Вы находитесь на странице: 1из 15

Multiconfiguration thermodynamic integration

T. P. Straatsma and J. A. McCammon



Citation: J. Chem. Phys. 95, 1175 (1991); doi: 10.1063/1.461148
View online: http://dx.doi.org/10.1063/1.461148
View Table of Contents: http://jcp.aip.org/resource/1/JCPSA6/v95/i2
Published by the American Institute of Physics.

Additional information on J. Chem. Phys.
Journal Homepage: http://jcp.aip.org/
Journal Information: http://jcp.aip.org/about/about_the_journal
Top downloads: http://jcp.aip.org/features/most_downloaded
Information for Authors: http://jcp.aip.org/authors
Downloaded 11 Apr 2012 to 129.215.149.92. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/about/rights_and_permissions
Multiconfiguration thermodynamic integration
T. P. Straatsma and J. A. McCammon
Chemistry Department, University of Houston, 4800 Calhoun Road, Houston, Texas 77204-5641
(Received 27 August 1990; accepted 10 April 1991 )
A modified thermodynamic integration technique is presented to obtain free energy differences
from molecular dynamics simulations. In this multiconfiguration thermodynamic integration
technique, the commonly employed single configuration (slow growth) approximation is not
made. It is shown, by analysis of the sources of error, how the multiconfiguration variant of
thermodynamic integration allows for a soundly based assessment of the statistical error in the
evaluated free energy difference. Since ensembles of configurations are generated for each
integration step, a statistical error can be evaluated for each integration step. By generating
ensembles of different lengths, the statistical error can be equally distributed over the
integration. This eliminates the difficult problem in single configuration thermodynamic
integrations of determining the best rate of change of the Hamiltonian, which is usually based
on equally distributing the free energy change. At the same time, this procedure leads to a
more efficient use of computer time by providing the possibility of added accuracy from
separate calculations of the same Hamiltonian change. The technique is applied to a simple but
illustrative model system consisting of a monatomic solute in aqueous solution. In a second
example, a combination of multiconfiguration thermodynamic integration and thermodynamic
perturbation is used to obtain the potentials of mean force for rotation of the sidechain
dihedral angles for valine and threonine dipeptides with restrained backbones in aqueous
solution.
I. INTRODUCTION
The calculation of free energies from molecular simula-
tions presents special problems as a result of the direct rela-
tion between free energy and the partition function, which is
prohibitively difficult to determine from simulation calcula-
tions. The origin of this difficulty is the inability to sample
the entire accessible phase space within any reasonable
amount of simulation time, even for rather simple systems.
To circumvent this problem, methods have been developed
to evaluate free energy differences rather than absolute free
energies. Two such methods, thermodynamic perturbation
and thermodynamic integration, have gained relatively wide
usage.
The method of thermodynamic integration best corre-
sponds to the general idea of how to measure a change in the
free energy of a system when some parameter specifying the
thermodynamic state of the system is slowly varied. When
carefully conducted, thermodynamic integration is an accu-
rate and straightforward method that does not suffer from
the problems encountered in many other methods in dealing
with systems that are large or have high densities.
The thermodynamic integration technique has become
one of the most frequently used procedures to obtain free
energy differences in a wide variety of systems. Unfortunate-
ly, the technique is often used without a proper error analy-
sis. Recent work has led to the notion that, in order to obtain
reliable free energy differences, even for relatively simple
systems, far longer 'simulations are necessary than is com-
monly assumed to be sufficient. 1.2 The purpose of this article
is to present a multiconfiguration thermodynamic integra-
tion (MCTI) technique in which statistical errors can be
determined straightforwardly by avoiding the usual single
configuration (SCTI) approximation. With this method the
reliability of calculated free energy differences is greatly im-
proved, and application of the method allows for more effi-
cient use of computer time. The method is illustrated in the
calculation of the free energy of hydration difference be-
tween neon and xenon, showing the most important aspects
of this new approach. Also, it is shown how a combination of
multiconfiguration thermodynamic integration and thermo-
dynamic perturbation can be used to obtain potentials of
mean force, illustrated for the rotation of the sidechain dihe-
dral angles for valine and threonine dipeptides with re-
strained backbones in aqueous solution.
II. THEORY
A. Thermodynamic integration
The most common application of thermodynamic inte-
gration calculations is the determination of the free energy
difference between two states of a system that differ in inter-
molecular or intramolecular interaction potentials. The in-
teraction potential is expressed as a function of some control
variable A that determines the state of the system. As a con-
sequence of the Hamiltonian being a function of this control
variable A, the partition function for the system is a function
of A as well. For an isothermal isobaric ensemble the parti-
tion function is
3
il(A) =-l-fffexp{ - H(A) +PV}dVd
pN
dqN
h
3N
N! kBT '
(1)
where N is the number of particles, h is Planck's constant,
H(A) is theA-dependent Hamiltonian,p is the pressure, Vis
J. Chern. Phys. 95 (2),15 July 1991 0021-9606/91/141175-14$03.00 1991 American Institute of Physics 1175
Downloaded 11 Apr 2012 to 129.215.149.92. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/about/rights_and_permissions
1176 T. P. Straatsma and J. A. McCammon: Multiconfiguration thermodynamic integration
the volume, kB is Boltzmann's constant, T is the absolute
temperature, and pN and qN are the momenta and positions
of the N particles.
The fundamental thermodynamic state function for the
isothermal isobaric ensemble is the Gibbs free energy G,
I
aG(A) = _ kBT [aa(A) ]
aA a (A) aA
which also is a function of A,
(2)
Differentiating G(A) with respect to A then gives
4

5
(3)
s s s (aH(A)laA)exp{ - [H(A) + pV IkBT ]}dV dpN dqN

(4)
S S Sexp{ - [H(A) +pVlkBT]}dVdpNdqN
= (aH(A) )
aA ....
which is an ensemble average of the mechanical quantity
aH(A)laA for a system with Hamiltonian H(A). This en-
semble average is readily obtained from molecular simula-
tions.
The Gibbs free energy difference between two states of a
system, described by the Hamiltonians H(A = 0) and
H (A = 1) can be obtained by integration of Eq. (5),
aG = G(A = 1) - G(A = 0) = t [aG(A)] dA
Jo aA A
= t (aH(A) ) dA. (6)
Jo aA A
Since molecular simulations are performed with dis-
crete steps, this integral has to be evaluated as a sum of en-
semble averages (MeT!),
aG=GI-G
o
=
a
o
(7)
(5)
where i counts over the number of different values of A, and
aA; is the difference between successive values of A. In gen-
eral the aA; do not have to be constant during a thermody-
namic integration.
In a very common approximation of thermodynamic
integrations, the sum over ensemble averages is replaced by a
sum over single configuration values for the derivative
(SCTI) ,
aG = I (aH) aA;.
; aA A
(8)
B. Thermodynamic perturbation
Thermodynamic perturbation is based on the expres-
sion, derived by Zwanzig,6 for the free energy difference be-
tween two systems with Hamiltonians Ho and
HI =Ho+aH,
(9)
= -kBTln S S Sexp{ - (HI +pVlkBT)}dVdpNdqN
S S Sexp{ - (Ho+pVlkBT)}dVdpNdqN
(10)
k
S S Sexp{ - (Ho + pV IkBT )}exp{ - (aH IkBT )}dV dpN dqN
= - B TIn ---------------------------------------------- (11 )
S S Sexp{ - (Ho+pVlkBT)}dVdpNdqN
= -kBTln(ex
p
{ -::r })o'
where the averaging is performed over the ensemble of con-
figurations obtained for Hamiltonian H
o
, the reference en-
semble. As was pointed out by several authors/'
s
the term
perturbation is somewhat misleading, since this expression is
exact and does not correspond to a perturbation theory in
the usual sense.
In multistep thermodynamic perturbation (MSTP) an
ensemble of configurations is generated for a range of values
for A, where A has the same function as in seT! or MeT!. A
perturbation free energy contribution is evaluated from each
ensemble generated at A; of
9
(12)
aG; = kB Tln(ex
p
{ _ H;_ 112 - H; }'
kBT /A,
- k B Tln( exp { - H; + H; } t, (13)
where H; = H(A; ),H;-1I2 = H(A; - HA.; - A.;_I})' and
H;+ 112 = H(A; + HA; - A;_I})' For the first value
11.1 =0,H;_1I2 =H1/2 =HI,andforthelastvalueA
N
= 1,
H N + 112 = H N' so that the first and last aG; have only one
term in Eq. ( 13). The totalfree energy difference can then be
found from
J. Chem. Phys., Vol. 95, No.2, 15 July 1991
Downloaded 11 Apr 2012 to 129.215.149.92. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/about/rights_and_permissions
T. P. Straatsma and J. A. McCammon: Multiconfiguration thermodynamic integration 1177
(14)
and the total statistical error from
( )
1/2
E(flG) = E; ,
(15)
where Ei = E(A.G;) is the statistical error of the ith pertur-
bation step.
c. Systematic, statistical, and sampling errors
Analyses of the applicability and limitations of thermo-
dynamic integrations using Eq. (8), in which such factors as
total simulation time, the use of cutoff radii, and the sizes of
the system and the time step were considered, have been
reported.
2
,7.10.11 The Hamiltonian time-lag problem, caused
by the strong correlation of successive dynamics steps,12 is a
major concern in SCT!.
The criterion for a correctly performed thermodynamic
integration is that for each value of A the ensemble of config-
urations [MCTI, Eq. (7)] or the single configuration
[SCTI, Eq. (8) ] is representative for the system with Ham-
iltonian H (). ). Several sources of error need to be consid-
ered. A systematic error is introduced by the fact that, after
each change of A, the configuration used in the next molecu-
lar dynamics step will not be representative for the Hamilto-
nian H(A). This error decreases with decreasing step size in
A. In SCTI the number of steps in A can be made quite large.
In practical calculations, however, the step size of A can not
be made infinitesimally small, and SCTI will lead to this
systematic error. To assess the magnitude of this error the
SCTI is usually performed in a forward and, after some equi-
libration of the final state, in a reverse direction. The differ-
ence between the two results obtained, the hysteresis, is
usually taken as a measure of this systematic error, assuming
that the systematic error in the forward direction does not
cancel the systematic error in the reverse direction. The ra-
tionale behind this assumption is that the free energy evalu-
ated from a configuration that is not representative for the
Hamiltonian used will be higher than the value obtained
from a representative configuration. Because this is true for
the SCTI in the reverse direction also, the systematic errors
from both directions do not cancel. In MCTI an ensemble of
configurations is generated for each value of A. To keep cal-
culations feasible, the number of steps in A must be chosen
considerably smaller than is possible in SCTI. Consequently,
the step size of A is larger. Generation of each ensemble
should therefore be preceded by equilibration.
The fluctuations in properties obtained from equilibri-
um molecular dynamics simulations lead to statistical er-
rors. In MCTI calculations the derivative in Eq. (7) for each
value of A is obtained as an ensemble average. Consequently,
a statistical error Ei = E [(aH)(A)laA >;] can be evaluat-
ed for the ensemble average (aH(A)laA >; at each value of
A;. Assuming that the statistical errors from successive steps
are independent, an estimate for the statistical error for the
entire MCTI can be found from
( )
112
E(A.G) = E;A.A.; .
(16)
Since SCTI simulations do not involve the calculation of
ensemble averages, a statistical error can not be evaluated.
There is no reason, however, to expect statistical errors not
to be important in SCT! simulations. This is a serious disa?-
vantage, since this not only affects the final free dif-
ference that is obtained, but also the hysteresis. ThIS makes
the hysteresis a very unreliable measure of the error in SCT!.
In a recent publication Mitchell and McCammon
2
discuss
the difficulties in obtaining precise free energy values from
SCTI and the limited value of hysteresis as a measure of
error.
Adequate sampling of phase space is a fundamental
problem in any molecular simulation. Unfortunately,
systematic and statistical errors, errors due to poor sampltng
are difficult to determine quantitatively. A general case in
which inadequate sampling presents a serious problem is
when mUltiple minima exist on the energy surface, separated
by energy barriers. A specific example is any molecule that
has distinct rotational isomeric states (e.g., the gauche and
trans conformations of butane). Since only one configura-
tion is used to obtain the derivative of the Hamiltonian, im-
portant restrictions have to be applied to the systems that
can be studied using SCT!. If more than just one stable con-
formation of the system can exist at any point during the
integration, a single configuration value for the derivative
will be a poor approximation for the value that would have
been obtained from a well sampled ensemble. The entropic
contribution due to the existence of other conformations of
the molecule would be completely neglected. But even if
there is just one valley through phase space from initial to
final state of the system, this valley may change in breadth.
This change in accessible phase space would be reflected in
the value obtained from an ensemble average, but is not in
the value obtained from a single configuration. Although
such contributions would be properly accounted for in an
extremely long SCTI calculation in which a fully representa-
tive sample is generated for every small interval in A, such
calculations will be prohibitively impractical in practice.
Where multiple minima occur, special techniques can be ap-
plied to combine all of their contributionsY-15 MCTI, how-
ever, has the advantage that the statistical error estimates
may provide indications of poor sampling and, as shown
below, allows for systematic improvements in sampling.
In MCT! simulations not all ensembles need to be of
equal size. Each ensemble can be made as large as needed to
make the statistical error roughly the same for each value of
A. Each part of the free energy curve obtained then has the
same statistical accuracy. Due to the strong correlation in
molecular dynamics simulations statistical error will have to
be evaluated using a correlation analysis approach.
Another advantage of MCTI simulations is that the lim-
ited number of values of A allows for recording of the final
configuration and some additional information for each val-
ue of A. The accuracy of the free energy difference result can
then easily be improved by increasing the size of generated
ensembles, without loss of CPU time spent in earlier calcula-
tions. For SCTI simulations entire runs have to be redone at
a longer total simulation time, essentially wasting the CPU
time spent in the shorter runs.
J. Chem. Phys., Vol. 95, No.2, 15 July 1991
Downloaded 11 Apr 2012 to 129.215.149.92. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/about/rights_and_permissions
1178 T. P. Straatsma and J. A. McCammon: Multiconfiguration thermodynamic integration
There is no need for a non-linear change of A. in MCT!
simulations if the length of each ensemble is determined by
the statistical error. A nonlinear change of A. is often needed
in SCT!, because regions with large fluctuations will require
more densely spaced values of A. than regions with relatively
smaller fluctuations in the value of the derivative. In MCTI
this problem is taken care of automatically by continuing to
generate configurations for each ensemble until the statisti-
cal error, which is determined by the fluctuations, becomes
less than a certain preset value.
In SCTI simulations, evaluating a hysteresis from a re-
verse simulation is a way to get at least some crude measure
of the error, but this makes it necessary to perform exactly
the same integration again in the reverse direction, thereby
doubling the CPU time needed. Another option is to per-
form multiple runs with equal integration lengths. This will
give an estimate of the statistical error, but the systematic
error is expected to be roughly the same in all runs because
the change in the Hamiltonian is made in the same simula-
tion time. Considering the fact that the use of the hysteresis
as a measure of error appears more questionable as the SCTI
technique is further analyzed, the MCT! technique is pro-
viding a more reliable estimate of the error.
D. Free energy calculations using restraining potentials
Since free energy is a thermodynamic state function the
path followed through Hamiltonian space in thermodynam-
ic integrations is in principle irrelevant for the resulting total
free energy difference. The particular choice of the func-
tional form may affect the error in the free energy difference
considerably in actual calculations, and a particular choice is
usually made in an attempt to decrease the error in the final
free energy difference result. This is true for free energy dif-
ference calculations between real states of the systems, but
also between states that include A-independent restraining
potentials to keep one or both systems in a fictitious state.
The thermodynamic perturbation method can be used to
evaluate the free energy difference between restrained and
unrestrained system. 13, 14 An example of a particular choice
of the path followed, in order to decrease the error in the
final result, is electrostatic decoupling.16
In certain cases one is interested in the intermediate
states during a thermodynamic integration in which A-de-
pendent restraining potentials are used. The change in the
system during the thermodynamic integration can, for ex-
ample, describe a postulated reaction coordinate. In that
case, the Hamiltonian H used for the thermodynamic inte-
gration consists of the A-independent Hamiltonian Ho de-
W(x) = - RT lnp(x)
S exp[ - (H /RT) ]dr'
- RT In
S exp[ - (H /RT) ]dr
I
scribing the normal interactions in the system and a A-de-
pendent restraining potential U that determines the desired
path through Hamiltonian space that describes the reaction
coordinate
H(p,q,/l.) = Ho(p,q) + U(q,A). (17)
Analogous to the result ofthe derivation in Sec. II A, the free
energy as a function of A is given by
aG(A) = r). ') dA', (18)
J.I.u aA ).'
where the integration is from some initial value A ' = Ao to
A'=A.
The free energy difference found using Eq. (18) repre-
sents the free energy of the system described by the Hamilto-
nian including the restraining potential. The free energy of
the unrestrained system when in the phase space defined by
the simulation using the Hamiltonian including the restrain-
ing potential can be found by correcting for the use of the
restraining potential by means of a thermodynamic pertur-
bation calculation
aG(A) = r' ') dA'
J)." aA ).'
]L.
(19)
The free energy difference given by Eq. (19) represents the
free energy difference of the unrestrained system between
different parts of phase space, namely those parts as obtained
from the simulations in which the restraining potential was
used.
Since two ensemble averages are needed for each value
of A in this expression, MCT! can be used to evaluate this
free energy difference. Note that the first term is an integral
and in a practical calculation will be an accumulated value,
while the second term in Eq. (19) is a single correction term
for each value of A. The SCTI approximation can be made
here also, leading to
aG(A = r' ')] dA' - U(A). (20)
L" aA ).'
Obviously, the discussion in the previous sections of the ap-
plicability and reliability of this approximation is valid here
as well.
If A can be uniquely associated with the value of some
internal coordinate in the system, the free energy difference
evaluated using Eq. (19) or Eq. (20) is identical to the po-
tential of mean force (PMF) Walong this internal coordi-
nate. In that case, the free energy difference expression Eq.
(19) is a good approximation for the PMF
I7
(21)
(22)
(23)
J. Chem. Phys., Vol. 95, No.2, 15 July 1991
Downloaded 11 Apr 2012 to 129.215.149.92. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/about/rights_and_permissions
T. P. Straatsma and J. A. McCammon: Multiconfiguration thermodynamic integration 1179
S exp( - [H + U(x)/RT] )dr' S exp[ - (H IRT) ]dr
= -RTln -U(x)+RTln---------- (24)
S exp( - [H + U(x)/RT] )dr S exp( - [H + U(x)/RT] )dr
= - RTlnp*(x) - U(x) - RTln(exp( - )),
(25)
where p is the probability density for the internal coordinate
in the unrestrained system and p* is the probability density
for the internal coordinate in the restrained system. If Ax is
the value for A that results in the value x for the internal
coordinate, then W(Ax)::::: W(x). The phase space integra-
tion S dr = S dr' + S dx is separated into an integration
along the dimension of the reaction coordinate S dx and the
remainder of phase space S dr'. For the exact one-to-one
correspondence between Ax, and x, the correction needed
for the use of the restraining potential is equal to this poten-
tial,
- U(.{x) = - U(x).
(26)
As an example of a restraining potential applied to an
internal coordinate, consider a proper dihedral potential
function with single multiplicity,
U(if;,A) = C,,{l + cos [if; - S(A) n, (27)
where S(A) is the A-dependent angle determining the angle
if; = S(A) + 1T oflowest energy. The function form of S(A)
could be a function such as
(28)
so that 8(A) changes from the initial value 8; to the final
value 8[ if A changes from 0 to 1. If this is the only change in
the thermodynamic integration, the derivative is found from
=C,,(S[-8;)(sin[if;-8(A')]),\,' (29)
\ aA ,t.
In order to obtain a small spread in the internal coordinate if;
around the average value if;x for the simulation at A = Ax,
the restraining force constant C", has to be chosen as large as
possible without causing numerical problems in the integra-
tion of the equations of motion. 18 It has also been suggested,
in addition to using a large force constant, to only count
contributions to the ensemble averages ofEq. (19) for which
the internal coordinate is equal to the desired value. 19 In the
example given above, only contributions would be counted
for which <P = 8(A) + 1T with a very small tolerance. In
practice this may prove to be inefficient, since many configu-
rations generated may have to be skipped. Also, with this
procedure the nature of the ensemble generated becomes un-
clear, and the validity of the ensemble average uncertain.
The use of high force constants for restraining potentials
presents a special problem in the choice of the time step.
Clearly, the time step should not be taken too large in order
to prevent the simulation from becoming unstable. Too
small a time step, however, will not lead to representative
ensembles of configurations, from which the ensemble aver-
ages have to be evaluated, within a reasonable amount of
computer time. Due to numerical errors, caused by the use of
finite time steps, the use of high force constants may lead to
I
configurations with high energies being over-represented in
the generated ensembles. This has a small effect on the en-
semble average of the derivative of the Hamiltonian as given
in Eq. (18). High restraint energies will have a large effect
on the ensemble average (exp[ U(A)/RT]) needed for the
free energy correction in Eq. (19). The large fluctuations in
the restraint energies are found to have a smaller influence
on this free energy connection if only the first order approxi-
mation to this correction is used,
aG(A) = dA'-(U(A,t. (30)
J,tO) aA ,t'
The potential of mean force evaluated using Eq. (30) or
( 19) will be an approximation due to the spread of the reac-
tion coordinate for each value of A, which is caused by the
fact that, in order to obtain stable dynamics, the force con-
stant of the applied restraining potential can not be chosen
too large. Fortunately, this spread is expected to be small in
the wells of potentials of mean force since there is a natural
tendency for the reaction coordinate to remain at the posi-
tion of the well, even without the use of the restraining poten-
tial. Consequently, the uncorrected free energy difference
between wells can be expected to be very similar to the cor-
rected result.
The procedure outlined in this section is an alternative
to the calculation of free energy differences along a reaction
coordinate, in which constraints
I5
,20.21 are used rather than
a restraining potential.
III. COMPUTATIONAL DETAILS
The MCT! technique is the standard free energy ther-
modynamic integration technique in the molecular dynam-
ics program package ARGOS,22 with which the simulations
described in the following sections were carried out. SCT! is
just a limiting case of MCT!, in which the minimum and
maximum number of configurations per step in A are both 1.
To be able to perform SCTI with nonlinear change in A this
option is available too, although this is not needed in MCT!.
The simulation volume is a cubic box, subject to period-
ic boundary conditions. To keep temperature and pressure at
298 K and 10
5
Pa, respectively, coordinates and velocities
were scaled every dynamics step, using a method to weakly
couple the system to an external heat and pressure bath.
23
Relaxation times used for the weak coupling are 0.4 and 0.5
ps for temperature and pressure, respectively.
The potential model used for water is the SPC/E water
model developed by Berendsen et al.
24
The Lennard-Jones
parameters for the interaction
Cl2 C6
V,u(r.) = __ 'J ---" (31)
l} 12 6
rij rij
between the noble gases and the water oxygen atoms 10 are
given in Table I. The intramolecular distances are fixed, us-
J. Chem. Phys., Vol. 95, No.2, 15 July 1991
Downloaded 11 Apr 2012 to 129.215.149.92. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/about/rights_and_permissions
1180 T. P. Straatsma and J. A. McCammon: Multiconfiguration thermodynamic integration
TABLE I. Lennard-Jones potential parameters for the noble gas interac-
tions with SPC/E oxygen.
Interaction
Ne-O
Xe-O
~ ~ 2
10-
3
kJnm
6
mol-! 10-
6
kJnm!2mol-!
1.12457 0.99806
8.92289 18.472 I
ing the coordinate resetting procedure SHAKE.
25
-
27
The GROMOS
28
force field was used to describe the inter-
actions of the dipeptide.
The statistical errors reported in this article have been
evaluated using a method by Straatsma et al.
7
IV. FREE ENERGY OF HYDRATION OF Ne AND Xe
The difference in free energy of hydration for Ne and Xe
provides a simple but illustrative problem to analyze the
AG I kJ mol-
1
6.
4.
2.
o.
-2.
O. 5.
(llG/Il"-) I kJ mol-
1
400.
300.
200.
100.
O.
o.
-100.
-200.
-300.
-400.
O. 5.
r
10. 15.
10. 15.
TABLE II. SCTI free energy difference results for the mutation Ne-Xe in
aqueous solution, using the functional form A = (t IT )'.
Time IlG- IlG- Hysteresis
ps
kJ mol-! kJ mol-! kJ mol-!
25 -2.9 2.4 -0.5
50 -0.7 3.5 2.8
100 1.1 1.9 3.0
250 -2.2 1.8 0.4
500 - 1.4 2.3 1.0
750 - 1.0 1.6 0.6
thermodynamic integration technique. Simulations were
performed in the isothermal isobaric ensemble of the nonpo-
lar monatomic solute and 128 SPC/E water molecules.
Intuitively, the most accurate final free energy differ-
ence result from a scn calculation is expected if the total
change in the free energy during the process is evenly distrib-
20.
20.
a
b
c
FIG. I. Single configuration thermody-
namic integration free energy Ne-Xe,
25. withA = (tIT)2in25ps. (a) IlGvstime,
forward (solid) and reverse (dashed);
(b) (aG laA) for forward SCT!; (c)
(aG I aA) for reverse SeT!.
25.
Time I ps
J. Chem. Phys . Vol. 95, No.2. 15 July 1991
Downloaded 11 Apr 2012 to 129.215.149.92. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/about/rights_and_permissions
T. P. Straatsma and J. A. McCammon: Multiconfiguration thermodynamic integration 1181
AG / kJ mol-
1
6.
4.
2.
o.
-2.
O.
(BG/BX) / kJ mol-
1
400.
300.
200.
100.
o.
o.
-100.
-200.
-300.
-400.
O.
250. 500.
250. 500.
a
FIG. 2. Single configuration thermody-
namic integration free energy Ne .... Xe.
750. with A = (t IT)2 in 750 ps. (a) t:.G vs
time. forward (solid) and reverse
(dashed); (b) (aG laA) for forward
SCTI; (c) (aG laA) for reverse SCT!.
b
c
750.
Time / ps
uted over the molecular dynamics time steps. This means
that the slope of the free energy change versus simulation
time should be as uniform as possible over the whole range.
By appropriately choosing the time dependence of the con-
trol variable A., steep slopes can often be avoided. For the
process of changing Ne into Xe in aqueous solution this can
be accomplished by a functional form

(32)
where t is the running simulation time, and T is the total
simulation time for the thermodynamic integration. With
this functional form for A. several SeT! calculations were
performed in forward and reverse directions with different
total simulation times. The free energy results and the hys-
tereses obtained are listed in Table II. The first thing to note
is the slow and poor convergence of the final free energy
values as a function of simulation time. Second, the hystere-
sis does not monotonically decrease with increasing total
simulation time. The cause of this problem becomes appar-
ent when the free energy change for a single thermodynamic
integration is plotted as a function of the simulation time,
together with the calculated derivatives. In Fig. 1 these prop-
erties are shown for the thermodynamic integration in 25 ps.
Figure 1 (a) gives the free energy result for the forward di-
rection as the full line, and the curve for the reverse direction
as the broken line. Figures 1 (b) and 1 (c) give the derivative
(JG / JA.) for the forward and reverse direction, respectively.
The free energy curves are smoother and more similar in the
latter part of the simulation, where the fluctuations in the
derivative are small. The first half of the thermodynamic
integration shows large fluctuations in the derivative and
larger differences between the forward and reverse free ener-
gy curves. It is important to observe that, at certain values of
J. Chem. Phys., Vol. 95. No.2. 15 July 1991
Downloaded 11 Apr 2012 to 129.215.149.92. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/about/rights_and_permissions
1182 T. P. Straatsma and J. A. McCammon: Multiconfiguration thermodynamic integration
TABLE III. SeT! free energy difference results for the mutation Ne ..... Xe in
aqueous solution. using the functional form ,1= (t IT )".
Time I::.G- I::.G- Hysteresis
ps kJmol-
'
kJ mol-I kJ mol-I
10 5.6 4.6 10.2
25 -0.3 -0.1 -0.4
50 0.2 4.0 4.1
100 - 1.4 0.8 -0.6
250 - 1.9 1.5 - 0.4
350 - 1.6 0.5 -1.1
500 -0.8 0.8 0.0
625 - 1.5 1.1 -0.3
750 -2.2 1.8 -0.3
A., the free energy curves for the forward and reverse direc-
tions, deviate more than the hysteresis, the difference at the
end of the simulations. Between 8 and 9 ps in Fig. 1 (a) there
is a difference between the two curves of about 2 kJ mol - 1 ,
while the final hysteresis is only of the order of 0.5 kJ mol - I .
AG j kJ mol-
1
6.
4.
2.
o.
-2.
o.
(8GjaX) j kJ mol-
1
400.
300.
200.
100.
o.
o.
-100.
-200.
-300.
-400.
O.
/
- - --
250. 500.
250. 500.
Yet at every time step the Hamiltonian is exactly the same
for the forward and reverse simulations. Therefore, the two
curves should be exactly the same, not only at beginning and
end, but over the whole range. The fact that the difference
does not increase monotonically, and the fact that the final
hysteresis can be found to be much smaller than the differ-
ence between the two curves at some other point during the
simulation, illustrate the limited value of the hysteresis as a
measure of the error.
For the longest thermodynamic integration (750 ps)
using the same time dependence of A., the corresponding
curves are plotted in Fig. 2. Forward and reverse free energy
changes are much smoother and resemble each other more
compared to the 25 ps simulation. Still, as expected, fluctu-
ations in the derivative behave in the same way, and, conse-
quently, the largest deviation between forward and reverse
free energy result is found in the first part of the calculation,
where high fluctuations in the derivative are found. For this
seemingly simple process 750 ps still appears to give an ap-
a
FIG. 3. Single configuration thermody-
namic integration free energy Ne ..... Xe.
750. with A = (t IT)" in 750 ps. (a) I::.G vs
time, forward (solid) and reverse
(dashed); (b) (oG laA) for forward
SeTI; (c) (oG laA) for reverse sen.
b
c
750.
Time j pH
J. Chem. Phys . Vol. 95, No.2, 15 July 1991
Downloaded 11 Apr 2012 to 129.215.149.92. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/about/rights_and_permissions
T. P. Straatsma and J. A. McCammon: Multiconfiguration thermodynamic integration 1183
TABLE IV. MCTI free energy difference results for the mutation Ne-Xe in aqueous solution.
Equilibration Minimum Maximum Total time
steps steps steps ps
100 100 100 20.000
100 100 500 31.188
100 100 1000 46.544
250 250 500 59.364
250 250 1000 85.048
100 100 5000 194.572
25 100 10000 363.880
100 250 17500 669.450
100 250 25000 776.529
preciable uncertainty in the final free energy result. The hys-
teresis is of the order of 30% of the calculated free energy
difference.
6.G
kJ mol-I
1.4
-0.4
-2.0
-2.7
-1.5
- 1.3
- 2.3
- 1.5
-2.3
Error.
kJ mol-I
0.9
0.2
0.2
0.2
0.1
0.2
0.2
0.2
0.2

(33)
The series of thermodynamic integrations was repeated
using a different time dependence of A,
This redistributes the stepsize in A, making the LUi smaller
in the range with large fluctuations in the derivative, and
AG / kJ mol-
1
5.0
2.5
0.0
-2.5
O.
(8G/8X) / kJ mol-
1
75.
50.
25.
o.
E(AG) / kJ mol-
1
0.20
0.15 I-
0.10
0.05
0.00
O.
250.
I
I
250.
J
I
500. 750.
b
I I
<:
-
-
I T
500.
Time / ps
750.
J. Chem. Phys., Vol. 95, No.2, 15 July 1991
FIG. 4. Multiconfiguration thermody-
namic integration free energy Ne-Xe, in
776ps. (a) 6.Gvstime; (b) (aGlaA); (c)
Statistical error vs time.
Downloaded 11 Apr 2012 to 129.215.149.92. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/about/rights_and_permissions
1184 T. P. Straatsma and J. A. McCammon: Multiconfiguration thermodynamic integration
AG / kJ mol-
t
r - - - - - - - - - - - - - - . - - - - - - - - - - - - - - r - - - - - - - - - - - - - ~ - - - - - - - - - - - - ~
40.
20.
o.
0.00
(IJG/IJA) / kJ mol-
t
250.
o.
-250.
E(AG) / kJ mol-
1
1.5
1.0
0.5
0.0
0.00
0.25 0.50
0.25 0.50
larger in the range with small fluctuations. The free energy
results and hystereses obtained for a series of calculations
with different total simulation times are given in Table III.
The forward and reverse results appear to converge more
rapidly with increasing total simulation time, compared to
the first series of calculations. The hysteresis, although
smaller than in the first series at comparable simulation
lengths, still does not monotonically decrease with increas-
ing total simulation time. Again, this lack of trend in the
magnitude of the hysteresis with increasing total simulation
time and the fact that negative values for the hysteresis are
found, which contradicts the argument that systematic er-
rors will not cancel between forward and reverse thermody-
namic integrations, lead to the observation that hysteresis is
a poor measure of the quality of this type of calculation.
The accumulated free energy change for the 750 ps sim-
ulation is shown in Fig. 3, together with the derivatives for
the forward and reverse simulation. In the large fluctuation
a
0.75
c
0.75
FIG. 5. Free energy difference for the
forced rotation around the X dihedral an-
gle in ifJ and", restrained valine dipeptide
in aqueous solution. (a) Free energy
change as a function of A after 2000 (dot-
ted line), 3000 (dash-<lot line), 4000
(dashed line), and 5000 (solid line) con-
1.00 figurations in each ensemble for A. (b)
1.00
Free energy derivative as a function of A
after 2000 (dotted line), 3000 (dash-<lot
line), 4000 (dashed line), and 5000 (solid
line) configurations in each ensemble for
A. (c) Accumulated statistical error in
free energy curve as a function of A after
2000 (dotted line), 3000 (dash-<lot line),
4000 (dashed line), and 5000 (solid line)
configurations in each ensemble for A.
part of the derivative, where the steps in A, are very small, as
well as in the last part of the curve where steps in A, are
relatively large but the fluctuations in the derivative are
small, the free energy curves have a very similar slope. It is in
the region in between where steps in A, are becoming larger,
that the fluctuations of the derivative are still large enough to
cause the free energy difference curves to show a noticeable
difference in slope. As a result the hysteresis of the total run
is still found to be appreciable, considering the length of the
simulation, 750 ps.
The results from the two series of SCTI simulations de-
scribed above show the difficulty of efficiently distributing
molecular dynamics time steps over the range of values of A,.
Table IV lists the results of a series ofMCTI simulations. For
each simulation the number of steps in A, was 100, equally
spaced from 0 to 1. This table lists the number of equilibra-
tion steps for each value of A" the minimum and maximum
number of steps for the evaluation of the ensemble average of
J. Chem. Phys., Vol. 95, No.2, 15 July 1991
Downloaded 11 Apr 2012 to 129.215.149.92. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/about/rights_and_permissions
T. P. Straatsma and J. A. McCammon: Multiconfiguration thermodynamic integration 1185
11' / kJ mol-
1
50. r - - - - - - - - - - - - . - - - - - - - - - - - - ~ - - - - - - - - - - _ , - - - - - - - - - - - - ~
40.
30.
20.
10.
o.
-11' o
AG ...... / J.cJ mol-
1
-2.
-3.
-4.
-5.
-6.
o
the derivative, the total simulation time, including the time
used for the equilibrations, and the accumulated Gibbs free
energy differences using Eq. (7), as well as the accumulated
statistical error evaluated using Eq. (16). The statistical er-
ror threshold was set to be 1.0 kJ mol - I in the ensemble
average of the derivative. The free energy difference seems to
converge much faster than in the SCT! simulations de-
scribed above. The variation in final free energy differences
is of the same order as the statistical errors found.
In Fig. 4 the accumulated free energy as a function of
simulation time is given for the longest of the MCT! simula-
tions in Table IV, together with the derivative in Fig. 4(b).
The derivative shows hardly any fluctuations, leading to a
very smooth free energy change. The free energy shows a
very interesting feature. In the last 50 ps of the simulation the
fluctuations in the derivative are so small that the minimum
number of steps were already sufficient to bring the statisti-
cal error below the preset threshold. The peculiar shape of
the free energy curve illustrates that distribution of the simu-
lation steps according to the fluctuations in the derivative
gives a more efficient and more accurate result than equal
distribution of free energy change over the molecular dy-
namics simulation.
At the long time scale of the simulations reported here,
an additional cause for lack of convergence may be the long
time fluctuations in an aqueous environment. 1.30
11'/2
x
x
FIG. 6. Potential of mean force for rota-
tion of dihedral angle X in valine dipeptide
(a) PMF as obtained from a MCT! free
energy calculation (dashed curve), fitted
to a sum of six sinusoidal functions (dot-
ted curve), and as obtained from the dis-
tribution of the angle X from a simulation
using the negative of the fitted function as
an umbrella potential. (b) Correction to
the MCT! free energy curve to obtain the
PMF.
V. POTENTIAL OF MEAN FORCE IN VALINE AND
THREONINE DIPEPTIDE
Multiconfiguration thermodynamic integrations have
been carried out to determine the X potential of mean force
for the and t/J restrained valine dipeptide and threonine
dipeptide in aqueous solution. This serves to illustrate the
use of this type offree energy evaluation in a case in which a
A-dependent restraining potential with a large force constant
is employed to obtain a potential of mean force. The back-
bone dihedral angles and t/J were restrained, first to avoid
the necessity of considering the multiple stable isomeric
states that would otherwise be possible, and second, to mod-
el the degree of freedom in X as it would occur in a protein in
which the backbone is less flexible. The angles and t/J were
restrained at - 2.617 99 and 2.61799 rad, respectively,
with sinusoidal restraining potentials with force constants
175 kJ mol- I . Simulations were performed on the dipeptide
in 198 SPC/E water molecules in a cubic periodic box, in the
isothermal isobaric ensemble. The restraining potential on
the X dihedral angle is a sinusoidal potential with a force
constant of 175 kJ mol-I. The reference angle of this poten-
tial is changed during the MCT! from a to 21T, so that the
angle X is changed from 1T to - 1T.
The complete change was made in 101 equally spaced
steps in A. At each value of A, 500 equilibration steps were
taken, followed by 5000 steps of data acquisition. To illus-
J. Chem. Phys., Vol. 95, No.2, 15 July 1991
Downloaded 11 Apr 2012 to 129.215.149.92. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/about/rights_and_permissions
1186 T. P. Straatsma and J. A. McCammon: Multiconfiguration thermodynamic integration
trate the convergence of the free energy curve with increas-
ing number of configurations generated for the evaluation of
the ensemble averages, and the ease with which additional
configurations can be generated for an already completed
simulation, the data acquisition was performed in four sepa-
rate calculations. For the valine case, the free energy change,
the ensemble average of the derivative and the accumulated
statistical error are shown in Figs. 5(a), 5(b), and 5(c),
respectively, as a function of A after 2000 (dotted line), 3000
(dash-dot line), 4000 (dashed line), and 5000 data acquisi-
tion steps (solid line). The first thing to note is the difference
between the initial and final values of the free energy curve,
which should be zero. The fact that the final value tends to
approach zero with increasing number of configurations in
each ensemble, and the fact that there is a converging change
of the entire free energy curve with increasing size of the
ensembles, points towards a systematic error that becomes
less when more configurations are generated for each value
of A. The most probable cause is that the equilibration with
500 steps for each ensemble is insufficiently long to fully
equilibrate the ensemble after the change in A. The ensemble
derivatives for the four cases in Fig. 5 (b) show no noticeable
differences. An important result is the extremely slow con-
vergence of the statistical error with increasing ensemble
size, as can be seen from Fig. 5 (c). It clearly shows the slow
convergence of the statistical error with increasing number
of data. A correlation analysis was performed in calculating
the statistical error. The errors were found to be more than
five times larger than the standard errors. Although the eval-
uation of the statistical error in MCTI using Eq. (16) may
represent a conservative estimate, the importance of a corre-
lation analysis is obvious. A second observation is that the
statistical error is monotonically increasing as the integra-
tion proceeds. Unlike in the example in the previous section,
it appears not to be necessary to let the statistical error deter-
mine the number of configurations in each ensemble. The
statistical error from the fairly long simulation of 500 ps is
1.4 kJ mol-I.
The potential of mean force for rotation of dihedral an-
gle X, evaluated using Eq. (30) with a force constant of
826.21 kJ mol - I for the restraint potentials, is shown by the
dashed curve in Fig. 6(a). The correction to the MCn free
energy to obtain this PMF is shown in Fig. 6(b). These
curves were evaluated using data obtained from a simulation
with 5000 steps per A. The correction applied is small com-
pared to the free energy values it is applied to. Consequently,
there is only a small difference between the free energy curve
and the evaluated potential of mean force. The small correc-
tion needed also indicates that, for each value of A, the reac-
tion coordinate X only deviates slightly from the minimum
in the restraining potential.
To make a comparison with the potential of mean force
as evaluated using Eq. (25) the PMF as obtained from the
MCn was fitted to the function
" / kJ mol-
t
35. ~ - - - - - - - - - - - r - - - - - - - - - - - - ~ - - - - - - - - - - - ' - - - - - - - - - - - - '
30.
25.
20.
15.
10.
5.
o.
-71"
AG
rr
/ kJ mol-
1
-2.
-3.
-4.
-5.
-6.
-71"
".. a
o 71"/2
b
o
J. Chem. Phys., Vol. 95, No.2, 15 July 1991
71"
FIG. 7. Potential of mean force for rota-
tion of dihedral angle X t in threonine di-
peptide (a) PMF as obtained from a
MCTI free energy calculation (dashed
curve), fitted to a sum of six sinusoidal
functions (dotted curve), and as obtained
from the distribution of the angle X from a
simulation using the negative of the fitted
function as an umbrella potential. (b)
Correction to the MCTI free energy curve
to obtain the PMF.
Downloaded 11 Apr 2012 to 129.215.149.92. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/about/rights_and_permissions
T. P. Straatsma and J. A. McCammon: Multiconfiguration thermodynamic integration 1187
TABLE V. Parameters obtained for the fit of the potential of mean force
around X in valine dipeptide with the function given in the text.
C, 8,
n,
kJ mol- I kJ mol-
I
1 7.73337 - 0.78235
2 10.36670 1.26350
3 9.43613 2.98581
4 1.25923 - 0.866 43
5 0.59568 1.540 57
6 0.24668 2.87143
6
<1>(x) = L C; [1 + cos(n;x - 0;)].
(34)
;= I
The parameters n;,e;, and 0; for this fit are given in Tables V
and VI, and the function is shown by the dotted curve in Fig.
6 (a). The distribution p* of the angle X was then obtained
from a series of molecular dynamics simulations using an
umbrella potential U(X) = - <1>(X)' With this umbrella
potential the rotation around X is expected to be almost free.
Each of the final configurations at each step in A. in the
MeT! was used as a starting configuration for a lOps molec-
ular dynamics simulation. Using the distribution p* ob-
tained from these simulations in Eq. (25) resulted in the
PMF given as the solid curve in Fig. 6 (a). The two curves for
the PMF are, considering the statistical error, in very good
agreement.
These calculations were repeated for the threonine di-
peptide, under the same conditions, again using a force con-
stant of 826.21 kJ mol- I for the restraint potentials. Figure
7 (a) shows the potential of mean force from a MeT! simula-
tion (dashed curve), the fit of this PMF to the function given
by Eq. (34) (dotted curve), and the PMF evaluated by use
in Eq. (2S) of the angle distribution p* as obtained from a
series of molecular dynamics simulations in which
UCr) = - <P(l') was used as an umbrella potential (solid
curve). Figure 7 (b) gives the correction that was applied to
the MeT! free energy curve to obtain the PMF. Again good
agreement is found between the two potentials of mean
force.
VI. DISCUSSION
The multiconfiguration thermodynamic integration
technique presented here and the multiple step perturbation
method are based on the same basic idea that the most reli-
TABLE VI. Parameters obtained for the fit of the potential of mean force
around XI in threonine dipeptide with the function given in the text.
C, 0,
n,
kJ mol-
I
kJmol-
1
1 1.799 10 0.55666
2 7.50140 1.08722
3 9.29020 3.05060
4 0.504 76 - 0.17836
5 0.35520 0.89194
6 0.26550 4.077 23
able way of obtaining properties is by averaging over ensem-
bles of configurations. The basic problem in perturbation
method calculations is that the reference ensemble has to be
representative for a system described by the perturbed Ham-
iltonian as well. This problem does not exist in multiconfi-
guration thermodynamic integration since the derivative of
the Hamiltonian should be evaluated from the generated en-
semble.
The MCTI technique makes a more reliable assessment
of the statistical error possible compared to the commonly
used single configuration thermodynamic integration in
which the error is usually obtained from the hysteresis. This
is important because the reliability of the hysteresis as a mea-
sure of error is questionable. The MeT! also allows adding
to the quality of previous simulations, without loss of pre-
viously obtained results. Moreover, it is possible, at the same
time, to increase the quality of free energy evaluations in
limited parts of the complete change of the system based
upon the contribution to the statistical error. This is not pos-
sible with SCT! simulations. MeT! simulations therefore
make more efficient use of computing resources, while pro-
viding more soundly based error estimates.
ACKNOWLEDGMENTS
This work has been supported in part by the National
Science Foundation, the Robert A. Welch Foundation, the
Texas Advanced Research Program, the Houston Ad-
vanced Research Center, and HNS Supercomputers. J. A.
M. is the recipient of the George H. Hitchings Award from
the Burroughs Wellcome Fund.
1M. H. Mazor and B. M. Pettitt, Mol. Simulation 6,1 (1991).
2M. J. Mitchell and J. A. McCammon, J. Compo Chern. 12, 271 (1991).
3D. A. McQuarrie, in Statistical Thermodynamics (Harper & Row, New
York,1973).
'P. H. Berens, D. H. J. Mackay, G. M. White, and K. R. Wilson, J. Chern.
Phys. 79, 2375 (1983).
50. Frenkel and A. J. C. Ladd, J. Chern. Phys. 81, 3188 (1984).
6R. W. Zwanzig, J. Chern. Phys. 22,1420 (1954).
7T. P. Straatsma, Ph.D. thesis, University ofGroningen, Groningen, 1987.
8M. Mezei, J. Chern. Phys. 86, 7084 (1987).
9W. L. Jorgensen and C. Ravimohan, J. Chern. Phys. 83, 3050 (1985).
lOT. P. Straatsma, H. J. C. Berendsen, andJ. P. M. Postma, J. Chern. Phys.
85,6720 (1986).
II T. P. Straatsma and H. J. C. Berendsen, J. Chern. Phys. 89, 5876 (1988).
12
0. A. Pearlman and P. A. Kollman, J. Chern. Phys. 91, 7831 (1989).
13T. P. Straatsma and J. A. McCammon, J. Chern. Phys. 90, 3300 (1989).
"T. P. Straatsma and J. A. McCammon, J. Chern. Phys. 91, 3631 (1989).
"D. J. Tobias, C. L. Brooks, and S. H. Fleishman, Chern. Phys. Lett. 156,
256 (1989).
16 P. A. Bash, U. C. Singh, R. Langridge, and P. A. Kollman, Science 36,
564 (1987).
11 J. A. McCammon and S. C. Harvey, in Dynamics of Proteins and Nucleic
Acids (Cambridge University, Cambridge, 1987).
18 J. van Eerden, W. J. Briels, S. Harkema, and D. Feil, Chern. Phys. Lett.
164,370 (1989).
19 W. F. van GUnsteren, in Computer Simulation of Biomolecular Systems.
Theoretical and Experimental Applications, edited by W. F. van Gun-
steren and P. K. Weiner (ESCOM, Leiden, 1989), p. 27.
20W. L. Jorgensen and J. K. Buckner, J. Phys. Chern. 90, 4651 (1986).
21 W. L. Jorgensen and J. K. Buckner, J. Phys. Chern. 91, 6083 (1987).
22T. P. Straatsma andJ. A. McCammon, J. Compo Chern. 11,943 (1990).
23H. J. C. Berendsen, J. P. M. Postma, W. F. van Gunsteren, A. DiNo1a,
and J. R. Haak, J. Chern. Phys. 81, 3684 (1984).
J. Chem. Phys., Vol. 95, No.2, 15 July 1991
Downloaded 11 Apr 2012 to 129.215.149.92. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/about/rights_and_permissions
1188 T. P. Straatsma and J. A. McCammon: Multiconfiguration thermodynamic integration
24H. J. C. Berendsen, J. R. Grigera, and T. P. Straatsma, J. Phys. Chern. 91,
6269 (1987).
2 ~ H. J. C. Berendsen and W. F. van Gunsteren, in The Physics ojSuperionic
Conductors and Electrode Materia/s, NATO ASI Ser. B: Physics, Vol. 92,
edited by J. W. Perram (Plenum, London, 1983), p. 221.
26 J. P. Ryckaert, G. Ciccotti, and H. J. C. Berendsen, J. Compo Phys. 23,
327 (1977).
21W. F. van Gunsteren and H. J. C. Berendsen, Mol. Phys. 34,1311 (1977).
28Groningen Molecular Simulation program manual, version GROMS7.
BIOMOS B. Y., Groningen, The Netherlands.
29T. P. Straatsma, H. J. C. Berendsen, and A. J. Starn, Mol. Phys. 57, 89
(1986).
301. Ohmine, H. Tanaka, and P. G. Wolynes, J. Chern. Phys. 89, 5852
(1988).
J. Chem. Phys., Vol. 95, No.2, 15 July 1991
Downloaded 11 Apr 2012 to 129.215.149.92. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/about/rights_and_permissions

Вам также может понравиться