Improvements to the APBS biomolecular solvation

arXiv:1707.00027v2 [q-bio.BM] 21 Aug 2017

software suite

Elizabeth Jurrus? , Dave Engel? , Keith Star? , Kyle Monson? , Juan Brandi? ,
Lisa E. Felberg† , David H. Brookes† , Leighton Wilson‡ , Jiahui Chen§ ,
Karina Liles? , Minju Chun? , Peter Li? , David W. Gohara◦ , Todd Dolinskyk ,
Robert Konecny∆ , David R. KoesB , Jens Erik Nielsen¶ ,
Teresa Head-Gordon† , Weihua Geng§ , Robert Krasny‡ , Guo-Wei Wei∇ ,
Michael J. Holst∆ , J. Andrew McCammon∆ , and Nathan A. BakerC
?
Pacific Northwest National Laboratory
†
University of California, Berkeley
‡
University of Michigan
§
Southern Methodist University
◦
St. Louis University
k
FoodLogiQ
B
University of Pittsburgh
¶
Protein Engineering, Novozymes A/S
∇
Michigan State University
∆
University of California San Diego
C
To whom correspondence should be addressed. Advanced Computing,
Mathematics, and Data Division; Pacific Northwest National Laboratory;
Richland, WA 99352, USA. Division of Applied Mathematics; Brown
University; Providence, RI 02912, USA. Email: [email protected]

August 23, 2017

 Abstract

The Adaptive Poisson-Boltzmann Solver (APBS) software was developed to solve the equa-
tions of continuum electrostatics for large biomolecular assemblages that has provided impact
in the study of a broad range of chemical, biological, and biomedical applications. APBS
addresses three key technology challenges for understanding solvation and electrostatics in
biomedical applications: accurate and efficient models for biomolecular solvation and elec-
trostatics, robust and scalable software for applying those theories to biomolecular systems,
and mechanisms for sharing and analyzing biomolecular electrostatics data in the scientific
community. To address new research applications and advancing computational capabilities,
we have continually updated APBS and its suite of accompanying software since its release
in 2001. In this manuscript, we discuss the models and capabilities that have recently been
implemented within the APBS software package including: a Poisson-Boltzmann analytical
and a semi-analytical solver, an optimized boundary element solver, a geometry-based ge-
ometric flow solvation model, a graph theory based algorithm for determining pKa values,
and an improved web-based visualization tool for viewing electrostatics.
1 Introduction
Robust models of electrostatic interactions are important for understanding early molecular
recognition events where long-ranged intermolecular interactions and the effects of solvation
on biomolecular processes dominate. While explicit electrostatic models that treat the so-
lute and solvent in atomic detail are common, these approaches generally require extensive
equilibration and sampling to converge properties of interest in the statistical ensemble of
interest [1]. Continuum approaches that integrate out important, but largely uninteresting
degrees of freedom, sacrifice numerical precision in favor of robust but qualitative accuracy
and efficiency by eliminating the need for sampling and equilibration associated with explicit
solute and solvent models.
While there is a choice among several implicit solvation models [1–6], one of the most
popular implicit solvent models for biomolecules is based on the Poisson-Boltzmann (PB)
equation [7–9]. The PB equation provides a global solution for the electrostatic potential
(φ) within and around a biomolecule by solving the partial differential equation
M
X
− ∇ · ∇φ − ci qi e−β(qi φ+Vi ) = ρ. (1)
i

The solvent is described by the bulk solvent dielectric constant εs as well as the properties
of i = 1, . . . , M mobile ion species, described by their charges qi , concentrations ci , and
steric ion-solute interaction potential Vi . The biomolecular structure is incorporated into the
equation through Vi , a dielectric coefficient function , and a charge distribution function ρ.
The dielectric  is often set to a constant value εm in the interior of the molecule and varies
sharply across the molecular boundary to the value εs which describes the bulk solvent.
The shape of the boundary is determined by the size and location of the solute atoms as
well as model-specific parameters such as a characteristic solvent molecule size [10]. The
charge distribution ρ is usually a sum of Dirac delta distributions which are located at
atom centers. Finally, β = (kT )−1 is the inverse thermal energy where k is the Boltzmann
constant and T is the temperature. The potential φ can be used in a variety of applications,
including visualization, other structural analyses, diffusion simulations, and a number of
other calculations that require global electrostatic properties. The PB theory is approximate
and, as a result, has several well-known limitations which can affect its accuracy, particularly
for strongly charged systems or high salt concentrations [7, 11]. Despite these limitations, PB
methods are still very important and popular for biomolecular structural analysis, modeling,
and simulation.
Several software packages have been developed that solve the Poisson-Boltzmann equa-
tions to evaluate energies, potentials, and other solvation properties. The most significant
(based on user base and citations) of these include CHARMM [12], AMBER [13], DelPhi [14],
Jaguar [15], Zap [16], MIBPB [17], and APBS [18]. However the APBS and associated soft-
ware package PDB2PQR has served a large community of ∼27,000 users by creating web
servers linked from the APBS website [19] that support preparation of biomolecular struc-
tures (see Section 2) and a fast finite-difference solution of the Poisson-Boltzmann equation
(see Section 3.1) that are further augmented with a set of analysis tools. Most APBS elec-
trostatics calculations follow the general workflow outlined in Figure 1. An even broader

1
 Typical APBS & PDB2PQR user workflow
APBS &
User PDB2PQR only APBS only
PDB2PQR

Start

Upload a Assign
Repair heavy
structure or titration
atoms
specify ID states

Add
hydrogens

Assign charge Calculate

& radius electrostatic
parameters properties

Offline
Web-based
visualization
visualization
& analysis

End

Figure 1: Workflow for biomolecular electrostatics calculations using the APBS-PDB2PQR

software suite. This workflow is supported by the APBS tool suite with specific support by
PDB2PQR for preparing biomolecular structures (see Section 2) and support by APBS for
performing electrostatics calculations (see Section 3).

range of features and more flexible configuration is available when APBS and PDB2PQR
are run from the command line on Linux, Mac, and Windows platforms, and which can
be run locally or through web services provided by the NBCR-developed Opal toolkit [20].
This toolkit allows for the computing load for processor intensive scientific applications to
be shifted to remote computing resources such as those provided by the National Biomedical
Computation Resource (NCBR). Finally, APBS can run through other molecular simulation
programs such as AMBER [13], CHARMM [12], NAMD [21], Rosetta [22] and TINKER [23].
General support for integration of APBS with 3rd -party programs is provided by the iAPBS
library [24, 25].
This article provides an overview of the new capabilities in the APBS software and
its associated tools since their release in 2001 [18, 26–28]. In particular new solutions
to the PB equation have been developed and incorporated into APBS: a fully analytical
model based on simple spherical geometries that treats full mutual polarization accurately,
known as PBAM (Poisson-Boltzmann Analytical Model) [29, 30] as well as its extension
to a semi-analytical PBE solver PB-SAM (Poisson-Boltzmann Semi-Analytical Model) that
treats arbitrary shaped dielectric boundaries [31, 32]. APBS-PDB2PQR also now includes

2
an optimized boundary element solver, a geometry-based geometric flow solvation model, a
graph theory based algorithm for determining pKa values, and an improved web-based visu-
alization tool for viewing electrostatics. We describe these new approaches in the remainder
of the paper, and more detailed documentation for these tools is available on the APBS
website [19]. As discussed in the following sections, APBS-PDB2PQR offers a wide range of
features for users across all levels of computational biology expertise.

2 Preparing biomolecular structures

Electrostatics calculations begin with specification of the molecule structure and parameters
for the charge and size of the constituent atoms. Constituent atoms are generally grouped
in types with charge and size values specified by atom type in a variety of force field files
developed for implicit solvent calculations [1]. APBS incorporates this information into
calculations via the “PQR” format. PQR is a file format of unknown origins used by several
software packages such as MEAD [33] and AutoDock [34]. The PQR file simply replaces the
temperature and occupancy columns of a PDB flat file [35] with the per-atom charge (Q)
and radius (R). There are much more elegant ways to implement the PQR functionality
through more modern extensible file formats such as mmCIF [36] or PDBML [37]; however,
the simple PDB format is still one of the most widely used formats and, therefore, continued
use of the PQR format supports broad compatibility across biomolecular modeling tools and
workflows.
The PDB2PQR software is part of the APBS suite that was developed to assist with the
conversion of PDB files to PQR format [38, 39]. In particular, PDB2PQR automatically sets
up, executes, and optimizes the structure for Poisson-Boltzmann electrostatics calculations,
outputting a PQR file that can be used with APBS or other modeling software. Some of the
key steps in PDB2PQR are described below.

Repairing missing heavy atoms Within PDB2PQR, the PDB file is examined to see
if there are missing heavy (non-hydrogen) atoms. Missing heavy atoms can be rebuilt using
standard amino acid topologies in conjunction with existing atomic coordinates to deter-
mine new positions for the missing heavy atoms. A debump option performs very limited
minimization of sidechain χ angles to reduce steric clashes between rebuilt and existing
atoms [38].

Optimizing titration states Amino acid titration states are important determinants of
biomolecular (particularly enzymatic) function and can be used to assess functional activity
and identify active sites. The APBS-PDB2PQR system contains several methods for this
analysis.

• Empirical methods. PDB2PQR provides an empirical model (PROPKA [40]) that

uses a heuristic method to compute pKa perturbations due to desolvation, hydrogen
bonding, and charge-charge interactions. PROPKA is included with PDB2PQR. The
empirical PROPKA method has surprising accuracy for fast evaluation of protein pKa
values [41].

3
 • Implicit solvent methods. PDB2PQR also contains two methods for using implicit
solvent (Poisson-Boltzmann) models for predicting residue titration states. The first
method uses Metropolis Monte Carlo to calculate titration curves and pKa values
(PDB2PKA); however, sampling issues
can be a major problem with Monte Carlo
methods when searching over the O 2N titration states of N titratable residues. The
second method is a new polynomial-time algorithm for the optimization of discrete
states in macromolecular systems [42]. The paper by Purvine et al. [42] describes the
performance of the new graph cut method compared to existing approaches. While the
new deterministic method offers advantages for large systems, the traditional stochastic
approach is often sufficient for small to moderate systems. This method transforms
interaction energies between titratable groups into a graphical flow network. The
polynomial-time O (N 4 ) behavior makes it possible to rigorously evaluate titration
states for much larger proteins than Monte Carlo methods.

Adding missing hydrogens The majority of PDB entries do not include hydrogen posi-
tions. Given a titration state assignment, PDB2PQR uses Monte Carlo sampling to position
hydrogen atoms and optimize the global hydrogen-bonding network in the structure [43].
Newly added hydrogen atoms are checked for steric conflicts and optimized via the debump-
ing procedure discussed above.

Assigning charge and radius parameters Given the titration state, atomic charges
(for ρ) and radii (for  and Vi ) are assigned based on the chosen force field. PDB2PQR cur-
rently supports charge/radii force fields from AMBER99 [44], CHARMM22 [45], PARSE [46],
PEOE PB [47], Swanson et al. [48], and Tan et al. [49]. Many of these force fields only pro-
vide parameters for amino acid biomolecules. The PEOE approach [47] provides algorithms
to parameterize ligands. However, we welcome contributions for other biomolecular force
fields, particularly for lipids and amino acids.

3 Solving the Poisson-Boltzmann and related solvation

equations
The APBS software was designed from the ground up using modern design principles to
ensure its ability to interface with other computational packages and evolve as methods and
applications change over time. APBS input files contain several keywords that are generic
with respect to the type of calculation being performed; these are described in Appendix A.1.
The remainder of this section describes the specific solvers available for electrostatic calcu-
lations, also described in more detail on the APBS website [19].

3.1 Finite difference and finite element solvers

The original version of APBS was based on two key libraries from the Holst research group.
FEtk is a general-purpose multi-level adaptive finite element library [50, 51]. Adaptive finite
element methods can resolve extremely fine features of a complex system (like biomolecules)

4
while solving the associated equations over large problem domain. For example, FEtk has
been used to solve electrostatic and diffusion equations over six orders of magnitude in length
scale [52]. The finite difference PMG solver [51, 53] trades speed and efficiency for the high-
accuracy and high-detail solutions of the finite element FEtk library. However, many APBS
users need only a relatively coarse-grained solution of φ for their visualization or simulation
applications. Therefore, most APBS users employ the Holst group’s finite difference grid-
based PMG solver for biomolecular electrostatics calculations. Appendix A.2 describes some
of the common configuration options for finite difference and finite element calculations in
APBS.

3.2 Geometric flow

Several recent papers have described our work on a geometric flow formulation of Poisson-
based implicit solvent models [54–58]. The components of this geometric model are described
in previous publications [56, 57]. The geometric flow approach couples the polar and non-
polar components of the implicit solvent model with two primary advantages over existing
methods. First, this coupling eliminates the need for an ad hoc geometric definition for
the solute-solvent boundary. In particular, the solute-solvent interface is optimized as part
of the geometric flow calculation. Second, the optimization of this boundary ensures self-
consistent calculation of polar and nonpolar energetic contributions (using the same surface
definitions, etc.), thereby reducing confusion and the likelihood of user error. Additional
information about the geometric flow implementation in APBS is provided in Appendix A.3.
This equation is solved in APBS using a finite difference method.

3.3 Boundary element methods

Boundary element methods offer the ability to focus numerical effort on a much smaller
region of the problem domain: the interface between the molecule and the solvent. APBS
now includes a treecode accelerated boundary integral PB solver (TABI-PB) developed by
Geng and Krasny to solve a linearized version of the PB equation (Eq. 1) [59]. A discussion
of the relative merits between finite difference/element and boundary integral PB methods
are provided in Geng et al. and Li et al. [59, 60]. In this method, two coupled integral
equations defined on the solute-solvent boundary define a mathematical relationship between
the electrostatic surface potential and its normal derivative with a set of integral kernels
consisting of Coulomb and screened Coulomb potentials with their normal derivatives [61].
The boundary element method requires a surface triangulation, generated by a program
such as MSMS [62] or NanoShaper [63], on which to discretize the integral equations. A
Cartesian particle-cluster treecode is used to compute matrix-vector products and reduce
the computational cost of this dense system from O(N 2 ) to O(N log N ) for N points on
the discretized molecular surface [60, 61]. A comparison of electrostatic potential with
PDB ID 3app from the APBS multi-grid solution method and the new TABI-PB solver are
shown in Figure A1. Additional information about the boundary element method and its
implementation in APBS is provided in Appendix A.4.

5
Figure 2: Electrostatic potential visualization of protein with PDB ID 3app for (A) APBS
multigrid and (B) TABI-PB. VMD [64] was used to generate the figure in (A), and VTK
to generate the figure in (B). The potentials are on a [−4, 4] red-white-blue color map in
units of kJ/mol/e. Calculations were performed at 0.15 M ionic strength in monovalent salt,
298.15 Kelvin, protein dielectric 2, and solvent dielectric 78.

3.4 Analytical and semi-analytical methods

Numerical solution methods tend to be computationally intensive, which has led to the
adoption of analytical approaches for solvation calculations such as generalized Born [65]
and the approaches developed by Head-Gordon and implemented in APBS. In particular,
the Poisson-Boltzmann Analytical Method (PB-AM), was developed by Lotan and Head-
Gordon in 2006 [29]. PB-AM produces a fully analytical solution to the linearized PB
equation for multiple macromolecules, represented as coarse-grained low-dielectric spheres.
This spherical domain enables the use of a multipole expansion to represent charge-charge
interactions and higher order cavity polarization effects. The interactions can then be used
to compute physical properties such as interaction energies, forces, and torques. An example
of this approximation, and the resulting electrostatic potentials from PB-AM, are shown in
Figure 3.
The Poisson Boltzmann Semi-Analytical method (PB-SAM) is a modification of PB-
AM that incorporates the use of boundary integrals into its formalism to represent a com-
plex molecular domain as a collection of overlapping low dielectric spherical cavities [31].
PB-SAM produces a semi-analytical solution to the linearized PB equation for multiple
macromolecules in a screened environment. This semi-analytical method provides a better
representation of the molecular boundary when compared to PB-AM, while maintaining
computational efficiency. An example of this approximation, and the resulting electrostatic
potentials from PB-SAM, are shown in Figure 3.
Because it is fully analytical, PB-AM can be used for model validation as well as for repre-

6
 (A) (B) (C)

(D) (E) (F)

Figure 3: Comparison of APBS, PB-AM and PB-SAM results (A-C) and electrostatic po-
tential visualization for APBS PB-AM and PB-SAM (D-E). VMD [64] isosurface of barnase
molecule generated using (A) APBS boundary element method, (B) APBS PB-AM method,
(C) APBS PB-SAM method. (A-C) atoms colored according to their charge and isosurfaces
are drawn at 1.0 (blue) and −1.0 (red) kT/e electrostatic potential. (D) APBS PB-AM
potential on the coarse-grain surface of the barnase molecule, (E) APBS PB-SAM potential
in a 2D plane surrounding the barstar molecule, (F) APBS PB-SAM potential over range
[−1, 1] on the coarse-grain surface of the barnase molecule (blue region is the location of
positive electrostatic potential and barstar association). All calculations were performed at
0.0 M ionic strength, 300 Kelvin, pH 7, protein dielectric 2, and solvent dielectric 78. All
electrostatic potentials are given in units of kT /e.

7
senting systems that are relatively spherical in nature, such as globular proteins and colloids.
PB-SAM, on the other hand has a much more detailed representation of the molecular sur-
face and can therefore be used for many systems that other APBS (numerical) methods are
currently used for. Through APBS, both programs can be used to compute the electro-
static potential at any point in space, report energies, forces, and torques of a system of
macromolecules, and simulate a system using a BD scheme [66]. Additional details about
these methods and their use in APBS are presented in Appendix A.5. In addition to these
advantages, the strengths and weaknesses of PB-AM and PB-SAM over existing methods
are discussion in Yap et al. [31] and Lotan et al. [29].

4 Using APBS results

4.1 Visualization
One of the primary uses of the APBS tools is to generate electrostatic potentials for use
in biomolecular visualization software. These packages offer both the ability to visualize
APBS results as well as a graphical interface for setting up the calculation. Several of these
software packages are thick clients that run from users’ computers, including PyMOL [67,
68], VMD [64, 69], PMV [70, 71], and Chimera [72, 73]. We have also worked with the
developers of Jmol [74, 75] and 3Dmol.js [76, 77] to provide web-based setup and visualization
of APBS-PDB2PQR calculations and related workflows. APBS integration with Jmol has
been described previously [78]. 3Dmol.js is a molecular viewer that offers the performance of
a desktop application and convenience of a web-based viewer which broadens accessibility for
all users. As part of the integration with 3Dmol.js, we implemented additional enhancements,
including extending our output file formats and creating a customized user interface. Data
from the APBS output file is used to generate surfaces, apply color schemes, and display
different molecular styles such as cartoons and spheres. An example of the 3Dmol.js interface
is shown in Figure 4. Examples of 3Dmol.js visualization options are shown in Figure 5.

4.2 Other applications

Besides visualization and the processes described in Section 2, there are a number of other
applications where APBS can be used. For example, during the past four years, the APBS-
PDB2PQR software has been used in the post-simulation energetic analyses of molecular
dynamics trajectories [79], understanding protein-nanoparticle interactions [80, 81], under-
standing nucleic acid-ion interactions [82], biomolecular docking [83] and ligand binding [84],
developing new coarse-grained protein models [85], setting up membrane protein simula-
tions [86], etc. APBS also plays a key role in PIPSA for protein surface electrostatics
analysis [87] as well as BrownDye [88] and SDA [89] for simulation of protein-protein as-
sociation kinetics through Brownian dynamics. As discussed above with PB-SAM, another
application area for implicit solvent methods is in the evaluation of biomolecular kinetics
where implicit solvent models are generally used to provide solvation forces (or energies) for
performing (or analyzing) discrete molecular or continuum diffusion simulations with APBS
in both of these areas [79, 89–94].

8
Figure 4: 3Dmol.js interface displaying a rendering of fasciculin-2 (1FAS) protein with
translucent, solvent accessible surface using a stick model and red-green-blue color scheme.

5 The future of APBS

To help with modularity and to facilitate extensibility, APBS was based on an object- ori-
ented programming paradigm with strict ANSI-C compliance. This “Clean OO C” [95] has
enabled the long-term sustainability of APBS by combining an object-oriented design frame-
work with the portability and speed of C. This object-oriented design framework has made
it relatively straightforward to extend APBS functionality and incorporate new features.
The Clean OO C design has served APBS well for the past 17 years. This design still
forms the basis for many modules of APBS and PDB2PQR. Additionally, we have begun to
explore a framework for integrating components that exploits the common workflows used by
APBS-PDB2PQR users and maps naturally to cloud-based resources. Our vision for APBS
is to build the infrastructure that can enable our users to implement their own models and
methods so that they can run on a common system. Our goal is to have a well-designed,
well-tested and well-documented industry-grade framework that will integrate the APBS-
PDB2PQR capabilities and make straightforward the incorporation of new features and new
workflows.

Acknowledgments
The authors gratefully acknowledge NIH grant GM069702 for support of APBS and PDB2-
PQR. PNNL is operated by Battelle for the U.S. DOE under contract DE-AC05-76RL01830.
THG and DHB were supported by the Director, Office of Science, Office of Basic Energy
Sciences, of the U.S. Department of Energy under contract DE-AC02-05CH11231. LEF was
supported by the National Science Foundation Graduate Research Fellowship under Grant

9
Figure 5: Renderings of three different proteins: actinidin (2ACT) (top), fasciculin-2 (1FAS)
(center), and pepsin-penicillium (2APP) (bottom). To demonstrate the different visualiza-
tion options. From left to right: solvent-accessible surface, solvent-excluded surface, van der
Waals surface, and cartoon models are shown all using red-white-blue color scheme (exclud-
ing cartoon model), where red and blue correspond to negative and positive electrostatic
potentials, respectively.

10
No. DGE 110640. JAM and RK were supported by NIH grants GM31749 and GM103426.
LWW was supported by the Department of Defense through the National Defense Science &
Engineering Graduate Fellowship (NDSEG) Program. WG and RK were supported by NSF
grant DMS-1418966.

11
A Appendix
These appendices provide basic information about configuring APBS electrostatics calcula-
tions. Rather than duplicating the user manual on the APBS website [19], we have only
focused on input file keywords that directly relate to the setup and execution of solvation
calculations.

A.1 General keywords for implicit solvent calculations

This section contains information about the general keywords used to configure implicit
solvent calculations in the ELEC section of APBS input files. These keywords are used in all
of the solver types described in this paper:
• ion charge hchargei conc hconci radius hradiusi: This line can appear multiple
times to specify the ionic species in the calculation. This specifies the following terms
in Eq. 1: hchargei gives qi in units of electrons, hconci gives ci in units of M, and
hradiusi specifies the ionic radius (in Å) used to calculate Vi .

• lpbe|npbe: This keyword indicates whether to solve the full nonlinear version of Eq. 1
(npbe) or a linearized version (lpbe).

• mol hid i: Specify the ID of the molecule on which the calculations are to be performed.
This ID is determined by READ statements which specify the molecules to import.

• pdie and sdie: These keywords specify the dielectric coefficient values for the bio-
molecular interior (pdie) and bulk solvent (sdie). A typical value for sdie is 78.54;
values for pdie are much more variable and often range from 2 to 40.

• temp htempi: The temperature (in K) for the calculation. A typical value is 298 K.
Additionally, READ statements in APBS input files are used to load molecule information,
parameter sets, and finite element meshes. More detailed information about these and other
commands can be found on the APBS website [19].

A.2 Finite element and finite difference calculations in APBS

The finite difference and finite element methods used by APBS have been described exten-
sively in previous publications [18, 26–28, 53]; this section focuses on the configuration and
use of these methods in APBS.

A.2.1 Finite difference calculation configuration

APBS users have several ways to invoke the PMG finite difference solver [53] capabilities of
APBS through keywords in the APBS input file ELEC block. The different types of finite
difference calculations include:
• mg-manual: This specifies a manually configured multigrid finite difference calculation
in APBS.

12
 • mg-auto: This specifies an automatically configured multigrid finite difference calcula-
tion in APBS, using focusing [96] to increase the resolution of the calculation in areas
of interest.

• mg-para: This specifies a parallel version of the multigrid finite difference calculat-
ing, using parallel focusing to increase the resolution of the calculation in areas of
interest [18].

These different types of calculations have several keywords described in detail on the APBS
website [19]. Some of the most important settings are described below.

Boundary conditions Boundary conditions must be imposed on the exterior of the fi-
nite difference calculation domain. In general, biomolecular electrostatics calculations use
Dirichlet boundary conditions, setting the value of the potential to an asymptotically correct
approximation of the true solution. There are several forms of these boundary conditions
available in APBS with approximately equal accuracy given a sufficiently large calculation
domain (see below). The only boundary condition which is not recommended for typical
calculations is the zero-potential Dirichlet condition.

Grid dimensions, center, and spacing Finite difference calculations in APBS are per-
formed in rectangular domains. The key aspects of this domain include its length Li and
number of grid points ni in each direction i. These parameters are related to the spacing hi
of the finite difference grid by hi = Li /(ni − 1). Grid spacings below 0.5 Å are recommended
for most calculations. The number of grid points is specified by the dime keyword. For
mg-manual calculations, the grid spacing can be specified by either the grid or the glen
keywords, which specify the grid spacing or length, respectively. For mg-auto or mg-para
calculations, the grid spacing is determined by the cglen and fglen keywords, which specify
the lengths of the coarse and fine grids for the focusing calculations. Grid lengths should
extend sufficient distance away from the biomolecule so that the chosen boundary condition
is accurate. In general, setting the length of the coarsest grid to approximately 1.5 times
the size of the biomolecule gives reasonable results. However, as a best practice, it is impor-
tant to ensure that the calculated quantities of interest do not change significantly with grid
spacing or grid length. The center of the finite difference grid can be specified by the gcent
command for mg-manual calculations or by the cgcent and fgcent keywords for the coarse
and fine grids (respectively) in mg-auto or mg-para focusing calculations. These keywords
can be used to specify absolute grid centers (in Cartesian coordinates) or relative centers
based on molecule location. Because of errors associated with charge discretization, it is
generally a good idea to keep the positions of molecules on a grid fixed for all calculations.
For example, a binding calculation for rigid molecules should keep all molecules in the same
positions on the grid.

Charge discretization As mentioned above, atomic charge distributions are often mod-
eled as a collection of Dirac delta functions or other basis functions with extremely small
spatial support. However, finite difference calculations performed on grids with finite spacing,
requiring charges to be mapped across several grid points. This mapping creates significant

13
dependence of the electrostatic potential on the grid spacing which is why it is always im-
portant to use the same grid setup for all parts of an electrostatic calculation. The chgm
keyword controls the interpolation scheme used for charge distributions and includes the
following types of discretization schemes:
• spl0: Traditional trilinear interpolation (linear splines). The charge is mapped onto
the nearest-neighbor grid points. Resulting potentials are very sensitive to grid spacing,
length, and position.

• spl2: Cubic B-spline discretization as described by Im et al. [97]. The charge is

mapped onto the nearest- and next-nearest-neighbor grid points. Resulting potentials
are somewhat less sensitive (than spl0) to grid spacing, length, and position; however,
this discretization can often require smaller grid spacings for accurate representation
of charge positions.

• spl4: Quintic B-spline discretization as described by Schnieders et al. [98]. Similar to

spl2, except the charge/multipole is additionally mapped to include next-next-nearest
neighbors (125 grid points receive charge density). This discretization results in less
sensitivity to grid spacing and position but nearly always requires smaller grid spacings
for accurate representation of charge positions.

Surface definition APBS provides several difference surface definitions through the srfm
keyword:
• mol: The dielectric coefficient is defined based on a molecular surface definition. The
problem domain is divided into two spaces. The “free volume” space is defined by the
union of solvent-sized spheres (size determined by the srad keyword) which do not
overlap with biomolecular atoms. This free volume is assigned bulk solvent dielectric
values. The complement of this space is assigned biomolecular dielectric values. With a
non-zero solvent radius (srad), this choice of coefficient corresponds to the traditional
definition used for PB calculations. When the solvent radius is set to zero, this corre-
sponds to a van der Waals surface definition. The ion-accessibility coefficient is defined
by an “inflated” van der Waals model. Specifically, the radius of each biomolecular
atom is increased by the radius of the ion species (as specified with the ion keyword).
The problem domain is then divided into two spaces. The space inside the union of
these inflated atomic spheres is assigned an ion-accessibility value of 0; the complement
space is assigned bulk ion accessibility values.

• smol: The dielectric and ion-accessibility coefficients are defined as for mol (see above).
However, they are then “smoothed” by a 9-point harmonic averaging to somewhat
reduce sensitivity to the grid setup [99].

• spl2: The dielectric and ion-accessibility coefficients are defined by a cubic-spline sur-
face [97]. The width of the dielectric interface is controlled by the swin parameter.
These spline-based surface definitions are very stable with respect to grid parameters
and therefore ideal for calculating forces. However, they require substantial reparam-
eterization of the force field [100].

14
 • spl4: The dielectric and ion-accessibility coefficients are defined by a 7th -order poly-
nomial. This surface definition has characteristics similar to spl2, but provides higher
order continuity necessary for stable force calculations with atomic multipole force
fields (up to quadrupole) [98].

A.2.2 Finite elements calculation configuration

Users invoke the FEtk finite element solver [50] in APBS by including the fe-manual keyword
in the ELEC section of the input file. Many aspects of the finite element configuration
closely follow the finite difference options described above. This section only describes the
configuration options which are unique to finite element calculations in APBS.
Finite element calculations begin with an initial mesh. This mesh can be imported from
an external file via the usemesh keyword or generated by APBS. APBS generates the initial
mesh from a very coarse 8-tetrahedron mesh of size domainLength which is then refined
uniformly or selectively at the molecular surface and charge locations, based on the value of
the akeyPRE keyword. This initial refinement occurs until the mesh has targetNum vertices
or has reached targetRes edge length (in Å).
As described previously [26, 27], APBS uses FEtk in a solve-estimate-refine iteration
which involves the following steps:

1. Solve the problem with the current finite element mesh.

2. Estimate the error in the solution as a function of position on the mesh. The method
of error estimation is determined by the ekey keyword which can have the values:

• simp: Per-simplex error threshold; simplices with error above this limit are flagged
for refinement.
• global: Global (whole domain) error limit; flag enough simplices for refinement
to reduce the global error below this limit.
• frac: The specified fraction of the simplices with the highest amount of error are
flagged for refinement.

3. Adaptively refine the mesh to reduce the error using the error metric described by
ekey.

This iteration is repeated until a target error level etol is reached or a maximum number
of solve-estimate refine iterations (maxsolve) or vertices (maxvert) is reached.

A.3 Geometric flow calculations in APBS

This section contains additional information about the geometric flow equation implemen-
tation in APBS introduced in Section 3.2. The geometric flow methods used by APBS have
been described extensively in previous publications [54–57, 101]; this section focuses on the
configuration and use of these methods in APBS.

15
A.3.1 Geometric flow calculation configuration
Users invoke the geometric flow solver in APBS by including the geoflow-auto keyword
in the ELEC section of the input file. Because the geometric flow solver is based on finite
difference solvers, many of the keywords for this section are similar to those described in
Section A.2.1. Three additional parameters are needed for geometric flow calculations to
specify how nonpolar solvation is linked to the polar implicit solvent models:

• gamma htensioni: Specify the surface tension of the solvent in units of kJ mol−1 Å−2 .
Based on Daily et al. [57], a recommended value for small molecules is 0.431 kJ mol−1
Å−2 .

• press hpressurei: Specify the internal pressure of the solvent in units of kJ mol−1 Å−3 .
Based on Daily et al. [57], a recommended value for small molecules is 0.104 kJ mol−1
Å−3 .

• bconc hconcentrationi: Specify the bulk concentration of solvent in Å−3 . The bulk
density of water, 0.0334 Å−3 , is recommended.

• vdwdisp hbool i: Indicate whether van der Waals interactions should be included in
the geometric flow calculation through hbool i (1 = include, 0 = exclude). If these
interactions are included, then a force field with van der Waals terms must be included
through a READ statement in the APBS input file.

A.4 Boundary element method implementation

This appendix provides additional information about the boundary element method intro-
duced in Section 3.3.

A.4.1 Boundary element method background

This section provides additional background on the TABI-PB boundary element solver [59],
introduced in Section 3.3. As described earlier, this method involves solving two coupled
integral equations defined on the solute-solvent boundary which define a mathematical re-
lationship between the electrostatic surface potential and its normal derivative with a set
of integral kernels consisting of Coulomb and screened Coulomb potentials with their nor-
mal derivatives. The boundary element method requires a surface triangulation, generated
by a program such as MSMS [62] or NanoShaper [63], on which to discretize the integral
equations. Figure A1 shows different types of surface discretizations as well as example
electrostatic potential output.
The coupled second kind integral equations employed by TABI-PB for calculating the
surface potential φ and its normal derivative [61] are:
Z  
1 ∂φ (y)
(1 + ε) φ (x) = K1 (x, y) + K2 (x, y) φ (y) dSy + S1 (x), x ∈ Γ,
2 Γ ∂ν
  Z   (2)
1 1 ∂φ (x) ∂φ (y)
1+ = K3 (x, y) + K4 (x, y) φ (y) dSy + S2 (x), x ∈ Γ
2 ε ∂ν Γ ∂ν

16
 (A) (B) (C)

(D) (E) (F)

Figure A1: APBS TABI-PB electrostatic potential results for PDB ID 1a63. Surfaces were
generated with (A) 20228 triangles via MSMS [62], (B) 20744 triangles via NanoShaper
SES, and (C) 21084 triangles via NanoShaper Skin [63]. These discretizations resulted in
surface potentials (with units kT /e) of (D) MSMS in range [−8.7, 8.6], (E) NanoShaper SES
in range [−13.4, 7.5], and (F) NanoShaper Skin in range [−33.8, 8.0]. All calculations were
performed at 0.15 M ionic strength in 1:1 salt, with protein dielectric 1, solvent dielectric
80, and temperature 300K. Red and blue surface colors correspond to negative and positive
electrostatic surface potentials, respectively.

17
where ε = εm /εs , the ratio of the dielectric constant in the solute region and the dielectric
constant in the solvent region. The integral kernels K1 , K2 , K3 , K4 are defined in Eq. 3.

K1 (x, y) = G0 (x, y) − Gκ (x, y) ,

∂Gκ (x, y) ∂G0 (x, y)
K2 (x, y) = ε − ,
∂νy ∂νy
∂G0 (x, y) 1 ∂Gκ (x, y) (3)
K3 (x, y) = ε − ,
∂νx ε ∂νx
∂ 2 Gκ (x, y) 1 ∂ 2 G0 (x, y)
K4 (x, y) = ε − ,
∂νx ∂νy ε ∂νx ∂νx

where G0 and Gκ are the Coulomb and screened Coulomb potentials defined as:
1
G0 (x, y) = ,
4π |x − y|
(4)
e−κ|x−y|
Gκ (x, y) = .
4π |x − y|

The normal derivatives of the potential kernels G are:

3
∂G (x, y) X
= νn (y) ∂yn G (x, y) ,
∂νy n=1
3
∂G (x, y) X
=− νn (y) ∂xn G (x, y) , (5)
∂νx n=1
3 X 3
∂ 2 G (x, y) X
=− νm (x) νn (y) ∂xn ∂yn G (x, y)
∂νx ∂νy m=1 n=1

for the three spatial components n of the normal direction. Additionally, the source terms
S1 and S2 in Eq. 2 are:
Nc
1 X
S1 (x) = qk G0 (x, yk ) ,
εm k=1
Nc
(6)
1 X ∂G0 (x, yk )
S2 (x) = qk ,
εm k=1 ∂νx

where Nc is the number of atoms in the solute molecule, and qk is the charge of the kth
atom. Note that S1 is a linear superposition of the point charge electrostatic potentials, and
S2 is a linear superposition of the normal derivatives of the potentials.
Given a surface triangularization with N elements – where xi and Ai are the centroid

18
and area, respectively, of the ith triangle – the integral equations are discretized as:
N  
1 X ∂φ (xj )
(1 + ε) φ (xi ) = K1 (xi , xj ) + K2 (xi , xj ) φ (xj ) Aj + S1 (xi ),
2 j=1
∂ν
j6=i
  N   (7)
1 1 ∂φ (xi ) X ∂φ (xj )
1+ = K3 (xi , xj ) + K4 (xi , xj ) φ (xj ) AJ + S2 (xI ).
2 ε ∂ν j=1
∂ν
j6=i

The omission of the j = i term in the summation avoids the singularity of the kernels at
that point. Note that the right hand sides of these equations consist of sums of products of
kernels and the surface potential or its normal derivative. These are the analogues to the
N -body potential in the treecode, with the surface potential or its normal derivatives playing
the role of the charges. In the discretized form, the total electrostatic energy of solvation is
given by Eq. 8:
Nc N  
1X X ∂φ (xj )
Esol = qk K1 (xk , xj ) + K2 (xk , xj ) φ (xj ) Aj (8)
2 k=1 j=1 ∂ν

where qk is the charge on the kth atom of the solute molecule, and xk is its position.
The matrix-vector products involve evaluation of the integral kernels over the surface
elements. These evaluations effectively take the form of an N -body potential: a sum over a
set of N positions of products between a kernel and a “charge” at each position. In this case,
the locations of the N particles are the centroids of the surface triangularization elements.
A Cartesian particle-cluster treecode is used to compute matrix-vector products and reduce
the computational cost of this dense system from O(N 2 ) to O(N log N ) for N points on the
discretized molecular surface [60]. In particular, to rapidly evaluate the N -body potential
at the N particle locations, the treecode subdivides the particles into a tree-like hierarchical
structure of clusters. At each location, the potential contribution from nearby particles
is computed by direct sum, while for well-separated particle cluster interactions, a Taylor
approximation about the center of the cluster is used to evaluate the contribution. The
Taylor coefficients are calculated through recurrence relations. The resulting linear system
is then solved with GMRES iteration [102].
Because the integral equations are defined on the molecular boundary, the singular
charges are handled analytically and do not introduce the same issues present in grid-based
schemes. The integral equations also rigorously enforce the interface conditions on the sur-
face, and the boundary condition at infinity is exactly satisfied. Thus, the boundary integral
formulation can potentially be superior to other methods for investigating electrostatic po-
tential on the boundary.

A.4.2 Boundary element calculation configuration

APBS users can invoke TABI-PB with the bem-manual flag in the ELEC section of the input
file. Major options include:

19
 • tree order horder i: An integer indicating the order of the Taylor expansion for de-
termining treecode coefficients. Higher values of horder i will result in a more accurate
– but more expensive – calculations. A typical choice for this parameter is 3.
• tree n0 hnumber i: The maximum number of particles allowable in a leaf of the
treecode (clusters in the last level of the tree). A typical choice for this parameter
is 500.
• mac hcriterioni: Multipole acceptance criterion specifies the distance ratio at which
the Taylor expansion is used. In general, a higher value of hcriterioni will result in a
more accurate but more expensive computation; while a lower value causes more direct
summations and forces the particle-cluster interaction to descend to a finer cluster level.
A typical choice for this parameter is 0.8.
• mesh hflagi: The software used to mesh the molecular surface; 0 = MSMS, 1 = Nano-
Shaper’s SES implementation, and 2 = NanoShaper’s Skin implementation. See Fig-
ure A1 for an example of surface meshes.
• outdata hflagi: Type of output data file generated; 0 = APBS OpenDX format [103]
and 1 = ParaView format [104].
Additional information about parameter settings is provided via the APBS website [19].
TABI-PB produces output including the potential and normal derivative of potential for
every element and vertex of the triangularization, as well as the electrostatic solvation energy.
Examples of electrostatic surface potential on the protein 1a63 are shown in Figure 2 by using
MSMS and NanoShaper.

A.5 Analytical and semi-analytical method implementations

This appendix provides additional information about the analytical and semi-analytic meth-
ods [29, 30] introduced in Section 3.4.

A.5.1 Analytical method (PB-AM) background

The solution to the PB-AM model is represented as a system of linear equations:
A = Γ · (∆ · T · A + E), (9)
where A represents a vector of the effective multipole expansion of the charge distributions of
each molecule, E is a vector of the fixed charge distribution of all molecules, Γ is a dielectric
boundary-crossing operator, ∆ is a cavity polarization operator, and T is an operator that
transforms the multipole expansion from the global (lab) coordinates to a local coordinate
frame. The unknown A determined using the Gauss-Seidel iterative method and can then be
used to compute physical properties such as interaction energies, forces, and torques. The
interaction energy for molecule i, (Ω(i) ) is given in Eq. 10.
* N +
1 X
Ω(i) = T · A(j) , A(i) (10)
s j6=i

20
where s is the dielectric constant of the solvent and hM, N i denotes the inner product.
When energy is computed, forces follow as:
1
F(i) = ∇i Ω(i) = [h∇i T · A(i) , A(i) i + hT · A(i) , ∇i A(i) i] (11)
s
By definition, the torque on a charge in the molecule is the cross product of its position
relative to the center of mass of the molecule with the force it experiences. The total torque
on the molecule is a linear combination of the torque on all charges of the molecule, as
illustrated in Eq. 12.
1 x (i) y (i) z (i)  
τ (i) = × ∇i L(i)

H , H , H (12)
s
(i) P i (i) (i) (i) (i) n (i) (i)
where α Hn,m = M j=1 αj γn qj (ρj ) Yn,m (ϑj , ϕj ) , α = x, y, z, is a coefficient vector for
(i)
each of the charges in the molecule, Mhi is the numberi of charges in molecule i, qJ is the
(i) (i) (i) (i)
magnitude of the j th charge, and pj = ρj , ϑj , ϕj is its position in spherical coordinates.
For more details on the PB-AM derivation, see Lotan and Head-Gordon [29].

A.5.2 Semi-analytical method (PM-SAM) background

The derivation details of PB-SAM have been reported previously [31, 32], with the main
points being summarized in this section. The electrostatic potential (φr ) of the system at
any point r is governed by the linearized form of the PB equation:

− ∇ · ∇φ + κ2 φ = ρ, (13)

where κ is the inverse Debye length. Eq. 13 is a linearization of Eq. 1 for βqi φ  1. For the
case of spherical cavities, we can solve Eq. 13 by dividing the system into inner sphere and
outer sphere regions, and enforcing a set of boundary conditions that stipulate the continuity
of the electrostatic potential and the electrostatic field at the surface of each sphere. The
electrostatic potential outside molecule (I) is described by:
N 0
mol
e−κ|r−r | (I) 0 0
X  Z 
(i)
φout (r) = 4π h (r )dr (14)
I=1 dΩ(I) |r − r0 |

where h(r) is an effective surface charge that can be transformed into the unknown multipole
expansion H (I,k) with inside molecule I and sphere k. In a similar manner, the interior
potential is given as
(I)
NC (I) Z
(i)
X 1 qα 1 1
φin (r) = · + f (I) (r0 )dr0 (15)
α=1
(I)
|r − rα | in 4π dΩ(I) |r − r0 |

(I)
where N
h C is the number
i of charges in molecule I, qα is the magnitude of the α-th charge,
(I) (I) (I) (I)
rα = ρα , θα , φα is its position in spherical coordinates, and f (r) is a reactive surface

21
charge that can be transformed into the unknown multipole expansion F (I,k) . The reactive
multipole and the effective multipole, H (I,k) , are given as:
Z  (I,k) n+1
1 (I,k) 0 a (I,k)
(I,k)
Fn,m ≡ f (r ) Yn,m (θ0 , φ0 )dr0 (16)
4π dΩ(I,k) r0
Z  0 n
1 (I,k) 0 r (I,k)
(I,k)
Hn,m ≡ h (r ) în (κr0 )Yn,m (θ0 , φ0 )dr0 (17)
4π dΩ(I,k) a(I,k)
where Yn,m is the spherical harmonics, Yn,m is the complete conjugate, and a(I,k) is the radius
of sphere k of molecule I. These multipole expansions can be iteratively solved using:
(I,k) (I,k)
Fn,m = hIE,n,m , W F (I,k) i (18)
(I,k) (I,k)
Hn,m = hIE,n,m , W H (I,k) i (19)

where W F (I,k) and W H (I,k) are scaled multipoles computed from fixed charges and polariza-
(I,k)
tion charges from other spheres. IE,n,m is a matrix of the surface integrals over the exposed
surface: Z Z
(I,k) 1 (I,k) 0 (I,k)
IE,n,m ≡ Y (θ , φ0 )Yn,m (θ0 , φ0 ) sin θ0 dθ0 dφ0 (20)
4π φE θE l,s
Using the above formalism, physical properties of the system, such as interaction energy,
forces and torques can also be computed. The interaction energy of each molecule, (Ω(i) ), is
the product of the molecule’s total charge distribution (from fixed and polarization charges)
with the potential due to external sources. This is computed as the inner product between the
molecule’s multipole expansion, (H (I,k) ), and the multipole expansions of the other molecules
in the system, (LHN (I,k) ) as follows:
(I)
Nk
1 X
Ω(i) = hLHN (I,k) , H (I,k) i (21)
s k

which allows us to define the force which is computed as the gradient of the interaction
energy with respect to the position of the center of molecule I :
(I) (I)
Nk Nk
1 X 1 X
F (I) = −∇Ω(I) =− fI,k = − (h∇LHN (I,k) , H (I,k) i + hLHN (I,k) , ∇H (I,k) i) (22)
s k s k

As in the analytical PB-AM method, the torque on a charge in the molecule is the cross
product of its position relative to the center of mass of the molecule with the force it expe-
riences. For a charge at position P about the center of mass c(I) for molecule I, the torque
(I,k)
is given by the cross product of its position rP with respect to the center of mass and
(I,k)
the force on that charge fP . We can re-express rP as the sum of vectors from the center
of molecule I to the center of sphere k (c(I,k) ) and from the center of sphere k to point P
(I,k)
(rP ). The total torque on molecule I is then given by Eq. 23.
(I) (I)
Nk Nk
X XX (I,k)
τ (I) = c(I,k) × fI,k + rP × fP (23)
k k P ∈k

22
where fI,k is given in Eq. 22 and
(I)
Nk
1 X (I,k) (I,k)
fP = − (h∇I LHN (I,k) , HP i + hLHN (I,k) , ∇I HP i) (24)
s k

where
(I,k) (I,k)
HP,n,m = h(θp , φp )Yn,m (θp , φp ) (25)
(I,k) (I,k)
∇j HP,α,n,m = [∇j h(θp , φp )]α Yn,m (θp , φp ) (26)

where α = x, y, z. For the derivation of the PB-SAM solver please see the previous publica-
tions [31, 32] .

A.5.3 PB-AM and PB-SAM configuration in APBS

PB-AM and PB-SAM have been fully integrated into APBS, and is invoked using the keyword
pbam-auto or pbsam-auto in the ELEC section of an APBS input file. Major options include:

• runname hnamei: Desired name to be used for outputs of each run.

• pbc hlengthi: Size of the periodic simulation/calculation domain.

• runtype dynamics: Perform a Brownian Dynamics simulation.

• ntraj hnumber i: Number of Brownian Dynamics simulations to run.

• term htypei hvaluei hmol i: Allows the user to indicate conditions for the termination
of each BD trajectory. The following values of htypei are allowed:

– time htimei: A limit on the total simulation time.

– x or y or y or z or r and h>=i or h<=i: Represents the approach of two molecules
to a certain distance r or certain region of space given by x or y or y. The operators
>= and <= represent the corresponding inequalities.

The parameter hmol i is the molecular index that this condition applies. hmol i should
be 0 for time and for a termination condition of x, y or z, the molecule index that this
termination condition applies to.

• xyz hidx i hfpathi: Molecule index hidx i and file path hfpathi for the molecule starting
configurations. A starting configuration is needed for each molecule and each trajec-
tory. Therefore, jf there are m molecules and ntraj hni trajectories, then the input
file must contain m × n xyz entries.

• tolsp hval i: Modify the coarseness of the molecular description. hval i is the distance
(in Å) beyond the solvent-excluded surface that the coarse-grained representation ex-
tends. Increasing values of hval i leads to fewer coarse-grained spheres, faster calculation
times, but less accurate solutions. Typical values for hval i are between 1 and 5 Å.

23
The commands (keywords) not included in this list are used to specify system conditions,
such as temperature and salt concentration. These parameters are similar to those found
in the ELEC section of a usual APBS run and are documented on the Contributions portion
of the APBS website [19]. Additional information about parameter settings is provided via
the APBS website [19]. Examples of the electrostatic potentials produced from PB-AM and
PB-SAM are shown in Figure 3.

References
[1] P Ren et al. “Biomolecular electrostatics and solvation: a computational perspec-
tive”. In: Quarterly Reviews of Biophysics 45.4 (2012), pp. 427–91. doi: 10.1017/
S003358351200011X.
[2] M. E. Davis and J A McCammon. “Electrostatics in biomolecular structure and dy-
namics”. In: Chemical Reviews 90.3 (1990), pp. 509–521.
[3] M F Perutz. “Electrostatic effects in proteins”. In: Science 201.4362 (1978), pp. 1187–
91.
[4] K A Sharp and B Honig. “Electrostatic interactions in macromolecules: theory and
applications”. In: Annual Review of Biophysics and Biophysical Chemistry 19 (1990),
pp. 301–32. doi: 10.1146/annurev.bb.19.060190.001505.
[5] B Roux and T Simonson. “Implicit solvent models”. In: Biophysical Chemistry 78.1-2
(1999), pp. 1–20.
[6] A Warshel et al. “Modeling electrostatic effects in proteins”. In: Biochimica et Bio-
physica Acta 1764.11 (2006), pp. 1647–76. doi: 10.1016/j.bbapap.2006.08.007.
[7] M Fixman. “The Poisson-Boltzmann equation and its application to polyelectrolytes”.
In: Journal of Chemical Physics 70 (1979), pp. 4995–5005.
[8] P Grochowski and J Trylska. “Continuum molecular electrostatics, salt effects, and
counterion binding–a review of the Poisson-Boltzmann theory and its modifications”.
In: Biopolymers 89.2 (2008), pp. 93–113. doi: 10.1002/bip.20877.
[9] G Lamm et al. “The Poisson-Boltzmann Equation”. In: Reviews in Computational
Chemistry 19 (2003), pp. 147–365.
[10] B. Lee and F. M Richards. “The interpretation of protein structures: Estimation of
static accessibility”. In: Journal of Molecular Biology 55.3 (1971), pp. 379–400. doi:
10.1016/0022-2836(71)90324-X.
[11] R. R. Netz and H. Orland. “Beyond Poisson-Boltzmann: Fluctuation effects and corre-
lation functions”. In: European Physical Journal E: Soft Matter 1.2-3 (2000), pp. 203–
214.
[12] B R Brooks et al. “CHARMM: the biomolecular simulation program”. In: Journal of
Computational Chemistry 30.10 (2009), pp. 1545–614. doi: 10.1002/jcc.21287.
[13] D A Case et al. “The AMBER biomolecular simulation programs”. In: Journal of
Computational Chemistry 26.16 (2005), pp. 1668–88. doi: 10.1002/jcc.20290.

24
[14] L Li et al. “DelPhi: a comprehensive suite for DelPhi software and associated re-
sources”. In: BMC Biophysics 5 (2012), p. 9. doi: 10.1186/2046-1682-5-9.
[15] AD Bochevarov et al. “Jaguar: A high-performance quantum chemistry software pro-
gram with strengths in life and materials sciences”. In: International Journal of Quan-
tum Chemistry 113.18 (2013), pp. 2110–2142.
[16] JA Grant, BT Pickup, and A Nicholls. “A smooth permittivity function for Poisson-
Boltzmann solvation methods”. In: Journal of Computational Chemistry 22 (2001),
pp. 608–640.
[17] Y. C. Zhou, M. Feig, and G. W. Wei. “Highly accurate biomolecular electrostatics
in continuum dielectric environments”. In: Journal of Computational Chemistry 29.1
(2008), pp. 87–87. doi: 10.1002/jcc.20769.
[18] N A Baker et al. “Electrostatics of nanosystems: application to microtubules and the
ribosome”. In: Proceedings of the National academy of Sciences of the United States
of America 98.18 (2001), pp. 10037–41. doi: 10.1073/pnas.181342398.
[19] “APBS website”. http : / / www . poissonboltzmann . org/. url: http : / / www .
poissonboltzmann.org/.
[20] Sriram Krishnan et al. “Design and evaluation of Opal2: A toolkit for scientific soft-
ware as a service”. In: Services-I, 2009 World Conference on. IEEE. 2009, pp. 709–
716. url: http://nbcr-222.ucsd.edu/opal2/dashboard.
[21] J C Phillips et al. “Scalable molecular dynamics with NAMD”. In: Journal of Com-
putational Chemistry 26.16 (2005), pp. 1781–802. doi: 10.1002/jcc.20289.
[22] David Baker. “Rosetta website”. https://www.rosettacommons.org/. url: https:
//www.rosettacommons.org/.
[23] Jay Ponder. “TINKER website”. https : / / dasher . wustl . edu / tinker/. url:
https://dasher.wustl.edu/tinker/.
[24] R Konecny, N A Baker, and J A McCammon. “iAPBS: a programming interface to
Adaptive Poisson-Boltzmann Solver (APBS)”. In: Computational Science and Dis-
covery 5.1 (2012). doi: 10.1088/1749-4699/5/1/015005.
[25] “iAPBS website”. https://mccammon.ucsd.edu/iapbs/. url: https://mccammon.
ucsd.edu/iapbs/.
[26] M Holst, N Baker, and F Wang. “Adaptive multilevel finite element solution of the
Poisson-Boltzmann equation I. Algorithms and examples”. In: Journal of Computa-
tional Chemistry 21.15 (2000), pp. 1319–1342.
[27] N Baker, M Holst, and F Wang. “Adaptive multilevel finite element solution of the
Poisson-Boltzmann equation II. Refinement at solvent-accessible surfaces in biomolec-
ular systems”. In: Journal of Computational Chemistry 21.15 (2000), pp. 1343–1352.
[28] N A Baker et al. “The adaptive multilevel finite element solution of the Poisson-
Boltzmann equation on massively parallel computers”. In: IBM Journal of Research
and Development 45.3-4 (2001), pp. 427–438.

25
[29] Itay Lotan and Teresa Head-Gordon. “An Analytical Electrostatic Model for Salt
Screened Interactions between Multiple Proteins”. In: Journal of Chemical Theory
and Computation 2.3 (2006). PMID: 26626662, pp. 541–555. doi: 10.1021/ct050263p.
eprint: http://dx.doi.org/10.1021/ct050263p.
[30] Lisa E. Felberg et al. “PB-AM: An open-source, fully analytical linear Poisson-
Boltzmann solver”. In: Journal of Computational Chemistry 38.15 (2017), pp. 1275–
1282. issn: 1096-987X. doi: 10.1002/jcc.24528.
[31] E-H Yap and T Head-Gordon. “New and Efficient Poisson-Boltzmann Solver for In-
teraction of Multiple Proteins”. In: Journal of Chemical Theory and Computation 6.7
(2010), pp. 2214–2224. doi: 10.1021/ct100145f.
[32] E. H. Yap and T. Head-Gordon. “Calculating the Bimolecular Rate of Protein-Protein
Association with Interacting Crowders”. In: Journal of Chemical Theory and Com-
putation 9.5 (2013), pp. 2481–9. issn: 1549-9618 (Print) 1549-9618 (Linking). doi:
10.1021/ct400048q. url: http://www.ncbi.nlm.nih.gov/pubmed/26583736.
[33] Donald Bashford. “An Object-Oriented Programming Suite for Electrostatic Effects
in Biological Molecules”. In: Scientific Computing in Object-Oriented Parallel En-
vironments. Ed. by Yutaka Ishikawa et al. Vol. 1343. Lecture Notes in Computer
Science. Berlin: Springer, 1997, pp. 233–240.
[34] Garrett M. Morris et al. “AutoDock4 and AutoDockTools4: Automated docking with
selective receptor flexibility”. In: Journal of Computational Chemistry 30 (2009),
pp. 2785–2791. url: http://www3.interscience.wiley.com/cgi-bin/abstract/
122365050/ABSTRACT.
[35] “Protein Data Bank Contents Guide: Atomic Coordinate Entry Format Description.
Version 3.3”. http : / / www . wwpdb . org / documentation / file - format - content /
format33/v3.3.html. 2012. url: http://www.wwpdb.org/documentation/file-
format-content/format33/v3.3.html.
[36] “PDBx/mmCIF Dictionary Resources”. http : / /mmcif . wwpdb . org/. url: http :
//mmcif.wwpdb.org/.
[37] J. Westbrook et al. “PDBML: the representation of archival macromolecular struc-
ture data in XML”. In: Bioinformatics 21.7 (2005), pp. 988–992. doi: 10 . 1093 /
bioinformatics/bti082.
[38] Todd J Dolinsky et al. “PDB2PQR: an automated pipeline for the setup of Poisson–
Boltzmann electrostatics calculations”. In: Nucleic Acids Research 32.suppl 2 (2004),
W665–W667.
[39] T J Dolinsky et al. “PDB2PQR: expanding and upgrading automated preparation
of biomolecular structures for molecular simulations”. In: Nucleic Acids Research 35
(2007), W522–5. doi: 10.1093/nar/gkm276.
[40] Chresten R. Sondergaard et al. “Improved Treatment of Ligands and Coupling Effects
in Empirical Calculation and Rationalization of pKa Values”. In: Journal of Chemical
Theory and Computation 7.7 (2011), pp. 2284–2295.

26
[41] H. Li, A. D. Robertson, and J. H. Jensen. “Very fast empirical prediction and ratio-
nalization of protein pKa values”. In: Proteins 61 (2005), p. 704721.
[42] Emilie Purvine et al. “Energy Minimization of Discrete Protein Titration State Mod-
els Using Graph Theory”. In: Journal of Physical Chemistry 120 (2016), pp. 8354–
8360. doi: 10.1021/acs.jpcb.6b02059.
[43] J E Nielsen and G Vriend. “Optimizing the hydrogen-bond network in Poisson-
Boltzmann equation-based pKa calculations”. In: Proteins 43.4 (2001), pp. 403–12.
[44] Junmei Wang, Piotr Cieplak, and Peter A. Kollman. “How well does a restrained
electrostatic potential (RESP) model perform in calculating conformational energies
of organic and biological molecules?” In: Journal of Computational Chemistry 21.12
(2000), pp. 1049–1074. doi: 10 . 1002 / 1096 - 987X(200009 ) 21 : 12<1049 :: AID -
JCC3>3.0.CO;2-F.
[45] A.D. MacKerell Jr., M. Feig, and C.L. Brooks III. “Extending the treatment of back-
bone energetics in protein force fields: limitations of gas-phase quantum mechanics
in reproducing protein conformational distributions in molecular dynamics simula-
tions”. In: Journal of Computational Chemistry 25.11 (2004), pp. 1400–1415. doi:
10.1002/jcc.20065.
[46] Doree Sitkoff, Kim A. Sharp, and Barry Hong. “Accurate calculation of hydration
free energies using macroscopic solvation models”. In: Journal of Physical Chemistry
98.7 (1994), pp. 1978–1988.
[47] Paul Czodrowski et al. “Development, validation, and application of adapted PEOE
charges to estimate pKa values of functional groups in protein-ligand complexes”. In:
Proteins 65.2 (2006), pp. 424–437. issn: 1097-0134. doi: 10.1002/prot.21110.
[48] J M J Swanson, S A Adcock, and J A McCammon. “Optimized Radii for Poisson-
Boltzmann Calculations with the AMBER Force Field”. In: Journal of Chemical
Theory and Computation 1.3 (2005), pp. 484–93. doi: 10.1021/ct049834o.
[49] Chunhu Tan, Lijiang Yang, and Ray Luo. “How Well Does Poisson-Boltzmann Im-
plicit Solvent Agree with Explicit Solvent? A Quantitative Analysis”. In: Journal
of Physical Chemistry B 110.37 (2006), pp. 18680–18687. issn: 1520-6106. doi: 10.
1021/jp063479b.
[50] Michael Holst. “Adaptive numerical treatment of elliptic systems on manifolds”. In:
Journal of Computational Mathematics 15 (2001), pp. 139–191.
[51] “FEtk website”. http://www.fetk.org. url: http://www.fetk.org.
[52] Kaihsu Tai et al. “Finite element simulations of acetylcholine diffusion in neuromus-
cular junctions”. In: Biophysical Journal 84.4 (2003), pp. 2234–2241. doi: 10.1016/
S0006-3495(03)75029-2.
[53] Michael Holst and Faisal Saied. “Multigrid solution of the Poisson—Boltzmann equa-
tion”. In: Journal of Computational Chemistry 14.1 (1993), pp. 105–113. doi: 10.
1002/jcc.540140114.

27
[54] Z Chen, N A Baker, and G W Wei. “Differential geometry based solvation model I:
Eulerian formulation”. In: Journal of Computational Physics 229.22 (2010), pp. 8231–
8258. doi: 10.1016/j.jcp.2010.06.036.
[55] Z Chen, N A Baker, and G W Wei. “Differential geometry based solvation model II:
Lagrangian formulation”. In: Journal of Mathematical Biology 63.6 (2011), pp. 1139–
200. doi: 10.1007/s00285-011-0402-z.
[56] Z Chen et al. “Variational approach for nonpolar solvation analysis”. In: Journal of
Chemical Physics 137.8 (2012), p. 084101. doi: 10.1063/1.4745084.
[57] M D Daily et al. “Origin of parameter degeneracy and molecular shape relationships
in geometric-flow calculations of solvation free energies”. In: Journal of Chemical
Physics 139.20 (2013), p. 204108. doi: 10.1063/1.4832900.
[58] D G Thomas et al. “ISA-TAB-Nano: a specification for sharing nanomaterial research
data in spreadsheet-based format”. In: BMC Biotechnology 13 (2013), p. 2. doi:
10.1186/1472-6750-13-2.
[59] Weihua Geng and Robert Krasny. “A treecode-accelerated boundary integral Poisson-
Boltzmann solver for electrostatics of solvated biomolecules”. In: Journal of Compu-
tational Physics 247 (2013), pp. 62–78.
[60] Peijun Li, Hans Johnston, and Robert Krasny. “A Cartesian treecode for screened
Coulomb interactions”. In: Journal of Computational Physics 228 (2009), pp. 3858–
3868.
[61] André Juffer et al. “The electric potential of a macromolecule in a solvent: A funda-
mental approach”. In: Journal of Computational Physics 97.1 (1991), pp. 144–171.
[62] Michel Sanner, Arthur Olson, and Jean Claude Spehner. “Fast and robust compu-
tation of molecular surfaces”. In: Proc 11th ACM Symp Comp Geom. ACM. 1995,
pp. C6–C7.
[63] Sergio Decherchi and Walter Rocchia. “A general and robust ray-casting-based algo-
rithm for triangulating surfaces at the nanoscale”. In: PLoS ONE 8.4 (2013), e59744.
[64] W Humphrey, A Dalke, and K Schulten. “VMD: visual molecular dynamics”. In:
Journal of Molecular Graphics 14.1 (1996), pp. 33–8, 27–8.
[65] D. Bashford and D. A. Case. “Generalized Born models of macromolecular solvation
effects”. In: Annual Review in Physical Chemistry 51 (2000), pp. 129–152. doi: 10.
1146/annurev.physchem.51.1.129.
[66] Donald Ermak and J. A. McCammon. “Brownian dynamics with hydrodynamic in-
teractions”. In: Journal of Chemical Physics 69.4 (1978), pp. 1352–1360.
[67] Schrödinger, LLC. “The PyMOL Molecular Graphics System, Version 1.8”. PyMOL.
2015.
[68] “PyMOL website”. http://www.pymol.org. url: http://www.pymol.org.
[69] “VMD website”. http://www.ks.uiuc.edu/Research/vmd/.

28
[70] Graham T. Johnson et al. “ePMV Embeds Molecular Modeling into Professional
Animation Software Environments”. In: Structure 19.3 (2011), pp. 293–303. doi: 10.
1016/j.str.2010.12.023.
[71] “MGLTools (AutoDock, PMV, Vision) website”. http://mgltools.scripps.edu/.
url: http://mgltools.scripps.edu/.
[72] “Chimera website”. https://www.cgl.ucsf.edu/chimera/. url: https://www.
cgl.ucsf.edu/chimera/.
[73] E F Pettersen et al. “UCSF Chimera–a visualization system for exploratory research
and analysis”. In: Journal of Computational Chemistry 25.13 (2004), pp. 1605–12.
doi: 10.1002/jcc.20084.
[74] “Jmol website”. http://jmol.sourceforge.net/. url: http://jmol.sourceforge.
net/.
[75] Angel Herraez. “Biomolecules in the computer: Jmol to the rescue”. In: Biochemistry
and Molecular Biology Education 34.4 (2006), pp. 255–261.
[76] D Koes. “3Dmol.js”. http://3dmol.csb.pitt.edu/. url: http://3dmol.csb.
pitt.edu/ (visited on 2016).
[77] Nicholas Rego and David Koes. “3Dmol.js: molecular visualization with WebGL”. In:
Bioinformatics 31.8 (2015), pp. 1322–1324. url: http://3dmol.csb.pitt.edu/.
[78] S Unni et al. “Web servers and services for electrostatics calculations with APBS and
PDB2PQR”. In: Journal of Computational Chemistry 32.7 (2011), pp. 1488–91. doi:
10.1002/jcc.21720.
[79] RO Dror et al. “Structural basis for modulation of a G-protein-coupled receptor by
allosteric drugs”. In: Nature 503.7475 (2013), pp. 295–299.
[80] L Treuel et al. “Impact of Protein Modification on the Protein Corona on Nanopar-
ticles and Nanoparticle-Cell Interactions”. In: ACS Nano 8.1 (2013), pp. 503–513.
[81] Silvia H. De Paoli et al. “The effect of protein corona composition on the interaction
of carbon nanotubes with human blood platelets”. In: Biomaterials 35.24 (2014),
pp. 6182–6194. issn: 0142-9612. url: http://www.sciencedirect.com/science/
article/pii/S0142961214004657.
[82] J Lipfert et al. “Understanding nucleic acid-ion interactions”. In: Annual Review of
Biochemistry 83 (2014), pp. 813–41. doi: 10 . 1146 / annurev - biochem - 060409 -
092720.
[83] V A Roberts et al. “DOT2: Macromolecular docking with improved biophysical mod-
els”. In: Journal of Computational Chemistry 34.20 (2013), pp. 1743–58. doi: 10.
1002/jcc.23304.
[84] T. Evangelidis et al. “An integrated workflow for proteome-wide off-target identifica-
tion and polypharmacology drug design”. In: International Conference on Bioinfor-
matics and Biomedicine Workshops. IEEE. 2009, pp. 32–39. url: doi:%2010.1109/
BIBMW.2012.6470348.

29
[85] E Spiga et al. “Electrostatic-Consistent Coarse-Grained Potentials for Molecular Sim-
ulations of Proteins”. In: Journal of Chemical Theory and Computation 9.8 (2013),
pp. 3515–3526. doi: 10.1021/ct400137q.
[86] Phillip J Stansfeld et al. “MemProtMD: Automated Insertion of Membrane Protein
Structures into Explicit Lipid Membranes”. In: Structure 23.7 (2015), pp. 1350–1361.
[87] S Richter et al. “webPIPSA: a web server for the comparison of protein interaction
properties”. In: Nucleic Acids Res 36.Web Server issue (2008), W276–80. doi: 10.
1093/nar/gkn181.
[88] Gary A. Huber and J. Andrew McCammon. “Browndye: A software package for Brow-
nian dynamics”. In: Computer Physics Communications 181.11 (2010), pp. 1896–
1905. doi: 10.1016/j.cpc.2010.07.022.
[89] Michael Martinez et al. “SDA 7: A modular and parallel implementation of the sim-
ulation of diffusional association software”. In: Journal of Computational Chemistry
36.21 (2015), pp. 1631–45. doi: 10.1002/jcc.23971.
[90] Y Cheng et al. “Finite element analysis of the time-dependent Smoluchowski equa-
tion for acetylcholinesterase reaction rate calculations”. In: Biophysical Journal 92.10
(2007), pp. 3397–406. doi: 10.1529/biophysj.106.102533.
[91] Y Cheng et al. “Continuum simulations of acetylcholine diffusion with reaction-
determined boundaries in neuromuscular junction models”. In: Biophysical Chemistry
127.3 (2007), pp. 129–39. doi: 10.1016/j.bpc.2007.01.003.
[92] Y Song et al. “Continuum diffusion reaction rate calculations of wild-type and mutant
mouse acetylcholinesterase: adaptive finite element analysis”. In: Biophysical Journal
87.3 (2004), pp. 1558–1566.
[93] A H Elcock. “Molecular simulations of diffusion and association in multimacromolec-
ular systems”. In: Methods in Enzymology 383 (2004), pp. 166–98. doi: 10.1016/
S0076-6879(04)83008-8.
[94] P Mereghetti and RC Wade. “Atomic Detail Brownian Dynamics Simulations of Con-
centrated Protein Solutions with a Mean Field Treatment of Hydrodynamic Interac-
tions”. In: Journal of Physical Chemistry 116.29 (2012), pp. 8523–8533.
[95] Michael J Holst. “Clean Object-Oriented C”. http://fetk.org/codes/maloc/api/
html/index.html.
[96] Michael K. Gilson, Kim A. Sharp, and Barry H. Honig. “Calculating the electro-
static potential of molecules in solution: Method and error assessment”. In: Journal
of Computational Chemistry 9.4 (1988), pp. 327–335. doi: 10.1002/jcc.540090407.
[97] W. Im, D. Beglov, and B. Roux. “Continuum Solvation Model: Computation of elec-
trostatic forces from numerical solutions to the Poisson-Boltzmann equation”. In:
Computer Physics Communications 111 (1998), pp. 59–75.
[98] M. J. Schnieders et al. “Polarizable atomic multipole solutes in a Poisson-Boltzmann
continuum”. In: Journal of Chemical Physics 126 (2007), p. 124114. url: https :
//www.ncbi.nlm.nih.gov/pubmed/17411115.

30
 [99] Robert E. Bruccoleri et al. “Finite difference Poisson-Boltzmann electrostatic cal-
culations: Increased accuracy achieved by harmonic dielectric smoothing and charge
antialiasing”. In: Journal of Computational Chemistry 18.2 (1997), pp. 268–276.
[100] Mafalda Nina, Wonpil Im, and Benoit Roux. “Optimized atomic radii for protein
continuum electrostatics solvation forces”. In: Biophysical Chemistry 78.1-2 (1999),
pp. 89–96. doi: 10.1016/S0301-4622(98)00236-1.
[101] D G Thomas et al. “Parameterization of a geometric flow implicit solvation model”.
In: Journal of Computational Chemistry 34.8 (2013), pp. 687–95. doi: 10.1002/jcc.
23181.
[102] Y. Saad and M. Schultz. “GMRES: A generalized minimal residual algorithm for solv-
ing nonsymmetric linear systems”. In: Journal of Scientific and Statistical Computing
7 (1986), pp. 856–869.
[103] “OpenDX”. http://fetk.org/codes/maloc/api/html/index.html. url: http:
//www.opendx.org/.
[104] James Ahrens, Berk Geveci, and Charles Law. ParaView: An End-User Tool for Large
Data Visualization. Elsevier, 2005.

31