Deep learning potentials for hydration and protonation in biomolecular simulations: bond breaking is the goal and the problem

Biomolecular dynamics has long been known to be influenced by hydration shells. However, first principles calculations of hydration shell structure and dynamics, and the effects on the solvated biomolecule, have been historically out of reach due to computational costs: even the smallest simulations of fully solvated small biomolecular fragments using periodic boundary conditions can require over 500 atoms, and molecular dynamics on these systems are limited to sub-nanosecond simulation times. Using high-performance computing and highly optimized, scalable density functional theory (DFT) programs, we are now able to overcome some of these challenges to discover the effects of the hydration shell on the dynamics of biomolecular systems from a first principles perspective. These simulations still suffer from timescale restrictions, preventing the simulation of timescales relevant to most experimental validation. Recently, advances in deep neural network potentials (DNNPs) promise to extend the accuracy of first principles simulations to much larger systems and can be used to simulate hundreds of nanoseconds. Now the challenge becomes training and retraining the model to avoid incorrect behavior, especially across transition states. For solvated biomolecules, protonation state changes resulting from proton transfers are now accessible to the DNNP, while for classical empirical force fields these are generally not allowed. Unphysical transfer rates must be corrected by sampling transition state regions and teaching the model the correct energetic barriers for each proton, a difficult task. Here we discuss some of our results tackling these challenges for solvated nucleic acids, for which hydration structure plays an important role.