Fundamental understanding of water-mediated biomolecular interactions is one of the most challenging problems in science. Studies of biomolecular interactions using molecular dynamics simulations or experiments (e.g., atomic force microscope or other molecular tools) are extremely time consuming and costly. Recently it was shown that quantifying the collective properties of water near protein surfaces, specifically, the free energy of dehydration of surface patches using indirect umbrella sampling (INDUS) simulation technique, contains key information about biomolecular interactions.  However, despite the efficiency of INDUS, it is not feasible to sample the vast space of surfaces of all proteins. We believe that a new approach based on AI informed by data from selected INDUS simulations can provide the solution. In this work, we employed sparse INDUS sampling to generate dehydration free energies for a limited number of protein patches. This data was then used to augment a semi supervised variational autoencoder which maps structural and chemical features of protein patches to their hydrophobicity. Our deep learning model can predict context dependent hydrophobicity for protein patches that is several orders faster than INDUS or other techniques. We also observed fascinating insights on how contextual environment and topography shapes the hydrophobicity of protein surface patches. Further, this work can allow users to place a “quilt” on the protein of interest, dividing the surface into patches based on hydrophobicity, and using that information to predicting interaction of proteins with partners of interest.

Using deep learning and MD simulations to model protein interactions

ACS Fall 2021

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