Graph Machine Learning for (Bio)Molecular Modeling and Force Field Construction

Abstract

Force fields (FFs)---the (parametrized) mapping from the geometry of a physical state to potential energy and forces---are a crucial component of molecular dynamics (MD) simulations, whose associated Boltzmann-like target probability densities are sampled to estimate ensemble observables, to harvest quantitative insights of the system. Currently, state-of-the-art force fields are either fast (molecular mechanics, MM-based) or reliable (quantum mechanics, QM-based), but seldom both. Machine learning force fields strive to approach a balance between speed and accuracy by fitting simple neural functional forms to QM energies and forces, which, though approaching or surpassing chemical accuracy on some limited chemical spaces, are still orders of magnitude slower than MM force fields. In this thesis, leveraging graph-based machine learning and incorporating inductive biases crucial to chemical and biophysical modeling, we approach the lotusland from two angles---to make MM more accurate and to make machine learning force fields faster. Along the way, we design a plethora of useful gadgets, including the first unified force field for joint protein-ligand parametrization, an AM1-BCC surrogate charge model thousands-fold faster with error smaller than discrepancies among implementations, and a way to forecast the fate of dynamic systems before the simulation even starts.