Reinforced molecular dynamics: Physics-infused generative machine learning model simulates protein motion.

We propose a simple and practical machine learning-based desktop solution for analyzing biologically relevant protein motions. We termed our technology reinforced molecular dynamics (rMD) combining MD trajectory data and free-energy (FE) map data to train a dual loss function autoencoder network that can explore conformational space retroactively with no need for additional simulations. The key insight of rMD is that it replaces the latent space with an FE map for structure prediction, thus infusing the autoencoder network with physical context. The FE map is computed from a biased MD simulation over a low-dimensional collective variable (CV) space that captures some biological function. In the proposed rMD framework, the FE map can then be used directly to generate more protein structures in poorly sampled regions or to follow alternative paths in CV space to explore multiple conformational transitions. The rMD technology is entirely self-contained, does not rely on any pretrained model, and can be run on a GPU desktop computer. We present our rMD computations in a key area of molecular-glue degraders aimed at a deeper understanding of the structural transition from open to closed conformations of CRBN.