Convergence of replica exchange molecular dynamics

THE JOURNAL OF CHEMICAL PHYSICS 123, 154105 2005 Convergence of replica exchange molecular dynamics Wei Zhang and Chun Wu Department of Chemistry and Biochemistry, University of Delaware, Newark, Delaware 19716 and University of California Davis Genome Center and Department of Applied Science, University of California, Davis, California 95616 Yong Duan a University of California Davis Genome Center and Department of Applied Science, University of California, Davis, California 95616 Received 18 April 2005; accepted 16 August 2005; published online 18 October 2005 Replica exchange molecular dynamics REMD method is one of the generalized-ensemble algorithms which performs random walk in energy space and helps a system to escape from local energy traps. In this work, we studied the accuracy and efficiency of REMD by examining its ability to reproduce the results of multiple extended conventional molecular dynamics MD simulations and to enhance conformational sampling. Two sets of REMD simulations with different initial configurations, one from the fully extended and the other from fully helical conformations, were conducted on a fast-folding 21-amino-acid peptide with a continuum solvent model. Remarkably, the two REMD simulation sets started to converge even within 1.0 ns, despite their dramatically different starting conformations. In contrast, the conventional MD within the same time and with identical starting conformations did not show obvious signs of convergence. Excellent convergence between the REMD sets for T 300 K was observed after 14.0 ns REMD simulations as measured by the average helicity and free-energy profiles. We also conducted a set of 45 MD simulations at nine different temperatures with each trajectory simulated to 100.0 and 200.0 ns. An excellent agreement between the REMD and the extended MD simulation results was observed for T 300 K, showing that REMD can accurately reproduce long-time MD results with high efficiency. The autocorrelation times of the calculated helicity demonstrate that REMD can significantly enhance the sampling efficiency by 14.3± 6.4, 35.1± 0.2, and 71.5± 20.4 times at, respectively, 360, 300, and 275 K in comparison to the regular MD. Convergence was less satisfactory at low temperatures T 300 K and a slow oscillatory behavior suggests that longer simulation time was needed to reach equilibrium. Other technical issues, including choice of exchange frequency, were also examined. 2005 American Institute of Physics. DOI: 10.1063/1.2056540 I. INTRODUCTION ventional simulation methods. 3 The detailed description can be found in several methodology 1,4 and review papers. 3,5,6 Some of the powerful applications include studies on peptides and miniproteins with REMD. 7 17 REMD was also applied to compare explicit with implicit solvent models 18,19 and among different force fields. 20 REMD was designed to enhance the sampling efficiency and a single REMD simulation can give thermodynamic quantities at a range of temperatures which is also very attractive. Comparison of REMD with MD has unequivocally shown that REMD can sample more conformational space than regular MD simulations within the same time scales. 10 Sanbonmatsu and Garcia compared a 32 ns constanttemperature MD with the results of a 16-replica simulation each 2 ns in an explicitly solvated system. 10 They found that REMD sampled approximately five times more conformational space than the MD simulations of comparable aggregated simulation time. This was encouraging. However, questions still remain regarding the convergence of the method. In particular, one needs to assess the ability of REMD to reproduce the original free-energy landscapes dictated by the underlying force field and simulation parameters. To this end, one needs to know whether the REMDa Author to whom correspondence should be addressed. Fax: 530 754-9648. Electronic mail: duan@ucdavis.edu Conformational sampling has been a long-standing issue in computational sciences. For complex systems such as peptides and proteins, the enormity of the number of energy minima coupled with the time needed to escape from each of these sometimes not-so-minor conformational traps has rendered one of the greatest challenges in computational science. This acute problem is particularly debilitating for realistic simulations such as molecular dynamics MD. Thus, advanced techniques to enable efficient sampling of conformational space can be tremendously helpful. The replica exchange molecular dynamics REMD method was initially introduced by Sugita and Okamoto. 1 It combines the idea of multiple-copy simulation, simulated annealing, and Monte Carlo methods and is one of the generalized-ensemble algorithms 2,3 that performs random walk in energy space due to non-boltzmann weighting factors. The random walk allows a simulation to pass energy barriers and sample more conformational space than by con- 0021-9606/2005/123 15 /154105/9/$22.50 123, 154105-1 2005 American Institute of Physics

154105-2 Zhang, Wu, and Duan J. Chem. Phys. 123, 154105 2005 produced thermodynamic quantities can converge to the equilibrium distribution. Obviously, there are two ways to test the convergence of the REMD method: one is to compare the REMD-derived properties with experimentally determined results; the other is to compare with equilibrated MD simulation results. We chose the latter comparison, not only because of experimental complexity but also, more importantly, comparison between simulations with the same parameters and force fields eliminates many unknown bias which otherwise might invalidate the comparisons. In the present paper, we systematically tested the convergence of the REMD method using a 21-residue helical peptide Fs-21 Ace-AAAAA AAARA 3 A-Nme as the model system, 21,22 taking advantage of the rapid folding kinetics, marginal stability, and continuum solvent model. Two sets of REMD simulations were conducted with two extreme initial structures: the extended conformation and the fully helical conformation. Because the native state of the peptide is helical, we will henceforth refer to the simulations from the extended conformation as the folding set and refer to the other that started from the fully helical conformation as the unfolding set. In addition, regular MD simulations were performed at nine different temperatures. In these MD simulations, five trajectories were produced at each temperature for extensive conformational sampling and the duration of each trajectory was 100.0 to 200.0 ns. Comparison between the two REMD sets and between REMD and MD results showed a high sampling efficiency of the REMD method, and the convergence problem was discussed according to different comparison results. II. METHODS A. REMD The detailed description can be found in the work of Sugita and Okamoto. 1 Here, we briefly summarize the method. In the REMD simulations, independent simulations were conducted on a set of noninteracting replicas of the same molecular system over a range of temperatures. The system at a different temperature can exchange configurations at the fixed time intervals with a transition probability satisfying the detailed balance. The exchange takes place using the Metropolis criterion with an acceptance probability, P = exp i j E i E j, 1 where E i is the total potential energy of the replica i at temperature T i ; i =1/ k B T i and k B is the Boltzmann s constant. Obviously, the acceptance ratio depends on the temperature differences between the neighboring replicas. Since the potential energy typically varies linearly with the temperature for condensed-phase systems, the replica temperatures were chosen roughly exponentially distributed. We verified our choices by conducting several short trail MD simulations at different temperatures and calculated the average potential energies. In addition, because of finite system size, system temperatures can fluctuate around the target temperatures of the heat bath. We tested this by two sets of REMD simulations differing by the way the exchange temperatures were obtained. In one set, the target temperatures were used to calculate the P and in the other the real temperatures were used. The results showed that these two REMD sets had very similar exchange modes in terms of frequency of exchange and traverse across the different temperatures. We therefore chose the target temperatures for the calculation of P and for exchange. Finally, we set a target acceptance ratio at 40% by choosing the temperatures. The actual acceptance ratio during REMD was 37%. Both folding and unfolding REMD sets were simulated with 18 replicas and with temperatures ranging from 250 to 610 K at 13 30 K intervals. In each REMD set, the replicas started from the same structure but with different random number seeds for the initial velocity assignment and at different temperatures. The simulations started from a 1.0 ns regular MD run at the target temperatures. Replica exchanges were attempted between neighboring pairs every 500 integration steps 1.0 ps. The REMD was performed for 14.0 ns for each replica. We also conducted REMD simulations with 10-step 0.020 ps and 2500-step 5 ps exchange frequencies to examine the issue whether the exchange frequency would have any influence on the results. Nevertheless, our main results were based on simulations with the 500-step exchange frequency. B. Simulation protocol The AMBER 7 simulation package with REMD was used in all simulations. The Fs-21 A 5 AAARA 3 peptide was modeled using the force field of Duan et al. ff03. 23 All bonds connecting hydrogen atoms were constrained by SHAKE algorithm to allow an integration time step of 2.0 fs. Generalized Born GB implicit solvent model was applied to model the solvation effect, with an effective 0.2M salt concentration modeled based on Debye-Huckel theory. 24,25 Surface area terms were not explicitly modeled in the GB model for their contribution is typically small. The interior and solvent dielectric constants were set to, respectively, 1.0 and 78.5. Simulations were done at constant-temperature ensemble and the temperature was maintained by Berendsen thermostat. 26 For comparison, regular constant-temperature MD simulations were conducted at seven different temperatures 265, 360, 380, 400, 430, 460, and 500 K. Five trajectories were produced at each temperature with the identical GB model and parameters to those in the REMD simulations. Each simulation was run for 100.0 ns for an aggregate total of 3.5 s, starting from the fully extended conformation. The last 50 ns are used for the analyses. We also included two previously reported trajectory sets at 273 and 300 K in this analysis, 27 making a total of nine different temperatures. These two sets were five 200.0-ns trajectories at 273 K and 25 100.0-ns trajectories at 300 K, as described in detail before. The helical content was calculated based on the mainchain hydrogen bonds, defined as the O:H distance 2.8 Å and the angle CvO:H 120. The simulated snapshots were clustered using a heuristic method in which a snapshot was included into its closet cluster if the main-chain rootmean-square distance RMSD is smaller than 2.5 Å after

154105-3 Replica exchange molecular dynamics J. Chem. Phys. 123, 154105 2005 FIG. 1. Folding and unfolding curves of REMD and MD. A: From 14 ns REMD. B: From the 0 1 ns REMD. C: From 1.0 ns MD. D: Immediately after the initial 1.0 ns relaxation. rigid body alignment. The representative structures were further compared and the clusters were merged together if the representative structures were closer than 2.5 Å. III. RESULTS For condensed-phase systems, temperature equilibration means stabilization of both kinetic and potential energies and the latter often requires longer time than the former because of the time required for heat dissipation. In general, exchange should take place after the system temperature stabilizes as judged by the changes in the potential energy. As one may expect, system size plays a role in this regard; large systems require a long relaxation time following each temperature change. Short exchange time intervals, although desirable for data collection, may prevent the system from complete heat dissipation due to the temperature and kinetic energy changes. On the other hand, because exchange takes place only when two replicas have similar potential energies and their target temperatures are close to one another, frequent exchange is possible. Therefore, various exchange frequencies, ranging from 0.02 to 5 ps, have been reported in literatures. We compared the results between attempts of every ten steps 0.02 ps and every 500 steps 1.0 ps and observed similar average exchange rates. In both cases, each replica sampled a wide temperature range and most of the replicas traveled the whole temperature range for many times. However, for the exchange at every ten steps, two of the 18 replicas went through the entire temperature range only by fewer than two times over 14 ns. This is considerably fewer than other replicas that went through the entire temperature range for more than six times. We attributed this to the excessively short intervals between the exchanges which might have prevented the system to stabilize to the target temperature and, thus, made it easier for the system to exchange back. We also tested at the exchange frequency of 5.0 ps and found that the results were qualitatively similar to those with the 1.0 ps exchange frequency discussed later, suggesting that 1.0 ps is an adequate choice for this system. Thus, we chose to exchange the replicas every 500 steps 1.0 ps, which is less frequent than some of the exchange frequencies reported by others. A. Helical contents, folding, and unfolding curves A clear advantage of REMD is its ability to simulate the system simultaneously at multiple temperatures and it is an ideal approach to obtain the thermodynamic properties at multiple temperatures. One can readily obtain the folding/ unfolding behavior from one set of simulations. We thus calculated the helical contents at the simulated temperatures which mimics the peptide s melting curve. Figure 1 A shows the average helical contents of the three REMD sets over the last 12.0 ns. For replicas over 300 K, three smooth melting curves, including the folding and unfolding sets, converged very well to each other. They also agreed with the multiple equilibrated MD results. This showed the ability of REMD to simultaneously reproduce peptide s properties at multiple temperatures. The agreement with the multiple equilibrated MD simulations is particularly remarkable, indicating that indeed REMD can faithfully reproduce the results of conventional MD with significantly enhanced efficiency. Figure 1 B shows the average helical content within the first 1.0 ns of REMD immediately following the initial 1.0 ns MD relaxation leg. Remarkably, the folding and unfolding curves, including the low-temperature parts, started to show convergence even within 1.0 ns REMD simulations. Both curves started to show convergence to the melting curves obtained from multiple long-time MD simulations. The relatively smooth curves are also striking. Evidentially,

154105-4 Zhang, Wu, and Duan J. Chem. Phys. 123, 154105 2005 FIG. 2. Helical contents of representative REMD folding and unfolding trajectories. A: Folding at 361 K. B: Unfolding at 361 K. C: Folding at 308 K. D: Unfolding at 308 K. E: Folding at 277 K. F: Unfolding at 277 K. the peptide conformation started to converge even within the 1.0 ns REMD simulation. For comparison, we show the folding and unfolding curves from 1.0 ns MD runs Fig. 1 C and those immediately after the initial 1.0 ns MD relaxation phase Fig. 1 D. At the end of the 1.0 relaxation leg Fig. 1 D, interestingly, the unfolding set, which started from the fully helical conformation, started to resemble the melting curve. This was due to the fast unfolding at high temperatures. When it started from the straight chain conformation, the peptide folded partially at medium temperatures, and remained completely unfolded at low temperatures at the end of the relaxation leg because of the slow folding rates at these temperatures. In contrast to the 1.0 ns REMD curves Fig. 1 B, the 1.0 ns MD curves Fig. 1 C were similar to their respective starting curves and remained different from one another. Thus, not surprisingly, regular MD failed to show signs of convergence within 1.0 ns, further corroborating the observations by others that REMD is indeed an efficient and powerful method for conformational sampling. Notwithstanding these rather encouraging results, differences still persisted at low-temperature replicas T 300 K, Fig. 1 A, despite the relatively long simulations on a relatively simple systems in comparison to other studies. Figure 2 shows the helical contents of several representative replica trajectories. At 361 K Figs. 2 A and 2 B, the trajectories one is unfolding and the other is folding varied very frequently and, as a consequence, their helical contents converged well. This clearly illustrates the sampling power of the method which enables the peptide to sample many conformations within a short simulation time. In comparison, however, such transitions were less frequent at 308 K Figs. 2 C and 2 D. When the temperature further decreased to 277 K Figs. 2 E and 2 F, both sets showed considerably fewer transitions to the fully helical state. At lower temperatures, the variations were similar between unfolding and folding simulations. We calculated the autocorrelation functions of the helicity for the replicas and compared them with those calculated from the conventional MD simulations. Shown in Fig. 3 are some of the representative autocorrelation functions. A striking feature is the rapidly vanishing autocorrelation in the REMD simulations. We fitted these autocorrelation functions by single exponential functions to quantify the correlation times that are summarized in Table I. At 277 K, the correlation times were 0.127 and 0.219 ns for, respectively, REMD folding and unfolding trajectories. They were 0.105 ns at 308 K and decreased further to 0.041 and 0.016 ns at 361 K. These were considerably shorter than the correlation times observed in the MD simulations which were 12.34 ns at 273 K, 3.68 ns at 300 K, and 0.40 ns at 360 K. If one takes TABLE I. Correlation time measured from the helicity autocorrelation functions. REMD MD Folding Unfolding Average Folding Ratio T=277 K 0.127 0.219 0.173 12.34 273 K 71.5 T=308 K 0.105 0.105 0.105 3.68 300 K 35.1 T=361 K 0.041 0.016 0.028 0.40 360 K 14.3

154105-5 Replica exchange molecular dynamics J. Chem. Phys. 123, 154105 2005 FIG. 3. Representative autocorrelation functions obtained from REMD and conventional MD at the respective temperatures. For MD simulations at 273, 300, and 360 K the autocorrelation functions are averaged over multiple trajectories. these as measures of sampling efficiency, REMD is about 71.5± 20.4, 35.1± 0.2, and 14.3± 6.4 times more efficient than conventional MD at 275, 300, and 360 K, respectively. Interestingly, all three results suggested an effective reduction of the energy barrier by about 2 kcal/mol at their respective temperatures 1.9 kcal/mol at 360 K, 2.1 kcal/mol at 300 K, and 2.3 kcal/mol at 275 K. More discussions are presented later. We compared the melting curves at different time periods: every 2 ns, every 3 ns, and every 4 ns Fig. 4. Convergence was generally decent at a higher temperature range. However, differences between folding and unfolding curves were also obvious at a lower temperature range even at long time periods. The reason for the slow convergence might simply be because the slower kinetics at lower temperatures which sets the physical limits of the rate of conformational change even with the accelerated conformational sampling methods such as REMD. Also shown in Fig. 4 are the average helicity calculated from a REMD set with a 5 ps exchange rate. The curves from the 5 ps set were qualitatively consistent with those from the folding and unfolding sets which were done with 1 ps exchange intervals. FIG. 4. Melting curves from different time periods every 2, 3, and 4 ns, respectively. Both folding and unfolding sets with 1.0 ps exchange frequency and an unfolding set with 5 ps exchange frequency are shown.

154105-6 Zhang, Wu, and Duan J. Chem. Phys. 123, 154105 2005 FIG. 5. Contour plots of potential of mean force as a function of total number of hydrogen bonds and radius of gyration. A: Last 10 ns of folding REMD set with 1 ps exchange frequency. B: Last 10 ns of unfolding REMD set with 1 ps exchange frequency. C: last 10 ns of unfolding REMD set with 5 ps exchange frequency. D: last 50 ns of the equilibrated MD sets. B. Potential of mean force PMF Two-dimensional PMF profiles were generated along the main-chain hydrogen bonds and the radius of gyration R from the data of the last 10.0 ns snapshots, ignoring the first 4.0 ns. Figures 5 A and 5 B show the two-dimensional PMF profiles of the folding and unfolding sets. The representative structures, produced from clustering analysis data not shown, are also shown in Fig. 6. Overall, these two sets, including some of the fine details, are similar. Both PMF profiles provide a clear view of the conformational variation with temperatures. There are two free-energy minimum regions at the low temperatures. One of which is in the extended helix region with 14 main-chain hydrogen bonds and an R of about 10 Å and the other is in the compact structure regions with 8 main-chain hydrogen bonds. There is a free-energy barrier separating these two regions that decreases when temperature increases. The two regions merged together at 360 K. Further temperature increase beyond 360 K reduced the number of main chain hydrogen bonds even further which eventually approaches zero at very high temperature. Shown in Fig. 5 C are the PMFs of the 5 ps exchange REMD set. The high degree of similarity between this set and those with 1 ps exchange rate reinforced the notion that 1 ps exchange rate was sufficient for the system. Figure 5 D shows the PMF profiles from the last 50 ns of extended MD simulations. The qualitative resemblance between this and the REMD PMF profiles is readily evident. The same two low-energy regions and the same energy barrier variations were observed. The remarkable agreements with REMD results further showed the effectiveness and accuracy of the REMD method. C. Principle components analysis Sanbonmatsu and Garcia compared short-time replica simulation 2 ns with constant-temperature MD Ref. 10 and found that the aggregated sampling ability of REMD including those high-temperature replicas was superior than a 32 ns MD simulation with a comparable aggregated simulation time 32 ns. However, it is also important to assess its ability to sample the conformational space at a physiologically relevant temperature and to assess the conformational space sampled in REMD against that sampled in a fully equilibrated MD simulation set. Here, we further tested REMD sampling by comparing our folding and unfolding REMD simulations to a set of 25 100 ns MD trajectories at 300 K. The principal components were calculated from a data set that includes all trajectories subject to this comparison. The folding and unfolding REMD sets each 2.52 10 4

154105-7 Replica exchange molecular dynamics J. Chem. Phys. 123, 154105 2005 FIG. 6. Representative structures. snapshots: 14 ns each of the 18 replicas and the equilibrated MD total 2.5 10 5 snapshots trajectories were first aligned and the principle components were calculated according to the average configuration of all snapshots. The average configuration was at the origin of the two-dimensional 2D map. The position of each snapshot on the 2D map was represented by its projections on the two principal component vectors. Figures 7 A and 7 B illustrate the first nanoseconds of the folding and unfolding REMD sets, and comparably, the first 18 ns sampling of one representative MD trajectory Fig. 7 C. Consistent with the results of Sanbonmatsu and Garcia, 10 REMD sampled a much wider space, while the MD trajectory was limited at certain low-energy area. Figures 7 D and 7 E present the results of all 14 ns REMD each set 2.52 10 4 snapshots of 18 replicas. The total 25 trajectories of 100 ns MD simulations are shown in Fig. 7 F, constructed from 2.5 10 4 snapshots out of the 2.5 10 5 by picking one from every ten snapshots to keep a comparable number of samples. The distributions of the scatters of both REMD sets were similar to those from the equilibrated MD simulations. In Figs. 7 G 7 I, we compare the sampling of equilibrated MD with that of a replica at a similar temperature 292 K. The equilibrated MD PCA map was collected from the last 10 ns snapshots of all 25 trajectories total 2.5 10 4 snapshots. The folding and unfolding REMD plots showed sampling of 14 ns of the replicas at 292 K 1.4 10 3 snapshots each. Despite the differences in the number of snapshots which resulted in somewhat sparse distributions of REMD scatters, the high degree similarity between these FIG. 7. Principle component analysis comparison between REMD and constant-temperature 300 K MD.

154105-8 Zhang, Wu, and Duan J. Chem. Phys. 123, 154105 2005 three sets is readily apparent. We thus conclude that REMD sampled the same conformational space and showed similar sampling distribution as the equilibrated MD. Once again, this reinforces the notion that REMD results faithfully mimicked the properties of canonical ensembles. IV. DISCUSSION Like all generalized-ensemble simulations, REMD performs random walk in energy space therefore helping the system to escape from local energy minima and search wider conformational space. In the present work, we examine the convergence issue from several perspectives: the convergence of REMD with different simulation times, that of REMD from different initial structures, and between REMD and extended MD simulations. Encouraging results have been obtained from comparisons between REMD and MD results at different temperatures. Comparisons of the helical content, potential mean force, and principle component analysis all show consistent results, suggesting REMD is able to sample the conformational space more efficiently and is able to do so without changing the free-energy landscapes of the system. The folding and unfolding sets also showed consistent results. Even though these two sets started from completely different initial structures, the results were remarkably similar because of the very efficient sampling ability of REMD. These are very encouraging. Fs-21 is a fast-folding helical peptide. Williams et al. 28 showed, by laser-induced temperature-jump experiments and time-resolved infrared spectroscopy, that it can fold rapidly. In simulations, the folding time can be significantly accelerated by application of implicit solvent models which ignores solvent viscosity. 27 Another important feature of Fs-21 is its marginal stability. The experimentally observed helicity is 50% at 334 K, 28 consistent with the 50% experimental helicity of a Fs-21 variant. 29 Thus, its experimental folding free energy was close to 0.0 kcal/mole at around 300 K, consistent with our previous simulation which showed a folding free energy of 0.1 kcal/mole at 300 K. 27 The marginal stability and rapid folding kinetics allow reasonable sampling in conventional MD and simplify the comparison. Therefore, the 100 ns multiple trajectory MD simulations with continuum solvent model at different temperatures were sufficient to represent the equilibrated systems. One notable discrepancy is that the helical contents Figs. 2 and 3 clearly showed differences at the low temperatures 300 K, suggesting their lack of convergence. In addition, judging from the autocorrelation functions, a small fraction 20% of the helicity was oscillatory at low temperature with a periodicity of about 1.0 2.0 ns which is similar to the time needed for the replicas to traverse across the entire temperature range. However, because the oscillating components were essentially nonexistent at higher temperatures, the oscillating component can not be attributed solely to the temperature traverse. Instead, the explanation lies at the much reduced kinetics at the low temperatures. At these temperatures, for example, the folding rates are significantly reduced, as shown in our earlier study. The reduced kinetics slowed down the conformational transitions that were needed to reach equilibrium after the exchange took place. Although this might be a source of concern, the good convergence at temperatures above 300 K suggests that one can obtain reliable information from REMD for T 300 K. For T 300 K, still longer simulations might be needed which is particularly true for larger systems. Rhee and Pande 14 also reported the convergence problem with their multiplexed-remd MREMD method, which was a modified REMD to enable REMD on large number of heterogeneous computer clusters. The system they examined was a 23-amino-acid protein BBA5. They employed 20 with 200 multiplexed replicas at each temperature and 50 ns for each trajectory. Although their MREMD reached the folded state, the simulation was not long enough to reach a satisfactory thermodynamic convergence. One reason was that BBA5 was relatively slow system and a considerably longer simulation time is needed to reach thermodynamic convergence. Satisfactory convergence was only reached at high temperature 500 K and long simulation time 50 ns. The results were consistent with our observations. V. CONCLUSIONS Convergence of REMD simulation method is examined using a fast folding 21-amino-acid peptide Fs-21 as the model system. Two sets of REMD simulations with different initial configurations and nine constant-temperature MD simulations were conducted. Comparison between the two REMD sets and between REMD and regular MD suggested that REMD was able to sample the conformational space more efficiently and was able to do so without changing the free-energy landscapes of the system. The autocorrelation times of the calculated helicity demonstrate that REMD can significantly enhance the sampling efficiency by 14.3± 6.4, 35.1±0.2, and 71.5±20.4 times in comparison to the regular MD at 360, 300, and 275 K, respectively. Excellent agreement between the REMD and the extended MD simulation results was observed for T 300 K, showing that REMD can accurately reproduce equilibrated MD results with high efficiency. However, the observed slow oscillation of helical content at low temperatures 300 K indicated that REMD simulations did not converge within our simulation time 14 ns. ACKNOWLEDGMENTS This work has been supported by research grants from NIH GM64458 and GM67168 to one of the authors Y.D.. Usage of PYMOL and VMD graphics packages is gratefully acknowledged. 1 Y. Sugita and Y. Okamoto, Chem. Phys. Lett. 314, 141 1999. 2 U. H. E. Hansmann and Y. Okamoto, Curr. Opin. Struct. Biol. 9, 177 1999. 3 A. Mitsutake, Y. Sugita, and Y. Okamoto, Biopolymers 60, 96 2001. 4 Y. Okamoto, J. Mol. Graphics Modell. 22, 425 2004. 5 S. Gnanakaran, H. Nymeyer, J. Portman, K. Y. Sanbonmatsu, and A. E. Garcia, Curr. Opin. Struct. Biol. 13, 168 2003. 6 K. Tai, Biophys. Chem. 107, 213 2004.

154105-9 Replica exchange molecular dynamics J. Chem. Phys. 123, 154105 2005 7 A. E. Garcia and K. Y. Sanbonmatsu, Proteins 42, 345 2001. 8 R. Zhou, B. J. Berne, and R. Germain, Proc. Natl. Acad. Sci. U.S.A. 98, 14931 2001. 9 A. E. Garcia and K. Y. Sanbonmatsu, Proc. Natl. Acad. Sci. U.S.A. 99, 2782 2002. 10 K. Y. Sanbonmatsu and A. E. Garcia, Proteins 46, 225 2002. 11 A. E. Garcia and J. N. Onuchic, Proc. Natl. Acad. Sci. U.S.A. 100, 13898 2003. 12 S. Gnanakaran and A. E. Garcia, Biophys. J. 84, 1548 2003. 13 J. W. Pitera and W. Swope, Proc. Natl. Acad. Sci. U.S.A. 100, 7587 2003. 14 Y. M. Rhee and V. S. Pande, Biophys. J. 84, 775 2003. 15 R. Zhou, Proc. Natl. Acad. Sci. U.S.A. 100, 13280 2003. 16 A. K. Felts, Y. Harano, E. Gallicchio, and R. M. Levy, Proteins 56, 310 2004. 17 W. C. Swope, J. W. Pitera, F. Suits et al., J. Phys. Chem. B 108, 6582 2004. 18 H. Nymeyer and A. E. Garcia, Proc. Natl. Acad. Sci. U.S.A. 100, 13934 2003. 19 R. Zhou, Proteins 53, 148 2003. 20 T. Yoda, Y. Sugita, and Y. Okamoto, Chem. Phys. Lett. 386, 460 2004. 21 D. J. Lockhart and P. S. Kim, Science 260, 198 1993. 22 D. J. Lockhart and P. S. Kim, Science 257, 947 1992. 23 Y. Duan, S. Chowdhury, G. Xiong et al., J. Comput. Chem. 24, 1999 2003. 24 D. Bashford and D. A. Case, Annu. Rev. Phys. Chem. 51, 129 2000. 25 V. Tsui and D. A. Case, Biopolymers 56, 275 2000. 26 H. J. C. Berendsen, J. P. M. Postma, W. F. van Gunsteren, A. DiNola, and J. R. Haak, J. Comput. Phys. 81, 3684 1984. 27 W. Zhang, H. X. Lei, S. Chowdhury, and Y. Duan, J. Phys. Chem. B 108, 7479 2004. 28 S. Williams, T. P. Causgrove, R. Gilmanshin, K. S. Fang, R. H. Callender, W. H. Woodruff, and R. B. Dyer, Biochemistry 35, 691 1996. 29 P. A. Thompson, W. A. Eaton, and J. Hofrichter, Biochemistry 36, 9200 1997.