A Graph-Based Machine Learning Framework to Assign Empirical Interaction Parameters for Novel Molecules

Molecular dynamics (MD) simulations of novel molecules are essential to modern drug discovery and related fields that require large-scale sampling of chemical space. However, obtaining accurate and robust force field parameters for novel molecules remains a major challenge. One approach to this problem is to perform quantum chemical calculations on a per-molecule basis, but doing so on a large scale is prohibitively expensive. Here, we present a graph-based machine learning framework for empirical force field assignment that scales significantly more efficiently with molecule size than quantum chemical methods, giving speedups of up to seven orders of magnitude for larger (20+ atoms) molecules. We demonstrate the utility of our framework by applying it to predict harmonic bond force constants and show that the resulting model outperforms the popular Seminario method for force constant assignment when using experimental data as benchmarks. This performance gain is likely a result of the model's intrinsic homogenization effect, which counteracts the noise in the training data. Moreover, our framework overcomes several challenges in translating first-principles calculations to classical MD force field parameters, producing parameters with better transferability and interpretability compared to previous models of this class.