THERMODYNAMIC MODELING OF PROTEIN INTERACTIONS AND

PHASE BEHAVIOR

by

Leigh Jian Quang

A thesis submitted to the Faculty of the University of Delaware in partial

fulfillment of the requirements for the degree of Master of Chemical Engineering

Spring 2012

Copyright 2012 Leigh Jian Quang

All Rights Reserved
 THERMODYNAMIC MODELING OF PROTEIN INTERACTIONS AND

PHASE BEHAVIOR

by
Leigh Jian Quang

Approved: __________________________________________________________
Abraham M. Lenhoff, Ph.D.
Professor in charge of thesis on behalf of the Advisory Committee

Approved: __________________________________________________________
Stanley I. Sandler, Ph.D.
Professor in charge of thesis on behalf of the Advisory Committee

Approved: __________________________________________________________
Norman J. Wagner, Ph.D.
Chair of the Department of Chemical Engineering

Approved: __________________________________________________________
Babatunde A. Ogunnaike, Ph.D.
Interim Dean of the College of Engineering
Approved: __________________________________________________________
Charles G. Riordan, Ph.D.
Vice Provost for Graduate and Professional Education
 ACKNOWLEDGMENTS

There were many people who have helped me during my time at the

University of Delaware. My experience as a graduate student was perhaps the most

humbling period of my life, but the support from my peers was instrumental in making

this work possible. I am grateful and honored to have had the opportunity to

collaborate with so many diligent and exceptional people in the chemical engineering

department. I will always remember the unique camaraderie among everyone in the

graduate student community.

First, I would like to thank the members of the Lenhoff and Sandler

groups. I would like to acknowledge my advisers Professor Abraham Lenhoff and

Professor Stan Sandler for their critique of this thesis. I would like to thank Dr.

Rachael Lewus for her helpful experimental insights in protein phase behavior. She

was able to answer many of my conceptual questions during the early stages of my

work, which helped me tremendously in getting the big picture of my project. I also

thank Dr. Jaeoon Chang for discussions on the patch-antipatch model for proteins and

general aspects of molecular simulation. His down-to-earth personality and innate

ability to explain complicated concepts in simple terms helped me tremendously in

learning the technical aspects of simulation. I would like to thank Dr. Chris Gillespie

who provided me his source code for the atomistic B22 calculations. I like to

acknowledge the support from Steve Traylor, Anvar Samadzoda, and Dr. Kelley

Kearns. They were always there to discuss not only the technical aspects of my

research, but also a wide variety of nontechnical topics ranging from sports to politics.

iv
It was always refreshing to be able to talk about things outside of research.

Sometimes it is easy to forget that there is a world beyond graduate school. Russ

Burnett and Vinit Choudhary were both helpful for their constructive feedback on my

work. Although my project was outside of their area of expertise, they were still able

to provide me with some of their perspectives for my work.

This work would not have been possible without support from outside of

the chemical engineering department. I am grateful for the National Science

Foundation for providing the funding for this project. I also thank Dr. Sandeep Patel

in the Department of Chemistry for providing the computational resources that

allowed me to perform the B22 calculations.

Finally, I would like to acknowledge the support from my fellow

classmates at the University of Delaware. I would like to give special thanks to Nima

Nikbin and Stephen Edie for putting up with me during my last two months in

Delaware (thanks guys I really appreciate it!).

v
 TABLE OF CONTENTS

LIST OF TABLES ........................................................................................................ ix

LIST OF FIGURES ....................................................................................................... xi
ABSTRACT ................................................................................................................. xv

Chapter

1 INTRODUCTION .............................................................................................. 1

1.1 The Protein Phase Diagram ....................................................................... 2

1.2 Protein-Protein Interactions ....................................................................... 4
1.3 Theoretical and Simulation Studies ........................................................... 7
1.4 The Osmotic Second Virial Coefficient: B22 ........................................... 12
1.5 Objective and Thesis Outline .................................................................. 14

2 CONTINUUM THERMODYNAMIC MODELS OF PROTEIN

INTERACTIONS AND PHASE BEHAVIOR ................................................ 17

2.1 Motivation and Goal ................................................................................ 17

2.2 Modeling Structure .................................................................................. 20
2.3 Flory-Huggins Model .............................................................................. 22

2.3.1 Theory ......................................................................................... 22

2.3.2 Results ......................................................................................... 24

2.4 Haas-Drenth Model ................................................................................. 31

2.4.1 Theory ......................................................................................... 31

2.4.2 Results ......................................................................................... 32

2.5 Osmotic Virial Equation .......................................................................... 38

2.5.1 Theory ......................................................................................... 38

2.5.2 Results ......................................................................................... 43

2.6 Discussion................................................................................................ 54
2.7 Conclusions ............................................................................................. 55

vi
 3 PATCH-ANTIPATCH MODEL OF PROTEINS AND THE
CALCULATION OF B22 ................................................................................. 56

3.1 Introduction ............................................................................................. 56

3.1.1 Review of Patch Models.............................................................. 56

3.1.2 The “Patch-Antipatch” Model ..................................................... 59
3.1.3 Objective ..................................................................................... 61

3.2 Theory and Methods ................................................................................ 62

3.2.1 Determining “Patch-Antipatch” Pairs ......................................... 62

3.2.2 Interaction Energies ..................................................................... 63

3.2.2.1 Short-Range Interactions .............................................. 63

3.2.2.2 Electrostatic Interactions .............................................. 65

3.2.3 Calculation of B22 ........................................................................ 67

3.3 Results ..................................................................................................... 72

3.3.1 Identification of Patch-Antipatch Pairs ....................................... 72

3.3.2 Calculation of B22 ........................................................................ 87

3.4 Discussion.............................................................................................. 106

3.5 Conclusions ........................................................................................... 109

4 CONCLUSIONS AND RECOMMENDATIONS ......................................... 111

4.1 Conclusions ........................................................................................... 111

4.1.1 Continuum Thermodynamic Models......................................... 111

4.1.2 Patch-Antipatch Model of Proteins and the Calculation of
B22 .............................................................................................. 112

4.2 Recommendations and Future Directions ............................................. 114

4.2.1 The Calculation of B22 ............................................................... 114

4.2.2 Molecular Simulation of Protein Phase Behavior ..................... 115

REFERENCES ........................................................................................................... 117

vii
Appendix

A DERIVATION OF LIQUID-LIQUID EQUILBIRIUM FROM THE

OSMOTIC VIRIAL EQUATION .................................................................. 131
B CALCULATION OF THE OSMOTIC THIRD VIRIAL
COEFFICIENT B222 ....................................................................................... 136

viii
 LIST OF TABLES

Table 2.1: Physical properties of ribonuclease A used in the continuum

models...................................................................................................... 20

Table 2.2: Physical properties of water used in the continuum models. .................. 20

Table 3.1: Proteins studied for patch-antipatch analysis and their physical
properties. ................................................................................................ 62

Table 3.2: Log normal probability distribution function parameters for

lysozyme and chymosin B estimated from the 106 sampled
configurations. ......................................................................................... 75

Table 3.3: Absolute and relative frequencies of lysozyme well depths for
different sampling. ................................................................................... 78

Table 3.4: Absolute and relative frequencies of chymosin B well depths for
different sampling. ................................................................................... 79

Table 3.5: Ten most attractive angular configurations for short-range

interactions of lysozyme identified from 106 randomly sampled
orientations. ............................................................................................. 81

Table 3.6: Thirty-five most attractive angular configurations for short-range

interactions of chymosin B identified from 106 randomly sampled
orientations. ............................................................................................. 82

Table 3.7: Refined orientations for lysozyme identified from the local
sampling in 0.10 radian around the central orientation in Table
3.4. The resulting angular configurations are significantly more
attractive than the originally sampled orientations. ................................. 85

Table 3.8: Refined orientations for patch-antipatch pairs for chymosin B

identified from the local sampling in 0.10 radian around the
central orientation in Table 3.5. ............................................................... 86

ix
Table 3.9: B22 calculated from 106 randomly sampled configurations based on
excluded volume contribution and both excluded volume and
short-range attraction. The  value is the sphere equivalent
diameter determined from the empirical correlation of Neal and
Lenhoff (141). The error in the Monte Carlo estimate is calculated
from equation 3.16................................................................................... 88

Table 3.10: Contributions to B22 from individual patch-antipatch pairs for

lysozyme based on short-range non-electrostatic interaction
energies alone and with electrostatics at pH 7. The contributions
were determined by integrating within ±0.10 radian around the
central orientation using the DCUHRE integration routine. ................... 98

Table 3.11: B22 contributions from individual patch-antipatch pairs for

chymosin B based on short-range non-electrostatic interaction
energies alone and with electrostatic interactions at pH 5. The
contributions were determined by integrating within 0.10 radian
around the central orientation using the DCUHRE integration
routine. ..................................................................................................... 99

Table 3.12: Background B22 for lysozyme based on short-range non-

electrostatic interactions alone and with addition of electrostatics
at pH 7. Patch-antipatch pairs with  < -20 kT were excluded in
the Monte Carlo integration................................................................... 102

Table 3.13: Background B22 for chymosin B based on short-range interactions

alone and with addition of electrostatics at pH 5. Patch-antipatch
pairs with  < -20 kT were excluded in the Monte Carlo
integration. ............................................................................................. 102

Table 3.14: Background B22 for lysozyme based on short-range non-

electrostatics interactions alone and with addition of electrostatics
at pH 7. Patch-antipatch pairs with  < -7 kT were excluded in the
Monte Carlo integration. ....................................................................... 106

Table 3.15: Background B22 for chymosin B based on short-range non-

electrostatics interactions alone and with addition of electrostatics
at pH 5. Patch-antipatch pairs with  < -7 kT were excluded in the
Monte Carlo integration. ....................................................................... 106

x
 LIST OF FIGURES

Figure 1.1: Theoretical colloidal phase diagram adapted from Foffi et al. (6). .......... 3

Figure 1.2: Cartoon of simple isotropic model of proteins. Protein molecules

are represented as perfect spheres of diameter and the
interactions depend only on the center-to-center distance r. .................... 9

Figure 1.3: Schematic of the conceptual path from molecular structure to

thermodynamic solution properties of proteins, which includes the
osmotic second virial coefficient B22 and phase behavior. ..................... 16

Figure 2.1: () Binodal, () spinodal, and () B22 data for ribonuclease A in
ammonium sulfate system at 23°C, pH 7. The dotted rectangle
encloses the region where B22 and phase behavior data overlap.
Results were taken from Dumetz et al. (45, 85). .................................... 19

Figure 2.2: Schematic of the modeling pathways used to relate B22 and phase
behavior with the continuum models. .................................................... 21

Figure 2.3: Comparison of B22 predictions from the () Flory-Huggins model
with () experimental B22 data. .............................................................. 25

Figure 2.4: Critical point predicted by the Flory-Huggins model compared

with () experimental binodal data. The critical point is located at
a salt concentration of 1.22 M. The location of the critical point
demonstrates that the equilibrium phase boundary is located at
higher salt concentrations. ...................................................................... 26

Figure 2.5: B22 predictions from Flory-Huggins model calculated from values
of () m=503 and () m=615 compared with original predictions
from () m= 559 and () experimental B22 values. ............................... 27

Figure 2.6: Predicted critical point from the Flory-Huggins model for ()
m=503, ( ) m=559, and () m=603 relative to the ()
experimental binodal boundary. ............................................................. 29

Figure 2.7: Predicted binodal boundaries from the Flory-Huggins model for
() m=279 compared with () experimental results. ............................ 30

xi
Figure 2.8: Comparison of B22 predictions from the () Haas-Drenth model
with () experimental B22 data. .............................................................. 33

Figure 2.9: Phase behavior predictions from experimental B22 values with the
Haas-Drenth model compared with () experimental binodal
data. The Haas-Drenth model does predict () the binodal
boundary to exist within the overlap region. The predicted ()
critical point occurs at an ammonium sulfate concentration of 1.02
M and a protein concentration of 177 mg/ml. ........................................ 34

Figure 2.10: B22 predictions from the Haas-Drenth model calculated from
values of () m=503 and () m=615 compared with original
predictions from () m= 559 and () experimental B22 values. ........... 35

Figure 2.11: Predicted binodal boundaries from the Haas-Drenth model for
() () m=279 compared with the original predictions from
m=559 and the () experimental results. ............................................... 37

Figure 2.12: Plot of the Yukawa potential in reduced units for b* values of (–)
5, (–) 10, (–) 20, and (–) 30. The (–) hard-sphere repulsion occurs
at r*=1.0. Increasing b* corresponds to a decrease in the range of
attraction. ................................................................................................ 40

Figure 2.13: Comparison of the (–) 140-35 Lennard-Jones potential with the (–
) Yukawa potential for b* = 22. ............................................................. 41

Figure 2.14: Computed B222 from the Yukawa potential for b* values of ()
25, () 35, and () 45. .......................................................................... 44

Figure 2.15: Phase behavior predictions from the osmotic virial equation based
on B222 calculated from the Yukawa potential. Predictions were
made for b* values of () 25, () 35, and () 45 and compared
with the () experimental binodal data. ................................................. 45

Figure 2.16: Binodal predictions based on b*=35 compared with the ()
experimental binodal data. Osmotic virial predictions were made
from the experimental B22 values with an assumed error of ()
+2 10-4 mol ml/g2 and () -2 10-4 mol ml/g2. The results are
compared with the predictions from the () original B22 data set. ........ 47

xii
Figure 2.17: Binodal predictions based on b*=45 compared with the ()
experimental binodal data. Osmotic virial predictions were made
from the experimental B22 values with an assumed error of ()
+2 10-4 mol ml/g2 and () -2 10-4 mol ml/g2. The results are
compared with the predictions from the () original B22 data set. ........ 48

Figure 2.18: B222 computed from the square well potential for values of ()
1.05, () 1.20, () 1.50, and () 2.10. ................................................. 50

Figure 2.19: B222 computed from the 140-35 Lennard-Jones potential. ..................... 51

Figure 2.20: Plot of the ten Wolde-Frenkel potential for  values of (–) 10, (–)
20, (–) 30, and (–) 50. ............................................................................ 52

Figure 2.21: B222 computed from the ten Wolde-Frenkel potential for  values
of () 10, () 20, and () 30. ............................................................... 53

Figure 3.1: A cartoon of a “patch-antipatch” model of protein molecules. The

“patch” is colored in blue and the corresponding “antipatch” is
colored in red. Specific interactions occur only between unique
blue and red colored regions, which depend on the angles of
alignment m and n. .............................................................................. 61

Figure 3.2: Histograms of the distribution of the short-ranged interaction well

minima for 104 randomly sampled configurations for ()
lysozyme and () chymosin B. The inset histograms for each
protein are meant to magnify the tails of the distributions. .................... 73

Figure 3.3: Histograms of the distribution of the short-ranged interaction well

minima for 105 randomly sampled configurations for ()
lysozyme and () chymosin B. .............................................................. 76

Figure 3.4: Histograms of the distribution of the short-ranged interaction well

minima for 106 randomly sampled configurations for ()
lysozyme and () chymosin B. The largest well depth identified
was on the order of -20 kT. ..................................................................... 77

Figure 3.5: Comparison of relative frequencies of () lysozyme and ()

chymosin B absolute well depths with respective fits from (–) log
normal probability distribution function. The fits are based on
parameter values of  = 0.975,  = 0.652 for lysozyme and  =
1.067,  = 0.628 for chymosin B............................................................ 80

xiii
Figure 3.6: Well depth as a function of the angles for the orientation listed in
entry 1 of Table 3.5. The largest change occurs when is
decreased by -0.02 radian, which indicates that the originally
sampled orientation is not the optimum alignment. ............................... 84

Figure 3.7: Histogram of the computed Iin for the 106 randomly sampled
configurations for lysozyme. The inset shows an enlarged view
of the high-Iin tail of the distribution. ..................................................... 89

Figure 3.8: Histogram of the computed Iin for the 106 randomly sampled
configurations for chymosin B. The inset shows an enlarged view
of the high-Iin tail of the distribution. ..................................................... 90

Figure 3.9: Iconfig computed from the DCUHRE routine as a function of  for
lysozyme patch-antipatch pair 4 in Table 3.7. Iconfig increases
monotonically as  increases due to the increase in the
hypervolume of the integration. ............................................................. 92

Figure 3.10: Iconfig normalized by the volume of integration v0 as a function of

 for lysozyme patch-antipatch pair 4 in Table 3.7. The
normalized integral decreases as  is increased. .................................... 93

Figure 3.11: Plot of the localized configuration integration for lysozyme as a

function of the total well depth. Iconfig was computed based on (+)
short-range interactions alone and () short-range interactions
with electrostatics. Integration was performed within the limits of
0.10 radian around the central orientation of each patch-
antipatch configuration using the DCUHRE routine. The
regressed curve is F(x)=0.0200 exp(-0.805x), R2=0.9941...................... 95

Figure 3.12: Plot of the localized configuration integration for chymosin B as a

function of the total well depth. Iconfig was computed based on ( )
short-range interactions alone and () short-range interactions
with electrostatics. Integration was performed within the limits of
0.10 radian around the central orientation of the configurations.
The regressed curve is F(x)=0.0195 exp(-0.770x), R2=0.9806. ............. 96

xiv
 ABSTRACT

Protein phase behavior encompasses the formation of dense phases, which

include amorphous aggregates, gels, dense liquids, and crystals. The major solution

variables that dictate the type of dense phase that is formed are pH, temperature, type

of precipitant, precipitant concentration, and protein concentration. Because of the

large parameter space and rich variety of phase transitions possible, protein phase

behavior is a complex phenomenon. Fundamentally, macroscopic phase transitions

are governed by the molecular interactions between proteins in solution. One

promising way of quantifying protein-protein interactions and relating them to phase

behavior is through the osmotic second virial coefficient B22, a dilute-solution property

that characterizes two-particle interactions. The relationship of B22 to overall phase

behavior of proteins is explored in this work.

The goal of this thesis is to quantitatively relate protein-protein

interactions to protein phase diagrams in order to develop predictive models of phase

behavior under different solution conditions. A continuum-level approach is used

initially to relate experimental B22 data and phase diagrams of proteins by appealing to

existing thermodynamic models, with the expectation that a simple continuum model

could provide a useful mechanistic framework for predicting protein phase behavior.

The first approach attempted was to relate protein interactions and phase behavior

within the Flory-Huggins theory of polymer solutions. The second approach utilized

the model of Haas and Drenth, which is based on the free energy of mixing for hard

spheres. Finally, phase equilibrium was predicted from virial coefficients using the

xv
osmotic virial equation. A qualitative relationship was found between B22 and phase

behavior from these continuum models; however, quantitative agreement could not be

obtained. The isotropic assumption shared among these models in addition to the

orientationally-averaged nature of B22 suggests that the anisotropic character of protein

interactions cannot be neglected, demonstrating the need for more detailed molecular-

level models.

The role of anisotropy in protein interactions was explored through

analysis of “patch-antipatch” pairs in the computation of B22 in atomistic detail.

Patch-antipatch pairs represent highly attractive orientations resulting from geometric

complementarity between protein surfaces. Previous work used simple Monte Carlo

integration for the calculation of B22 from atomistic models of proteins. However, the

presence of patch-antipatch pairs led to significant numerical concerns. These

concerns warranted a reexamination of the numerical methods for computing B22.

A hybrid Monte Carlo/patch integration approach is utilized to calculate

B22 for lysozyme and chymosin B. This method involves a combination of numerical

integration techniques in an attempt to obtain better convergence in predicting B22.

The overall B22 for the proteins studied was separated into three components:

contributions from the excluded volume, from the patch-antipatch pairs, and from
background configurations. The excluded volume component was found to be

adequately determined using simple Monte Carlo integration. The contributions from

individual patch-antipatch pairs were accounted for by carefully integrating the

subregions of the configuration space occupied by these pairs using a globally

adaptive integration routine. The background component to B22 was also calculated by

xvi
simple Monte Carlo integration in which the regions of the configuration space

occupied by the patch-antipatch pairs were excluded.

The calculations performed that account for the full protein structure

emphasize the importance of several features of protein interactions. First, the

difference in the interaction behavior of the two proteins studied was found to be

largely attributed to the charge anisotropy of patch-antipatch pairs. However, the

relation of the results to experimental data is limited by the omission of accounting for

the specific hydration of proteins. Hydration effects are known to affect, and usually

attenuate, patch-antipatch configurations, and therefore would be expected to

significantly impact the accurate prediction of B22. Classical colloidal as well as

atomistic models that omit these important features are inadequate in providing a

quantitative representation of protein interactions for a wide range of solution

conditions.

xvii
 Chapter 1

INTRODUCTION

Protein phase behavior refers to the appearance of various condensed

phases that proteins can form in solution. It encompasses the formation of amorphous

aggregates, gels, dense liquids, and crystals. The types of dense phases formed are

sensitive to the solution conditions, which include pH, temperature, type of

precipitant, precipitant concentration, and protein concentration. Because of the large

parameter space and rich variety of phase transitions possible, protein phase behavior

is a complex phenomenon.

The phase behavior of proteins plays a key role in many

biopharmaceutical processes. Undesired phase separation of protein therapeutics can

occur during processing and storage, which can raise serious efficacy and safety

concerns. However, phase separation may be desired in other situations, and

operations such as precipitation can be employed in downstream protein purification

processes because of their low cost. One specific form of phase separation of wide

interest is protein crystallization, which is a prerequisite step for determining protein

structure by x-ray diffraction; however, the growth of high-quality crystals is often the

bottleneck. Crystallization is also of interest in the pharmaceutical industry because it

offers an advantageous way for delivery of doses of highly concentrated protein

therapeutics. Phase separation of proteins can have serious medical implications as

well. The onset of neurodegenerative diseases such as Alzheimer’s disease has been

attributed to the formation of fibrillar protein aggregates called amyloids.

1
 Understanding the conditions and mechanisms that lead to formation of

protein dense phases from solution is therefore crucial to predicting and ultimately

controlling phase transitions in these diverse systems. Such knowledge will be

invaluable in a wide variety of applications. This work aims to contribute towards this

endeavor by utilizing and evaluating quantitative models that construct the pathway

from molecular structure to the thermodynamic solutions properties of proteins.

1.1 The Protein Phase Diagram

Early experimental studies of liquid-liquid phase separation of globular

proteins have led to interpretion of protein phase behavior within the framework of

colloidal systems. Ishimoto and Tanaka first presented evidence of temperature-

induced liquid-liquid phase separation of aqueous lysozyme solutions (1). These

results were questioned by Phillies (2), but Taratuta et al. confirmed them by

performing extensive cloud-point temperature measurements at different pH values,

salt types, and salt concentrations (3). It was later observed that the liquid-liquid

phase separation was actually metastable with respect to the solid-liquid transition (4),

behavior characteristic of colloidal particles experiencing short-ranged attractions (5,

6). Similar phase behavior was also observed experimentally for several different -
crystallins (7–10). The similarities in the metastability of the liquid-liquid phase

separation for both proteins and colloids suggest that protein phase behavior follows

the same physics as the phase transitions of colloids.

The phase diagram of proteins can be theoretically interpreted within the

context of the theoretical phase diagram for colloidal systems experiencing short-

ranged interactions (5, 6) (Figure 1.1). The colloidal phase diagram is most directly

compared to experimental protein phase behavior results when presented in a two-

2
dimensional plane where the colloid concentration is plotted on the ordinate and the

reduced temperature on the abscissa. The direction of decreasing temperature

corresponds to increasing interparticle attraction, a trend that is qualitatively analogous

to increasing precipitant concentration under salting-out conditions.

Figure 1.1: Theoretical colloidal phase diagram adapted from Foffi et al. (6).

The main feature of this phase diagram is a solid-liquid equilibrium

region, which corresponds to protein crystals in equilibrium with a protein-lean

supernatant fluid. Within this region lies the metastable liquid-liquid coexistence

envelope (bounded by the binodal curve) in which the protein partitions into a dilute

and more concentrated liquid phases. Inside this liquid-liquid coexistence region is

3
the spinodal boundary beyond which phase separation occurs instantaneously through

spinodal decomposition. Thus, knowing the relative locations of the phase boundaries

as a function of the solution conditions is important in navigating through the phase

diagram.

The metastable nature of the liquid-liquid phase separation is an important

feature of protein phase behavior. A correlation exists between the metastable liquid-

liquid coexistence and crystallization (11). Liquid-liquid phase separation has been

suggested to significantly change the kinetic pathway for crystal nucleation (12, 13).

Specifically, the free energy barrier for crystal nucleation is drastically reduced near

the liquid-liquid critical point due to critical density fluctuations. Therefore, knowing

the location of the liquid-liquid coexistence region can have implications for selecting

optimal solution conditions for protein crystallization.

1.2 Protein-Protein Interactions

Protein phase behavior is governed fundamentally by the molecular

interactions between protein molecules, which are still not completely understood.

What is known is that the interactions include contributions from van der Waals

forces, electrostatics, hydration effects (including hydrophobic interactions), and

depletion effects (where relevant). Attractive van der Waals forces arise from three

different contributions: permanent dipole-induced dipole interactions (Debye

interaction), permanent dipole-permanent dipole interactions (Keesom interaction),

and induced dipole-induced dipole interactions (London dispersion interaction) (14,

15). These attractions are complemented by long-ranged electrostatic interactions due

to the charges that some amino acid residues carry on the protein surface.

4
 Solvation forces, which are associated with water structuring around the

protein surface, include effects that may be classified as hydrophobic or as hydration

effects (15, 16). The hydrophobic effect results from the presence of nonpolar patches

on the surface, with which water molecules are unable to form hydrogen bonds. To

minimize the free energy, the nonpolar regions associate with one another, driving

water molecules away from the surface to more extensive formation of hydrogen

bonds in the bulk of the solution. Hydration effects may occur in hydrophilic regions

on the solute molecules, where hydrogen bonding with adjacent water molecules may

result in a steric and hence effectively repulsive barrier to association with other solute

molecules. Solvation forces are still poorly understood and a quantitative explanation

is presently lacking.

Additional depletion attraction is induced due to the osmotic pressure

gradient caused by the addition of nonadsorbing polymers such as polyethylene glycol

(PEG) (17). Depletion interactions are a result of an entropic effect; when two protein

molecules are in sufficient proximity, the polymer-excluded volumes of the molecules

overlap. Consequently, the polymer molecules cannot penetrate into the space

between the protein molecules, resulting in effective attraction due to the osmotic

pressure of the polymer in the bulk solution.

In addition to the different contributions to protein-protein interactions, an

important feature is also their anisotropy, which is a consequence of the nonspherical

shape of the protein molecule and the heterogeneous properties of the protein surface.

The shape anisotropy of proteins has a profound effect on the van der Waals

attractions. Computations have shown that shape anisotropy has an appreciable effect

5
on the magnitude and orientational distribution of van der Waals interaction energies

compared to calculations based on the ideal sphere approximation (18).

There is a complex interplay among the different forces that govern

protein-protein interactions and solution conditions. For example, at low salt

concentrations protein interactions are dominated by long-ranged electrostatic forces,

which are usually repulsive. At high concentrations of salt, electrostatic forces are

screened and short-ranged van der Waals forces and hydrophobic interactions tend to

drive the precipitation of protein from solution. The phenomenon where proteins

become less soluble as more salt is added is known as salting-out. One explanation

for this behavior is that the salt ions alter the hydrogen bonding network of the layer of

water that shields the protein surface hydration layer (19). Consequently, protein

molecules interact less with water, resulting in an increase in the protein-protein

interactions. Different ions have been found to have varying impacts on protein phase

behavior. The salting-out effectiveness of different ions is reflected in the Hofmeister

series (20), which is an empirical ranking of the ability of different ions to disrupt the

hydration layer.

A quantitative understanding of the underlying mechanism of specific ion

effects on phase behavior is still incomplete, but much progress has been made in the
past fifteen years (21–26). It was long believed that the Hofmeister effect was due to

the influence of the ion on the hydrogen bonding network of bulk water, but

experimental results suggest that this influence does not extend beyond the first

hydration shell (27–29). Rather, direct ion-protein interactions appear to contribute

significantly to the ability of a specific ion to salt out proteins from solution.

6
 Two recent developments have emerged that have led to the development

of salt-specific models of protein interactions. The first cites significant contribution

from dispersion forces between ions and protein molecules (24, 30–33). The

polarizabiltity of the ion is the unique physical characteristic that determines the ion’s

specificity. Including ion dispersion contributions in calculating protein-protein

interactions has been shown to qualitatively capture the reverse Hofmeister behavior

exhibited by lysozyme (34, 35), where the salting-out trend follows the opposite order

of the Hofmeister ranking. These findings are consistent with experimental

observations (36).

The second development emphasizes the role of solvent-assisted ion

binding to the hydrophobic regions of the protein surface. Efforts to model this effect

have used molecular dynamics simulations of ions and hydrophobic interfaces to

determine the effective interactions between proteins (37–40). Models of protein

interactions incorporating effects of ion binding have also yielded predictions that are

qualitatively consistent with the salting-out behavior for lysozyme (41).

1.3 Theoretical and Simulation Studies

The interactions between protein molecules play a central role in defining
the macroscopic thermodynamic properties of protein solutions. The key to relating

the microscopic and macroscopic properties is the potential of mean force (PMF). The

PMF represents the effective interaction between two molecules in a system of n

molecules as a combination of the direct interactions between the molecules and

indirect forces from the other species, which include the solvent and ions (42, 43).

Another way of interpreting the PMF is that it is the free energy required to bring two

molecules in an n body system from infinite separation to a particular configuration in

7
solution. In principle, specifying the PMF fixes the thermodynamic properties of the

system. However, the intermolecular interactions of proteins are quite complex and

are strongly dependent on the solution conditions. Therefore, it is not possible at

present to determine an exact PMF model of protein-protein interactions. An

alternative approach is to start with simple idealized models with a few parameters

based on the physics of the system, and fit the parameters to correlate experimental

data.

Experimental evidence of a metastable liquid-liquid phase transition in

proteins has led to theoretical efforts to relate the phase behavior and intermolecular

interactions of proteins within the framework of model colloids (44, 45). It is well

known from theory (46), experiment (47), and simulation that the phase behavior of

colloidal dispersions is sensitive to the range of the interaction between the particles.

Colloidal particles experiencing an attraction that is short-ranged with respect to its

diameter exhibit a metastable liquid-liquid phase transition. Since proteins exhibit

similar phase behavior, idealized colloidal models can serve as a starting point for

describing the thermodynamics of protein solutions.

Early theoretical work attempted to approximate protein solutions as a

one-component system and model the effective interactions between proteins with a
simple form of the pair potential. Within this framework, the solvent is treated as a

continuum background and the protein molecules are represented as hard spheres

experiencing attraction that is a function only of the center-to-center distances (Figure

1.2).

Several simple potentials have been used to model protein interactions,

including the adhesive hard sphere (48–51), square-well (52–55), Yukawa (6, 56, 57),

8
and modified Lennard-Jones (12, 58) potentials. While the forms of these potentials

are different, each of these models has parameters for the interaction strength,

interaction range, and particle diameter. In these models, the interaction range is the

dominant parameter that controls the shape and location of the phase boundaries (58).

Large values of the range parameter lead to phase diagrams that have a stable vapor-

liquid phase separation, which is analogous to a liquid-liquid phase transition for the

colloidal system. As the range parameter becomes sufficiently small, this transition

becomes metastable with respect to the solid-liquid phase transition. In addition, it has

been shown that these isotropic potential models follow an extended law of

corresponding states (59). The consequence of this is that the thermodynamic phase

behavior of systems that interact through short-ranged attraction becomes insensitive

to the details of the interaction potential if they are scaled by the proper parameters.

Figure 1.2: Cartoon of simple isotropic model of proteins. Protein molecules are
represented as perfect spheres of diameter and the interactions
depend only on the center-to-center distance r.

A more complex formulation that has been used to characterize colloidal

interactions and has been applied to proteins comes from DLVO theory (14, 15, 60,

9
61). In this theory the particles are modeled as ideal spheres with a homogeneous

charge distribution that interact via short-ranged van der Waals attraction and long-

ranged Coulombic repulsion. The solvent is assumed to be a structureless continuum

with a uniform dielectric constant. The potential function that reflects this framework

is the sum of three contributions

1.1

where uHS is the hard-sphere excluded volume contribution to the potential, uvdW is the

contribution from the van der Waals attraction, and uelec is the contribution from the

electrostatic repulsion. Within the DLVO framework, the salt screens the electrostatic

interaction between protein molecules, thereby reducing the repulsive electrostatics.

The advantage of DLVO theory is that the model parameters can be directly related to

the solution conditions and physical properties of the system. The van der Waals

contribution is characterized by the Hamaker constant and the size and the separation

distance between the two spheres. The Coulombic repulsion is governed by pH,

which determines the net charge of the protein, and the solution ionic strength, as well

as the size and separation distance.

Studies using simple isotropic models have provided qualitative insight

into the relationship between interactions and the phase behavior of proteins.

Specifically, the metastable liquid-liquid transition exhibited by protein solutions can

be explained in terms of attraction that is short-ranged relative to the size of the

protein. The phase diagrams from these simple models are in qualitative agreement

with experimental observation; however, the model parameters cannot always be

physically related to the solution conditions (pH, temperature, ionic strength), which

are known to determine the protein-protein interactions and therefore phase behavior

10
(62, 63). Consequently, the parameters of these simple models cannot be determined

a priori and can only be used to fit experimental data (62, 63). In addition, the

metastable critical point for protein systems has been shown to be sensitive to the

solution conditions, and this sensitivity cannot generally be captured by spherically

symmetric potentials (64, 65). Thus, simple intermolecular potentials can only be

used as empirical models and cannot be used to predict protein phase behavior for a

wide range of solution conditions.

The phase diagrams predicted by the DLVO model qualitatively correlate

the experimental phase behavior data for lysozyme and -crystallin (66–68), but the

model is unable to quantitatively predict phase behavior that agrees with experiment.

There are several problems with DLVO theory that limits its predictive capability for

the phase behavior of proteins. First, the model does not account for other important

solvation forces that are known to be significant, such as hydrophobic interactions and

hydration effects. Omission of these forces is one of the reasons that DLVO theory

failed to describe phase behavior for some proteins such as apoferritin (69) and

hemoglobins (HbS and HbA) even qualitatively (70). In addition, DLVO theory does

not properly take into account specific ion effects because the theory treats ions as

point charges in solving the Poisson-Boltzmann equation (32, 60). Consequently, the
model cannot explain the varying salting-out abilities of different ions at high salt

concentrations. Another limitation of this theory is that it is not capable of explaining

the salting-in behavior of some proteins. This discrepancy is due to the assumption of

a uniform charge distribution, which inherently treats the electrostatics as always

repulsive. However, the distribution of charges carried by the titratable amino acids

can lead to attractive electrostatic interactions. The screening of the attractive

11
electrostatic interactions with increasing salt concentration leads to increasing stability

of the protein solution, resulting in salting-in behavior. Therefore, DLVO theory

provides an incomplete description of protein-protein interactions, and cannot be

expected to provide quantitatively accurate predictions of protein phase behavior.

More complex models that go beyond spherically symmetric potentials

have been used to predict the phase behavior of proteins. These models emphasize

different features of protein-protein interactions and have predicted phase behavior

with varying degrees of success. An embedded charge model has been used to

account for the charge anisotropy of proteins (71), but the phase diagrams predicted

from this representation were found to agree only qualitatively with experiment (72).

One class of models that have been used for colloids and have been used to represent

the anisotropy of the short-ranged attractions of proteins are patch models (73–78).

Patch models represent protein molecules as spheres that carry attractive regions on

the surface to account for orientationally local strong interactions. Patch models have

been shown to provide more accurate quantitative representations of protein phase

diagrams (73, 75, 79).

1.4 The Osmotic Second Virial Coefficient: B22

One method for characterizing the effective protein-protein interactions is

through the osmotic second virial coefficient, B22. B22 is a dilute solution property that

is a measure of effective two-body interactions in solution, and it provides a link to the

PMF via the statistical mechanical expression for B22 which, accounting for

orientation dependence is given as (42, 80)

12
 1.2

Here W is the PMF, r12 is the center-to-center distance,  and  are the spherical

angles representing the location of the second molecule relative to the first, and , ,

are the Euler angles denoting the rotation of the second molecule (81).
Extensive computations explicitly accounting for the full structural details

of proteins have shown that due to the Boltzmann weighting of the PMF, a few highly

attractive complementary configurations contribute disproportionately to the

calculation of B22 (82, 83). That protein-protein interactions can be dominated by a

few highly attractive configurations has been demonstrated experimentally by a single

point mutation of a crystal contact for T4 lysozyme (84). These attractive

configurations that control bulk solution properties are characteristic of molecular

recognition stemming from the geometric complementarity of apposing regions.

These attractive regions can serve as contacts for crystal formation, which suggest a

plausible correlation with crystallization (85) and protein phase behavior (45, 86–88).

For example, when B22 is positive, the protein molecules on balance repel one another

and remain stable in solution. As B22 becomes negative, the protein interactions are

net attractive and may lead to the formation of condensed phases. The region of

slightly negative B22 values known as the crystallization slot (-1×10-4 to -8×10-4 mol

ml/g2) was identified empirically by George and Wilson to be conducive to the

formation of protein crystals (88). If B22 is too negative, the strong attractions may

prevent the protein molecules from rearranging and forming the specific contacts that

lead to a crystalline lattice, resulting in amorphous precipitates and gels.

13
1.5 Objective and Thesis Outline
The objective of this thesis is to quantitatively relate protein-protein

interactions to protein phase diagrams in order to develop predictive models of protein

phase behavior at different solution conditions. The motivation stems from the need

for providing a rational methodology for the design and optimization of bioseparation

processes. Developing such rational strategies can be significantly aided by

knowledge of the phase diagram. Experimental measurements of phase diagrams for

proteins is a nontrivial task; it can be expensive in terms of time, labor, and supply of

protein due to the difficulty of characterizing the various dense phases, time to attain

true equilibrium, and the wide range of possible solution conditions. In addition, if

crystallization conditions are not known for a protein, measurement of the solid-liquid

phase boundaries is not possible since crystals are needed. Consequently, complete

phase diagrams have been measured for only a few proteins (89). Therefore,

developing predictive models of protein phase behavior is essential and can have

significant industrial and scientific benefits.

The following chapters aim to elucidate the path from molecular structure

to the thermodynamic properties of proteins. Proteins can be represented at various

levels of coarse-graining, from simple spheres to a full atomistic structure. The level
of structural representation directly impacts the ability to model protein interactions

(PMF), which ultimately allows the prediction of bulk solution properties such as B22

and phase behavior. The conceptual path from molecular structure to thermodynamic

properties is illustrated in Figure 1.3. In this thesis, models that represent proteins at

various levels are explored to evaluate their capability of providing the link between

protein interactions and phase behavior.

14
 In Chapter 2, an attempt is made to model protein solutions within the

framework of existing continuum thermodynamic models that have been established

for polymer and colloidal systems. This was done by quantitatively evaluating the

relationship of measured B22 and phase behavior data for a model globular protein

using these models. In addition, phase equilibrium is modeled from the osmotic virial

equation derived from McMillan-Mayer solution theory. Chapter 3 focuses on the

anisotropy of protein-protein interactions on the molecular level and how this feature

impacts the prediction of B22. The anisotropy arising from the shape complementarity

between protein surfaces and the charge distribution is analyzed within the context of

the “patch-antipatch” representation of protein interactions. Further, the numerical

technique for computing B22 from atomistic descriptions of proteins is reexamined,

and a new approach is proposed and outlined. This thesis is concluded by

summarizing the findings from this work and recommendations are made for future

directions.

15
Figure 1.3: Schematic of the conceptual path from molecular structure to
thermodynamic solution properties of proteins, which includes the
osmotic second virial coefficient B22 and phase behavior.

16
 Chapter 2

CONTINUUM THERMODYNAMIC MODELS OF PROTEIN

INTERACTIONS AND PHASE BEHAVIOR

2.1 Motivation and Goal

Unlike previous studies that have taken a molecular approach to develop

predictive models (12, 53, 56, 66–68), a continuum-level approach is explored here to

relate protein-protein interactions and phase behavior. Previous work has

demonstrated a correlation between the osmotic second virial coefficient, B22, and

phase diagrams found from experiments (45, 86, 87). Therefore, there is evidence to

suggest that B22 may offer a reasonable quantitative measure of the effective

interactions between protein molecules. That is, the evidence suggests that the overall

effects of the solution conditions (pH, temperature, and precipitant concentration) on

protein-protein interactions can collectively be represented by B22. The aim of this

chapter is to use experimental B22 data to quantitatively predict the phase diagrams of

proteins by utilizing existing classical thermodynamic models. The hope is that a

simple continuum model with few parameters can provide a useful mechanistic

framework for predicting the phase behavior of protein solutions.

The models investigated in this work were the Flory-Huggins model (90),

the Haas-Drenth model (91–93), and the osmotic virial equation (94). These models

were used to calculate values of B22 and the phase behavior for ribonuclease A in

ammonium sulfate solutions at pH 7 and 23°C and to compare the calculated values to

experimental data obtained by Dumetz et al. (45, 86) (Figure 2.1). The B22 values for

this system were measured using self-interaction chromatography (95, 96) and the

17
dilute boundary of the liquid-liquid coexistence region of the phase diagram was

obtained by a microbatch technique. This particular system was chosen because the

B22 values partially overlap the binodal curve over the range of 0.90 M to 1.25 M

ammonium sulfate (Figure 2.1). It was therefore possible to make a direct correlation

of B22 data with phase behavior within this salt range using the models listed above. It

should be noted that B22 is often difficult to measure at salt concentrations in which

phase separation is observed since the attractions are very strong. As a result, B22 data

rarely overlap phase behavior data for proteins over the same range of salt

concentration, if at all. This difficulty makes relating complementary sets of data with

the above models challenging.

This chapter of the thesis is organized as follows. The modeling structure

used for relating B22 and phase behavior within the framework of the continuum

models is briefly described. For each model, the theoretical foundation is introduced

and the equations that govern phase equilibrium are presented. These equations

provide the modeling structure used for relating B22 and phase behavior. Next, results

of the correlations between experimental B22 and phase behavior data for ribonuclease

A for each of the models are presented and discussed. Finally, conclusions are drawn

on the capability of these continuum models to relate protein interactions and phase
behavior based on the results for ribonuclease A.

18
Figure 2.1: () Binodal, () spinodal, and () B22 data for ribonuclease A in
ammonium sulfate system at 23°C, pH 7. The dotted rectangle
encloses the region where B22 and phase behavior data overlap.
Results were taken from Dumetz et al. (45, 86).

19
2.2 Modeling Structure
The physical parameters for ribonuclease A and the solvent (water) used

for this study are presented in Table 2.1 and Table 2.2, respectively. The specific

volume of ribonuclease A used in these calculations is a value generally used in the

literature for globular proteins (91), which is v2=0.735 ml/g.

Table 2.1: Physical properties of ribonuclease A used in the continuum models.

Property Value Ref

pI 9.6 (45)
MW (g/mol) 13700 (45)
 (nm) 3.1 (86)
2 (ml/g) 0.735 (91)

Table 2.2: Physical properties of water used in the continuum models.

Property Value
MW (g/mol) 18.02
 (g/ml) 0.998
V (ml/mol) 18.06

A schematic of the modeling structure that was followed for relating B22

and phase behavior using the three continuum models is shown in Figure 2.2. With

the Flory-Huggins and Haas-Drenth models, the phase behavior can be predicted

directly from B22 values at each salt concentration, or vice versa. In using phase

behavior to calculate B22, the larger salt range available in the phase behavior data can

be taken advantage of to predict B22. By following this path, B22 predictions can be

made for salt concentrations beyond the experimental range of the B22 data. However,

using measured B22 data to find phase behavior predictions is restricted to the window

20
of conditions for which both B22 and phase diagrams are available. Consequently,

phase behavior calculations cannot be made at higher salt concentrations due to the

lack of B22 data.

For the osmotic virial equation, only one approach was followed, in which

experimental B22 data were used as inputs to predict the corresponding phase behavior.

The reason this path was chosen is that B22 data are needed to compute higher virial

coefficients necessary for the osmotic virial equation to relate interactions and phase

behavior. The reverse approach of using experimental phase behavior data to

calculate virial coefficients could not be utilized because of the need to calculate both

B22 as well as at least one higher virial coefficient.

Figure 2.2: Schematic of the modeling pathways used to relate B22 and phase
behavior with the continuum models.

21
2.3 Flory-Huggins Model

2.3.1 Theory
The Flory-Huggins model of polymer solutions is derived from simple

lattice theory for fluids and has historically been used to predict the phase behavior of

polymer-solvent systems (90). In this model, the system is considered to be composed

of uniform lattice sites that can be occupied by either one solvent molecule or a

polymer subunit (monomer). The polymer is represented as a linear, flexible chain of

interconnected subunits in which the chain is free to adopt any configuration. Each

monomer is allowed to occupy any one lattice site as long as the monomers remain

interconnected. Because of its simplicity and few parameters, the Flory-Huggins

model was deemed worth exploring. The excess free energy of the system has been

derived as (97)

2.2

where 1 and 2 refer to water and protein, respectively, xi is the mole fraction of species

i, and i is the volume fraction of species i. The m parameter reflects the degree of

polymerization of the polymer relative to the size of the solvent molecule and can be

calculated as the ratio of the molar volumes of the polymer to the solvent

2.3

where Vi is the molar volume of species i. Based on the physical parameters listed in

Tables 2.1 and 2.2, a value of 559 was set for the m parameter. The  parameter is an

adjustable parameter that represents the effective interaction between the solute and

solvent.

22
 For the Flory-Huggins model, it can be shown that liquid-liquid phase

equilibrium is modeled by

2.4

2.5

The interaction parameter  is linearly related to B22 by (14)

2.6

where 2 is the specific volume of the protein in units of volume/mass. The spinodal

region of the phase diagram can be determined by the criterion for thermodynamic

stability given as (97)

2.7

From the criterion given by equation 2.7, the spinodal curve from the Flory-Huggins

model is

2.8

The critical point of the phase diagram occurs at the maximum of the spinodal curve

and therefore can be determined by setting the derivative of equation 2.8 with respect

to the volume fraction 2 equal to zero. From this procedure it can be shown that the

critical volume fraction and  are related to the m parameter by

2.9

23
 2.10

The critical B22 is obtained by substituting equation 2.10 into equation 2.6, which

yields

2.11

2.3.2 Results
The B22 predictions made from experimental phase behavior data using the

Flory-Huggins model are presented in Figure 2.3. The model predictions correctly

capture the qualitative trend of decreasing B22 values with increasing salt

concentration, but the results differ quantitatively from the experimental values. The

Flory-Huggins model predicts stronger protein-protein attractions than those reflected

by the experimental measurements for the entire B22-phase behavior overlap region

(0.90 M to 1.25 M). In addition, the steep slope observed in the experimental B22 data

is an important feature not captured by the model.

The reverse path in which phase behavior was predicted from

experimental B22 data was also followed using the Flory-Huggins model. The model

predicts phase separation to occur at higher salt concentrations, as suggested by its

critical point, which is located at 1.22 M ammonium sulfate (Figure 2.4). The critical

point represents the threshold for phase coexistence, with phase separation not

observed at salt concentrations less than that at the critical point. The critical point

predicted by the model indicates that phase separation occurs only at salt

concentrations beyond the overlap region for which both B22 and phase behavior data

are not available. As a result, binodal calculations could not be performed using the

24
model due to the lack of experimental B22 values at the higher salt concentrations.

Despite this limitation, the location of the critical point suggests that the Flory-

Huggins model does not predict phase equilibrium over the same salt range as the

experimental data, and hence does not adequately describe the phase behavior.

Figure 2.3: Comparison of B22 predictions from the () Flory-Huggins model
with () experimental B22 data.

25
Figure 2.4: Critical point predicted by the Flory-Huggins model compared with
() experimental binodal data. The critical point is located at a salt
concentration of 1.22 M. The location of the critical point
demonstrates that the equilibrium phase boundary is located at
higher salt concentrations.

A sensitivity analysis was performed to probe the response of the B22

predictions to small perturbations in the Flory-Huggins model parameters. The

parameter that was perturbed was the m parameter, which characterizes the size of the

protein, which is directly related to the specific volume of the protein 2. The m
parameter was adjusted by 10% from the original value of m=559 used for

ribonuclease A and B22 values were then recalculated from the model. The predicted

26
B22 values for the small adjustments in m are presented in Figure 2.5. Perturbing the m

parameter causes a shift in the predictions; decreasing m causes the B22 predictions to

shift to higher values whereas increasing m causes them to shift to lower values.

However, tuning the m parameter does not lead to a significant change in the slope of

the B22 predictions. Thus, based on the results from the sensitivity analysis, there does

not appear to be a value of m that would lead to B22 predictions that match the

experimental data more convincingly.

Figure 2.5: B22 predictions from Flory-Huggins model calculated from values of
() m=503 and () m=615 compared with original predictions from
() m= 559 and () experimental B22 values.

27
 The effect on the phase behavior predicted from B22 data using adjusted

values of the m parameter was also investigated. As previously mentioned, the model

yielded a critical point that was located at higher salt concentrations than the

experimental data suggest. For the Flory-Huggins model, the critical point location

depends solely on the value of the m parameter, as shown in equations 2.9 and 2.10.
Figure 2.6 shows the effect of changing m by 10% from the original value (m=559)

on the location of the critical point. Increasing m shifts the critical point towards

higher salt concentrations and lower protein concentrations whereas decreasing m

shifts the critical point towards lower salt concentrations and higher protein

concentrations. Based on this trend, m should be decreased in order for phase

separation to be predicted at the lower salt concentrations found in the experimental

data. The m parameter was subsequently decreased by 50% of the original value used,

to a value of 279. The predicted binodal curve was calculated since phase coexistence

shifted to lower salt concentrations, for which B22 interaction data were available

(Figure 2.7). However, it does not appear that there is a physically reasonable value of

m that would lead to predicted phase behavior from B22 data that matches the

experimental results. Consequently, extension of this analysis to lower m values was

not performed.

28
Figure 2.6: Predicted critical point from the Flory-Huggins model for () m=503,
( ) m=559, and () m=603 relative to the () experimental binodal
boundary.

29
Figure 2.7: Predicted binodal boundaries from the Flory-Huggins model for ()
m=279 compared with () experimental results.

30
2.4 Haas-Drenth Model

2.4.1 Theory
The Haas-Drenth model is based on the free energy of mixing for hard

spheres in a solvent and has been used to describe the protein-water phase diagram.

The free energy per unit volume in this model is given as (91–93, 98)

2.12

where  is the molecular volume of the protein, m represents the size of the protein

relative to the solvent molecule, k is the Boltzmann constant, T is the absolute

temperature, represents the interaction between protein molecules in solution, 2 is

the volume fraction of protein, and c is the protein volume fraction in the crystal

(usually taken to be 0.50). Based on the physical parameters listed in Tables 2.1 and

2.2, a value of 559 was set for the m parameter. The first term of equation 2.12

represents the enthalpic contribution from protein-protein interactions. The second

and third terms together represent the contribution from the entropy of mixing for hard

spheres (14, 15). Liquid-liquid coexistence for this model is obtained from the two

equilibrium conditions

2.13

2.14

The interaction parameter is linearly related to B22 by

2.15

31
where MW is the molecular weight of the protein.

The spinodal curve can be derived by applying the condition for

thermodynamic stability given by equation 2.7 to the free energy model in equation

2.12. The critical point occurs at the maximum of the spinodal curve and therefore, by

taking the derivative of the equation for the spinodal curve, it can be shown that the

Haas-Drenth free energy model predicts a critical volume fraction and as

2.16

2.17

The critical B22 is obtained by substituting equation 2.17 into equation 2.15, which

yields

2.18

2.4.2 Results
The B22 predictions made from experimental phase behavior data using the

Haas-Drenth model are presented in Figure 2.8. Similar to the predictions from the

Flory-Huggins model, the Haas-Drenth model predictions correctly capture the

qualitative trend of decreasing B22 values with increasing salt concentration.

However, the predictions are still quantitatively different from the experimental

values. The model predicts values of B22 that are on the same order of magnitude as

the data, but seems to underpredict the attractions at higher salt concentrations. In

addition, the steep slope observed in the experimental B22 data is an important feature

not captured by the model.

32
Figure 2.8: Comparison of B22 predictions from the () Haas-Drenth model with
() experimental B22 data.

To fully explore the prediction capability of the Haas-Drenth model, the

reverse path in which phase behavior was predicted from experimental B22 data was

also followed. The predicted equilibrium binodal boundary and critical point are

presented in Figure 2.9. The Haas-Drenth model appears to provide a better

description of the phase behavior than the Flory-Huggins model. Phase separation is

predicted to occur in the small overlap salt range of the experimental data. The

predicted binodal phase boundary also decays more sharply when compared with the

experimental data. While the phase behavior results appear to be reasonable, the

33
experimental data show that the actual critical point occurs at a lower salt

concentration than that predicted by the model, which is at 1.02 M ammonium sulfate.

Therefore, it would seem that the Haas-Drenth model does not predict phase

equilibrium over the same salt range as the experimental results.

Figure 2.9: Phase behavior predictions from experimental B22 values with the
Haas-Drenth model compared with () experimental binodal data.
The Haas-Drenth model does predict () the binodal boundary to
exist within the overlap region. The predicted () critical point
occurs at an ammonium sulfate concentration of 1.02 M and a protein
concentration of 177 mg/ml.

The sensitivity of the B22 predictions to perturbations in the model

parameters was explored. The m parameter was adjusted by 10% from the original

34
value of m=559 used for ribonuclease A and B22 values were then recalculated. The

predicted B22 values for the adjusted values of m are presented in Figure 2.10. Similar

to the Flory-Huggins model, decreasing m causes the B22 predictions to shift to higher

values whereas increasing m causes the predictions to shift to lower values. However,

the B22 predictions seem to be insensitive to small adjustments in m. In addition,

tuning the m parameter does not lead to a change in the slope of the B22 predictions.

Thus, there does not appear to be a value of m that would lead to B22 predictions that

match the experimental data.

Figure 2.10: B22 predictions from the Haas-Drenth model calculated from values
of () m=503 and () m=615 compared with original predictions
from () m= 559 and () experimental B22 values.

35
 The effect on the phase behavior predicted by the Haas-Drenth model for

adjustments in the value of the m parameter was also explored. The location of the

critical point of this model depends primarily on the specific volume of the protein 2,

which in turn is directly related to the m parameter. Increasing m shifts the critical

point towards higher salt concentrations and lower protein concentrations, whereas

decreasing m shifts the critical point towards lower salt concentrations and higher

protein concentrations. Since the original calculations did not predict phase separation

to occur at the lower salt concentrations seen experimentally, m had to be decreased in

order to shift the predicted boundary in the direction of lower salt concentration. The

m parameter was subsequently decreased by 50% of the original value used to a value

of 279. The resulting binodal curve from this adjustment compared with the original

prediction is shown in Figure 2.11. The adjustment did not shift the boundary

significantly enough to capture the correct phase behavior over the entire salt range.

Thus, it appears that there is no physically realistic value of m that would lead to phase

behavior predictions consistent with the experimental data.

36
Figure 2.11: Predicted binodal boundaries from the Haas-Drenth model for ()
() m=279 compared with the original predictions from m=559 and
the () experimental results.

37
2.5 Osmotic Virial Equation

2.5.1 Theory
The osmotic virial equation was derived by McMillan and Mayer to

describe the nonideality of dilute solutions and is given by (94)

2.19

where  is the osmotic pressure, R is the molar gas constant, and c is the protein

concentration in units of mass/volume. Equation 2.19 is frequently truncated at the

B22 term; however, in order for the model to predict phase separation, it must include

at least the term in the osmotic third virial coefficient B222, which represents three-

body interactions in solution. The governing equations for liquid-liquid coexistence

for the osmotic virial equation can be derived as (see Appendix A)

2.20

2.21

Predicting fluid phase equilibrium requires that the third virial coefficient

B222 be specified. B222 can be theoretically calculated using a simple model of the

potential of mean force (PMF) for protein-protein interactions. The hard-core

attractive Yukawa potential was chosen as the PMF model because it has been used to

describe colloidal interactions (6, 56). The Yukawa potential was originally derived

as a screened Coulombic PMF model to capture long-ranged repulsive electrostatic

interactions (101), but it has been modified to model the short-ranged attractions that

38
dominate phase separation in protein solutions. This Yukawa potential consists of a

hard-sphere contribution and an attractive tail and is given by (2, 17)

2.22

where r is the center-to-center intermolecular distance,  is the interaction well depth,

 is the particle diameter, and b is a parameter that characterizes the range over which
the attraction occurs in units of inverse length. The potential can be rewritten in terms

of reduced variables, in which the parameters of the potential are scaled by

characteristic values. If reduced variables are defined as b*=b, r*=r/, and u*=u/,

the Yukawa potential can be expressed as

2.23

A plot of the Yukawa potential for different values of the range parameter b* is shown

in Figure 2.12.

B222 was calculated from experimental B22 data for different values of b*.

To determine a starting value for these calculations, b* was chosen such that the

Yukawa potential approximately overlapped with the 140-35 Lennard-Jones potential.

The 140-35 Lennard-Jones potential was empirically found to adequately describe

short-ranged non-electrostatic interactions between protein molecules based on more

extensive atomistic simulations of protein-protein interactions (78). From this

procedure, a value of b*=22 for the Yukawa potential was found to approximately

overlay with the 140-35 Lennard-Jones potential (Figure 2.13).

39
Figure 2.12: Plot of the Yukawa potential in reduced units for b* values of (–) 5,
(–) 10, (–) 20, and (–) 30. The (–) hard-sphere repulsion occurs at
r*=1.0. Increasing b* corresponds to a decrease in the range of
attraction.

40
Figure 2.13: Comparison of the (–) 140-35 Lennard-Jones potential with the (–)
Yukawa potential for b* = 22.

The procedure for calculating B222 from B22 involved the following steps:

1) The reduced interaction range parameter b* was fixed and the molecular diameter,

 of ribonuclease A was taken to be the sphere-equivalent value of 3.1 nm (86).

2) The  parameter was determined from the isotropic model for the second virial

coefficient, which is given by (42)

2.24

41
where NA is Avogadro’s number and u(r) is the Yukawa potential. The experimental

B22 values were used as inputs into equation 2.24 and the corresponding  values were

calculated at each salt concentration.

3) Values of the  parameter from step 2) were used to calculate B222 values from the

equation for third virial coefficients (102–104)

2.25

where rij is the intermolecular separation between particle i and j and f(rij) is the Mayer

cluster function defined as (42)

2.26

The model for the third virial coefficient represented by equation 2.25 assumes that

the molecules are spherically symmetric, the interactions are pairwise additive and

multibody interactions are neglected. To compute B222 using the Yukawa potential in

equation 2.25, the method of Alder and Pople (103) for calculating third virial

coefficients for potentials with hard-sphere cores was utilized. This method was

previously used by Graben and Present to calculate third virial coefficients for the

Sutherland potential (104). The computed third virial coefficients were compared

with the results of Naresh and Singh (105), who utilized the Mayer sampling

technique to calculate the virial coefficients for the Yukawa potential. All calculations

and data analysis were performed using appropriate numerical tools in Matlab (see

Appendix B for actual code).

42
2.5.2 Results
The computed B222 values are presented in Figure 2.14. Different trends

were observed in the behavior of B222 for different b* values as a function of salt. For

b*=25, B222 initially increases with increasing salt concentration but then sharply

decreases to negative values. This trend has been observed in the behavior of third

virial coefficients for other potential models (102). For b*=35, B222 is positive

throughout the salt range of interest, but the values appear to plateau around 1.25 M,

which suggests that the values would begin to decrease at higher salt concentrations.

When b* is further increased to a value of 45, B222 increases over the entire salt range.

Thus, it can be inferred that as b* increases, B222 predicted from the Yukawa potential

increases over a wider range of salt concentration.

The phase diagram predictions from the experimental B22 data using the

osmotic virial equation are presented in Figure 2.15. Phase separation was predicted

by the model for B222 values calculated using b* values of 22 or greater in the Yukawa

potential, which corresponds to an interaction range that is 1/22 of the particle

diameter , or less than 1.5 Å. It was determined from these calculations that positive

B222 values are needed in order for the osmotic virial equation to predict phase

separation. Phase behavior calculations with the model were made for b* values of
25, 35, and 45, corresponding to decreasing interaction distances. Calculations were

not performed for higher values of b* because such short interaction ranges were

considered physically unrealistic. The phase behavior predictions follow the correct

qualitative trend of decreasing solubility with increasing salt concentrations; however,

none of the values of b* yield phase behavior that matches the experimental data. In

each case, phase separation is not predicted at lower salt concentrations, for which it is

observed experimentally.

43
Figure 2.14: Computed B222 from the Yukawa potential for b* values of () 25,
() 35, and () 45.

44
Figure 2.15: Phase behavior predictions from the osmotic virial equation based
on B222 calculated from the Yukawa potential. Predictions were
made for b* values of () 25, () 35, and () 45 and compared with
the () experimental binodal data.

45
 The sensitivity of the phase equilibrium predictions from the osmotic

virial equation to experimental errors in B22 was also explored. An approximate

estimate of 2 10-4 mol ml/g2 was assumed for the error in the experimental B22 data.

The original experimental B22 values were adjusted by either +2 10-4 mol ml/g2 or

-2 10-4 mol ml/g2 and corresponding B222 values were recalculated for fixed values

b*. The phase diagrams were subsequently recalculated from the osmotic virial

model. The resulting phase diagram predictions for b* values of 35 and 45 are
presented in Figures 2.16 and 2.17, respectively. Decreasing the B22 values by -2 10-4

mol ml/g2 causes the predicted equilibrium phase boundary to fall back to lower

protein concentrations, which essentially means that the binodal curve becomes

broader. This trend is to be expected because decreasing B22, i.e., increasing the

strength of attraction, would lead to the protein being less soluble in solution and

therefore would push the equilibrium phase boundary to lower concentrations.

However, it does not appear that the predictions are sufficiently sensitive to B22

changes that errors in B22 would account for the discrepancy between the predicted

binodal and the experimental phase behavior results.

46
Figure 2.16: Binodal predictions based on b*=35 compared with the ()
experimental binodal data. Osmotic virial predictions were made
from the experimental B22 values with an assumed error of ()
+2 10-4 mol ml/g2 and () -2 10-4 mol ml/g2. The results are
compared with the predictions from the () original B22 data set.

47
Figure 2.17: Binodal predictions based on b*=45 compared with the ()
experimental binodal data. Osmotic virial predictions were made
from the experimental B22 values with an assumed error of ()
+2 10-4 mol ml/g2 and () -2 10-4 mol ml/g2. The results are
compared with the predictions from the () original B22 data set.

Because the predicted phase coexistence computed using the Yukawa

potential did not match experiment, other potential of mean force models were

explored to determine if they could lead to better phase behavior results. These

models included the square-well potential, 140-35 Lennard-Jones potential, and the

ten Wolde-Frenkel potential. The square-well potential was chosen because of its

simplicity and the fact that since the interactions between proteins are very short-

48
ranged, a potential of mean force model of this form may be adequate to describe the

interaction. The 140-35 Lennard-Jones potential (78) was used because this isotropic

potential provides a reasonable approximation to detailed atomistic simulations of

protein-protein interactions. The ten Wolde-Frenkel potential was also investigated

because it has been used to account for both direct and for solvent-induced interactions

between globular proteins (12). The same methodology as that for calculating B222

from the Yukawa potential was utilized for these potential models.

The square-well potential is the simplest attractive potential and is given

in terms of reduced variables as

2.27

where  is the parameter that characterizes the range of attraction. B222 for the square-

well potential has the analytical form (106, 107)

2.28

where x is defined as

2.29

A plot of B222 values calculated from equation 2.28 for different values of  is

presented in Figure 2.18. For low values of , B222 initially remains fairly flat but then

decreases sharply as the salt concentration increases. When  is sufficiently large, B222

instead increases monotonically within the salt range of interest. The results suggest

49
that the B222 values determined from higher values of  would be more suitable for

calculating phase equilibrium since positive values of B222 are needed for the osmotic

virial equation to predict phase separation.

Figure 2.18: B222 computed from the square well potential for values of () 1.05,
() 1.20, () 1.50, and () 2.10.

B222 was also computed from the 140-35 Lennard-Jones potential, which is

given by

2.30

50
A plot of the resulting B222 values is presented in Figure 2.19. The computed B222

values are mostly negative over the range of salt concentration and decrease sharply at

higher salt concentrations.

Figure 2.19: B222 computed from the 140-35 Lennard-Jones potential.

The ten Wolde-Frenkel potential is a generalized Lennard-Jones potential

and is given as

2.31

51
where  is the parameter that controls the range of attraction. As  increases, the

interaction range decreases and vice versa. A plot of the potential for different values

of  is shown in Figure 2.20. An  value of 50 was determined previously to

qualitatively reproduce the phase behavior of proteins (12), so this value was used as a

starting point in the B222 calculations. The  parameter was then adjusted and the

corresponding B222 values were computed, as shown in Figure 2.21.

Figure 2.20: Plot of the ten Wolde-Frenkel potential for  values of (–) 10, (–) 20,
(–) 30, and (–) 50.

52
Figure 2.21: B222 computed from the ten Wolde-Frenkel potential for  values of
() 10, () 20, and () 30.

Phase behavior calculations were subsequently performed using B222

values calculated from the above potential of mean force models. However, the

resulting B222 values did not lead to phase separation with the osmotic virial equation.

The inability to predict phase coexistence suggests that these potential models are not

adequate for describing the interactions for ribonuclease A. However, the problem

may arise from the need to include higher-order virial coefficients to obtain better

phase coexistence predictions.

53
2.6 Discussion
The discrepancies between the model predictions and experimental data

may be the result of the simplifying assumptions inherent in the continuum models.

The lattice representation used by Flory-Huggins theory allows the polymer chain to

adopt any random configuration, which is unrealistic for proteins since they are known

to have a preferred native conformation. In addition, the interaction parameter

assumes that each monomer interacts equally with the solvent molecules and that the

magnitude is dependent on the number of contacts. This assumption effectively

represents the interactions as being isotropic since the theory does not account for

strongly attractive regions of the polymer chain. The Haas-Drenth model treats

protein molecules as interacting hard spheres where with no regions that display

strong attractions. The framework of hard spheres also implies the assumption of

isotropic interactions. The inability of the osmotic virial equation predictions to match

experimental phase behavior could be partially due to the assumptions made in

calculating B222 from B22 data. B222 was determined by assuming pairwise additivity

and neglecting multibody interactions in addition to the isotropic assumption.

Another possible reason for the inability to predict phase behavior could

be the form of the potential of mean force used to calculate B222. The Yukawa
potential is a simple isotropic potential model that would not be expected to

realistically capture the complex anisotropic interactions of proteins in solution. The

other potential models that were utilized to calculate B222, which included the square-

well potential, 140-35 Lennard-Jones potential, and ten Wolde-Frenkel potential, did

not result in B222 values that led to prediction of phase equilibrium. Furthermore,

improvements in predicting phase behavior with the osmotic virial equation may

require inclusion of higher-order virial coefficient terms; however, calculating such

54
virial coefficients from PMF models is a difficult task and therefore may not be an

efficient path to follow for future work.

Another issue that may explain the quantitative disagreement between

prediction and experiment could be the orientation-averaged nature of B22. While B22

is dominated by a few attractive configurations due to the Boltzmann weighting of the

PMF (82), the orientational averaging essentially washes out the molecular details of

the interactions. Because of this averaging, B22 provides an incomplete representation

of protein-protein interactions and thus would be expected to be limited in its ability to

quantitatively predict phase behavior.

2.7 Conclusions
Isotropic interactions are commonly assumed in the models used, and may

be the reason for the limited quantitative capabilities of the models to predict protein

phase behavior. Therefore, based on the work that has been done with the continuum

models, it can be concluded that the anisotropic character of protein-protein

interactions should be taken into account to quantitatively predict protein phase

behavior. In order to account for anisotropy of protein interactions to predict phase

behavior, molecular-level modeling methods will be needed.

One class of models that has been used to account for the anisotropy of

colloidal particle interactions and has recently been applied to proteins are patch

models (74). These models have been used to simulate colloidal phase behavior and

the results from these models have been shown to be different from those resulting

from isotropic models. The use of patch models to model protein-protein interactions

is the subject of the next chapter.

55
 Chapter 3

PATCH-ANTIPATCH MODEL OF PROTEINS AND THE CALCULATION

OF B22

3.1 Introduction

3.1.1 Review of Patch Models

The interactions of protein molecules are inherently anisotropic. The

physical basis of this anisotropy stems from the nonuniform charge distribution,

nonspherical shape, rough local topography, and heterogeneous functionality on the

protein surface. Anisotropic interactions are responsible for the wide range of solution

phenomena observed in proteins, which include the formation of clusters, gels,

glasses, and crystal nucleation. Incorporating this feature in modeling protein-protein

interactions is important in the simulation of protein phase behavior. However, direct

molecular simulations of phase behavior using models of proteins represented in full

atomistic detail with explicit solvent are not presently computationally tractable. A

more feasible approach entails utilizing a simplified coarse-grained representation of

protein molecules that captures the essential physics of protein interactions. One such

approach is the use of patch models.

Several classes of patch models have been proposed to account for the

orientation dependence in protein-protein interactions. One of the earliest patch

models applied for proteins was the aeolotopic model developed by Lomakin et al. for

describing the phase behavior of -crystallin (75). In this model, the protein molecule

56
is represented as a sphere of which the surface includes a number of attractive spots.

Neighboring protein molecules are said to make contact when the interactions are

between these spots, with the interaction modeled by a square-well potential.

Lomakin et al. found that by including this directionality in modeling the interactions,

the predicted fluid-fluid coexistence curve broadens and more closely matches

experimental measurements than does the isotropic square-well model. Sear (76) later

developed a conical site model for globular proteins in which the particle is modeled

as a hard sphere and interactions occur between paired sites. Using Wertheim

perturbation theory to predict phase behavior, he found that this model was capable of

predicting a metastable fluid-fluid phase transition. Kern and Frenkel (73) proposed a

patch model for colloidal particles in which the specific directional interactions

between patches depend on the relative orientations of two interacting protein

molecules. Their model offered greater flexibility in the number of patches, patch

coverage, and range of patch-patch interactions. Liu et al. (79) later extended the Kern

and Frenkel model by adding a background isotropic square-well attraction in addition

to the patch interactions to simulate protein phase behavior.

The various patch models that have been used to represent proteins have

different characteristics, but they do share common features. They generally represent
each protein molecule as a hard sphere with specific attractive regions on the surface.

In most of the models studied, each patch on one sphere can interact equally with all

other patches on surrounding spheres. The interactions between these regions are

strongly attractive and short-ranged with respect to the particle size. Most work with

patch models has described the patch-patch attractions using the square-well potential,

but there have been studies that have used different potential models (77, 78, 108).

57
Patch models involve a larger parameter set, which includes the number of patches,

patch size, patch arrangement, and range and strength of patch-patch interactions.

Thus, patch models offer greater flexibility and potential for describing the rich variety

of dense phases and their phase boundaries observed experimentally for proteins.

The phase behavior predicted by patch models is both qualitatively and

quantitatively different from the behavior predicted by isotropic potential models.

Similar to isotropic models, patch models are able predict a metastable fluid-fluid

transition region for short-ranged interactions (73, 76, 77, 109, 110). However, the

number of patches is one of the key parameters that controls the phase diagram (79,

111–115). For patch models, it has been shown by theory and simulation (79, 109,

113, 114) that decreasing the number of patches shifts the critical point to lower

temperatures and densities. When compared in reduced units, the fluid-fluid

coexistence curve is broader and the form is in quantitative agreement with

experimental data for lysozyme and -crystallin (73, 75, 109). This result indicates the

importance of patchiness in modeling phase behavior for proteins. Patch models also

open the possibility of describing competing crystalline phases that can be

orientationally ordered or disordered (108, 110, 116–120). The types of stable crystal

structures predicted depend on the compatibility of the patch arrangement on the

surface. Most studies of patch models so far have focused on symmetric arrangements

of patches, although the distribution of attractive patches on actual protein surfaces is

certainly non-uniform. Incorporating anisotropy through patch models has been useful

in describing the rich phase behavior known experimentally for proteins.

Current patch models provide a useful framework for incorporating

anisotropy; however, they are inadequate in representing some of the unique

58
molecular details of proteins. First, these models assume that the protein molecules

are spheres, which is a major simplification since even globular proteins are not

perfectly spherical. In addition, most studies of patch models have assumed symmetric

patch distributions where each patch can interact equally with all other patches. In

reality, the attractive regions are non-uniformly distributed on the protein surface.

Also, specific interactions between geometrically complementary regions are known

to be an intrinsic feature of protein-protein interactions (121). It is these highly

specific attractions that play an important part in determining solution properties, such

as the osmotic second virial coefficient B22, and that control protein crystallization

(82). Therefore, current patch-patch models cannot truly describe the structure of any

real protein solid phase. So far, there have been no studies that have explored the

effect of non-uniform patch distribution on the structure of crystals and other dense

phases. Theoretical examination of protein phase behavior with patch models could

be improved by realistically representing the structural details of protein shape, patch

distribution, and specific interactions between geometrically complementary regions.

One model that takes a step in this direction is the patch-antipatch model developed by

Hloucha et al. (78), which is explored in this chapter.

3.1.2 The “Patch-Antipatch” Model

The patch-antipatch model explicitly accounts for highly specific

interactions that arise from complementary regions on the protein surface. In this

model, the protein is represented as a sphere decorated with patches and corresponding

antipatches on the surface (Figure 3.1). In addition to the weak, distance-dependent

isotropic interaction, there are strong interactions that occur only between patch-

59
antipatch pairs, which reflect the specific pairwise attractions between geometrically

complementary regions. The interaction potential that describes this framework is

3.1

in which

3.2

where Np is the number of patch-antipatch pairs and uiso and uPAm,n are the isotropic

and patch-antipatch contributions to the interaction potential, respectively. Each

unique patch m interacts only with its corresponding antipatch n. The patch-antipatch

interactions are modulated by the scaling function S(m,n), which is dependent on

the relative orientations of the patch and antipatch m andn, respectively, and on the

size of the patches p. The virtue of this model over other patch models is that it

explicitly represents the molecular recognition phenomenon characteristic of protein-

protein interactions.

60
Figure 3.1: A cartoon of a “patch-antipatch” model of protein molecules. The
“patch” is colored in blue and the corresponding “antipatch” is
colored in red. Specific interactions occur only between unique blue
and red colored regions, which depend on the angles of alignment m
and n.

3.1.3 Objective
The objectives of this work are to 1) identify patch-antipatch pairs for

specific model proteins and determine the physical patch-antipatch parameters and 2)

analyze the impact of these highly attractive patch-antipatch pairs on the computation

of B22 at the atomistic level. Understanding the role of these patch-antipatch pairs will

give insight into the anisotropic nature of protein-protein interactions and the influence

it has on the solution properties of proteins, which includes phase behavior.

The unique patch-antipatch parameters for individual proteins are

determined using a hybrid atomistic/continuum methodology for calculating

interaction energies between protein molecules (82, 122). This method involves

simulation of two protein molecules modeled in full in atomistic detail and calculating

the energy of interactions for different angular configurations. This method is capable

61
of capturing the effects of geometric complementarity between the surfaces of protein

molecules. The parameters are:

1) The number of patch-antipatch pairs Np

2) The location and arrangement of the patch-antipatch pairs specified by the

translation angles ,  and rotation angles , , 

3) The size of the patch-antipatch pairs p

4) The strength of patch-antipatch interactions p

These patch-antipatch parameters are determined for two model proteins that exhibit

different solution behavior: lysozyme and chymosin B. Lysozyme has been

extensively studied in the literature and is known to exhibit salting-out behavior for a

wide range of pH values (123). Chymosin B, on the other hand, exhibits salting-in

behavior at pH values near its pI, which is thought to be due to the anisotropy of its

charge distribution (124). The structural data from the PDB files and physical

properties of these model proteins are presented in Table 3.1.

Table 3.1: Proteins studied for patch-antipatch analysis and their physical
properties.

Protein PDB ID MW (g/mol) Residues pI

Lysozyme 4LYZ 14300 129 11
Chymosin B 1CMS 35673 323 4.6

3.2 Theory and Methods

3.2.1 Determining “Patch-Antipatch” Pairs

Patch-antipatch pairs were determined from calculations of short-ranged

attractions (primarily van der Waals interactions) between two protein molecules

62
represented in full atomistic detail. Protein structures were obtained from the atomic

coordinates contained in PDB files from the RCSB Protein Data Bank (www.pdb.org).

The protein molecules were assumed to be rigid bodies in their native conformation.

Because of the short-ranged nature of van der Waals attractions, these interactions are

sensitive to the local geometry of the surfaces. In fact, the level of detail used for the

protein structure has a profound effect on the magnitude of the van der Waals

attraction (18). Thus, regions where there is geometric complementarity of apposing

surfaces lead to stronger attraction. Shape complementarity of the surfaces plays an

important role in the “lock and key” mechanism that is intrinsic in the biological

specificity of protein-protein interactions.

The attraction between two protein molecules was quantified by

calculation of the interaction energies, which depends on the relative orientations of

the molecules. The interactions for a unique angular configuration were determined

by fixing one protein molecule at the origin and translating the second molecule

towards the first in fixed steps, with the interaction energy calculated at each step. For

orientations where attractions are strong, this leads to a potential with a larger well

depth. Thus, a “patch-antipatch” pair is characterized by angular orientations that lead

to interaction potentials with particularly deep wells.

3.2.2 Interaction Energies

3.2.2.1 Short-Range Interactions

A hybrid atomistic/continuum method was used to calculate the short-

ranged interactions (82, 122); a brief review of the model formulation is given here.

In this formulation, interactions between two protein molecules are calculated as a

63
sum of the pairwise interactions between the atoms of the proteins, with solvent

effects implicitly taken into account. The interactions between atom pairs are modeled

based on the separation distances between the atoms. For atom pairs separated by a

center-to-center distance of more than 6 Å, the interactions are determined by the

continuum Lifshitz-Hamaker formulation of van der Waals interactions given by

3.3

where rij is the center-to-center separation distance between two volume elements dVi

and dVj, and AH is the Hamaker constant, which for protein-water-protein interactions

has been determined to be 3.1 kT (125). If two atoms are represented as spheres of

radii Ri and Rj, it can be shown from equation 3.3 that the interaction potential is (126)

3.4

The total free energy of interaction in the Lifshitz-Hamaker approach, WLH, is the sum

of all atom-atom pair interactions described by equation 3.4

3.5

For atom pairs separated by a center-to-center distance of less than 6 Å,

the continuum approximation for the solvent breaks down, and therefore equation 3.4

cannot be used. For this situation, the atomistic Lennard-Jones formulation is used to

determine the short-range interaction

64
 3.6

where Uij is the interaction energy, rij is the center-to-center distance between two

atoms, ij is the size parameter, and ij is the strength of interaction parameter. The

total free energy of interaction in the Lennard-Jones approach, WLJ, is the sum of all

atom-atom pair interactions described by equation 3.6

3.7

The parameters for the Lennard-Jones model were taken from the OPLS-AA force

field (127). The issue with utilizing equation 3.6 to model the dispersion interaction is

that the solvent molecules need to be included explicitly. To take into account effects

from solvation forces, an empirical parameter is introduced to correct the magnitude

of the Lennard-Jones contribution. It has been shown that a value of 0.50 for 

provides a reasonable adjustment when compared with experimental binding free

energies for proteins (128). The total non-electrostatic interaction Wne is the sum of

both free energy contributions from the Lifshitz-Hamaker and Lennard-Jones

approaches

3.8

in which the Lennard-Jones contribution WLJ is scaled by the empirical factor =

0.50.

3.2.2.2 Electrostatic Interactions

The electrostatic interactions are a result of the charges carried by

titratable amino acid residues and partial charges of the atoms. A pairwise screened

65
Coulombic potential was utilized in computing the electrostatic contribution to the

interactions between protein molecules Welec

3.9

where qi and qj are the charges on the two interacting atoms, 0 is the permittivity of

free space, r is the dielectric constant of the solvent, and  is the Debye parameter

that characterizes the length scale for screening of electrostatic interactions by the free

ions in solution. The Debye parameter is related to the ionic strength by

3.10

where ci is the concentration of the ion i, e is the elementary charge, and zi is the

valence of the ion i. Within this framework, the solvent is treated as a structureless

continuum in which its effects are characterized solely by its dielectric constant r,

which for water was taken to be approximately 80.

The partial charges carried by each atom were taken from the OPLS-AA

force field. The effects of pH are reflected in the distribution of charges that are

assigned to ionizable amino acid residues. The protonation state of these residues for

a given pH depends on the pKa values of the titratable groups on the amino acids.

Because the folding of the protein places ionizable residues in environments different

from the solvent-exposed one typical of free amino acids, the local electrochemical

environment of such residues may be altered. These effects can alter the pKa values of

the side chains relative to those for the corresponding free amino acid. To address

these effects, pKa values were determined from the web server propKa (129, 130)

66
(http://propka.ki.ku.dk). From these pKa values and the known pH, the magnitudes of

the charges were computed.

Previous methods of accounting for electrostatics involved solving the

Poisson-Boltzmann equation using a finite-difference method or a boundary-element

approach (131, 132). The advantage of the method used in this work is that it provides

a simple and computationally faster method to account for the effect of protein shape

on the electrostatic interactions.

3.2.3 Calculation of B22

The calculation of B22 for proteins involves sampling angular

configurations between two protein molecules, calculating the interaction energies

using the models described above, and integrating over all possible configurations.

Similar calculations using atomistic models of proteins have been made with more

elaborate approaches (133, 134). However, the emphasis in this work is on simulating

two protein molecules because B22 is by definition a dilute solution property that

characterizes the interactions between two molecules. Allowing for the relative

orientation of two anisotropic molecules, B22 is given as (42, 80)

3.11

where W is the PMF, r12 is the center-to-center distance,  and  are the spherical

angles representing the translation of the second molecule relative to the first

molecule, and , , are the Euler angles denoting the rotation of the second

molecule. The potential of mean force W represents the interactions between the

molecules and is modeled as the sum of the contributions from the non-electrostatic

67
(van der Waals and solvation forces) Wne and electrostatic Welec contributions. It is a

function of both the relative orientation and center-to-center distance between two

molecules. Equation 3.11 can be rearranged by decomposing the radial integral and

consolidating the angular integrals to give

3.12

where the orientation angles are collectively represented by . The first term in

brackets in Equation 3.12 represents the excluded volume contribution to B22 and is

dependent on rc, the center-to-center distance at contact. The center-to-center distance

at contact rc was approximated by linearly interpolating between the points in the PMF

in which the interaction energy transitioned from negative (attraction) to positive

(repulsion due to overlap). The second term is the distance-dependent integral, which

is referred to as the inner integral Iin, and is a direct measure of the energetic

contribution to B22 due to the interactions between molecules for a specific set of

orientation angles . To calculate Iin for a given , the cubic splines method was

used to interpolate the points of the PMF and then a one-dimensional Gaussian

quadrature from the Fortran subroutine library QUADPACK was used to perform the

integration. By designating the distance integral as Iin, Equation 3.12 can be

represented as

3.13

The key challenge is evaluating the two configuration integrals in equation

3.13. Previous work (82, 84, 135–137) utilized Monte Carlo integration to compute

these integrals due to the irregular nature of the integrands. In this approach, N

68
different orientations are randomly sampled from the global configuration space. For

each of these configurations, unique values rc and Iin are computed. B22 is calculated

as the arithmetic average of all rc and Iin computed for all N sampled configurations

(138)

3.14

where V is the hypervolume of the configuration space that is explored. The full

angular space for the configuration integral in equation 3.12 is

3.15

The error in B22 is determined by the variance of the integrand f as

3.16

The angle brackets in equation 3.16 denote the arithmetic mean of the integrand over

the N sample points.

Equation 3.16 demonstrates that the rate of convergence using Monte
Carlo integration is proportional to 1/ N. However, the irregular, highly peaked

nature of Iin due to the Boltzmann weighting of the PMF combined with its

multidimensionality make the reliability of Monte Carlo integration questionable.

Because there are highly peaked regions in the Iin landscape due to highly attractive

patch-antipatch interactions, much of the sampling should focus on these regions since

they make the most significant contribution to B22. However, Monte Carlo integration

69
approximates the integral by determining the mean value of Iin and multiplying it by

the domain of integration. The linear averaging involved in this scheme weights the

contributions from each sampled individual configuration equally. Thus, the few tall

peaks that occupy a small fraction of the global configuration space skew the linear

average, and consequently the integrated value is overestimated. In addition, the

configuration space of Iin is very large due to its high dimensionality, and therefore

there is always uncertainty in the identification of all the peaked regions. The

numerical concerns associated with Monte Carlo integration warrant a reexamination

of the numerical method for computing the configuration integral, and ultimately B22.

To address these numerical concerns, a hybrid Monte Carlo/patch

integration method is proposed and utilized to compute B22. In this approach, B22 is

broken into the sum of three contributions

3.17

where is the excluded volume contribution, is the contribution from the

patch-antipatch interactions, and is the contribution from the non-patch-

antipatch interactions. Configurations with well depths more attractive than -20 kT

were considered to be patch-antipatch pairs and were included in . Conversely,

configurations with less attractive wells were incorporated in .

The excluded volume contribution is computed using the Monte Carlo

integration method described above

3.18

The landscape of the integrand rc is expected to be relatively flat since physically the

center-to-center distance at contact is expected to have a limited range. Therefore,

70
random sampling of the global space is sufficient to yield adequate convergence using

Monte Carlo integration.

The energetic contribution to B22 from the patch-antipatch interactions

requires a more careful and detailed integration procedure. Suppose there are Np

known unique patch-antipatch pairs and the central orientation for the ith patch-

antipatch pair is i = {iiiii}. Furthermore, suppose that the boundary of the

subregion that each patch-antipatch pair occupies is  around its respective central

orientation, which represents the size of the patch. The localized patch integration

3.19

is then performed where the domain of local integration is PA = [i-,i+] [i-

,i+] [i-,i+] [i-,i+] [i-,i+] . The integration over this subregion

is computed using the globally adaptive multidimensional integration routine

DCUHRE (139, 140).

Equation 3.19 represents the contribution to B22 from an individual patch-

antipatch configuration, and the integral in the equation is referred to in what follows

as the configurational integral Iconfig. If all patch-antipatch pairs occupy distinct, non-

overlapping subregions, then the total contribution to B22 of all Np patch-antipatch

pairs is obtained simply by summation

3.20

To complete the calculation of the overall B22, the background

contributions from the non-patch-antipatch configurations are accounted for by Monte

Carlo integration. However, instead of sampling the entire global configuration space,

71
the subregions that the patch-antipatch pairs occupy are excluded from the sampling.

By excluding the configurations that fall in the peaked regions and retaining the ones

that are low to moderately peaked, the Iin landscape is presumably flatter and therefore

Monte Carlo integration would be suitable for calculating this contribution

3.21

where T is the size of the hypervolume occupied by the non patch-antipatch

interactions.

3.3 Results

3.3.1 Identification of Patch-Antipatch Pairs

The histograms representing the distributions of the well depths for non-

electrostatic interactions from 104 randomly sampled configurations for lysozyme and

chymosin B are shown in Figure 3.2. For each protein, most of the configurations

sampled are weakly to moderately attractive, with the modes of the histograms

occurring between -1 kT and -2 kT. The linear average of the well depths for each

protein is about -3 kT. However, for each of the proteins, relatively few strongly
interacting configurations were found, with the strongest configuration identified for

chymosin B having a well minimum of -16.6 kT. However, the question that arises is

whether all high complementary configurations have been identified from the initial

orientation sampling.

72
Figure 3.2: Histograms of the distribution of the short-ranged interaction well
minima for 104 randomly sampled configurations for () lysozyme
and () chymosin B. The inset histograms for each protein are meant
to magnify the tails of the distributions.

The effect of more extensive orientation sampling on patch-antipatch

identification was subsequently explored. The number of random configurations

sampled for each protein was increased by an order of magnitude to 105

configurations. The histograms showing the distributions of the well depths for those

sampled orientations are presented in Figure 3.3. The shapes of the distributions are

similar to those for the 104 configurations initially sampled for each protein; however,

upon closer inspection of the tails of the histograms, more highly attractive

73
configurations were identified as a result of the increase in sampling. The results

indicate that not all patch-antipatch pairs were identified from the initial 104 random

configurations sampled. A further increase in the number of orientations sampled was

therefore needed to adequately explore the configuration space and identify patch-

antipatch pairs. A sampling of 106 random configurations was performed and the

distributions of the well depths are shown in Figure 3.4. Once again more attractive

angular configurations were detected, with the largest well depths being -22.3 kT for

lysozyme and -27.1 kT for chymosin B. Thus, the challenge in the orientation

sampling is to be able to properly and effectively sample the tails of the well depth

distribution. It is interesting to note that more configurations with < -20 kT were

identified for chymosin B than for lysozyme even though the number of

configurations sampled was the same for both proteins. The configurations with < -

20 kT for each protein are presented in Tables 3.5 and 3.6.

The relative frequencies of the well depth distribution for both proteins

appear to be independent of the degree of sampling, with the exception of the

histogram tails (Tables 3.3 and 3.4). The shapes of the histograms suggest that the

well depth variable for both proteins follows approximately a log normal probability

distribution. The probability distribution function that describes such a variable is

given by (141)

3.22

where f is the probability of observation, x is the random variable,  is the location

parameter, and  is the scale parameter. The  and  parameters characterize the log

normal distribution and are computed from the unbiased estimators (141)

74
 3.23

Because the log normal distribution is defined only for positive values of x, the well

depth random variable is defined as the absolute value of /kT

3.24

The parameters for both proteins were determined from well depths computed from

the 106 sampled configurations and are shown in Table 3.2. The fits of the

distributions from these parameters are shown in Figure 3.5. For both proteins, the log

normal probability distribution provides a reasonable description of the relative

frequencies of the well depths in the peaked region of the histograms. However, this

distribution does not provide a good distribution of the tails of the histograms since the

tails do not provide a representative sample of the population of very strongly

attractive configurations.

Table 3.2: Log normal probability distribution function parameters for lysozyme
and chymosin B estimated from the 106 sampled configurations.

Protein  
LYZ 0.975 0.652
CMS 1.067 0.628

75
Figure 3.3: Histograms of the distribution of the short-ranged interaction well
minima for 105 randomly sampled configurations for () lysozyme
and () chymosin B.

76
Figure 3.4: Histograms of the distribution of the short-ranged interaction well
minima for 106 randomly sampled configurations for () lysozyme
and () chymosin B. The largest well depth identified was on the
order of -20 kT.

77
 Table 3.3: Absolute and relative frequencies of lysozyme well depths for different
sampling.

104 Configurations 105 Configurations 106 Configurations

Bin Frequency Rel. Freq. Frequency Rel. Freq. Frequency Rel. Freq.
(0,-1) 766 7.65 10-2 7155 7.15 10-2 74269 7.43 10-2
(-1,-2) 2479 2.48 10-1 25450 2.55 10-1 253823 2.54 10-1
(-2,-3) 2248 2.25 10-1 22552 2.26 10-1 225918 2.26 10-1
(-3,-4) 1694 1.69 10-1 16298 1.63 10-1 163073 1.63 10-1
(-4,-5) 1106 1.11 10-1 11249 1.12 10-1 109809 1.10 10-1
(-5,-6) 724 7.24 10-2 7042 7.04 10-2 70141 7.01 10-2
(-6,-7) 415 4.15 10-2 4327 4.33 10-2 43472 4.35 10-2
(-7,-8) 249 2.49 10-2 2543 2.54 10-2 25850 2.59 10-2
(-8,-9) 152 1.52 10-2 1482 1.48 10-2 15077 1.51 10-2
(-9,-10) 83 8.30 10-3 882 8.82 10-3 8567 8.57 10-3
(-10,-11) 44 4.40 10-3 477 4.77 10-3 4665 4.67 10-3
(-11,-12) 23 2.30 10-3 266 2.66 10-3 2482 2.48 10-3
(-12,-13) 8 8.00 10-4 127 1.27 10-3 1360 1.36 10-3
(-13,-14) 7 7.00 10-4 71 7.10 10-4 723 7.23 10-4
(-14,-15) 2 2.00 10-4 35 3.50 10-4 394 3.94 10-4
(-15,-16) 0 0 23 2.30 10-4 188 1.88 10-4
(-16,-17) 0 0 9 9.00 10-5 88 8.80 10-5
(-17,-18) 0 0 5 5.00 10-5 50 5.00 10-5
(-18,-19) 0 0 3 3.00 10-5 27 2.70 10-5
(-19,-20) 0 0 1 1.00 10-5 15 1.50 10-5
(-20,-21) 0 0 2 2.00 10-5 5 5.00 10-6
(-21,-22) 0 0 1 1.00 10-5 1 1.00 10-6
(-22,-23) 0 0 0 0 2 2.00 10-6
(-23,-24) 0 0 0 0 1 1.00 10-6

78
Table 3.4: Absolute and relative frequencies of chymosin B well depths for
different sampling.

104 Configurations 105 Configurations 106 Configurations

Bin Frequency Rel. Freq. Frequency Rel. Freq. Frequency Rel. Freq.
(0,-1) 498 4.98 10-2 4700 4.70 10-2 47522 4.75 10-2
(-1,-2) 2502 2.50 10-1 24660 2.47 10-1 247585 2.48 10-1
(-2,-3) 2261 2.26 10-1 23024 2.30 10-1 229386 2.29 10-1
-1
(-3,-4) 1675 1.68 10 16587 1.66 10-1 166303 1.66 10-1
-1
(-4,-5) 1129 1.13 10 11564 1.16 10-1 114250 1.14 10-1
(-5,-6) 787 7.87 10-2 7420 7.42 10-2 74522 7.45 10-2
-2
(-6,-7) 467 4.67 10 4717 4.72 10-2 47298 4.73 10-2
-2
(-7,-8) 284 2.84 10 2967 2.97 10-2 29520 2.95 10-2
-2
(-8,-9) 164 1.64 10 1802 1.80 10-2 17750 1.78 10-2
(-9,-10) 92 9.20 10-3 1066 1.07 10-3 10703 1.07 10-2
-3
(-10,-11) 68 6.80 10 637 6.37 10-3 6611 6.61 10-3
-3
(-11,-12) 37 3.70 10 373 3.73 10-3 3783 3.78 10-3
-3
(-12,-13) 22 2.20 10 190 1.90 10-3 2077 2.08 10-3
(-13,-14) 4 4.00 10-3 103 1.03 10-3 1217 1.22 10-3
-4
(-14,-15) 4 4.00 10 79 7.90 10-4 644 6.44 10-4
-4
(-15,-16) 3 3.00 10 42 4.20 10-4 381 3.81 10-4
(-16,-17) 0 0 21 2.10 10-4 202 2.02 10-4
(-17,-18) 0 0 6 6.00 10-5 112 1.12 10-4
(-18,-19) 0 0 11 1.10 10-4 78 7.80 10-5
(-19,-20) 0 0 2 2.00 10-5 21 2.10 10-5
(-20,-21) 0 0 1 1.00 10-5 16 1.60 10-5
(-21,-22) 0 0 1 1.00 10-5 14 1.40 10-5
(-22,-23) 0 0 0 0 2 2.00 10-6
(-23,-24) 0 0 1 1.00 10-5 1 1.00 10-6
(-24,-25) 0 0 0 0 0 0
(-25,-26) 0 0 0 0 0 0
(-26,-27) 0 0 0 0 1 1.00 10-6
(-27,-28) 0 0 0 0 1 1.00 10-6

79
Figure 3.5: Comparison of relative frequencies of () lysozyme and () chymosin
B absolute well depths with respective fits from (–) log normal
probability distribution function. The fits are based on parameter
values of  = 0.975,  = 0.652 for lysozyme and  = 1.067,  = 0.628
for chymosin B.

80
Table 3.5: Ten most attractive angular configurations for short-range
interactions of lysozyme identified from 106 randomly sampled
orientations.

      kT
1 1.191 0.975 5.309 1.505 4.201 -22.28
2 2.213 1.292 2.358 0.763 1.862 -20.70
3 3.513 1.002 4.286 0.681 1.800 -20.05
4 3.301 1.228 0.607 1.424 2.635 -23.32
5 5.211 0.575 4.674 1.283 5.123 -20.62
6 1.110 1.145 4.566 1.696 4.923 -21.09
7 3.276 1.010 2.392 0.831 2.138 -20.18
8 2.338 1.019 0.422 0.110 2.240 -22.36
9 1.776 2.791 4.207 1.594 4.237 -20.94
10 1.565 2.830 2.183 2.465 0.545 -19.98

81
Table 3.6: Thirty-five most attractive angular configurations for short-range
interactions of chymosin B identified from 106 randomly sampled
orientations.

#      kT
1 4.298 2.105 1.138 2.030 1.845 -23.63
2 4.074 1.347 4.930 2.051 2.096 -26.12
3 4.299 0.977 5.746 0.751 0.217 -20.97
4 3.909 1.157 5.566 0.448 5.185 -20.43
5 4.399 2.511 0.031 1.757 5.561 -20.94
6 4.084 0.746 0.395 2.584 5.010 -20.35
7 3.607 1.295 2.076 1.657 0.177 -21.27
8 5.884 1.458 3.317 0.997 5.614 -20.03
9 5.888 1.565 1.271 2.518 2.597 -20.58
10 3.599 2.140 5.977 0.252 0.706 -21.38
11 6.025 1.461 0.227 1.556 0.962 -20.77
12 6.233 1.303 4.536 1.702 1.704 -20.68
13 0.048 2.491 2.023 0.158 0.812 -20.59
14 3.026 1.342 2.839 2.825 0.116 -27.36
15 3.016 1.126 6.121 2.775 3.114 -22.30
16 1.168 2.841 3.861 1.119 2.824 -20.16
17 2.722 2.274 0.056 2.153 0.649 -21.12
18 2.694 0.945 6.057 2.047 5.737 -20.93
19 0.567 0.818 0.489 1.655 5.215 -21.06
20 0.994 2.640 3.606 2.428 4.930 -21.95
21 0.447 1.413 3.793 1.225 5.585 -21.55
22 0.512 1.594 2.939 1.809 0.381 -21.99
23 0.577 1.249 2.706 1.075 5.117 -21.63
24 1.122 2.518 1.715 1.153 0.248 -21.99
25 2.476 2.103 1.461 0.973 5.029 -20.36
26 0.870 2.331 3.964 1.604 3.330 -21.08
27 2.096 2.440 3.645 1.933 2.479 -21.26
28 2.495 1.400 0.534 1.085 4.471 -20.62
29 1.164 2.290 6.260 2.271 3.240 -20.56
30 1.802 2.312 4.456 1.532 3.731 -20.83
31 2.204 2.035 1.788 1.511 0.062 -20.00
32 1.644 2.268 0.733 1.815 2.108 -21.42
33 0.969 1.238 4.187 2.865 5.352 -21.18
34 1.084 1.130 2.555 1.397 3.987 -22.94
35 1.744 2.180 1.114 1.999 4.534 -21.30

82
 Another issue that arises in the random orientation sampling is whether the

patch-antipatch pairs identified are sampled at the optimally aligned orientation that

represents the true well depth. To explore this, the first highly attractive configuration

for lysozyme in Table 3.5 (entry 1) was selected as a test case. Each individual
orientation angle for this configuration was varied within 0.20 radian around the

central orientation while holding the other angles fixed at their respective central

values. The response in the well depth to these variations is shown in Figure 3.6. If

the central orientation were at the true optimum alignment, the minimum in the

interaction energy profile would occur at a variation of 0 radian. However, Figure 3.6

indicates that this configuration is in fact not the optimally aligned one. A similar

result was obtained by Hloucha et al. (see Figure 2 in reference (78)). Thus, it can be

inferred that the orientations that are identified in the random sampling may not, in

general, be at their energy minima.

To approximate the optimal alignment that leads to the energy minimum,

a local sampling around the central orientation for each patch-antipatch configuration

was performed. This local sampling entailed sampling 105 random orientations
confined within the limits of 0.10 radian around each angle of the central orientation.

This local sampling procedure was performed for several of the patch-antipatch pairs
for lysozyme and chymosin B and the fined-tuned orientations are shown in Tables 3.7

and 3.8, respectively. When compared with the initial sampled configurations, the

differences in each of the angles are small, yet these small variations can lead to

substantial changes in the well depths. The most significant case is for the first patch-

antipatch pair for lysozyme, in which a configuration with a well depth on the order of

-40 kT was identified. Similar results can be seen for the chymosin B patch-antipatch

83
configurations in which the refinement led to a well-depth as high as -39 kT. The

interactions between patch-antipatch pairs for proteins are quite sensitive to small

perturbations in orientation; a slight change in the alignment can lead to a significant

change in the attraction. This sensitivity indicates that the region representing a given

patch-antipatch interaction is only a minute fraction of the global angular space and

that these patch regions are very small.

Figure 3.6: Well depth as a function of the angles for the orientation listed in
entry 1 of Table 3.5. The largest change occurs when is decreased
by -0.02 radian, which indicates that the originally sampled
orientation is not the optimum alignment.

84
Table 3.7: Refined orientations for lysozyme identified from the local sampling in
0.10 radian around the central orientation in Table 3.4. The resulting
angular configurations are significantly more attractive than the
originally sampled orientations.

#      kT Initial kT

1 1.184 0.907 5.274 1.540 4.158 -40.48 -22.28
2 2.203 1.273 2.269 0.747 1.897 -24.30 -20.70
3 3.478 1.050 4.245 0.773 1.856 -26.34 -20.05
4 3.244 1.190 0.612 1.369 2.537 -29.06 -23.32
5 5.260 0.597 4.640 1.189 5.123 -25.05 -20.62
6 1.085 1.125 4.504 1.754 4.905 -24.97 -21.09
7 3.244 0.963 2.332 0.919 2.235 -26.04 -20.18
8 2.355 0.987 0.505 0.134 2.163 -22.88 -22.36
9 1.808 2.754 4.295 1.621 4.184 -24.76 -20.94
10 1.533 2.818 2.122 2.469 0.608 -24.60 -19.98

85
Table 3.8: Refined orientations for patch-antipatch pairs for chymosin B
identified from the local sampling in 0.10 radian around the central
orientation in Table 3.5.

#      kT Initial kT

1 4.322 2.148 1.212 2.105 1.919 -26.88 -23.63
2 4.028 1.379 4.833 2.064 2.040 -39.11 -26.12
3 4.311 0.940 5.784 0.708 0.192 -26.90 -20.97
4 3.921 1.144 5.648 0.446 5.110 -25.38 -20.43
5 4.480 2.548 0.106 1.806 5.660 -25.22 -20.94
6 4.110 0.764 0.494 2.622 5.109 -24.98 -20.35
7 3.555 1.243 1.987 1.715 0.193 -25.83 -21.27
8 5.853 1.452 3.268 1.022 5.663 -28.84 -20.03
9 5.930 1.563 1.355 2.525 2.589 -22.89 -20.58
10 3.593 2.122 5.890 0.179 0.769 -24.69 -21.38
11 5.994 1.448 0.177 1.526 0.999 -25.66 -20.77
12 6.214 1.347 4.517 1.614 1.754 -27.06 -20.68
13 0.036 2.515 1.924 0.195 0.861 -25.32 -20.59
14 2.972 1.340 2.831 2.771 0.208 -35.80 -27.36
15 2.954 1.126 6.109 2.861 3.213 -33.06 -22.30
16 1.267 2.860 3.836 1.106 2.823 -24.05 -20.16
17 2.691 2.312 -0.007 2.146 0.650 -25.77 -21.12
18 2.756 0.896 6.132 2.146 5.686 -25.16 -20.93
19 0.580 0.805 0.538 1.705 5.241 -23.76 -21.06
20 0.950 2.620 3.512 2.428 4.868 -28.02 -21.95
21 0.453 1.426 3.781 1.324 5.642 -28.02 -21.55
22 0.483 1.567 2.881 1.763 0.360 -26.92 -21.99
23 0.536 1.249 2.618 1.087 5.129 -27.48 -21.63
24 1.073 2.518 1.677 1.091 0.186 -27.09 -21.99
25 2.451 2.091 1.411 0.955 5.128 -22.51 -20.36
26 0.872 2.310 3.968 1.562 3.338 -28.43 -21.08
27 2.090 2.402 3.629 1.895 2.504 -28.19 -21.26
28 2.495 1.432 0.521 1.022 4.446 -23.66 -20.62
29 1.065 2.278 6.198 2.259 3.325 -23.99 -20.56
30 1.727 2.350 4.358 1.544 3.830 -25.02 -20.83
31 2.230 1.979 1.713 1.610 -0.037 -24.60 -20.00
32 1.743 2.293 0.832 1.828 2.027 -27.89 -21.42
33 0.918 1.250 4.125 2.766 5.301 -25.59 -21.18
34 1.097 1.074 2.557 1.297 3.937 -27.08 -22.94
35 1.744 2.137 1.120 2.057 4.596 -38.82 -21.30

86
3.3.2 Calculation of B22
The excluded volume contribution to B22 was first calculated using the

Monte Carlo integration approach, where the global configuration space was randomly

sampled and the center-to-center contact distance was calculated for each

configuration sampled. The final numerical values are presented and compared with

the theoretical hard sphere B22 values based on the equivalent sphere diameter of each

protein (Table 3.9). The values are comparable in magnitude, but the value is higher

when the protein is modeled atomistically than when it is represented as a hard sphere.

This result is consistent with the findings of Neal and Lenhoff (142), who

demonstrated that the magnitude of the excluded volume contribution depends on the

level of structural detail of the protein considered. They found that when protein

molecules are resolved atomistically, the excluded volume contribution can be as

much as 40% greater than the result obtained when the protein is modeled as an ideal

sphere. The disparity in the excluded volume B22 predicted by the two representations

is attributed to the roughness of the protein surface. In the atomistic case, groups of

atoms that protrude on the surface limit the closeness of approach when two protein

molecules come in contact. Consequently, there is a greater effective center-to-center

distance leading to a greater excluded volume contribution to B22 than for the ideal
sphere representation.

The contributions from the molecular interactions were subsequently

included in the B22 calculations. As a starting point, only the short-ranged non-

electrostatic interactions were considered in the calculation. The B22 values based on

the excluded volume and van der Waals interactions calculated using Monte Carlo

integration are presented in Table 3.9. There are two symptoms of computational

challenges evident in these results. First, the B22 values computed for each protein do

87
not appear to converge even after 106 configurations are sampled and second, the final

computed values are orders of magnitude greater than typical experimentally

measured values. To understand why the magnitudes of the computed B22 value are so

large, the histograms of the computed Iin values for the 106 configurations sampled for

lysozyme and chymosin B are shown in Figures 3.7 and 3.8, respectively. The range
of Iin computed spans many orders of magnitude, from as low as 1 101 Å3 to as high

as 1 1014 Å3. This wide range is due to the nature of the integrand in Iin;

configurations with large well depths are magnified as a result of the Boltzmann

weighting of the PMF. Because the Monte Carlo integration method takes the linear

average of all Iin values, the few configurations that have extremely large Iin values

contribute disproportionately to the mean value of Iin and consequently the integral is

overestimated. Thus, the linear averaging used in Monte Carlo integration cannot

provide an accurate numerical estimate of the configurational integral to compute B22.

Table 3.9: B22 calculated from 106 randomly sampled configurations based on
excluded volume contribution and both excluded volume and short-
range attraction. The  value is the sphere equivalent diameter
determined from the empirical correlation of Neal and Lenhoff (142).
The error in the Monte Carlo estimate is calculated from equation
3.16.

 Total B22
4 2 4 2
Protein ( 10 ml mol/g ) (Å) ( 10 ml mol/g ) (ml mol/g2)
LYZ 2.997 ± 0.001 32.0 2.02 -0.24 ± 0.07
CMS 1.3990 ± 0.0005 42.9 0.783 -0.95 ± 0.59

88
Figure 3.7: Histogram of the computed Iin for the 106 randomly sampled
configurations for lysozyme. The inset shows an enlarged view of the
high-Iin tail of the distribution.

89
Figure 3.8: Histogram of the computed Iin for the 106 randomly sampled
configurations for chymosin B. The inset shows an enlarged view of
the high-Iin tail of the distribution.

90
 To address the numerical issues associated with Monte Carlo integration, a

hybrid Monte Carlo/patch integration method was utilized to compute B22. Different

configurations of varying strengths for each protein, which included the highly
attractive patch-antipatch pairs, were selected and integrated within the limits of 

radians around the central orientations using the DCUHRE integration routine. The 

parameter directly determines the size of the subregion over which the interaction

energy is strongly attractive before it decays, and therefore represents the patch size.

However, increasing  enlarges the hypervolume of the integration and consequently

Iconfig increases monotonically. This can be seen in Figure 3.9 in which Iconfig is

computed using the integration routine for the fourth lysozyme patch-antipatch pair

listed in Table 3.7. Thus, the size of a patch-antipatch pair cannot be determined

directly from the dependence of Iconfig on . To estimate an appropriate value for ,

Iconfig was normalized by the hypervolume of integration v0 and the result was plotted

against . The integration hypervolume v0 is directly related to  through equation

3.15 with the appropriate integration limits. This normalization filters the effect of the

hypervolume size in the integration and gives some indication of how the rate of

growth of the integral decreases due to the decay of the interaction potential. Figure

3.10 shows that at approximately  = 0.10 radian the normalized integral decreases by
a factor of about 10 from its highest value. This significant decay would indicate that

the effects of the attraction on Iconfig dissipates and therefore = 0.10 radian

characterizes the patch size for this particular orientation. Although the patch size for

each patch-antipatch pair may be different,  = 0.10 radian was chosen as a universal

value for all integration of patch-antipatch pairs in order to simplify further

computations.

91
Figure 3.9: Iconfig computed from the DCUHRE routine as a function of  for
lysozyme patch-antipatch pair 4 in Table 3.7. Iconfig increases
monotonically as  increases due to the increase in the hypervolume
of the integration.

92
Figure 3.10: Iconfig normalized by the volume of integration v0 as a function of 
for lysozyme patch-antipatch pair 4 in Table 3.7. The normalized
integral decreases as  is increased.

The localized integration was performed for the cases where only short-

range interactions were considered and where both short-range and electrostatic

interactions were included. The local integration represents the contribution of these
individual patch-antipatch regions to B22. The results of the integration, in units of Å3,

are plotted as a function of the total well depth (short-range and electrostatics) of the

configuration with the lowest energy minimum for each patch-antipatch pair in Figure

3.11 for lysozyme and in Figure 3.12 for chymosin B. The results were fitted using an

93
exponential regression model. There is some scatter in the results, which is to be

expected since there is variability in the shapes of the energy profiles and uncertainty

in the energy minima, and not all patch-antipatch interactions may be confined
completely within 0.10 radians. Despite this scatter, there is a consistent, well-

behaved relationship in which the computed localized configurational integral Iconfig

increases as the strength of interaction increases. The correlations for both proteins are

similar; both regressions have comparable preexponential factors, but the regressed fit

for chymosin B possesses a less negative exponent. The trends demonstrate that

configurations with stronger attractions make a greater contribution to B22.

Electrostatics can tune a configuration’s contribution to B22 by either reducing its

attraction through addition of repulsion or enhancing its attraction. Repulsive

electrostatics result in decreasing Iconfig whereas attractive electrostatics corresponds to

increasing Iconfig. By knowing a patch-antipatch pair’s well depth, Figures 3.11 and

3.12 can be used to empirically estimate its contribution to B22 without the need for

performing a detailed integration using the DCUHRE routine.

94
Figure 3.11: Plot of the localized configuration integration for lysozyme as a
function of the total well depth. Iconfig was computed based on (+)
short-range interactions alone and () short-range interactions with
electrostatics. Integration was performed within the limits of 0.10
radian around the central orientation of each patch-antipatch
configuration using the DCUHRE routine. The regressed curve is
F(x)=0.0200 exp(-0.805x), R2=0.9941.

95
Figure 3.12: Plot of the localized configuration integration for chymosin B as a
function of the total well depth. Iconfig was computed based on ( )
short-range interactions alone and () short-range interactions with
electrostatics. Integration was performed within the limits of 0.10
radian around the central orientation of the configurations. The
regressed curve is F(x)=0.0195 exp(-0.770x), R2=0.9806.

The B22 contributions for the individual lysozyme pairs are presented in

Table 3.10, which include contributions resulting from the short-range interactions

alone and from both short-range and electrostatic interactions. The results show that

inclusion of electrostatics significantly reduces the attractions for most of the patch-

antipatch pairs, as indicated by the increased contribution to B22 as the ionic strength

increases. This indicates the electrostatic interactions in the configurations

96
represented by the patch-antipatch pairs for lysozyme are predominantly repulsive.

However, two of the ten patch-antipatch pairs with  < -20 kT (entries 1 and 8 in Table

3.10) appear to have an increased B22 contribution when electrostatics are

incorporated, which suggests that this particular pair possesses attractive electrostatics.

It is interesting to note that the configuration in entry 1 of Table 3.10, which is the -40

kT patch identified after local sampling refinement, has a contribution that is orders of

magnitude greater than those of the other pairs. The contribution by this lone patch-

antipatch pair would therefore overwhelm the contributions from the other patches if

they were summed together.

The trends in the B22 contributions from the chymosin B patch-antipatch

pairs appear to show different behavior. Although some of the pairs exhibit repulsive

electrostatics, some of the configurations show the opposite trend. In fact, almost one-

third of the patch-antipatch pairs with  < -20 kT were found to have increased B22

contributions when electrostatics are incorporated. Thus, chymosin B was not only

found to have a larger number of strongly attractive patch-antipatch pairs than

lysozyme, but also a greater percentage of the attractive patch-antipatch pairs possess

attractive electrostatics. For a few configurations, the differences are within the errors

of integration and may be statistically insignificant, but the results indicate that in
general there are smaller repulsive electrostatic effects for chymosin B than for

lysozyme. Given that the local integration for each pair occurs over a very small

portion of the global configuration space, the B22 contributions from these patch-

antipatch pairs are quite remarkable.

97
 Table 3.10: Contributions to B22 from individual patch-antipatch pairs for
lysozyme based on short-range non-electrostatic interaction energies
alone and with electrostatics at pH 7. The contributions were
determined by integrating within ±0.10 radian around the central
orientation using the DCUHRE integration routine.

Non-electrostatics pH 7, 0.10M pH 7, 0.20M pH 7, 0.30M

/kT (mol ml/g2) /kT (mol ml/g2) /kT (mol ml/g2) /kT (mol ml/g2)
-40.48 -4.37 101 -40.82 -8.00 101 -41.27 -1.06 102 -41.31 -1.13 102
-24.31 -5.54 10-5 -22.00 -8.51 10-6 -22.71 -1.60 10-5 -23.01 -2.10 10-5
-26.34 -6.06 10-4 -22.62 -1.48 10-5 -23.69 -4.03 10-5 -24.21 -6.65 10-5
-29.77 -2.01 10-2 -25.89 -1.12 10-3 -26.57 -2.17 10-3 -27.05 -3.35 10-3
-25.05 -2.09 10-4 -24.77 -9.72 10-5 -25.13 -2.01 10-4 -25.21 -2.13 10-4
-24.97 -9.14 10-5 -23.45 -2.58 10-5 -23.95 -4.45 10-5 -24.12 -5.32 10-5
-26.05 -7.08 10-4 -22.86 -2.92 10-5 -23.70 -6.90 10-5 -24.14 -1.05 10-4
-22.36 -3.02 10-5 -23.39 -5.02 10-5 -23.82 -6.68 10-5 -23.85 -6.90 10-5
-24.76 -2.69 10-4 -23.05 -5.60 10-5 -23.82 -1.11 10-4 -24.11 -1.47 10-4
-24.60 -3.69 10-4 -20.66 -7.46 10-6 -21.93 -2.45 10-5 -22.55 -4.40 10-5
-19.09 -1.43 10-6 -17.70 -1.19 10-6 -18.09 -1.58 10-6 -18.16 -1.74 10-6
-18.65 -1.17 10-6 -16.87 -2.12 10-7 -17.73 -4.24 10-7 -18.02 -5.67 10-7
-15.77 -1.24 10-7 -16.27 -1.85 10-7 -16.33 -2.26 10-7 -16.39 -2.30 10-7
-12.91 -1.62 10-8 -11.69 -5.02 10-9 -12.24 -8.39 10-9 -12.46 -1.03 10-8
-11.32 -1.84 10-9 -10.46 -1.00 10-9 -10.98 -1.54 10-9 -11.16 -1.74 10-9
-9.37 -7.58 10-10 -9.09 -5.72 10-10 -9.39 -7.64 10-10 -9.46 -8.16 10-10
-7.72 -3.30 10-10 -5.92 -4.95 10-11 -6.54 -1.01 10-10 -6.84 -1.35 10-11

98
 Table 3.11: B22 contributions from individual patch-antipatch pairs for chymosin
B based on short-range non-electrostatic interaction energies alone
and with electrostatic interactions at pH 5. The contributions were
determined by integrating within 0.10 radian around the central
orientation using the DCUHRE integration routine.

Non-electrostatics pH 5, 0.10M pH 5, 0.30M pH 5, 0.40M

/kT (mol ml/g2) /kT (mol ml/g2) /kT (mol ml/g2) /kT (mol ml/g2)
-26.88 -2.10 10-5 -27.66 -5.65 10-5 -27.12 -3.11 10-5 -27.00 -2.93 10-5
-39.11 -1.08 100 -37.43 -2.67 10-1 -37.87 -4.17 10-1 -37.98 -4.57 10-1
-26.90 -2.97 10-5 -25.74 -1.23 10-5 -25.87 -1.39 10-5 -25.94 -1.49 10-5
-25.38 -1.05 10-5 -24.41 -4.56 10-6 -24.85 -6.87 10-6 -24.92 -7.33 10-6
-25.22 -1.99 10-5 -23.83 -5.72 10-6 -24.64 -8.89 10-6 -24.68 -1.18 10-5
-24.89 -1.04 10-5 -25.15 -1.63 10-5 -25.14 -1.68 10-5 -25.09 -1.49 10-5
-25.83 -2.80 10-5 -24.56 -1.03 10-5 -25.15 -1.63 10-5 -25.23 -1.75 10-5
-28.84 -3.06 10-4 -28.71 -9.31 10-5 -28.80 -1.14 10-4 -28.83 -1.21 10-4
-22.89 -2.71 10-6 -23.30 -3.77 10-6 -23.19 -3.45 10-6 -23.19 -3.45 10-6
-24.69 -1.49 10-5 -22.58 -2.57 10-6 -23.76 -8.13 10-6 -23.94 -9.64 10-6
-25.66 -9.85 10-6 -25.78 -1.19 10-5 -25.84 -1.21 10-5 -25.80 -1.17 10-5
-27.06 -3.58 10-5 -23.60 -4.77 10-6 -25.27 -1.79 10-5 -25.59 -2.32 10-5
-25.32 -1.86 10-5 -25.27 -1.79 10-5 -25.53 -2.20 10-5 -25.54 -2.23 10-5
-35.50 -1.66 10-1 -35.73 -2.06 10-1 -35.72 -2.03 10-1 -35.68 -1.93 10-1
-33.06 -1.02 10-2 -32.66 -7.20 10-3 -32.83 -8.67 10-3 -32.84 -8.84 10-3
-24.05 -1.63 10-5 -23.40 -1.14 10-5 -24.02 -1.65 10-5 -24.05 -1.69 10-5
-25.77 -1.35 10-5 -25.96 -2.29 10-5 -25.74 -1.84 10-5 -25.65 -1.72 10-5
-25.16 -1.54 10-5 -24.65 -1.48 10-5 -24.72 -1.46 10-5 -24.71 -1.43 10-5
-23.76 -6.76 10-6 -22.68 -2.30 10-6 -23.42 -4.11 10-6 -23.49 -4.38 10-6
-28.02 -2.40 10-5 -25.70 -3.67 10-6 -26.99 -7.07 10-5 -27.20 -1.59 10-5
-28.02 -9.45 10-5 -26.44 -2.24 10-5 -27.22 -4.96 10-5 -27.32 -5.43 10-5
-27.12 -7.85 10-5 -25.34 -1.91 10-5 -26.46 -4.62 10-5 -26.65 -5.39 10-5
-27.48 -1.04 10-4 -26.47 -2.07 10-5 -27.37 -9.54 10-5 -27.45 -1.02 10-4
-27.09 -5.72 10-5 -28.20 -1.33 10-4 -27.94 -1.05 10-4 -27.83 -9.44 10-5
-22.51 -5.98 10-6 -21.88 -3.98 10-6 -22.17 -5.17 10-6 -22.21 -5.40 10-6
-28.43 -1.04 10-4 -27.31 -5.51 10-5 -27.73 -8.20 10-5 -27.79 -8.60 10-5
-28.19 -2.44 10-4 -30.39 -1.05 10-3 -30.08 -8.20 10-4 -29.89 -7.09 10-4
-23.66 -5.84 10-6 -21.99 -1.71 10-6 -22.67 -3.29 10-6 -22.81 -3.73 10-6
-23.99 -3.46 10-6 -24.60 -1.05 10-5 -24.30 -8.28 10-6 -24.22 -7.79 10-6
-25.02 -2.44 10-5 -23.55 -5.71 10-6 -24.22 -1.15 10-5 -24.35 -1.31 10-5
-24.60 -9.13 10-6 -24.72 -7.61 10-6 -24.75 -7.72 10-6 -24.73 -7.70 10-6

99
-27.89 -1.36 10-4 -28.63 -3.81 10-4 -28.16 -2.27 10-4 -28.04 -2.01 10-4
-25.59 -5.35 10-5 -23.72 -8.40 10-6 -24.32 -1.50 10-5 -24.41 -1.72 10-5
-27.15 -8.03 10-5 -26.21 -3.81 10-5 -26.52 -4.84 10-5 -26.54 -4.94 10-5
-38.82 -3.58 10-1 -37.64 -1.19 10-1 -38.08 -1.86 10-1 -38.19 -2.07 10-1
-19.34 -3.31 10-7 -18.35 -1.59 10-7 -18.78 -2.50 10-7 -18.83 -2.61 10-7
-16.16 -9.54 10-9 -16.71 -1.09 10-8 -16.70 -1.07 10-8 -16.73 -1.11 10-8
-14.78 -2.40 10-9 -14.32 -1.78 10-9 -14.54 -2.29 10-9 -14.56 -2.24 10-9
-13.06 -1.54 10-9 -12.76 -1.13 10-9 -12.88 -1.27 10-9 -12.91 -1.33 10-9
-10.33 -2.69 10-10 -9.88 -2.56 10-10 -9.93 -2.58 10-10 -9.98 -2.62 10-10

100
 The background contribution from the non-patch-antipatch configurations

was accounted for using Monte Carlo integration. The subregions occupied by the

patch-antipatch pairs with  < -20 kT were excluded in the orientation sampling. The

calculated background contributions to B22 are shown in Table 3.12 and 3.13. The

results show that that the two proteins have differing background B22 trends. The

background B22 is reduced when electrostatics are incorporated for lysozyme, but this

contribution increases as the electrostatics are screened. On the other hand, the

background B22 for chymosin B is enhanced when electrostatics are included. Typical

experimentally measured B22 values are on the order of 10-5 to 10-3 mol ml/g2.

However, the magnitude of the background contribution is still beyond experimentally

measured values despite the exclusion of the strongly attractive patch-antipatch

configurations. This is due to the cumulative contribution from the configurations that

display moderately attractive well depths. These configurations still

disproportionately impact the mean value Iin in the Monte Carlo averaging. It suggests

that Monte Carlo integration may not be suitable even for these moderate

configurations and that the same detailed integration procedure used for the patch-

antipatch pairs is needed. To obtain a better numerical estimate, the background

contribution to B22 was broken down into two components: weak interactions and

moderate interactions.

101
Table 3.12: Background B22 for lysozyme based on short-range non-electrostatic
interactions alone and with addition of electrostatics at pH 7. Patch-
antipatch pairs with  < -20 kT were excluded in the Monte Carlo
integration.

(mol ml/g2)
Non-electrostatics -0.10 ± 0.01
0.10M -0.046 ± 0.003
0.20M -0.060 ± 0.003
0.30M -0.078 ± 0.003

Table 3.13: Background B22 for chymosin B based on short-range interactions

alone and with addition of electrostatics at pH 5. Patch-antipatch
pairs with  < -20 kT were excluded in the Monte Carlo integration.

(mol ml/g2)
Non-electrostatics -0.055 ± 0.003
0.10M -0.070 ± 0.005
0.30M -0.062 ± 0.004
0.40M -0.061 ± 0.004

The configurations that are considered to display weak interactions were

identified by specifying a cutoff well depth cutoff and determining the contributions to

B22 from the configurations with  > cutoff by the Monte Carlo method. To estimate

cutoff, a useful benchmark is the unweighted average well depth from all the sampled
106 configurations, i.e. ≈ -3 kT, which corresponds to Iin 3.9×104 Å3. This value

can be seen as the baseline in the Iin landscape around which most of the peaks from

the weakly attractive configurations are scattered. Monte Carlo integration estimates
an integral I by the product of the mean value of a function f and the region of

integration v

3.25

102
For a function that is characterized as a peak within a domain, the mean value is the

average height that is located somewhere between the apex and base of the peak. The

Iin landscape can be viewed as being composed of discrete peaks in a

multidimensional domain, with each peak having its own mean value Iin .

Configurations with stronger attractions will have taller peaks in the Iin landscape and
therefore will have a greater Iin . Because there are peaks that disproportionately

influence the overall mean value of the Iin function, those peaks should be excluded in

the Monte Carlo averaging. Furthermore, because the baseline Iin defines the

boundary between weak and moderate configurations in the Iin landscape, it is

postulated that configurations with Iin that is equal to or less than the baseline Iin are

appropriate for inclusion in the Monte Carlo integration.

From the general expression in equation 3.25, Iin for a configuration is

determined by scaling its local integrated peak Iconfig by the subdomain of integration

v0

3.26

For = 0.10 radian, the average hypervolume occupied by a patch-antipatch pair is v0
= 1.97 10-4. Furthermore, the empirical correlations of Iconfig from Figures 3.11 and

3.12 provide the functional relationship between Iconfig and /kT for both proteins

3.27

3.28

Substituting equations 3.27 and 3.28 into equation 3.26 gives the relationship between

a configuration’s mean value in the Iin landscape and /kT for a configuration. Using

103
the postulated condition that the weak configurations are ones that have Iin

3.9 104 Å3, it was determined that  -7 kT satisfies this criterion. Therefore, it was

concluded that configurations that have  > cutoff =-7 kT were appropriate for inclusion

in the Monte Carlo averaging.

The contribution from configurations with stronger interaction < cutoff =

-7 kT had to be determined by the patch integration method. However, there is a large

number of such configurations and performing integration for individual patches

would be computationally expensive. Rather than perform a detailed local integration

for each orientation, therefore, this explicit approach was reserved for the strongest

patch-antipatch pairs with < -20 kT (Tables 3.10 and 3.11) and the remaining

integrals with moderate interactions were evaluated using the empirical correlations

for Iconfig for both proteins from equations 3.27 and 3.28. The individual contributions

were summed together, achieving a significant savings in computing time.

The number of moderate configurations was determined using the

observation that the relative frequency of the well depths as a function of  is

relatively consistent even as the degree of sampling increases (Figures 3.2 to 3.4). It

was postulated that the relative frequency of the well depth distribution f provides an

estimate of the fraction of the configuration space that is occupied by configurations in

a particular range of well depths dThereforethe size of the hypervolume V

occupied by configurations with well depth can be represented as

3.29

such that the total hypervolume is conserved

3.30

104
where V is the total hypervolume of the global space equal to 323. As mentioned

previously, the relative frequency of /kT for both proteins is adequately described by

a log normal probability distribution function (Figure 3.5). By knowing the

hypervolume occupied by configurations in a particular range of  and the

hypervolume of a single patch (which is set by the  parameter), the absolute number
of moderate configurations N in this range can be approximated by

3.31

The total B22 contribution from the moderate configurations that have < cutoff is

3.32

calculated using the procedure outlined above at different

ionic strengths for lysozyme at pH 7 and chymosin B at pH 5 is shown in Table 3.14

and Table 3.15, respectively. For lysozyme, decreases as the ionic

strength increases. The background B22 contributions from the weak and moderate

configurations are similar when only short-range interactions are considered.

However, the addition of electrostatics leads to decreased contributions from both

components and the overall background contribution decreases. At each level of ionic

strength, configurations with weak interactions actually contribute more than the

moderate interactions. In other words, the moderate configurations display more

repulsive electrostatics and these electrostatic effects have slightly more impact on B22

for lysozyme.
for chymosin B exhibits different behavior. The addition of

electrostatics increases the overall attraction. Although the weak contribution shows a

trend of increasing attraction as the ionic strength increases, the magnitude of the

105
moderate contribution increases when electrostatics are incorporated, indicating that it

is the moderately attractive configurations that are most responsible for this effect.

The net result is increased attraction due to offsetting the effect of repulsion by the

moderate configurations with attractive electrostatics.

Table 3.14: Background B22 for lysozyme based on short-range non-electrostatics

interactions alone and with addition of electrostatics at pH 7. Patch-
antipatch pairs with  < -7 kT were excluded in the Monte Carlo
integration.

( 104mol ml/g2)
Weak Moderate Total
Non-electrostatics -10.3 -10.4 -20.7
0.10M -5.62 -5.45 -11.1
0.20M -7.76 -6.99 -14.8
0.30M -8.66 -8.58 -17.2

Table 3.15: Background B22 for chymosin B based on short-range non-

electrostatics interactions alone and with addition of electrostatics at
pH 5. Patch-antipatch pairs with  < -7 kT were excluded in the
Monte Carlo integration.

( 104mol ml/g2)
Weak Moderate Total
Non-electrostatics -3.66 -1.73 -5.39
0.10M -3.58 -2.04 -5.62
0.30M -3.65 -1.98 -5.63
0.40M -3.65 -1.76 -5.41

3.4 Discussion
The identification of patch-antipatch pairs for proteins is strongly

dependent on the density of orientation sampling of the two protein molecules.

106
However, the degree of sampling is limited by the available computational power.

Present computing speeds readily allow sampling of order 106 angular configurations

to be performed for identifying patch-antipatch pairs, which is two orders of

magnitude greater than the sampling reported previously by Hloucha et al. for bovine

chymotrypsinogen (78). In their work, the strongest patch-antipatch pairs that were

identified possessed well depths between -13 kT and -15 kT. From this work, most of

the strongest configurations were on the order of at least -20 kT.

The presence of strongly attractive patch-antipatch configurations with

well depths on the order of -20 kT led to significant numerical issues when Monte

Carlo integration was used. While this integration approach was found to be

appropriate for determining the excluded volume contribution to B22, it fails when the

energetic contributions are included. The convergence issues were addressed using a

hybrid Monte Carlo/patch integration method. By performing a detailed integration

over the patch-antipatch subregions using a globally adaptive integration routine, a

functional relationship between the strength of attraction for a configuration and its

contribution to B22 was obtained. However, calculation of B22 could not be completed

for both proteins due to the skewed contributions from anomalously strong patch-

antipatch pairs that possessed well minima that were on the order of -30 kT to -40 kT
(Tables 3.10 and 3.11). This indicates that other effects need to be accounted for in

the calculation of B22.

To make complete and accurate predictions of B22, the effect of hydration

must be accounted for explicitly. In these interaction calculations, an implicit solvent

assumption was made where the effects of the solvent were accounted for by the

Hamaker constant in the Lifshitz-Hamaker model and by an empirical factor in the

107
Lennard-Jones model. The hybrid method is capable of capturing surface

complementarity well, but the effect of strongly bound water molecules is lost.

Explicit inclusion of hydration effects can lead to elimination of high complementarity

protein-protein configurations (patch-antipatch pairs) (135, 136). Such an effect

would be expected to play an important role in accurately calculating B22. This is

illustrated by the contribution of a -40 kT patch-antipatch pair for lysozyme, which

was found to have a dominant B22 contribution that was orders of magnitude greater

than what is experimentally measured. Such an anomalously attractive configuration

could very well be eliminated by one or more strongly bound water molecules. Thus,

hydration effects cannot be neglected and are crucial in the accurate prediction of B22.

The calculations of B22 using the proposed hybrid method include

additional potential sources of uncertainty. First, there is the issue of whether all

patch-antipatch pairs can be identified by a random sampling. The configuration

space is large due to the degrees of freedom in defining the relative orientation of two

protein molecules. Another uncertainty is whether the identified patch-antipatch pairs

are in the optimally aligned orientation. While refinement was attempted by a local
sampling within 0.10 radian around the central orientations of the strongest

configurations, a different energy minimum may or may not be detected if the limits
are expanded. Once again, this issue is due to the size of the configuration space. The

uncertainties in the optimal alignment and absolute population of patch-antipatch pairs

are a cause for concern in calculating B22 using atomistic models. However, the

statistical distribution of well depths from a finite sampling provides valuable

information on the fraction of the global angular space occupied by the different

configurations. This is because the relative frequency of orientations with various

108
levels of attractions is independent of the sample size taken, as shown by the

consistent distribution of well depths for different levels of sampling performed.

Furthermore, the distribution was found to be adequately described by an ideal log

normal probability distribution function. This information significantly aided in the

computation of the background contribution to B22. Another significant uncertainty is

the size of the patch-antipatch pairs. A simplification was made in which all
configurations were assumed to be confined within 0.10 radian around their

respective central orientations. The localized patch integration calculations were

performed using this assumption. Although the B22 contributions from the patch

integration approach are very much dependent on what is chosen for , it is has been

shown that the interactions of the patch-antipatch configurations may be quite

sensitive to perturbations in orientation. Therefore, it can be inferred that these patch-

antipatch pairs occupy only a small fraction of the global configuration space. For =

0.10 radian, the subregion of the configurational hypervolume occupied by each patch-
antipatch pair is on average approximately 1.97 10-4, which is a very small portion of

the global configurational space.

3.5 Conclusions
A detailed numerical approach for computing B22 from atomistic models

of proteins was proposed and carried out. The issues this work has attempted to

address illustrate the difficulty and uncertainties of computing B22 at the atomistic

scale. The results highlight the influence of structure on the anisotropy of protein-

protein interactions and therefore the solution properties of proteins. The

heterogeneity of the Iin function makes calculation of the B22 integral particularly

challenging. The proposed hybrid approach appears to provide a better method of

109
calculating B22 at the atomistic scale when compared to the Monte Carlo approach.

However, the identification of anomalously attractive configurations with well depths

~ -30 kT to -40 kT strongly emphasizes the importance and necessity of incorporating

hydration effects to make accurate predictions of B22, which will require the use of

molecular dynamics simulations.

A particular aspect of this work is accounting for the charge distribution of

the protein surface in representing protein interactions. The effect of electrostatics on

the B22 trends for lysozyme and chymosin B was captured by incorporating a screened

Coulombic potential contribution in the pairwise atomistic interaction calculations.

This approach provided a simple and computationally efficient way of accounting not

only for the charge distribution, but also the effect of shape on the electrostatic

interactions. By incorporating the effects of electrostatics, the qualitative trends in B22

were distinguished for the two proteins studied that agreed with experimental

observations. Simple colloidal models that do not account for this important feature

are unable to qualitatively predict differences in the solution behavior of proteins.

110
 Chapter 4

CONCLUSIONS AND RECOMMENDATIONS

The central theme of this thesis is elucidation of the relation between

the molecular structure of proteins and their continuum thermodynamic properties.

The structural properties of proteins are complex, and it is this complexity that has a

profound impact on the molecular interactions that ultimately dictate their

macroscopic solution properties. The osmotic second virial coefficient, B22, provides

a promising qualitative link between the protein-protein interactions and phase

behavior of proteins. This relationship was explored quantitatively, which yielded

significant insights into which essential aspects of protein interactions must be

incorporated in thermodynamic models in order to lead to accurate predictions of the

solution behavior for proteins. In this chapter, the findings and conclusions from this

work are summarized and recommendations for future investigations are put forth.

4.1 Conclusions

4.1.1 Continuum Thermodynamic Models

In Chapter 2, an attempt was made to quantitatively relate experimental

B22 values with the phase diagrams for the model protein ribonuclease A. Several

continuum models derived from classical theories for polymers and colloids were

explored in an effort find a mechanistic framework for protein solutions. While a

qualitative correlation between B22 and phase behavior was found, quantitative

agreement could not be obtained using the continuum models. Phase equilibrium was

111
also predicted from osmotic virial coefficients using the osmotic virial equation

derived from McMillan-Mayer solution theory. Although theoretically calculated

third virial coefficients along with B22 values were used in the phase equilibrium

calculations, the phase diagram predicted from this model only qualitatively agreed

with experimental results.

The discrepancy of the results from the continuum models may be due to

the isotropic assumption inherent in each model and also the molecular nature of B22.

The orientationally-averaged character of B22 provides an incomplete description of

protein-protein interactions and therefore B22 is limited in its ability to predict phase

behavior. It is clear that the anisotropic character of protein-protein interactions

cannot be neglected, providing the most likely explanation for the inadequacy of the

continuum models studied. This finding justified the need for exploring molecular-

level models that incorporate the anisotropy as an essential feature.

4.1.2 Patch-Antipatch Model of Proteins and the Calculation of B22

Patch models have proven to be useful in providing a coarse-grained

representation of proteins to model their anisotropic interactions. They have been

shown to provide a better description of protein phase behavior and much progress has
been made in understanding the phase diagrams predicted from these models.

Accounting for the specific interactions through incorporation of patches can explain

features of protein phase behavior not possible with isotropic models. However,

because of the large parameter space that is possible for different combinations of

patch parameters, much of the work in the literature has focused on simplified patch

representations that provide only a caricature of proteins. The parameters of patch

models should be faithful to the physical and structural attributes of proteins.

112
 The patch-antipatch parameters for two proteins were determined from

extensive atomistic simulations. Several new findings are reported in this thesis

beyond the work of Neal et al. (82) and Hloucha et al. (78). First, the orientational

sampling for identifying patch-antipatch pairs was increased from 104 configurations

to 106 configurations. As a consequence of this increase in sampling, a larger number

of patch-antipatch pairs with deeper attractive wells were identified. Local

refinements were performed to find the approximate energy minima, which in some

cases led to configurations with well depths as a high as -40 kT. The large size of the

global configurational space leads to uncertainties in identifying all patch-antipatch

pairs. However, the relative frequency distribution of interaction well depths was

found to be independent of the degree of orientation sampling and thus provides an

estimate of the fraction of the configuration space that is occupied by the different

patch-antipatch pairs.

It has been shown in this work that the patch-antipatch pairs contribute

significantly to the calculation of B22. Due to these patch-antipatch configurations, the

Monte Carlo integration approach was found to be unsuitable for evaluating the

multidimensional integral necessary for computing B22. The hybrid Monte

Carlo/patch integration approach that is proposed in Chapter 3 addresses some of the

numerical issues in calculating B22. However, the presence of anomalously strong

patch-antipatch pairs gave rise to B22 contributions that were orders of magnitude

greater than typical experimental values. This discrepancy may largely be attributed

to the fact that hydration effects were neglected in the calculations. Hydration effects

are known to attenuate the interactions of patch-antipatch pairs and will have a

significant impact on the calculation of B22. Thus, further refinements are needed in

113
order to make accurate predictions of B22, which will allow for meaningful

comparisons with experimentally measured values.

4.2 Recommendations and Future Directions

4.2.1 The Calculation of B22

Accurate predictions of B22 from atomistic models will require accounting

for the specific hydration of proteins, which is known to affect high-complementarity

configurations (patch-antipatch pairs) (135, 136). The hybrid continuum/atomistic

approach used to calculate the short-range non-electrostatic interaction energies

captures the effect of surface complementarity between proteins, but the effect of

strongly bound water molecules is not generally accounted for. These water

molecules strongly bound to the protein provide additional steric hindrance that can

effectively inhibit the interactions between patch-antipatch pairs. Accounting for such

effects will require the use of such methods as molecular dynamics simulations where

water molecules are explicitly included.

Another direction that should be explored is the calculation of electrostatic

interactions. The addition of a pairwise screened Coulomb contribution to the

interaction energies provided a simple way of accounting for the shape anisotropy of

the charge distribution in the calculation of B22. This approach was able to

qualitatively reflect the impact of this anisotropy on B22 for lysozyme and chymosin B.

However, this model is simplistic in its treatment of the electrostatics since it does not

explicitly account for the local dielectric boundaries. It would be necessary to

compare how the magnitude of the electrostatic energies computed from the pairwise

screened Coulomb potential approach compares with the energies determined from

114
solving the Poisson-Boltzmann equation for the full protein geometry. Several

software packages that utilize finite-difference or boundary-element approaches for

solving the Poisson-Boltzmann equation can be used for such a comparison.

Once these further refinements in the calculation procedure of B22 are

performed and affirmed, the methodology can be extended to determining cross-

interaction virial coefficients B23, which characterize the association of two different

protein molecules. Experimental B23 data for protein mixtures have previously been

measured and reported in the literature (123, 143). Quantitative predictions for the

cross-interactions of proteins can have relevant applications in industrial protein

separation processes.

4.2.2 Molecular Simulation of Protein Phase Behavior

Another avenue that should be explored further is predicting phase

behavior from the “patch-antipatch” representation of proteins. The patch-antipatch

parameters determined from the atomistic simulations performed in this work should

be utilized in the simulation of the phase behavior for these proteins. Although

atomistic models of the kind used in Chapter 3 are, in principle, suitable for describing

crystalline phases (144), such computations would be very challenging based on

current computational capabilities. However, patch models based on spheres may be

appropriate for simulating phases such as liquids since these disordered states do not

require a very specific 3-D molecular packing structure.

The logical route towards this endeavor is through the use of molecular

simulation (145, 146). The patch-antipatch parameters specific to the protein of

interest can be determined a priori from atomistic calculation methods described in

Chapter 3. There are various standard molecular simulation techniques for predicting

115
phase behavior, which include the Gibbs ensemble Monte Carlo (147–150) and grand

canonical Monte Carlo/histogram reweighting techniques (150, 151). An in-depth

discussion of these techniques is beyond the scope of this work, but it is clear that

because of the strong attractions due to the specific patch interactions, particles may

be trapped by the very strong attractions in some configurations (152). As a result,

sampling of the entire configuration space becomes prohibitively long and the system

becomes nonergodic. Advanced simulation methods for strongly associating fluids

will need to be implemented. Several biased methods have been reported to address

such issues, which include the bond-biased Monte Carlo method (153), association-

biased Monte Carlo method (154), and aggregation volume bias Monte Carlo method

(155). These methods improve upon standard simulation methods by biasing the

acceptance criterion for accepting trial moves that allow for more efficient sampling of

the configuration space. These biased methods should be utilized to simulate the

equilibrium phases of proteins.

116
 REFERENCES

1. Ishimoto, C., and T. Tanaka. 1977. Critical behavior of a binary mixture of

protein and salt water. Physical Review Letters. 39: 474–477.

2. Phillies, G. 1985. Comment on “Critical behavior of a binary mixture of protein

and salt water.”Physical Review Letters. 55: 1341-1341.

3. Taratuta, V.G., A. Holschbach, G.M. Thurston, D. Blankschtein, and G.B.

Benedek. 1990. Liquid-liquid phase separation of aqueous lysozyme solutions:
effects of pH and salt identity. Journal of Physical Chemistry. 94: 2140–2144.

4. Broide, M., T. Tominc, and M. Saxowsky. 1996. Using phase transitions to

investigate the effect of salts on protein interactions. Physical Review E. 53:
6325-6335.

5. Asherie, N., A. Lomakin, and G. Benedek. 1996. Phase diagram of colloidal

solutions. Physical Review Letters. 77: 4832-4835.

6. Foffi, G., G. McCullagh, A. Lawlor, E. Zaccarelli, K. Dawson, F. Sciortino, P.

Tartaglia, D. Pini, and G. Stell. 2002. Phase equilibria and glass transition in
colloidal systems with short-ranged attractive interactions: application to
protein crystallization. Physical Review E. 65: 1-17.

7. Broide, M.L., C.R. Berland, J. Pande, O.O. Ogun, and G.B. Benedek. 1991.
Binary-liquid phase separation of lens protein solutions. Proceedings of the
National Academy of Sciences of the United States of America. 88: 5660-5664.

8. Schurtenberger, P., R. Chamberlin, G. Thurston, J. Thomson, and G. Benedek.

1989. Observation of critical phenomena in a protein-water solution. Physical
Review Letters. 63: 2064-2067.

9. Thomson, J.A., P. Schurtenberger, G.M. Thurston, and G.B. Benedek. 1987.

Binary liquid phase separation and critical phenomena in a protein/water
solution. Proceedings of the National Academy of Sciences of the United States
of America. 84: 7079-7083.

117
10. Berland, C.R., G.M. Thurston, M. Kondo, M.L. Broide, J. Pande, O. Ogun, and
G.B. Benedek. 1992. Solid-liquid phase boundaries of lens protein solutions.
Proceedings of the National Academy of Sciences of the United States of
America. 89: 1214-1218.

11. Muschol, M., and F. Rosenberger. 1997. Liquid–liquid phase separation in

supersaturated lysozyme solutions and associated precipitate
formation/crystallization. The Journal of Chemical Physics. 107: 1953-1962.

12. ten Wolde, P.R., and D. Frenkel. 1997. Enhancement of protein crystal
nucleation by critical density fluctuations. Science. 277: 1975-1978.

13. ten Wolde, P.R., and D. Frenkel. 1999. Enhanced protein crystallization around
the metastable critical point. Theoretical Chemistry Accounts: Theory,
Computation, and Modeling. 101: 205–208.

14. Hiemenz, P., and R. Rjagopalan. 1997. Principles of colloid and surface
chemistry. 3rd ed. New York: Marcel Dekker.

15. Israelachvili, J. 1992. Intermolecular surfaces and forces. 2nd ed. Academic
Press.

16. Leckband, D., and J. Israelachvili. 2001. Intermolecular forces in biology.

Quarterly Reviews of Biophysics. 34: 105-267.

17. Kulkarni, A.M., A.P. Chatterjee, K.S. Schweizer, and C.F. Zukoski. 2000.
Effects of polyethylene glycol on protein interactions. The Journal of Chemical
Physics. 113: 9863-9873.

18. Roth, C.M., B.L. Neal, and A.M. Lenhoff. 1996. Van der Waals interactions
involving proteins. Biophysical Journal. 70: 977-987.

19. Collins, K.D. 2004. Ions from the Hofmeister series and osmolytes: effects on
proteins in solution and in the crystallization process. Methods. 34: 300-311.

20. Hofmeister, F. 1888. Zur Lehre von der Wirkung der Salze. Arch. Exp. Pathol.
Pharmakol. 24: 247-260.

21. Kunz, W. 2010. Specific ion effects. 1st ed. World Scientific.

22. Kunz, W. 2010. Specific ion effects in colloidal and biological systems. Current
Opinion in Colloid & Interface Science. 15: 34-39.

118
23. Wilson, E.K. 2007. A renaissance for Hofmeister. Chemical & Engineering
News. 85: 47–49.

24. Kunz, W., P. Lo Nostro, and B.W. Ninham. 2004. The present state of affairs
with Hofmeister effects. Current Opinion in Colloid & Interface Science. 9: 1-
18.

25. Tavares, F.W., M. Boström, E.R.A. Lima, and E.C. Biscaia Jr. 2010. Ion-
specific thermodynamic properties of colloids and proteins. Fluid Phase
Equilibria. 296: 99-105.

26. Lima, E.R.A., E.C. Biscaia Jr., M. Boström, and F.W. Tavares. 2010. Ion-
specific thermodynamical properties of aqueous proteins. Anais da Academia
Brasileira de Ciências. 82: 109-126.

27. Omta, A.W., M.F. Kropman, S. Woutersen, and H.J. Bakker. 2003. Influence of
ions on the hydrogen-bond structure in liquid water. The Journal of Chemical
Physics. 119: 12457-12461.

28. Omta, A.W., M.F. Kropman, S. Woutersen, and H.J. Bakker. 2003. Negligible
effect of ions on the hydrogen-bond structure in liquid water. Science. 301:
347-349.

29. Zhang, Y., and P.S. Cremer. 2006. Interactions between macromolecules and
ions: The Hofmeister series. Current Opinion in Chemical Biology. 10: 658-
663.

30. Ninham, B.W., and V. Yaminsky. 1997. Ion binding and ion specificity: the
Hofmeister effect and Onsager and Lifshitz theories. Langmuir. 13: 2097-2108.

31. Boström, M., D.R.M. Williams, and B.W. Ninham. 2003. Specific ion effects:
why the properties of lysozyme in salt solutions follow a Hofmeister series.
Biophysical Journal. 85: 686-694.

32. Boström, M., D. Williams, and B. Ninham. 2001. Specific ion effects: why
DLVO theory fails for biology and colloid systems. Physical Review Letters.
87: 1-4.

33. Lima, E.R.A., E.C. Biscaia, M. Bostrom, F.W. Tavares, and J.M. Prausnitz.
2007. Osmotic second virial coefficients and phase diagrams for aqueous
proteins from a much-improved Poisson-Boltzmann equation. Journal of
Physical Chemistry C. 111: 16055-16059.

119
34. Lettieri, S. 2010. Specific ion effects and the phase diagram of lysozyme.
Physics Procedia. 6: 46-51.

35. Lettieri, S., X. Li, and J. Gunton. 2008. Hofmeister effect and the phase
diagram of lysozyme. Physical Review E. 78: 1-6.

36. Ries-Kautt, M.M., and A.F. Ducruix. 1989. Relative effectiveness of various
ions on the solubility and crystal growth of lysozyme. The Journal of Biological
Chemistry. 264: 745-748.

37. Lund, M., and P. Jungwirth. 2008. Patchy proteins, anions and the Hofmeister
series. Journal of Physics: Condensed Matter. 20: 494218.

38. Lund, M., and P. Jungwirth. 2008. Ion specific protein assembly and
hydrophobic surface forces. Physical Review Letters. 100: 1-4.

39. Lund, M., R. Vacha, and P. Jungwirth. 2008. Specific ion binding to
macromolecules: effects of hydrophobicity and ion pairing. Langmuir. 24:
3387-3391.

40. Horinek, D., and R.R. Netz. 2007. Specific ion adsorption at hydrophobic solid
surfaces. Physical Review Letters. 99: 48-51.

41. Lettieri, S., X. Li, and J. Gunton. 2009. Ion specific effects on phase transitions
in protein solutions. Physical Review E. 79: 1-6.

42. McQuarrie, D.A. 2000. Statistical mechanics. Mill Valley, CA: University
Science Books.

43. Curtis, R., and L. Lue. 2006. A molecular approach to bioseparations: protein–
protein and protein–salt interactions. Chemical Engineering Science. 61: 907-
923.

44. Piazza, R. 2000. Interactions and phase transitions in protein solutions. Current
Opinion in Colloid & Interface Science. 5: 38-43.

45. Dumetz, A.C., A.M. Chockla, E.W. Kaler, and A.M. Lenhoff. 2008. Protein
phase behavior in aqueous solutions: crystallization, liquid-liquid phase
separation, gels, and aggregates. Biophysical Journal. 94: 570-583.

46. Gast, A.P., C.K. Hall, and W.B. Russel. 1983. Polymer-induced phase
separations in nonaqueous colloidal suspensions. Journal of Colloid and
Interface Science. 96: 251-267.

120
47. Ilett, S., A. Orrock, W. Poon, and P. Pusey. 1995. Phase behavior of a model
colloid-polymer mixture. Physical Review E. 51: 1344-1352.

48. Rosenbaum, D., P. Zamora, and C. Zukoski. 1996. Phase behavior of small
attractive colloidal particles. Physical Review Letters. 76: 150-153.

49. Rosenbaum, D., and C. Zukoski. 1996. Protein interactions and crystallization.
Journal of Crystal Growth. 169: 752-758.

50. Miller, M.A., and D. Frenkel. 2004. Simulating colloids with Baxter’s adhesive
hard sphere model. Journal of Physics: Condensed Matter. 16: S4901-S4912.

51. Miller, M.A., and D. Frenkel. 2004. Phase diagram of the adhesive hard sphere
fluid. The Journal of Chemical Physics. 121: 535-545.

52. Liu, H., S. Garde, and S. Kumar. 2005. Direct determination of phase behavior
of square-well fluids. The Journal of Chemical Physics. 123: 174505.

53. Lomakin, A., N. Asherie, and G.B. Benedek. 1996. Monte Carlo study of phase
separation in aqueous protein solutions. The Journal of Chemical Physics. 104:
1646-1656.

54. Pagan, D.L., and J.D. Gunton. 2005. Phase behavior of short-range square-well
model. The Journal of Chemical Physics. 122: 184515.

55. Duda, Y. 2009. Square-well fluid modeling of protein liquid-vapor coexistence.

The Journal of Chemical Physics. 130: 116101.

56. Hagen, M.H.J., and D. Frenkel. 1994. Determination of phase diagrams for the
hard-core attractive Yukawa system. The Journal of Chemical Physics. 101:
4093.

57. Shukla, K.P. 2000. Phase equilibria and thermodynamic properties of hard core
Yukawa fluids of variable range from simulations and an analytical theory. The
Journal of Chemical Physics. 112: 10358.

58. Lutsko, J.F., and G. Nicolis. 2005. The effect of the range of interaction on the
phase diagram of a globular protein. The Journal of Chemical Physics. 122:
244907.

59. Noro, M.G., and D. Frenkel. 2000. Extended corresponding-states behavior for
particles with variable range attractions. The Journal of Chemical Physics. 113:
2941.

121
60. Curtis, R.A., J.M. Prausnitz, and H.W. Blanch. 1998. Protein-protein and
protein-salt interactions in aqueous protein solutions containing concentrated
electrolytes. Biotechnology and Bioengineering. 57: 11-21.

61. Gunton, J.D., A. Shiryayev, and D.L. Pagan. 2007. Protein condensation. 1st
ed. New York: Cambridge University Press.

62. Li, J., R. Rajagopalan, and J. Jiang. 2010. Molecular modeling and simulation
for phase behavior of protein solutions. Biomolecular Engineering. 661: 75-
130.

63. Li, J. 2008. Phase separations in protein solutions: a Monte Carlo simulation
study. Ph.D. Thesis, National University of Singapore.

64. Rosenbaum, D.F., A. Kulkarni, S. Ramakrishnan, and C.F. Zukoski. 1999.

Protein interactions and phase behavior: sensitivity to the form of the pair
potential. The Journal of Chemical Physics. 111: 9882.

65. Vliegenthart, G.A., and H.N.W. Lekkerkerker. 2000. Predicting the gas–liquid
critical point from the second virial coefficient. The Journal of Chemical
Physics. 112: 5364.

66. Pellicane, G., D. Costa, and C. Caccamo. 2004. Theory and simulation of short-
range models of globular protein solutions. Journal of Physics: Condensed
Matter. 16: S4923-S4936.

67. Pellicane, G., D. Costa, and C. Caccamo. 2003. Phase coexistence in a DLVO
model of globular protein solutions. Journal of Physics: Condensed Matter. 15:
375-384.

68. Pellicane, G., D. Costa, and C. Caccamo. 2004. Microscopic determination of

the phase diagrams of lysozyme and gamma-crystallin solutions. The Journal of
Physical Chemistry B. 108: 7538–7541.

69. Petsev, D.N., and P.G. Vekilov. 2000. Evidence for non-DLVO hydration
interactions in solutions of the protein apoferritin. Physical Review Letters. 84:
1339-1342.

70. San Biagio, P.L., and M.U. Palma. 1991. Spinodal lines and Flory-Huggins
free-energies for solutions of human hemoglobins HbS and HbA. Biophysical
Journal. 60: 508-512.

122
71. Carlsson, F., M. Malmsten, and P. Linse. 2001. Monte Carlo simulations of
lysozyme self-association in aqueous solution. Journal of Physical Chemistry B.
105: 12189-12195.

72. Rosch, T.W., and J.R. Errington. 2007. Investigation of the phase behavior of
an embedded charge protein model through molecular simulation. The Journal
of Physical Chemistry B. 111: 12591-12598.

73. Kern, N., and D. Frenkel. 2003. Fluid–fluid coexistence in colloidal systems
with short-ranged strongly directional attraction. The Journal of Chemical
Physics. 118: 9882.

74. Bianchi, E., R. Blaak, and C.N. Likos. 2011. Patchy colloids: state of the art and
perspectives. Physical Chemistry Chemical Physics. 13: 6397-6410.

75. Lomakin, A., N. Asherie, and G.B. Benedek. 1999. Aeolotopic interactions of
globular proteins. Proceedings of the National Academy of Sciences of the
United States of America. 96: 9465-9468.

76. Sear, R.P. 1999. Phase behavior of a simple model of globular proteins. The
Journal of Chemical Physics. 111: 4800.

77. Gögelein, C., G. Nägele, R. Tuinier, T. Gibaud, A. Stradner, and P.

Schurtenberger. 2008. A simple patchy colloid model for the phase behavior of
lysozyme dispersions. The Journal of Chemical Physics. 129: 085102.

78. Hloucha, M., J.F.M. Lodge, A.M. Lenhoff, and S.I. Sandler. 2001. A patch–
antipatch representation of specific protein interactions. Journal of Crystal
Growth. 232: 195-203.

79. Liu, H., S.K. Kumar, F. Sciortino, and G.T. Evans. 2009. Vapor-liquid
coexistence of fluids with attractive patches: an application of Wertheim’s
theory of association. The Journal of Chemical Physics. 130: 044902.

80. Zimm, B.H. 1946. Application of the methods of molecular distribution to

solutions of large molecules. The Journal of Chemical Physics. 14: 164-179.

81. Goldstein, H. 1950. Classical mechanics. Cambridge: Addison-Wesley Press

Inc.

82. Neal, B.L., D. Asthagiri, and A.M. Lenhoff. 1998. Molecular origins of osmotic
second virial coefficients of proteins. Biophysical Journal. 75: 2469-2477.

123
83. Neal, B., D. Asthagiri, O.D. Velev, A.M. Lenhoff, and E.W. Kaler. 1999. Why
is the osmotic second virial coefficient related to protein crystallization? Journal
of Crystal Growth. 196: 377-387.

84. Chang, R.C., D. Asthagiri, and A.M. Lenhoff. 2000. Measured and calculated
effects of mutations in bacteriophage T4 lysozyme on interactions in solution.
Proteins. 41: 123-132.

85. Guo, B. 1999. Correlation of second virial coefficients and solubilities useful in
protein crystal growth. Journal of Crystal Growth. 196: 424-433.

86. Dumetz, A.C., A.M. Snellinger-O’Brien, E.W. Kaler, and A.M. Lenhoff. 2007.
Patterns of protein protein interactions in salt solutions and implications for
protein crystallization. Protein Science. 16: 1867-1877.

87. Lewus, R.A., P.A. Darcy, A.M. Lenhoff, and S.I. Sandler. 2011. Interactions
and phase behavior of a monoclonal antibody. Biotechnology Progress. 27:
280-289.

88. George, A., and W.W. Wilson. 1994. Predicting protein crystallization from a
dilute solution property. Acta Crystallographica. Section D, Biological
Crystallography. 50: 361-365.

89. Asherie, N. 2004. Protein crystallization and phase diagrams. Methods. 34: 266-
272.

90. Flory, P.J. 1953. Principles of polymer chemistry. 1st ed. Ithaca, NY: Cornell
University Press.

91. Haas, C., and J. Drenth. 1998. The protein−water phase diagram and the growth
of protein crystals from aqueous solution. The Journal of Physical Chemistry B.
102: 4226-4232.

92. Haas, C., and J. Drenth. 2000. The interface between a protein crystal and an
aqueous solution and its effects on nucleation and crystal growth. The Journal
of Physical Chemistry B. 104: 368-377.

93. Haas, C., and J. Drenth. 1999. Understanding protein crystallization on the basis
of the phase diagram. Journal of Crystal Growth. 196: 388-394.

94. McMillan Jr, W.G., and J.E. Mayer. 1945. The statistical thermodynamics of
multicomponent systems. The Journal of Chemical Physics. 13: 276-305.

124
95. Tessier, P.M., S.D. Vandrey, B.W. Berger, R. Pazhianur, S.I. Sandler, and A.M.
Lenhoff. 2002. Self-interaction chromatography: a novel screening method for
rational protein crystallization. Acta Crystallographica Section D Biological
Crystallography. 58: 1531-1535.

96. Tessier, P.M., A.M. Lenhoff, and S.I. Sandler. 2002. Rapid measurement of
protein osmotic second virial coefficients by self-interaction chromatography.
Biophysical Journal. 82: 1620-1631.

97. Sandler, S.I. 2006. Introduction to chemical, biochemical, and engineering

thermodynamics. 4th ed. John Wiley & Son.

98. Haas, C., and J. Drenth. 1995. The interaction energy between two protein
molecules related to physical properties of their solution and their crystals and
implications for crystal growth. Journal of Crystal Growth. 154: 126-135.

99. Overbeek, J.T.G. 1978. The first Rideal lecture. Microemulsions, a field at the
border between lyophobic and lyophilic colloids. Faraday Discussions of the
Chemical Society. 65: 7–19.

100. Carnahan, N.F., and K.E. Starling. 1970. Thermodynamic properties of a rigid-
sphere fluid. The Journal of Chemical Physics. 53: 600-603.

101. Rowlinson, J.S. 1989. The Yukawa potential. Physica A: Statistical Mechanics
and its Applications. 156: 15-34.

102. Sherwood, A., and J. Prausnitz. 1964. Third virial coefficient for the Kihara,
exp-6, and square-well potentials. The Journal of Chemical Physics. 41: 413-
428.

103. Alder, B.J., and J.A. Pople. 1957. Third virial coefficient for intermolecular
potentials with hard sphere cores. The Journal of Chemical Physics. 26: 325-
328.

104. Graben, H., and R. Present. 1964. Third virial coefficient for the Sutherland (∞,
ν) potential. Reviews of Modern Physics. 36: 1025-1033.

105. Naresh, D.J., and J.K. Singh. 2009. Virial coefficients of hard-core attractive
Yukawa fluids. Fluid Phase Equilibria. 285: 36-43.

106. Kihara, T. 1953. Virial coefficients and models of molecules in gases. Reviews
of Modern Physics. 25: 831-843.

125
107. Hirschfelder, J.O., C.F. Curtiss, and R.B. Bird. 1954. Molecular theory of gases
and liquids. New York: John Wiley & Sons Inc.

108. Chang, J., A.M. Lenhoff, and S.I. Sandler. 2004. Determination of fluid--solid
transitions in model protein solutions using the histogram reweighting method
and expanded ensemble simulations. The Journal of Chemical Physics. 120:
3003-3014.

109. Liu, H., S.K. Kumar, and F. Sciortino. 2007. Vapor-liquid coexistence of
patchy models: relevance to protein phase behavior. The Journal of Chemical
Physics. 127: 084902.

110. Li, X., J.D. Gunton, and A. Chakrabarti. 2009. A simple model of directional
interactions for proteins. The Journal of Chemical Physics. 131: 115101.

111. Bianchi, E., J. Largo, P. Tartaglia, E. Zaccarelli, and F. Sciortino. 2006. Phase
diagram of patchy colloids: towards empty liquids. Physical Review Letters. 97:
1-4.

112. Fantoni, R., D. Gazzillo, A. Giacometti, M.A. Miller, and G. Pastore. 2007.
Patchy sticky hard spheres: analytical study and Monte Carlo simulations. The
Journal of Chemical Physics. 127: 234507.

113. Bianchi, E., P. Tartaglia, E. Zaccarelli, and F. Sciortino. 2008. Theoretical and
numerical study of the phase diagram of patchy colloids: ordered and
disordered patch arrangements. The Journal of Chemical Physics. 128: 144504.

114. Giacometti, A., F. Lado, J. Largo, G. Pastore, and F. Sciortino. 2010. Effects of
patch size and number within a simple model of patchy colloids. The Journal of
Chemical Physics. 132: 174110.

115. Giacometti, A., F. Lado, J. Largo, G. Pastore, and F. Sciortino. 2009. Phase
diagram and structural properties of a simple model for one-patch particles. The
Journal of Chemical Physics. 131: 174114.

116. Song, X. 2002. Role of anisotropic interactions in protein crystallization.

Physical Review E. 66: 1-4.

117. Romano, F., E. Sanz, and F. Sciortino. 2009. Role of the range in the fluid-
crystal coexistence for a patchy particle model. The Journal of Physical
Chemistry B. 113: 15133-15136.

126
118. Romano, F., E. Sanz, and F. Sciortino. 2010. Phase diagram of a tetrahedral
patchy particle model for different interaction ranges. The Journal of Chemical
Physics. 132: 184501.

119. Noya, E.G., C. Vega, J.P.K. Doye, and A.A. Louis. 2010. The stability of a
crystal with diamond structure for patchy particles with tetrahedral symmetry.
The Journal of Chemical Physics. 132: 234511.

120. Noya, E.G., C. Vega, J.P.K. Doye, and A.A. Louis. 2007. Phase diagram of
model anisotropic particles with octahedral symmetry. The Journal of Chemical
Physics. 127: 054501.

121. Jones, S., and J.M. Thornton. 1996. Principles of protein-protein interactions.
Proceedings of the National Academy of Sciences. 93: 13-20.

122. Asthagiri, D., B. Neal, and A. Lenhoff. 1999. Calculation of short-range

interactions between proteins. Biophysical Chemistry. 78: 219–231.

123. Moon, Y.U., R.A. Curtis, C.O. Anderson, H.W. Blanch, and J.M. Prausnitz.
2000. Protein – protein interactions in aqueous ammonium sulfate solutions.
Lysozyme and bovine serum albumin ( BSA ). Journal of Solution Chemistry.
29: 699-717.

124. Maurer, R.W., S.I. Sandler, and A.M. Lenhoff. 2011. Salting-in characteristics
of globular proteins. Biophysical Chemistry. 156: 72-78.

125. Roth, C., and A.M. Lenhoff. 1996. Improved parametric representation of water
dielectric data for Lifshitz theory calculations. Journal of Colloid and Interface
Science. 179: 637-639.

126. Hamaker, H. 1937. The London—van der Waals attraction between spherical
particles. Physica. 4: 1058-1072.

127. Jorgensen, W.L., and J. Tirado-Rives. 1988. The OPLS potential functions for
proteins. Energy minimizations for crystals of cyclic peptides and crambin.
Journal of the American Chemical Society. 110: 1657–1666.

128. Horton, N., and M. Lewis. 1992. Calculation of the free energy of association
for protein complexes. Protein Science. 1: 169-181.

129. Li, H., A.D. Robertson, and J.H. Jensen. 2005. Very fast empirical prediction
and rationalization of protein pKa values. Proteins. 61: 704-721.

127
130. Olsson, M.H.M., C.R. Søndergaard, M. Rostkowski, and J.H. Jensen. 2011.
PROPKA3: Consistent treatment of internal and surface residues in empirical
pKa predictions. Journal of Chemical Theory and Computation. 7: 525-537.

131. Yoon, B.J., and A.M. Lenhoff. 1992. Computation of the electrostatic
interaction energy between a protein and a charged surface. The Journal of
Physical Chemistry. 96: 3130–3134.

132. Yoon, B.J., and A.M. Lenhoff. 1990. A boundary element method for molecular
electrostatics with electrolyte effects. Journal of Computational Chemistry. 11:
1080-1086.

133. McGuffee, S.R., and A.H. Elcock. 2006. Atomically detailed simulations of
concentrated protein solutions: the effects of salt, pH, point mutations, and
protein concentration in simulations of 1000-molecule systems. Journal of the
American Chemical Society. 128: 12098-110.

134. Mereghetti, P., R.R. Gabdoulline, and R.C. Wade. 2010. Brownian dynamics
simulation of protein solutions: structural and dynamical properties.
Biophysical Journal. 99: 3782-3791.

135. Paliwal, A., D. Asthagiri, D. Abras, A.M. Lenhoff, and M.E. Paulaitis. 2005.
Light-scattering studies of protein solutions: role of hydration in weak protein-
protein interactions. Biophysical Journal. 89: 1564-1573.

136. Asthagiri, D., A. Paliwal, D. Abras, A.M. Lenhoff, and M.E. Paulaitis. 2005. A
consistent experimental and modeling approach to light-scattering studies of
protein-protein interactions in solution. Biophysical Journal. 88: 3300-3309.

137. Gillespie, C. 2011. Protein solution thermodynamics: polymorphic

crystallization, liquid-liquid phase separation and osmotic second virial
coefficients. Ph.D. Thesis, University of Delaware.

138. Press, W.H., B.P. Flannery, S.A. Teukolosky, and W.T. Vetterling. 1986.
Numerical recipes: the art of scientific computing. Cambridge: Cambridge
University Press.

139. Berntsen, J., T.O. Espelid, and A. Genz. 1991. Algorithm 698; DCUHRE: an
adaptive multidemensional integration routine for a vector of integrals. ACM
Transactions on Mathematical Software. 17: 452-456.

128
140. Berntsen, J., T.O. Espelid, and A. Genz. 1991. An adaptive algorithm for the
approximate calculation of multiple integrals. ACM Transactions on
Mathematical Software. 17: 437-451.

141. Ogunnaike, B.A. 2010. Random phenomena: fundamentals of probability and

statistics for engineers. Boca Raton: CRC Press.

142. Neal, B.L., and A.M. Lenhoff. 1995. Excluded volume contribution to the
osmotic second virial coefficient for proteins. AIChE Journal. 41: 1010-1014.

143. Cheng, Y.-C., C.L. Bianco, S.I. Sandler, and A.M. Lenhoff. 2008. Salting-out
of lsozyme and ovalbumin from mixtures: predicting precipitation performance
from protein−protein interactions. Industrial & Engineering Chemistry
Research. 47: 5203-5213.

144. Chang, J., A.M. Lenhoff, and S.I. Sandler. 2005. The combined simulation
approach of atomistic and continuum models for the thermodynamics of
lysozyme crystals. Journal of Physical Chemistry B. 109: 19507-19515.

145. Allen, M.P., and D.J. Tildesley. 1987. Computer simulation of liquids. 1st ed.
Oxford University Press.

146. Frenkel, D., and B. Smit. 2002. Understanding molecular simulation: from
algorithms to applications. 2nd ed. Academic Press.

147. Panagiotopoulos, A.Z., V. Wong, and M.A. Floriano. 1998. Phase equilibria of
lattice polymers from histogram reweighting Monte Carlo simulations.
Macromolecules. 31: 912-918.

148. Panagiotopoulos, A. 1987. Direct determination of phase coexistence properties

of fluids by Monte Carlo simulation in a new ensemble. Molecular Physics.
100: 237-246.

149. Panagiotopoulos, A.Z. 1995. Gibbs ensemble techniques. NATO ASI Series C
Mathematical and Physical Sciences-Advanced Study Institute. 460: 463–502.

150. Panagiotopoulos, A. 2000. Monte Carlo methods for phase equilibria of fluids.
Journal of Physics: Condensed Matter. 25: R25-R52.

151. Ferrenberg, A.M., and R.H. Swendsen. 1988. New Monte Carlo technique for
studying phase transitions. Physical Review Letters. 61: 2635-2638.

129
152. Hloucha, M. 1999. Theoretical investigation of protein interactions.
Unpublished Report, University of Delaware.

153. Tsangaris, D.M., and J.J. de Pablo. 1994. Bond-bias simulation of phase
equilibria for strongly associating fluids. The Journal of Chemical Physics. 101:
1477.

154. Busch, N.A., M.S. Wertheim, Y.C. Chiew, and M.L. Yarmush. 1994. A Monte
Carlo method for simulating associating fluids. The Journal of Chemical
Physics. 101: 3147.

155. Chen, B., and J.I. Siepmann. 2000. A novel Monte Carlo algorithm for
simulating strongly associating fluids: applications to water, hydrogen fluoride,
and acetic acid. The Journal of Physical Chemistry B. 104: 8725-8734.

156. Dobert, F., A. Pfennig, and M. Stumpf. 1995. Derivation of the consistent
osmotic virial equation and its application to aqueous poly(ethylene glycol)-
dextran two-phase systems. Macromolecules. 28: 7860-7868.

130
 Appendix A

DERIVATION OF LIQUID-LIQUID EQUILBIRIUM FROM THE OSMOTIC

VIRIAL EQUATION

The derivation of the equations for describing the liquid-liquid equilibrium

of protein solutions from the osmotic virial equation is presented in this appendix.

This derivation is consistent with the one reported by Döbert et al. (156). The protein

solution is modeled as a binary mixture consisting of solvent (1) + protein (2) species.

It is assumed that there is no repartitioning of the salt species in the mixture. The

criterion for liquid-liquid equilibrium is the equality of the chemical potentials in the

mixture of both the solvent, 1, and the protein, 2, in each phase:

(B.1)

where I and II designate the light and dense phases, respectively. To determine the

chemical potentials for the species, it can be shown that the chemical potential of the

solvent is directly related to the osmotic pressure  (14)

(B.2)

where is the chemical potential of the pure solvent, and is the partial molar

volume of the solvent. The osmotic virial equation in terms of protein mass

concentration c and truncated at the third virial term is

(B.3)

131
where R is the molar gas constant, MW is the molecular weight of the protein, and B2

and B3 are the second and third virial coefficients, respectively. Substituting equation

(B.3) into equation (B.2) yields the expression for the chemical potential of the solvent

(B.4)

To determine the chemical potential of the protein species, the Gibbs-Duhem relation

(97) is invoked

(B.5)

where N1 and N2 are the number of moles of the solvent and protein species,

respectively. Rearranging equation (B.5) in terms of 2 leads to

(B.6)

where C is the constant of integration. To evaluate the derivative inside the integral of

equation (B.6), it is convenient to write 1 in equation (B.4) in terms of N1 and N2.

The mass concentration of protein c is related to its mole fraction x2 by

(B.7)

where is the molar volume of the solution. For a binary system, is the sum of the

partial molar volumes of the solvent and protein weighted by their respective mole

fractions x1 and x2

(B.8)

Furthermore, a species mole fractions is the ratio of the number of moles of the

species to the total number of moles of the system

132
 (B.9)

By appropriate substitutions of equations (B.7), (B.8), and (B.9) into equation (B.4),

the chemical potential of the solvent 1 can be rewritten in terms of N1 and N2

(B.10)

Inserting equation (B.10) into equation (B.6), differentiating µ1 with respect to N1, and
integrating the result with respect to N2 yields the expression for µ2

(B.11)

The initial condition that when is used to find the integration constant

(B.12)

where is the chemical potential of the pure protein species. The integration
constant C is inserted back into equation (B.11), and 2 becomes

(B.13)

133
The chemical potentials can be expressed in terms of the protein volume fraction  and

the osmotic virial coefficients. The protein volume fraction is

(B.14)

The osmotic virial coefficients are defined as

(B.15)

Using these relationships, the chemical potential of the solvent in equation (B.10) and

the chemical potential of the protein in equation (B.13) can be rewritten in terms of ,

B22, and B222

(B.16)

(B.17)

For dilute concentrations of protein, it can be assumed that the mixture is ideal enough
that the partial molar volume of the protein is equal to its molar volume, = .

Applying the conditions represented in equation (B.1) yields the model for liquid-

liquid equilibrium

(B.18)

134
 (B.19)

in which I and II are the concentration of protein in the light and dense liquid phases,

respectively. Equations (B.18) and (B.19) can be solved simultaneously to obtain the

equilibrium concentrations in the light and dense liquid phases, provided that the
physical properties of the protein (MW and ) are specified and the virial coefficients

B22 and B222 at known solution conditions are used as inputs

135
 Appendix B

CALCULATION OF THE OSMOTIC THIRD VIRIAL COEFFICIENT B222

This appendix contains the MATLAB source code for the calculation of

B222 from the Yukawa potential that is discussed in Chapter 2. The program utilizes

MATLAB’s numerical integration toolbox. In order to run this code, the numerical

integration toolbox needs to be in the same directory as the program files during

execution. This toolbox contains the functions implemented in the code for the

numerical integration of the third virial coefficient, which can be obtained from the

website http://www2.math.umd.edu/~jmr241/mfiles/nit/.

The file names and descriptions are given below. The range parameter b is

set at a value 35 in line 65 of the main driver file B222_Yukawa.m; however, this

value can be adjusted by the user. The user is also free to adjust the values of the

parameters in the file parameters.inp. Attached is a copy of the input files that are

needed. In principle, this code can be modified for any potential that possesses hard-

sphere repulsion with an attractive tail. It is left to the user to make such

modifications.

136
File: parameters.inp – Input file which contains physical parameters

Temp 296
R 8.314
k 1.38E-23
Na 6.023E23
e 1.60E-19
epsr 78.54
eps0 8.85E-12
MW1 18.02
rho1 0.998
MW2 13700
rho2 4.5
sigma 3.1E-9
phi_c 0.5

File: B22_Data.dat – Input file which contains the experimental B22 data

-1.27E-04
-1.55E-04
-1.88E-04
-2.24E-04
-2.65E-04
-3.10E-04
-3.61E-04
-4.17E-04
-4.79E-04
-5.48E-04
-6.24E-04
-7.07E-04
-7.98E-04
-8.97E-04
-1.01E-03
-1.12E-03
-1.25E-03
-1.39E-03

137
File: B222_Yukawa.m – Main driver file for B222 calculation

% This MATLAB program is used to calculate the third virial

% coefficient from the Yukawa potential using the method of
% Alder and Pople, J. Chem Phys., 1957. To run this file,
% the numerical integration toolbox needs to be in the directory.
%
% INPUT
% parameters.inp
% File with necessary physical parameters for calculations
%
% B22_Data
% File with the experimental B22 data. This data is used to
% fit the epsilon parameter of the Yukawa potential for a
% fixed value of the range parameter b*.
%
% OUTPUT
% T*
% Reduced temperature T/epsilon
% B222
% Osmotic third virial coefficient
%
% Authors: Leigh J. Quang
% Abraham Lenhoff
% Stanley Sandler
%
% Last Modified: 11/29/2011
% University of Delaware
% Department of Chemical Engineering

clc
clear all
close all
clear global
addpath nit
global T k Na e epsr eps0 MW1 V1 rho1 phi_c MW2 m s

%% Experimental Data
% Read in data
fid = importdata('parameters.inp');

% Constants
T = fid.data(1); % Temperature (K)
R = fid.data(2); % Molar gas constant (J/molK)
k = fid.data(3); % Boltzmann constant (J/K)
Na = fid.data(4); % Avogadro's number (1/mole)
e = fid.data(5); % Elementary charge (C)

138
epsr = fid.data(6); % Dielectric constant of solvent
eps0 = fid.data(7); % Permitivity constant

% Water Properties
MW1 = fid.data(8); % Molecular weight (g/mol)
rho1 = fid.data(9); % Density (g/ml)
V1 = MW1/rho1; % Molar volume (ml/mol)

% Protein Properties
MW2 = fid.data(10); % Molecular weight (g/mol)
rho2 = fid.data(11); % g/cm^3
V2 = MW2/rho2; % Molar volume (ml/mol)
s = fid.data(12); % Diameter (m)
m = (rho1*MW2)/(rho2*MW1);
phi_c = fid.data(13);

%% B22 Data
ydata = importdata('B22_Data.dat');
n = length(ydata);

%% Range Parameter
b = [35];
m = length(b);

%% B222 Calculation
syms x y r z
epsilon = zeros(n,m);
B3 = zeros(n,m);
Temp = zeros(n,m);
rmax = 5;
for j = 1:1:m
for i = 1:1:n
% Yukawa Epsilon Parameter Optimization
resid = @(eps) norm(B2(eps,b(j))- ydata(i));
epsilon(i,j) = fminsearch(resid,500);

% Third Virial Coefficient from Yukawa

Tr = T/epsilon(i,j);
wx = -exp(-b(j)*(x-1))/x;
wy = -exp(-b(j)*(y-1))/y;
wz = -exp(-b(j)*(z-1))/z;
fx = exp(-wx/Tr)-1;
fy = exp(-wy/Tr)-1;
fz = exp(-wz/Tr)-1;

C1 = 5/8;
C2 = -12*numint2((1-0.75*x+0.0625*x^3)*x^2*fx,x,1,2,y,0,1);
C3 = 36*[numint3(x*y*z*fx*fy,z,x-y,1,y,1,x,x,1,2) + ...
numint3(x*y*z*fx*fy,z,x-y,1,y,x-1,x,x,2,rmax)];
C4 = -12*[numint3(x*y*z*fx*fy*fz,z,1,x+y,y,1,x,x,1,2) + ...
numint3(x*y*z*fx*fy*fz,z,x-y,x+y,y,1,x-1,x,2,rmax) + ...

139
 numint3(x*y*z*fx*fy*fz,z,1,x+y,y,x-1,x,x,2,rmax)];
B3(i,j) = C1+C2+C3+C4;
Temp(i,j) = Tr;
end
end

%% Output results
format short
disp('T*:')
disp(Temp)
disp(' ')
format short eng
disp('3rd Virial Coefficient B222 (cm^6*mol/g^3):')
disp(' ')
B222 = B3.*[((2*pi*Na*s^3)/3).^2.*100^6.*(1/MW2^3)];
disp(B222)

File: B2.m – Integral equation for B22 in terms of reduced variables

function F = B2(eps,b)

% Function file for solving the B22 integral equation to determine

% epsilon. The range parameter b must be passed to this function.

global T k Na e epsr eps0 MW1 V1 rho1 phi_c MW2 m s

F = ((2.*pi.*Na.*s.^3)./3).*(1 + quadgk(@(r) f(r,eps,b),1,Inf));

end

File: f.m – Mayer cluster function for the Yukawa potential

function F = f(r,eps,b)

% Function file that contains the Mayer cluster function for the
% Yukawa potential. This function is using in the function file B2.m
% for MATLAB's quadgk integration function.

global T k Na e epsr eps0 MW1 V1 rho1 phi_c MW2 m s

w = -eps.*exp(-b.*(r.^(1/3)-1))./r.^(1/3);
F = 1 - exp(-w/T);

end

140