R-CCS Cafe 第229回 第2部

Coarse grained molecular dynamics (CG-MD) simulation, in which one degree of freedom for each residue is considered, has succeeded in reproducing various biological processes (dynamics) for many macro molecules qualitatively with lower calculation cost compared to all atom molecular dynamics (AA-MD) simulation. However, CG-MD simulations still remains several problems. In particular, CG models sensitively depend on various parameters such as inter-molecule interaction. Therefore, searching suitable parameter sets which succeeds in reproducing a target biological process qualitatively is quite important. Because exhaustive investigation of all candidate parameters for CG-MD simulation needs high computational costs, it is essential to identify efficiently successful parameters that can reproduce biological function. In this work, we proposed a efficient searching method to explore the successful parameter sets for CG model by combining two machine learning techniques: Bayesian optimization and active Learning method.Concretely, we evaluated its performance using a biological rotary motor. We successfully identified the successful region with lower computational costs (about 10%) without sacrificing accuracy compared to exhaustive search.