BPCPPDYSOE21 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

BP: Fachverband Biologische Physik

BP 11: Poster A: Single Molecule, Multicellular, Bioimaging, Focus Sessions, etc.

BP 11.4: Poster

Montag, 22. März 2021, 16:30–19:00, BPp

Deep reinforcement learning of molecular mechanisms — •Roberto Covino¹, Hendrik Jung², Arjun Wadhawan³, Peter G. Bolhuis³, and Gerhard Hummer^2,4 — ¹Frankfurt Insitute for Advanced Studies, Frankfurt am Main, Germany — ²Max Planck Institute of Biophysics, Frankfurt am Main, Germany — ³Van 't Hoff Institute for Molecular Sciences, University of Amsterdam, Amsterdam, The Netherlands — ⁴Institute of Biophysics, Goethe-University Frankfurt, Frankfurt, Germany

We present a deep reinforcement learning artificial intelligence (AI) that learns the molecular mechanism from computer simulations. The AI simulates molecular reorganizations and progressively learns how to predict their outcome. We integrate path theory, transition path sampling (TPS), and deep learning. TPS is a Markov Chain Monte Carlo method to sample the rare trajectories connecting metastable states. Using reinforcement learning, we iteratively train a deep neural network on the outcomes of TPS simulation attempts. In this way, we increase the rare-event sampling efficiency while gradually revealing the underlying mechanism. At convergence, the AI learns the rare events' committor function, encoded in the trained neural network. By using symbolic regression, we distill simplified quantitative models that reveal mechanistic insight in a human-understandable form. Our innovative AI enables the sampling of rare events by autonomously driving many parallel simulations with minimal human intervention and aids their mechanistic interpretation.