Intrinsic map dynamics exploration for uncharted effective free-energy landscapes

Significance Direct simulations explore the dynamics of physical systems at their natural pace. Molecular dynamics (MD) simulations (e.g., of macromolecular folding) extensively revisit typical configurations until rare and interesting transition events occur. Biasing the simulator away from regions already explored can, therefore, drastically accelerate the discovery of features. We propose an enhanced sampling simulation framework, where MD and machine learning adaptively bootstrap each other. Machine learning guides the search for important configurations by processing information from previous explorations. This search proceeds iteratively in an algorithmically orchestrated fashion without advance knowledge of suitable collective variables. Applied to a molecular sensor of lipid saturation in membranes, a helix dissociation pathway not seen in millisecond simulations is discovered at the second iteration.