Convergence and Heterogeneity in Peptide Folding with Replica Exchange Molecular Dynamics

Replica exchange molecular dynamics (REMD) techniques have emerged as standard tools for simulating peptides and small proteins, in part, to evaluate the accuracy of modern classical force fields for polypeptides. However, it often remains a challenge to unambiguously discriminate force field flaws from simulations that do not reach equilibrium convergence. Here, we examine closely the convergence behavior of REMD runs for 14 test alpha and beta peptide systems, using an AMBER force field with a generalized Born/surface area implicit solvation model. Somewhat surprisingly, we find that convergence times can be quite large compared to the time scales reached in many earlier REMD efforts, with some short peptides requiring up to 60 ns of run time (per replica). Moreover, we detect a high degree of run-to-run heterogeneity, finding that REMD runs of the same peptide seeded with different initial velocities can exhibit a range of fast- and slow-folding behavior. By increasing the number of swap attempts per REMD cycle, we are able to reduce heterogeneity by diminishing the presence of slower-folding trajectories. Finally, we notice that convergence often can be signaled by a spike in the population of the most populated configurational cluster - a metric that is independent of the native structure. These results suggest that the systematic application of long runs, multiple trials, and convergence indicators may be important in future folding studies and in force field development efforts.