Exploring Protein Phase Spce using Machine Learning based Enhanced Sampling

Abstract

Molecular Dynamics simulations of Condensed matter systems are known to have very slow dynamics and thus often leading to poor sampling of the phase space. This can lead to wrong interpretation of system’s macroscopic properties properties. In this paper, we would like to introduce a machine learning based approach that can efficiently explore the phase space of a system. This is done in a two-step process: (1) classify metastable states into unique clusters (classifier model) (ii) develop a non-linear relation within each cluster (regressor model). We apply this method on a simple system like alanine tetrapeptide and show that there is a drastic improvement in sampling compared to a benchmark study with already established collective variables. We also apply this method on a larger peptide sequence, namely a Chameleon sequence. The model is able to reach all the secondary structures with reasonable sampling.