Unsupervised SE(3) Disentanglement for in situ Macromolecular Morphology Identification from Cryo-Electron Tomography

TL;DR

This work presents an unsupervised SE(3) disentanglement framework (Harmony) with a novel multi-choice learning module to identify in situ macromolecular morphologies from cryo-ET subtomograms. By separating SE(3) transformations from morphology in a latent space and using a generator to recover template morphologies, the method overcomes limitations of EM-based approaches in capturing rare conformations. Across simulated and real subtomogram datasets, it achieves superior clustering and morphology recovery, discovering morphologies previously unidentified by existing pipelines. The approach reduces dependence on manual hyperparameters and is amenable to integration with downstream refinement steps, enabling high-resolution, automated exploration of macromolecular diversity in cellular contexts.

Abstract

Cryo-electron tomography (cryo-ET) provides direct 3D visualization of macromolecules inside the cell, enabling analysis of their in situ morphology. This morphology can be regarded as an SE(3)-invariant, denoised volumetric representation of subvolumes extracted from tomograms. Inferring morphology is therefore an inverse problem of estimating both a template morphology and its SE(3) transformation. Existing expectation-maximization based solution to this problem often misses rare but important morphologies and requires extensive manual hyperparameter tuning. Addressing this issue, we present a disentangled deep representation learning framework that separates SE(3) transformations from morphological content in the representation space. The framework includes a novel multi-choice learning module that enables this disentanglement for highly noisy cryo-ET data, and the learned morphological content is used to generate template morphologies. Experiments on simulated and real cryo-ET datasets demonstrate clear improvements over prior methods, including the discovery of previously unidentified macromolecular morphologies.

Figures (10)

• Figure 1: Schematic overview of our method. Input subtomograms and their augmentations are encoded into disentangled latent factors: transformation parameters ($\theta$) and morphology factors ($z$). Decoders reconstruct the input under multiple transformations, with a winner-takes-all Sum of Squared Error (SSE) loss selecting the best reconstruction. A similarity loss enforces consistency between augmented pairs, encouraging separation of transformation and morphology.
• Figure 2: Results on realistic simulated data (SNR $0.01$ and $30^\circ$ Missing Wedge Angle).
• Figure 3: Our method recognizes the morphology of membrane proteins and several membrane-bound enzyme complexes in thylakoid membrane region of Chlamydomonas reinhardtii. A. A central slab (slice across depth axis) of Chlamydomonas reinhardtii tomogram, where subtomograms are extracted from thylakoid membrane region. B. Schematic representation of membrane-bound protein complexes involved in photosynthesis in the thylakoid membrane. C. A few of the morphologies (initial 3D models) generated by RELION and DISCA. D. The UMAP visualization of the morphology latent factor by our method along with the morphology groups obtained with GMM (K=8). E. A few of the morphologies (decoded outputs of median of certain morphology classes) obtained by our method.
• Figure 4: The image shows why projection-image-based reconstruction methods (cryoDRGN powell2024learning and its variants) are not suitable for highly heterogeneous 3D structure identification.
• Figure 5: (a) The existing cryo-ET image processing pipeline and our method's positioning, (b)Schematic diagram of identifying refined macromolecular templates from 3D cellular cryo-ET tomogram