Accuracy of small molecule force fields lies at the crux of computer aided drug design. While a wide range of additive force fields have been remarkably successful in addressing the key findings of drug design projects, their limitations impede their accuracy. One of the key limitations is electronic polarizability, which is addressed by polarizable force fields, such as CHARMM’s Drude Polarizable Force Field. Drude General Force Field (DGenFF) is CHARMM’s polarizable general force field for small organic molecules. To be as complete as CHARMM General Force Field (CGenFF), DGenFF requires 43 new atom types. Introduction of new atom types involves several stages of parameter optimization, most challenging one being the optimization of Lennard-Jones (LJ) parameters. The LJ parameters are optimized to reproduce reference properties involving target model compounds. The reference data includes experimental thermodynamic properties, like enthalpy of vaporization (ΔHvap) molecular volume (VMol), dielectric constant, and viscosity, among others. To perform LJ parametrization while accounting for the high dimensionality of parameters we harness the predictive ability of Deep Learning (DL) by combining it with an initial design algorithm, Latin Hypercube Design (LHD). LHD extensively explores the multidimensional LJ parameter space for multiple atom types simultaneously, to identify a subset of distributed parameter sets. ΔHvap and VMol are then calculated via MD simulations for such parameter sets for selected model compounds creating the training dataset for DL model. The trained DL model is then used to identify optimal sets of parameters closer to the target experimental data, which are further refined and ranked based on QM interactions with rare gases, water, and other experimental observables. The developed DL-based parametrization framework represents a high throughput LJ parameter optimization based on experimental and QM properties. Application of the DL LJ optimization approach to several new atom types will be presented.

Harnessing the power of deep learning in force field development: Lennard-Jones parameter optimization for Drude General Force field (DGenFF)

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.