JAMUN: Transferable Molecular Conformational Ensemble Generation with Walk-Jump Sampling

Conformational ensembles of protein structures are immensely important both to understanding protein function, and for drug discovery in novel modalities such as cryptic pockets. Current techniques for sampling ensembles are computationally inefficient, or do not transfer to systems outside their training data. We present walk Jump Accelerated Molecular ensembles with Universal Noise (JAMUN), a step towards the goal of efficiently sampling the Boltzmann distribution of arbitrary proteins. By extending Walk-Jump Sampling to point clouds, JAMUN enables ensemble generation at orders of magnitude faster rates than traditional molecular dynamics (MD) as well as state-of-the-art ML methods. JAMUN relies on learning a "smoothed" manifold, and then using the same learned score to denoise to the physical manifold. The smoothed manifold allows for Langevin dynamics that is faster as it does not need to overcome thermodynamic traps, while retaining enough structure and information from the data to result in much faster and more reliable denoising. We demonstrate the efficacy of JAMUN on two datasets consisting of 2 amino-acid long polypeptides, and show that it is able to predict the stable basins of small peptides that were not seen during training. We quantify our sampling with Shannon-Jensen divergences of the full distribution as well as for MSM populations trained on the MD data.