Parallel Sampling of Protein-Ligand Dynamics

Abstract

AbstractMolecular dynamics (MD) simulations of protein-ligand complexes are essential for computer-aided drug design. In particular they enable the calculation of free energies and thus binding affinities. However, these simulations require significant computational resources and can take days to weeks to achieve relatively short timescales compared to biologically relevant timescales. To address this issue, we introduce a method for non-sequential generation of MD samples using a generative deep neural network trained on a large corpus of protein-ligand complex simulations. The method generates accurate protein-ligand complexes with full protein and ligand flexibility and is able to recapitulate the conformation space sampled by MD simulations with high coverage. This development is a step forward towards one-shot molecular sampling that can be utilized in the calculation of protein-ligand free energies.3