Unsupervised SE(3) Disentanglement for in situ Macromolecular Morphology Identification from Cryo-Electron Tomography
Mostofa Rafid Uddin, Mahek Vora, Qifeng Wu, Muyuan Chen, Min Xu
TL;DR
This paper introduces a deep learning framework for unsupervised disentanglement of SE(3) transformations from macromolecular morphology in cryo-electron tomography, improving identification of diverse structures.
Contribution
It presents a novel multi-choice learning approach for disentangling transformations from morphology in noisy cryo-ET data, enabling discovery of new macromolecular structures.
Findings
Improved morphology identification over prior methods
Discovery of previously unidentified macromolecular morphologies
Effective disentanglement of transformations from morphology
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…
Peer Reviews
Decision·Submitted to ICLR 2026
- The classification of subtomograms is a very important, yet difficult problem. Instead of optimizing subtomogram poses and classes at the same time (as in many traditional methods), the paper tried to directly disentangle SE(3) transformation and classify/cluster in the latent space. - The introduction of the MCL module is interesting and greatly improved the performance.
- The presentation of the results can be much improved. The slice-by-slice visualization is hard to see or interpret and I would suggest the authors to only use the isosurface visualization of the 3D density maps. With the current presentations, it is extremely difficult to evaluate the performance of the propose method. I would increase my score if the presentation is improved and demonstrates a reasonable performance of the proposed method. - The paper also lacks novelty from the method persp
1. Important problem domain: fully automated, unsupervised morphology identification from in-cell tomograms. 2. Empirical robustness: large ARI/AUC-FSC gains under extreme noise conditions. 3. Elegant MCL integration: computationally light mechanism improving convergence. 4. Comprehensive evaluation: simulated + real datasets; thorough appendix for reproducibility.
1. Limited novelty – no new theoretical insight beyond adapting Harmony + MCL. The proposed model inherits almost the entire architecture and loss design from Harmony (CVPR 2022) and Multi-Choice Learning (Guzman-Rivera 2012; Kohl 2018). The paper provides no new theoretical analysis of SE(3) disentanglement or identifiable latent factors. Equations (4)–(7) merely restate known forms with SE(3) transforms substituted for generic transformations. There is no formal proof that the new objective ac
1. The paper constructs multiple simulated subtomogram datasets of macromolecular mixtures that should be valuable for subsequent research. 2. The work reveals previously unidentified macromolecular morphologies, underscoring its practical significance. 3. Compared with traditional methods, the proposed approach demonstrates clear performance gains and addresses a decade-long unsolved problem in structural biology.
1. Limited novelty. The method comprises two key components: unsupervised SE(3) disentanglement and multiple-choice learning. The reconstruction and regularization loss used for unsupervised SE(3) disentanglement are common loss. Moreover, Shekarforoush et al. have previously applied the MCL idea to estimate SO(3), so I suppose using MCL here to estimate SE(3) templates does not show enough novelty. 2. Missing baselines. Beyond the classical maximum-likelihood method RELION, this paper does not
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdvanced Electron Microscopy Techniques and Applications · Electron and X-Ray Spectroscopy Techniques · Enzyme Structure and Function