Simon Boothroyd

We report a new charge model and a new general small molecule force field. Here, we address the development and benchmarking of both the Open Force Field (OpenFF) AshGC charge model, as well as the Sage 2.3.0 small molecule force field for drug-like molecules. AshGC is a graph neural network-based method for efficient assignment of charges to both small and large molecules, designed to produce charges of semi- empirical quality at low cost. We describe the model, its architecture, featurization, training process, and performance. AshGC largely generates charges within the range of other accepted AM1-BCC backends such as OpenEye’s oequacpac and AmberTools’ sqm, but we note areas of poor performance such as in smaller molecules and in charged molecules, as well as areas of chemistry underrepresented in the training set. We fur- ther describe the development and performance of Sage 2.3.0, where Lennard-Jones and valence parameters are re-trained with neural network charges for the first time. Our benchmark suite of common force field tasks ranges from optimizing molecular structures in vacuum to protein-ligand binding free energies. Overall, Sage 2.3.0 performs comparably to previous Sage force fields, although we report minor improvements in the prediction of condensed phase properties and a minor decrease in performance in the prediction of non-aqueous solvation free energies. As with other OpenFF force fields, Sage 2.3.0 was validated in protein-ligand benchmarks to be compatible with Amber protein force fields. All data are publicly available, along with scripts, environments, and complete details for reproducing the training and benchmarking of AshGC and Sage 2.3.0 at https://github.com/openforcefield/ashgc-v1.0-fit and github.com/openforcefield/ash-sage-rc2 respectively.

Common methods for assigning atom-centered partial charges in computational chemistry, such as RESP and AM1-BCC, rely on quantum mechanical or semi-empirical calculations of the molecule of interest, which are expensive to compute and dependent on the choice of input molecular conformer(s). Graph neural network (GNN) based continuous atom embeddings have been shown to be a fast and flexible solution for partial charge assignment, but those developed so far for condensed phase modeling have usually been trained to reproduce AM1-BCC charges, which themselves seek to reproduce the HF/6-31G(d) molecular electrostatic potential. Here, we investigate the suitability of various common charge assignment schemes, including ESP and atoms-in-molecule (AIM) based approaches, as training targets for new GNN based charge models. We show that the strengths of both approaches can be combined by co-training GNN models to AIM charges and molecular dipoles and electrostatic potentials. We collect a dataset of quantum mechanical AIM properties computed at a high level of theory (wB97X-D/def2-tzvpp), in both vacuum and implicit solvent, and train new GNN charge models to each. Charges can be scaled between the vacuum and solvated charge sets, and combined with Lennard-Jones parameters optimized using the Open Force Field infrastructure, to yield force fields that are suitably polarized for condensed phase modeling. We further demonstrate that the charge models may be applied to explore electrostatics-driven structure-activity relationships in medicinal chemistry. The charge models are freely available at: https://github.com/cole-group/nagl-mbis/.

A wide range of density functional methods and basis sets are available to derive the electronic structure and properties of molecules. Quantum mechanical calculations are too computationally intensive for routine simulation of molecules in the condensed phase, prompting the development of computationally efficient force fields based on quantum mechanical data. Parametrizing general force fields, which cover a vast chemical space, necessitates generating sizable quantum mechanical datasets with optimized geometries and torsion scans. To achieve this efficiently, it is crucial to choose a quantum mechanical method that balances computational cost and accuracy. In this study we seek to assess the accuracy of quantum mechanical theory for specific properties such as conformer energies, and torsion energetics. To comprehensively evaluate various methods, we focus on a representative set of 59 diverse small molecules, comparing approximately 25 combinations of functional and basis sets against the reference level coupled cluster calculations at complete basis set limit.

Software to more rapidly and accurately predict protein--ligand binding affinities is of high interest for early-stage drug discovery, and physics-based methods are among the most widely used technologies for this purpose. The accuracy of these methods depends critically on the accuracy of the potential functions they use. Potential functions are typically trained against a combination of quantum chemical and experimental data. However, although binding affinities are among the most important quantities to predict, experimental binding affinities have not to date been integrated into the experimental dataset used to train potential functions. In recent years, the use of host--guest complexes as simple and tractable models of binding thermodynamics has gained popularity due to their small size and simplicity, relative to protein--ligand systems. Host--guest complexes can also avoid ambiguities that arise in protein--ligand systems, such as uncertain protonation states. Thus, experimental host--guest binding data are an appealing additional data type to integrate into the experimental dataset used to optimize potential functions. Here, we report the extension of the Open Force Field Evaluator framework to enable the systematic calculation of host--guest binding free energies and their gradients with respect to force field parameters, coupled with the curation of 126 host--guest complexes with available experimental binding free energies. As an initial application of this novel infrastructure, we optimized generalized Born (GB) cavity radii for the OBC2 GB implicit solvent model against experimental data for 36 host--guest systems. This refitting led to a dramatic improvement in accuracy for both the training set and a separate test set with 90 additional host--guest systems. The optimized radii also showed encouraging transferability from host--guest systems to 59 protein-ligand systems. However, the new radii are significantly smaller than the baseline radii and lead to excessively favorable hydration free energies (HFE). Thus, users of the OBC2 GB model currently may choose between GB cavity radii that yield more accurate binding affinities or GB cavity radii that yield more accurate HFEs. We suspect that achieving good accuracy on both will require more far-reaching adjustments to the GB model. We note that binding free energy calculations using the OBC2 model in OpenMM gain about a 10x speedup relative to corresponding explicit solvent calculations, suggesting a future role for implicit solvent absolute binding free energy (ABFE) calculations in virtual compound screening. This study proves the principle of using host--guest systems to train potential functions that are transferrable to protein--ligand systems, and provides an infrastructure that enables a range of applications.