Introduction to Free Energy

Methods
 Why Do This?
• The free energy of a system is perhaps the most
important thermodynamic quantity, and is usually
taken as the Helmholtz or Gibb’s free energy
• Techniques to calculate the free energy (or
relative free energy) of a system are very useful
studying phase transitions, critical phenomena
or other transformations
• We can never calculate absolute free energies
(since we don’t have an appropriate reference
state), however relative free energies can be
found using several different computational
techniques
 Calculating Free Energies
• We know from statistical mechanics that we can
calculate the free energy (here the Helmholtz
free energy) by evaluating integrals like
 
N N
 N N

A = kB T ln dp dr exp −βH(p , r )

where H is the Hamiltonian.

• In practice it is very difficult to evaluate such

integrals using MC or MD since we do not
adequately sample high energy regions
Calculation of Free Energy Differences
• Although our simulation methods cannot
give us absolute free energies, free energy
differences are much more tractable
• Consider two states X and Y
• Since the free energy is a state function,
the difference in energy between these
two states is simply
   
dpN drN exp −βHY (pN , rN )
ΔA = −kB T ln 
dpN drN exp [−βHX (pN , rN )]
 Free Energy Differences
• If we multiply the numerator by the factor
exp(βHX ) exp(−βHX ) ≡ 1
we get

  N N

dp dr exp [−βHY ] exp [βHX ] exp [−βHX ]
ΔA = −kB T ln 
dpN drN exp [−βHX ]
  N N 
dp dr exp [−β(HY − HX )] exp [−βHX ]
= −kB T ln 
dpN drN exp [−βHX ]
 Free Energy Difference
• Since we are clever, we notice that this is
nothing more than an ensemble average
taken over the state X

ΔA = −kB T lnexp [−β(HY − HX )X

• Equivalently we could write the reverse

process
ΔA = −kB T lnexp [−β(HX − HY )Y
 Overlapping States
• In order to evaluate an ensemble average
like
ΔA = −kB T lnexp [−β(HY − HX )X
we could run a simulation either state X or
Y and collect statistics
• Problems arise however when the states X
and Y do not overlap such that simulating
one state does a poor job of sampling the
other
 Intermediate States
• If the energy difference between the two
states is large
|HX − HY |  kB T
we can introduce an intermediate state
between X and Y
ΔA = A(Y ) − A(X)
= (A(Y ) − A(I)) + (A(I) − A(X))


Q(Y ) Q(I)
= −kB T ln ×
Q(I) Q(X)
 Intermediate States
• We can obviously extend this treatment to
include multiple intermediate states with
increasing overlap

ΔA = A(Y ) − A(X)
= (A(Y ) − A(N )) + (A(N ) − A(N − 1)) + · · ·
+ (A(2) − A(1)) + (A(1) − A(X))


Q(Y ) Q(N ) Q(2) Q(1)
= −kB T ln × ··· ×
Q(N ) Q(N − 1) Q(1) Q(X)
 Intermediate States
• One key to this method is that
intermediate states do not need to
correspond to actual physical states
(consider changing ethane to ethanol)
H H H H

H C C H H C C O H

H H H H
• Using molecular mechanics we can
smoothly interpolate between these two
states
 Implementation
• If we have an empirical force field (like we
do in molecular mechanics) we can write
all of the force field terms as a linear
combination of the values for X and Y
– Bonds: k() = k(Y) + (1-)k(X)
lo() = lo(Y) + (1-)lo(X)
– Angles: k
() = k
(Y) + (1-)k
(X)

o() = 
o(Y) + (1-)
o(X)
– Charges: q() = q(Y) + (1-)q(X)
– VDW: () = (Y) + (1-)(X)
() = (Y) + (1-)(X)
– etc.
 Coupling Parameter
• As we change the coupling parameter

from 0 to 1, we move from state X to Y
• At each intermediate step
i we perform a
simulation (Monte Carlo or MD) by first
performing a short equilibration run (since
our point of equilibrium has changed) and
then a “production” run where we calculate

ΔA(λi → λi+1 ) = kB T lnexp(−βΔHi )


(
i+1)
Free Energy Perturbation

A

0 1

 Free energy perturbation example
• Oostenbrink C, van
Gunsteren WF. Proteins 54
(2) 234-246, 2004.
• Poly-chlorinated biphenyl
binding to estrogen receptor
• “Fast” FEP on 17
compounds
• Good agreement with
experiment
• Insight into structural and
dynamic aspects of ligand
binding
GROMOS ∆Gsolv in Water
AMOEBA Binding Free Energies
+2 +2
EF Hand: Relative Ca /Mg Binding Affinity

Wild Type: ~10 4 x (expt)

6.6 kcal/mol (calc)

Glu -> Asp: ~10 x (expt)

1.3 kcal/mol (calc)

Trypsin-Benzamidine: Absolute Binding Affinity

6.3 to 7.3 kcal/mol (expt)

6.7 +/- 0.6 kcal/mol (calc)

Pengyu Ren, U Texas

 Thermodynamic Integration
• Instead of evaluating the difference in the
free energy between subsequent states,
we could also calculate the derivative of
the Hamiltonian
 λ=1 
∂H
ΔA = dλ
λ=0 ∂λ λ

• In this case, the free energy difference is

the area under the curve
 Thermodynamic Intergration

H/


A

0 1

 Slow Growth Method
• If the changes in the system are gradually
made such that the Hamiltonian is nearly
constant, we can expand the exponential
and ln terms to get

ΔA = −kB T ln exp (−β [H(λi+1 ) − H(λi )])

 −kB T ln 1 − β [H(λi+1 ) − H(λi )]

 [H(λi+1 ) − H(λi )]
 More Reading
• Many references and papers that cover
these topics. In the texts for this class
consider:
– Leach Chapter 11 (watch for errors!!)
– Frenkel & Smit Chapter 7
Umbrella Sampling and
Histogram Methods
 The Sampling Problem
• By now you realize that the major problem in
simulations is that of sampling
• We have an exact method of computing a
partition function and associated thermodynamic
quantities, however this is dependent on us
accurately sampling the entire conformational
space
• In general (i.e. the way most people run
simulations) MD simulations do not do an
adequate jobs of sampling configurational space
unless run for a very, very long time
 Let’s Force the System to Sample
• The basic idea behind Umbrella Sampling
is that we can bias or force the system to
sample a particular region(s) (based on
some reaction coordinate)

• If we were interested in the free energy

difference between two systems X and Y,
we should sample the conformational
space associated with both conformations
 Free Energy Perturbation
• Recall from our discussion of FEP that the free energy
difference between two systems can be expressed as
  N 
dr exp [−βUY ]
ΔU = −kB T ln  N
dr exp [−βUx ]
or equivalently
 N
dr exp [−βUY ]
exp(−βΔU )X =  N
dr exp [−βUx ]
 A New Weight Function
• In order to sample both X and Y spaces, we now
introduce a new weight function
(rN) to replace the
Boltzmann factor
 N
dr π(rN ) exp [−βUY ] /π(rN )
exp(−βΔU )X =  N
dr π(rN ) exp [−βUx ] /π(rN )
which using our shorthand notation becomes

exp(−βUY )/π(rN )π

exp(−βΔU )X =
exp(−βUX )/π(rN )π
Umbrella Sampling Considerations
• In order that both the numerator and
denominator are non-zero, the weight function

(rN) should have considerable overlap between
the spaces of X and Y
• This property gives rise to the name Umbrella
Sampling
• Although it appears we could sample the entire
space with a single choice of
(rN), this is not
optimal. It is still best to perform several
sampling runs using overlapping windows
 Choosing a Weight Function
• In order for Umbrella Sampling to work well we need to
make a good choice for
(rN) – it is not know a priori
• A common choice is to make the biasing potential
quadratic
U  (rN ) = U (rN ) + W (rN )
= U (rN ) + kw (rN − roN )2
so that the biasing potential is simply
 
π(rN ) = exp −βU  (rN )
 Weighted Histogram Analysis
Method (WHAM)
• Umbrella Sampling is valid in theory, but the
implementation is often difficult since the
“windows” of overlap must be carefully chosen to
minimize the error (since the errors from the
individual simulations add quadratically)

• WHAM is a useful method for combining sets of

simulations with different biasing potentials in a
manner such that the unbiased potential of
mean force (PMF) can be found
Periodic Box Simulation
(alanine dipeptide and
206 water molecules)

Stochastic Dynamics
(576 trajectories of
200 picoseconds each)

Free Energies via

Umbrella Sampling
and 2D-WHAM:
Nw

∑n ρ i wi (φ ,ψ )
ρ (φ ,ψ ) = Nw
i =1

∑n e
i=1
i
−wi ( φ ,ψ ) − Fi ) / kbT

 
Fi = − kb T ln ∑ ∑e − wi ( φ ,ψ ) / kbT
ρ (φ ,ψ )
φ ψ 

∆G (φ ,ψ ) = − kb T ln ρ (φ ,ψ )
 AMBER ff99 CHARMM27
180 180

120 120

3.2
60 2.8 60
2.4
0 2 0
1.6
1.2
-60 -60
0.8
0.4
-120 -120

-180 -180
-180 -120 -60 0 60 120 180 -180 -120 -60 0 60 120 180

OPLS-AA
180

120

Solvated 60 Free
Alanine Energy
Psi

Dipeptide -60
Surfaces
-120

-180
-180 -120 -60 0 60 120 180
Phi
 Conformational Populations
Alpha Pass Beta Other
Amber ff94 68 5 26 1
Amber ff99 77 10 13 1
CHARMM27 46 2 52 0
OPLS-AA 13 9 75 3
OPLS-AA/L 23 8 65 4
SCCDFTB (Amber) 27 16 48 9
SCCDFTB (CHARMM) 33 14 48 4
SCCDFTB (CEDAR) 27 12 61 0
AMOEBA (Polar Water) 29 16 54 1
AMOEBA (Fixed Water) 32 13 54 1
Examples and Further Reading
• Leach has some details on Umbrella
Sampling in Ch. 11

• Frenkel & Smit discusses Umbrella

Sampling and WHAM (disguised as the
self-consistent histogram method) in Ch. 7

• There are many papers using these

methods. (See Ron Levy paper that uses
both techniques)
Jarzynski’s Method