Stochastic Policies Enhance Molecular Transition Path Sampling Robustness

This paper addresses a critical challenge in molecular transition path sampling (TPS), a technique vital for understanding biomolecular mechanisms. Traditional machine learning-based control methods for TPS, which learn forces during molecular dynamics (MD) rollouts, often suffer from instability and strong dependence on initial conditions. While these methods produce physically plausible trajectories, their reliability has been a concern. The researchers propose a solution by recasting rollout-based control as learning a path-space proposal distribution and introducing controlled stochasticity. They developed two novel stochastic policies: FS-TPS, which directly parameterizes a state-dependent Gaussian distribution for control policy output, and LaS-TPS, which samples a compact latent control variable to decode structured, cross-atom-correlated force variations. Extensive multi-seed experiments on various biomolecular systems, including alanine dipeptide, chignolin, and the fast-folding protein BBL, demonstrated the effectiveness of these stochastic policies. They consistently improved transition success rates and path quality compared to deterministic policy baselines. Crucially, they also substantially reduced sensitivity to random initialization, making the TPS process more robust and reliable for scientific discovery.