A Generative Foundation Model for Cryo-EM Densities

Abstract

Single-particle cryo-electron microscopy (cryo-EM) enables structure determination of macromolecular complexes, yet reconstructions are often degraded by noise, anisotropic sampling and reconstruction artifacts. Correcting these effects is an ill-posed inverse problem that requires strong prior knowledge in addition to experimental data. Here we introduce cryoFM, a generative foundation model for cryo-EM densities that unifies data-driven structural priors with dataset-specific constraints within a Bayesian inference framework. CryoFM is trained in an unsupervised manner on thousands of high-quality cryo-EM maps using flow matching, learning a generalizable prior over macromolecular density distributions. Combined with explicit likelihood models describing experimental degradations, cryoFM enables flow posterior sampling, an inference-only procedure that performs denoising and restoration and refinement while remaining explicitly constrained by dataset-derived statistics. We show that this framework improves density reconstruction and refinement across diverse experimental settings, including preferred-orientation datasets and cases with strong spatial heterogeneity in signal-to-noise ratio, without introducing hallucinated features. In addition, cryoFM can be fine-tuned into conditional generative models for density post-processing, yielding maps with improved interpretability and fewer artifacts compared to existing supervised approaches. Together, cryoFM establishes generative foundation models as a principled and controllable framework for cryo-EM density reconstruction and modification.