OPUS-DSD2: Disentangling Dynamics and Compositional Heterogeneity for Cryo-EM/ET

Abstract

Cryo-electron microscopy and tomography (cryo-EM/ET) capture structural heterogeneities in macromolecules, ranging from dynamic motions to compositional changes-- key to understanding biological mechanisms. While the deep learning framework, OPUS-DSD, advanced heterogeneity analysis for cryo-EM, it conflates different types of heterogeneities, and remains incompatible with cryo-ET which can elucidate macromolecular functions in their native cellular environments. Here, we present OPUS-DSD2, a unified framework for disentangling structural heterogeneity in both cryo-EM and cryo-ET data. OPUS-DSD2 augments a 3D convolutional neural network with a multi-layer perceptron based rigid-body dynamics model, advocating the separation of subunit level rigid-body dynamics from other heterogeneities. Tests on real datasets demonstrate that OPUS-DSD2 effectively captures large-scale subunit motions while isolating spatially localized structural variations into distinct principal components of the composition latent space. Critically, OPUS-DSD2 enables direct analysis of noisy cryo-ET template-matching results, bypassing labor-intensive subtomogram classification and averaging and unlocking high-throughput visual proteomics. Backward-compatible with OPUS-DSD, OPUS-DSD2 is available at https://github.com/alncat/opusDSD.