Using structure to evaluate the accuracy and transferability of machine learning potentials of biomolecules.

Molecular simulation is a powerful tool to describe chemical and physical processes across length scales. Fundamental tradeoffs between the accuracy of quantum mechanics and speed of classical models makes simulating biosystems very difficult. Machine learning potentials (MLPs) offer a potential solution which combines the accuracy of electronic structure theory with the computational scaling of classical approaches. Applying MLPs to biosystems could unlock new developments in drug discovery, biological mechanisms and interactions. Despite this promise, rigorous benchmarks are needed to determine whether MLPs reliably reproduce first-principles counterparts. It is commonplace to use energies and forces as the primary performance metrics for MLPs. While essential, we show that models with among the lowest energy errors can simultaneously produce the largest structural errors. Through simulations of two isomeric biomolecules, this work shows that structural information must be added as a metric to assess both the overall accuracy and transferability of an MLP. Here, we show that when structural quantities are included alongside traditional metrics, MLPs can achieve high fidelity to first-principles reference data and remain transferable across related chemical motifs examined. These findings underscore the importance of structure-based validation and support the continued use of MLPs for biomolecular simulations.