University of Heidelberg

Molecular Biotechnology (Summer 2004)

Bachelor Thesis:

Basic Principles
in
Molecular Modeling

By Barmak Mostofian (2185360) and Filipp Frank (2203674)

Contents

1 Introduction 1

2 Classical Mechanics 2
2.1 Newton’s Second Law . . . . . . . . . . . . . . . . . . . . . . . . 2
2.2 Integration Algorithms . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2.1 Verlet Algorithm . . . . . . . . . . . . . . . . . . . . . . . 3
2.2.2 Leap-Frog Algorithm . . . . . . . . . . . . . . . . . . . . . 4
2.2.3 Velocity Verlet Algorithm . . . . . . . . . . . . . . . . . . 5

3 Statistical Mechanics 6
3.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.2 Ensemble Averages and Time Averages . . . . . . . . . . . . . . 6

4 CHARMM 8
4.1 Force Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
4.1.1 General Features of
Molecular Mechanics Force Fields . . . . . . . . . . . . . . 8
4.1.2 Bond Streching . . . . . . . . . . . . . . . . . . . . . . . . 9
4.1.3 Angle Bending . . . . . . . . . . . . . . . . . . . . . . . . 11
4.1.4 Torsional Terms . . . . . . . . . . . . . . . . . . . . . . . 12
4.1.5 Improper Torsions / Out-of-Plane Bending . . . . . . . . 13
4.1.6 Electrostatic Interactions . . . . . . . . . . . . . . . . . . 14
4.1.7 Van der Waals Interactions . . . . . . . . . . . . . . . . . 15
4.1.8 The CHARMM Force Field – A Simple Molecular Me-
chanics Force Field . . . . . . . . . . . . . . . . . . . . . . 16
4.2 Data Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.2.1 Residue Topology File (RTF) . . . . . . . . . . . . . . . . 18
4.2.2 Parameter File (PARAM) . . . . . . . . . . . . . . . . . . 18
4.2.3 Protein Structure File (PSF) . . . . . . . . . . . . . . . . 19
4.2.4 Coordinate File (CRD) . . . . . . . . . . . . . . . . . . . 19

5 Energy Minimization 20
5.1 Energy Minimization:
Statement of the Problem . . . . . . . . . . . . . . . . . . . . . . 20
5.2 Derivative Minimization Methods . . . . . . . . . . . . . . . . . . 20
5.2.1 First-Order Minimization Methods . . . . . . . . . . . . . 21
5.2.2 A Second-Order Minimization Method - The Newton-Raphson
Method (NR) . . . . . . . . . . . . . . . . . . . . . . . . . 23

i
6 Molecular Dynamics (MD) and Normal Mode Analysis (NMA) 25
6.1 Molecular Dynamics - Running A Molecular Dynamics Simulation 25
6.1.1 Starting Structure . . . . . . . . . . . . . . . . . . . . . . 25
6.1.2 Modification of the Starting Structure . . . . . . . . . . . 25
6.1.3 Energy Minimization . . . . . . . . . . . . . . . . . . . . . 26
6.1.4 Heating Dynamics . . . . . . . . . . . . . . . . . . . . . . 26
6.1.5 Equilibration and Rescaling Velocities . . . . . . . . . . . 28
6.1.6 Production Dynamics . . . . . . . . . . . . . . . . . . . . 29
6.2 Normal Mode Analysis . . . . . . . . . . . . . . . . . . . . . . . . 31

7 What Else is Possible 37

A Acknowledgements 39

ii
List of Figures

4.1 Variation in bond energy with interatomic separation . . . . . . . 10

4.2 Comparison of the simple harmonic potential (Hooke’s Law) with
the Morse curve. . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
4.3 Bond Angle θ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
4.4 Variation in energy with rotation of the carbon-carbon bond in
ethane. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
4.5 A torsion angle (dihedral angle) A-B-C-D is defined as the angle
Phi between the planes (ABC) and (BCD). A torsion angle can
vary through 360 degrees. . . . . . . . . . . . . . . . . . . . . . . 13
4.6 The improper dihedral term is designed to maintain planarity
about certain atoms. The potential is described by a harmonic
function. α is the angle between the plane formed by the central
atom and two peripheral atoms and the plane formed by the
peripheral atoms only.) . . . . . . . . . . . . . . . . . . . . . . . . 14
4.7 The Lennard-Jones potential. The collision parameter, σ, is
shown along with the well depth, epsilon. rn ull is the point
of minimum energy. The dashed curves represent Paulirepulsion
and van der Waals attraction. . . . . . . . . . . . . . . . . . . . . 16
4.8 Example of the RTF for the Alanine residue. . . . . . . . . . . . 18

5.1 Quadratic Approximation at the Minimum. . . . . . . . . . . . . 21

5.2 Steepest Descent . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
5.3 Line Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
5.4 SD on a narrow valley . . . . . . . . . . . . . . . . . . . . . . . . 23
5.5 Minimization No. 4 . . . . . . . . . . . . . . . . . . . . . . . . . . 24

6.1 Aligned BPTI Structures shown in Cartoon Representation

(Blue: Before Minimization; Red: After Minimization) . . . . . . 26
6.2 Temperature vs. Time. . . . . . . . . . . . . . . . . . . . . . . . . 27
6.3 Total Energy vs. Time. . . . . . . . . . . . . . . . . . . . . . . . 27
6.4 Kinetic Energy vs. Time . . . . . . . . . . . . . . . . . . . . . . . 28
6.5 Potential Energy vs. Time . . . . . . . . . . . . . . . . . . . . . . 28
6.6 Total Energy vs. Time during Equilibration . . . . . . . . . . . . 28
6.7 C-S-S Angle of Disulfide Bridge between CYS 5 and CYS 14 . . 29
6.8 C-S-S-C Dihedral Angle of Disulfide Bridge between CYS 5 and
CYS 14 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
6.9 RMSD during Production Run for different Parts of the Protein 30
6.10 Summary of the proceeding of a Molecular Dynamics simulation 31

iii
6.11 As a molecule consisting of three atoms, water has three normal
modes which are presented here. Experimental (and calculated)
frequencies are shown. The first one has a significantly lower
q
frequency ν than the others. Quantity ν is proportional to νk
with force constant k and reduced mass µ . . . . . . . . . . . . . 32
6.12 The motion of a molecule around an energy minimum can be
approximately described by a parabolic energy profile. This is the
reason why one has to generate the energy-minimized structure
(green ball), which is located around a minimum of the energy
surface, before starting a normal mode calculation . . . . . . . . 33
6.13 Vibrations in 2-dimensional space. In reality one more dimension
comes into play . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
6.14 A linear triatomic molecule like CO2 . The vectors (here: scalars)
x1 , x2 , x3 define displacements of the corresponding atoms. . . . 35
6.15 Results of normal mode calculation for a linear triatomic molecule.
λi , ηi and xi describe eigenvalues, eigenvectors and amplitudes. . 36
6.16 Protein molecules are the most examined type of molecule with
respect to vibrational motion. Obtaining the normal modes of
motion, one can notice a difference between the motional fre-
quency of bigger (global) parts and smaller (local) parts of the
molecule, e.g. a whole domain or even just an atomic link be-
tween two distinct atoms. Global motions of a protein are often
specific to it and can be related to its function. . . . . . . . . . . 36

7.1 A saddle point of the multidimensional reaction path corresponds

to the transition state between reactants and products. . . . . . . 37

iv
Chapter 1

Introduction

Molecular modeling is concerned with ways to describe the behavior of molecules

and molecular systems. Computational techniques have revolutionized molecu-
lar modeling to the extent that calculations could not be performed without the
use of a computer molecular modeling is invariably associated with computer
modeling or computational chemistry. This discipline encompasses not only
quantum mechanics (QM) but also molecular mechanics (MM), conformational
analyses and other computer-based methods for understanding and predicting
the behavior of molecular systems.
In this article we describe the approach to molecular modeling with means of
classical mechanics and the usefulness of empirical energy functions for inves-
tigating the physical and chemical properties of a wide variety of molecules.
The CHARMM program [1] is explained in more detail since it presents a
state-of-the-art computer program which can efficiently handle all aspects of
computations with energy functions (force fields). In fact, MM-methods are
also called force field methods. The potential energy of a molecular system is
calculated and the changes of the system in time can be determined by means
of integration algorithms. These computations are the core of a molecular dy-
namics (MD) simulation. After such a simulation, investigating the structures
can start, e.g. calculations of deviation or fluctuation.
We would like to present the mathematical background of a MD simulation
like calculations of the potential energy, of a minimized value for this energy,
of the systems proceeding during a dynamic simulation or of normal modes of
the molecular systems vibration. Some results of these operations dealing with
a small protein called bovine pancreatic trypsin inhibitor (BPTI) are also pre-
sented as far as discussed before. Our goal is to make the way these methods
work easier to understand with the help of our results on BPTI. We also explain
the physical assumptions that help setting up a force field and the most impor-
tant definitions and equations of statistical mechanics which mark a correlation
between the analyses at a microscopic level and thermodynamic properties of
a system. Finally, we will provide an overview of some other techniques that
could be performed within CHARMM but we did not apply on BPTI.

1
Chapter 2

Classical Mechanics

2.1 Newton’s Second Law

Molecular Dynamics simulations are based on Newton’s second law, the equation
of motion [2, 3]:

dvi d2 vi
Fi = mi · a = mi · = mi · 2 . (2.1)
dt dt
It describes the motion of a particle of mass mi along the coordinate xi with
Fi being the force on mi in that direction. This is used to calculate the motion
of a finite number of atoms or molecules, respectively, under the influence of
a force field that describes the interactions inside the system with a potential
energy function, V (~x), where ~x corresponds to the coordinates of all atoms in
the system. The relationship of the potential energy function and Newton’s
second law is given by

F (~xi ) = −∇i V (~x), (2.2)

with F (~xi ) being the force acting on a particle due to a potential, V (~x). Com-
bining these two equaions gives

dV (~x) d2 xi
= −mi · , (2.3)
dxi dt2
which relates the derivative of the potential energy to the changes of the
atomic coordinates in time. As the potential energy is a complex multidimen-
sional function this equation can only be solved numerically with some approx-
imations.
1
With the acceleration being a = − m · dV
dx we can then calculate the changes
of the system in time by just knowing (i) the potential energy V (~x), (ii) initial
coordinates xi,0 and (iii) an initial distribution of velocities, vi,0 . Thus, this
method is deterministic, meaning we can predict the state of the system at any
point of time in the future or the past.
The initial distribution of velocities is usually randomly chosen from a Gaussian
or Maxwell-Boltzmann distribution [3], which gives the probability of atom i
having the velocity in the direction of x at the temperature T by:

2
  21 2
!
1 mi vi,x

mi
p(vi,x ) = · exp − . (2.4)
2πkb T 2 kb T
Velocities are then corrected so that the overall momentum of the system
equals a zero vector:
N
X
P = mi~vi = ~0. (2.5)
n=1

2.2 Integration Algorithms

The solution of the equation of motion given above is a rather simple one which
is only sufficiently good over a very short period of time, in which the velocities
and accelerations can be regarded as constant. So algorithms were introduced
repeatedly performing small time steps, thus propagating the system’s proper-
ties (positions, velocities and accelerations) in time. Time steps are typically
chosen in the range of 1 fs [2]. It is necessary to use such a small time step, as
many molecular processes occur in such small periods of time that they cannot
be resolved with larger time steps. A time series of coordinate sets calculated
this way is referred to as a trajectory and a single coordinate set as a frame.
Today trajectories of about 10 ns, thus consisting of 102 frames, can be calcu-
lated.

2.2.1 Verlet Algorithm

All algorithms assume that the system’s properties can be approximated by a
Taylor series expansion [4]:
1
~x(t + δt) = ~x(t) + δt · ~v (t) + δt2 · ~a(t) + ... (2.6)
2
1
~v (t + δt) = ~v (t) + δt · ~a(t) + δt2 · ~b(t) + ... (2.7)
2
1
~a(t + δt) = ~a(t) + δt · ~b(t) + δt2 · ~c(t) + ..., (2.8)
2
with ~x, ~v and ~a being the positions, the velocities and the accelerations of
the system.The series expansion is usually truncated after the quadratic term.
Probably the most widely used algorithm for integrating the equations of motion
in MD simulations is the Verlet algorithm (Verlet 1967)[2, 3]. It can be derived
by simply summing the Taylor expressions for the coordinates at the time (t+δt)
and (t − δt):
1
~x(t + δt) = ~x(t) + δt · ~v (t) + δt2 · ~a(t) + ... (2.9)
2
1
~x(t − δt) = ~x(t) − δt · ~v (t) + δt2 · ~a(t) − ... (2.10)
2

⇒ ~x(t + δt) = 2~x(t) − ~x(t − δt) + δt2 · ~a(t). (2.11)

3
 Thus, it uses the position ~x(t) and acceleration ~a(t) at time t and the posi-
tions from the previous step ~a(t − δt) to calculate new positions ~x(t + δt).
In this algorithm velocities are not explicitly calculated but can be obtained in
several ways. One is to calculate mean velocities between the positions ~x(t + δt)
and ~x(t) − δt].
1
~v (t) = · [~x(t + δt) − ~x(t − δt)]. (2.12)
2δt
The advantages of this algorithm are that it is straightforward and has mod-
est storage requirements, comprising only two sets of positions [~x(t)and~a(t−δt)]
and the accelerations ~a(t). The disadvantage, however, is its moderate preci-
sion, because the positions are obtained by adding a small term [δt2 ·~a(t)] to the
difference of two much larger terms [2~a(t) − ~x(t − δt)].This results in rounding
errors due to numerical limitations of the computer.
Furthermore, this is obviously not a self-starting algorithm. New positions
~x(t + δt) are obtained from the current positions ~x(t) and the positions at
the previous step ~x(t − δt). So at t = 0 there are no positions for t − δt) and
therefore it is necessary to provide another way to calculate them. One way is
to use the Taylor expansion truncated after the first term:
1
~x(t + δt) = ~a(t) + δt · ~v (t) + δt2 · ~a(t) + ... (2.13)
2

⇒ ~a(−δt) = ~x(0) − δt · ~v (0). (2.14)

2.2.2 Leap-Frog Algorithm

There are several variations of the Verlet algorithm trying to avoid its disad-
vantages. One example is the leap-frog algorithm[2, 3]. It uses the following
equations:
1 1
~v (t + δt) = ~v (t − δt) + δt · ~a(t) (2.15)
2 2
1
~x(t + δt) = ~x(t) + δt · ~v (t + δt), (2.16)
2
where ~a(t) is obtained using
1 dV
~a(t) = − · . (2.17)
m d~x
First, the velocities ~v (t + 21 δt) are calculated from the velocities at (t − δt)
and the accelerations ~a(t). Then the positions ~x(t + δt) are deduced from the
velocities just calculated and the positions at time t. In this way the velocities
first ‘leap-frog’ over the positions and then the positions leap over the velocities.
The leap-frog algorithm’s advantages over the Verlet algorithm are the inclusion
of the explicit velocities and the lack of the need to calculate the differences
between large numbers.
An obvious disadvantage, however, is that the positions and velocities are not
synchronized. This means it is not possible to calculate the contribution of the
kinetic energy (from the velocities) and the potential energy (from the positions)
to the total energy simultaneously.

4
2.2.3 Velocity Verlet Algorithm
The velocity Verlet algorithm (Swope et al. 1982)[2, 3] yields positions, velocities
and accelerations at time t and does not compromise precision:
1
~x(t + δt) = ~x(t) + δt · ~v (t) + δt2~a(t) (2.18)
2
1
~v (t + δt) = ~v (t) + δt[~a(t) + ~a(t + δt)]. (2.19)
2
For this algorithm more than two calculations have to be done for a single
time step. This is due to the fact that calculation of the velocities ~v (t + δt)
requiires acceleration values at (t) and (t + δt). So first the positions at (t + δt)
are calculated; then the velocities at time (t + δt) are computed using
1 1
~v (t + δt) = ~v (t δt) + δt · ~a(t + δt). (2.20)
2 2

Summary
We have now learned four examples of integration algorithms. But what could
make us prefer one over another? As for any other computer algorithm the ideal
method should be fast, which means computationally efficient, require as little
memory as possible and be easy to program. These however are not the main
features you should examine. They are of rather secondary interest for most
MD simulations because most algorithms do not demand significant storage
amount and calculations for the integration are rather fast in comparison to
other calculations in a simulation such as the calculation of the force acting on
every single atom in the system. Thus, other features are considered first: The
algorithm should conserve overall momentum and energy, be time-reversible and
permit a long time step without great loss of precision.
The choice of the integration step size, in fact, is very important. One must
weigh the increased accuracy of using a small step size against the longer real
time that can be simulated when a larger step size is used.

5
Chapter 3

Statistical Mechanics

MD simulations provide information at the microscopic level. Statistical me-

chanics are then required to convert this microscopic information to macroscopic
observables such as pressure, energy, heat capacities, etc. Statistical mechanics
relate these macroscopic observables to the distribution of molecular positions
and motions. Therefore, time independent statistical averages are introduced.
For a better understanding some definitions are reviewed here [3, 5]:

3.1 Definitions
The mechanical or microscopic state of a system is defined by the atomic
positions xi and the momenta pi = mi · vi . They can be considered as a mul-
tidimensional space with 6N coordinates, for which they both contribute 3N
coordinates. This space is called phase space.

The thermodynamic or macroscopic state of a system is defined by a

set of parameters that completely describes all thermodynamic properties of
the system. An example would be the temperature T , the pressure P , and the
number of particles N . All other properties can be derived from the fundamental
thermodynamic equations.

An ensemble is the collection of all possible systems which have different

microscopic states but have the same macroscopic or thermodynamic state.
Ensembles can be defined by fixed thermodynamic properties as already stated
before. Examples for ensembles with different characteristics are:
N V E, N V T, N P T, µV T ,
(E = total energy, P = pressure, V = volume, µ = chemical potential)

3.2 Ensemble Averages and Time Averages

In an experiment one examines a macroscopic sample with an enormously high
number of atoms or molecules respectively. So the measured thermodynamic
properties reflect an extremely large number of different conformations of the
system, representing a subset of the ensemble. We have to say subset, because

6
an ensemble is the complete collection of microscopic systems and a macroscopic
sample can only consist of a finite number of systems. A sufficiently big sample,
however, can be seen as good approximation to an ensemble. That is why
statistical mechanics defines averages corresponding to experimentally measured
thermodynamic properties as ensemble averages [3, 5].
The ensemble average is given by:
ZZ
hAiensemble = pN d~xN A(~
d~ pN , ~xN )ρ(~
pN , ~xN ), (3.1)

where hAi is the measured observable, which is stated as a function of the

momenta p~i and the positions ~xi . Quantity ρ(~pN , ~xN ) is the probability density
for the ensemble and the integration is performed over all momenta and positions
of the system d~ pN , d.~xN . So, the ensemble average is the average value of an
observable weighted with its probability.
This integral is extremely difficult to calculate as it involves calculating all
possible states of the system.
In an MD simulation an extremely large number of conformations is generated
sequentially in time. To calculate an ensemble average the simulation has to
cover all possible conformations corresponding to the ensemble, at which the
simulation takes place. The time average [3] is given by
Zτ
1
hAitime = lim pN , ~xN )dt,
A(~ (3.2)
τ →∞ τ
t=0

where τ is the simulation time.

This expression is well approximated by an average over a simulation performed
over a sufficiently long period of time and so representing a sufficient amount of
phase space:

Zτ M
1 1 X
hAitime = lim pN , ~xN )dt =
A(~ pN
A(~ xN
M,~ M ), (3.3)
τ →∞ τ M i=1
t=0

where M is the number of frames and A(~ pN

M,~xN
M ) the values of the observable
A in frame M .
The idea is based on the Ergodic Hypothesis [5], one of the most fundamental
axioms of statistical mechanics, which states that the time average equals the
ensemble average:

hAiensemble = hAitime . (3.4)

The idea, as already indicated above, is to simulate the system for a rel-
atively long time, so that it will pass through an extremely high number of
conformations, which can then be referred to as a representative subset of an
ensemble.

7
Chapter 4

CHARMM

The CHARMM (Chemistry at HARvard Molecular Mechanics) program is a

general purpose molecular mechanics simulation program. Besides energy min-
imization, dynamics simulation, vibrational analysis and thermodynamic calcu-
lations which are all performed with the use of the empirical potential energy
function (the force field), there are interfaces to several quantum mechanical
programs allowing mixed QM and MM calculations. The program can treat
systems ranging in size from small individual organic molecules to large pro-
teins and DNA molecules either isolated, in solutions or in crystalline solids.
In this section we explicitly describe the core of the CHARMM program, the
force field, and some important files which the program works with. We start
with an introduction into characteristics of force fields before dealing with the
single portions of the function.

4.1 Force Fields

4.1.1 General Features of
Molecular Mechanics Force Fields
Force fields enable the potential energy of a molecular system V to be calcu-
lated rapidly. The energy is a function of the atomic positions of all the atoms in
the system which are usually expressed in term of Cartesian coordinates. Unlike
quantum mechanical methods that deal with the electrons in a system, force field
techniques calculate the energy of a system only as a function of the nuclear posi-
tions. This is legitimated by the Born-Oppenheimer-Approximation [6]. Thus,
Molecular Mechanics are invariably used to perform calculations on systems
containing a significant number of atoms which would bring enormous quantum
mechanical calculations with them.
A typical force field represents each atom in the system as a single point
and energies as a sum of two-, three-, and four-atom interactions such as bond
stretching and angle bending. Although, simple functions (e.g. Hooke’s Law)
are used to describe these interactions, the force field can work quite well. The
potential energy of a certain interaction is described by an equation which in-
volves the positions of the particles and some parameters (e.g force constants
or reference values) which have been determined experimentally or by quantum
mechanical calculations.

8
Several types of force fields exist. Two of those may use an identical functional
form yet have very different parameters and thus bring about different energies
for the same system. Moreover, force fields with the same functional form but
different parameters, and force fields with different functional forms, may give
close results. A force field should be considered as a single entity; it doesn’t need
to be correct to divide the energy into its individual components or even to take
some of the parameters from one force field and mix them with parameters from
another one.
An important point that one shouldn’t forget is that no ‘correct’ form for a force
field exists. If one functional form performs better than another, that form will
be favored. Most of the force fields commonly used do have a very similar form –
we will discuss this particular form in more detail later on – but it should always
be kept in mind that there may be better functional forms, particularly when
developing a force field for new classes of molecules. molecular mechanics force
fields are often a compromise between accuracy and computational efficiency;
the most accurate ones may often be unsatisfactory for efficient computation.
As the performance of computers increases, it becomes possible to incorporate
more sophisticated models.
A concept that is common to most force fields is that of an atom type. For a
quantum mechanics calculation it is usually necessary to specify the charge of
the nuclei, together with the geometry of the system and the overall charge and
spin multiplicity. For a force field calculation, however, the overall charge and
spin multiplicity are not explicitly required, but it is usually necessary to assign
an atom type to each atom in the system. This contains information about
its hybridization state and sometimes the local environment. For example, it is
necessary to distinguish between carbon atoms which adopt a tetrahedral geom-
etry (sp3 -hybridized), which are trigonal (sp2 - hybrdized) and carbons which are
linear (sp-hybridized). The corresponding parameters are expressed in terms of
these atom types, so that the reference angle Θ0 for a tetrahedral carbon atom
would be about 109.5 ◦ and that for a trigonal carbon near 120. For example,
the MM2 [7], MM3 [8, 9, 10] and MM4 [11, 12, 13, 14, 15] force fields of Allinger
and co-workers, that are widely used for calculations on ”small” molecules,
distinguish the following types of carbon atoms: sp3 , sp2 , sp, carbonyl, cyclo-
propane, radical, cyclopropene and carbonium ion. The value of the potential
energy V is calculated as a sum of internal or bonded terms, which describe
the bonds, angles and bond rotations in a molecule, and a sum of external or
non-bonded terms, which account for interactions between non-bonded atoms
or atoms separated by three or more covalent bonds. So it is:

V (~x) = Vbonded (~x) + Vnon−bonded (~x), (4.1)

Let us now discuss the individual contributions to a molecular mechanics
force field, giving a selection of the various functional forms that are in common
use. We shall then have a look at the CHARMM force field which is used for
our calculations on BPTI.

4.1.2 Bond Streching

The potential energy curve (Morse potential) for a typical bond has the shape
shown in Fig.4.1. This potential has the form:

9
 Figure 4.1: Variation in bond energy with interatomic separation

VB (t) = De {1 − exp[−a(l − l0 )]}2 . (4.2)

q
µ
De is the depth of the potential energy minimum and a = ω · 2D e
, where µ
is the reduced mass and ω is the frequency of the bond vibration. q The frequency
ω is related to the stretching constant of the bond kl , by ω = kµl , where kl
determines the strength of the bond. The length l0 is the reference bond length.
It is the value that the bond adopts when all other terms in the force field are
zero. Both l0 and kl are specific for each pair of bound atom. Values of kl are
often evaluated from experimental data such as infrared stretching frequencies
or from quantum mechanical calculations. Values of l0 can be inferred from
high resolution crystal structures or microwave spectroscopy data.
The Morse potential is not usually used in molecular mechanics force fields. It is
computationally demanding the curve describes a wide range of behavior, from
the strong equilibrium to dissociation. Normally, this is not necessary for Molec-
ular Mechanics calculations where we are more interested in slight deviations
of bonds from their equilibrium values. Consequently, simpler expressions are
often used. The most elementary approach is to use a Hooke’s Law formula in
which the energy varies with the square of the displacement from the reference
bond length l0 :
1
VB (l) = kl · (l − l0 )2 (4.3)
2
The Hooke’s Law functional form is a reasonable approximation to the shape
of the potential energy curve at the bottom of the potential well, at distances

10
Figure 4.2: Comparison of the simple harmonic potential (Hooke’s Law) with
the Morse curve.

that correspond to bonding in ground-state molecules. It is less accurate away

from equilibrium (Fig.4.2).
To approximate the Morse curve more accurately, cubic and higher terms
can be included which give a better model close to the equilibrium structure
than the quadratic form but also create a potential passing through maxima
further away from l0 . This can lead to a catastrophic lengthening of bonds.

4.1.3 Angle Bending

The deviation of valence angles from their reference values (Fig.4.3) is also
frequently described using a harmonic potential:
1
VA (Θ) = · kΘ · (Θ − Θ0 )2 (4.4)
2
The force constant kΘ and the reference value Θ0 depend on the chemical
type of atoms constituting the angle. Rather less energy is required to distort an
angle away from equilibrium than to stretch or compress a bond, and thus force

11
 Figure 4.3: Bond Angle θ

constants are smaller here. As with the bond-stretching terms, the accuracy of
the force field can be improved by the incorporation of higher-order terms.
These two terms, the bond-stretching and angle-bending, describe the deviation
from an ideal geometry; effectively, they are penalty functions and the sum of
them should be close to zero in a perfectly optimized structure. These two terms
are often regarded as hard degrees of freedom, in that quite substantial ener-
gies are required to cause significant deformations from their reference values.
Most of the variation in structure and relative energies is due to torsional and
non-bonded contributions.

4.1.4 Torsional Terms

This term represents the torsion angle potential function which models the pres-
ence of steric barriers between atoms separated by three covalent bonds. The
existence of barriers to rotation about chemical bonds is fundamental to un-
derstanding the structural properties of molecules and conformational analysis.
The three minimumenergy staggered conformations and the three maximum-
energy eclipsed structures of ethane (Fig.4.4) are a classic example of the way
in which the energy changes with a bond rotation [16]. Quantum mechanical
calculations suggest that the rotation-barrier arises from antibonding interac-
tions between the H-atoms on opposite ends of the molecule; they are minimized
when the conformation is staggered. Not all molecular mechanics force fields
use torsional potentials; it may be possible to rely on non-bonded interactions
between the atoms at the end of each torsion angle. However, most force fields
for organic molecules do use explicit torsional potentials with a contribution
from each bonded quartet of atoms A-B-C-D in the system (Fig.4.5). Thus,
there would be nine individual torsional terms for ethane. Torsional potentials
are almost always expressed as a cosine series expansion. One functional form
is:
1
VT (Φ) = kΦ · [1 + cos(nΦ − δ)] (4.5)
2
The quantity kΦ is often referred to as the barrier height, but to do so is
misleading, obviously so when more than one term is present in the expansion.
Moreover, other terms in the force field equation contribute to the barrier height

12
Figure 4.4: Variation in energy with rotation of the carbon-carbon bond in
ethane.

Figure 4.5: A torsion angle (dihedral angle) A-B-C-D is defined as the angle
Phi between the planes (ABC) and (BCD). A torsion angle can vary through
360 degrees.

as a bond is rotated, especially the non-bonded interactions between atom A and

D. The value of kΦ does, however, give a qualitative indication of the relative
barriers to rotation; for example kΦ for an amide bond (A-C=N-D) will be
larger than for a bond between two sp3 carbon atoms (A-C-C-D). n is the
periodicity; its value gives the number of minimum points in the function as the
bond is rotated through 360 ◦ C. δ (the phase factor) sets the minimum energy
angle.

4.1.5 Improper Torsions / Out-of-Plane Bending

Several chemical groups involve arrangements of 4 or more atoms in a plane. For
these groups it is sometimes advantageous to have an additional term enforc-
ing planarity. For example, it is found experimentally that the oxygen atom of
cyclobutanone remains in the plane of the ring. This is because the π-bonding
energy, which is maximized in the planar arrangement, would be much reduced
if the oxygen were bent out of the plane. Using a force field containing just stan-

13
Figure 4.6: The improper dihedral term is designed to maintain planarity about
certain atoms. The potential is described by a harmonic function. α is the angle
between the plane formed by the central atom and two peripheral atoms and
the plane formed by the peripheral atoms only.)

dard terms we already know, the equilibrium structure would have the oxygen
atom located out of the plane formed by the adjoining carbon atom and the two
carbon atoms bonded to it. The simplest way to achieve the desired geometry
is to use an out-of-plane bending term. One approach is to treat the four atoms
as an improper torsion angle i.e., a torsion angle in which the four atoms are
not bonded in the sequence A-B-C-D. Another way involves a calculation of the
angle between a bond from the central atom and the plane defined by the central
atom and the other two atoms (Fig.4.6). A value of 0 ◦ corresponds to all four
atoms being planar. With these definitions the deviation of the out-of-plane
coordinate can be modeled using a harmonic potential of the form:
1
ν(α) = kα · α2 . (4.6)
2
The improper torsion or improper dihedral definition is more widely used as
it can then be easily included with the ‘proper’ torsional terms in the force field.
However, the other form may be better to implement out-of-plane bending in
the force field.
After learning the most important bonded terms of the energy function we are
now going to have a look at the non-bonded terms which consist at least of the
electrostatic and the van der Waals interactions in the system.

4.1.6 Electrostatic Interactions

Interactions between atoms due to their permanent dipole moments are de-
scribed approximately by treating the charged portions as point charges. Then
we use the Coulomb potential for point charges to estimate the forces between
the charged atoms. The Coulomb potential is an effective pair potential that
describes the interaction between two point charges. It acts along the line con-
necting the two charges. It is given by the equation:

14
 qi qj
VE (i, j) = . (4.7)
4π0 r0
rij is the distance between qi and qj , the electric charge in coulombs carried
by charge i and j respectively, and 0 is the electrical permittivity of space.
Alternative approaches to the calculation of electrostatic interactions, e.g. the
central multipole expansion which is based on the electric moments, may provide
more exact solutions to the electrostatic interactions [17, 18].

4.1.7 Van der Waals Interactions

The van der Waals interaction between two atoms arises from a balance be-
tween repulsive and attractive forces. The repulsive force arises at short dis-
tances where the electron-electron interaction is strong (Pauli repulsion). The
attractive force, also referred to as dispersion force, arises from fluctuations in
the charge distribution in the electron clouds. The fluctuation in the electron
distribution on one atom gives rise to an instantaneous dipole which, in turn,
induces a dipole in a second atom giving rise to an attractive interaction. Each
of these two effects is equal to zero at infinite atomic separation and becomes
significant as the distance decreases. The attractive interaction is longer range
than the repulsion but as the distance becomes short, the repulsive interaction
becomes dominant. This gives rise to a minimum in the energy (see Fig.4.7).
For a force field we require a means to model the interatomic potential curve ac-
curately, using a simple empirical expression that can be rapidly calculated. The
best known of the van der Waals potential functions is the Lennard-Jones 12-
6 function, which takes the following form for the interaction between two atoms:
  
σ 12  σ 6
VvdW (r) = 4 · − . (4.8)
r r
The Lennard-Jones 12-6 potential contains just two adjustable parameters: the
collision diameter σ (the separation for which the energy is zero) and the well
depth . These parameters are graphically illustrated in Fig.4.7.
The need for a function that can be rapidly evaluated is a consequence of the
large number of van der Waals interactions that must be determined in many
of the systems that we would like to model. The 12-6 potential is widely used,
particularly for calculations on large systems, as r-12 can be quickly calculated
by squaring the r-6 term. The r-6 term can also be calculated from the square
of the distance without having to perform a computationally expensive square
root calculation. Different powers have also been used for the repulsive part of
the potential [19]; values of 9 or 10 give a less steep curve and are used in some
force fields.
The 6-12 term between hydrogen-bonding atoms is replaced by an explicit hydro-
genbonding term in some force fields, which is often described using a Lennard-
Jones 10-12 potential. This function is used to model the interaction between
the donor hydrogen atom and the heteroatom acceptor atom. Its use is intended
to improve the accuracy with which the geometry of hydrogen-bonding systems
is predicted.
The most time consuming part of a molecular dynamics simulation is the cal-
culation of the non-bonded terms in the potential energy function. In principle,
these energy terms should be evaluated between every pair of atoms; in this

15
Figure 4.7: The Lennard-Jones potential. The collision parameter, σ, is shown
along with the well depth, epsilon. rn ull is the point of minimum energy. The
dashed curves represent Paulirepulsion and van der Waals attraction.

case, the number increases as the square of the number of atoms for a pair-wise
model. To speed up the computation two approaches are applied. In the first
approach the interactions between two atoms separated by a distance greater
than a pre-defined distance, the cutoff distance, are ignored. The interactions
are simply set to zero for interatomic distances greater than the cutoff distance
(Truncation-Method) or the entire potential energy surface is modified such
that at the cutoff distance the interaction potential is zero (Shift-Method). The
other way is to reduce the number of interaction sites. The simplest way to do
this is to subsume some or all of the atoms (usually just the hydrogen atoms)
into the atoms to which they are bonded (United-Atom-Method). Considerable
computational savings are possible; for example, if butane is modeled as a four-
site model rather than one with twelve atoms, the van der Waals interaction
between all the atoms involves the calculation of six terms rather than 78.

4.1.8 The CHARMM Force Field – A Simple Molecular

Mechanics Force Field
Now that we know how the different contributions to a force field can be
described, let us look at the simplest form such a force field can have, the
CHARMM force field:

16
 1 X 1 X
Vtot = kl · (l − l0 )2 + kΘ · (Θ − Θ0 )2 +
2 2
bonds angles
1 X 1 X
+ kΦ · [1 + cos(nΦ − δ)] + kα · α2 +
2 torsions 2 impropers (4.9)
" 12  6 #
X qi qj X σij σij
+ + 4i j − .
i,j
4π0 rij ij
rij rij

Let us consider how this simple force field would be used to calculate the
energy of a conformation of a simple molecule. Propane is one that is most
popular for this task [2] it has more terms than ethane for example, and it is
not as complicated as butane (butane has far more non-bonded and torsional
energy terms. As it is explained by Leach, propane has ten bonds: two C-C
and eight C-H bonds. The C-C bonds are symmetrically equivalent but the
C-H bonds fall into two classes, one group corresponding to the two hydrogens
bonded to the central methylene (CH2 ) carbon and one group corresponding to
the six hydrogens bonded to the methyl carbons. In some sophisticated force
fields different parameters would be used for these two different types of C-H
bond, but in the CHARMM force field the same bonding parameters (i.e. kl
and l0 ) would be used for each of the eight C-H bonds. This is a good example
for transferability since the same parameters can be used for a wide variety of
molecules. There are 18 different valence angles in propane, comprising one
C-C-C angle, ten C-C-H angles and seven H-C-H angles. Note that all angles
are included in the CHARMM force field even though some of them may not
be independent of the others. There are 18 torsional terms: twelve H-C-C-H
torsions and six H-C-C-C torsions. They are modeled with a cosine series ex-
pansion having minima at the trans and gauche conformations. The improper
dihedral term is dropped out for propane. Finally, there are 27 non-bonded
terms to calculate, comprising 21 H-H interactions and six H-C interactions. A
sizeable number of terms are thus included in the CHARMM model, even for
a molecule as simple as propane. Even so, the number of terms, namely 73 is
much less than the number of integrals that would be involved in an equivalent
quantum mechanical calculation.
The force field equation given above is only one variant of the many CHARMM
force fields. There are other potential energy functions which contain more
terms for a more precise calculation, implemented in the CHARMM program,
e.g. the extended electrostatics model [20] or the fast multipole method [21]
for treating long-range electrostatic interactions with a multipole approxima-
tion. Our calculations on BPTI, however, were performed with the force field
presented above.

4.2 Data Structures

Data Structures include information about the molecule, its composition, its
chemical connectivity, certain atomic properties and parameters for the energy
function and more. This information for a particular class of molecules, e.g.
proteins or nucleic acids, is contained in the topology file and the parameter file.

17
 Figure 4.8: Example of the RTF for the Alanine residue.

For a specific molecule, the necessary data is stored in the Protein Structure
File and the Coordinate File, respectively.

4.2.1 Residue Topology File (RTF)

The RTF contains local information about atoms, bonds, angles etc. for each
possible type of monomer unit (residue) that can be used in building a particular
type of macromolecule. For each residue the covalent structure is defined, i.e.,
how the atoms are connected to one another to form amino acids, DNA bases
or lipid molecules. Fig.4.8 depicts the RTF for the amino acid Alanine.

4.2.2 Parameter File (PARAM)

The parameter File is associated with the RTF file as it contains all the nec-
essary parameters for calculating the energy of the molecule(s). These include
the equilibrium bond lengths and angles for bond stretching, angle bending and
dihedral angle terms in the potential energy function as well as the force con-
stants and the Lennard-Jones 12-6 potential parameters. As already mentioned,
these parameters are associated with particular atom types.

18
4.2.3 Protein Structure File (PSF)
The PSF is the most fundamental data structure used in CHARMM. It is gener-
ated for a specific molecule or molecules and contains the detailed composition
and connectivity of the molecule(s). It describes how molecules are divided
into residues and molecular entities (segments), which can range from a single
macromolecular chain to multiple chains solvated by explicit water molecules.
The PSF must be specified before any calculations can be performed on the
molecule. The PSF constitutes the molecular topology but does not contain
information regarding the bond lengths, angles, etc., so it is necessary to read
in the parameter file to add the missing information.

4.2.4 Coordinate File (CRD)

The CRD file contains the Cartesian coordinates of all atoms in the system.
Those are most often obtained from x-ray crystal structures or from NMR
structures. Missing coordinates can be built within the CHARMM program; in
addition, hydrogen atoms which are not present in x-ray crystal structures of
proteins can be placed using CHARMM. Two sets of coordinates can also be
accommodated in the program which is useful for a variety of calculations, e.g.
the RMSD (Root Mean Square Deviation) between the experimental structure
and a structure from the simulation (see Chapter 6.1).

19
Chapter 5

Energy Minimization

5.1 Energy Minimization:

Statement of the Problem
Given is a multidimensional function V which depends on several variables
x1 , x2 , x3 , . . . , xi . The problem is to find the values of these variables where
V gives minimum values. Minimum values are those points in an energy land-
scape where you will reach higher positions in any direction.
So the first derivative of the function with respect to every single variable is
zero:
∂V
= 0, (5.1)
∂xi
with xi and xj being the coordinates of the simulated atoms; and the matrix
∂V
∂xi is positive definite.
According to this, minimization methods can be classified into two main groups:
Those algorithms which use derivatives and those which do not [2]. Derivatives
provide useful information about the shape of the energy landscape and so can
significantly accelerate the search process. That is why methods which use
derivatives are more appropriate to minimize the complex functions describing
the energy landscape of a molecule.
For this reason we will solely concentrate on these methods in the following.

5.2 Derivative Minimization Methods

The direction of the first derivative, the gradient, of a function tells us in which
way we will find a minimum and its magnitude indicates the steepness of the
slope at a given point of the function. The second derivative provides informa-
tion about the curvature of the energy landscape.
To describe derivative methods it is useful to write the energy function as a Tay-
lor expansion around point xk (that we write first for the 1-dimensional case)
[4]:

20
 Figure 5.1: Quadratic Approximation at the Minimum.

1
V (x) = V (xk ) + (x − xk ) · V 0 (xk ) + [(x − xk )2 · V 00 (xk )] + . . . , (5.2)
2
where V 0 is the first derivative and V 00 the second derivative of the energy
function V .
In the case of a multidimensional function the variable x corresponds to a vector
~x and the derivatives are replaced by matrices: For a system with N atoms V (~x)
is a function of 3N coordinates [2]. So ~x has 3N components and the gradient,
~g = V 0 (~x)T , accordingly is a vector with 3N dimensions as well, with each
element being the partial derivative of V with respect to a single coordinate,
∂V 00
∂xi . The second derivative V (~ x) is a (3N × 3N )-matrix. Every element (i, j)
corresponds to the partial second derivate of V with respect to the coordinates
2
xi and xj , ∂x∂i ∂x
V
j
. This is a symmetric matrix which is called Hessian Matrix.
Thus the multidimensional Taylor expansion is written as follows [22]:

1
V (~x) = V (~xk ) + V 0 (~xk ) · (~x − ~xk ) + [(~x − ~xk )T · V 00 (~xk ) · (~x − ~xk ) + . . . . (5.3)
2
This is a quadratic function and thus can only be seen as an approximation
for an energy function. However, the area close to a minimum is well approxi-
mated by this Taylor expansion as can be seen in Fig.5.1
In the following we will discuss the most important and most frequently used
methods for energy minimization in molecular modelling. They can be classified
according to the highest order derivative used.

5.2.1 First-Order Minimization Methods

First-Order minimization methods use the information of the first derivative,
the gradient ~gk , to find local minima of the function of interest.

Steepest Descent (SD)

The steepest descent method moves along the negative gradient −~gk downhill
the energy landscape beginning from a starting point of interest. As all these

21
 Figure 5.2: Steepest Descent Figure 5.3: Line Search

methods are iterative methods the search is done stepwise:

Calculate the gradient ~gk at a starting point ~xk and take a step in the direction
of the negative gradient −~gk , the point of arrival being the starting point for
the next iteration.
Knowing the direction of the gradient the next thing to do is to determine the
length of the step to take. This is achieved by performing a line search in the
direction of the gradient [2]. Imagine a cross-section through the energy function
in direction of ~gk and you see that there will be a (one-dimensional) minimum,
which is the optimal starting point for the next step, ~gk+1 .
If you take the next step from this minimum point, the directions, ~νk , and
the gradients, ~gk , of successive steps will always be orthogonal [2]:

~gk · ~gk+1 = 0 (5.4)

~νk · ~νk−1 = 0 (5.5)

Finding the minimum via line search can be achieved in several ways. One
possibility, for example, is to perform a one dimensional quadratic approxima-
tion in the direction of the gradient.

Conjugate Gradient Method

The SD approach encounters severe problems when used with functions other
than quadratic functions especially when they have the shape of a narrow valley.
In this case it gives undesirable behaviour [22] taking only very short steps and
thus extremely increasing computation (see Fig.5.4).
That is why another, similar approach has been introduced: the conjugate gra-
dient method. In contrast to SD, in this method the gradients of successive
steps are not orthogonal. Instead the directions at each point ~vk are orthogo-
nal to gradients of the following steps ~gk provided that a line search has been
performed:

22
 Figure 5.4: SD on a narrow valley

~νk · ~gk+1 = 0. (5.6)

For every direction ~νk the direction of the previous step ~νk−1 is taken into
account:

~νk = −~gk + γk · νk−1 , (5.7)

with γ being a scalar constant given by

~gk · ~gk
γk = . (5.8)
~gk−1 · ~gk−1
Thus, new directions are linear combinations of the current gradient ~gk and
the previous direction ~νk−1 . As there is no previous direction for the first step,
the conjugate gradient methods starting direction is the same as for SD.

5.2.2 A Second-Order Minimization Method - The Newton-

Raphson Method (NR)
For this method lets have a look at the Taylor expansion in the 1-dimensional
case again:

1
V (x) = V (xk ) + (x − xk ) · V 0 (xk ) + [(x − xk )2 · V 00 (xk )] + . . . (5.9)
2
The first derivative of this function is

V 0 (x) = V 0 (xk ) + (x − xk ) · V 00 (xk ). (5.10)

At the minimum x = xM in the first derivative is Zero (V 0 (xM in ) = 0) and so

V 0 (xk )
xM in = xk − . (5.11)
V 00 (xk )
The expression for a multidimensional function is

23
 Minimization 1. Method 2. Method Final Energy / kcal·mol−1
1. 600 Steps SD - -875
2. 600 Steps NR - -1142
3. 100 Steps NR 500 Steps SD -1069
4. 100 Steps SD 500 Steps NR -1189

Table 5.1: Results of different Energy Minimizations performed on BPTI.

~xM in = ~xk − (V 00 )−1 (~xk )V 0 (~xk ), (5.12)

00 −1
where (V ) (~xk ) is the inverse of the Hessian matrix. Calculation of it can
be computationally very expensive, especially for complex energy functions of
large molecules [1].
Like the SD approach NR is an iterative method. For a quadratic function this
method finds the minimum in one step from any starting point. In practice,
however, a quadratic function is just an approximation for an energetic land-
scape and you have to perform more steps to get to a minimum. In this case
~xM in is used as the starting point (~xk ) for the next iteration.
Besides the methods covered here, there are a lot of other procedures used to
minimize the energy in molecular dynamics simulations. For example, there is
a number of methods called quasi-Newton methods, which aim to reduce com-
putation by eliminating the need to calculate the Hessian matrix.
Having discussed several minimization methods used in molecular modelling we
now want to know which method should be applied. Since we start minimiza-
tion from experimentally obtained structures with very high energy, SD or the
conjugate method will be first applied as they reach lower energies very fast due
to their low demands on computation. However, when reaching regions of lower
energy, where slopes are less steep, other, more sophisticated methods such as
Newton-Raphson can be applied to speed up the search for low energy points.
To examine the performance of the two most frequently used methods, SD and
NR, four different minimizations were performed on BPTI the results confirmed
the assumptions stated above.
Minimization 4, which had the best performance, is shown in Fig.5.4 where
the energy is plotted against the number of time steps. When the minimization
method switches to NR, the number of time steps is set to Zero again.

Figure 5.5: Minimization No. 4

24
Chapter 6

Molecular Dynamics (MD)

and Normal Mode Analysis
(NMA)

6.1 Molecular Dynamics - Running A Molecular

Dynamics Simulation
6.1.1 Starting Structure
Before starting a molecular dynamics simulation one needs an initial set of
coordinates meaning a configuration of the molecule being simulated. In general
structures retrieved from X-ray scattering or NMR experiments are used [23].
These can be obtained from the Brookhaven Protein Data Bank (http://www.
rcsb.org/pdb/). The starting structure is of extreme importance. As molecular
dynamics simulations are computationally very expensive, it is only possible to
calculate a few nanoseconds within an appropriate period of time. Dynamic
molecular processes like protein folding, however, take much longer; protein
folding for example can require a few milliseconds up to 10 seconds in vivo.
Thus, the starting structure should be close to the state you want to simulate.

6.1.2 Modification of the Starting Structure

Structures retrieved from a protein databank lack hydrogen atoms, because they
cannot be resolved properly. So these have to be added to the file. In addition
there are water molecules immanent in the structure. A decision has to be
made whether to include water molecules in the simulation or not. Then you
can either add additional water molecules or remove them. This depends on
what kind of data you want to produce. People who want to simulate crystal
states will exclude water. Others, examining ligand binding for example, will
include water, as they simulate a process that occurs in a biological environment.

25
Figure 6.1: Aligned BPTI Structures shown in Cartoon Representation
(Blue: Before Minimization; Red: After Minimization)

6.1.3 Energy Minimization

As mentioned before NMR structures and X-ray crystal structures tend to have
high energy interactions like Pauli repulsions. That is due to the fact that
the methods to retrieve molecular structures are not perfect and especially in
x-ray-structures there are crystal contacts, which lead to a compaction of the
molecules. Furthermore, hydrogen atoms are added to relatively arbitrary po-
sitions near their neighbors. Thus there are atoms lying too close together so
that the Pauli repulsion outweighs the dispersion attraction and the energy is
raised high above natural energy levels. These high energy interactions lead to
local distortions which result in an unstable simulation. They can be released
by minimizing the energy of the structure before starting a run.
The minimization results in a structure with an energy near the lowest possible
energy the system can have. Thus, this state corresponds to a temperature near
0 K, where no motion can be seen [2]. You can easily imagine that there is no
motion meaning no forces on the atoms in the system of a minimized structure:
In an energy minimum the gradient of the potential energy equals a zero vector,
∂V (~x)
∂~x . This means the derivative of V (~ x) with respect to any coordinate of
any atom equals zero. And so, with the force being the negative gradient of
V (~x), there is no net force acting on any atom in any direction and therefore no
motion will be seen inside the system. As you can see in Fig.6.1, the structures
before and after minimization can vary extremely.

6.1.4 Heating Dynamics

To raise the temperature from 0 K to the desired value, you first have to assign
initial velocities to the atoms corresponding to a Gaussian distribution for a
certain temperature to provide the system with energy [1].
Then for the first time in the course of the MD procedure Newton’s equations

26
 Figure 6.2: Temperature vs. Time. Figure 6.3: Total Energy vs. Time.

of motion are integrated to propagate the system in time. This is done for a
certain period of time to let the system equilibrate in the new thermodynamic
state, giving the energy time to evenly distribute throughout the system. In the
next step the velocities are scaled to values corresponding to a slightly higher
temperature and another equilibration phase is carried out. You can reach the
desired temperature by simply repeating this process.
Typical steps for raising the temperature are about 5 K and the short equilibra-
tion period lasts about 0.3-1.0 s depending on the size of the simulated system.
So a heating process with 5 K every 0.3 ps, for example, would raise the temper-
ature from 0 K to 300 K in 20 ps (as shown for BPTI in Fig6.2), which is in fact
very quick. An even more rapid heating, though, would result in high-energy
motions and interactions that are physically not feasible. Energy density for
example would be increased locally to a level, when bonds break or, if quantum
mechanics were neglected, they would elongate far beyond natural dissociation
lengths. So every molecule with a complex three dimensional structure made up
by hydrogen bonds or other non-bonded interactions would denature and thus
be useless for any further simulation.
In Figures 6.2-6.5 it is shown how some properties of the system vary in time
during a heating process. As you can see the kinetic energy and the temperature
show exactly the same behavior. This is because they are related to each other
by following equation:
1 3
Ekin = mh~v 2 i = N kZ, (6.1)
2 2
where N is the number of atoms and k the Boltzmann constant.
Having raised the total energy of the system in a heating step, this amount of
energy is evenly distributed throughout the system as already stated before. On
the one hand this distribution occurs in space, meaning very high and very low
velocities of single atoms get closer to the mean value. On the other hand the
distribution occurs in terms of different energies: kinetic and potential energy.
As you can see the total energy is the only property that is constant during
equilibration. Kinetic and potential energy, however, behave irregularly, show-
ing that the total energy distributes between them in a vibrational kind of way.

27
Figure 6.4: Kinetic Energy vs. Time Figure 6.5: Potential Energy vs. Time

6.1.5 Equilibration and Rescaling Velocities

After having heated the system so quickly, the structure might be unstable and
the temperature may drop too low [3]. That is why you have to equilibrate your
system properly before running the real dynamics simulation, the production
run. As stated before, equilibration is the process where the kinetic energy and
the potential energy evenly distribute themselves throughout the system.
This is done by simulating the system whilst monitoring important properties
such as temperature, the different energy terms, structure etc.. At constant
periods of time the velocities are rescaled to the values for the desired temper-
ature. This can be seen as vertical lines in Fig.6.6.
This is done until the simulation becomes stable with respect to time, which
means till thermodynamic terms like temperature and energy are retained in
a certain, small interval for a sufficiently long time. Only when the average
temperature of the system stabilizes one can collect the trajectory information
for analysis.
You may wonder now why the energy or temperature drops at all. In natural
systems the energy is conserved, of course. In a simulation, however, energy

Figure 6.6: Total Energy vs. Time during Equilibration

28
 Figure 6.8: C-S-S-C Dihedral Angle of
Figure 6.7: C-S-S Angle of Disulfide
Disulfide Bridge between CYS 5 and
Bridge between CYS 5 and CYS 14
CYS 14

conservation may be violated due to several reasons: The force field may ne-
glect some critical effects (for example due to a badly selected cutoff) or the
numerical integration method may be not precise enough due to numerical lim-
itations of the computer resulting from the binary coding of numbers.

6.1.6 Production Dynamics

After all these preparations, which can take up to a couple of weeks, the actual
molecular dynamics simulation can be started by integrating Newton’s equa-
tions of motion of the system for the desired period of time. This can be from
several hundred picoseconds up to some nanoseconds.
All the coordinates, velocities, accelerations and momenta generated during the
production run are saved and used for analysis. Thus, for example, average
structures can be calculated and compared to experimental structures and even
more important time dependent properties such as different energy terms, an-
gles and dihedral angles or distances between atoms or whole selections of atoms
can be displayed and interpreted.
For example the time dependent behavior of the C-S-S angles and C-S-S-C di-
hedral angles in a disulfide bridge of BPTI during a 50 ps simulation is shown
in Figures 6.7 and 6.8.

As disulfide bridges are tertiary structure elements and thus play a role in
keeping the three-dimensional structure alive, they should be conserved during
simulation. A flip between totally different levels of angles would indicate a
conformational change of the disulfide bridge, which was obviously not the case
during our simulation.
Another important time dependent property of dynamic systems is the root
mean square deviation (RMSD). RMSD indicates how much two structures vary
in terms of differences between the coordinates of the structures and is calculated
with

29
 Figure 6.9: RMSD during Production Run for different Parts of the Protein

v
u
u1 X N
β 2 1
DRM S α
= h(~xi − ~xi ) i = t
2 (~xα − ~xβi )2 (6.2)
N i=1 i

where ~xi is the coordinate of atom i and α and β correspond to the different
structures. Calculation of RMSD for a time series of coordinate sets compared
to a reference structure yields graphs like shown in Fig.6.9, where RMSD values
of BPTI are displayed for (i) the whole protein, (ii) the side chains and (iii) the
backbone of the protein compared to BPTI crystal structure.

As you can see, the side chains RMSD is higher compared to the backbone,
which means that they are more flexible, whereas the backbone seems rather
rigid. The whole protein RMSD obviously has values between these two, because
for this calculation both atom selections (side chain and backbone) are taken
into account at once.
A special case of RMSD is the root mean square fluctuation (RMSF) where the
reference structure is an average structure over the whole trajectory:
v
u
u1 X N
α Average 2 1
FRM S = h(~xi − ~xi ) i =t
2 (~xα − ~xAverage )2 . (6.3)
N i=1 i i

RMSF can be related to experimental data by comparing with so-called B

factors, Bi , which are measured in X-ray experiments:
8 2
Bi = π (FRM S,i )2 (6.4)
3
RMSF calculation can also give better information on how structures fluc-
tuate during a simulation, while RMSD is more appropriate to show that sim-
ulations are performed close to experimental structures to convince scientists
with a rather critical view on reliability of molecular dynamics simulations.
Thus, combined with molecular graphics programs which can display molecular
structures in a time dependent way, molecular dynamics simulations provide a
powerful tool to visualize and understand conformational changes involved in

30
Figure 6.10: Summary of the proceeding of a Molecular Dynamics simulation

ligand binding, catalysis or other functions of biological macromolecules at an

atomic level.
To give an overview of the whole procedure described in this chapter, it is sum-
marized in Fig.6.10.

6.2 Normal Mode Analysis

The Normal Mode Analysis (NMA) calculates features of the normal vibration
modes of molecules. Obtaining them, scientists try to deduce certain properties
of molecules (e.g. protein function) just by observing the overall vibrational mo-
tion of the examined molecule or by comparing the theoretical to spectroscopic
results. Furthermore, comparisons to experiments can be used in the parametri-
sation of a force field. NMA is performed at a hypothetical, motionless state
at 0 K. However, experimental measurements are made on molecules at a finite
temperature when the molecules undergo all kinds of motion. To compare theo-
retical and experimental results it is necessary to make appropriate corrections
to allow for these motions. These corrections are calculated using standard sta-
tistical mechanics formulae.
The vibrational contribution (Uvib ) to the internal energy at a temperature T
requires knowledge of the actual vibrational frequencies (ni ). Constituent Uvib

31
Figure 6.11: As a molecule consisting of three atoms, water has three normal
modes which are presented here. Experimental (and calculated) frequencies are
shown. The first one has a significantly
q lower frequency ν than the others.
k
Quantity ν is proportional to ν with force constant k and reduced mass µ

equals the difference in the vibrational enthalpy at the temperature T and at

0 K and is given by:
 
X hνi
Uvib (T ) = Nnm (6.5)
i=1
exp(hνi /kb T ) + 1
Nnm is the number of normal vibrational modes for the system. Even the
zero-point energy Uvib (0), obtained by summing 21 hνi for each normal mode,
can be quite substantial, amounting to about 100 kcal/mol for a C6 molecule.
Other thermodynamic quantities such as entropies and free energies may also
be calculated from the vibrational frequencies using the relevant statistical me-
chanics expressions.
Normal modes of vibrations are oscillations about an energy minimum, charac-
teristic of a system’s structure and its energy function. For a purely harmonic
energy function, any motion can be exactly expressed as a superposition of nor-
mal modes (anharmonic potentials can be approximated by harmonic potentials
at sufficiently low temperatures of the system). Thus, normal modes are useful
because they describe collective motions of the atoms in a coupled system that
can be individually excited.
The three normal modes of water are schematically illustrated in Fig.6.11; a
nonlinear molecule with N atoms has 3N 6 normal modes [24].
A normal mode calculation is based on the assumption that the energy
surface is quadratic in the vicinity of an energy minimum (Fig.6.12).
This means that each normal mode acts as a simple harmonic oscillator with

d2 x
m = F = −kx. (6.6)
dt2
Since series expansions are useful for approximating functions (4) we can
rewrite the energy function in one dimension like this:
1
V (x) = V (0) + V 0 (0) · x + V 00 (0) · x2 + . . . . (6.7)
2
Making the function vary in a quadratic fashion the series is truncated after
the second derivative. If zero corresponds to a minimum of the energy landscape
it is

32
Figure 6.12: The motion of a molecule around an energy minimum can be ap-
proximately described by a parabolic energy profile. This is the reason why
one has to generate the energy-minimized structure (green ball), which is lo-
cated around a minimum of the energy surface, before starting a normal mode
calculation

1
V (x) = V (0) + V 00 (0) · x2 , (6.8)
2
and with

d2 x
m = F = −V 0 (x) (6.9)
dt2
we obtain

F = −V 00 (0) · x. (6.10)
The solution of the second-order differential equation

d2 x V 00 (0)
= x (6.11)
dt2 m
k
is x = A · exp(iωt)with the angular velocity ω = m and the amplitude A.
Fig.6.11 actually illustrates the vibrational motion of a particle in one dimension
(red path). Fig.6.13 shows this kind of motion in two dimensions. To picture it
in three dimensions is more difficult.
In three dimensions the frequencies of the normal modes together with the
displacements of the individual atoms may be calculated from a molecular me-
chanics force field using the Hessian matrix of second derivatives (V). The
Hessian must first be converted to the equivalent force-constant matrix in mass-
weighted coordinates (F), as follows:
1 1
F = M 2 V00 9M 2 . (6.12)

33
Figure 6.13: Vibrations in 2-dimensional space. In reality one more dimension
comes into play

M is a diagonal matrix of dimension 3N ×3N , containing the atomic masses. All

elements of M are zero except those on the diagonal; M1,1 = M2,2 = M3,3 = m1 ,
M4,4 = M5,5 = M6,6 = m2 , ..., M3N −2,3N −2 = M3N −1,3N −1 = M3N,3N = mN .
1
Each non-zero element of M 2 is thus the inverse square root of the mass of the
appropriate atom. The masses of the atoms must be taken into account because
a force of a given magnitude will have a different effect upon a larger mass than
a smaller one. For example, the force constant for a bond to a helium atom is,
to a good approximation, the same as to a hydrogen, yet the different mass of
the helium gives a different motion and a different zero-point energy. The use
of mass-weighted coordinates takes care of these problems.
Strictly mathematically, the above term for the force-constant matrix comes
from a coordinate transformation where the valid term

Mẍ = −V00 x (6.13)

− 21
has been matrix multiplied with the square matrix M , thus leading
1 1 − 21 1
M 2 ẍ = M 2 V00 M M 2 x. (6.14)
We next have to obtain the eigenvalues and eigenvectors of the matrix F
which then have to be retransformed to the original coordinate system in or-
der to correspond to the desired modes of motion. Calculating eigenvalues and
eigenvectors is usually performed using matrix diagonalization [25]. If the Hes-
sian is defined in terms of Cartesian coordinates, then six of these eigenvalues
will be zero as they correspond to translational and rotational motion of the
entire system. The frequency of each normal mode is then calculated from the
eigenvalues λi using the relationship
√
λi
νi = , (6.15)
2π
√ k
as λ corresponds to m and since the solution of the second-order differential
ω
equation is an oscillating function, the frequency ν equals 2π . The ratio of the
amplitudes (xi ) of each normal mode are given by the eigenvectors ηi which
themselves characterize the mode of motion.

34
Figure 6.14: A linear triatomic molecule like CO2 . The vectors (here: scalars)
x1 , x2 , x3 define displacements of the corresponding atoms.

As a simple example of a normal mode calculation consider the linear triatomic

system in Fig.6.14. We shall just consider motion along the long axis of the
molecule and that all three atoms have the same masses (m1 = m2 = m3 ).
The displacements of the atoms from their equilibrium positions along this axis
are denoted by xi . It is assumed that these are small compared to the values
of the equilibrium bond lengths and the system is harmonic with bond force
constants k. The potential energy is given by:
1 1
V = k · (x1 − x2 )2 + · (x2 − x3 )2 . (6.16)
2 2
We next calculate the first derivatives of the potential energy with respect
to the three coordinates x1 , x2 , x3 :
∂V
mẍ1 = − = −k · (x1 − x2 ); (6.17)
∂x1
∂V
mẍ2 = − = −k · (x2 − x2 ) − k · (x2 − x3 ); (6.18)
∂x2
∂V
mẍ3 = − = −k · (x3 − x2 ); (6.19)
∂x3
And the Hessian
 
k −k 0
V00 =  −k 2k −k  , (6.20)
0 −k k
the mass-weighted matrix
 
m 0 0
M =  0 m 0 , (6.21)
0 0 m
and the force-constant matrix
k −k
 
m m 0
−k 2k −k
F= m m m
. (6.22)
−k k
0 m m

The eigenvalues and eigenvectors of F dont need to be retransformed in this

example because of the shape of M (all atoms have the same masses). The
eigenvalues, each corresponding to a different mode of motion, are:

35
Figure 6.15: Results of normal mode calculation for a linear triatomic molecule.
λi , ηi and xi describe eigenvalues, eigenvectors and amplitudes.

Figure 6.16: Protein molecules are the most examined type of molecule with
respect to vibrational motion. Obtaining the normal modes of motion, one can
notice a difference between the motional frequency of bigger (global) parts and
smaller (local) parts of the molecule, e.g. a whole domain or even just an atomic
link between two distinct atoms. Global motions of a protein are often specific
to it and can be related to its function.

k k
λ1 = 0, λ2 = andλ3 = 3 (6.23)
m m
Now each of the three frequencies can be obtained as shown above. They
correspond to modes of a translation, a symmetric stretch and an asymmetric
stretch respectively (Fig.6.2). In this triatomic example the three-dimensional
eigenvector of each mode determines the displacement of each atom.
The motions of larger segments of an analyzed protein, for example, are ade-
quate for studying the molecules function. These motions are represented by
lowfrequency modes. High frequency modes correspond to motions of smaller
molecule parts (Fig.6.16).
The harmonic approximation to the energy surface is found to be appropriate
for welldefined energy minima such as the intramolecular degrees of freedom of
small molecules. For larger systems the harmonic approximation breaks down.
Such systems also have an extraordinarily large number of minima on the energy
surface. In these cases it is not possible to calculate accurately thermodynamic
properties using normal mode analysis. Rather, molecular dynamics simula-
tions or other methods must be used to sample the energy surface from which
properties can be derived.

36
Chapter 7

What Else is Possible

Other analyses exploring the energy surface focus on determining reaction path-
ways or transition structures of molecules. Since the minimum points of the
energy landscape may correspond to the reactants or products of a chemical re-
action or two important conformations of a molecule (Fig.7.1), the path between
those two minima (the ‘reaction path’ or ‘pathway’) might be of interest. The
transition structure is the point of highest potential energy along the reaction
pathway.
As one can imagine many methods have been worked out for elucidating re-

Figure 7.1: A saddle point of the multidimensional reaction path corresponds

to the transition state between reactants and products.

37
action pathways and finding transition structures. These structures correspond
to saddle points with one negative eigenvalue of the Hessian matrix, where the
energy passes through a maximum for movement along the reaction path. Cal-
culational methods for locating transition structures do so by searching along
the reaction pathway, e.g. by using minimization algorithms when provided with
an initial structure close to the wanted one. Conversely, methods for finding
the reaction pathway start from the transition structure and move downwards
using minimization. Yet other methods determine both simultaneously from
the two minima bordering the reaction path. In particular, the conjugate peak
refinement [26] is a good method for locating transition structures for systems
with many atoms where a number of such states between two conformations
may be.
Besides the molecular dynamics simulation, there is another widely spread com-
puter simulation technique we would like to mention here: The Monte Carlo
(MC) method which differs in some ways from the MD method. The Monte
Carlo simulation method also provides a picture of the system in different con-
formations. However, it does not show how the system switches between these
since the behavior of atomic and molecular systems cannot be determined with
respect to processing time. A Monte Carlo simulation generates configurations
of a system by randomly changing the positions of the atoms present and so the
outcome of each calculation depends only on its preceding one. Furthermore, the
total energy is determined directly from the potential energy function. These
two points are in contrast to a molecular dynamic simulation where Newtons
equations of motion are the basis. Nonetheless, thermodynamic quantities can
be derived using appropriate statistical mechanics formulae.

38
Appendix A

Acknowledgements

We would like to thank the group of ‘Biocomputing’ at the ‘Interdisciplinary

Center for Scientific Computing’, Heidelberg, and especially Lars Meinhold and
Dr. Vandana Kurkal for organizing the F-Practical (‘Vacuum Simulations of
BPTI’) and supporting our work.
Moreover, we would like to express our gratitude towards Dr. Moritz Diehl, who
supervised us during the seminar ‘Mathematical Methods in Bioinformatics’
which provided the basis for this report.

39
Bibliography

[1] Brooks B.R., Bruccoleri R.E., CHARMM: A Program for

Macromolecular Energy, Minimization, and Dynamics Calcu-
lations, Journal of Computational Chemistry, 1983: 4, 187
[2] Leach A.R., Molecular Modelling, Pearson Prentice Hall, Har-
low, GB, 2001
[3] http://www.ch.embnet.org/MD tutorial/index.html
[4] Diehl M., Fischer T., Skript zur Mathematik fuer die Moleku-
lare Biotechnologie, Universitaet Heidelberg, D, 2003, Chapter
3.8

[5] McQuaerrie D.A., Statistical Mechanics, University Science

Books, Sansalito, USA, 2000
[6] Born M., Oppenheimer R., Zur Quantentheorie der Molekeln,
Annalen der Physik, 1927, 84: 457
[7] Allinger N.L., MM2: A Hydrocarbon Force Field Utilizing V1
and V2 Torsional Terms, Journal of the American Chemical
Society, 1977, 99: 8127
[8] Allinger N.L., Li F., Yan L., The MM3 Force Field for Alkenes,
Journal of Computational Chemistry, 1990, 11: 848
[9] Allinger N.L., Li F., Yan L., Molecular Mechanics (MM3) Cal-
culations on Conjugated Hydrocarbons, Journal of Computa-
tional Chemistry, 1990, 11: 868
[10] Allinger N.L., Yuh Y.H., Lii J.-J., The MM3 Force Field for
Hydrocarbons I, Journal of the American Chemical Society,
1989, 111: 8551
[11] Allinger N.L., Chen K., Lii J.-H., An Improved Force Field
(MM4) for Saturated Hydrocarbons, Journal of Computational
Chemistry, 1996, 17: 642
[12] Allinger N.L., Chen K., Katzenelenbogen J.A., Wilson S.R.,
Anstead G.M., Hyperconjugative Effects on Carbon-Carbon
Bond Lengths in Molecular Mechanics (MM4), Journal of
Computational Chemistry, 1996, 17: 747

40
[13] Nevins N., Chen K., Allinger N.L., Molecular Mechanics
(MM4) Calculations on Alkenes, Journal of Computational
Chemistry, 1996, 17: 669
[14] Nevins N., Chen K., Allinger N.L., Molecular Mechanics
(MM4) Calculations on Conjugated Hydrocarbons, Journal of
Computational Chemistry, 1996, 17: 695
[15] Nevins N., Chen K., Allinger N.L., Molecular Mechanics
(MM4) Vibrational Frequency Calculations for Alkenes and
Conjugated Hydrocarbons, Journal of Computational Chem-
istry, 1996, 17: 730
[16] Vollhardt K.P.C., Schore N.E., Organic Chemistry, Freeman,
1999, Chap: 2.5-2.7
[17] Cox S.R., Williams D.E., Representation of Molecular Electro-
static Potential by a New Atomic Charge Model, Journal of
Computational Chemistry, 1981, 2: 304
[18] Fowler P.W., Buckingham A.D., Central or Distributed Mul-
tipole Moments? Electrostatic Models of Aromatic Dimers,
Chemical Physics Letters, 1991, 176: 11
[19] Halgren T.A., Representation of van der Waals (vdW) Inter-
actions in Molecular Mechanics Force Fields: Potential Form,
Combination Rules, and vdW Parameters, Journal of the
American Chemical Society, 1992, 114: 7827

[20] Stote R.H., States D.J., On the Treatment of Electrostatic

Interactions in Biomolecular Simulation, Journal de Chimie
Physique, 1991, 88: 2419
[21] Greengard L., Rokhlin V., A Fast Algorithm For Particle Sim-
ulations, Journal of Computational Chemistry, 1987, 73: 325
[22] Krabs W., Modern Methods of Optimization, Springer, Hei-
delberg/Berlin, D, 1990
[23] Bhm H.-J., Klebe G., Kubinyi H., Wirkstoffdesign, Spektrum
Heidelberg, D, 1996
[24] Atkins P.W., Physikalische Chemie, Wiley-VCH, Berlin, D,
2001,Chapter 16.4
[25] Diehl M., Fischer T., Skript zur Mathematik fuer die Moleku-
lare Biotechnologie, Universitaet Heidelberg, D, 2003, Chapter
5.2, 5.3

[26] Fischer S., Karplus M., Conjugate Peak Refinement: An Al-

gorithm for Finding Reaction Paths and Accurate Transition
States in Systems with Many Degrees of Freedom, Chemical
Physics Letters, 1992, 194: 252

41