A new AI framework builds atomic-resolution protein structures straight from raw cryo-EM images, skipping the blurry density map that normally sits in between.
Researchers call it Fold'EM. The standard cryo-electron microscopy pipeline stitches together many particle images into an electrostatic potential map, then fits an atomic model to that map after the fact. That two-step process needs a lot of particle images and gets harder as map resolution drops, especially for samples that shift between multiple shapes. Fold'EM instead feeds priors from protein generative models directly into the raw particle images at inference time, producing an atomic model without ever building the intermediate map. The team tested it on synthetic and real cryo-EM datasets, both when particle orientations were already known and when the model had to infer them from scratch.
The bigger claim is on heterogeneous samples: proteins that exist in several conformations in the same batch of particles. Fold'EM reportedly pulls distinct conformational states out of a mixed particle population without reconstructing and modeling each state separately. That targets exactly the cases where cryo-EM is slowest and most expensive today - flexible proteins and low-population states that don't have enough particles to build a clean map the old way.
This is the same move AlphaFold made popular - using a model's learned sense of plausible protein shapes as a shortcut - but applied earlier, fused directly with experimental measurements instead of bolted on after a map is already built. If it holds up, it could mean usable structures from far fewer particle images, which is the real bottleneck in cryo-EM cost and throughput.
Worth remembering: this is a preprint, not yet peer reviewed, and generative priors are good at producing structures that look right whether or not they are. A method that skips the sanity check of a full density map needs more scrutiny before anyone trusts it on a novel drug target.