Generative AI models can design new proteins, but they often produce physically unrealistic structures because they lack knowledge of real-world molecular forces. Traditional physics-based forcefields, mathematical descriptions of atomic interactions, can catch these errors, but they are too slow to run on graphics cards and don't work smoothly inside machine learning tools. This limits their use in training or refining AI models.
The authors built tmol, an open-source PyTorch library that reimplements the RosettaGenFF forcefield as fast, parallel operations on a GPU. It scores entire batches of protein structures, computes energy gradients automatically, and includes key algorithms like side-chain optimization and structure relaxation, all directly usable within PyTorch pipelines.
When used as a reward signal during generative protein interface design, tmol improved the number of stabilizing hydrogen bonds and water bridges between designed proteins and their targets. The library achieved up to 362 times higher throughput than previous implementations, making it practical to incorporate physics-based feedback into machine learning training loops for the first time.
