Discovering mechanism by which biomolecules function is crucial to understanding cellular processes. As a small change at one part of a protein, for example, could lead to a significant change of a much larger spatial scale at another part, the mechanism is generally multi-steps, multi-domains, and multi-scales. While molecular dynamics (MD) simulations have been routinely used to study mechanisms of proteins, no analysis method, to the best of our knowledge, could automatically explain all three of these aspects simultaneously.<br/><br/>Multi-layers models, including artificial neural networks and hierarchical probabilistic models, are potentially suitable to explain the mutli-scales dynamics that drive the function of proteins. However, designing a multi-layer model to explain mechanism would require overcoming unique challenges. Firstly, deep neural networks are highly over-parameterized. Many sets of very different parameters can fit and predict the motion of a protein in a molecular dynamics simulation equally well. Therefore, fitting the parameters of a deep neural network generally does not directly give a reproducible and unique explanation of protein dynamics. Secondly, multi-layer models have a large number of parameters and a large parameter space. Therefore, multi-layer models require more efficient sampling methods to obtain more reproducible results.<br/><br/>We propose unified multi-layer models to explain contribution of interactions between residues to the domain dynamics described by coupled hidden Markov models. We will also present two strategies to improve fitting of multi-layer models. Firstly, we applied enhanced sampling methods, which are traditionally used in MD simulations, to fitting of the model. Secondly, we devise a method to control the size of the parameter space by growing the multi-layer models incrementally while ensuring the model remain explainable.

Breaking into the molecular dynamics of protein residues with coupled multi-layer machine learning and enhanced fitting methods

ACS Fall 2020 Virtual Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.