< <

The Flying Ice Cube: Velocity

Rescaling in Molecular Dynamics
Leads to Violation of
Energy Equipartition
STEPHEN C. HARVEY,
1
ROBERT K.-Z. TAN,
1
THOMAS E. CHEATHAM III
2,
*
1
Department of Biochemistry and Molecular Genetics, University of Alabama at Birmingham,
Birmingham, Alabama 35294
2
Department of Pharmaceutical Chemistry, University of California, San Francisco, California 94143
Received 31 July 1997; accepted 4 December 1997
ABSTRACT: This article describes an unexpected phenomenon encountered
during MD simulations: velocity rescaling using standard protocols can
.
systematically change the proportion of total kinetic energy KE found in
motions associated with the various degrees of freedom. Under these conditions,
the simulation violates the principle of equipartition of energy, which requires a
mean kinetic energy of RTr2 in each degree of freedom. A particularly
pathological form of this problem occurs if one does not periodically remove the
.
net translation of and rotation about the center of mass. In this case, almost all
of the kinetic energy is converted into these two kinds of motion, producing a
system with almost no kinetic energy associated with the internal degrees of
freedom. We call this phenomenon the flying ice cube. We present a
mathematical analysis of a simple diatomic system with two degrees of freedom,
to document the origin of the problem. We then present examples from three
kinds of MD simulations, one being an in vacuo simulation on a diatomic
system, one involving a low resolution model of DNA in vacuo, and the third
using a traditional all-atom DNA model with full solvation, periodic boundary
*Present address: National Institutes of Health, Bethesda, MD
20892-5626
Correspondence to: S. C. Harvey; e-mail: harvey@
neptune.cmc.uab.edu
Contractrgrant sponsor: National Institutes of Health; con-
tractrgrant number: HL-34343, CA-25644
Contractrgrant sponsor: Pittsburgh Supercomputing Cen-
ter; contractrgrant number: MCA935017P
( ) Journal of Computational Chemistry, Vol. 19, No. 7, 726740 1998
1998 John Wiley & Sons, Inc. CCC 0192-8651 / 98 / 070726-15
FLYING ICE CUBE
conditions, and the particle mesh Ewald method for treating long-range
electrostatics. Finally, we discuss methods for avoiding the problem.
1998 John Wiley & Sons, Inc. J. Comput Chem 19: 726740, 1998
Keywords: molecular dynamics; energy equipartition; macromolecular
simulations; constant temperature simulations
Introduction
.
1, 2
olecular dynamics MD simulations are
Mwidely used to investigate questions of
macromolecular structure and dynamics. Properly
parameterized, they can provide important kinetic
and thermodynamic information on a wide range
of problems.
The MD algorithm numerically integrates New-
tons equations of motion. Once the trajectory is
begun, a free MD simulation should correspond to
an ensemble that has constant total energy. There
are two problems with such simulations. First, the
gradual accumulation of unavoidable numerical
errors may lead to drifts in the total energy. Sec-
ond, experiments are usually conducted at con-
stant temperature and pressure, which corre-
sponds to the isothermicisobaric ensemble, not a
constant energy ensemble.
To correct for these problems, standard MD
packages offer different tools for maintaining con-
stant temperature. Among these are periodic ve-
locity rescaling and other methods that uniformly
modify the atomic velocities. For periodic velocity
rescaling, the total kinetic energy is determined at
regular intervals, and the instantaneous effect tem-
perature, T, is calculated by noting that the total
kinetic energy is 3NRTr2, where N is the number
of atoms, R is the universal gas constant, and the
factor of three arises from the fact that each atom

has three degrees of freedom. Strictly speaking,

this relationship only holds for the average kinetic
energy, calculated over a sufficiently long time
that the system represents the appropriate thermo-
.
dynamic ensemble. The ratio between the target
temperature and the instantaneous temperature is
used to calculate a factor by which all velocities
are multiplied. One can use a factor that instanta-
neously adjusts the temperature to the target value,
or one can use some fraction of this factor, treating
the system as if it were coupled to a temperature
bath with a characteristic relaxation time for the
transmission of energy between the bath and the
system.
3
We have found that periodic velocity rescaling
leads to an unexpected problem: a gradual bleed-
ing of KE from high frequency motions such as
bond stretching and angle bending into low-
frequency motions. This represents a violation of
the principal of equipartition of energy, which
requires that each degree of freedom has the same
mean kinetic energy. We made this discovery in-
dependently of one another on two very different
kinds of systems. In one, AMBER
4
was used to
simulate an all-atom model with full solvation,
whereas, in the other, in vacuo simulations were
carried out on a low-resolution model of DNA
using YAMMP.
5
A similar phenomenon was inde-

pendently observed at Wesleyan University M. A.

Young and D. L. Beveridge, personal communica-
.
tion during all-atom simulations with full solva-
tion. We therefore concluded that this problem is
inherent in velocity rescaling.
This article has three sections. In the first, we
present an analytical treatment of the effects of
velocity rescaling on a very simple model system.
This provides an introduction to the phenomenon
and a rigorous explanation for its cause. The sec-
ond section presents some manifestations of this
problem that we have observed during different
kinds of MD simulations. The final section dis-
cusses precautionary measures that can be taken to
avoid or minimize problems associated with veloc-
ity rescaling.
Analysis of Problem
Let us consider what happens during an MD
simulation with periodic velocity rescaling in a
simple one-dimensional system with two particles,
which we can treat analytically. The system has
two degrees of freedom. One of these corresponds
to the translational motion of the center of mass,
which has kinetic energy E . The other corre-
0
sponds to the vibrational motion.
Let us suppose that the potential energy of the
single internal degree of freedom, x, depends
quadratically on the deviation of x from its ideal
JOURNAL OF COMPUTATIONAL CHEMISTRY 727
HARVEY, TAN, AND CHEATHAM
value, x . The most common example of this is a
0

bond stretching term with force constant k. Fol-

lowing convention, we have incorporated the
.
Hookes law factor of 0.5 into k. We denote the
potential energy of this degree of freedom by U:
2
. . .
U x sk x yx 1
0
The kinetic energy for the internal degree of free-
dom is denoted by E , where the subscript indi-
i
cates internal. x and its derivatives are oscilla-
tory, and so is E . We denote the angular fre-
i
quency of the kinetic energy oscillations by . It is
easily shown that:
. .
E t s 1 qsin t
i
where indicates the mean kinetic energy of the
: .
internal degree of freedom, E . We use E t
i
.
without a subscript to indicate the total kinetic
energy:
. . .
E t sE qE t sE q 1 qsin t
0 i 0
and we designate the average kinetic energy by
E sE q. The purpose of periodically reassign-
0
ing or rescaling velocities is, of course, to remove
drifts in E, so that, over the duration of the MD
simulation, the mean kinetic energy is close to the
value required by the temperature specified for
the simulation. This requires that:
n
E s RT
/
2

where n is the number of degrees of freedom two

.
in the present example , and the double bar indi-
cates an average over the duration of the MD
simulation.
In the discussion that follows, we will assume
that rescaling follows the original procedure devel-
oped about 20 years ago and used in many pro-
grams. In this protocol, the instantaneous kinetic
energy is calculated, and then velocities are re-
'
scaled by a factor , where is equal to the
desired mean kinetic energy divided by the instan-
taneous kinetic energy. In the present situation,
with two degrees of freedom:
sRTrE
Note that, because E is time-dependent, also
depends on time.
The results described in what follows are no
different if, rather than instantaneously jumping
the kinetic energy all the way to its target value,
rescaling is more gradual and is based on coupling
to a temperature bath.
3
The rescaling factor in the
coupling protocol is only a fraction of the value of
specified here, but the direction of rescaling
.
increasing or decreasing velocities is the same,
and the effect over very long simulations is also
the same.
The effect of rescaling is to immediately give
the system an instantaneous kinetic energy corre-
sponding to the target temperature. However, even
in the absence of numerical errors, the mean ki-
netic energy during the interval of free MD from
this rescaling to the next rescaling will not be
equal to RT. This is because rescaling can happen
at any point in the cycle of the intramolecular
motion, and is calculated from the instantaneous
value of the kinetic energy. The system has con-
stant total energy, but kinetic energy oscillates. If
the rescaling occurs when the internal coordinate
is far from its ideal value, the kinetic energy will
.
be small and will be larger , whereas, if rescal-
ing happens when the internal coordinate is near
its ideal value, the kinetic energy will be large
.
and will be smaller .
One would expect that, if the kinetic energy
oscillates about a value near RT, there should be
about a 50% probability that the instantaneous
kinetic energy is less than RT, so that )1, and
the probability that -1 would also be nearly
50%. Furthermore, because E and E are scaled
i 0
by the same amount, one would intuitively expect
that there should be no change in the fraction of
total kinetic energy that is found in the internal
degrees of freedom versus the kinetic energy asso-
ciated with the center of mass motion. In other
words, one would expect the principal of equipar-
tition of energy to be preserved. The purpose of
this work is to show that this is not the case.
This unexpected result arises from the fact that
the rescaling factor depends on the instanta-
neous value of the kinetic energy at the time of
rescaling, and has different effects on the two
kinds of kinetic energy. Let us use primes to desig-
nate the kinetic energies after rescaling, and angu-
lar brackets to denote expected values. First, the
expected value of the kinetic energy for the center
of mass motion is

X
: : : .
E s E s E 2
0 0 0
where the removal of E from the angular brackets
0
is possible because the center of mass motion is
uniform throughout the oscillatory cycle. In con-
trast, the kinetic energy associated with the inter-
VOL. 19, NO. 7 728
FLYING ICE CUBE
nal degree of freedom has expectation value:

X
: : .
E s E 3
i i
and E cannot be taken outside the angular brack-
i
ets, because it is time-dependent.
depends on the point in the cycle at which
rescaling takes place. So, to calculate expectation
values, we consider to be a time-dependent
variable and derive the expectation values by av-
eraging over one full cycle of the bond vibration.
Thus:
RT RT
.
t s s
.
E q 1 qsin t E q sin t
0
RT 1
.
s 4
1 qc sin t E
where we have introduced the abbreviation c s
rE. Using t as the variable of integration and
averaging over one cycle, we get:
1 RT 1
2
: .
s d t
H
.
2 1 qcsin t E 0
RT
y1r2
2
.
s 1 yc
E
.
Combining this result with eq. 2 gives, on rear-
ranging:

X
:
E RT
y1r2 0
2
. .
s 1 yc 5
E E
0
This gives a hint at the problem: if the mean
.
kinetic energy is exactly on target i.e., if E sRT ,
we would anticipate that the kinetic energy of

X
: .
translation should be preserved i.e., E rE s1 ,
0 0
but this is not so. The scaling factor, which should

2
.
y1r2
be equal to one in this case, is actually 1 yc .
Because c is always less than one, the expectation

X
:
value after rescaling, E , is larger than E , at
0 0
least when E sRT. This results in an increase in
the kinetic energy of translation.
Let us now calculate the expectation value for
the kinetic energy associated with the internal de-
. .
gree of freedom. Substituting eqs. 1 and 4 into
.
eq. 3 gives:
.
RT 1 qsin t
X
: :
E s E s
i i
;
1 qc sin t E
Dividing through by the mean kinetic energy ,
and calculating the average over one full cycle
gives:

X
: : .
E E 1 RT 1 qsin t
2
i i
.
s s d t
H
: .
E 2 1 qc sin t E 0
i
.
6
We are unable to evaluate this integral analyti-
cally, but it can be integrated numerically using

Mathematica Wolfram Research, Inc., Champaign,

.
IL . When divided by the normalization factor 2,
a value less than one is always obtained. If the
.
mean energy were right on target i.e., if E sRT ,
then we would expect that the kinetic energy asso-
ciated with the internal degree of freedom should
: .
remain unchanged i.e., E r s1 , but this is
i
not the case: this component of the kinetic energy
would actually be reduced.
Table I compares the effects of rescaling on the
.
translational kinetic energy, eq. 5 , with its effect
on the kinetic energy of the internal degree of
.
freedom, eq. 6 . The independent variable is the
fraction of the total kinetic energy associated with
the internal degree of freedom over the time inter-
val since the previous rescaling, c srE. For all
TABLE I.
Effect of Rescaling on Distribution of Kinetic Energy
As a Function of Fraction of Total Kinetic Energy
Associated with Internal Degree of Freedom
( )
c = // /// E .
X X
: :
E E
i 0
c
:
E E
i 0
0.01 0.99 1.00005
0.1 0.95 1.005
0.2 0.90 1.02
0.3 0.85 1.05
0.4 0.79 1.09
0.5 0.73 1.15
0.6 0.67 1.25
0.7 0.59 1.40
0.8 0.50 1.67
0.9 0.37 2.29
0.99 0.13 7.09
On rescaling, the kinetic energy of the internal degree of
( ) freedom, E , is decreased column 2 , whereas the kinetic
i
( energy of the overall translation, E , is increased column
0
) 3 . Unprimed and primed values are prior to and immedi-
ately after rescaling respectively. Angular brackets indicate
( ) either the mean value prior to rescaling or the expectation
( ) : value after rescaling . E is not oscillatory, so E sE .
0 0 0
The boldface values at c =0.5 indicate what would happen
if equipartition were instantaneously satisfied prior to rescal-
ing. See text for discussion of this case.
JOURNAL OF COMPUTATIONAL CHEMISTRY 729
HARVEY, TAN, AND CHEATHAM
values of c over the physically accessible range,
0 -c -1, there is a net conversion of kinetic en-
ergy from the internal degree of freedom to the
translation of the center of mass.
With repeated rescalings, there will be a grad-
ual loss of vibrational kinetic energy, and a corre-
sponding increase in the translational kinetic en-
ergy. Eventually, the amplitude of the vibrations
shrinks to nearly zero, as does the effective tem-
perature for the internal motions, and the molecule
becomes a flying ice cube. This phenomenon is
demonstrated in the simulation on the model di-
atomic system in the next section.
In more complex systems, repeated velocity
rescaling will produce a gradual loss of kinetic
energy associated with the high-frequency mo-
tions, and a concomitant increase in kinetic energy

associated with the zero frequency motions global

.
translation and rotation . Other degrees of freedom
may also experience a gradual buildup of kinetic
energy, if the periods of their characteristic vibra-
tions are long by comparison to the time interval
between successive rescalings. For such low fre-
quency motions, the expectation value for the rela-
tive energy after rescaling is given by an expres-
.
sion resembling eq. 6 , whereas the expectation
value for higher frequency motions resembles eq.
.
5 . Thus, a given degree of freedom will see a net
gain or loss of kinetic energy after many rescal-
ings, depending on whether it is a low frequency
or high frequency motion, relative to the frequency
of rescaling.
Examples from MD Simulations
MODEL DIATOMIC SYSTEM: ETHANE
( )
UNITED ATOM MODEL
We have carried out MD simulations on a united
atom model for ethane using the YAMMP pack-
age.
5
Because the simulations were done in three-
dimensional space, the system has six degrees of
freedom. As a consequence, the mathematical
analysis done in the previous section does not
apply exactly, but the net transfer of kinetic energy
from the high-frequency vibrational motion to the
zero frequency motions is clearly demonstrated.
The model consists of two pseudoatoms, each
representing a methyl group and having a mass of
.
15 amu. The force field parameters of eq. 1 are
.
x s1.54 A ideal bond length and k s240
0

2
.
kcalrmol A bond force constant . With a target
temperature of 300 K and six degrees of freedom,
the total target kinetic energy is 6RTr2 s1.8
kcalrmol.
The period of free vibration for this system is:
' s2 rk

where is the reduced mass, sm m r m q

1 2 1
.
m smr2, and msm sm is the mass of the
2 1 2
methyl group. Substituting the above values into
these relationships gives s38.4 fs.
A timestep of 1 fs was used in the simulations,
and velocity rescaling was done at intervals of 100
fs. This interval is longer than the period of vibra-
tion and is not simply related to the period, so
that, in a very long simulation, rescaling occurs at
all possible points in the oscillation cycle. The
effect of rescaling on the vibrational kinetic energy
.
is therefore described by eq. 6 .
Figures 13 show the time dependence of the
bond length, kinetic energy, and potential energy,
respectively. If we examine only the kinetic en-
ergy, it appears that rescaling is doing exactly
what it should: depending on the exact point in
the cycle where rescaling occurs, the mean kinetic
energy during the next free interval of MD may be
larger or smaller than in the previous interval, but
the mean kinetic energy, averaged over the total
trajectory, will converge to 1.8 kcalrmol after many
rescalings. On the other hand, the energy associ-

ated with the internal degree of freedom bond

.
stretching is gradually reduced, as is seen in the
declining amplitude of the bond length vibration
.
Fig. 1 and in the decaying mean potential energy
.
Fig. 3 . Note that there are occasions when the
rescaling causes the internal energy to increase
.
slightly at t s0.1, 0.2, 0.5, 0.9 ps , but when it
.
drops, it often does so sharply at t s0.3, 0.8 ps .
Lest there be any doubt that the bond vibra-
tional energy continues to decline, we have run
another simulation with identical parameters, but
a different initial random number seed, and saved
the structures over five 200-fs windows, with suc-
cessive windows separated by 100 ps. A plot of
.
bond length versus time for this simulation Fig. 4
shows that the amplitude of the bond vibration is
sharply reduced after only 100 ps, and it is essen-
tially zero after about 300 ps. There is no vibra-
tional kinetic energy, and all the kinetic energy is
in the translation of the center of mass and rotation
about the center of mass. Note that the mean bond
.
length 1.543 A , is larger than the ideal bond
.
length 1.540 A , because of the centrifugal force of
the molecular rotation.
VOL. 19, NO. 7 730
FLYING ICE CUBE
FIGURE 1. Time dependence of the bond length of the ethane model with two united atoms, during an MD simulation
with periodic rescaling. The gradual decay of the vibrational amplitude is evident. The period of the oscillation is 38.4 fs.
The overall molecular rotation produces a centrifugal force that tends to stretch the bond, so the mean bond length is

slightly greater than the ideal bond length of 1.54 A.

FIGURE 2. Time dependence of the kinetic energy of the ethane model. The KE goes through two peaks and two
(
valleys during each oscillatory cycle. The maxima occur when the potential energy goes to zero i.e., the bond length is
)
at its ideal value . The minima in KE occur at the extreme values of the bond length. Conservation of angular
momentum causes the rotational kinetic energy to be larger when the bond is fully compressed than when it is fully
( )
stretched. Compare the valleys in this figure to the peaks and valleys in Fig. 1.
JOURNAL OF COMPUTATIONAL CHEMISTRY 731
HARVEY, TAN, AND CHEATHAM
FIGURE 3. Time dependence of the potential energy of the ethane model. The PE goes through two peaks and two
( )
valleys during each oscillatory cycle. PE =0 when the bond length has its ideal value i.e., when x =x . The
0
centrifugal force of overall rotation causes the mean bond length to be greater than the ideal length, so the potential
(
energy is larger when the bond is stretched than it is when it is compressed. Compare the peaks in this figure to the
)
peaks and valleys in Fig. 1.
FIGURE 4. Time dependence of the bond length of the ethane model in an MD simulation covering 400.2 ps, with
periodic velocity rescaling. Five 200-fs windows at intervals of 100 ps are shown, and the rescaling event at the center
of each window is indicated. The vibrational energy is severely damped by 100 ps and is essentially zero by 300 ps.
VOL. 19, NO. 7 732
FLYING ICE CUBE
MORE COMPLEX ELASTIC SYSTEM:
SUPERCOILED DNA
In closing the section Analysis of the Problem,
we argued that, in addition to the buildup of
kinetic energy in zero frequency motions, energy
might also increase in other motions whose charac-
teristic frequencies are smaller than the frequency
of velocity rescaling. We have encountered inter-
esting examples of this phenomenon in simula-
tions on supercoiled DNA.
Double-helical DNA molecules in the size range
of several hundred to several thousand basepairs
.
bp are elastic polymers. Given the elastic moduli
of DNA, it is easier to bend the molecule than it is
to twist it. As a consequence, when twisting defor-
mations are applied to DNAs in this size range,
they are inclined to relieve the torsional stresses by
bending, producing nonplanar writhed configura-
tions.
A relaxed closed circular DNA is formed when
the molecule is covalently closed into a circle in
such a way that no torsional stress is introduced;
the equilibrium conformation is circular, and the

linking number the number of times the two

.
strands wrap around one another has its optimal
value, Lk . If the molecule is closed into a circle
0
.
with any other linking number i.e., if Lk /Lk ,
0
then the equilibrium conformation is supercoiled. It
resembles a figure eight, except that there are
multiple crossings. The number of crossings is
approximately equal to the writhe, Wr, which is
proportional to the linking deficit. For a molecule

with the elastic properties of DNA, Wr f0.7 Lk y

.
Lk .
0
We have developed a reduced representation
model for double-helical DNA that is useful for
studying the structure of supercoiled molecules.
6
The model, called 3DNA, uses three pseudoatoms
to define the plane of each basepair. Successive
basepairs are held together by pseudobonds, and

all of the terms in the force field are harmonic the

.
energy depends quadratically on the coordinate .
These include bonds, angles, improper torsions,
and nonbonded volume exclusion terms. Force
constants for the bonds, angles, and torsions were
chosen to mimic the known elastic moduli for
DNA bending and twisting.
This is thus a complex elastic model, and the
many degrees of freedom have a wide range of
characteristic frequencies. If velocities are rescaled
often enough, the analysis of the previous section
would predict a gradual transfer of kinetic energy
from the high-frequency vibrational motions to the
zero-frequency and low-frequency degrees of free-
dom.
We have carried out a series of long molecular
.
dynamics simulations 500 ns or more on large
closed circular DNAs. For both relaxed and super-
coiled molecules, a wide range of conformations is
observed when the simulations are run using free
MD, MD with infrequent velocity rescalings, or
MD with infrequent velocity reassignments. Any
of these procedures generates a thermodynamic
ensemble of structures whose average properties
generally differ significantly from those of the
equilibrium structure.
7
In particular, relaxed
molecules almost never show an open, circular
conformation, and interwound molecules are rarely
as extended as their equilibrium structures.
The behavior changes drastically when frequent
velocity rescaling is used, as the kinetic energy
from hundreds or thousands of high-frequency
vibrational modes is bled into the zero-frequency
and low-frequency motions. Here we describe re-
sults of MD with coupling to a constant tempera-
ture bath, following the algorithm of Berendsen et
al.
3
Because of the large masses and small force
constants of the 3DNA model, the simulations
described here use a timestep of 31.25 fs. Berend-
sen rescaling is done at every timestep, with a
characteristic time for coupling to the temperature
.
bath of 1 ps 32 timesteps .
For a relaxed closed circular molecule, the
buildup of rotational kinetic energy causes the
structure to open into a spinning circular confor-
mation, like a wheel. The behavior of underwound
molecules is determined partly by the magnitude

of the linking deficit how tightly the molecule is

.
interwound , and partly by chance the relation-
ship between the initial distribution of velocities
.
and the gradually accumulating centrifugal force .
Sometimes these molecules open into a spinning
circular conformation, like the relaxed molecules;
this is particularly likely if the linking deficit is
small. More highly underwound molecules are
prone to developing extended interwound confor-
mations, closely resembling the equilibrium struc-
ture. An example of this is shown in Figure 5.
With 1500 bp and 4500 pseudoatoms, the
molecule in Figure 5 has 13,500 degrees of freedom
.
df . After several hundred nanoseconds, the ki-
netic energy associated with all of these is mostly
converted into global translational and rotational
motions, which comprise only 6 df. But there is
one other zero-frequency mode that is also excited
JOURNAL OF COMPUTATIONAL CHEMISTRY 733
HARVEY, TAN, AND CHEATHAM
FIGURE 5. Evolution of the structure of model supercoiled DNAs in long MD simulations with coupling to a constant
temperature bath. The model has 1500 bp. The expected writhe is about 70% of the linking deficit. This is indeed the
( )
case after a long time for two cases Lk =y9 and Lk =y12 , but the rotational motion of the molecule with
Lk =y4 causes it to open into a circular form. After several hundred nanoseconds, almost all the kinetic energy in all
three models is associated with the translation of the center of mass and rotational motion about the center of mass.
by the rescaling. This motion, often called slither-
ing, is best described by considering one of the
.
interwound configurations at t s500 ns Fig. 5 .
Although the conformation appears fixed, the po-
sition of a given basepair within the structure is
time-dependent, and slithering refers to the mo-
tion of a given basepair around the track defined
by the interwound conformation. A formal way of
following the slithering motion is the loop tip pro-
file, which plots the numbers identifying the two
basepairs at the two ends of the interwound struc-
ture, as a function of time.
8
Loop tip motions in real DNA will have an
irregular, diffusive character. Loop tips appear and
disappear, due to changes in the global folding,
and the motion of a given loop tip will be diffusive
as long as the loop persists. Such behavior is seen
in MD simulations using our model if velocities
are only occasionally rescaled, or if velocities are
reassigned.
8, 9
Repeated velocity rescaling changes
.
this behavior, however Fig. 6 . After many rescal-
ings, there is sufficient kinetic energy associated
with the slithering degree of freedom that the
motion becomes persistent and monotonic. The
basepairs race around the interwound configura-
tion like a long toy train on an interwound track.
We have observed another interesting artifac-
tual behavior in interwound DNA models with
600 bp, when the bending elastic modulus of a
120-bp segment is substantially reduced relative to
that of normal DNA. This model mimics a closed
circular DNA with a 120-bp insert of a triplet
repeat sequence, because such sequences have a

smaller persistence length lower bending modu-

.
10
lus than normal DNA. Two examples are shown
in Figure 7. In each of them, the discontinuity in
elastic properties at the boundary between the
insert and the rest of the DNA creates an obstacle
VOL. 19, NO. 7 734
FLYING ICE CUBE
FIGURE 6. Slithering motion in the model supercoiled DNAs of Figure 5. The initially random, diffusive slither
eventually gives way to persistent, monotonic slithering motion along the interwound configurations for Lk =y9 and
Lk =y12. There are no defined loop tips for Lk =y4 after the molecule transforms into an open circle.
to the slithering motion. Slithering proceeds mono-
tonically, as in Figure 6, until this barrier is hit, at
which point the direction of slither reverses. The
reversal is classified as a low-frequency motion,
because it has a characteristic time on the order of
10 ns, which is much larger than the characteristic
time used to couple the system to the temperature
.
bath 1 ps .
JOURNAL OF COMPUTATIONAL CHEMISTRY 735
HARVEY, TAN, AND CHEATHAM
FIGURE 7. Slithering motions in interwound closed circular DNAs with 600 bp, containing a more flexible 120-bp

insert. The persistence length of the 480 bp of normal DNA is 500 A, whereas that of the insert is 300 A. After an initial
period characterized by slithering motions with a random, diffusive character, energy and momentum build up in this
degree of freedom. The molecule slithers monotonically and with essentially constant velocity until the boundary
between the two domains of differing elasticity reaches the loop, at which point the slithering usually reverses direction.
ALL-ATOM MOLECULAR DYNAMICS
SIMULATIONS: CONDENSED PHASE
SIMULATIONS WITH PERIODIC BOUNDARY
CONDITIONS
Despite an increase in the number of interacting
particles and therefore a greater number of means
to couple energy between internal degrees of free-
dom, all-atom molecular dynamics simulations of
condensed phase systems still offer examples of
the flying ice cube syndrome. As discussed in the
previous sections, if velocity scaling occurs at an
interval longer than the period of any given vibra-
tion in the system, energy from these high-
frequency modes may be transferred into lower
frequency modes. This is probably not a major
problem in the simulation of condensed phase
systems as long as the energy of these modes can
couple back into the system; for example, through
collisions among the molecules and if there is a
large enough number of low-frequency modes that
the drain of energy into each mode is very small.
Unfortunately, many of the methods employed
in the simulation of condensed phase systems do
not conserve energy in common usage unless par-
ticular care is taken. Specific examples include
infrequent pairlist updates in simulations where
the electrostatic interactions are subject to a cutoff,
cooling due to integration timesteps that are too
VOL. 19, NO. 7 736
FLYING ICE CUBE
large, SHAKE tolerances that are not low enough,
and Berendsen pressure coupling,
3
among other
possible sources. Cutoffs applied to the van der
Waals interactions and infrequent pairlist updates
can also lead to heating or cooling.
Small energy drains in the system can lead to
disastrous effects on the expected dynamics when
any method that uniformly scales velocities is ap-
plied to control temperature. Gradually over the
course of the simulation, the drain into low-
frequency modes continues until essentially all of
the motion is stored in these modes. This problem
is particularly acute if the center of mass transla-
tional motion is not periodically removed. Because
the center of mass translational energy cannot cou-
ple back into the system, the center of mass trans-
lational energy will grow with repeated rescaling.
This leads to the flying ice cube and the trapping
of the system in a local energy minimum.
We have previously reported all-atom MD sim-
w x
ulations on the DNA duplex d CCAACGTTGG
2
run with 18 sodium ions in water,
11, 12
and we
show here that energy equipartition can be vio-
lated with frequent velocity rescaling. The simula-
tions were run using the particle mesh Ewald
.
PME method within AMBER 4.1, with repeating
boundary conditions. The simulation was run with
cubic B-spline interpolation, a grid spacing of ;1

A, an Ewald coefficient of ;0.31, SHAKE on

.
hydrogens tolerance 0.0005 , a pairlist update ev-

ery 10 steps, and a cutoff of 9 A. Constant temper-

ature and pressure were maintained by coupling
.
to a bath coupling time 0.2 ps, 1 atm, 300 K ,
following the Berendsen method.
3
The simulation
represents approximately 9300 atoms in a periodic

3
cell of ;55 =41 =41 A .
Figure 8 shows the atomic fluctuations over 100
ps windows for all the DNA atoms in the simula-

tion. The fluctuations over the first 100 ps dotted

.
line and the 100-ps interval after the first nanosec-
.
ond not shown are comparable. However, after
2.3 ns, the fluctuations on most of the atoms drop
.
by nearly half solid line , and much of the kinetic
energy is associated with translation of the center
of mass. Note that not all atoms have their motion
damped: four peaks are evident for atoms whose
motion is relatively unaffected. These represent
the hydrogens on the thymine methyl groups.
Methyl rotation is opposed by only relatively weak
forces, so the frequency of rotational oscillations of
methyl groups is low. As the energy gets trans-
ferred into low-frequency modes, the methyl
groups spin faster and faster.
FIGURE 8. Root-mean-square amplitude of fluctuations
in DNA atomic positions, as a function of atom number,
for an all-atom MD simulation with periodic velocity
rescaling. The DNA molecule has 632 atoms. Heating
and equilibration are completed prior to time t =0. The
fluctuations over the first 100 ps of the simulation after
equilibration are represented by the dashed line, whereas
the solid line indicates the fluctuations over a 100-ps
interval that begins at t =2.3 ns. Note the loss of kinetic
energy in these degrees of freedom.
This system shows the flying ice cube phe-
nomenon in a traditional MD simulation with all-
atom representation. It provides examples of en-

ergy drains into a zero-frequency motion center of

.
mass translation and into a low-frequency mode
.
rotation of the methyl groups . It should be noted
that the effect is not due to some peculiarity in the
interaction between the macromolecule and the
solvent, because simulations on 125 TIP3P water
molecules lead to a complete drain into the center
of mass translational kinetic energy in ;1.5 ns
.
T. E. Cheatham and P. A. Kollman, unpublished .
Discussion
The MD algorithm numerically integrates New-
tons equations of motion. It is widely agreed that,
to avoid instabilities, dissipation, and systematic
JOURNAL OF COMPUTATIONAL CHEMISTRY 737
HARVEY, TAN, AND CHEATHAM
drifts, the Hamiltonian must be conservative, and
the integration method should be reversible and
symplectic.
13 16
The most common integrators,
such as the leapfrog
17
and velocity Verlet
18
algo-
rithms, meet these conditions, so that these prob-
lems are avoided, at least in principle, in the limit
of very small timesteps.
Nevertheless, there are a number of problems
that emerge in MD simulations. Despite small
timesteps, the gradual accumulation of unavoid-
able numerical errors can lead to drifts in the total
energy over very long simulations. More insidious
effects come from drifts in the total energy due to
systematic errors in the simulation. These gener-
ally relate to methodological inaccuracies, such as
improper truncation of the long-range inter-
actions,
19
and to tradeoffs that arise in methods
designed to speed up the calculations, such as the
use of low-order multipole expansions for long-
range electrostatics.
20
In addition, the experiments
with which simulations are compared are nor-
mally conducted at constant temperature and pres-
sure, which corresponds to the isothermicisobaric
ensemble, not the constant energy ensemble of a
free MD simulation.
These problems are usually attacked in MD
simulations by making occasional adjustments to
the atomic velocities. Velocity rescaling is one ap-
proach, whether by instantaneously adjusting the
temperature to the target temperature,
1, 2
or by
coupling to a temperature bath,
3
which represents
a gradual relaxation back toward the target tem-
perature. In many laboratories, these procedures
are done routinely, following historical protocols,
without concern about rigor or possible side ef-
fects. With regard to rigor, rescaling is often done
whenever the instantaneous temperature departs
from the target value by some margin; that margin
is often set without calculating the largest fluctua-
tion that would be statistically expected to occur
for the system being modeled. With regard to side
effects, the analysis and examples presented here
demonstrate that velocity rescaling can introduce
artifacts in the structural, dynamic, and energetic
properties of the system.
The phenomenon that we have described is
distantly related to the temperature echoes first
described by Grest et al.
21
and investigated by a
number of others.
22 25
In an equilibrated MD sim-
ulation, if the atomic velocities are instantaneously
.
set to zero quenched at time t , the phases of the
0
vibrational modes are all identical as the system
resumes its motion after the quench. If the dynam-
ics are allowed to evolve freely for a time t , and
1
.
then a second quench a quench echo is applied,
all vibrational modes lose kinetic energy except
those whose phase is zero or at the time of the
second quench. The unaffected modes are those
whose frequencies satisfy the relationship s
i i
nrt , where n is an integer. The first quench
1
creates coherent vibrations, whereas the second
removes kinetic energy fromall modes except those
whose periods are submultiples of t . As a conse-
1
quence, at times t st q2t , t st q3t , t s
0 1 0 1
t q4t , . . . there will be a dip in the kinetic en-
0 1
.
ergy a temperature echo . The approach has been
generalized to the inclusion of both cooling and
heating pulses,
23
and random but correlated sets of
velocities can be used for the two pulses to gener-
ate the phase correlations without perturbing the
system temperature.
25
Temperature echoes can be
used to examine normal modes, the density of
states, anharmonic effects, and other dynamic
properties in LennardJones glasses
21, 22
and in
macromolecular systems.
23, 24
The temperature echo phenomenon manifests
itself after only two velocity reassignments. Tradi-
tionally, zero velocities are reassigned,
21, 24
but
random velocities may be used as long as the two
sets of velocities are correlated.
25
In contrast, the
phenomenon reported here occurs after repeated
velocity rescalings, and there is a permanent shift
in the distribution of kinetic energy among differ-
ent modes, rather than any kind of echo. Our
analysis and examples show that this redistribu-
tion occurs whether the rescaling interval is quite

long or as small as one timestep i.e., when the

.
Berendsen temperature bath protocol is used .
Conclusions and Protective Measures
The gradual drain of kinetic energy from high-
frequency motions into zero- and low-frequency
motions is an unavoidable artifact of repeated ve-
locity rescaling. It has been observed in all-atom
simulations and in simulations using reduced rep-
.
resentations pseudoatoms , and the origin of the
problem is now understood. As seen in the simula-
tions described, this can greatly affect the struc-
tural, kinetic, and thermodynamic properties of
the system.
There are several possible approaches to pre-
vent the artifacts caused by periodic velocity
rescaling. Here we discuss three.
VOL. 19, NO. 7 738
FLYING ICE CUBE
VELOCITY REASSIGNMENT
Although velocity rescaling is very popular,
constant temperature can be maintained equally
well by other methods. Andersen
26
discussed pro-
tocols in which atomic velocities are periodically
reassigned from a Maxwellian distribution with
the appropriate temperature. This procedure, of-
fered in many MD packages, is equivalent to sub-

jecting each atom to a stochastic collision. A dis-

cussion of the subtle differences in the thermody-
namic ensembles defined by the velocity rescaling
and reassignment algorithms is given in Appendix
1
.
3 of the monograph by McCammon and Harvey.
Maintaining temperature by reassigning veloci-
ties, rather than rescaling, is one way to eliminate
the problems described in this study. Simulations
on supercoiled DNA with the 3DNA model do not
suffer from buildup of energy in low-frequency
modes if velocity reassignment is used to maintain

temperature D. Sprous, R. K.-Z. Tan and S. C.

.
Harvey, unpublished . As another example, simu-
lations of 125 TIP3P waters with velocity reassign-
ments showed no appreciable artifact during 10 ns

of simulation T. E. Cheatham and P. A. Kollman,

.
unpublished . We also note that frequent velocity
reassignment during the heating and equilibration
phases of MD simulations also helps to distribute
energy throughout the structure, avoiding poten-
tial hot spots due to nonuniform distribution of
structural distortions in the initial structure.
REMOVAL OF MOTION OF CENTER
OF MASS
Because velocity rescaling converts much of the
kinetic energy of the system into motion of the
center of mass, artifactual behavior can be reduced
by periodically removing the translational motion
of the center of mass, and rotation about it. Unfor-
tunately, this still leaves the simulation subject to
violation of energy equipartition, because low-
frequency modes can still be excited by repeated
velocity reassignments, as shown in both the all-
atom and reduced representation simulations on
DNA, described earlier. Furthermore, if only trans-
lational motion of the center of mass is removed,
the excitation of the zero-frequency rotational
mode around the center of mass can cause sub-
stantial distortion of the structure, due to the
growing centrifugal force. This is demonstrated by
the distortion of the mean bondlength in the di-
.
atomic molecule simulation Fig. 1 , by the asym-
metric character of the bondlength vibration in the
.
same simulation Figs. 24 , and by the regulariza-
tion of the structure of the supercoiled DNA
.
Fig. 5 .
We regard removal of translational and rota-
tional motions of the entire system as a desirable,
but not sufficient, precaution.
MODIFICATIONS TO ALGORITHMS THAT
USE VELOCITY RESCALING
Because velocity rescaling has a long history,
and those who wish to continue to use it should
consider modifying it. One possibility would be to
measure the mean temperature over some defined
interval, then use that value to calculate a rescal-
ing factor. Current protocols measure the instanta-
neous kinetic energy and calculate the correspond-
ing temperature. By averaging over a suitable in-
terval, much smaller scaling factors would result,
and one would uncouple the direction of scaling
from the instantaneous kinetic energy, which is the
source of the problem.
Within current algorithms, the problem can be
reduced by much less frequent rescaling, or,
equivalently, by the use of very long coupling
times in the Berendsen temperature bath protocol.
3
Rescaling should probably be supplemented by
occasionally reassigning velocities. If motion of the
center of mass and global rotational motion are
also periodically removed, this combination should
substantially ameliorate the problems identified
here.
Acknowledgments
Matt Young and David Beveridge have de-
scribed similar problems with velocity rescaling to
.
us personal communication ; we thank them,
Dennis Sprous, Tom Darden, and Bernard Brooks
for stimulating discussions. We thank Tamar
Schlick for reading a preliminary version of the
manuscript and offering several useful sugges-
tions, in particular for urging us to include discus-
sions on symplectic integrators and quench echo
dynamics.
References
1. J. A. McCammon and S. C. Harvey, Dynamics of Proteins and
Nucleic Acids, Cambridge University Press, Cambridge,
U.K., 1987.
2. C. L. Brooks, M. Karplus, and B. M. Pettit, Proteins: A
JOURNAL OF COMPUTATIONAL CHEMISTRY 739
HARVEY, TAN, AND CHEATHAM
Theoretical Perspective of Dynamics, Structure and Thermody-
namics, Wiley-Interscience, New York, 1988.
3. H. J. C. Berendsen, J. P. M. Postma, W. F. van Gunsteren, A.
. DiNola, and J. R. Haak, J. Chem. Phys., 81, 3684 1984 .
4. P. K. Weiner and P. A. Kollman, J. Comput. Chem., 2, 287
. 1981 .
5. R. K.-Z. Tan and S. C. Harvey, J. Comput. Chem., 14, 455
. 1993 .
. 6. R. K.-Z. Tan and S. C. Harvey, J. Mol. Biol., 205, 573 1989 .
7. D. Sprous, R. K.-Z. Tan, and S. C. Harvey, Biopolymers, 39,
. 243 1996 .
8. R. K.-Z. Tan, D. Sprous, and S. C. Harvey, Biopolymers, 39,
. 259 1996 .
. 9. D. Sprous and S. C. Harvey, Biophys. J., 70, 1893 1996 .
10. A. Bacolla, R. Gellibolian, M. Shimizu, S. Amirhaeri, S.
Kang, K. Ohshima, J. E. Larson, S. C. Harvey, B. D. Stollar,
. and R. D. Wells, J. Biol. Chem., 272, 16783 1997 .
11. T. E. Cheatham III and P. A. Kollman, J. Mol. Biol., 259, 434
. 1996 .
12. T. E. Cheatham III and P. A. Kollman, J. Am. Chem. Soc.,
. 119, 4805 1997 .
13. M. Tuckerman, B. J. Berne, and G. J. Martyna, J. Chem.
. Phys., 97, 1990 1992 .
14. J. J. Biesiadecki and R. D. Skeel, J. Comp. Phys., 109, 318
. 1993 .
15. J. M. Sanz-Serna and M. P. Calvo, Numerical Hamiltonian
Problems, Chapman & Hall, London, 1994.
16. T. Schlick, E. Barth, and M. Mandziuk, Annu. Rev. Biophys.
. Biomol. Struct., 26, 181 1997 .
. 17. L. Verlet, Phys. Rev., 159, 98 1967 .
18. W. C. Swope, H. C. Andersen, P. H. Berens, and K. R.
. Wilson, J. Comp. Phys., 76, 637 1982 .
19. P. J. Steinbach and B. R. Brooks, J. Comput. Chem., 15, 667
. 1994 .
20. T. C. Bishop, R. D. Skeel, and K. Schulten, J. Comput. Chem.,
. 18, 1785 1997 .
21. G. S. Grest, S. R. Nagel, and A. Rahman, Solid State Com-
. mun., 36, 875 1980 .
22. S. R. Nagel, A. Rahman, and G. S. Grest, Phys. Rev. Lett.,
. 47, 1665 1981 .
23. O. M. Becker and M. Karplus, Phys. Rev. Lett., 70, 3514
. 1983 .
24. D. Xu, K. Schulten, O. M. Becker, and M. Karplus, J. Chem.
. Phys., 103, 3112 1995 .
. 25. D. Xu and K. Schulten, J. Chem. Phys., 103, 3124 1995 .
. 26. H. C. Andersen, J. Chem. Phys., 72, 2384 1980 .
VOL. 19, NO. 7 740