Title:Scalable Hybrid Deep Neural Networks/Polarizable Potentials Biomolecular Simulations including long-range effects

Abstract:We present Deep-HP, a scalable multi-GPUs simulation platform dedicated to machine learning potentials. As part of Tinker-HP, it allows users to efficiently use their Pytorch/TensorFlow Deep Neural Networks (DNNs) models (ANI, DeePMD ...) to perform molecular dynamics (MD) simulations on million-atoms systems. Thanks to its MPI setup, Deep-HP extends DNNs simulation timescales while offering the possibility of coupling them to any classical (FFs) and polarizable (PFFs) force fields. We present a novel hybrid MD simulation protocol coupling the AMOEBA PFF to the ANI-2X DNN where solvent-solvent and solvent-solute interactions are computed using AMOEBA while the solute-solute interactions are computed by ANI-2x. It allows for the explicit inclusion of physical long-range effects (multipolar electrostatics & polarization) via efficient periodic boundary conditions. The DNNs/PFFs partition can be user-defined allowing hybrid simulations to include key ingredients such as polarizable solvents, polarizable counter ions... To reduce DNNs computational performance gap with FFs, we rely on extensive multiple timestepping & focus on the models contributions to low frequency modes of nuclear forces. The approach primarily evaluates the PFF forces while including the DNN ones only via a correction step resulting in an overall 30-fold acceleration over standard integration. Thanks to this strategy, accessing $\mu$s simulations with hybrid DNN/PFF models becomes possible. It allowed us to evaluate the accuracy of the ANI2X/AMOEBA model by calculating solvation free energies of various ligands in four different solvents as well as binding free energies of 13 ligand-protein complexes. The approach is shown to be accurate even for charged ligands, outperforming AMOEBA for solvation free energies in non-aqueous solvent and for ligand binding free energies.