Abstract:Structural heterogeneity due to the dynamic nature of macromolecules in situ presents a significant challenge to structural determination by cryo-electron tomography (cryo-ET). In this paper, we present OPUS-TOMO, a deep learning framework for analyzing structural heterogeneity in cryo-ET data. The method adopts a convolutional Encoder-Decoder architecture that adeptly maps real-space subtomograms onto a smooth low-dimensional latent space, which captures the complete landscape of compositional and conformational variations of macromolecules in cryo-ET data. OPUS-TOMO also incorporates algorithms, including a per-particle 3D CTF model and a pose correction network, specifically for handling cryo-ET data. Applications of OPUS-TOMO to multiple real cryo-ET datasets confirm the outstanding capacities of the new method in characterizing structural heterogeneity. The software is available at https://github.com/alncat/opusTOMO.

OPUS-TOMO: Deep Learning Framework for Structural Heterogeneity Analysis in Cryo-electron Tomography