CHAPTER THREE

Computational Biochemistry—
Enzyme Mechanisms Explored
Martin Culka2, Florian J. Gisdon2, G. Matthias Ullmann1
Computational Biochemistry, University of Bayreuth, Bayreuth, Germany
1
Corresponding author: e-mail address: [email protected]

Contents
1. Introduction 78
2. Structural Models 81
2.1 Continuum-Electrostatics Models 82
2.2 Quantum-Mechanical Models 83
2.3 Empirical Molecular-Mechanical Models 84
2.4 Hybrid QM/MM Models 86
2.5 Pseudo-Atomistic Models 89
3. Calculating Enzymatic Mechanisms 91
3.1 Thermodynamic Properties of Biological Systems 91
3.2 Kinetic Properties of Biochemical Systems 93
3.3 Mechanic and Energetic Properties of Molecular Reaction Paths 97
3.4 Path Search Strategy 104
4. Going Beyond the Exploration of the Reaction Paths 104
References 105

Abstract
Understanding enzyme mechanisms is a major task to achieve in order to comprehend
how living cells work. Recent advances in biomolecular research provide huge amount
of data on enzyme kinetics and structure. The analysis of diverse experimental results
and their combination into an overall picture is, however, often challenging. Micro-
scopic details of the enzymatic processes are often anticipated based on several hints
from macroscopic experimental data. Computational biochemistry aims at creation of a
computational model of an enzyme in order to explain microscopic details of the cat-
alytic process and reproduce or predict macroscopic experimental findings. Results of
such computations are in part complementary to experimental data and provide an
explanation of a biochemical process at the microscopic level. In order to evaluate
the mechanism of an enzyme, a structural model is constructed which can be analyzed
by several theoretical approaches. Several simulation methods can and should be com-
bined to get a reliable picture of the process of interest. Furthermore, abstract models of

2
These authors contributed equally.

Advances in Protein Chemistry and Structural Biology, Volume 109 # 2017 Elsevier Inc. 77
ISSN 1876-1623 All rights reserved.
http://dx.doi.org/10.1016/bs.apcsb.2017.04.004
78 Martin Culka et al.

biological systems can be constructed combining computational and experimental

data. In this review, we discuss structural computational models of enzymatic systems.
We first discuss various models to simulate enzyme catalysis. Furthermore, we review
various approaches how to characterize the enzyme mechanism both qualitatively
and quantitatively using different modeling approaches.

1. INTRODUCTION
“Certainly no subject or field is making more progress on so many
fronts at the present moment than biology, and if we were to name the most
powerful assumption of all, which leads one on and on in an attempt to
understand life, it is that all things are made of atoms, and that everything
that living things do can be understood in terms of the jigglings and wig-
glings of atoms” (Feynman, 1964). These words were said by the well-
known physicist Richard Feynman in his famous Lectures on Physics now
more than 50 years ago. Until today, this sentence has not lost its validity
and is the basis for much of the biomolecular research, maybe today even
more than 50 years ago. Today, we are beginning to understand how the
interplay of atoms and molecules lead to the complex processes that we find
in living systems. Recent advances in genomics and systems biology help to
gain more and more insights into the molecular organizations of the cell,
which are nevertheless still too complex to allow a full overview of all
the biochemical processes at atomic detail. New techniques in structural
biology such as the free-electron laser allow to analyze structurally the kinet-
ics of molecular processes at an atomic level (Nango et al., 2016; Pande et al.,
2016). On the other hand, modern electron microscopy allows us to gain
insights into large molecular assemblies that were not accessible to structural
investigations a few years ago (Bartesaghi et al., 2015; K€ uhlbrandt, 2014).
These new techniques complement traditional techniques of structural biol-
ogy such as X-ray crystallography and NMR. In the past, a structural char-
acterization of a protein was considered the final goal of an investigation.
Nowadays to gain a deep understanding of an enzymatic mechanism, a
structure is just the beginning. The structural information needs to be com-
pleted by other experimental information for instance from spectroscopy,
kinetics, or electrochemistry. Often, all these different aspects are difficult
to merge. Thus, it is important to approach the enzymatic mechanism also
from a theoretical side. With the help of modern methods from computa-
tional chemistry, it is possible to gain insights into enzymatic mechanisms
and to complete the picture.
Computational Biochemistry—Enzyme Mechanisms Explored 79

In order to understand experimental data, we build models of reality and

use these models in our theoretical considerations. Models are not only used
for theoretical calculations, but also the interpretation of experimental data
relies on models of the studied system. A model is a generalized hypothetical
description used for analyzing or explaining a system. It is a simplified repre-
sentation of a real system intended to enhance our ability to understand, pre-
dict, and control the behavior of the real system. When a model is made, there
are always approximations required. Therefore a model is always an idealized
representation of the real system. A model reproduces only certain aspects of
the real system, only those that are relevant for the properties under study.
Other aspects of the same real system may not be described equally well by
the model, since it was originally made for another purpose. A model should
be able to explain experimental data and make predictions about the outcome
of new experiments. In order to be able to make predictions, a model has to be
complicated enough to represent all important aspects of a real system at an
appropriate level of description. However, the more complicated and complex
a model is, the more difficult it becomes to interpret the results of the model. It
is therefore required that the model is as complicated as necessary and not more
complicated in order to give insights into the behavior of the real system. Most
importantly, a model should promote our understanding of nature.
A model of a real system is generally constructed in several steps (see
Fig. 1). The first and probably most important step is the construction of
the conceptual model. In this step, the real system is translated into an ide-
alized model system. Having in mind which properties of the system are
interesting, the features of the system that are important to reproduce these
properties are selected and it is decided how to use them to describe the
desired aspects of the system. This first step requires a detailed inspection
of the system that should be modeled in order to decide which details are
required for the desired representation. When only a qualitative understand-
ing is requested, the conceptual model is often enough to picture the system
and thus it represents the final goal of the investigation. Instead, when a
quantitative or semiquantitative understanding of the system is desired,
the conceptual model needs to be translated into a mathematical model.
A concrete physical theory is required to forge the mathematical model
designed in the conceptual model. If this mathematical model can be solved
analytically, the goal is reached and the behavior of the model system can be
compared with the behavior of the real system. However, the mathematical
models are often too complex to be solved analytically. In this case, the
mathematical model needs to be translated into a discrete mathematical
80 Martin Culka et al.

Physics, Biology Formal theory

Real Conceptual Mathematical
system Chemistry model Mathematics model

Experimental
data Numerical
mathematics
Analytical
Simulation solution
data

Model Simulation Computer Algorithms Discrete

system model model
Calculation Programming

Fig. 1 Building theoretical models of real systems. This scheme depicts the stages of
building theoretical models that can help to analyze real systems. These theoretical
models are abstractions of the real systems that help to understand the behavior of
the system. Mathematical modeling allows a quantitative comparison between the real
system and the model.

model using methods from numerical mathematics. This discrete model can
then be implemented in computer programs. The computer model can then
be used to perform computer simulations and calculations using sets of
parameters. The resulting simulation data can be compared with experimen-
tal data and so the quality of the model can be judged. The quantities that
have been calculated from the model can lead to a better understanding of
the real system. Moreover, the analysis of the effect of well-defined changes
of the initial parameter can provide deeper insights into the behavior of the
real system.
One of the most basic model that is widely used in the description of
chemical systems is the Born–Oppenheimer approximation according to
which one can separate the motion of nuclei from the motion of the elec-
trons. This approximation allows us to describe molecular structures by
defining the nuclear coordinates and to speak about different electronic
states of a molecule. One fundamental concept that is used in molecular bio-
physics is the molecular energy landscape which describes the energy of a
molecular system with N atoms in dependence of its 3N coordinates. This
energy landscape contains minima which represent stable states of the sys-
tems, (first-order) saddle points which represent transition states of chemical
reactions, and minimum energy paths which represent the trajectory along
which chemical reactions occur. A large part of theoretical biochemistry is
concerned with exploring this energy landscape and to extract different kind
Computational Biochemistry—Enzyme Mechanisms Explored 81

of data from it. In this article, we review some of the most important
methods to explore this energy landscapes and to explain how they can
be used to explore enzymatic mechanisms.

2. STRUCTURAL MODELS
The most computational enzymology techniques construct a structural
model of the enzyme in question, usually using its experiment-derived 3D
structure as a starting point. A structural model provides a way to calculate
energy of the given structure, by using various approximations of the real
picture to stay computationally feasible. Based on energy differences
between different states within structural computer models of enzymes, ther-
modynamics, and kinetics of enzyme catalysis can be addressed. With the use
of structural models (Fig. 2), researchers aim to track the events within the

Implicit
Continuum-electrostatics

Pseudo-atomistic mechanics
Structural models

Molecular mechanics

Reactive force fields

Quantum mechanics
Explicit
ns

es

ls
es
om

el
tro

ul

ul

ex

C
ec

ec
At
ec

pl
ol

ol

om
El

C
ro
ac
M

Fig. 2 Structural models for the simulation of catalytic events within enzymes. The ver-
tical axis represents an ascending gradient from explicit to implicit models. The system
description lose the detailed (explicit) character in favor of averaged (implicit) descrip-
tion of physical properties. The horizontal axis shows the information that can be
obtained from the models, and together illustrates the size of the biological system,
which is computationally feasible to treat. The model bars indicate possible applications.
The fading areas show the potential of future developments enabled by improvements
of algorithmic and computational power. All models can be combined into the so-called
hybrid or multiscale models, which treat different parts of the biological system with the
appropriate method, to enhance the effectiveness of the simulation and the quality of
results.
82 Martin Culka et al.

enzyme that lead to catalysis. Models used to study enzyme mechanisms

can be atomistic, i.e., they explicitly describe the behavior of electrons
and nuclei, or pseudo-atomistic, i.e., they treat groups of atoms as one
entity. In contrast to these particle-based models, continuum models
describe a system in terms of continuous properties assigned to the space.
Typical continuum models in enzymology are implicit solvent models.
Their aim is to describe the average properties of a solvent environment,
instead of discrete contributors of its parts. In this section, we briefly discuss
various types of structural models and their combinations to simulate
enzymatic activities.

2.1 Continuum-Electrostatics Models

Since most of the effects in biochemistry are dominated by electrostatics,
a pure electrostatic model can be used to satisfactory describe many features
of biomolecular systems while being computationally efficient. The most
commonly used continuum-electrostatics model in biochemistry relies on
the Poisson–Boltzmann equation. For a recent review the reader is referred
to Ullmann and Bombarda (2014). The basic idea of this continuum-
electrostatics model is to describe the protein as a low-dielectric region,
which is embedded in (aqueous) solvent, described as a high-dielectric
region. The charge distribution of the protein is described by a fixed charge
distribution in the low-dielectric region, which is given by the molecular
structure of the protein. The charge distribution of the protein is modeled
by (fractional) point charges that are placed at the center of the atoms. The
dissolved ions are represented by a Boltzmann-distributed charge density.
The boundary of the low-dielectric region is defined by the solvent acces-
sible surface of the protein.
The Poisson–Boltzmann equation is usually solved numerically
(Honig & Nicholls, 1995; Warwicker & Watson, 1982). The solution of
the Poisson–Boltzmann equation can be expressed as a potential ϕ(r), that
is, composed of two contributions:
XM
qi
ϕðrÞ ¼ + ϕrf ðrÞ (1)
i¼1
4πεp jr  r0i j

First, the Coulomb potential at the position r caused by M point charges qi at

positions r0i in a medium with a permittivity εp, and second, the reaction field
potential ϕrf(r), arising from the M point charges qi and the dielectric
Computational Biochemistry—Enzyme Mechanisms Explored 83

boundary between the protein and the solvent, as well as from the distribu-
tion of ions in the solution. Electrostatic energy can be obtained by integrat-
ing the potential distribution over the space.
Continuum-electrostatics models such as the above described Poisson–
Boltzmann model have various applications. The most straightforward is cal-
culation of solvation energies and visualization of electrostatic potentials
(Baker, Sept, Joseph, Holst, & McCammon, 2001). More advanced appli-
cations allow to calculate energies of different protonation and oxidation
microstates of the protein, as well as ligand binding energies and many other
applications (Ullmann et al., 2008). Furthermore, the Poisson–Boltzmann
model or other popular continuum models such as COSMO (Klamt &
Sch€ €rmann, 1993) are often combined with more detailed models to sim-
uu
ulate solvent effects in a computationally affordable way (Chen, Noodleman,
Case, & Bashford, 1994; Li, Nelson, Peng, Bashford, & Noodleman, 1998;
Liu et al., 2004).

2.2 Quantum-Mechanical Models

The most detailed structural models that treat both electrons and nuclei
explicitly are based on quantum mechanics (QM). In QM models, the
energy of the structure is derived from solving an approximate Schr€ odinger
equation. One way is to employ ab initio molecular orbital wave function
approaches such as the single-determinant Hartree–Fock (HF) method
(Cramer, 2004). The major drawback of the HF method is the mean-field
treatment of electron repulsion referred to as the electron correlation prob-
lem. Møller–Plesset theory (Møller & Plesset, 1934) addresses the electron
correlation effects by means of Rayleigh–Schr€ odinger perturbation theory.
The most common second-order version is called MP2 (Head-Gordon,

Pople, & Frisch, 1988). In coupled cluster (CC) methods (Cı́žek, 1966),
multielectron wave-functions are constructed using the exponential cluster
operator to account for the electron correlation. The latter two more accu-
rate approaches are, however, still very costly for larger systems. For com-
plex biological systems, fast semiempirical (SE) approximations of the
Hartree–Fock theory have become very popular in the past and are still used
and developed up to now (Stewart, 2013; Řezáč & Hobza, 2012;
Yilmazer & Korth, 2015). The SE methods achieve considerable calculation
speed-up by parameterizing various parts of the HF theory in order to repro-
duce experimental or high-level ab initio QM results.
84 Martin Culka et al.

Density functional theory (DFT) is an alternative ab initio approach that

has become very popular for biological systems due to its favorable price/
performance ratio (van Mourik, B€ uhl, & Gaigeot, 2014). Instead of
searching for a multidimensional wave-function that describes the position
of every electron, the whole problem is solved from the point of view of
total electron distribution (density) in the space. The electron correlation
is taken into account in DFT methods, albeit on approximate level. In addi-
tion, hybrid DFT–HF methods (Becke, 1993; Zhao & Truhlar, 2008)
combine the best of the both worlds and have turned out to be very suc-
cessful (Bryantsev, Diallo, Van Duin, & Goddard, 2009). Also DFT has a
parameterized alternative called density-functional tight-binding (DFTB).
DFTB is designed to reproduce DFT results rather than fit empirical data
like SE methods (Elstner, 2006). QM methods play an inevitable role in
computational studies of enzyme-catalyzed reactions, since some amount
of quantum-mechanical treatment or QM-based parameterization is always
needed in the active site.
Even with the enormous growth of computational power in last decades,
biological systems such as enzymes can hardly be treated in the full extend by
QM methods. Successful attempts have been made to simulate whole pro-
teins purely by QM methods (Cole & Hine, 2016; Todorovic, Bowler,
Gillan, & Miyazaki, 2013). Nevertheless, in order to make the calculations
feasible, researches often reduce the QM model of the enzyme to the active
site residues only, by constructing a so-called cluster model. Continuum sol-
vent is used to mimic the protein as a low-dielectric environment beyond
the shell of the active site residues. Cluster models of enzyme active sites
have been used to elucidate the mechanisms of various enzymes, including
difficult cases such as radical enzymes (Feliks & Ullmann, 2012) or metallo-
enzymes (Li & Ryde, 2014; Manta, Raushel, & Himo, 2014). Although a
rather simple approach, due to the absence of various sources of errors that
more complex methods may introduce, the cluster models retain their role
in computational enzymology (Georgieva & Himo, 2010). Considering the
influence of the protein environment, the continuum solvent might be rep-
laced by the protein residues simulated on an empirical level, see QM/MM
approaches later.

2.3 Empirical Molecular-Mechanical Models

Due to the enormous computational cost of the QM approaches, alternative
empirical methods that use principles of classical physics have been
Computational Biochemistry—Enzyme Mechanisms Explored 85

developed for treatment of complex systems. These methods are generally

referred to as molecular mechanics (MM). In classical MM models, atoms
are represented as spheres connected by springs. Empirical parameters have
been derived to reproduce the expected behavior of biomolecules. These
include spring constants for bond lengths, angles, and torsion angles as well
as nonbonded interaction parameters such as van der Waals radii and partial
atomic charges. The whole set of MM parameters is called force field (FF).
The total potential energy of a system is calculated as a sum of all bonded and
nonbonded contributions:

! X X X
V ðR Þ ¼ Kb ðb  b0 Þ2 + Kθ ðθ  θ0 Þ2 + Kφ ð1 + cos ðnφ  δÞÞ
bonds angles torsions
8 9
>
> >
>
>
> >
>
>
> >
>
>
> 2 >
>
>
> !12 !6 3 >
>
X X < Rijmin Rijmin =
εijmin 4 5 qi qj
+ Kω ðω  ω0 Þ2 + 2 +
>
> >
0 εrij >
r r
nonbonded> >
ij ij 4πε
improper >
> |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl
ffl {zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl
ffl } |fflfflfflffl{zfflfflfflffl} >
>
>
> >
>
>
>
pairs > >
>
torsions
: LennardJones Coulomb >
;
potential law

(2)

where b is the bond length, θ is the bond angle, ϕ the torsion angle, ω the
improper torsion angle. The 0 indices represent the equilibrium values of
given parameters and K the respective spring constants. Van der Waals
interactions are represented by the Lennard–Jones potential where Rijmin is
the distance energy minimum position, rij is the distance between two
atoms and εminij is the interaction constant. Electrostatic interactions are
described by the Coulomb law where qi and qj are the partial charges of
the interacting atoms, rij is their distance and ε0ε is the environment
permittivity.
Several empirical MM force fields for biomolecules have been devel-
oped over the years, e.g., CHARMM (MacKerell et al., 1998), AMBER
(Cornell et al., 1996), GROMOS (Oostenbrink, Villa, Mark, & Van
Gunsteren, 2004), or OPLS (Jorgensen, Maxwell, & Tirado-Rives,
1996). MM models proved to be very useful especially for studying protein
dynamics by molecular dynamics (MD) simulations. In context of enzyme
catalysis, pure MM models can be used to study conformational changes or
substrate binding dynamics (Costa, Batista, Bisch, & Perahia, 2015;
86 Martin Culka et al.

Gilson & Zhou, 2007). When one wants to model details of a catalytic
mechanism in the enzyme active site, however, the conventional MM
models are insufficient. They are, however, often used to model the non-
reactive parts of the enzyme in hybrid QM/MM approaches, as described
below.
Nevertheless, attempts have been made to allow for chemical reactivity
within the empirical force fields without employing QM principles. These
approaches are referred to as reactive force fields. One of the most widely
used reactive force field, that has been also utilized for biological systems,
is called ReaxFF (Senftle et al., 2016; van Duin, Dasgupta, Lorant, &
Goddard III, 2001). In ReaxFF, the empirical energy equation (Eq. 2) is
modified to follow a more complex bond order formalism rather than the
balls-on-springs formalism. Although reactive force fields have been origi-
nally developed for material chemistry applications (Liang et al., 2013), first
attempts have been made to model peptides, small proteins (Golkaram,
Shin, & van Duin, 2014; Monti et al., 2013), and DNA (Verlackt et al.,
2015) using ReaxFF.
Empirical force fields are popular because of their speed and their often
realistic description of molecules, which is a consequence of extensive
parameterization to reproduce experimentally derived structures and behav-
iors. It has been shown that MM methods are able to reproduce experimen-
tally derived molecular structures, except for nonstandard regions (Kulik,
Luehr, Ufimtsev, & Martı́nez, 2012). In such regions, ab initio QM methods
consistently perform better. Despite this benefit, as already mentioned
above, full ab initio studies of proteins are restricted to rather small systems
(Kulik et al., 2012). Instead, hybrid approaches (referred to as QM/MM)
have been developed to combine accurate ab initio models and fast empirical
models.

2.4 Hybrid QM/MM Models

In QM/MM, the active site, where the reaction takes place, is treated on
QM level, while the rest of the enzyme is modeled by an empirical MM
force field. Since its development in 1976 (Warshel & Levitt, 1976),
QM/MM has become the most popular approach to study enzymatic reac-
tions (Senn & Thiel, 2007, 2009).
The key feature that QM/MM methods have to deal with is the inter-
action between QM and MM regions. For the description of a QM/MM
system, two energy schemes evolved, a subtractive scheme and an additive
Computational Biochemistry—Enzyme Mechanisms Explored 87

scheme. The energy in the subtractive QM/MM scheme (Eq. 3) is obtained

by an MM calculation of the entire system (S) with the inner part (I) cut out
and replaced by a QM calculation.
sub
EQM=MM ¼ EMM ðSÞ + EQM ðIÞ  EMM ðIÞ (3)

The QM energy of the quantum mechanically treated inner part (EQM(I)) is

added to the MM energy of the entire system (EMM(S)), but the MM energy
of the inner part (EMM(I)) has to be subtracted to avoid double counting.
The most widely used approach that employs the subtractive scheme is
ONIOM (Maseras & Morokuma, 1995; Svensson et al., 1996), which is
capable of combining n layers of any implemented QM or MM approach.
The potential drawback of the subtractive scheme is that a proper MM
description of the active site region is required and this is often difficult
to achieve for substrate or enzyme cofactors. In the additive QM/MM
scheme (Field, 2007), the MM energy is only calculated on the outer system
(O) instead of S (Eq. 4). The QM energy is added and additionally, a cou-
pling term (EQM/MM(I, O)) is introduced, which treats the interaction of the
QM and the MM part.
add
EQM=MM ¼ EMM ðOÞ + EQM ðIÞ + EQM=MM ðI,OÞ (4)

The EQM/MM(I, O) term itself is composed of van der Waals, electrostatic,

and bonded contributions. Van der Waals interactions are fully described on
the MM level. The common description with the Lennard–Jones potential is
also applied to QM atoms. Therefore it is necessary to have suitable param-
eters for all QM atoms. Since the main impact of van der Waals interactions
occurs in a short range, atoms near the boundary should be well para-
metrized. Thus, changes of QM atom properties during catalysis have no
meaningful influence on EQM/MM(I, O), as long as they appear not close
to the boundary. In proteins, important catalytic contributions are mediated
by electrostatic interactions. All charges and partial charges cause electro-
static forces. Within one model, the interactions of charges are well defined,
but the coupling between two models needs to be adapted.
The electrostatic term within EQM/MM(I, O) describes the coupling
between the QM charge density and the MM charge model. One efficient
but simple method to treat electrostatic interaction is mechanical embedding
(Bakowies & Thiel, 1991), where the MM model is directly applied to the
QM region. Since the charge density of the QM region is then represented
88 Martin Culka et al.

within the MM model, the MM region can react to it. A major drawback is,
however, that the QM calculation is performed in the absence of the elec-
trostatic MM environment, thus the atoms in the QM part cannot react to
their full environment. But especially this electrostatic environment is cru-
cial for enzymatic catalysis (Zhang, 2013). The electrostatic embedding
approach (Bakowies & Thiel, 1991) treats this important interaction by pro-
viding the electrostatic environment for the QM calculation. The electro-
static environment appears as a one-electron term in the QM Hamiltonian.
Thus the charge distribution in the QM region is polarized according to the
MM charges. Despite the higher accuracy in the calculation, charge leakage
effects can occur at the boundary of the QM region, where the QM charge
density is polarized in immediate proximity by MM charges. In both embed-
ding schemes the MM charges are rigid and do not react to the QM charge
density. In a polarized embedding scheme the polarization happens in both,
the QM region direction and the MM region direction. For the polarization
of the MM region, a polarizable force field has to be applied on the MM level
(Thompson & Schenter, 1995). This approach of polarized embedding is in
general the most accurate, though computational demanding one.
In the majority of QM/MM simulations of enzymes, one has to deal with
covalent bonds on QM/MM boundary (e.g., bond between an amino acid
side chain and the protein backbone) in addition to the nonbonded inter-
actions. The boundary QM atom valency needs to be saturated to allow
for proper electronic structure calculation. The simplest approach to treat
the QM/MM boundary is the link atom approach (Field, Bash, &
Karplus, 1990; Singh & Kollman, 1986), where the QM atom is capped
by an auxiliary atom (usually hydrogen) that is constraint in the direction
of the QM/MM bond. This generates a problem of QM density over-
polarization near the boundary that has to be treated. Alternative approaches
are based on frozen hybrid orbitals (Amara, Field, Alhambra, & Gao, 2000;
Thery, Rinaldi, Rivail, Maigret, & Ferenczy, 1994). A frontier atom is cho-
sen in the QM/MM boundary and a set of suitably oriented localized orbitals
is placed on it. This treatment allows to converge on more proper electronic
structure in the boundary region. On the other hand frozen orbital appro-
aches are more technically demanding and require calibration for the specific
bond and QM method.
A conceptually different QM/MM approach that has been used to study
enzyme catalysis is the empirical valence bond (EVB) (Kamerlin & Warshel,
2010, 2011; Warshel, 2003). EVB uses the valence-bond (VB) approach for
quantum description of the enzyme active site. The QM methods discussed
Computational Biochemistry—Enzyme Mechanisms Explored 89

so far are based on molecular orbital theory that combines the atomic orbitals
into a molecular orbital wave-function. The VB methods instead describe
the system as a linear combination of all possible states where electrons
occupy localized orbitals. In EVB, the stationary points along the reaction
path are described by an MM force field, while the transitions between them
are treated by an SE valence-bond QM approach (Shurki, Derat, Barrozo, &
Kamerlin, 2015). Thus for every reaction step simulated by EVB, a set of
empirical parameters has to be derived. This is usually done on a model reac-
tion in solution or in gas phase, where the parameters are fitted to reproduce
the experimental data or ab initio QM results (Åqvist & Warshel, 1993).
Once well calibrated, extensive conformational sampling along the reaction
pathway can be achieved in reasonable time with EVB in order to get a
proper free energy landscape (see Section 3.3.2). On the other hand, the
major disadvantage of EVB in comparison to common QM and QM/
MM methods is the need of specific calibration for every reaction step. In
fact, prior knowledge of the reaction mechanism is required to perform
an EVB simulation, while unknown mechanism alternatives can be discov-
ered within conventional ab initio QM and QM/MM models.

2.5 Pseudo-Atomistic Models

Although the computational and algorithmic power is increasing, modeling
of larger protein systems remains difficult. Bridging between molecular
behavior and biological system function requires different levels of abstrac-
tion to manage the huge amount of data in a reasonable time. Pseudo-
atomistic models provide one level of abstraction by introducing pseudo
atoms, which comprise groups of atoms, several amino acids or even whole
molecules. These pseudo atoms are represented as one entity with all atoms
within such an entity considered to be frozen. They are modeled to simulate
the essential or averaged behavior of such entities and their interactions. Pure
pseudo-atomistic models contain merely homogeneous or heterogeneous
pseudo atom species. Besides these pure models, multiscale models are deve-
loped, which combine atomistic resolution of some molecular parts with
pseudo-atomistic simplification. Because of the simplification, to reduce
multiple individual atoms to single entities, comparable to coarser grains,
pseudo-atomistic models are referred to as coarse-grained (CG) models.
The advantage of a CG model is, that the degrees of freedom are decreased.
Therefore larger systems can be studied and longer time scales can be reached
than using classical atomistic models. In the first application of such a
90 Martin Culka et al.

pseudo-atomistic model on proteins (Levitt & Warshel, 1975), it is pointed

out that this simplification brings additional benefit in the interpretation of
results. With the reduction in degrees of freedom, the energy landscape for
the system is smoothened. That is, less important movements or details are
averaged, and essential features can be focused. Since CG models allow the
treatment of bigger biological systems, and can simplify the analysis, their
development was increasing in the last years, as some recent reviews show
(Ingólfsson et al., 2014; Kar & Feig, 2014; Kmiecik et al., 2016; Meier et al.,
2013; Noid, 2013; Riniker, Allison, & van Gunsteren, 2012; Saunders &
Voth, 2013).
The concept of a coarse-grained model is a reduction of degrees of free-
dom. This reduction is achieved by replacing several individual atoms by
pseudo atoms. For an optimal simulation it is important that the coarse-
grained system keeps the overall character of the all-atom system. Therefore
it is necessary to provide rules for the coarse-grained particles to behave.
Analogue to the all-atom MM force fields described above, the behavior
is evaluated by an empirical potential energy function. Therefore a protein
is represented as an elastic network of coarse-grained beads connected to
each other by elastic springs. In a simple case, each bead represents one
amino acid. Such a representation is useful to investigate domain motions
in large systems, where longer time scales have to be achieved (R€ ucker,
Wieninger, Ullmann, & Sticht, 2012). A harmonic bond potential is applied
to each pair of beads, to allow them to move with respect to their surround-
ing. The potential energy function (Eq. 5) is a power series expansion near a
minimum structure r0, represented as a 3N dimensional Cartesian coordi-
nate vector.
1
V ðrÞ ¼ V ð0Þ + rV ðrÞ + rT Hr (5)
2
All movements of the system can be evaluated relative to that minimum. The
constant term V (0) describes the energy at the minimum position and can be
set to zero. The first derivative of the potential V (r) is the gradient, which is
zero at a minimum. The elastic network potential simplifies to the second-
order term, which is a sum of pairwise potentials, with the second derivative
matrix H providing the force constants.
In recent years more and more complex models have been developed,
which better and better represent the nature of biological systems
(Tozzini, 2005). A further development in the treatment of big biological
systems is the combination with more accurate methods, such as MM or
Computational Biochemistry—Enzyme Mechanisms Explored 91

QM/MM. Such so-called multiscale or multiresolution models are suitable

for the description of catalytic effects within a biological relevant surround-
ing. In a recent study (Sokkar, Boulanger, Thiel, & Sanchez-Garcia, 2015)
chorismate mutase and p-hydroxybenzoate hydroxylase were investigated
with a QM/MM/CG approach, by modeling the catalytically relevant part
of the enzyme in QM, the remaining amino acids in MM and the sur-
rounding solvent in CG. In such multiscale models it is important to define
efficient data exchange between the different potentials. A proper exchange
between the different resolutions becomes even more important in the adap-
tive resolution multiscale models (Heyden & Truhlar, 2008; Shen & Hu,
2014; Zavadlav, Melo, Marrink, & Praprotnik, 2015). Here, specific regions
are defined, similarly to the QM/MM/CG approach, but with connecting
buffer zones, which enable the entering particles to change their resolution.
Such approaches incorporate all benefits of the multiscale models and the
flexibility to adapt to major changes within the modeled biological system.
A detailed list of CG models and available programs with an extensive
description of the current state of the art can be found in a recent review
(Kmiecik et al., 2016).

3. CALCULATING ENZYMATIC MECHANISMS

As discussed in the introduction, in computational enzymology one is
interested in explaining macroscopic thermodynamic and kinetic data
derived from experiment by microscopic models. In Section 2, we reviewed
types of structural model environments to study the enzyme catalysis. In this
section, we review how to derive thermodynamic and kinetic parameters
within these models.

3.1 Thermodynamic Properties of Biological Systems

Each simulation of biological systems requires the analysis of its thermo-
dynamic equilibrium states. At equilibrium, no driving forces act on the sys-
tem. Thus, these states are thermodynamically stable. The relative stability of
an initial state and a final state, for instance, allows to predict, if a reaction is
endergonic or exergonic.
The direction in which a system changes can be evaluated by changes in
free energy. Dependent on the simulation, the Gibbs free energy ΔG (Eq. 6)
at constant temperature T and pressure P, or the Helmholtz free energy ΔA
(Eq. 7) at constant temperature and volume V can be calculated. ΔA is
comprised of the internal energy ΔU of the system, and temperature scaled
92 Martin Culka et al.

entropy ΔS, a measure for thermally accessible configurations. For ΔG the

system energy is the enthalpy ΔH, which is comprised of internal energy ΔU
of a system and the work for its volume expansion PV at adjusted pressure.
Since experiments are mainly performed at constant pressure, we will here-
after refer to free energy as the Gibbs free energy ΔG.

ΔG ¼ ΔH  T ΔS ¼ ΔU + PV  T ΔS (6)
ΔA ¼ ΔU  T ΔS (7)

The computational prediction of thermodynamic properties is based on the

analysis of ensembles. An ensemble is a large number of virtual copies of a
system with identical macroscopic properties. To obtain reasonable ensem-
bles, sampling methods such as MD approaches or Monte Carlo (MC)
approaches are applied (Paquet & Viktor, 2015). While MD approaches
are usually based on Newtonian mechanics, MC methods are based on
repeated random sampling using certain thermodynamics principles. MD
simulations are used to investigate the time evolution of biological systems.
Applying Newton’s second law of motion (Eq. 8) the movements of atoms
in a system are described with the classical equation of motion.

Fi ¼ ri E ¼ mi ai (8)

The force Fi acting on atom i with mass mi determines its acceleration ai,
while the force is defined by the negative gradient of the potential energy
function riE with respect to the coordinates of atom i. By integrating
the equations of motion for all atoms over small time steps, it is possible
to obtain their time dependent positions, just providing the initial positions
and the initial velocities. The initial positions are directly given by the coor-
dinates of the atoms, and the initial velocities are usually distributed ran-
domly with a certain probability distribution. Only the acceleration is
needed, which can be obtained by the gradient of the potential energy func-
tion and the atomic mass (Eq. 8). The velocities of the atoms define the
temperature, which is an important thermodynamic property in MD simu-
lations. An equilibration simulation is performed until the system reaches an
equilibrium state, which is a global minimum on the energy surface. During
this process it is important that the system has enough time and energy
(i.e., temperature) to escape local minima. The system can in some cases any-
way remain kinetically trapped in a local minimum. Annealing simulations
start at higher temperature to overcome such energy barriers and gradually
Computational Biochemistry—Enzyme Mechanisms Explored 93

decrease it to the desired temperature, allowing the system to reach an equi-

librium state.
In contrast to MD simulations, statistical approaches, such as Monte
Carlo sampling, are independent of force evaluation. The sampling is
based on random movements of the system from a set of possible move-
ments, while new configurations are accepted in the case of lower energy,
ΔE < 0. Configurations with higher energy are accepted according to the
Metropolis criterion, which ensures importance sampling of the microstates
by a Boltzmann factor exp ðΔE =kB T Þ. Monte Carlo simulations create
ensembles by energetically evaluating random movements of a system with
the possibility to permit less favorable movements. In contrast to MD,
Monte Carlo moves are not bound to small motions that can happen during
the time integration step. Thus in Monte Carlo, it is often easier to over-
come energy barriers, which prevents the system to be kinetically trapped.
On the other hand in standard Metropolis Monte Carlo, dynamic properties
of the system are not accessible, since the time dimension is not considered in
Monte Carlo. If only sampling is considered, the same level of convergence
is achieved faster by Monte Carlo simulations than with MD, provided
efficient Monte Carlo moves are chosen (Jorgensen & Tirado-Rives,
1996). Monte Carlo sampling can also predict thermodynamic properties
such as preferred protonation states or redox states under specific environ-
mental conditions such as pH, solution redox potential, or membrane
potential (Bombarda, Becker, & Ullmann, 2006; Calimet & Ullmann,
2004; Ullmann, 2000; Ullmann & Ullmann, 2012). These are equally
important properties to define a state, which can be used for further kinetic
analysis of biological systems.

3.2 Kinetic Properties of Biochemical Systems

Once thermodynamics of the stable conformations and protonation or oxi-
dation states of the enzyme are identified, the interest of a computational
enzymologist turns to kinetic properties. Enzyme kinetics aims at tracking
the rates of chemical or physical processes in the enzyme. In computational
enzymology, discrete reaction steps are first addressed within a structural
model before the rate of overall reactions can be related to experimental
measurements.
Reaction rate is characterized by a temperature-dependent rate constant,
which can be viewed as a probability factor for overcoming a free energy
94 Martin Culka et al.

A B

Fig. 3 (A) Reaction coordinate for interconversion between two stable states is thermo-
dynamically characterized by reaction free energy ΔG° and kinetically by the free acti-
vation energy ΔG6¼ . (B) Marcus model of two parabolas representing the two stable
electronic configurations of the system. Reaction free energy ΔG° and the reorganiza-
tion energy λ used for calculation of the free activation energy ΔG6¼ of the electron
transfer are highlighted.

barrier between two stable states of the system (Fig. 3A). The reaction barrier
can be related to the reaction rate by the Arrhenius law:
 
kðT Þ ¼ A exp βΔG6¼ (9)

where ΔG6¼ is the free activation energy, β is 1 =kB T (kB is the Boltzmann
constant, T is the absolute temperature), and A is a preexponential factor.
Eq. (9) is further adopted to theoretically describe rates of different processes
involved in the enzyme function.
The first category are long-range electron transfer reactions, which can
be regarded as nonadiabatic processes. A Marcus model of two harmonic
potentials representing an initial and a final electronic configuration
(Fig. 3B) can be used to describe the reaction coordinate. The free activation
energy can be calculated from reaction free energy ΔG° and reorganization
energy λ, which is the energy needed to perform the structural changes
within the system:
ðΔG° + λÞ2
ΔG6¼ ¼ (10)
4λ
The reaction rate constant (9) can then be adopted for electron transfers
using Marcus theory for the activation energy and Fermi’s golden rule
(Marcus & Sutin, 1985) for the preexponential factor:
!
2π 2 1 ðΔG° + λÞ2
kET ðT Þ ¼ HDA pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi exp β (11)
ℏ 4πβ1 λ 4λ
Computational Biochemistry—Enzyme Mechanisms Explored 95

where ℏ is the reduced Planck constant and HDA 2

is the electronic coupling
between the reactant state and the product state, which, to a good appro-
ximation, decreases exponentially with the distance between donor and
acceptor (Gamow, 1928; Gray & Winkler, 2005). Marcus theory for
instance in the form of the Moser–Dutton ruler (Moser, Keske,
Warncke, Farid, & Dutton, 1992; Page, Moser, Chen, & Dutton, 1999)
can be combined with a microstate model used for calculating microscopic
redox potentials (Ullmann, 2000), and an electrostatic model for calculating
reorganization energies (Sharp, 1998) in order to calculate charge transfer
energies in complex systems. This approach was successfully applied within
continuum-electrostatics models to describe electron transfer reactions in
complex systems such as photosynthetic center (Becker, Ullmann, &
Ullmann, 2007; Bombarda & Ullmann, 2011) which was recently reviewed
(Ullmann, Mueller, & Bombarda, 2016).
The second category, which will be discussed more extensively here, is
chemical reactions occurring in the enzyme active site that can be described
by transition state theory (TST). TST assumes that the free activation energy
ΔG6¼ can be derived from quasi-equilibrium between the reactant and tran-
sition states. Thus, intermediate structures along the reaction coordinate
(Fig. 3A) including the transition state structure are explicitly analyzed.
The reaction rate constant (Eq. 9) can be expressed by Eyring–Polanyi
equation:
 
kðT Þ ¼ ζβ1 h1 exp βΔG6¼ (12)

where h is the Planck constant and ζ in the transmission coefficient which

corresponds to the probability of being reactive once the transition state is
reached. Even though the transmission coefficient has to be considered in
general (i.e., for quantum tunneling effects), the lowering of ΔG6¼ is a dom-
inant effect of enzymes in catalysis (Gao et al., 2006) and majority of com-
putational enzymology studies focus on it. To get the rate-limiting energy
barrier of a chemical process, one has to investigate the reaction path
between substrate and product state. On the basis of the models described
in the previous section, potential energy of any conformation in a mole-
cular structure (here enzyme–substrate complex) can be calculated. When
searching for the reaction path, the 3N-dimensional energy function (with
N the number of atoms of the system) is portrayed as a potential energy sur-
face (PES). A PES can be viewed as a landscape with valleys and mountain
passes, which correspond to stable states and easiest transition pathways
between them. A reaction coordinate of an elementary step follows the path
96 Martin Culka et al.

High
Pro
Energy

Energy
In In

Sub
Pro
Sub
Low
Reaction coordinate
Fig. 4 Reaction path can be imagined as a trail through an energy landscape where
basins represent the stable states and the passes represent the transition states. The
path from substrate (Sub) to product (Pro) usually involves several intermediate (In)
states. The energy change plotted against the position on the path is referred to as reac-
tion profile.

from a minimum along the shallowest ascent toward the transition state and
from there along the deepest descent toward the product state. Most of the
chemical reactions are composed of several elementary steps forming
together the minimum energy path (MEP) from reactant to product (Fig. 4).
Many methods have been developed to identify reaction paths and to
distinguish among mechanism variants based on corresponding energy
profiles. In general, one has to first identify the mechanistic options and then
get corresponding accurate energy profiles. At the beginning of a computa-
tional kinetic study, optimization techniques are usually used to find minima
and saddle points on the PES, which together define MEPs. The obtained
energy profile corresponds to the enthalpic part of the free energy (Eq. 6). If
the enthalpic differences among mechanism variants are big, one can already
make qualitative conclusions about the preferred mechanism. The entropic
part of Eq. (6) has to be considered to obtain proper reaction free energy
profile at finite temperature in general, although the effect on the reaction
barrier height is not always that profound (Kazemi, Himo, & Åqvist, 2016).
The entropy-corrected PES is referred to as free energy surface at certain
temperature. To get a quantitative free energy surface, one needs to perform
extensive sampling along the reaction path, which can be extremely costly or
even prohibitive in complex structural models. Therefore, the sampling is
usually performed only when the most probable mechanism is identified
on PES.
In the next part, we first discuss methods to identify a reaction path on
PES. We then turn to the sampling methods that aim to get full free energy
reaction profiles. In the end, we return to the preexponential transmission
coefficient of Eq. (12) and discuss the role of nuclear quantum effects.
Computational Biochemistry—Enzyme Mechanisms Explored 97

3.3 Mechanic and Energetic Properties of Molecular

Reaction Paths
Computational investigation of an enzyme mechanism begins with a struc-
tural mechanistic model of the enzyme. As described earlier, stable experi-
mentally determined structures corresponding to minima on the PES are
commonly used as starting states. Approaches for searching reaction paths
can be categorized as single ended or double ended. Double-ended appro-
aches require the knowledge of two states, usually the initial state and the
final state, to search for mechanistic possibilities in between. In most cases,
however, one first explores the PES with a single-ended method and then
loads the first path estimate into a double-ended method for further
refinement.

3.3.1 Methods to Explore and Investigate Mechanistic Possibilities

3.3.1.1 Single-Ended Reaction Path Search Methods
The simplest single-ended approach is the adiabatic mapping of the PES.
Based on chemical intuition and experimental data, a set of movements that
lead toward a desired final state is applied. A function characterizing a desired
movement is called collective variable (CV). For instance, in a cysteine pro-
tease, the catalytic cysteine has to be deprotonated by a neighboring histidine
before it can attack the peptide bond of the substrate. Therefore, a first col-
lective variable to be investigated is the distance between the thiol hydrogen
of the cysteine and the closest nitrogen of the histidine. During adiabatic
mapping along such a CV, the distance is shrunk in discrete steps. The
enzyme structure is optimized at each step, while the CV is biased at given
position by a harmonic potential. The mapping finishes by reaching the
desired product state and the energies of the discrete scanning steps represent
a rough potential energy reaction profile. The energy maxima and minima
along this reaction profile should be further optimized without bias. Minima
can be optimized by standard gradient-based energy minimization routines
(Leach, 2001). Maxima can be refined to first-order saddle points using
the eigenvector following method (Cerjan, 1981) that requires calculation
of the Hessian matrix of second derivatives. The largest negative eigenvector
of the Hessian points in direction of the reaction path and thus can be
followed uphill to reach the saddle point. Since the saddle point is hard
to reach in one step, the Hessian has to be updated. Although various
approximate techniques exist (Munro & Wales, 1999; Wales & Walsh,
1998), the usage of purely single-ended eigenvector following for big sys-
tems with many degrees of freedom may be both unreliable and computa-
tionally unfeasible.
98 Martin Culka et al.

Alternatively, the single-ended growing string method (Zimmerman &

Bowman, 2016) can be used to locate the nearest saddle point and next stable
intermediate state. This method, like double-ended methods discussed fur-
ther below, generates a chain of structures to sample the reaction coordinate.
Unlike its double-ended sister, which uses both initial and final state struc-
tures for the chain generation, the single-ended growing string method uses
a set of collective variables to indicate the initial direction from the initial
state. Once, a saddle point estimate is located, the eigenvector following
method is used to refine it. Interestingly, the single-ended growing string
method can suggest the CVs automatically and thus explore various reaction
possibilities from a given initial state.
In contrast to the above described PES mapping approaches, the reac-
tion mechanism can be also explored at finite temperature without knowl-
edge of the final state structure. In principle, sufficiently long MD
simulations of the initial state at physiological temperature should lead
to a reaction. Current computer power is, however, not sufficient to per-
form such a long simulation, especially for models with quantum essence
that are required in most cases. Fortunately, MD simulations can be driven
along predefined CVs in a similar way as in adiabatic PES mapping. One
popular technique is called umbrella sampling (US) (Torrie & Valleau,
1977). A harmonic (umbrella) bias potential is applied in order to restrain
the MD simulation around a certain value of the CV. The value of the CV
can be changed in discrete steps, while a sufficiently long biased MD sim-
ulation is performed for each value till the desired final state is reached.
Another elegant method how to sample reaction possibilities at finite tem-
perature in accessible time is called metadynamics (Laio & Parrinello,
2002). In this method, Gaussian bias potential “hills” are added to the
PES in certain time intervals along the values of a CV that has been already
visited in the MD trajectory. In this fashion, the MD is discouraged to
revisit already sampled CV values, and thus the simulation proceeds
toward the desired final state. The bias introduced by metadynamics brings
bigger flexibility compared to the US bias on one hand, but potentially
slower convergence on the other.
Although the single-ended finite temperature methods provide reason-
able conformational freedom and entropic aspects in one simulation, their
use with higher QM models can be computationally unfeasible. Further-
more, while the choice of collective variables in stiff adiabatic PES mapping
is usually trivial, in more flexible bias MD methods it can be rather elusive.
Therefore, double-ended methods reviewed in the next paragraphs are often
Computational Biochemistry—Enzyme Mechanisms Explored 99

preferred. The entropic aspects can also be accessed a posteriori once the
PES path is established as we shall see below.

3.3.1.2 Double-Ended Reaction Path Search Methods

Once the final state structure is discovered by a single-ended method or
when it is known from an experimental structure or guessed by chemical
intuition, so-called double-ended methods can be employed to construct
a path between them. Alternatively, an estimated path derived, e.g., from
adiabatic PES mapping described above can be loaded in and refined.
The major double-ended method category that will be described here com-
prises of chain-of-states (COS) methods. A COS method discretizes the
reaction path between the initial and final state into a set of intermediate
structures and optimizes them in a connected manner. Once this procedure
achieves convergence, the resulting path should represent the minimal
energy pathway (MEP) between the initial and final state.
One popular COS approach is nudged elastic band (NEB) (Jonsson,
Mills, & Jacobsen, 1998). NEB approximates the reaction path by a set of
structures that are connected by springs into a chain. The parallel component
of the spring force keeps the images distributed along the whole path, while
the perpendicular component helps to push images to the MEP valley.
Because NEB does not directly seek for the saddle point structures, a
climbing image variant of NEB (CI-NEB) (Henkelman, Uberuaga, &
Jónsson, 2000) has been developed. The converged NEB path is usually
loaded into CI-NEB. The parallel component of the NEB spring potential
acting on the highest image of the path is inversed to drive it uphill, while the
perpendicular component is kept unchanged in order to keep the image on
the path. Once the path is converged again, the highest image should rep-
resent the saddle point structure.
The string method (SM) (E, Ren, & Vanden-Eijnden, 2002) uses a concept
on equal image distribution to cover the path instead of introducing spring
constants between the images. The images are equally redistributed along a
spline fit of the reaction path every optimization cycle. The path is con-
verged once the redistribution does not significantly change the image posi-
tion. An SM variant called growing string method (Peters, Heyden, Bell, &
Chakraborty, 2004) construct the chain of states gradually from the initial
state toward the final state and thus reduces the bias potentially introduced
by initial path guess (e.g., linear interpolation). Another SM variant
optimizes the COS at finite temperature, allowing for a better confor-
mational relaxation of the MEP (Vanden-Eijnden & Venturoli, 2009;
100 Martin Culka et al.

E, Ren, & Vanden-Eijnden, 2005). Note that the single-ended string method
described above is a variant of the growing string method with even less
initial bias.
The above described COS methods do not a priori search for the tran-
sition state structures. Although CI-NEB makes a step in this direction, it is
not actually guaranteed that it will reach a first-order saddle point. One can,
of course, employ the above described eigenvector following approaches to
refine the maxima of a COS path to saddle points. However, these methods
are both computationally costly and also unreliable once the input structure
is not close to the actual saddle point. An elegant solution is the conju-
gate peak refinement (CPR) (Fischer & Karplus, 1992; Gisdon, Culka, &
Ullmann, 2016) method that gradually constructs chain of states between
the initial and final states while it aims at locating the first-order saddle
points. In contrast to NEB or SM, the number of states along the path is
not fixed and thus the sampling in the saddle point region can be increased
to facilitate its proper location. The CPR method is based on the fact that in
the vicinity of a saddle point, there is one direction, which points to an ener-
getic maximum, while all others lead to a minimum. The CPR algorithm
picks the highest energy structure along the discretized path and performs a
line maximization along the corresponding tangential path vector. This
corrected maximum is then minimized in conjugate space similar to the con-
jugated gradient minimization method. By staying conjugate to the original
path, falling to the neighboring minimum is prevented. The optimized
structure is added into the chain of states. Gradually the saddle point is
approached by providing more sampling in the transition region and occa-
sionally the conjugate optimization procedure can converge to locate the
first-order saddle point. A successful run ends, when all maxima along the
reaction path between the initial and final state are identified and optimized
to first-order saddle points.

3.3.1.3 Path-Based Reaction Path Search Methods

Once a reaction path is found, it is sometimes necessary to correct it for
conformational flexibility, since a biological system has many possible
degrees of freedom. Path-based methods include the conformational flex-
ibility to evaluate alternative reaction paths or a set of possible reactive coor-
dinates. These methods usually require an initial path, which does not
necessarily have to be properly refined. Any path generated by double-
ended or single-ended methods can be used. The concept of metadynamics,
introduced as a single-ended search method, can also be applied for
Computational Biochemistry—Enzyme Mechanisms Explored 101

path-based approaches. Instead of using a trivial CV, such as bond distances

or dihedral angles, a function that describes the whole reaction path is used
(Branduardi, Gervasio, & Parrinello, 2007; Bernardi, Melo, & Schulten,
2015). The application of the path CV allows to dynamically sample the
conformational space based on the input path estimate. This provides infor-
mation on the conformational flexibility of the input intermediates and
transition states, which is often crucial for reaction paths obtained from
static approaches, such as adiabatic mapping, NEB, or CPR. If the distance
from the initial path is taken as a second CV to be biased, even alternative
paths can be identified and energetically evaluated.
A conceptually different approach is followed in transition path sampling
(TPS) (Bolhuis, Dellago, & Chandler, 1998; Dellago, Bolhuis, Csajka, &
Chandler, 1998). The aim of TPS is to connect two stable states by a col-
lection of all likely transition pathways, which represent the transition path
ensemble. As mentioned above, an input trajectory does not necessarily have
to be properly refined, thus it may have a low weight in the transition path
ensemble. Therefore, for a TPS simulation the initial trajectory has to be
evaluated and equilibrated toward a representative transition path trajectory.
Unlike US ensembles or metadynamics ensembles, a TPS simulation creates
unbiased trajectories, since it does not enhance the sampling of rare events by
bias potentials (Swenson & Bolhuis, 2014). TPS, however, is also dependent
on a collective variable, yet it is not used to drive the simulation, but rather to
discriminate the stable states of the system, and to monitor the progress along
the trajectory. This makes the CV a crucial quantity in a TPS simulation.
The enhancement of sampling rare events in TPS is achieved by importance
sampling of trajectory space. All dynamic movements are preformed by a
Monte Carlo approach, while a new trajectory is created from an existing
one in the ensemble. For that, specific procedures are available to drive
movements in trajectory space (Bolhuis, Chandler, Dellago, & Geissler,
2002; Rowley & Woo, 2009). TPS has been successfully applied to simulate
biological systems (Dellago & Bolhuis, 2007) and is also recently used to
study enzyme catalysis, such as the hydride transfer in a dihydrofolate reduc-
tase (Wang, Antoniou, Schwartz, & Schramm, 2016).

3.3.2 Methods to Obtain Free Energy

So far, we have been concentrating on the qualitative aspects of the reaction
mechanism. Now we turn to the methods to get the free energy profile esti-
mates. As noted earlier (Eq. 6), the free energy is composed of the potential
energy (enthalpic) part that is calculated directly by the model, and the
102 Martin Culka et al.

entropic part, for which extensive sampling at finite temperature along the
reaction path is needed. In case of US and metadynamics the sampling is
already included in the path search procedure, so entropic influence is
explicitly included and thus free energy profile can be directly estimated.
For US, weighted histogram analysis method (WHAM) (Kumar,
Rosenberg, Bouzida, Swendsen, & Kollman, 1992) can be used to remove
the umbrella bias and to integrate the simulation windows into a free energy
profile. In case of metadynamics, the sum of added Gaussian hills plotted
against the collective variable directly represents the free energy profile.
TPS collected pathways are dynamic trajectories, thus kinetic information,
such as rate constants, can be extracted (Dellago, 2007). But since only tra-
jectories are considered that connect certain regions in configuration space,
the configurations are not distributed according to the equilibrium distribu-
tion of the system. To determine free energy profiles, one has to obtain
equilibrium-distributed configurations. One possibility is to apply biasing
procedures, such as US variants, to rarely visited states, and divide the
CV into overlapping windows (Dellago, 2007).
In case of static methods such as adiabatic PES mapping or chain-of-states
methods, the entropic aspect has to be added by sampling method (MD or
Monte Carlo) to account for entropic part of the free energy profile. One
option is to use these static paths as an input to above described path-based
approaches. A possible obstacle can be that achieving extensive sampling
when high-level QM methods are used to treat the active site may be com-
putationally prohibitive. In the same time, usage of, e.g., SE QM models can
introduce additional bias into the results. A way out of this dilemma offers
the free energy perturbation (FEP) methods. For instance, one can find the
reaction path by a chain-of-states method within QM/MM model (K€astner,
Senn, Thiel, Otte, & Thiel, 2006). Subsequently, an MD simulation is per-
formed for every state of the PES reaction path with QM region kept frozen.
Perturbation energy is calculated as energy for moving one step forward in
the PES QM reaction path while staying in the same MM conformational
ensemble:

ΔEpert ¼ EQM=MM ðriQM , rMM Þ  EQM=MM ðriQM , riMM Þ

+1 i
(13)

where i and i + 1 are the indices of adjacent path steps and r are the coor-
dinate vectors. Free energy for the i ! i + 1 step is then calculated by
averaging over the MM ensemble at step i using Zwanzig equation
(Zwanzig, 1954):
Computational Biochemistry—Enzyme Mechanisms Explored 103

1
ΔGi!i + 1 ¼  ln h exp ðβΔEpert Þii (14)
β

The reaction free energy profile is calculated by adding QM energy

corrected for zero-temperature vibrations to the perturbation-derived free
energy. Another example where FEP concept is exploited are the FEP/
US studies in EVB model (Warshel, 1991).

3.3.3 Methods to Simulate Nuclear Quantum Effects

The majority of methods commonly used to locate and sample the enzyme
reaction rely on atomistic models that treat the nuclei as classical objects.
Even in the common QM methods, just the electrons are treated as quantum
objects. This approximation is necessary to make the calculations on bigger
systems computationally feasible and in many cases also justified, since the
heavy nuclei hardly show any quantum behavior. The quantum nature
becomes in practice significant in case of proton transfer processes, which
are often part of enzyme-catalyzed reaction coordinates. If a proton transfer
is the rate-limiting step, the nuclear quantum-mechanical effects (NQM)
play important role in the overall reaction rate. NQM effect are most pro-
nounced when comparing protium and deuterium variants and calculating
kinetic isotope effects.
Although small number of protons can be treated as quantum particles
using nuclear-electronic orbital (Webb, Iordanov, & Hammes-Schiffer,
2002) methods, most of the approaches that deal with NQM in enzymes
can be classified as correction methods that are applied, e.g., on US path
ensembles. One direction is the ensemble-averaged variational transition state
theory with multidimensional tunneling (EA-VTST/MT) (Garcia-Viloca,
Alhambra, Truhlar, & Gao, 2001; Truhlar et al., 2002) that corrects the free
energy barrier ΔG6¼ for quantum-mechanical vibrations. EA-VTST/MT in
addition also calculates the transmission coefficient ζ (see Eq. 12) that corrects
mainly for quantum-mechanical tunneling through the free energy barrier.
Another family of approaches to NQM is based on Feynman path integrals.
Quantum nature of the nuclei is approximated by transforming the classical
spheres into rings of quasiparticles connected by springs. Path integral MD
simulation techniques include quantized classical path (QCP) (Hwang &
Warshel, 1993), centroid molecular dynamics (CMD) (Cao & Voth,
1994a, 1994b), or ring-polymer molecular dynamics (RPMD) (Braams &
Manolopoulos, 2006; Craig & Manolopoulos, 2004).
104 Martin Culka et al.

3.4 Path Search Strategy

Taken together, enormous amount of methods to locate reaction path and
calculate its energy profile has been developed. It is not easy to objectively
find the one and only correct strategy, although different groups certainly
have their preferred approaches. The choice of the method is also influenced
by the primary question that the researcher is asking. If the task is to find the
correct catalytic mechanism, many potential variants of the reaction mech-
anism need to be tested in a reasonable time. If the individual steps of the
mechanism are already known, the task might be to get proper rate constant
to relate the model to experimental parameters. In many cases, big energy
barriers on PES can rule the unfeasible mechanism variants out and find
the most promising set of reaction steps in fraction of time in more accurate
QM models compared to direct usage of sampling approaches in approxi-
mate models. The PES path can be further optimized and corrected by much
more demanding sampling methods, or a reaction-specific SE potential
(e.g., EVB) can be constructed based on the previous PES investigation
in high-level QM models.
The kinetic and thermodynamic parameters determined by the afore-
described methods should be combined in order to get a more reliable pic-
ture. The different methods should not be viewed as competitive approaches
but rather as complementary and one should search for synergy among dif-
ferent methods with limitations of the models in mind. The limitations and
synergy should also be considered when comparing computational results
with experimental data.

4. GOING BEYOND THE EXPLORATION OF THE

REACTION PATHS
In order to understand the biological systems, it is not enough to
understand the mechanism of an enzyme. It is required to analyze the
enzyme in its physiological context and how the rate of catalysis is influenced
by parameters such as pH or metabolite concentration. Today’s systems biol-
ogy is using kinetic models with stretched exponential or noninteger stoi-
chiometry in order to describe complex metabolic networks. Even if such
models are of certain practical use to solve some research problems, they
are not satisfactory from a theoretical point of view, since mass and energy
conservation is not guaranteed. Other models are taking kinetic parameters
from databases. However, such parameters were usually determined under
Computational Biochemistry—Enzyme Mechanisms Explored 105

specific circumstances and cannot account for all possible effects. Consider-
ing the constantly increasing number of complete genomes and partially
reconstructed metabolisms, it comes more and more important to get a more
realistic view of the metabolic reaction in its context. The challenge for
computational biochemistry today and in the future is to derive enzymatic
parameters from structure models by using methods that we reviewed in this
article. However, it will be required to go beyond such information. The
kinetic parameters can be combined in master equation approaches
(Becker et al., 2007; Bombarda & Ullmann, 2011) or kinetic Monte Carlo
simulations (Till, Becker, Essigke, & Ullmann, 2008) in order to simulate
complete catalytic cycles that are influenced by environmental parameters
such as pH or membrane potential. To carry the approach further, it might
be possible to model the whole cellular context in reaction–diffusion equa-
tions, which may allow in the future to model complex biochemical reac-
tions. Combining structural biology and systems biology may thus be a
promising direction, especially considering the pace in which both fields
make progress in recent years. With the help of computer models that rely
on a solid experimental basis, we may more and more understand how the
jigglings and wigglings of atoms leads to the complex phenomenon we
call life.

REFERENCES
Amara, P., Field, M. J., Alhambra, C., & Gao, J. (2000). The generalized hybrid orbital
method for combined quantum mechanical/molecular mechanical calculations: Formu-
lation and tests of the analytical derivatives. Theoretical Chemistry Accounts, 104, 336–343.
Åqvist, J., & Warshel, A. (1993). Simulation of enzyme reactions using valence bond force
fields and other hybrid quantum/classical approaches. Chemical Reviews, 93, 2523–2544.
Baker, N. A., Sept, D., Joseph, S., Holst, M. J., & McCammon, J. A. (2001). Electrostatics of
nanosystems: Application to microtubules and the ribosome. Proceedings of the National
Academy of Sciences, 98, 10037–10041.
Bakowies, D., & Thiel, W. (1991). MNDO study of large carbon clusters. Journal of the
American Chemical Society, 113, 3704–3714.
Bartesaghi, A., Merk, A., Banerjee, S., Matthies, D., Wu, X., Milne, J. L. S., &
Subramaniam, S. (2015). 2.2 Å resolution cryo-EM structure of Î2-galactosidase in com-
plex with a cell-permeant inhibitor. Science, 348, 1147–1151.
Becke, A. (1993). Density functional thermochemistry. III. The role of exact exchange. The
Journal of Chemical Physics, 98, 5648–5652.
Becker, T., Ullmann, R. T., & Ullmann, G. M. (2007). Simulation of the electron transfer
between the tetraheme subunit and the special pair of the photosynthetic reaction center
using a microstate description. The Journal of Physical Chemistry. B, 111, 2957–2968.
Bernardi, R. C., Melo, M. C. R., & Schulten, K. (2015). Enhanced sampling techniques in
molecular dynamics simulations of biological systems. Biochimica et Biophysica Acta, 1850,
872–877.
106 Martin Culka et al.

Bolhuis, P. G., Chandler, D., Dellago, C., & Geissler, P. L. (2002). Transition path sampling:
Throwing ropes over rough mountain passes, in the dark. Annual Review of Physical
Chemistry, 53, 291–318.
Bolhuis, P. G., Dellago, C., & Chandler, D. (1998). Sampling ensembles of deterministic
transition pathways. Faraday Discussions, 110, 421–436.
Bombarda, E., Becker, T., & Ullmann, G. M. (2006). The influence of the membrane poten-
tial on the protonation of bacteriorhodopsin: Insights from electrostatic calculations into
the regulation of proton pumping. Journal of the American Chemical Society, 128,
12129–12139.
Bombarda, E., & Ullmann, G. M. (2011). Continuum electrostatic investigations of charge
transfer processes in biological molecules using a microstate description. Faraday
Discussions, 148, 173–193.
Braams, B. J., & Manolopoulos, D. E. (2006). On the short-time limit of ring polymer molec-
ular dynamics. The Journal of Chemical Physics, 125, 124105.
Branduardi, D., Gervasio, F. L., & Parrinello, M. (2007). From A to B in free energy space.
The Journal of Chemical Physics, 126, 054103. http://dx.doi.org/10.1063/1.2432340.
Bryantsev, V. S., Diallo, M. S., Van Duin, A. C. T., & Goddard, W. A. (2009). Evaluation of
B3LYP, X3LYP, and M06-class density functionals for predicting the binding energies of
neutral, protonated, and deprotonated water clusters. Journal of Chemical Theory and
Computation, 5, 1016–1026.
Calimet, N., & Ullmann, G. M. (2004). The influence of a transmembrane pH gradient on
protonation probabilities of bacteriorhodopsin: The structural basis of the back-pressure
effect. Journal of Molecular Biology, 339, 571–589.
Cao, J., & Voth, G. A. (1994a). The formulation of quantum statistical mechanics based on
the Feynman path centroid density. I. Equilibrium properties. The Journal of Chemical
Physics, 100, 5093–5105.
Cao, J., & Voth, G. A. (1994b). The formulation of quantum statistical mechanics based on
the Feynman path centroid density. II. Dynamical properties. The Journal of Chemical
Physics, 100, 5106–5117.
Cerjan, C. J. (1981). On finding transition states. The Journal of Chemical Physics, 75, 2800.
Chen, J. L., Noodleman, L., Case, D., & Bashford, D. (1994). Incorporating solvation effects
into density functional electronic structure calculations. The Journal of Physical Chemistry,
98, 11059–11068.

Cı́žek, J. (1966). On the correlation problem in atomic and molecular systems. Calculation of
wavefunction components in Ursell-type expansion using quantum-field theoretical
methods. The Journal of Chemical Physics, 45, 4256–4266.
Cole, D. J., & Hine, N. D. M. (2016). Applications of large-scale density functional theory in
biology. Journal of Physics. Condensed Matter, 28, 393001.
Cornell, W. D., Cieplak, P., Bayly, C. I., Gould, I. R., Merz, K. M., Ferguson, D. M., …
Kollman, P. A. (1996). A second generation force field for the simulation of proteins,
nucleic acids, and organic molecules. Journal of the American Chemical Society, 118,
2309–2309.
Costa, M. G. S., Batista, P. R., Bisch, P. M., & Perahia, D. (2015). Exploring free energy
landscapes of large conformational changes: Molecular dynamics with excited normal
modes. Journal of Chemical Theory and Computation, 11, 2755–2767.
Craig, I. R., & Manolopoulos, D. E. (2004). Quantum statistics and classical mechanics: Real
time correlation functions from ring polymer molecular dynamics. The Journal of Chemical
Physics, 121, 3368–3373.
Cramer, C. J. (2004). Essentials of computational chemistry: Theories and models (2nd ed.).
Chichester: Wiley.
Dellago, C. (2007). Transition path sampling and the calculation of free energies. In C. Chipot
& A. Pohorille (Eds.), Free energy calculations (pp. 249–276). Berlin, Heidelberg: Springer.
Computational Biochemistry—Enzyme Mechanisms Explored 107

Dellago, C., & Bolhuis, P. G. (2007). Transition path sampling simulations of biological sys-
tems. In M. Reiher (Ed.), Atomistic approaches in modern biology (Vol. 268, pp. 291–317).
Berlin, Heidelberg: Springer.
Dellago, C., Bolhuis, P. G., Csajka, F. S., & Chandler, D. (1998). Transition path sampling
and the calculation of rate constants. The Journal of Chemical Physics, 108, 1964.
E, W., Ren, W., & Vanden-Eijnden, E. (2002). String method for the study of rare events.
Physical Review B, 66, 052301.
E, W., Ren, W., & Vanden-Eijnden, E. (2005). Finite temperature string method for the
study of rare events. The Journal of Physical Chemistry B, 109, 6688–6693.
Elstner, M. (2006). The SCC-DFTB method and its application to biological systems. The-
oretical Chemistry Accounts, 116, 316–325.
Feliks, M., & Ullmann, G. M. (2012). Glycerol dehydratation by the B12-independent
enzyme may not involve the migration of a hydroxyl group: A computational study.
The Journal of Physical Chemistry. B, 116, 7076–7087.
Feynman, R. (1964). The Feynman lectures on physics. New York, USA: Basic Books.
Retrieved from http://www.feynmanlectures.caltech.edu/.
Field, M. J. (2007). A practical introduction to the simulation of molecular systems (2nd ed.).
Cambridge, UK: Cambridge University Press.
Field, M. J., Bash, P. A., & Karplus, M. (1990). A combined quantum mechanical and molec-
ular mechanical potential for molecular dynamics simulations. Journal of Computational
Chemistry, 11, 700–733.
Fischer, S., & Karplus, M. (1992). Conjugate peak refinement: An algorithm for finding reac-
tion paths and accurate transition states in systems with many degrees of freedom. Chem-
ical Physics Letters, 194, 252–261.
Gamow, G. (1928). Zur quantentheorie des atomkernes. Zeitschrift fr Physiotherapie, 51,
204–212.
Gao, J., Ma, S., Major, D. T., Nam, K., Pu, J., & Truhlar, D. G. (2006). Mechanisms and free
energies of enzymatic reactions. Chemical Reviews, 106, 3188–3209.
Garcia-Viloca, M., Alhambra, C., Truhlar, D. G., & Gao, J. (2001). Inclusion of quantum-
mechanical vibrational energy in reactive potentials of mean force. The Journal of Chemical
Physics, 114, 9953–9958.
Georgieva, P., & Himo, F. (2010). Quantum chemical modeling of enzymatic reactions: The
case of histone lysine methyltransferase. Journal of Computational Chemistry, 31,
1707–1714.
Gilson, M. K., & Zhou, H.-X. (2007). Calculation of protein-ligand binding affinities.
Annual Review of Biophysics and Biomolecular Structure, 36, 21–42.
Gisdon, F. J., Culka, M., & Ullmann, G. M. (2016). PyCPR—A python-based implemen-
tation of the Conjugate Peak Refinement (CPR) algorithm for finding transition state
structures. Journal of Molecular Modeling, 22, 242.
Golkaram, M., Shin, Y. K., & van Duin, A. C. T. (2014). Reactive molecular dynamics study
of the pH-dependent dynamic structure of α-helix. The Journal of Physical Chemistry. B,
118, 13498–13504.
Gray, H. B., & Winkler, J. R. (2005). Long-range electron transfer. Proceedings of the National
Academy of Sciences, 102, 3534–3539.
Head-Gordon, M., Pople, J. A., & Frisch, M. J. (1988). MP2 energy evaluation by direct
methods. Chemical Physics Letters, 153, 503–506.
Henkelman, G., Uberuaga, B. P., & Jónsson, H. (2000). A climbing image nudged elastic
band method for finding saddle points and minimum energy paths. The Journal of Chemical
Physics, 113, 9901.
Heyden, A., & Truhlar, D. G. (2008). Conservative algorithm for an adaptive change of res-
olution in mixed atomistic/coarse-grained multiscale simulations. Journal of Chemical
Theory and Computation, 4, 217–221.
108 Martin Culka et al.

Honig, B., & Nicholls, A. (1995). Classical electrostatics in biology and chemistry. Science,
268, 1144–1149.
Hwang, J. K., & Warshel, A. (1993). A quantized classical path approach for calculations of
quantum mechanical rate constants. The Journal of Physical Chemistry, 97, 10053–10058.
Ingólfsson, H. I., Lopez, C. A., Uusitalo, J. J., de Jong, D. H., Gopal, S. M., Periole, X., &
Marrink, S. J. (2014). The power of coarse graining in biomolecular simulations. Wiley
Interdisciplinary Reviews: Computational Molecular Science, 4, 225–248.
Jonsson, H., Mills, G., & Jacobsen, K. W. (1998). Nudged elastic band method for finding
minimum energy paths of transitions. In B. J. Berne, G. Ciccotti, & D. F. Coker (Eds.),
Classical and quantum dynamics in condensed phase simulations (pp. 385–404). Singapore:
World Scientific.
Jorgensen, W. L., Maxwell, D. S., & Tirado-Rives, J. (1996). Development and testing of the
OPLS all-atom force field on conformational energetics and properties of organic liquids.
Journal of the American Chemical Society, 118, 11225–11236.
Jorgensen, W. L., & Tirado-Rives, J. (1996). Monte Carlo vs molecular dynamics for con-
formational sampling. The Journal of Physical Chemistry, 100, 14508–14513.
Kamerlin, S. C. L., & Warshel, A. (2010). The EVB as a quantitative tool for formulating
simulations and analyzing biological and chemical reactions. Faraday Discussions, 145,
71–106.
Kamerlin, S. C. L., & Warshel, A. (2011). The empirical valence bond model: Theory and
applications. Wiley Interdisciplinary Reviews: Computational Molecular Science, 1, 30–45.
Kar, P., & Feig, M. (2014). Recent advances in transferable coarse-grained modeling of pro-
teins. Advances in Protein Chemistry and Structural Biology, 96, 143–180.
K€astner, J., Senn, H. M., Thiel, S., Otte, N., & Thiel, W. (2006). QM/MM free-energy
perturbation compared to thermodynamic integration and umbrella sampling: Applica-
tion to an enzymatic reaction. Journal of Chemical Theory and Computation, 2, 452–461.
Kazemi, M., Himo, F., & Åqvist, J. (2016). Enzyme catalysis by entropy without Circe effect.
Proceedings of the National Academy of Sciences of the United States of America, 113,
2406–2411.
Klamt, A., & Sch€ €rmann, G. (1993). COSMO: A new approach to dielectric screening in
uu
solvents with explicit expressions for the screening energy and its gradient. Journal of the
Chemical Society, Perkin Transactions 2 (799–805).
Kmiecik, S., Gront, D., Kolinski, M., Wieteska, L., Dawid, A. E., & Kolinski, A. (2016).
Coarse-grained protein models and their applications. Chemical Reviews, 116,
7898–7936.
K€uhlbrandt, W. (2014). Cryo-EM enters a new era. eLife, 3, e03678.
Kulik, H. J., Luehr, N., Ufimtsev, I. S., & Martı́nez, T. J. (2012). Ab initio quantum chem-
istry for protein structures. The Journal of Physical Chemistry. B, 116, 12501–12509.
Kumar, S., Rosenberg, J. M., Bouzida, D., Swendsen, R. H., & Kollman, P. A. (1992). The
weighted histogram analysis method for free-energy calculations on biomolecules. I. The
method. Journal of Computational Chemistry, 13, 1011–1021.
Laio, A., & Parrinello, M. (2002). Escaping free-energy minima. Proceedings of the National
Academy of Sciences of the United States of America, 99, 12562–12566.
Leach, A. R. (2001). Molecular modelling: Principles and applications (2nd ed.). New Jersey:
Prentice Hall.
Levitt, M., & Warshel, A. (1975). Computer simulation of protein folding. Nature, 253,
694–698.
Li, J., Nelson, M. R., Peng, C. Y., Bashford, D., & Noodleman, L. (1998). Incorporating
protein environments in density functional theory: A self-consistent reaction field calcu-
lation of redox potentials of [2Fe2S] clusters in ferredoxin and phthalate dioxygenase
reductase. The Journal of Physical Chemistry. B, 102, 6311–6324.
Computational Biochemistry—Enzyme Mechanisms Explored 109

Li, J., & Ryde, U. (2014). Comparison of the active-site design of molybdenum oxo-transfer
enzymes by quantum mechanical calculations. Inorganic Chemistry, 53, 11913–11924.
Liang, T., Shin, Y. K., Cheng, Y.-T., Yilmaz, D. E., Vishnu, K. G., Verners, O., … van
Duin, A. C. (2013). Reactive potentials for advanced atomistic simulations. Annual
Review of Materials Research, 43, 109–129.
Liu, T., Han, W.-G., Himo, F., Ullmann, G. M., Bashford, D., Toutchkine, A., …
Noodleman, L. (2004). Density functional vertical self-consistent reaction field theory
for solvatochromism studies of solvent-sensitive dyes. The Journal of Physical Chemistry.
B, 108, 11157–11169.
MacKerell, A. D., Bashford, D., Bellott, M., Dunbrack, R. L., Evanseck, J. D., Field, M. J.,
… Karplus, M. (1998). All-atom empirical potential for molecular modeling and dynam-
ics studies of proteins. The Journal of Physical Chemistry. B, 102, 3586–3616.
Manta, B., Raushel, F. M., & Himo, F. (2014). Reaction mechanism of zinc-dependent
cytosine deaminase from Escherichia coli: A quantum-chemical study. The Journal of Phys-
ical Chemistry. B, 118, 5644–5652.
Marcus, R., & Sutin, N. (1985). Electron transfers in chemistry and biology. Biochimica et
Biophysica Acta, Reviews on Bioenergetics, 811, 265–322.
Maseras, F., & Morokuma, K. (1995). IMOMM: A new integrated ab initio + molecular
mechanics geometry optimization scheme of equilibrium structures and transition states.
Journal of Computational Chemistry, 16, 1170–1179.
Meier, K., Choutko, A., Dolenc, J., Eichenberger, A. P., Riniker, S., & Van
Gunsteren, W. F. (2013). Multi-resolution simulation of biomolecular systems:
A review of methodological issues. Angewandte Chemie, International Edition, 52,
2820–2834.
Møller, C., & Plesset, M. S. (1934). Note on an approximation treatment for many-electron
systems. Physics Review, 46, 618–622.
Monti, S., Corozzi, A., Fristrup, P., Joshi, K. L., Shin, Y. K., Oelschlaeger, P., …
Bourne, P. E. (2013). Exploring the conformational and reactive dynamics of biomol-
ecules in solution using an extended version of the glycine reactive force field. Physical
Chemistry Chemical Physics, 15, 15062.
Moser, C. C., Keske, J. M., Warncke, K., Farid, R. S., & Dutton, P. L. (1992). Nature of
biological electron transfer. Nature, 355, 796–802.
Munro, L. J., & Wales, D. J. (1999). Defect migration in crystalline silicon. Physical Review B,
59, 3969–3980.
Nango, E., Royant, A., Kubo, M., Nakane, T., Wickstrand, C., Kimura, T., … Iwata, S.
(2016). A three-dimensional movie of structural changes in bacteriorhodopsin. Science,
354, 1552–1557.
Noid, W. G. (2013). Perspective: Coarse-grained models for biomolecular systems. The Jour-
nal of Chemical Physics, 139, 090901.
Oostenbrink, C., Villa, A., Mark, A. E., & Van Gunsteren, W. F. (2004). A biomolecular
force field based on the free enthalpy of hydration and solvation: The GROMOS
force-field parameter sets 53A5 and 53A6. Journal of Computational Chemistry, 25,
1656–1676.
Page, C. C., Moser, C. C., Chen, X., & Dutton, P. L. (1999). Natural engineering principles
of electron tunneling in biological oxidation-reduction. Nature, 402, 47–52.
Pande, K., Hutchison, C. D. M., Groenhof, G., Aquila, A., Robinson, J. S., Tenboer, J., …
Schmidt, M. (2016). Femtosecond structural dynamics drives the trans/cis isomerization
in photoactive yellow protein. Science, 352, 725–729.
Paquet, E., & Viktor, H. L. (2015). Molecular dynamics, Monte Carlo simulations, and
Langevin dynamics: A computational review. BioMed Research International, 2015,
183918.
110 Martin Culka et al.

Peters, B., Heyden, A., Bell, A. T., & Chakraborty, A. (2004). A growing string method for
determining transition states: Comparison to the nudged elastic band and string methods.
The Journal of Chemical Physics, 120, 7877–7886.
Riniker, S., Allison, J. R., & van Gunsteren, W. F. (2012). On developing coarse-grained
models for biomolecular simulation: A review. Physical Chemistry Chemical Physics, 14,
12423.
Rowley, C. N., & Woo, T. K. (2009). New shooting algorithms for transition path sampling:
Centering moves and varied-perturbation sizes for improved sampling. The Journal of
Chemical Physics, 131, 234102.
R€ucker, P., Wieninger, S. A., Ullmann, G. M., & Sticht, H. (2012). pH-dependent molec-
ular dynamics of vesicular stomatitis virus glycoprotein G. Proteins, 80, 2601–2613.
Saunders, M. G., & Voth, G. A. (2013). Coarse-graining methods for computational biology.
Annual Review of Biophysics, 42, 73–93.
Senftle, T. P., Hong, S., Islam, M. M., Kylasa, S. B., Zheng, Y., Shin, Y. K., … van
Duin, A. C. T. (2016). The ReaxFF reactive force-field: Development, applications
and future directions. Computational Materials, 2, 15011.
Senn, H. M., & Thiel, W. (2007). QM/MM methods for biological systems. In M. Reiher
(Ed.), Atomistic approaches in modern biology (Vol. 268, pp. 173–290). Berlin, Heidelberg:
Springer.
Senn, H. M., & Thiel, W. (2009). QM/MM methods for biomolecular systems. Angewandte
Chemie, International Edition, 48, 1198–1229.
Sharp, K. E. (1998). Calculation of electron transfer reorganization energies using the Finite
Difference Poisson-Boltzmann model. Biophysical Journal, 73, 1241–1250.
Shen, L., & Hu, H. (2014). Resolution-adapted all-atomic and coarse-grained model for bio-
molecular simulations. Journal of Chemical Theory and Computation, 10, 2528–2536.
Shurki, A., Derat, E., Barrozo, A., & Kamerlin, S. C. L. (2015). How valence bond theory
can help you understand your (bio) chemical reaction. Chemical Society Reviews, 44,
1037–1052.
Singh, U. C., & Kollman, P. A. (1986). A combined ab initio quantum mechanical and
molecular mechanical method for carrying out simulations on complex molecular sys-
tems: Applications to the CH3Cl + Cl exchange reaction and gas phase protonation
of polyethers. Journal of Computational Chemistry, 7, 718–730.
Sokkar, P., Boulanger, E., Thiel, W., & Sanchez-Garcia, E. (2015). Hybrid quantum
mechanics/molecular mechanics/coarse grained modeling: A triple-resolution approach
for biomolecular systems. Journal of Chemical Theory and Computation, 11, 1809–1818.
Stewart, J. J. P. (2013). Optimization of parameters for semiempirical methods VI: More
modifications to the NDDO approximations and re-optimization of parameters. Journal
of Molecular Modeling, 19, 1–32.
Svensson, M., Humbel, S., Froese, R. D. J., Matsubara, T., Sieber, S., & Morokuma, K.
(1996). ONIOM: A multilayered integrated MO + MM method for geometry
optimizations and single point energy predictions. A test for Diels–Alder reactions
and Pt(P(t-Bu)3)2 + H2 oxidative addition. The Journal of Physical Chemistry, 100,
19357–19363.
Swenson, D. W. H., & Bolhuis, P. G. (2014). A replica exchange transition interface sam-
pling method with multiple interface sets for investigating networks of rare events. The
Journal of Chemical Physics, 141, 044101.
Thery, V., Rinaldi, D., Rivail, J.-L., Maigret, B., & Ferenczy, G. G. (1994). Quantum
mechanical computations on very large molecular systems: The local self-consistent field
method. Journal of Computational Chemistry, 15, 269–282.
Thompson, M. A., & Schenter, G. K. (1995). Excited ates of the bacteriochlorophyll b dimer
of Rhodopseudomonas viridis: A QM/MM study of the photosynthetic reaction center
that includes MM polarization. The Journal of Physical Chemistry, 99, 6374–6386.
Computational Biochemistry—Enzyme Mechanisms Explored 111

Till, M. S., Becker, T., Essigke, T., & Ullmann, G. M. (2008). Simulating the proton transfer
in Gramicidin A by a sequential dynamical Monte Carlo method. The Journal of Physical
Chemistry. B, 112, 13401–13410.
Todorovic, M., Bowler, D. R., Gillan, M. J., & Miyazaki, T. (2013). Density-functional the-
ory study of gramicidin A ion channel geometry and electronic properties. Journal of The
Royal Society Interface, 10, 20130547.
Torrie, G., & Valleau, J. (1977). Nonphysical sampling distributions in Monte Carlo free-
energy estimation: Umbrella sampling. Journal of Computational Chemistry, 23, 187–199.
Tozzini, V. (2005). Coarse-grained models for proteins. Current Opinion in Structural Biology,
15, 144–150.
Truhlar, D. G., Gao, J., Alhambra, C., Garcia-Viloca, M., Corchado, J., Sánchez, M. L., &
Villà, J. (2002). The incorporation of quantum effects in enzyme kinetics modeling.
Accounts of Chemical Research, 35, 341–349.
Ullmann, G. M. (2000). The coupling of protonation and reduction in proteins with multiple
redox centers: Theory, computational method, and application to cytochrome c3. The
Journal of Physical Chemistry. B, 104, 6293–6301.
Ullmann, G. M., & Bombarda, E. (2014). Continuum electrostatic analysis of proteins.
In G. Naray-Szabo (Ed.), Protein modelling (pp. 135–163). Cham, Switzerland:
Springer International Publishing.
Ullmann, G. M., Kloppmann, E., Essigke, T., Krammer, E.-M., Klingen, A. R., Becker, T.,
& Bombarda, E. (2008). Investigating the mechanisms of photosynthetic proteins using
continuum electrostatics. Photosynthesis Research, 97, 33–53.
Ullmann, G. M., Mueller, L., & Bombarda, E. (2016). Theoretical analysis of
electron transfer in proteins: From simple proteins to complex machineries. In
W. Cramer (Ed.), Cytochrome complexes: Evolution, structures, energy transduction, and
signaling (Vol. 37, pp. 99–127). Dordrecht, Netherlands: Springer International
Publishing.
Ullmann, R. T., & Ullmann, G. M. (2012). GMCT: A Monte Carlo simulation package for
macromolecular receptors. Journal of Computational Chemistry, 33, 887–900.
Vanden-Eijnden, E., & Venturoli, M. (2009). Revisiting the finite temperature string
method for the calculation of reaction tubes and free energies. The Journal of Chemical
Physics, 130, 194103.
van Duin, A. C. T., Dasgupta, S., Lorant, F., & Goddard, W. A., III. (2001). ReaxFF:
A reactive force field for hydrocarbons. The Journal of Physical Chemistry. B, 105,
9396–9409.
van Mourik, T., B€ uhl, M., & Gaigeot, M.-P. (2014). Density functional theory across chem-
istry, physics and biology. Philosophical Transactions. Series A, Mathematical, Physical, and
Engineering Sciences, 372, 20120488.
Verlackt, C. C. W., Neyts, E. C., Jacob, T., Fantauzzi, D., Golkaram, M., Shin, Y.-K., …
Bogaerts, A. (2015). Atomic-scale insight into the interactions between hydroxyl radicals
and DNA in solution using the ReaxFF reactive force field. New Journal of Physics, 17,
103005.
Řezáč, J., & Hobza, P. (2012). Advanced corrections of hydrogen bonding and dispersion for
semiempirical quantum mechanical methods. Journal of Chemical Theory and Computation,
8, 141–151.
Wales, D. J., & Walsh, T. R. (1998). Theoretical study of the water pentamer. American Insti-
tute of Physics, 105, 6957–6971.
Wang, Z., Antoniou, D., Schwartz, S. D., & Schramm, V. L. (2016). Hydride transfer in
DHFR by transition path sampling, kinetic isotope effects, and heavy enzyme studies.
Biochemistry, 55, 157–166.
Warshel, A. (1991). Computer modeling of chemical reactions in enzymes and solutions. New York:
Wiley & Sons.
112 Martin Culka et al.

Warshel, A. (2003). Computer simulations of enzyme catalysis: Methods, progress, and

insights. Annual Review of Biophysics and Biomolecular Structure, 32, 425–443.
Warshel, A., & Levitt, M. (1976). Theoretical studies of enzymic reactions: Dielectric, elec-
trostatic and steric stabilization of the carbonium ion in the reaction of lysozyme. Journal
of Molecular Biology, 103, 227–249.
Warwicker, J., & Watson, H. C. (1982). Calculation of the electrostatic potential in the active
site cleft due the α-Helix dipols. Journal of Molecular Biology, 186, 671–679.
Webb, S. P., Iordanov, T., & Hammes-Schiffer, S. (2002). Multiconfigurational nuclear-
electronic orbital approach: Incorporation of nuclear quantum effects in electronic struc-
ture calculations. The Journal of Chemical Physics, 117, 4106–4118.
Yilmazer, N. D., & Korth, M. (2015). Enhanced semiempirical QM methods for biomolec-
ular interactions. Computational and Structural Biotechnology Journal, 13, 169–175.
Zavadlav, J., Melo, M. N., Marrink, S. J., & Praprotnik, M. (2015). Adaptive resolution
simulation of polarizable supramolecular coarse-grained water models. The Journal of
Chemical Physics, 142, 244118.
Zhang, Y. (2013). Electrostatic interaction of the electrostatic-embedding and mechanical-
embedding schemes for QM/MM calculations. Communications in Computational Chemis-
try, 1, 109–117.
Zhao, Y., & Truhlar, D. G. (2008). The M06 suite of density functionals for main group
thermochemistry, thermochemical kinetics, noncovalent interactions, excited states,
and transition elements: Two new functionals and systematic testing of four M06-class
functionals and 12 other function. Theoretical Chemistry Accounts, 120, 215–241.
Zimmerman, M., & Bowman, G. (2016). Chapter nine—How to run FAST simulations.
Methods in Enzymology, 578, 213–225.
Zwanzig, R. W. (1954). High-temperature equation of state by a perturbation method. I.
Nonpolar gases. The Journal of Chemical Physics, 22, 1420–1426.