Finding low dimensional representation of data from long-timescale trajectories of biomolecular processes, such as, protein folding, or protein-ligand binding, is of fundamental importance. In this contribution, we combine ideas from Variational approach to Markov processes (VAMP) and Graph Neural Networks to generate an end-to-end framework to efficiently learn feature transformations for building a high-resolution Markov model. Our GraphVAMPNet method is an enhancement of the previous VAMPNet approach and uses graph message passing to define the structural features. Presenting the molecules as graphs has the advantages of rotational, permutational and translational invariance that is built into graph neural networks. We have implemented several graph neural networks that differ in how the messages are propagated through nodes and edges of molecular graphs and tested their efficiency in learning molecular kinetics for multiple long-timescale protein folding trajectories. Our method performs better than the previous VAMPNet with faster convergence of timescales and more interpretable few-state kinetic models. Moreover, since we use general graph embeddings for molecular representations, our method is inherently transferable across molecular space which can theoretically reduce the amount of data for learning molecular kinetics.

GraphVAMPNet: Learning molecular kinetics using graph neural networks and variational approach to Markov processes

ACS Spring 2022

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