Explainable Machine Learning Guided Enhanced Sampling of Protein Conformational Transition in HSP90

Abstract

Elucidating the thermodynamics, kinetics, and mechanisms of protein conformational transitions remains a longstanding challenge in molecular dynamics simulations. We employ enhanced sampling simulations using explainable machine-learning (ML)-based collective variables (CVs) to efficiently explore the free-energy landscape of the millisecond-time scale ATP-lid conformational transition in heat shock protein 90 (HSP90). Our simulations yielded relative free energies consistent with nuclear magnetic resonance (NMR) experiments at modest computational cost. The interpretability of the surrogate model CV enabled the identification of key residues in the ATP-lid transition, providing mechanistic insight at atomistic resolution. The ML-CV trained on the wild-type system shows transferability to two mutant variants, successfully reproducing experimental population shifts. We could also obtain accurate kinetic rates of the conformational transition using an integrated sampling strategy that combined biased enhanced sampling with an unbiased weighted ensemble algorithm. Our framework provides detailed thermodynamic, kinetic, and mechanistic insights into the complex conformational transitions of the N-terminal domain of HSP90, highlighting the potential of explainable machine-learning-based collective variables for investigating complex conformational landscapes in biomolecular systems and informing future inhibitor design efforts.