Seminar on AI for Molecular Mechanism Discovery by Gianmarco Lazzeri

Rare events, where molecular systems transition between metastable states, are common in nature. Studying them with molecular dynamics simulations presents two main challenges: reducing the computational cost of sampling transitions and gaining insights into the underlying mechanisms. AI for molecular mechanism discovery (AIMMD) addresses both tasks simultaneously. It optimizes a reaction coordinate (RC) that guides the sampling while also using the RC and the sampling to access unbiased free energies, transition mechanism, and rates. In this active learning scheme, no external biasing forces are applied. Instead, AIMMD generates short, physically unbiased paths, reweights them, and assembles them to mimic long-timescale equilibrium behavior. Enhancing the sampling reduces to selecting the paths’ starting conditions from already obtained configurations. This minimal intervention provides robustness and accelerates convergence. I will present applications of AIMMD to complex biological systems, including transmembrane protein dimerization and lipid uptake by P116.