In replica-exchange sampling, replicas of the protein system evolve in parallel at dif- ferent temperatures. At certain times, replicas are exchanged between neighboring temperatures Ti and Tj = i + 1 with a probability,
w( Cold → Cnew) = min(1, exp(−βiE(Cj) − βjE(Ci) + βiE(Ci) + βjE(Cj)))
= min(1, exp(∆β∆E))
(5.1) with β = 1/kBT . The resulting random walk through temperature yields an en-
hanced exploration of configurations at low temperatures. However, despite many successful applications, replica-exchange sampling is often restricted by severe lim- itations. Take as an example simulations of proteins in explicit solvent where the number of required replicas increases rapidly with protein size. As the time to sam- ple independent configurations increases quadratically with the number of replicas, it follows that both many replicas and long trajectories are required to generate sufficient statistics at temperatures of interest. While exchange schemes have been developed that target specifically this problem,167, 168 all-atom folding simulations in
explicit solvent are still restricted to rather short proteins (of ≈ 50 − 80 residues). Note also that the rapid growth of replica number is not restricted to explicit solvent simulation but will appear for all sufficiently large systems.
Here, and in many other practical applications, the use of replica-exchange sam- pling is held back because the exchange move leads to a proposed state Cnew of the
multiple replica system that is exponentially suppressed. However, once such an exchange move is accepted, the two replicas will quickly evolve and the compound system will assume a new state that has a weight comparable to that before the ex- change move. Hence, the problem is to “tunnel” through the unfavorable “transition state” generated by the exchange move.
We propose to tackle this problem by combining replica-exchange molecular dy- namics with ideas from hybrid MC/MD.161 In the latter technique, one starts with a configuration qi and velocities vi corresponding to the selected temperature. A
short molecular dynamics run leads to a configuration qio, vio that is accepted or
rejected by a Metropolis step. As the time reversibility of the Verlet algorithm guarantees detailed balance, the Monte Carlo step ensures that the sampled config- urations are distributed according to the chosen temperature. Utilizing in a similar way the time reversibility of the Verlet algorithm our RET replaces a configuration A by ˆB at temperature T1, and B by ˆA at temperature T2,
T1 : A −→ A0 &% B00−→ ˆB,
T2 : B −→ B0 %& A00−→ ˆA,
where A = (qA, vA) denotes a state characterized by coordinates qA of all its atoms
Hence, the RET move consists of four parts:
1. In the first part, the configuration A (B) evolves by a short microcanonical molecular dynamics run to a configuration A0 = (qA0 , v0A) (B0) with total energy Etot = EP ot+ Ekin= E1 (E2).
2. In the second step, the two replicas are provisionally exchanged, and at the same time their velocities rescaled such that the total energies at the two temperatures stay the same:
Etot(B00) = E1 and Etot(A00) = E2. (5.2)
Here, A00 = (q0A, vA00) and B00 = (qB0 , v00B). This is achieved by rescaling the velocities according to169, 170 v00A= vA0 s E2− Epot(qA0 ) Ekin(v0A) and vB00 = vB0 s E1− Epot(qB0 ) Ekin(v0B) . (5.3)
3. The above exchange move generates a “transition state” in the multiple replica system where the unfavorable potential energies at the two temperatures are compensated by the rescaled velocities. In the third step, each of the config- urations A00(B00) evolve again by a short microcanonical molecular dynamics run to a configuration ˆA = (ˆqA, ˆvA)( ˆB) where the velocity distribution cor-
responds now again to the target temperatures, and the potential energies are comparable to the ones found at the respective temperatures before the exchange move.
set A, B and is accepted with probability
w( Cold→ Cnew) = min(1, exp(−β
1(Epot(ˆqB) − Epot(qA))
−β2(Epot(ˆqA) − Epot(qB)))).
(5.4)
If rejected, the molecular dynamics simulations will continue at T1(T2) with
configuration A(B). In both cases, new velocities are drawn from a distribution corresponding to the respective temperatures. Again, the time reversibility of the trajectories A −→ A0 (B −→ B0) and A00 −→ ˆA (B00 −→ ˆB) ensures detailed balance and convergence to the correct distribution.
In order to see that RET leads to an increased exchange rate, we write the acceptance probability of Eq. (5.4) as
w( Cold → Cnew) = min(1, exp (∆β∆E) × exp(−β
1(Epot(ˆqA) − Epot(qA))
−β2(Epot(ˆqB) − Epot(qB)))),
(5.5) where ∆E = Epot(ˆqB) − Epot(ˆqA). The first factor is the acceptance rate for regular
replica-exchange sampling. Hence, the acceptance probability is enhanced by wRET
wREM D = exp (−β1(Epot(ˆqA) − Epot(qA)) − β2(Epot(ˆqB) − Epot(qB))) (5.6)
where Epot(ˆqA) − Epot(qA) and Epot(ˆqB) − Epot(qB) have opposite sign. Assuming
both terms to be similar in magnitude, the enhancement factor can be approximated as
wRET
wREM D ≈ exp (−∆βδE) (5.7)
∆β = β2 − β1 < 0. As this enhancement factor is always larger or equal one, it
follows that the acceptance rate will be always better than in traditional replica- exchange moves, but will depend on the length of microcanonical step which controls how much the potential energies Epot(ˆqA) (Epot(ˆqB)) differ from Epot(qA) (Epot(qB)).
Our test case is the designed 20-residue Trp-cage miniprotein17, 162 (Protein Data
Bank Identifier 1L2Y). As one of the smallest proteins with a defined tertiary struc- ture it is often used to evaluate new sampling schemes.171, 172 Using the same force
field, implicit solvent, and temperature distribution, we compare the results from our RET simulations with previous simulations173 that rely on regular REMD.85–87, 163 Out of the many Trp-cage replica-exchange studies,164–166 these are chosen by us
because their setup leads to a melting temperature of 400 K, closer to the exper- imental values of 315 K162 than found in other implicit solvent simulations. We use the molecular dynamics program package GROMACS,174 version 4.6.5, either
in its original version or modified to implement our RET approach. The modified version is available as supplementary material175 and from the authors. Interactions between the atoms in the protein are described by the Amber force field 94,176 and
the interaction of the protein with the surrounding solvent is approximated by a generalized Born surface area implicit solvent.177 The N- and C-termini are capped with methyl groups. We use the LINCS algorithm78 to constrain hydrogen atoms
to their bonded heavy atoms. van der Waals and Coulomb energies are calculated using twin range cutoffs. The equations of motion are integrated with a time step of 1 fs for RET and 2 fs for regular REMD. We use either 22, or 12 replicas distributed over a range of temperatures from 250 to 605 K. For 22 replicas, the selected tem- peratures are the same as in Ref.,173 i.e., 250, 255, 260, 265, 273, 284, 298, 315, 333, 353, 373, 393, 413, 433, 454, 473, 493, 513, 533, 555, 580, and 605 K. The thermo-
stat temperature is controlled by the v-rescale method.132 Exchanges are attempted
every 200 ps, which for RET includes two segments of 1 ps where the system evolves at constant energy. All simulations start from unfolded configurations and physical quantities are calculated after discarding the first 50 ns.