#35 · Structural bio & protein design

Learned models reshape cryo-EM reconstruction and electron microscopy acquisition

The dominant thread is machine learning embedded in structural and volumetric imaging pipelines, applied either to reconstruction and heterogeneity analysis or to acquisition efficiency. cryoFM, a flow-matching generative foundation model for cryo-EM densities, is combined with likelihood models in Bayesian posterior sampling to denoise and refine maps. A benchmark of five cryo-EM heterogeneity methods (3DFlex, 3DVA, cryoDRGN, ManifoldEM, Bayesian reweighting) on public TRPV1 data, alongside molecular dynamics simulation, provides a comparative reference point for conformational analysis. SmartEM applies machine learning during real-time SEM acquisition, rescanning only needed regions for roughly seven-fold faster connectomic imaging of nematode and mouse samples. The items differ in modality and scale, so the shared element is methodological rather than a common mechanism. What is new is the use of learned models as generative priors for density reconstruction and as acquisition-time controllers for electron microscopy.