ignoring the cooperativities of the point mutations within each chain (already seen in Table 6.1). Most measured energies were not equal to the sum of the component energies, either by residue or chain. However, the WFGMT-phage measured energy was similar to the sum of the residue energies, and the G28T-phagel02R measured energy was similar to the sum of the energies of the mutants from each chain. In the case of the latter, this indicated that the a and f} chains could in fact be additive given the phagel02R mutant on the p chain. Unfortunately D26W-phagel02R, WFGMS-phagel02R, and

WFGMT-phagel02R did not fold adequately to produce measurable data. In the cases of G28T- phage and D26W-phage, the relative kinetics and binding energies between these two combinations were similar to those between G28T and D26W alone, in that the association rate was higher for G28T-phage, and the dissociation rate and overall binding energy were superior for D26W-phage.

Additivity of Kinetic Rates for Combinations of Mutations

Based on the results presented in Tables 6.1 and 6.3, the energies of the combinations of TCR mutations show a clear nonadditivity. Figure 6.3 shows these results in a scatter plot, with the additive AAG plotted against the actual AAG. Some of the mutations are on or close to the line, however the majority of them, particularly the beta and the alpha+beta combinatioins, are above the line. These represent measured energies that are better than expected from additivity. Three a chain combinations are below the line, with energies worse than expected from additivity: 118

WFTMT, WRTMT, and WRMMT. These, as discussed previously in this chapter, have steric hindrances responsible for the binding becoming less favorable than expected from additivity of the point mutations.

To explore the basis of the nonadditivity, we divided the energy into association and dissociation rate components and analyzed the energetic additivity in terms of these components (Figure 6.4 and 6.5). Based on these figures, it is clear that the dissociation rates are far more additive than the association rates for these mutations. Even the alpha+beta and beta mutations, which are not additive for the total energy (Figure 6.3), show additivity for the dissociation. Outliers of the dissociation plot (below the line) are WFTMT, WRTMT, WRMMT, and the WFGMT-phage alpha+beta combination. For some mutations, however, association additivity can be seen, as some alpha combination mutations are on the line in Figure 6.5. But some of the alpha chain mutations, and all of the alpha+beta and beta mutations, are not on the line. It should also be noted that the scale of the association rate energy change is smaller, reflecting the small association rate changes seen for this system, compared with dissocation rate.

The correlations for the total additivity, association additivity and dissociation additivity plots are

0.57, 0.46, and 0.87 respectively. Analyzing just the alpha chain combination mutations, these correlations become: 0.48, 0.66, and 0.83. Interestingly, the association additivity correlation is greater that the total energy additivity in this case, but still the highest correlation is for the dissociation.

This analysis indicates that the dissociation rates of the component mutations can be a good predictor of the dissociation rate of a combined mutation. The results from the point mutations in 119 the previous chapter (Figure 5.1) indicated that the dissociation rate was largely responsible for the energetic improvement of those mutants. This dataset also indicates that the dissociation rate is largely responsible for the energetic improvement (comparing the x-axis ranges between

Figures 6.4 and 6.5). More importandy, it shows the additive properties of the dissociation rate as well, and that it is much more additive than the total energy or the association rate.

DISCUSSION

In order to effectively model affinity-enhancing protein mutations, it is essential to understand how these mutations combine. Even mutations derived from in vitro selection methods require combinations of several mutated residues to achieve large (> 100-fold) binding energy improvement [94,95,97].

In this study, we have explored the reasons behind nonadditivity seen in the A6 T cell receptor system. The a chain mutations that had significantly less than additive energies (WFTMT,

WRTMT, and WRMMT) were found to have some steric hindrance between the side chains in the models, despite the Rosetta program packing the side chains in that manner. While in reality the structures of the bound mutant TCRs in these cases may not have exhibited this clash, some relief to the clash would have come at a similar or equal energetic cost. Crystallizing these mutants in bound form would be quite useful to determine exactly how the structures would behave.

These particular nonadditive mutations, while binding less than expected, provided useful information for the development of the ZAFFI predictive algorithm. Though including the intra- 120 chain in the van der Waals could lead to excessive noise in the scoring function, it would be useful to utilize such a term as a filter. In fact, the false positive ZAFFI prediction in Chapter 5,

(3G100I, appears to have some intra-chain side chain clash that would have been filtered by such a means.

In addition to analyzing these nonadditive cases, we have also analyzed the energetic additivity of all the combination mutations, both in terms of total energy and kinetic components. It was seen that the dissociation rate is in fact more additive than the association rate or the total energy when combining mutations. This result held for mutations within chains and involving both alpha and beta chains.

The reasons behind this are not yet clear. Several previous studies have investigated the kinetics of TCR binding and have determined that the recognition is flexible and involves a multiple stage process [107,123]. Therefore, since TCR association is clearly a complex process, it is likely that the association of the combinations of mutations would not follow a simple additive trend. As many of the association rates were more favorable than expected from additivity (above the dotted line in Figure 6.4), this could be cooperativity during the association stage preventing the association rate from falling significantly.

Future work includes working on an effective means to computationally model such additivity.

As mentioned, a filter should be utilized in ZAFFI to avoid small steric clashes within chains for multiple mutants. It would be very informative to have crystal structures of a subset of these complexes, in order to learn more about the structural reasons behind these energetic trends and to more efficiently model them. 121

Figure 6.1. Structural models of WRTMT (top) and WRMMT (bottom) TCR mutants. MHC, peptide, and TCR|3 are colored cyan, TCRa is colored green, residue a26 is colored blue, and residue cc28 is colored red. 122

25

20 ^7**%^ w* i.fyfr*y •

15 ^%^%^v^v^^^

10

^^^^^

-200 200 400 600 800 1000 1200 1400 1600 Time, seconds

Figure 6.2. SPR sensorgram for combination mutant D26W-phage, with raw data (points) and data fit to 1:1 Langmuir model (solid lines). Tax/MHC concentrations are 30 ^ig/ml, 15 Hg/ml, and 7.5 ng/ml. RU = response units. Measured KD for this mutant is 1.96 nM. 123

Additivity of AAG

I •— 1 1 i i i i Alpha • Beta ± • Alpha + Beta • x • - • -

•

• - • • • • • • • -

.,.-••""• -

i i i i i i i i -4.5 -3.5 -3 -2.5 -2 -1.5 -0.5 Measured AAG, kcal/mol

Figure 6.3. Additive AAG versus measured AAG for combinations of TCR point mutations. Dotted line represents perfect additivity. 124

Additivity of Association Rate

2.5 1— •• I 1 1 "-T 1 Alpha • Beta * • Alpha + Beta • x • A 1.5 • ,.••'*' o • | _..•-''

• * ..•••' ~ 2 1 c A ,.-••''' o • < 0.5 < ..•-•" • £ 3 CO "

-0.5

i i i 1 f 1 -1 -0.5 0 0.5 1 1.5 2.5 Measured AAG k^, kcal/mol

Figure 6.4. Association rate additivity for combination TCR mutations, shown as sum of the k™ component of AAG for the point mutations versus the measured k,,,, component of the AAG for the combination mutation. Dotted line, representing perfect additivity, shown for reference. 125

Additivity of Dissociation Rate

i i i i i Alpha • Beta ± V*'"" Alpha + Beta •

*?••••'•'"""""

• •

•....•••"" •

• -6 Ll l l l i r -6 -5 -4 -3 -2 -1

Measured AAG koff, kcal/mol

Figure 6^. Dissociation rate additivity for combination TCR mutations, shown as sum of the 1^ component of AAG for the point mutations versus the measured k^ component of the AAG for the combination mutation. Dotted line, representing perfect additivity, shown for reference. 126

Chapter 7

Exploring the In Silico Designability Protein Complexes,

and Affinity Enhancement of CD4/gpl20

ABSTRACT

Computational protein interface design has many uses and its applications are near limitless. In the previous two chapters, we have described the development and testing of a design algorithm,

ZAFFI, on the TCR/pepMHC complex. This led to significant improvement of the affinity of the

TCR for the pepMHC. Although the predictive results on the TCR/pepMHC system were

successful, it remained to be explored how this algorithm behaves when utilized to score other interfaces. Here we utilize ZAFFI to perform systematic mutagenesis of the interfaces from a large set of transient protein-protein complexes. With this large dataset of results, we are able to determine the in silico trends in the mutation propensities of certain residues. By investigating interfaces with regard to the highest number of residue positions that are predicted to increase interaction affinity, we identified the interface of human CD4 and HIV gpl20 as a strong candidate. We compared our energy function used for the T cell receptor with other candidate functions using known CD4 mutation data from the literature. In addition, as the HIV gpl20 protein as a whole has a high propensity to mutate, we then explored the sequence and structural data for this protein to determine the optimal CD4 mutations for targeting the conserved portions of the gp!20 structure. 127

INTRODUCTION

Computational protein design is a field that is being actively explored as a means to improve interaction affinity and modify interactions. Recent studies include improvement of antibody- antigen interaction affinity [104, 105], altering specificity of a protein-protein interaction [101], and design of a protein-DNA interaction [124].

Previously, we have described development of a protein design algorithm, ZAFFI, and its application to affinity enhancement of a T cell receptor. The scoring function weights were trained based on mutations from an enzyme-inhibitor system, and the final function was selected from candidate functions with various terms based on its correlation with known data for the

TCR. In addition, we developed an electrostatics filter to remove false positive predictions from the scoring function, optimized using data from our TCR dataset.

Having developed an effective scoring function, one question is how it behaves on a large set of protein interfaces and what trends might be seen. Additionally, given another protein-protein interface, it would be useful to know whether the ZAFFI function as applied to the TCR would be useful to similarly enhance its affinity via point mutagenesis, or if some reformulation of the function would be necessary.

This study explores these questions. We first utilized ZAFFI to analyze a dataset of nonredundant protein-protein interaction structures from Mintseris et al. [73], systematically mutating all residues to determine which point mutations improve affinity in silico. Based on this analysis, we 128 selected a candidate protein-protein interaction structure from this dataset that had relatively high potential for affinity enhancement: the human CD4 and HIV gpl20 interaction.

The CD4-gpl20 interaction has been studied for well over 15 years. Early studies included scanning mutagenesis of the CD4 protein to determine its binding site with the gpl20 protein

[125-127]. While the CD4 protein was initially considered to be a candidate for HIV inhibition, it was found to not be as effective as hoped in vivo, requiring much more CD4 protein than expected to neutralize the virus [128]. Since then, the CD4 protein has been crystallized in complex with several isolates of the gpl20 protein [129, 130], and the interaction has been extensively studied using surface plasmon resonance [131-134]. Recent focus on gpl20 inhibition has included monoclonal antibodies that target the CD4 binding site [135] and a scorpion toxin protein that was rationally designed to mimic the CD4 protein [136,137]. There were some early attempts to modify the CD4 protein to better bind the gpl20 protein, with limited success [138,

139]. It is possible that this has not been pursued as extensively because HIV would mutate to avoid CD4 that was extensively optimized for a particular gpl20 clone, in addition to the lack of success of wild-type soluble CD4.

In this study we present our analysis of the CD4 protein to predict mutations to improve its binding to HIV gpl20. In selecting these mutations, we explored the various crystallized strains of gpl20 and also data from multiple sequence alignments of known gpl20 sequences. We also explored possible scoring models in addition to the ZAFFI function, by comparing predicted energies with those of characterized CD4 mutations. Finally, we present initial results for binding affinities of our predicted CD4 mutants against the gp!20 protein. 129

METHODS

Protein Modeling and Simulation

Mutant proteins were modeled and scored as in Chapters 5 and 6, using Rosetta to produce the models and ZAFFI to score the models. In the case of exploring candidate energy functions,

weights were produced by training using the enzyme-inhibitor dataset from Prof. Laskowski, as in Chapter 5. All wild-type protein structures were downloaded from the PDB [16].

Protein Production and Affinity Measurement

To test the wild type and mutant CD4 binding to HIV gpl20, the first 183 amino acids (domains

1 and 2) of the CD4 protein, with His-tags cloned into the C-term, were expressed using leaky expression in E. coli [140]. The expressed CD4-His protein was then purified from cells via nickel column purification, and high pressure liquid chromatography was used to separate the folded protein from the aggregate CD4. The recombinant gpl20 protein was from the YU2 clone of HIV [141], obtained in frozen form from ImmunoDiagnostics, Inc.

Biacore surface plasmon resonance was then used to measure the CD4-gpl20 interaction.

Approximately 200 response units of CD4 protein was immobilized on the cell of a CM5 chip.

Recombinant gpl20 was then injected over the flow cells at 100 \iVmin, and the surface was regenerated after dissociation using two 10 second injections of 15 mM HC1. Analysis of the binding was performed using Biaevaluation software using a 1:1 Langmuir model, after double referencing [142], i.e. subtracting the signal from a reference cell and the signal from an injection over the CD4 cell with running buffer. 130

RESULTS

In Silico Systematic Mutagenesis of Transient Protein-Protein Interfaces

For all structures in the dataset of transient complex structures from Mintseris et al. [73], we performed in silico point mutagenesis of all interface residues from each side of the interface, followed by scoring and filtering of the mutations with ZAFFI. The dataset included 241 structures, of which 177 were classified as enzyme/inhibitor or other, and 64 were classified as antibody/antigen. This led to an average of approximately 22 residues mutated for each side of an interface, approximately 10,500 residues mutated, and 200,000 simulated mutations.

With the ZAFFI scores of all mutations of this dataset, we then performed a residue-based analysis of which mutations were predicted to improve binding. To classify mutants as improving binding, a ZAFFI score cutoff of -0.5 and ZAFFI filter cutoff of 0.06 were used; if both scores were below these thresholds for a mutation, then it was predicted to improve binding. We generated a color map to indicate the residue-based frequency of mutations that were predicted to improve binding, shown in Figure 7.1. Frequencies are defined as the number of the residue- residue transitions predicted to improve binding divided by the total number of those residue- residue transitions evaluated. Consistent with our ZAFFI analysis of TCR mutations, mutations to and from proline and to glycine were ignored, as these were likely to affect backbone conformation and dynamics, as discussed by Kortemme et al. [9].

Several interesting trends can be seen in Figure 7.1. The most notable is that for both datasets the affinity improvements are largely due to substitutions to larger, more hydrophobic residues (e.g. 131 tryptophan, phenylalanine), based on the high frequencies (hotter colors) on the bottom rows.

This may be in part due to the ZAFFI scoring function which consists of weighted terms for statistical contact potential, desolvation, and van der Waals. However, experimental evidence from the T cell receptor (Chapter 5) indicates that such mutations can indeed improve the binding energetics, for example the point mutations aR27F and aD26W. In addition, the importance of these large hydrophobic residues has been noticed in other studies. An in vitro antibody maturation experiment with a four amino acid code found that tyrosine residues were critically important interface residues [143], and a statistical analysis of protein interface residue conservation found that tryptophan, methionine and phenylalanine residues were the most conserved of all amino acids at protein-protein interfaces [116].

Other notable trends are apparent when comparing frequencies between the two datasets in Figure

7.1. Two of the most frequent mutations that improve binding for the antibody/antigen set, histidine to phenylalanine and histidine to tyrosine, are not as frequent in the enzyme/inhibitor and others set. In addition, the enzyme/inhibitor and other dataset includes a higher frequency of mutations from alanine and transitions from large hydrophobic residues to other hydrophobic residues (lower left cells). Such results may be due to the different physicochemical makeup of the classes of interfaces, as well as the interface shapes and complementarity. For instance, higher shape complementarities are seen in enzyme/inhibitor interfaces than for antibody/antigen interfaces [53].

With the datasets of binding improvements from in silico mutagenesis, we went from a residue- based analysis to a structure-based analysis to determine which proteins had greatest interface designability. Each protein from the complex was considered separately, designated as receptor or 132 ligand arbitrarily in the case of the enzyme/inhibitor or others dataset, and for the antibody/antigen dataset the antibody (or immunoglobulin) protein was designated receptor. The proteins were sorted by the number of residues with at least one mutation predicted to improve binding affinity. The top 20 proteins for each dataset are given in Tables 7.1 and 7.2.

Table 7.1. In silico designability of nonredundant transient enzyme/inhibitor and other complexes from PDB. Top 20 proteins shown. PDB ID Description Rec/Lig Improve1 Total2 Fraction3 1K3Z TRANSCRIPTION r 25 53 0.47 1K3Z TRANSCRIPTION 1 21 47 0.45 1K90 TOXIN ,LYASE/METAL_BINDING_PROTEIN r 20 66 0.30 1DHK COMPLEXJHYDROLASE/INHIBITOR) 1 20 41 0.49 1I7W CELL.ADHESION r 19 75 0.25 1FIN COMPLEX JTRANSFERASE/CYCLIN) r 18 47 0.38 1JCH RIBOSOMEJNHIBITOR^HYDROLASE 1 17 46 0.37 1F83 HYDROLASE/MEMBRANE_PROTEIN r 17 68 0.25 1F34 HYDROLASE/HYDROLASEJNHIBITOR r 17 37 0.46 1KXP CONTRACTILE_PROTEIN/PROTEIN_BINDING r 16 36 0.44 1ITB COMPLEX JIMMUNOGLOBULIN/RECEPTOR) r 16 41 0.39 1K90 TOXIN ,LYASE/METAL_BINDING_PROTEIN 1 15 64 0.23 1I4D SIGNALING_PROTEIN r 15 26 0.58 1F83 HYDROLASE/MEMBRANE_PROTEIN 1 15 32 0.47 1F51 TRANSFERASE r 15 32 0.47 1EFU COMPLEX JTWO_ELONGATION_FACTORS) r 15 53 0.28 1DN1 ENDOCYTOSIS/EXOCYTOSIS r 15 49 0.31 1WQ1C0MPLEX_(GTP-BINDING/GTPASE_ACTIVATI0N) r 14 36 0.39 1WQ1C0MPLEX_(GTP-BINDING/GTPASE_ACTIVATI0N) 1 14 28 0.50 1T0C COMPLEXJHYDROLASE/INHIBITOR) r 14 40 0.35 'Number of positions on protein with at least one mutation predicted to improve binding. 2Number of total interface positions tested on protein. 3Fraction of interface positions with improving mutations.

Table 7.2. In silico designability of nonredundant transient antibody/antigen complexes from PDB. Top 20 proteins shown. PDB ID Description Rec/Lig Improve' Total2 Fraction3 HQDIMMUNE_SYSTEM/BLOOD_CLOTTING r 11 35 0.31 1GC1 COMPLEX_(HIV_ENVELOPE_PROTEIN/CD4/FAB) r 11 24 0.46 133

1FBI COMPLEX_(ANTIBODY/ANTIGEN) 1 9 24 0.38 1EZV OXIDOREDUCTASE/ELECTRON_TRANSPORT r 9 22 0.41 1A07 COMPLEX_(MHC/VIRAL_PEPTIDE/RECEPTOR) 1 9 25 0.36 1SBB IMMUNE_SYSTEM r 8 14 0.57 1QKZ IMMUNE_SYSTEM r 8 19 0.42 1I9R CYTOKINE/IMMUNE_SYSTEM r 8 22 0.36 1HEZ ANTIBODY r 8 18 0.44 1E6J HIV_CAPSID_PROTEIN_(P24) 1 8 13 0.62 1BJ1 COMPLEXJANTIBODY/ANTIGEN) r 8 25 0.32 1AKJ COMPLEX_(MHC_I/PEPTIDE/CD8) 1 8 32 0.25 1NSN COMPLEX_(IMMUNOGLOBULIN/HYDROLASE) 1 7 23 0.30 1KXT HYDROLASEJMMUNE_SYSTEM 1 7 23 0.30 1I9R CYTOKINE/IMMUNE_SYSTEM 1 7 24 0.29 1FSK IMMUNE_SYSTEM r 7 25 0.28 1BQHIMMUNE_SYSTEM r 7 16 0.44 2JEL COMPLEXJANTIBODY/ANTIGEN) 1 6 17 0.35 1SBB IMMUNE_SYSTEM 1 6 13 0.46 1HEZ ANTIBODY 1 6 15 0.40 'Number of positions on protein with at least one mutation predicted to improve binding. 2Number of total interface positions tested on protein. 3Fraction of interface positions with improving mutations.

Both datasets include many candidates for mutagenesis to improve the interactions. For the enzyme/inhibitor or other dataset, the top two proteins are the two components of the complex from PDB entry 1K3Z, designated as "Transcription" in the PDB header. This is in fact the complex of the NF-KB transcription factor and IKBP, an important complex in transcription activation [144]. Other proteins listed in the top 20 of this dataset include the Ras-RasGAP complex (1WQ1), a well-known signaling protein complex that is implicated in many cancers.

The antibody/antigen structures include fewer residues that can be mutated to improve binding, though this is a result of the smaller interface size for these complexes. The similar ranges of values in the "Fraction" column of Tables 7.1 and 7.2 support this. The A6 TCR (1A07 ligand) is seen among the top antibody/antigen proteins for numbers of residues that can improve binding, with 9 residues out of 25 residues in the interface that are predicted to improve binding when mutated. The top two proteins in this set include the 1GC1 receptor, which is the human CD4 134 protein in the complex with HIV gpl20. It had 11 of 24 total interface residues with mutations predicted to improve binding. Considering this high degree of designability, along with a possible impact on HIV research and therapeutics, we decided to further explore this protein complex as a target for structure-based design.

Exploring Candidate Functions for Modeling CD4 Mutations

Before producing point mutations of the CD4 protein to test experimentally, we chose to evaluate the scoring function being used in the energetic modeling. This was performed by scoring point mutations of the CD4 protein that have been characterized for gpl20 binding in the literature. A total of 28 binding energy changes were obtained from several literature sources [126,127,134], and were simulated and scored. Weighted candidate scoring functions were produced for all combinations of terms and evaluated based on performance against the CD4 dataset. Results for two functions, the ZAFFI function and a function that uses intra-residue clash, short-range electrostatics, and backbone tytyprobabilit y terms, are shown in Figure 7.2.

The results in Figure 7.2 indicate possible difficulties scoring this system. The top plot, obtained using the ZAFFI scoring function, shows several false positive points, and the correlation is not as high as for the TCR mutations. In addition, it is clear from both plots that the filter, optimized for the TCR mutations, is not applicable for the CD4 system, as it leads to some false negatives and does not remove all false positives. The function with intra-residue clash, short-range electrostatics, and backbone tyty probability terms has a better correlation with these points, including fewer false positives. The short-range electrostatics term may be beneficial in this case 135 because electrostatics plays a larger role in the energetics for this complex versus the TCR system

(where including electrostatics in the scoring reduced the correlation).

It should be noted that this dataset of CD4 mutations is limited in several ways, and the results should be noted with this in mind. For one thing, the vast majority of the mutations in the set are alanine scanning mutations, limiting the character of the interface changes. More importantly, many of the energetic measurements were taken using methods with a large capacity for error, relative to surface plasmon resonance. This is indicated by some significant differences between the results for the same CD4 mutants for two different studies [126,127].

In addition to the above analysis, we analyzed the four CD4-gpl20 complex structures available in the PDB [16], representing three gpl20 clones (two from primary isolates): 1GC1, 1G9M,

1G9N, and 2B4C. Although their structures share the same binding orientation, there are some differences in the side chain positions and identities (in the case of the gpl20) that affect the in silico mutagenesis results. By screening for all four structures simultaneously, as well as for various candidate scoring functions, we were able to identify CD4 mutations with greater likelihood for improved binding across multiple gpl20 clones.

Analysis of gpl20 Conserved Residues

Although we had candidate mutations produced by our scoring function and several gpl20 structures, we also verified the conservation of the HIV gpl20 in the vicinity of the mutations to ensure that these affinity improvements would carry across all or most strains of gpl20. To do this, we downloaded the multiple sequence alignment of the HIV gp!20 (env) protein from the 136

LANL database ("http://www.hiv.lanl.gov/). This provided 975 HIV gpl20 sequences. We then mapped these sequences onto the gpl20 protein residues in the crystal structure, and colored each residue of the crystal structure based on the percentage of sequences with that residue identity conserved in the multiple sequence alignment. We used the gpl20 YU2 structure for this analysis; it is shown in Figure 7.3.

Overall, it can be seen in Figure 7.3 that the CD4 binding site of gpl20 is highly conserved (blue or green), as noted by other studies. The two unconserved (red) regions of gpl20 in the lower left and right are due to an artifact of the crystal structure, because the gpl20 sequence was changed from the full-length protein in these regions to allow for crystallization. The phenylalanine 43 residue of CD4, well known to be key to gpl20 binding, is shown in spacefill and is seen in a conserved pocket deep in the interface. It can also be seen that the CD4 helix on the left contacts a highly conserved arm of the gpl20. This region represents a good target for CD4 mutagenesis.

Preliminary Binding Affinity Results for a CD4 Mutation

Based on the above analysis, we produced a set of candidate CD4 mutations to test experimentally. They were expressed and measured as described in the Methods. We have measured the wild type CD4 and one CD4 mutant; sensorgrams for their gpl20 binding are shown in Figure 7.4. The wild type CD4 protein binds the YU2 gpl20 with a measured binding constant of 1.74 nM, an association rate of 9.77xl04 NT's"1, and a dissociation rate of 1.7xl04 s"1.

These values are close to previously published values for these proteins [131, 145]. The H27Y mutant of CD4 binds the gpl20 approximately 2.6-fold more strongly, with KD of 0.67 nM, association rate of 9.25xl04 M"V1, and dissociation rate of 6.16xl0"5 s"1. As with many of the 137 point mutations for the A6 TCR system (Chapter 5), the binding affinity improvement seen for this mutant is due to slower dissociation. The H27Y point mutation was also identified in a previous study using phage display to improve CD4 binding to another gpl20 clone [139].

Testing of more point mutation predictions will determine whether further binding improvement can be made, and verify the utility of the prediction protocol for this system.

DISCUSSION

After developing the protein design function ZAFFI and applying it to TCR design, we have utilized it to analyze other protein-protein interactions. This includes a large set of nonredundant transient protein-protein interactions, where we analyzed the frequency of residue-residue transitions for in silico improved binding. This led to interesting results regarding the overall transitions, and the differences between the antibody/antigen and enzyme/inhibitor/other sets gave insight into the different makeups of these different classes of interfaces. These differences may stem from the evolution of these complexes; enzyme/inhibitor complexes have undergone selection through evolution, whereas antibodies mature against the antigens in the thymus and are selected for high affinity and no cross-reactivity.

We then used the results for this dataset to determine the designability of the protein-protein interfaces in this dataset. Many of the interfaces had over 10 residues on a particular protein where mutations were predicted to improve binding. In addition, several interfaces had both partners appear with a large amount of possible mutations; this is most likely due to interfaces with some holes or gaps in the core or periphery where either partner can make substitutions to 138 fill these pockets. By looking at the top designable proteins, we selected the CD4-gpl20 interface as a target for design.

We then produced candidate mutations for the CD4 protein, using our existing scoring function and several candidate functions, as well as the gpl20 structural and sequence information available. This approach has several advantages over phage display; in particular the structure- based design allows selection of mutations that will ensure more breadth of binding improvement across all gpl20 clones. Our approach led to at least one predicted mutant that improves binding nearly three-fold by slowing the dissociation phase of binding. More mutations are currently being tested; with more data from these point mutations, we can combine mutations that improve binding, and also refine the protein design function if necessary.

Future work includes further testing of the CD4 mutations, and if significant binding improvement is seen for the YU2 gpl20 being tested we will also test the CD4 mutant against other gpl20 proteins. We can also identify other protein-protein interactions that can be improved

(Tables 7.1 and 7.2), and via experimental validation can explore how and to what extent we can design those interfaces as well. 139

Mutants with Improved Binding, Transient El/Other 0.3

0.25

0.2

0.15

0.1

0.05

W I F L C M V Y PATHGSQN E D K R WT Residue

Mutants with Improved Binding, Transient AA 0.3

0.25

0.2

0.15

0.1

0.05

W I F L C M VYPATHGSQN E D K R WT Residue Figure 7.1. Residue transition frequencies for in silico improvement of binding affinity for a dataset of enzyme/inhibitor and other structures (top) and antibody/antigen structures (bottom). Residue identities are given in one letter code, and are sorted by the hydrophobicity scale of Fauchere and Pliska [2]. Mutations to and from proline, and to glycine, were ignored. 140

CD4 Mutations, ZAFFI Function

• 3

2.5 - - • 2 - _.,. - . -""' 1.5 • • -

1 ..---"'" -

0.5 • 0

-0.5 i 1

-1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5 Experimental AAG CD4 Mutations, Intra+SRelec+BBprob Function

3 —1 1 1 T 1 1 1 1 —1

2.5 _ •

2 - • _..- _,..--""" - o 1.5 • - • 1 • 0.5 • - • o • • 0 &:Q-7Z ° •o •• -0.5 .-••x>#o o -1 - o

-1.5 1 . 1 • 1 I ' • ' -1 -0.5 0.5 1 1.5 2 2.5 35 Experimental AAG Figure 7.2. Predicted versus experimentally measured binding energy changes for CD4 mutants binding the gpl20 protein, using the ZAFFI scoring function (top) and a candidate function with intra-residue clash, short-range electrostatics, and backbone (jnjj probability terms (bottom). Mutations filtered using the ZAFFI filter are shown as open circles. Experimental values are in kcal/mol. 141

**n i

L

Figure 73. Crystal structure of the HIV gpl20 protein bound to the CD4 protein. The gpl20 protein is colored according to observed frequency of mutation of the residue (blue is most conserved). The CD4 protein is shown in gray cartoon, with the key interface residue Phe 43 shown in spacefill. 142

-100 100 200 300 400 500 600 700 800 900 Time, seconds CD4 H27Y

150 h

50 h

-100 300 400 500 900 Time, seconds Figure 7.4. Binding sensorgrams for CD4 wild type (top) and CD4 mutant H27Y (bottom) proteins to YU2 gpl20. Concentrations of injected gpl20 were 120 nM, 100 nM, and 80 nM. Sensorgrams were produced by modeling the data with a Langmuir 1:1 model using Biaevaluation software. 143

Chapter 8

Summary and Future Directions

The work in this thesis has spanned various aspects of computational protein modeling, including

predicting the structures of protein-protein interactions, improving upon these predictions, and

altering the protein interactions to improve binding by structure-based design. We have been able

to verify these predictions against docking benchmarks (for the protein docking work) as well as

experimentally (for the TCR design work). And perhaps most importantly, we have been able to

apply these programs toward modeling disease-related proteins (KSHV LANA and HIV gpl20),

potentially contributing to efforts to understand and control the diseases of which these proteins

are key components.

The protein-protein docking work presented in this thesis (Chapters 2, 3, 4) can be extended in

various ways. For instance, in collaboration with Kevin Wiehe, we have planned a modification

of the RosettaDock structural refinement where only select side chain positions are repacked

(rather than all side chains). Preliminary results using this method proved promising, with

improved running time and structural predictions.

In addition, the symmetric docking program M-ZDOCK can be applied to predict other

symmetric protein interactions of interest. One researcher has been collaborating with us

regarding using M-ZDOCK to predict the structure of a symmetric astrovirus protein. There are a

variety of other viruses that rely on symmetric multimers to function (e.g. HIV), and some of these multimers exhibit Cn symmetry. 144

One more direction of the docking work is that the docking programs in general can be further tested with regard to docking protein homology models. Dealing with homology models is sometimes necessary in docking research (e.g. with the LANA protein), and it has also been the part of some CAPRI targets. It would be useful to determine under what conditions the docking programs developed in our lab (M-ZDOCK, ZDOCK and ZRANK) perform in the context of homology models, and how dependent this performance is on the quality of the model.

The protein design research (Chapters 5, 6,7) has much promise and it is a very exciting field to be involved in. This is in part due to the fact that the predictions made by the modeling can be directly verified in the lab, and in most cases our lab is the first to characterize these mutations experimentally. In this regard, the CD4-gpl20 project is underway and more testing of the predicted mutations is necessary. As mentioned at the end of Chapter 7, the results from this initial testing of point mutations can lead to combination of point mutants that improve binding

(in the case that more than one mutation improves binding), or retraining of the CD4 mutant scoring function if necessary to better model the initial data.

We also have ambitions to further characterize the structures of the TCR mutations discussed in

Chapter 6. Even the point mutations we combined in the a chain led to cooperativities of up to approximately 0.8 kcal/mol, yet the source of this nonadditivity is not yet clear. Based on analysis of the kinetics, the on-rate is responsible for this nonadditivity, but there must be a structural manifestation of this binding energy. Computational modeling may give some possible clues regarding this, however crystallization should give a more conclusive answer. Modeling and/or crystallization can also be used to elucidate the structure of the strongly nonadditive CDR3P mutant from phage display, however it seems that this structure has been crystallized [146] 145

(though not yet released or described). With crystal the structure of this mutant, it should be possible to model mutations to combine with it that will further improve the binding affinity of the A6 TCR, perhaps to picomolar range. 146

LIST OF JOURNAL ABBREVIATIONS

AIDS Rev AIDS Reviews

Conf Proc IEEE Eng Med Biol Soc Conference Proceedings of the Institute of Electrical and Electronics Engineers Engineering in Medicine and Biology Society

Cold Spring Harb Symp Quant Biol Cold Spring Harbor Symposium on Quantitative Biology

Curr Opin Chem Biol Current Opinion in Chemical Biology

Curr Opin Struct Biol Current Opinion in Structural Biology

Curr Top Med Chem Current Topics in Medicinal Chemistry

Curr Top Microbiol Immunol Current Topics in Microbiology and Immunology

EMBOJ European Molecular Biology Organization Journal

FEBS Lett Federation of European Biochemical Societies Letters

Genome Inform Genome Informatics

J Biol Chem Journal of Biological Chemistry

J Exp Med Journal of Experimental Medicine

J Immunol Journal of Immunology

J Immunol Methods Journal of Immunological Methods

J Med Chem Journal of Medicinal Chemistry

J Mol Bio Journal of Molecular Biology

J Mol Med Journal of Molecular Medicine

J Mol Methods Journal of Molecular Methods

J Mol Recognit Journal of Molecular Recognition

J Struct Biol Journal of Structural Biology

J Virol Journal of Virology

Int J Pept Protein Res International Journal of Peptide and Protein Research 147

Nat Biotechnol Nature Biotechnology

Nat Struct Biol Nature Structural Biology

Nat Struct Mol Biol Nature Structural and Molecular Biology

Nucleic Acids Res Nucleic Acids Research

Proc Natl Acad Sci U S A Proceedings of the National Academy of Sciences of the United States of America

Prot Eng Des Sel Protein Enginering Design and Selection

Protein Eng Protein Engineering

Protein Sci Protein Science

Proteins Proteins: Structure, Function and Bioinformatics

Trends Biotechnol Trends in Biotechnology 148

BIBLIOGRAPHY

1. Delano, W.L., The PyMOL Molecular Graphics System. 2002.

2. Fauchere, J.L., et al., Amino acid side chain parameters for correlation studies in biology and pharmacology. Int J Pept Protein Res, 1988.32(4): p. 269-78.

3. Srinivasan, V., et al., Definition of sequence requirements for latency-associated nuclear antigen 1 binding to Kaposi's sarcoma-associated herpesvirus DNA. J Virol, 2004. 78(24): p. 14033-8.

4. Chen, R., L. Li, and Z. Weng, ZDOCK: an initial-stage protein-docking algorithm. Proteins, 2003.52(1): p. 80-7.

5. Garber, A .C., et al., DNA binding and modulation of gene expression by the latency- associated nuclear antigen of Kaposi's sarcoma-associated herpesvirus. J Virol, 2001. 75(17): p. 7882-92.

6. Mintseris, J., et al., Protein-Protein Docking Benchmark 2.0: an update. Proteins, 2005. 60(2): p. 214-6.

7. Gray, J J., et al., Protein-protein docking with simultaneous optimization of rigid-body displacement and side-chain conformations. J Mol Biol, 2003.331(1): p. 281-99.

8. Janin, J., et al., CAPRI: a Critical Assessment of PRedicted Interactions. Proteins, 2003. 52(1): p. 2-9.

9. Kortemme, T. and D. Baker, A simple physical model for binding energy hot spots in protein-protein complexes. Proc Natl Acad Sci USA, 2002.99(22): p. 14116-21.

10. James, J A., et al., Structure of adeno-associated virus type 2 Rep40-ADP complex: insight into nucleotide recognition and catalysis by superfamily 3 helicases. Proc Natl Acad Sci USA, 2004.101(34): p. 12455-60.

11. Qamra, R., V. Srinivas, and S.C. Mande,Mycobacterium tuberculosis GroEL homologues unusually exist as lower oligomers and retain the ability to suppress aggregation of substrate proteins. J Mol Biol, 2004.342(2): p. 605-17. 149

12. Korkhov, V .M., et al., Oligomerization of the GAB A transporter-1 is driven by an interplay of polar and hydrophobic interactions in transmembrane helix II. J Biol Chem, 2004.

13. Katchalski-Katzir, E., et al., Molecular surface recognition: determination of geometric fit between proteins and their ligands by correlation techniques. Proc Natl Acad Sci U S A, 1992.89(6): p. 2195-9.

14. Berchanski, A. and M. Eisenstein, Construction of molecular assemblies via docking: modeling of tetramers with D2 symmetry. Proteins, 2003. 53(4): p. 817-29.

15. Eisenstein, M., et al., Modeling supra-molecular helices: extension of the molecular surface recognition algorithm and application to the protein coat of the tobacco mosaic virus. J Mol Biol, 1997. 266(1): p. 135-43.

16. Berman, H.M., et al., The Protein Data Bank. Nucleic Acids Res, 2000.28(1): p. 235-42.

17. Viejo-Borbolla, A. and T.F. Schulz, Kaposi's sarcoma-associated herpesvirus (KSHV/HHV8): key aspects of epidemiology and pathogenesis. AIDS Rev, 2003.5(4): p. 222-9.

18. Verma, S.C., K. Lan, and E. Robertson, Structure and function of latency-associated nuclear antigen. Curr Top Microbiol Immunol, 2007.312: p. 101-36.

19. Zhang, C, et al., Determination of atomic desolvation energies from the structures of crystallized proteins. J Mol Biol, 1997.267: p. 707-26.

20. Chen, R., et al., ZDOCK predictions for the CAPRI challenge. Proteins, 2003. 52(1): p. 68-73.

21. Goldstein, H., Classical mechanics. 2d ed. 1980, Reading, Mass.: Addison-Wesley Pub. Co. xiv, 672.

22. Tilton, R.F., Jr., J.C. Dewan, and G A. Petsko, Effects of temperature on protein structure and dynamics: X-ray crystallographic studies of the protein ribonuclease-A at nine different temperatures from 98 to 320 K. Biochemistry, 1992.31(9): p. 2469-81.

23. Liu, Y., et al., Structures of the two 3D domain-swapped RNase A trimers. Protein Sci, 150

2002.11(2): p. 371-80.

24. Segelke, B .W., et al., Structures of two novel crystal forms ofNaja naja naja phospholipase A2 lacking Ca2+ reveal trimericpacking. J Mol Biol, 1998. 279(1): p. 223-32.

25. Scott, D.L., et al., Interfacial catalysis: the mechanism of phospholipase A2. Science, 1990. 250(4987): p. 1541-6.

26. Rey, F .A., et al., The envelope glycoprotein from tick-borne encephalitis virus at 2 A resolution. Nature, 1995.375(6529): p. 291-8.

27. Bressanelli, S., et al., Structure of a flavivirus envelope glycoprotein in its low-pH- induced membrane fusion conformation. Embo J, 2004. 23(4): p. 728-38.

28. Wlodawer, A., et al., Structure of bovine pancreatic trypsin inhibitor. Results of joint neutron and X-ray refinement of crystal form II. J Mol Biol, 1984.180(2): p. 301-29.

29. Hamiaux, C., et al., The BPT1 decanter observed in acidic pH crystal forms pre-exists as a stable species in solution. J Mol Biol, 2000. 297(3): p. 697-712.

30. Bates, P.A., et al., Enhancement of protein modeling by human intervention in applying the automatic programs 3D-JIGSAWand 3D-PSSM. Proteins, 2001. Suppl 5: p. 39-46.

31. Bochkarev, A., et al., Crystal structure of the DNA-binding domain of the Epstein-Barr virus origin-binding protein EBNA I. Cell, 1995. 83(1): p. 39-46.

32. Fiser, A. and A. Sali, ModLoop: automated modeling of loops in protein structures. Bioinformatics, 2003.19(18): p. 2500-1.

33. Pierce, B. and Z. Weng, ZRANK: reranking protein docking predictions with an optimized energy function. Proteins, 2007. 67(4): p. 1078-86.

34. Bochkarev, A., et al., Crystal structure of the DNA-binding domain of the Epstein-Barr virus origin-binding protein, EBNA1, bound to DNA. Cell, 1996. 84(5): p. 791-800.

35. Havranek, J.J., CM. Duarte, and D. Baker, A simple physical model for the prediction 151

and design of protein-DNA interactions. J Mol Biol, 2004.344(1): p. 59-70.

36. Morera, S., et al., Mechanism of phosphate transfer by nucleoside diphosphate kinase: X- ray structures of the phosphohistidine intermediate of the enzymes from Drosophila and Dictyostelium. Biochemistry, 1995.34(35): p. 11062-70.

37. Gonin, P., et al., Catalytic mechanism of nucleoside diphosphate kinase investigated using nucleotide analogues, viscosity effects, and X-ray crystallography. Biochemistry, 1999.38(22): p. 7265-72.

38. Garavaglia, S., et al., Structure of human NMN adenylyltransferase. A key nuclear enzyme for NAD homeostasis. J Biol Chem, 2002.277(10): p. 8524-30.

39. Werner, E., et al., Crystal structure of human nicotinamide mononucleotide adenylyltransferase in complex with NMN. FEBS Lett, 2002.516(1-3): p. 239-44.

40. Taylor, T .C. and I. Andersson, Structure of a product complex of spinach ribulose-1,5- bisphosphate carboxylase/oxygenase. Biochemistry, 1997.36(13): p. 4041-6.

41. Gulbis, J.M., et al., Structure of the cytoplasmic beta subunit-Tl assembly of voltage- dependent K+ channels. Science, 2000.289(5476): p. 123-7.

42. Gulbis, J. M., S. Mann, and R. MacKinnon, Structure of a voltage-dependent K+ channel beta submit. Cell, 1999.97(7): p. 943-52.

43. Wang, Y.F., et al., Channel specificity: structural basis for sugar discrimination and differential flux rates in maltoporin. J Mol Biol, 1997.272(1): p. 56-63.

44. Auerbach, G., et al., Zinc plays a key role in human and bacterial GTP cyclohydrolase I. Proc Natl Acad Sci USA, 2000.97(25): p. 13567-72.

45. Kitov, P.I., et al., Shiga-like toxins are neutralized by tailored multivalent carbohydrate ligands. Nature, 2000.403(6770): p. 669-72.

46. Hunt, J.F., et al., Structural adaptations in the specialized bacteriophage T4 co- chaperonin Gp31 expand the size of the Anfinsen cage. Cell, 1997.90(2): p. 361-71. 152

47. Aloy, P., et al., The relationship between sequence and interaction divergence in proteins. J Mol Biol, 2003.332(5): p. 989-98.

48. Garber, A.C., J. Hu, and R. Renne, Latency-associated nuclear antigen (LANA) cooperatively binds to two sites within the terminal repeat, and both sites contribute to the ability of LANA to suppress transcription and to facilitate DNA replication. J Biol Chem, 2002. 277(30): p. 27401-11.

49. Wong, L.Y. and A.C. Wilson, Kaposi's sarcoma-associated herpesvirus latency- associated nuclear antigen induces a strong bend on binding to terminal repeat DNA. J Virol, 2005. 79(21): p. 13829-36.

50. Kelley-Clarke, B., et al., Determination of Kaposi's sarcoma-associated herpesvirus C- terminal latency-associated nuclear antigen residues mediating chromosome association and DNA binding. J Virol, 2007.81(8): p. 4348-56.

51. Smith, G.R. and MJ. Sternberg, Prediction of protein-protein interactions by docking methods. Curr Opin Struct Biol, 2002.12(1): p. 28-35.

52. Halperin, I., et al., Principles of docking: An overview of search algorithms and a guide to scoring functions. Proteins, 2002.47(4): p. 409-43.

53. Jones, S. and J.M. Thornton, Principles of protein-protein interactions. Proc Natl Acad Sci USA, 1996. 93(1): p. 13-20.

54. Gabb, HA., R.M. Jackson, and MJ. Sternberg, Modelling protein docking using shape complementarity, electrostatics and biochemical information. J Mol Biol, 1997.272(1): p. 106-20.

55. Vakser, IA., Protein docking for low-resolution structures. Protein Eng, 1995. 8(4): p. 371-7.

56. Ritchie, D.W. and G J. Kemp, Protein docking using spherical polar Fourier correlations. Proteins, 2000.39(2): p. 178-94.

57. Abagyan, R., 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. Journal of Computational Chemistry, 1994.15(5): p. 488-506. 153

58. Fischer, D.,etal., A geometry-based suite ofmolecular docking processes. J Mol Biol, 1995.248(2): p. 459-77.

59. Comeau, S.R., et al., ClusPro: an automated docking and discrimination method for the prediction of protein complexes. Bioinformatics, 2004.20(1): p. 45-50.

60. Mandell, J.G., et al., Protein docking using continuum electrostatics and geometric fit. Protein Eng,2001.14(2): p. 105-13.

61. Jackson, R.M., H.A. Gabb, and M J. Sternberg, Rapid refinement of protein interfaces incorporating solvation: application to the docking problem. J Mol Biol, 1998.276(1): p. 265-85.

62. Wiehe, K., et al., ZDOCK and RDOCKperformance in CAPRI rounds 3,4, and 5. Proteins, 2005.60(2): p. 207-13.

63. Li, L., R. Chen, and Z. Weng, RDOCK: refinement of rigid-body protein docking predictions. Proteins, 2003. 53(3): p. 693-707.

64. Brooks, B .R., et al., CHARMM: A program for macromolecular energy, minimization, and dynamics calculations. Journal of Computational Chemistry, 1983.4: p. 187-217.

65. Neria, E., S. Fischer, and M. Karplus, Simulation of activation free energies in molecular systems. The Journal of Chemical Physics, 1996.105(5): p. 1902-1921.

66. Zhang, C, et al., Determination of atomic desolvation energies from the structures of crystallized proteins. J Mol Biol, 1997. 267(3): p. 707-26.

67. Chen, R. and Z. Weng, A novel shape complementarity scoring function for protein- protein docking. Proteins, 2003.51(3): p. 397-408.

68. Chen, R.,et al., A protein-protein docking benchmark. Proteins, 2003. 52(1): p. 88-91.

69. Press, W.H., Numerical recipes in C : the art of scientific computing. 2nd ed. 2002, Cambridge [Cambridgeshire] ; New York: Cambridge University Press, xxvi, 994 p.

70. Doerr, T.P. and Y.-K. Yu, Electrostatics of charged dielectric spheres with application to 154

biological systems. Physical Review E (Statistical, Nonlinear, and Soft Matter Physics), 2006. 73(6): p. 061902-14.

71. Rajamani, D., et al., Anchor residues in protein-protein interactions. Proc Natl Acad Sci USA, 2004.101(31): p. 11287-92.

72. Smith, G.R., M J. Sternberg, and PA. Bates, The relationship between the flexibility of proteins and their conformational states on forming protein-protein complexes with an application to protein-protein docking. J Mol Biol, 2005.347(5): p. 1077-101.

73. Mintseris, J. and Z. Weng, Optimizing protein representations with information theory. Genome Inform Ser Workshop Genome Inform, 2004.15(1): p. 160-9.

74. Fanelli, F. and S. Ferrari, Prediction of MEF2A-DNA interface by rigid body docking: a tool for fast estimation of protein mutational effects on DNA binding. J Struct Biol, 2006. 153(3): p. 278-83.

75. Bonvin, AM., Flexible protein-protein docking. Curr Opin Struct Biol, 2006.16(2): p. 194-200.

76. Chandonia, J.M. and S.E. Brenner, The impact of structural genomics: expectations and outcomes. Science, 2006.311(5759): p. 347-51.

77. Gray, J J., High-resolution protein-protein docking. Curr Opin Struct Biol, 2006.16(2): p. 183-93.

78. Chen, R. and Z. Weng, Docking unbound proteins using shape complementarity, desolvation, and electrostatics. Proteins, 2002.47(3): p. 281-94.

79. Fernandez-Recio, J., M. Totrov, and R. Abagyan, ICM-DISCO docking by global energy optimization with fully flexible side-chains. Proteins, 2003.52(1): p. 113-7.

80. Tong, W. and Z. Weng, Clustering protein-protein docking predictions. Conf Proc IEEE Eng Med Biol Soc, 2004.4: p. 2999-3002.

81. Duan, Y., B .V. Reddy, and Y .N. Kaznessis, Physicochemical and residue conservation calculations to improve the ranking of protein-protein docking solutions. Protein Sci, 2005.14(2): p. 316-28. 155

82. Wiehe, K., et al., The performance ofZDOCK and ZRANK in rounds 6-11 of CAPRI. Proteins, 2007.

83. Lorenzen, S. and Y. Zhang, Identification of near-native structures by clustering protein docking conformations. Proteins, 2007.68(1): p. 187-94.

84. Daily, M.D., et al., CAPRI rounds 3-5 reveal promising successes and future challenges for RosettaDock. Proteins, 2005.60(2): p. 181-6.

85. Mintseris, J., et al., Integrating statistical pair potentials into protein complex prediction. Proteins, 2007.

86. Wang, C, O. Schueler-Furman, and D. Baker, Improved side-chain modeling for protein- protein docking. Protein Sci, 2005.14(5): p. 1328-39.

87. Schueler-Furman, O., C. Wang, and D. Baker, Progress in protein-protein docking: atomic resolution predictions in the CAPRI experiment using RosettaDock with an improved treatment of side-chain flexibility. Proteins, 2005. 60(2): p. 187-94.

88. Zhang, C, J. Chen, and C. DeLisi, Protein-protein recognition: exploring the energy funnels near the binding sites. Proteins, 1999.34(2): p. 255-67.

89. Tovchigrechko, A. and LA. Vakser, How common is the funnel-like energy landscape in protein-protein interactions? Protein Sci, 2001.10(8): p. 1572-83.

90. Tsai, C J., et al., Folding funnels, binding funnels, and protein function. Protein Sci, 1999.8(6): p. 1181-90.

91. Andrusier, N., R. Nussinov, and H J. Wolf son, FireDock: Fast interaction refinement in molecular docking. Proteins, 2007.

92. Liang, S., et al., A simple reference state makes a significant improvement in near-native selections from structurally refined docking decoys. Proteins, 2007.

93. Janeway, C, Immunobiology 5 : the immune system in health and disease. 2001, Garland Pub.: New York. 156

Holler, P.D., et al., In vitro evolution of a T cell receptor with high affinity for peptide/MHC. Proc Natl Acad Sci USA, 2000. 97(10): p. 5387-92.

Li, Y., et al., Directed evolution of human T-cell receptors with picomolar affinities by phage display. Nat Biotechnol, 2005.23(3): p. 349-54.

Dunn, S.M., et al., Directed evolution of human T cell receptor CDR2 residues by phage display dramatically enhances affinity for cognate peptide-MHC without increasing apparent cross-reactivity. Protein Sci, 2006.15(4): p. 710-21.

Chlewicki, L.K., et al., High-affinity, peptide-specific Tcell receptors can be generated by mutations in CDR1, CDR2 or CDR3. J Mol Biol, 2005.346(1): p. 223-39.

Dufner, P., L. Jermutus, and R.R. Minter, Harnessing phage and ribosome display for antibody optimisation. Trends Biotechnol, 2006.24(11): p. 523-9.

Selzer, T., S. Albeck, and G. Schreiber, Rational design of faster associating and tighter binding protein complexes. Nat Struct Biol, 2000.7(7): p. 537-41.

Song, G., et al., Rational design of intercellular adhesion molecule-1 (ICAM-1) variants for antagonizing integrin lymphocyte function-associated antigen-1-dependent adhesion. J Biol Chem, 2006.281(8): p. 5042-9.

Joachimiak, L.A., et al., Computational design of a new hydrogen bond network and at least a 300-fold specificity switch at a protein-protein interface. J Mol Biol, 2006. 361(1): p. 195-208.

Kortemme, T., et al., Computational redesign of protein-protein interaction specificity. Nat Struct Mol Biol, 2004.11(4): p. 371-9.

Cole, D.K., et al., Computational design and crystal structure of an enhanced affinity mutant human CD8 alphaalpha coreceptor. Proteins, 2007.67(1): p. 65-74.

Clark, LA., et al., Affinity enhancement of an in vivo matured therapeutic antibody using structure-based computational design. Protein Sci, 2006.15(5): p. 949-60.

Lippow, S.M., K.D. Wittrup, and B. Tidor, Computational design of antibody-affinity improvement beyond in vivo maturation. Nat Biotechnol, 2007.25(10): p. 1171-6. 157

106. Baker, B.M., et al., Structural, biochemical, and biophysical studies of HLA-A2Ialtered peptide ligands binding to viral-peptide-specific human T-cell receptors. Cold Spring Harb Symp Quant Biol, 1999. 64: p. 235-41.

107. Davis-Harrison, R.L., F.K. Insaidoo, and B.M. Baker, Tcell receptor binding transition states and recognition of peptide/MHC. Biochemistry, 2007.46(7): p. 1840-50.

108. Garboczi, D .N., et al., Structure of the complex between human T-cell receptor, viral peptide andHLA-A2. Nature, 1996.384(6605): p. 134-41.

109. Bruccoleri, R.E., Application of Systematic Conformational Search to Protein Modeling. Molecular Simulation, 1993.10(2): p. 151 -174.

110. Bruccoleri, R.E. and M. Karplus, Prediction of the folding of short polypeptide segments by uniform conformational sampling. Biopolymers, 1987.26(1): p. 137-68.

111. Zhou, H. and Y. Zhou, Distance-scaled, finite ideal-gas reference state improves structure-derived potentials of mean force for structure selection and stability prediction. Protein Sci, 2002.11(11): p. 2714-26.

112. Garboczi, D .N., et al., Assembly, specific binding, and crystallization of a human TCR- alphabeta with an antigenic Tax peptide from human T lymphotropic virus type 1 and the class I MHC molecule HLA-A2. J Immunol, 1996.157(12): p. 5403-10.

113. Garboczi, D.N., D.T. Hung, and D.C. Wiley, HLA-A2-peptide complexes: refolding and crystallization of molecules expressed in Escherichia coli and complexed with single antigenic peptides. Proc Natl Acad Sci USA, 1992.89(8): p. 3429-33.

114. Ding, Y.H., et al., Four A6-TCRJpeptidelHLA-A2 structures that generate very different T cell signals are nearly identical. Immunity, 1999.11(1): p. 45-56.

115. Looger, L.L., et al., Computational design of receptor and sensor proteins with novel functions. Nature, 2003. 423(6936): p. 185-90.

116. Ma, B. and R. Nussinov, Trp/Met/Phe hot spots in protein-protein interactions: potential targets in drug design. Curr Top Med Chem, 2007.7(10): p. 999-1005.

117. Cole, D.K., et al., Human TCR-binding affinity is governed by MHC class restriction. J 158

Immunol, 2007.178(9): p. 5727-34.

118. Moza, B., et al., Long-range cooperative binding effects in a T cell receptor variable domain. Proc Natl Acad Sci USA, 2006.103(26): p. 9867-72.

119. Laugel, B., et al., Design of soluble recombinant T cell receptors for antigen targeting and Tcell inhibition. J Biol Chem, 2005. 280(3): p. 1882-92.

120. Kortemme, T. and D. Baker, Computational design of protein-protein interactions. Curr Opin Chem Biol, 2004.8(1): p. 91-7.

121. Wells, JA.,Additivity of mutational effects in proteins. Biochemistry, 1990.29(37): p. 8509-17.

122. Reichmann, D., et al., The molecular architecture of protein-protein binding sites. Curr Opin Struct Biol, 2007.17(1): p. 67-76.

123. Wu, L.C., et al., Two-step binding mechanism for T-cell receptor recognition of peptide MHC. Nature, 2002.418(6897): p. 552-6.

124. Ashworth, J., et al., Computational redesign of endonuclease DNA binding and cleavage specificity. Nature, 2006.441(7093): p. 656-9.

125. Bowman, M.R., et al., Identification and structural analysis of residues in the VI region ofCD4 involved in interaction with human immunodeficiency virus envelope glycoprotein gpl20 and class II major histocompatibility complex molecules. Proc Natl Acad Sci USA, 1990.87(22): p. 9052-6.

126. Moebius, U., et al., The human immunodeficiency virus gp!20 binding site on CD4: delineation by quantitative equilibrium and kinetic binding studies of mutants in conjunction with a high-resolution CD4 atomic structure. J Exp Med, 1992.176(2): p. 507-17.

127. Ashkenazi, A., et al., Mapping the CD4 binding site for human immunodeficiency virus by alanine-scanning mutagenesis. Proc Natl Acad Sci USA, 1990.87(18): p. 7150-4.

128. Daar, E.S., et al., High concentrations of recombinant soluble CD4 are required to neutralize primary human immunodeficiency virus type I isolates. Proc Natl Acad Sci U 159

S A, 1990.87(17): p. 6574-8.

129. Kwong, P.D., et al., Structure of an HIV gp!20 envelope glycoprotein in complex with the CD4 receptor and a neutralizing human antibody. Nature, 1998.393(6686): p. 648-59.

130. Kwong, P.D., et al., Structures ofHIV-1 gp!20 envelope glycoproteins from laboratory- adapted and primary isolates. Structure, 2000.8(12): p. 1329-39.

131. Zhang, W., et al., Antibody 17b binding at the coreceptor site weakens the kinetics of the interaction of envelope glycoprotein gp!20 with CD4. Biochemistry, 2001.40(6): p. 1662-70.

132. Myszka, D.G., et al., Energetics of the HIV gp!20-CD4 binding reaction. Proc Natl Acad Sci USA, 2000.97(16): p. 9026-31.

133. Martin-Garcia, J., et al., Interaction with CD4 and antibodies to CD4-induced epitopes of the envelope gpHOfrom a microglial cell-adapted human immunodeficiency virus type 1 isolate. J Virol, 2005. 79(11): p. 6703-13.

134. Wu, H., et al., Kinetic and structural analysis of mutant CD4 receptors that are defective in HlVgpUO binding. Proc Natl Acad Sci USA, 1996.93(26): p. 15030-5.

135. Zhou, T., et al., Structural definition of a conserved neutralization epitope on HIV-I gp!20. Nature, 2007.445(7129): p. 732-7.

136. Vita, C, et al., Rational engineering of a miniprotein that reproduces the core of the CD4 site interacting with HIV-I envelope glycoprotein. Proc Natl Acad Sci U S A, 1999. 96(23): p. 13091-6.

137. Huang, C .C., et al., Scorpion-toxin mimics of CD4 in complex with human immunodeficiency virus gpl20 crystal structures, molecular mimicry, and neutralization breadth. Structure, 2005.13(5): p. 755-68.

138. Tsui, P., et al., An efficient phage plaque screen for the random mutational analysis of the interaction ofHIV-1 gp!20 with human CD4. J Biol Chem, 1992. 267(13): p. 9361-7.

139. Krykbaev, R., J. McKeating, and I. Jones, Mutant CD4 molecules with improved binding to HIV envelope protein gp!20 selected by phage display. Virology, 1997.234(2): p. 160

196-202.

140. Osburne, M.S., et al., Production of secreted, soluble human two-domain CD4 protein in Escherichia coli. J Immunol Methods, 1999.224(1-2): p. 19-24.

141. Li, Y., et al., Molecular characterization of human immunodeficiency virus type 1 cloned directly from uncultured human brain tissue: identification of replication-competent and -defective viral genomes. J Virol, 1991.65(8): p. 3973-85.

142. Myszka, D.G., Improving biosensor analysis. J Mol Recognit, 1999.12(5): p. 279-84.

143. Fellouse, FA., C. Wiesmann, and S.S. Sidhu, Synthetic antibodies from afour-amino- acid code: a dominant role for tyrosine in antigen recognition. Proc Natl Acad Sci U S A, 2004.101(34): p. 12467-72.

144. Malek, S., et al., X-ray crystal structure of an IkappaBbeta x NF-kappaB p65 homodimer complex. J Biol Chem, 2003.278(25): p. 23094-100.

145. Mechulam, A., et al., Highly conserved beta!6/betal7 beta-hairpin structure in human immunodeficiency virus type 1 YU2 gpl20 is critical for CCR5 binding. J Mol Med, 2005.83(7): p. 542-52.

146. Sami, M., et al., Crystal structures of high affinity human T-cell receptors bound to peptide major histocompatibility complex reveal native diagonal binding geometry. Protein Eng Des Sel, 2007. 20(8): p. 397-403. 161

CURRICULUM VITAE

Brian Gregory Pierce [email protected] http://zlab .bu .edu/bpierce 26 Willow Street, Apt. 3 Newton Centre, MA 02459 (617) 320-3715

Education Boston University, Boston, Massachusetts Bioinformatics Ph.D. Program, August 2002-present. Anticipated completion: December 2007. Relevant Coursework: physical biochemistry, computational genomics, biological sequence analysis, molecular biology, biological databases Cumulative GPA: 3.75 on a 4.0 scale Thesis Advisor: Zhiping Weng

Duke University, Durham, North Carolina B.S. Physics, A.B. Computer Science, May 2000 Relevant Coursework: program design, algorithms, operating systems, computational physics, intermediate mechanics, electricity and magnetism, general chemistry, thermal physics, quantum physics, calculus, differential equations, and linear algebra Dean's List, Fall 1998, Spring 1999, Fall 1999, Spring 2000 Major GPA: 3.73 on a 4.0 scale

Work and Research Experience Protein Design Researcher. Boston University, Thesis Work, May 2005-present. • Utilized bioinformatics approaches to develop a new scoring function for protein affinity optimization • Predicted and verified several T cell receptor (TCR) point mutations that improved binding affinity; achieved 98-fold binding affinity improvement for peptide-MHC • Performed Biacore SPR, protein expression, refolding, and purification to verify predictions • Currently applying protein design protocol for modifying TCR specificity, and design of CD4-gpl20 interface (ongoing work)

Protein-Protein Docking Researcher. Boston University, Thesis Work, June 2003- present. • Implemented protein docking program M-ZDOCK for predicting structure of Cn symmetric multimers • Developed protein docking scoring function ZRANK for reranking initial-stage docking predictions • Participated in several rounds of group's effort in CAPRI international protein docking experiment 162

Software Testing Engineer. Lucent Software, Cambridge, MA. October 2000-July 2002. • Tested billing software dealing with placing orders and customer information in databases • Headed the system test group's automation effort • Designed and implemented a C++ and XML based automated testing program, and wrote a series of automated test plans for it to run

Program Developer/Researcher. General Atomics, La Jolla, California. Summer/Fall 1999. Selected for National Undergraduate Fellowship Program in Plasma Science and Fusion Engineering Developed and integrated a program library for running simulations of fusion plasmas on massively parallel computers, under the supervision of Dr. Yuri Omelchenko Wrote test simulations and profiled them on parallel clusters at Duke and at General Atomics Physics Researcher. Triangle Universities Nuclear Laboratory, Durham, NC. Summer 1998-Spring 2000. • Worked with the lab's director and two undergraduates on an experiment regarding the solar fusion reaction, 12C(a, y )160 , performed at the Duke Free Electron Laser Laboratory • Tested plastic track detectors for various alpha particle energies and incident angles, and determined the setup for the free electron laser run based on these results Awards and Accomplishments

• Recipient, National Undergraduate Fellowship in Plasma Physics, 1998 • Recipient, Boston University Graduate Research Scholarship, 2003-2007 • Developed two protein modeling programs that are licenced by Accelrys, Inc. Teaching Experience

• Lecturer and Teaching Assistant. BE 703, BME Numerical Methods, Boston University. Fall 2004, Spring 2004, Fall 2005, Fall 2007. • Teaching Assistant. CPS 6, Introductory Computer Programming, Duke University. Fall 1998, Spring 1999.

Publications and Conference Presentations Pierce B, Hourai Y, Weng Z. (Submitted) "ZDOCK 2.3.1 and ZDOCK 3.0.1: Using a New 3D Convolution Library to Enhance Docking Efficiency".

Haidar JN*, Pierce B*, Yu Y,Tong W, Li M, Weng Z. (Submitted) "Structure-Based Design of a T Cell Receptor Leads to Nearly 100-Fold Improvement in Binding Affinity for pepMHC". *joint first authors 163

Pierce B, Weng Z (Accepted). "A Combination of Rescoring and Refinement Significantly Improves Protein Docking Performance". Proteins. Bastas G, Sompuram SR, Pierce B, Vani K, Bogen SA (2007) "Bioinformatic Requirements for Protein Database Searching Using Predicted Epitopes from Disease- Associated Antibodies". Mol Cell Proteomics, In Press. Wiehe K, Pierce B, Tong W, Hwang H, Mintseris J, Weng Z (2007) "The Performance of ZDOCK and ZRANK in Rounds 6-11 of CAPRI". Proteins 69(4): 719-725. Mintseris J, Pierce B, Wiehe K, Anderson R, Chen R, and Weng Z (2007) "Integrating Statistical Pair Potentials into Protein Complex Prediction". Proteins 69(3): 511-520.

Pierce B, Weng Z (2007). "ZRANK: Reranking Protein Docking Predictions with an Optimized Energy Function". Proteins 67(4): 1078-1086.

Pierce B, Phillips AT, Weng Z (2007). "Structure Prediction of Protein Complexes". In Xu Y, Xu D & Liang J. (Eds.), Computational Methods for Protein Structure Prediction and Modeling Volume 2: Structure Prediction, (pp. 109-134). New York: Springer.

Pierce B, Tong W, Weng Z (2005). "M-ZDOCK: A Grid-based Approach for Cn Symmetric Multimer Docking". Bioinformatics 21(8), 1472-1476. Pierce B, Weng Z (2004). "Determining an Optimal Scoring Function for Protein Complex Predictions". Fourth International Workshop on Bioinformatics and Systems Biology, Kyoto. Pierce B, Omelchenko Y (1999). "An Object-Based, Parallel Framework for Particle-in- Cell Simulations in Fortran and MPI". APS 41s' annual meeting of the Division of Plasma Physics, Seattle.

Pierce B, Ram A, Sorrenti M, Tornow W (1998). "The 16<9(y ,a)12C Reaction at Stellar Energies". TUNL Progress ReportXXXVII.

Skills Rosetta, CHARMM, BLAST, C++, Perl, Java, Pascal, Fortran, SQL, Biacore, protein refolding and purification, parallel programming, MPI, Monte Carlo methods, object- oriented programming, Windows, Unix, Linux, MacOS, Latex, RCS, GDB, MS Office, PowerPoint, Adobe Illustrator.