Accurate Molecular Mechanics Force Fields through Data-driven Parameter Type Definitions
Biophysics
Computer-Aided Design of Materials and Simulation of Materials Behaviour from Atomic to Microscopic Scale
Final Report Abstract
Molecular mechanics forcefields describe how atoms and molecules interact with each other in a very efficient manner. Therefore, molecular mechanics forcefields are commonly used potential energy functions for representing molecular interactions in computer simulations and enable researcher to model the interactions of biomolecules and drugs. Most of the forcefields that are currently used are physically-motivated and contain a large number of physical parameter values, which are optimized against high-level reference calculations or experimental data. In addition to the physical parameter values, forcefields contain a set of mapping instructions, called parameter type definitions, that define how exactly the physical parameter values are mapped to the molecules and their substructures (atoms, bonds, angles, etc.). These parameter type definitions are manually drawn by a human expert and will impact the optimized physical parameter values. This manual definition of parameter type definitions causes all sorts of problem regarding accuracy and transferability of the forcefields (i.e. how well will the forcefield perform for molecules it hasn’t seen before). In the present work, we developed an approach for automatically discovering the underlying parameter type definitions and the corresponding physical parameter values only based on the data used for force field optimization. We successfully developed and implemented a method that is able to infer a complete forcefield from scratch, automatically adjusting the number of parameters, and thus the complexity of the forcefield, based on the data and statistically meaningful hyperparameters. We demonstrate that our method works by generating training data from an existing forcefield and then letting our method reconstruct the forcefield from scratch. We found that the forcefield is reconstructed with slightly different parameter type definitions and physical parameter values than the original forcefield. However, we found that our forcefield has the same accuracy as the original forcefield, but requires fewer parameters. Thus, our approach is not only able to generate forcefields from scratch, it is also able to generate forcefields that are even less complex than existing forcefields.
