Accurate estimation of solvation and binding free energies is a cornerstone of modern drug discovery, yet conventional force fields are still tuned empirically. We introduce Deep Origin's small-molecule neural network potential (NNP), trained on high-level quantum data and natively coupled with classical molecular mechanics in our proprietary MD engine. This hybrid MM/NNP framework supports practical free energy perturbation (FEP) simulations without the typical computational costs.Benchmarking against FreeSolv solvation database and BRD4 binding complex demonstrates high-level accuracy for intramolecular interactions and absolute binding ΔG predictions. Our CPU-optimized implementation achieves ~600 ns/day for 50 K-atom complexes—an order of magnitude faster than comparable GPU-based NNP-FEP methods—making it practical for large scale drug discovery projects. We show improved ΔG on example systems which we theorize is partially due to improved sampling of hydroxyl groups.These results establish NNP-augmented FEP as a scalable, accurate tool for accelerating drug design. Ongoing work will expand benchmarks to more challenging chemical spaces and release our models and workflows through Deep Origin's computational platform.
Originally presented at the Free Energy Workshop 2025