#34 · Structural bio & protein design

Deep learning extracts molecular structure from single-molecule and cryo-EM images

The dominant thread is computational recovery of structural information from sparse, noisy, or conformationally heterogeneous imaging data. Cryo-IEF, pre-trained on about 65 million cryo-EM particle images, supports classification, pose clustering, and quality assessment. The SMLM study applies a geometric deep learning autoencoder to perform template-free 3D particle averaging of sparse point clouds through latent-space averaging. The AFM study uses unsupervised deep learning to register and cluster image stacks by conformation, extending localization AFM to flexible membrane proteins. NanoporeDB differs in kind: it is a structural resource rather than an image-analysis method, providing over 6,600 multimeric protein nanopore models generated with AlphaFold-Multimer and AlphaFold3. What appears new is the shift toward template-free, latent-space approaches for sparse or conformationally variable targets, alongside large pretrained models applied to particle-level data.