(585e) Applying Deep Learning to Accelerate Molecular Dynamics Simulation-Based Structural Properties Prediction for Biomolecules.

Molecular dynamics (MD) simulations have been applied to shed light on the mechanisms of protein folding, conformational change, and protein-protein interactions. However, the computational time for large biomolecules, such as antibodies, is notoriously slow. This work combines high-throughput MD simulations and deep learning to develop surrogate models to predict antibody structural properties using only antibody sequences. These structural properties include solvent-accessible surface area, surface charge distribution, and surface hydrophobicity distribution. An extensive set of antibody sequences were retrieved from public databases. The homology models for these antibodies were built to run MD simulations and calculate the structural properties. The deep learning algorithm uses the antibody sequences as input and the structural properties as output for model training. Eventually, efficient surrogate models were developed to predict the structural properties of novel antibodies without the need to perform computationally expensive MD simulations. These surrogate models will facilitate the integration of computer simulations and data science for biomolecules.