Generative Modeling of Protein Conformational Ensembles With Cryo-EM Density Map Diffusion

Knowledge of a protein's conformational ensemble is critical to determining its function, yet state-of-the-art ensemble prediction models are limited by lack of high-quality conformational data from simulation or experiment. Recent advances in heterogeneous reconstruction for cryo-electron microscopy (cryo-EM) have enabled scientists to visualize ensembles of conformational density maps for larger proteins and complexes not typically accessible through simulation, but building atomic models into these maps remains a challenge. In this work, we propose a diffusion model, CryoSampler, that learns to generate atomic conformational ensembles while only being trained on cryo-EM maps and static atomic reference structures. By framing the optimization objective as a map denoiser that internally consists of an atomic predictor followed by a differentiable forward model of the cryo-EM volume rendering process, we show that CryoSampler properly estimates atomic coordinates for the training set and generalizes to new proteins at inference time via unconditional sampling, without needing any density maps. We demonstrate these capabilities on a synthetic dataset of calmodulin proteins simulated with a bimodal distribution.