Combining physics and machine learning for computer-aided drug design

Computational modeling has the potential to significantly decrease costs associated with drug development. Equations from physics can provide accurate property predictions, though they are often impractical to calculate in a reasonable time frame. The rise in machine learning (ML) applications to biology presents an exciting opportunity to improve predictive power of molecular models, though lack of constraints can lead to unphysical results. In this talk, I will outline recent progress using weighted ensemble enhanced sampling simulations of SARS-CoV-2 spike to reveal a glycan critical for spike opening, emphasizing the significance of molecular pathways for uncovering key mechanistic details. Next, I will share results using ML and deep mutational scanning data to characterize the spike mutational fitness landscape, and our ongoing efforts to combine ML and physics for improving virtual screening using data from DNA-encoded library screens.