RevGraphVAMP: A protein molecular simulation analysis model combining graph convolutional neural networks and physical constraints

Abstract

Molecular simulation (MD) is an important research area in the field of life sciences, focusing on understanding the mechanisms of biomolecular interactions at atomic scales. Protein simulation, as a critical subfield of life science, has frequently adopted MD for implementation, where the trajectory data play an important role in drug discovery. With the advancement of high-performance computing and deep learning technology, machine-prediction of protein properties from enormous trajectory data becomes popular and critical, which puts challenges on how to extract useful data features from the complicated simulation data and reasonably reduce the dimensionality. At the same time, in order to better study the Protein system, it is necessary to provide a meaningful explanation of biological mechanism for dimensionality reduction. In order to address this issue, a new unsupervised model RevGraphVAMP is proposed to intelligently analyze the simulation trajectory. RevGraphVAMP is based on the Markov variation method (VAMP) and innovatively integrates graph convolutional neural networks and physical constraint optimization to improve the learning performance of the model. Besides, the attention mechanism is introduced to calculate the importance of protein molecules, leading to interpretation of molecular mechanism. Compared with other VAMPNets models, the new model presented in this paper has achieved the highest VAMP scores and better state transition prediction accuracy in two public datasets. Additionally, it has higher dimensionality reduction discrimination ability for different substates and provides interpretable results for protein structural characterization.