Highly parallelizable path sampling with minimal rejections using asynchronous replica exchange and infinite swaps

• 成果类型: Article
• 署名作者: Zhang, Daniel T.; Baldauf, Lukas; Roet, Sander; Lervik, Anders; van Erp, Titus S.
• 署名单位: Norwegian University of Science & Technology (NTNU); Utrecht University
• 刊物名称: PROCEEDINGS OF THE NATIONAL ACADEMY OF SCIENCES OF THE UNITED STATES OF AMERICA
• ISSN/ISSBN: 0027-14952
• DOI: 10.1073/pnas.2318731121
• 发表日期: 2024-02-13

摘要

Capturing rare yet pivotal events poses a significant challenge for molecular simulations. Path sampling provides a unique approach to tackle this issue without altering the potential energy landscape or dynamics, enabling recovery of both thermodynamic and kinetic information. However, despite its exponential acceleration compared to standard molecular dynamics, generating numerous trajectories can still require a long time. By harnessing our recent algorithmic innovations-particularly subtrajectory moves with high acceptance, coupled with asynchronous replica exchange featuring infinite swaps-we establish a highly parallelizable and rapidly converging path sampling protocol, compatible with diverse high-performance computing architectures. We demonstrate our approach on the liquid-vapor phase transition in superheated water, the unfolding of the chignolin protein, and water dissociation. The latter, performed at the ab initio level, achieves comparable statistical accuracy within days, in contrast to a previous study requiring over a year. Significance Current molecular dynamics simulations have the capability to faithfully describe chemical reactions, nucleation, and protein folding events, providing essential information to drive the progress of technologies like catalysis and drug discovery. Nevertheless, even with the fastest supercomputers, a significant portion of these phenomena remains beyond the reach of standard simulations due to waiting times that can exceed billions of CPU years. While enhanced sampling techniques like path sampling allow for exponentially faster study of these events, obtaining converged results in experimentally relevant systems still typically spans from months to years. Here, we apply four innovative algorithmic enhancements that transform the state-of-the-art path sampling algorithm, replica exchange transition interface sampling (RETIS), into infinity RETIS, allowing convergence in a matter of days.