OPUS-TOMO: Deep Learning Framework for Structural Heterogeneity Analysis in Cryo-electron Tomography
Structural heterogeneity in biomolecules, arising from compositional and conformational variability, poses a major challenge for structure determination by cryo-electron tomography (cryo-ET). Here, we introduce OPUS-TOMO, a deep learning framework for resolving multiscale structural heterogeneity across the cryo-ET workflow. Our approach employs a convolutional encoder-decoder architecture and a rigid-body dynamics model to encode subtomograms into two separate low-dimensional latent spaces, realizing hierarchical modelling of compositional diversity and conformational dynamics. OPUS-TOMO adeptly resolved specie-level heterogeneity in template matching results and enabled sub-nanometer reconstructions for both Chlamydomonas reinhardtii ATP synthase dimer and Schizosaccharomyces pombe 80S ribosome, with an improvement of resolution up to 3 [A] compared to expert-annotated datasets. OPUS-TOMO also resolved in situ dynamics for these biomolecules by reconstructing functionally relevant structural continua. The software is available at https://github.com/alncat/opusTOMO.