Fold'EM uses the structural priors from protein generative models like AlphaFold to infer atomic structures directly from cryo-EM particle images, skipping the usual intermediate step of reconstructing a 3D electrostatic potential map. This allows accurate structure determination from far fewer particles than conventional methods: less than 1% with known particle orientations, about 3% ab initio, and roughly 5% for heterogeneous samples. The framework also resolves distinct conformational states from mixed particle populations without separate reconstructions.
The work is interesting because it addresses a major bottleneck in cryo-EM: the high sample complexity and cost associated with needing large numbers of particle images for density reconstruction. By leveraging strong sequence-derived priors, Fold'EM opens the possibility of structure determination for scarce samples and low-population conformational states that are often inaccessible. The approach redefines the role of experiment and prior in structure determination.