Physically Interpretable Biomolecular Conformation Generation with A Deep Probabilistic Framework

Abstract

The structure of proteins is highly dynamic, and their ability to undergo conformational changes is crucial for their biological function. While deep learning has shown success in predicting static 3D structures of proteins from their amino acid sequence, efficiently sampling various conformations and transition states between them remains a challenge. To address this issue, we propose a novel framework that combines theories and methods from coarse-graining and deep learning communities to generate biomolecular conformations in an efficient and physically interpretable manner. Our model learns a coarse-grained mechanical representation of alanine dipeptide as an elastic network of interconnected harmonic springs and generates novel but realistic conformations from this representation. We demonstrate that our model also preserves the essential physics of the molecular system by learning the correspondence of dihedral angles between coarse-grained and fully-atomistic conformations. Finally, we use our model to explore interpolating between two distant metastable conformations of alanine dipeptide by mechanically manipulating the coarse-grained conformations, generating smooth and physically plausible transition pathways implied by the forcefield but not present in our training data. Our work contributes to the field of molecular modeling by providing an efficient and parsimonious framework to sample from the conformational landscape of complex biomolecules, which can help in understanding the mechanisms of biological processes, accelerating drug discovery, and advancing the field of protein design.