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

BP: Fachverband Biologische Physik

BP 3: Computational Biophysics I

BP 3.9: Vortrag

Montag, 27. März 2023, 12:00–12:15, BAR 0106

Artificial Intelligence for Molecular Mechanism Discovery — •Hendrik Jung¹, Roberto Covino², A Arjun³, Christian Leitold⁴, Peter G Bolhuis³, Christoph Dellago⁴, and Gerhard Hummer¹ — ¹Max Planck Institute of Biophysics, Frankfurt, Germany — ²Frankfurt Institute for Advanced Studies, Frankfurt, Germany — ³University of Amsterdam, Amsterdam, The Netherlands — ⁴University of Vienna, Vienna, Austria

We present a machine learning algorithm to extract the mechanism of collective molecular phenomena from computer simulations. The algorithm combines transition path sampling (TPS), deep learning (DL), and statistical inference to simultaneously enhance the sampling and understanding of complex molecular reorganizations without human intervention. TPS is a Markov Chain Monte Carlo method in trajectory space that samples the rare transition trajectories connecting meta-stable states. In our algorithm a DL model is selecting the configurations from which the new trial trajectories are generated using shooting moves, i.e., the trajectories are propagated according to the physical model of the simulated system. By iteratively training on the outcomes of the shooting moves, the model simultaneously increases the efficiency of the rare-event sampling and gradually reveals the underlying mechanism of the transition. In a second step we distill the knowledge about the transition encoded in the DL model into a simplified mathematical expression. With this algorithm we study a diverse set of molecular systems ranging from the association of ions in solution to the oligomerization of a transmembrane alpha helix dimer.