Advanced Review

Umbrella sampling
Johannes Kästner∗

The calculation of free-energy differences is one of the main challenges in com-

putational biology and biochemistry. Umbrella sampling, biased molecular dy-
namics (MD), is one of the methods that provide free energy along a reaction
coordinate. Here, the method is derived in a historic overview and is compared
with related methods like thermodynamic integration, slow growth, steered MD,
or the Jarzynski-based fast-growth technique. In umbrella sampling, bias po-
tentials along a (one- or more-dimensional) reaction coordinate drive a system
from one thermodynamic state to another (e.g., reactant and product). The in-
termediate steps are covered by a series of windows, at each of which an MD
simulation is performed. The bias potentials can have any functional form. Often,
harmonic potentials are used for their simplicity. From the sampled distribution
of the system along the reaction coordinate, the change in free energy in each
window can be calculated. The windows are then combined by methods like the
weighted histogram analysis method or umbrella integration. If the bias potential
is adapted to result in an even distribution between the end states, then this whole
range can be spanned by one window (adaptive-bias umbrella sampling). In this
case, the free-energy change is directly obtained from the bias. The sampling in
each window can be improved by replica exchange methods; either by exchange
between successive windows or by running additional simulations at higher
temperatures. C 2011 John Wiley & Sons, Ltd. WIREs Comput Mol Sci 2011 1 932–942 DOI:
10.1002/wcms.66

INTRODUCTION tive constant to Q, which can be ignored. Then, Q is

obtained as
T he calculation of free-energy differences is a cen-
tral task in computational science. The free-
energy difference is the driving force of any pro-

Q = exp[−β E(r )]d Nr (1)
cess, such as a chemical reaction. Transition state with β = 1/(kB T), kB being the Boltzmann’s constant,
theory1–3 can be used to calculate reaction rates from T being the absolute temperature, and N being the
energy barriers, more exactly free-energy barriers.4 number of degrees of freedom of the system.
The free energy contains the entropy, a measure for The free (Helmholtz) energy A is related to Q
the available space. To map the available space in a via A = −1/β ln Q. The canonical partition func-
system bigger than a few atoms, extensive sampling is tion involves a constant number of particles, constant
required.5,6 Techniques are regularly being reviewed volume, and a constant temperature. If the pressure,
in the literature, with a few recent ones given in rather than the volume, is kept constant, the Gibbs
J Comput Chem7 and in a themed issue of J Comput free energy (usually denoted as G) is obtained. Apart
Chem in 2009.8,9 Applications range from the solid from the change in the ensemble, the following for-
state, catalytic reactions, biochemical processes to ra- malisms and derivations are equivalent for A and G.
tional drug design. In the condensed phase, which is relevant for most
The canonical partition function Q of a sys- applications, the systems are hardly compressible; so
tem can be calculated via an integral over the whole A and G are numerically very similar.
phase space, i.e., configuration space and momentum In chemical reactions, one is generally interested
space. If the potential energy E is independent of the in free-energy differences between two states. If the
momentum, the integral over the latter is a multiplica- two states differ by geometry (like a reactant and
product of a reaction) then the integration in Eq. (1)
∗
Correspondence to: [email protected] is done over a part of the coordinate space for each
Computational Biochemistry Group, Institute of Theoretical state.
Chemistry, University of Stuttgart, Stuttgart, Germany In many cases, a reaction coordinate (ξ ), a
DOI: 10.1002/wcms.66 continuous parameter which provides a distinction

932 
c 2011 John Wiley & Sons, Ltd. Volume 1, November/December 2011
 WIREs Computational Molecular Science Umbrella sampling

between two thermodynamic states, can be defined.

Any order-parameter is possible, even a change in the
energy expression (the Hamiltonian). The reaction
coordinate can be one or more dimensional. Often,
ξ is defined on geometric grounds, such as distance,
torsion, or the difference between root mean square
deviations from two reference states.
With ξ defined, the probability distribution of
the system along ξ can be calculated by integrating
out all degrees of freedom but ξ :

δ[ξ (r ) − ξ ] exp[(−β E)d Nr ]
Q(ξ ) =  . (2)
exp[(−β E)d Nr ]

Q(ξ ) dξ can be interpreted as the probability of find-

ing the system in a small interval dξ around ξ . Conse-
quently, this allows the calculation of the free energy
F I G U R E 1 | Separation of the reaction coordinate (dashed line)
along the reaction coordinate; A(ξ ) = −1/β ln Q(ξ ).
between two states (here represented by two minima on the potential
A(ξ ) is also called potential of mean force (PMF). If energy surface) into distinct windows. The system is mainly sampled
ξ is a general, i.e., non-Cartesian, coordinate, or a perpendicular to the reaction coordinate in each window.
set of those, a Jacobian term enters Eq. (2). As long
as the integration
 is performed in Cartesian coordi-
nates, Q(ξ ) = δ[ξ (r) − ξ ] exp(−βE) d Nr/Q. If the out this article, the term distribution P(ξ ) refers to
integration is done in a different set of coordinates, the normalized frequency of finding the system in the
q, ξ being one of those, an explicit Jacobian determi- vicinity of a given value of ξ . If P(ξ ) was obtained
nant |J(q)|, with Jij = dqi /drj , has to be taken into from an exact ensemble average rather than a sampled
account: Q(ξ ) = exp(−βE)|J(q)|d N−1 q/Q, where quantity, P(ξ ) would refer to a probability density.
the integration is performed over all coordinates However, simulations are only run for finite
except ξ . time. Regions in configuration space around a min-
In computer simulations, the direct phase-space imum in E(r) are typically sampled well, whereas
integrals used in Eqs. (1) and (2) are impossible to cal- regions of higher energy are sampled rarely. For
culate. However, if the system is ergodic, i.e., if every rare events, those with an energy barrier significantly
point in phase space is visited during the simulation, larger than kB T, direct sampling is infeasible. To ob-
Q(ξ ) is equal to tain a profile A(ξ ), however, also those high-energy
regions, those rare events, are required.
t Different techniques have been developed to
1
P(ξ ) = lim ρ[ξ (t  )]dt  (3) sample such rare events. One can broadly distinguish
t→∞ t
0 three different families of methods: (1) methods that
sample the system in equilibrium, (2) nonequilibrium
that is, the ensemble average Q(ξ ) becomes equal to sampling techniques, and (3) methods that introduce
the time average P(ξ ) for infinite sampling in an er- additional degrees of freedom, along which the free
godic system. In Eq. (3), t denotes the time and ρ energy is calculated. The third family includes λ-
simply counts the occurrence of ξ in a given inter- dynamics9–12 and metadynamics.13 The latter is cov-
val (of infinitesimal width in the exact equation and ered in a different contribution in this series and will
of finite width when calculating a histogram). So, in not be discussed here.
principle, A(ξ ) can be directly obtained from molec- In the remaining two families of methods, global
ular dynamics (MD) simulations by monitoring P(ξ ), sampling can be approximated by two techniques,
the distribution of the system along the reaction co- schematically illustrated in Figure 1. On the one hand,
ordinate. the path is split into windows. Each window covers
Note that the terms distribution, distribution only a small part of the range of ξ . The windows are
function, frequency, probability density, and possi- sampled individually. In postprocessing, the results
bly a few more are sometimes used in the literature of of the different windows are combined to result in a
chemistry and physics in different contexts. Through- global free-energy profile A(ξ ). On the other hand,

Volume 1, November/December 2011 

c 2011 John Wiley & Sons, Ltd. 933
Advanced Review wires.wiley.com/wcms

one can run multiple simulations. In each of those, The force on the frozen reaction coordinates is sam-
the system is driven from one state of interest (A) to pled. The resulting mean force is the derivative of the
the other state (B), taking a different path each time. free energy with respect to the reaction coordinate.
The postprocessing in this case includes averaging Integration of the mean force results in the PMF.
over the different simulations. It should be noted that there is some confusion
This review is organized as follows. In Accel- in the older literature over the term PMF.24 Especially
erated Sampling Techniques, the context of umbrella in the field of thermodynamic integration, one often
sampling is described by briefly introducing differ- referred to PMF as a quantity directly obtained by
ent techniques to sample free-energy profiles in MD integrating the mean force, which differs from the free
simulations. In Umbrella Sampling: Method, the me- energy by neglect of a correction of the metric tensor.
thodical details of umbrella sampling are derived and If the reaction coordinate is a spatial coordinate or a
described. Bias Potentials deals with different bias po- combination of those, constraining it to a fixed value
tentials, whereas Sampling Techniques mentions tech- also changes the momentum sampling. There is one
niques to accelerate the MD sampling itself. Methods component of the momentum canonically conjugated
to Analyze Umbrella Sampling Simulations discusses to each component of the spatial coordinates. If ξ is
the most common methods to analyze umbrella sam- frozen, the associated momentum is frozen (zero) as
pling simulations, i.e., to extract a free-energy profile well. This can lead to a change in the metric tensor of
from the sampled data. Conclusion finally summa- the system. For simple reaction coordinates, metric-
rizes the topic. tensor corrections have been derived.21,25–31
Rather than keeping the reaction coordinate
fixed in a number of windows, one can also vary
ACCELERATED SAMPLING a constraint slowly from one state to another in an
TECHNIQUES approach termed slow growth.32–34 Average and in-
tegration of the force on the constraint results in the
Accelerated Sampling Techniques Based free energy.
on Equilibrium Properties In umbrella sampling,14,15 the reaction coor-
All of the acceleration techniques discussed in this re- dinate is not constrained, but only restrained and
view have the goal of calculating the free-energy dif- pulled to a target value by a bias potential. There-
ference, an equilibrium property. However, in each fore, the full momentum space is sampled. Usually,
simulation, the system can either be in equilibrium, umbrella sampling is done in a series of windows,
or one can measure the response to some pertur- which are finally combined either with the weighted
bation and derive the free-energy change from that. histogram analysis method (WHAM)35,36 or using
The former case will be discussed here, the latter in umbrella integration.37
Free-Energy Differences from Nonequilibrium Simu- The bias potential can be varied to pull the sys-
lations. tem from one state to another rather than keeping
To drive a system over an energy barrier, one it fixed. If that variation is slow as compared with
can either (1) modify the energy expression in order the relaxation time of the system, the analysis can
to reduce the barrier, or (2) restrict the sampling space be performed by assuming an equilibrium state of the
to all degrees of freedom, but the reaction coordinate system, i.e., the mean force on the reaction coordinate
describing the transition over the barrier. The former can be sampled and integrated. This approach gained
is known as biased MD or umbrella sampling.14,15 popularity under the name of steered MD (SMD).38–41
Because this is the main focus of the present review, SMD directly simulates the influence of an atomic-
it will be discussed in detail in Umbrella Sampling: force microscope cantilever acting, e.g., on a protein.
Method.
In thermodynamic integration,16–21 a tech-
nique sometimes also referred to as blue moon Free-Energy Differences from
sampling,18,22 the transition over a barrier is simu- Nonequilibrium Simulations
lated by freezing the reaction coordinate at differ- Jarzynski42 demonstrated the equivalence of the free-
ent values in a number of windows and sampling energy change and an exponential average over the
the system perpendicular to ξ . The constraint freez- work W along nonreversible paths originating from a
ing the reaction coordinate has to be implemented in canonic ensemble:
an energy-conserving manner. Generally, the method
of Lagrange multipliers (Shake algorithm23 ) is used. exp(−βA) = exp(−βW). (4)

934 
c 2011 John Wiley & Sons, Ltd. Volume 1, November/December 2011
 WIREs Computational Molecular Science Umbrella sampling

This can be exploited in practical simulations by tion coordinate:

moving a constraint on the reaction coordinate rel-
atively fast from an equilibrated system to the tar- E b (r ) = E u (r ) + wi (ξ ). (6)
get system.43,44 This method became known as
fast growth. It is related to Bennett’s acceptance The superscript ‘b’ denotes biased quantities, whereas
ratio method.45 The changes in the energy along the superscript ‘u’ denotes unbiased quantities. Quan-
these paths are averaged according to Eq. (4). The tities without superscripts are always unbiased.
computational tradeoff is that the faster the constraint In order to obtain the unbiased free energy Ai (ξ ),
is varied, the larger is the statistical spread, and thus, we need the unbiased distribution, which is, according
more trajectories have to be calculated. to Eq. (2):
Free-energy perturbation46 (FEP) can be re- 
garded as a limiting case of methods based on exp[−β E(r )] δ[ξ  (r ) − ξ ] d Nr
Pi (ξ ) =
u  . (7)
Jarzynski’s equation: exp[−β E(r )] d Nr

exp(−βA) = exp(−βE)a (5) MD simulation of the biased system provides the bi-
ased distribution along the reaction coordinate Pib .
with E being the difference of the initial state and
Assuming an ergodic system,
the final state, and the ensemble averaged over the
initial state a. It should be noted, of course, that FEP 
exp{−β[E(r ) + ωi (ξ  (r ))]}δ[ξ  (r ) − ξ ]d Nr
was proposed and used many decades before the more Pi (ξ ) =
b  .
general Eq. (4). In FEP, the instantaneous change from exp{−β[E(r ) + ωi (ξ  (r ))]}d Nr
one state to another is sampled over a canonical en- (8)
semble. Thus, it corresponds to fast growth with the
constraint immediately moved to the target value. The Because the bias depends only on ξ and the in-
exponential average of the change results in the free- tegration in the enumerator is performed over all de-
energy difference. The term ‘perturbation’ is mislead- grees of freedom but ξ ,
ing because the method is exact and does not corre-
spond to a perturbation theory in the usual sense. Pib (ξ ) = exp[−βωi (ξ )]
A special challenge for free-energy simula- 
exp[−β E(r )]δ[ξ  (r ) − ξ ]d Nr
tions are quantum mechanics/molecular mechanics × . (9)
exp{−β[E(r ) + ωi (ξ  (r ))]}d Nr
(QM/MM) setups, in which a small part of the system
is described by comparatively expensive QM calcula- Using Eq. (7) results in
tions, whereas most of the system is handled by classi-
cal force fields (MM).47 A variant of FEP48–50 can be Piu (ξ ) = Pib (ξ ) exp[βωi (ξ )]
used to restrict the sampling to the computationally 
cheaper force field part. exp {−β [E(r ) + ωi (ξ (r ))]} d Nr
× 
exp [−β E(r )] d Nr
= Pib (ξ ) exp[βωi (ξ )]
UMBRELLA SAMPLING: METHOD 
exp[−β E(r )] exp{−β ωi [ξ (
r )]}d Nr
Umbrella sampling was developed by Torrie and × 
exp[−β E(r )]d Nr
Valleau14,15 based on related previous work.51,52 A
bias, an additional energy term, is applied to the sys- = Pib (ξ ) exp[βωi (ξ )] exp[−βωi (ξ )]. (10)
tem to ensure efficient sampling along the whole re-
action coordinate. This can either be aimed at in one From Eq. (10), Ai (ξ ) can be readily evaluated. Pib (ξ )
simulation or in different simulations (windows), the is obtained from an MD simulation of the bi-
distributions of which overlap. The effect of the bias ased system, wi (ξ ) is given analytically, and Fi =
potential to connect energetically separated regions in −(1/β) lnexp[−β ωi (ξ )] is independent of ξ :
phase space gave rise to the name umbrella sampling.
In this section, the formalism of recovering unbi- Ai (ξ ) = −(1/β) ln Pib (ξ ) − wi (ξ ) + Fi . (11)
ased free-energy differences from biased simulations
will be discussed. The next section describes different This derivation is exact. No approximation enters
forms of bias potentials used in the literature. apart from the assumption that the sampling in each
The bias potential wi of window i is an addi- window is sufficient. This is facilitated by an appro-
tional energy term, which depends only on the reac- priate choice of umbrella potentials wi (ξ ).

Volume 1, November/December 2011 

c 2011 John Wiley & Sons, Ltd. 935
Advanced Review wires.wiley.com/wcms

known; it is what we aim to calculate with umbrella

sampling. Therefore, two main families of bias po-
tentials have emerged: harmonic biases in a series of
windows along ξ , and an adaptive bias, which is ad-
justed to match −A(ξ ) in only one window spanning
the whole range of ξ .

Harmonic Bias Potentials

To ensure sampling in all regions of ξ , the range of
interest of ξ is split into a number of windows. In each
window, a bias function is applied to keep the system
close to the reference point ξ iref of the respective win-
dow i. Often, a simple harmonic bias of strength K is
F I G U R E 2 | Global free energy (thick solid curve) and the used:
contributions Ai of some of the windows (thin dashed curves). For
 2
clarity, only every third window is shown. At the bottom, the biased ωi (ξ ) = K/2 ξ − ξiref . (13)
distributions Pib as obtained from the simulation are shown (thin solid
curves). Relatively few bins (100) have been used to generate this After the simulations, the free-energy curves are
figure. combined with techniques discussed in Methods to
Analyze Umbrella Sampling Simulations (typically
WHAM or umbrella integration). The form of the
As long as one window spans the whole range
bias given in Eq. (13) is appealing because it contains
of ξ to be studied, Eq. (11) is sufficient to unbias the
only few parameters: K (which in principle can be
simulation. A(ξ ) is in any case only defined up to an
window dependent), the number of images, and ξ iref .
additive constant; so in this case, Fi can be chosen
The latter are usually chosen uniformly distributed
arbitrarily.
along ξ . The higher the number of images, the smaller
If the free-energy curves Ai (ξ ) of more windows
is generally the statistical error relative to CPU time.53
are to be combined to one global A(ξ ), see Figure 2,
However, the CPU time needed for equilibration, on
the Fi have to be calculated. They are associated
the contrary, increases with the number of images.
with introducing the bias potential and connect the
The MD simulations of the images are completely in-
free-energy curves Ai (ξ ) obtained in the different
dependent and thus, can run in parallel.
windows:
The choice of K, the strength of the bias, is the
exp(−β Fi ) = exp[−β ωi (ξ )] only critical decision. It has to be made before simula-
 tions are run. By contrast, additional windows could
= P u (ξ ) exp[−β ωi (ξ )]dξ always be inserted if the first series of windows results
 in too large gaps between the distributions. Overall,
= exp{−β[A(ξ ) + ωi (ξ )]}dξ (12) K has to be large enough to drive the system over the
barrier. Too large K, however, will cause very nar-
with Pu (ξ ) being the global unbiased distribution. The row distributions Pib (ξ ). Sufficient overlap between
Fi cannot directly be obtained from sampling. Meth- the distributions is required for WHAM, whereas it
ods to Analyze Umbrella Sampling Simulations will is not required, but still advantageous in umbrella
deal with methods to calculate them, i.e., to com- integration.54 Increasing K at constant time step also
bine the results of different windows in umbrella leads to increasing errors in the numerical integra-
sampling. tion of the equations of motions. If the time step is
too large (or K is too large), configurations with high
energies will be overrepresented.20
BIAS POTENTIALS For umbrella integration analysis, analytic ex-
pressions for the statistical error can be derived, which
Ideally, the bias potential is chosen such that sampling allow an estimate of an ideal K based on quantities,
along the whole range of the reaction coordinate ξ which can often be estimated prior to sampling.54
is uniform. Therefore, the optimal bias potential is It has also been suggested that the location of the
wopt = −A(ξ ). This would lead to a truly uniform next window to be sampled (ξ i+1 ref ) can be chosen
distribution Pib (ξ ). However, A(ξ ) is obviously not from the location and the widths of the previous

936 
c 2011 John Wiley & Sons, Ltd. Volume 1, November/December 2011
 WIREs Computational Molecular Science Umbrella sampling

window to match their estimated half maxima.55 An A too low barrier may be the result of discontinuities
alternative is to use data from the experiment to define in the path: the change from one window i to the next
the most promising bias parameters.56 window i + 1 may be reflected by only a small change
in ξ , but a larger change in other degrees of freedom
which are not included in ξ . Such artificial behavior
Adaptive Bias Umbrella Sampling results in jumps in the root mean square difference be-
The aim of adaptive bias umbrella sampling13,57–60 is tween the average structures of subsequent umbrella
to cover the whole range of interest of the reaction sampling windows.8
coordinate ξ in one simulation. In principle, this can
be achieved by choosing a bias w(ξ ) = −A(ξ ). This
exactly flattens the energy surface and leads to a uni- METHODS TO ANALYZE UMBRELLA
form sampling along ξ . Because A(ξ ) is, of course, not
SAMPLING SIMULATIONS
known a priori, one typically starts out with an initial
guess of w(ξ ) and iteratively improves it to achieve a Weighted Histogram Analysis Method
uniform distribution. (WHAM)
Numerous methods have been proposed for an es-
timation of Fi ,68,69 a promising one being70 the
Specialized Umbrella Potentials
WHAM.35,36 It aims to minimize the statistical er-
The local elevation method61 adds, similar to meta-
ror of Pu (ξ ). The global distribution is calculated by a
dynamics, a history-dependent (and thus, time-
weighted average of the distributions of the individual
dependent) bias to the potential energy. This has
windows:
recently62,63 been combined with umbrella sampling
by building up a local elevation bias in a compara- 
windows

tively short simulation and then sampling the distribu- P u (ξ ) = pi (ξ )Piu (ξ ). (14)
tion in that bias to reconstruct the free energy. Other i

special forms of umbrella potentials were used.64 The weights pi are chosen in order to minimize the
statistical error of Pu :
∂σ 2 (P u )
SAMPLING TECHNIQUES =0 (15)
∂ pi
In each window of an umbrella sampling run, the 
phase space has to be sampled as good as possible. under the condition pi = 1. This leads to35,36 :
Overlap between windows is required for WHAM ai
pi =  , ai (ξ ) = Ni exp[−β ωi (ξ ) + β Fi ] (16)
analysis (see below) and is desirable for umbrella inte- j aj
gration. The quality of the sampling can be enhanced
by Hamiltonian replica exchange.65–67 In specified in- with Ni being the total number of steps sampled for
tervals, the geometry of window i is used to calculate window i. The Fi are calculated by Eq. (12):

the total biased energy of a neighboring window j (i.e.,
the bias wj of window j is used), and additionally, the exp(−β Fi ) = P u (ξ ) exp[−βwi (ξ )] dξ. (17)
energy of geometry j with the bias wi is calculated.
If the sum of these energies is smaller than the sum Because Pu enters Eq. (17) and Fi enters Eq. (14) via
of the original energies, the two sets of coordinates Eq. (16), these have to be iterated until convergence.
are exchanged. If the sum is larger, exchange is still For many bins, this convergence can be slow.
possible based on the Metropolis criterion. Then the
simulations continue. This is done at regular intervals Umbrella Integration
with all pairs of images. Replica exchange between An alternative to WHAM for combining the windows
umbrella sampling windows enhances the quality of in umbrella sampling simulations with harmonic bi-
the sampling without additional computational cost. ases is umbrella integration.37 The problem of calcu-
The better the sampling, the more important is lating Fi is avoided by averaging the mean force rather
a proper choice of the reaction coordinate. If the reac- than the distribution P. The unbiased mean force is
tion coordinate misses important structural changes, independent of the Fi :
it can lead to artificial lowering or raising of the result
obtained by umbrella sampling.8 A too high barrier ∂ Aiu 1 ∂ ln Pib (ξ ) dwi
=− − . (18)
may be the result of an unfavorable path being taken. ∂ξ β ∂ξ dξ

Volume 1, November/December 2011 

c 2011 John Wiley & Sons, Ltd. 937
Advanced Review wires.wiley.com/wcms

The distribution Pib is expanded into a cumulant ex-

pansion, which is truncated after the second term (i.e.,
approximating Pib by a normal distribution). This is
equivalent to truncating a power series of Aiu (ξ ) after
the quadratic term. Because A(ξ ) can be assumed to
be smooth and because each window is only supposed
to cover a small part of ξ , truncating such a power
expansion is well justified:
⎡
2 ⎤
1 1 ξ − ξ b
Pib (ξ ) = b √ exp ⎣− i ⎦ . (19)
σi 2π 2 σib

Thus, and with a bias in the form of Eq. (13), Eq. (18)
now reads

∂ Aiu 1 ξ − ξb  
=  2i − K ξ − ξiref (20)
∂ξ β σb F I G U R E 3 | Weights of weighted histogram analysis method
i
(WHAM) and umbrella integration of three windows in a real
simulation of the enzyme para-hydroxybenzoate hydroxylase (PHBH).49
which only depends on the mean value ξib and the vari-
A maximum in the free energy is found between the second and third
ance (σ ib )2 of ξ in each window. These two quantities windows.
can easily be sampled. For one window, before com-
bining the different windows, its integration yields:
 2
 WHAM. (3) The (non-normalized) weights for com-
ξ − ξib 1
Aiu (ξ ) = bining the windows are different: ai (ξ ) = Ni exp(−β
 2 − K
2 β σib wi (ξ ) + βFi ) = Ni Pib /Piu in WHAM and ai (ξ ) = Ni Pib
    in umbrella integration.
+ ξ − ξib K ξiref − ξib + Ci . (21) The first point is the main difference between
the methods. The second difference can be changed
ξib shifts Aiu (ξ ) along the ξ axis and determines its in either of the methods. If umbrella integration is
applied on the whole distribution, its noise level in-
slope, whereas (σ ib )2 determines the curvature of Aiu ,
creases generally above the one obtained by WHAM.
and Ci is just the integration constant.
The additional differentiation adds to the noise. Also,
The curves of the mean forces of the different
its convergence properties with the bin width (num-
windows can directly be averaged to result in a global
ber of bins) are lost. On the contrary, WHAM was
mean force:
meanwhile used with Pib approximated by normal
∂A 
windows
∂ Au distributions.37,71 This leads to free-energy profiles as
= pi (ξ ) i . (22) smooth as those obtained from umbrella integration.
∂ξ ∂ξ
i The weights used by the different methods are
somewhat difficult to transform between the methods
This is conveniently done with (normalized) weights
because they weight different quantities. However,
proportional to Pib :
weights used in real simulations can be compared as
ai depicted in Figure 3. In both cases, analytic quanti-
pi (ξ ) =  , ai (ξ ) = Ni Pib (ξ ). (23) ties, not directly dependent on histograms, are used
j aj
in ai . Thus, the weights are smooth curves even if
The resulting global mean force can be numerically the distributions are noisy. It is clear from Figure 3
integrated. that the weights used in WHAM are broader than the
The difference between WHAM and umbrella ones used in umbrella integration. Using the weights
integration is threefold: (1) The unbiased distribu- of umbrella integration ai (ξ ) = Ni Pib in WHAM is
tions of the images are averaged in WHAM, whereas possible, but it results in noisier curves because Pib is
the mean force is averaged in umbrella integration. directly obtained from histograms in WHAM. It also
(2) The biased distributions are approximated by nor- leads to slightly deteriorated free-energy profiles as
mal distributions in umbrella integration, but not in the windows effectively overlap less. Strong overlap

938 
c 2011 John Wiley & Sons, Ltd. Volume 1, November/December 2011
 WIREs Computational Molecular Science Umbrella sampling

between the windows is more important in WHAM Estimation of the Sampling Error Bar
than in umbrella integration. The error bar from finite sampling can be estimated
As a special case of umbrella integration, one using umbrella integration analysis.54 This, in turn,
can truncate the power series of A(ξ ) already after allows to set up guidelines as to how the necessary
the linear term19,37,72–75 : simulation parameters should be set. The most im-
∂ Aiu   portant parameter is K, the strength of the bias. In
= −K ξib − ξiref . (24) general, K should be chosen as small as possible to
∂ξ allow for much overlap between the images. Let us
Comparison with Eq. (20) shows that this is accurate introduce κ as the negative second derivative of the
free energy with respect to ξ at the main barrier. Then,
for ξ = ξib . However, it has also been used for ξ =
K > κ is necessary to ensure a unimodular distribu-
ξ iref .
tion in all images. This is necessary in umbrella inte-
The expressions of umbrella integration allow
gration because these distributions are approximated
for an estimate of the statistical error in A from
by normal distributions. In WHAM, sampling over
MD simulation data.54 This, in turn, can be used to
the barrier is necessary, resulting in K > κ − kB T. Of
chose the parameters of the simulation, such as the
course, κ is not known a priori, but sometimes it can
strength of the bias K and the number of windows, in
be estimated.54
order to minimize the statistical error while keeping
In general, it is preferable to sample many
the requirement for CPU time at bay.
windows for shorter times than fewer windows for
Umbrella integration can also be performed
longer.79 This leads to a smaller statistical error be-
in multidimensional reaction coordinates.76,77 How-
cause of the better overlap between the windows and
ever, the necessary integration step becomes more dif-
is better parallelizable.
ficult (and prone to statistical error) in higher dimen-
sions, whereas the alternative WHAM analysis can
more straight forwardly be extended to more dimen-
CONCLUSION
sions.
The main advantages of umbrella integration The question whether umbrella sampling or one of its
over WHAM are the independence of the number of related methods discussed in Accelerated Sampling
grid points (bins) and the availability of an error es- Techniques is to be used cannot be answered in gen-
timate. The fact that only ξ and σ 2 enter the analysis eral. It may depend on the particular system. Some
of umbrella integration can be used to test the MD authors have compared the applicability of some of
runs for equilibration78 of these two quantities. This these methods.44,80–82 Although umbrella sampling
cannot directly be done for WHAM, where the whole might be preferred over thermodynamic integration
distribution enters the analysis. However, in princi- because of errors in the integration80 (which reduce
ple, one could test for equilibration of ξ and σ 2 , and with more windows), the additional free parameter
when these are equilibrated, assume that the whole K, which has to be chosen in umbrella sampling, was
distribution is equilibrated. Additionally, umbrella in- used as an argument against the latter.82 Additionally,
tegration is noniterative, which speeds up the analysis. the availability of a correction of the metric tensor for
However, the CPU time required for the analysis is, in the particular choice of the reaction coordinate might
general, negligible as compared with the time needed be an argument in favor of umbrella sampling, where
to acquire the MD sampling data. The reduction of such a correction is unnecessary.
Ai (ξ ) to second order in ξ reduces noise significantly. Overall, umbrella sampling is meanwhile a ma-
For cases with very few windows, however, this can ture and broadly accepted method for calculating
become a source of inaccuracies. free-energy differences.

ACKNOWLEDGMENTS

The author thanks the German Research Foundation (DFG) for financial support within the
Cluster of Excellence in Simulation Technology (EXC 310/1) and grant SFB716/C.6, both at
the University of Stuttgart.

Volume 1, November/December 2011 

c 2011 John Wiley & Sons, Ltd. 939
Advanced Review wires.wiley.com/wcms

REFERENCES
1. Arrhenius S. Über die Reaktionsgeschwindigkeit bei 19. van Gunsteren WF. Methods for calculation of free en-
der Inversion von Rohrzucker durch Säuren. Z Phys ergies and binding constants: successes and problems.
Chem (Leipzig) 1889, 4:226–248. In: van Gunsteren WF, Weiner PK, eds. Computer Sim-
2. Eyring H. The activated complex in chemical reactions. ulation of Biomolecular Systems. Vol. 1 Leiden: ES-
J Chem Phys 1935, 3:107–115. COM; 1989, 27.
3. Evans MG, Polanyi M. Some applications of the tran- 20. Straatsma TP, McCammon JA. Multiconfiguration
sition state method to the calculation of reaction veloc- thermodynamic integration. J Chem Phys 1991,
ities, especially in solution. Trans Faraday Soc 1935, 95:1175–1188.
31:875–894. 21. Sprik M, Ciccotti G. Free energy from constrained
4. Chandler D. Statistical mechanics of isomerization dy- molecular dynamics. J Chem Phys 1998, 109:7737–
namics in liquids and the transition state approxima- 7744.
tion. J Chem Phys 1978, 68:2959–2970. 22. Ciccotti G, Ferrario M. Blue moon approach to rare
5. Kollman P. Free energy calculations: Applications events. Mol Sim 2004, 30:787–793.
to chemical and biochemical phenomena. Chem Rev 23. Ryckaert JP, Ciccotti G, Berendsen HJC. Numerical
1993, 93:2395–2417. integration of the Cartesian equations of motion of
6. Beveridge DL, DiCapua FM. Free energy via molecular a system with constraints: Molecular dynamics of n-
simulation: Applications to chemical and biomolecu- alkanes. J Comput Phys 1977, 23:327–341.
lar systems. Annu Rev Biophys Biophys Chem 1989, 24. Mülders T, Krüger P, Swegat W, Schlitter J. Free energy
18:431–492. as the potential of mean constraint force. J Chem Phys
7. Christ CD, Mark AE, van Gunsteren WF. Basic ingre- 1996, 104:4869–4870.
dients of free energy calculations: A review. J Comput 25. Jain A. Compensating mass matrix potential for con-
Chem 2010, 31:1569–1582. strained molecular dynamics. J Comput Phys 1997,
8. Rosta E, Woodcock HL, Brooks BR, Hummer G. Ar- 136:289–297.
tificial reaction coordinate “tunneling” in free-energy 26. den Otter WK, Briels WJ. The calculation of
calculations: The catalytic reaction of RNase H. J Com- free-energy differences by constrained molecular-
put Chem 2009, 30:1634–1641. dynamics simulations. J Chem Phys 1998, 109:4139–
9. Knight JL, Brooks III CL. λ-Dynamics free energy simu- 4146.
lation methods. J Comput Chem 2009, 30:1692–1700. 27. den Otter WK. Thermodynamic integration of the free
10. Kong X, Brooks CL III. λ-Dynamics: A new ap- energy along a reaction coordinate in cartesian coordi-
proach to free energy calculations. J Chem Phys 1996, nates. J Chem Phys 2000, 112:7283–7292.
105:2414–2423. 28. Darve E, Pohorille A. Calculating free energies using
11. Liu Z, Berne BJ. Method for accelerating chain folding average force. J Chem Phys 2001, 115:9169–9183.
and mixing. J Chem Phys 1993, 99:6071–6077. 29. Schlitter J, Klähn M. A new concise expression for the
12. Tidor B. Simulated annealing on free energy surfaces free energy of a reaction coordinate. J Chem Phys 2003,
by a combined molecular dynamics and Monte Carlo 118:2057–2060.
approach. J Phys Chem 1993, 97:1069–1073. 30. Schlitter J, Klähn M. The free energy of a reaction coor-
13. Laio A, Parrinello M. Escaping free-energy minima. dinate at multiple constraints: A concise formulation.
Proc Natl Acad Sci U S A 2002, 99:12562–12566. Mol Phys, 2003, 101:3439–3443.
14. Torrie GM, Valleau JP. Monte Carlo free energy esti- 31. Hénin J, Chipot C. Overcoming free energy barriers
mates using non-Boltzmann sampling: Application to using unconstrained molecular dynamics simulations.
the sub-critical Lennard-Jones fluid. Chem Phys Lett J Chem Phys 2004, 121:2904–2914.
1974, 28:578–581. 32. Straatsma TP, Berendsen HJC, Postma JPM. Free en-
15. Torrie GM, Valleau JP. Nonphysical sampling distribu- ergy of hydrophobic hydration: A molecular dynam-
tions in Monte Carlo free-energy estimation: Umbrella ics study of noble gases in water. J Chem Phys 1986,
sampling. J Comput Phys 1977, 23:187–199. 85:6720–6727.
16. Born M. Volumen und Hydratationswärme der Ionen. 33. Beveridge DL, DiCapua FM. Free energy via molecular
Z Phys 1920, 1:45–48. simulatin: a primer. In: van Gunsteren WF, Weiner PK,
17. Kirkwood JG. Statistical mechanics of fluid mixtures. eds. Computer Simulation of Biomolecular Systems.
J Chem Phys 1935, 3:300–313. Vol. 1. Leiden, The Netherlands: ESCOM; 1989, 1–
18. Carter EA, Ciccotti G, Hynes JT, Kapral R. Con- 26.
strained reaction coordinate dynamics for the simu- 34. Hermans J. Simple analysis of noise and hysteresis in
lation of rare events. Chem Phys Lett 1989, 156:472– (slow-growth) free energy simulations. J Phys Chem
477. 1991, 95:9029–9032.

940 
c 2011 John Wiley & Sons, Ltd. Volume 1, November/December 2011
 WIREs Computational Molecular Science Umbrella sampling

35. Kumar S, Rosenberg JM, Bouzida D, Swendsen RH, from QM/MM free-energy simulations. Can J Chem
Kollman PA. The weighted histogram analysis method 2009, 87:1322–1337.
for free-energy calculations on biomolecules. I. The 51. McDonald IR, Singer K. Machine calculation of ther-
method. J Comput Chem 1992, 13:1011–1021. modynamic properties of a simple fluid at supercritical
36. Souaille M, Roux B. Extension to the weighted his- temperatures. J Chem Phys 1967, 47:4766–4772.
togram analysis method: Combining umbrella sam- 52. McDonald IR, Singer K. Examination of the ade-
pling with free energy calculations. Comput Phys Com- quacy of the 12–6 potential for liquid argon by means
mun, 2001 135:40–57. of Monte Carlo calculations. J Chem Phys 1969,
37. Kästner J, Thiel W. Bridging the gap between thermo- 50:2308–2315.
dynamic integration and umbrella sampling provides 53. Frenkel D, Smit B. Understanding Molecular Simula-
a novel analysis method: “Umbrella integration”. J tion: From Algorithms to Applications. 2nd ed. San
Chem Phys 2005, 123:144104(1–5). Diego, CA: Academic Press; 2002.
38. Grubmüller H, Heymann B, Tavan P. Ligand 54. Kästner J, Thiel W. Analysis of the statistical error
binding and molecular mechanics calculation of in umbrella sampling simulations by umbrella integra-
the streptavidin-biotin rupture force. Science 1996, tion. J Chem Phys 2006, 124:234106(1–7).
271:997–999.
55. Grossfield A, Woolf TB. Interaction of tryptophan
39. Izrailev S, Stepaniants S, Balsera M, Oono Y, Schul- analogs with POPC lipid bilayers investigated by
ten K. Molecular dynamics study of unbinding of molecular dynamics calculations. Langmuir 2002,
the avidin-biotin complex. Biophys J 1997, 72:1568– 18:198–210.
1581.
56. Mills M, Andricioaei I. An experimentally guided um-
40. Evans E, Ritchie K. Dynamic strength of molecular brella sampling protocol for biomolecules. J Chem
adhesion bonds. Biophys J 1997, 72:1541–1555. Phys 2008, 129:114101(1–13).
41. Balsera M, Stepaniants S, Izrailev S, Oono Y, Schul- 57. Mezei M. Adaptive umbrella sampling: Self-consistent
ten K. Reconstructing potential energy functions from determination of the non-Boltzmann bias. J Comput
simulated force-induced unbinding processes. Biophys Phys 1987, 68:237–248.
J, 1997, 73:1281–1287.
58. Hooft RWW, van Eijck BP, Kroon J. An adaptive um-
42. Jarzynski C. Nonequilibrium equality for free energy brella sampling procedure in conformational analysis
differences. Phys Rev Lett 1997, 78:2690–2693. using molecular dynamics and its application to glycol.
43. Hummer G. Fast-growth thermodynamic integration: J Chem Phys 1992, 97:6690–6694.
Error and efficiency analysis. J Chem Phys 2001, 59. Bartels C, Karplus M. Multidimensional adaptive um-
114:7330–7337. brella sampling: Applications to main chain and side
44. Zuckerman DM, Woolf TB. Theory of a systematic chain peptide conformations. J Comput Chem 1997,
computational error in free energy differences. Phys 18:1450–1462.
Rev Lett 2002, 89:180602–180606. 60. Bartels C, Karplus M. Probability distributions for
45. Bennett CH. Efficient estimation of free energy differ- complex systems: Adaptive umbrella sampling of the
ences from Monte Carlo data. J Comput Phys 1976, potential energy. J Phys Chem B 1998, 102:865–
22:245–268. 880.
46. Zwanzig RW. High-temperature equation of state by a 61. Huber T, Torda AE, van Gunsteren WF. Local eleva-
perturbation method. I. Nonpolar gases. J Chem Phys tion: A method for improving the searching properties
1954, 22:1420–1426. of molecular dynamics simulation. J Comput Aided
47. Hu H, Yang W. Free energies of chemical reactions Mol Design 1994, 8:695–708.
in solution and in enzymes with ab initio quantum 62. Hansen HS, Hünenberger PH. Using the local elevation
mechanics/molecular mechanics methods. Annu Rev method to construct optimized umbrella sampling po-
Phys Chem 2008, 59:573–601. tentials: Calculation of the relative free energies and in-
48. Zhang Y, Liu H, Yang W. Free energy calculation on terconversion barriers of glucopyranose ring conform-
enzyme reactions with an efficient iterative procedure ers in water. J Comput Chem 2010, 31:1–23.
to determine minimum energy paths on a combined ab 63. Hansen HS, Hünenberger PH. Ball-and-stick local el-
initio QM/MM potential energy surface. J Chem Phys evation umbrella sampling: Molecular simulations in-
2000, 112:3483–3493. volving enhanced sampling within conformational or
49. Kästner J, Senn HM, Thiel S, Otte N, Thiel W. alchemical subspaces of low internal dimensionali-
QM/MM free-energy perturbation compared to ther- ties, minimal irrelevant volumes, and problem-adapted
modynamic integration and umbrella sampling: Appli- geometries. J Chem Theory Comput 2010, 6:2622–
cation to an enzymatic reaction. J Chem Theory Com- 2646.
put 2006, 2:452–461. 64. Wu D. An efficient umbrella potential for the accurate
50. Senn HM, Kästner J, Breidung J, Thiel W. Finite- calculation of free energies by molecular simulation. J
temperature effects in enzymatic reactions—Insights Chem Phys 2010, 133:044115(1–9).

Volume 1, November/December 2011 

c 2011 John Wiley & Sons, Ltd. 941
Advanced Review wires.wiley.com/wcms

65. Sugita Y, Okamoto Y. Replica-exchange molecular dy- 74. Maragliano L, Fischer A, Vanden-Eijnden E, Ciccotti
namics method for protein folding. Chem Phys Lett G. String method in collective variables: Minimum free
1999, 314:141–151. energy paths and isocommittor surfaces. J Chem Phys
66. Sugita Y, Kitao A, Okamoto Y. Multidimensional 2006, 125:024106(1–15).
replica-exchange method for free-energy calculations. 75. Maragliano L, Vanden-Eijnden E. Single-sweep meth-
J Chem Phys 2000, 113:6042–6051. ods for free energy calculations. J Chem Phys 2008,
67. Curuksu J, Zacharias M. Enhanced conformational 128:184110(1–10).
sampling of nucleic acids by a new hamiltonian replica 76. Kästner J. Umbrella integration in two or more reaction
exchange molecular dynamics approach. J Chem Phys coordinates. J Chem Phys 2009, 131:034109(1–8).
2009, 130:104110(1–8). 77. Kästner J, Sherwood P. The ribosome catalyses peptide
68. Ferrenberg AM, Swendsen RH. New Monte Carlo bond formation by providing high ionic strength. Mol
technique for studying phase transitions. Phys Rev Lett Phys 2010, 108:293–306.
1988, 61:2635–2638. 78. Schiferl SK, Wallace DC. Statistical errors in molecu-
69. Ferrenberg AM, Swendsen RH. Optimized Monte lar dynamics averages. J Chem Phys 1985, 83:5203–
Carlo data analysis. Phys Rev Lett 1989, 63:1195– 5209.
1198. 79. Beutler TC, van Gunsteren WF. The computation of a
70. Roux B. The calculation of the potential of mean force potential of mean force: Choice of the biasing potential
using computer simulations. Comput Phys Commun in the umbrella sampling technique. J Chem Phys 1994,
1995, 91:275–282. 100:1492–1497.
71. Chakravorty DK, Kumarasiri M, Soudackov AV, 80. Czaplewski C, Rodziewicz-Motowidlo S, Liwo A,
Hammes-Schiffer S. Implementation of umbrella in- Ripoll DR, Wawak RJ, Scheraga HA. Molecular sim-
tegration within the framework of the empirical va- ulation study of cooperativity in hydrophobic associa-
lence bond approach. J Chem Theory Comput 2008, tion. Protein Sci 2000, 9:1235–1245.
4:1974–1980. 81. Shirts MR, Pande VS. Comparison of efficiency
72. Van Eerden J, Briels WJ, Harkema S, Feil D. Po- and bias of free energies computed by exponential av-
tential of mean force by thermodynamic integration: eraging, the Bennett acceptance ratio, and thermody-
Molecular-dynamics simulation of decomplexation. namic integration. J Chem Phys 2005, 122:144107(1–
Chem Phys Lett 1989, 164:370–376. 15).
73. Billeter SR, van Gunsteren WF. Computer simulation 82. Trzesniak D, Kunz APE, van Gunsteren WF. A com-
of proton transfers of small acids in water. J Phys Chem parison of methods to compute the potential of mean
A 2000, 104:3276–3286. force. Chem Phys Chem 2007, 8:162–169.

942 
c 2011 John Wiley & Sons, Ltd. Volume 1, November/December 2011