CryoSiam

CryoSiam is a self-supervised deep learning framework for the automated analysis of cryo-electron tomograms (cryo-ET), developed at the European Molecular Biology Laboratory (EMBL) by researchers spanning the Mahamid, Kreshuk, and Zaugg groups and posted as a bioRxiv preprint in November 2025. Cryo-ET captures the molecular architecture of cells in near-native, frozen-hydrated states at nanometer resolution, but the resulting tomograms are notoriously difficult to interpret: they suffer from extreme noise, a missing-wedge of information from limited tilt angles, and dense molecular crowding. Extracting biological meaning has historically required laborious manual annotation or task-specific supervised models trained on scarce labeled data.

CryoSiam reframes this problem around representation learning. Rather than training a separate supervised model for each downstream task, it learns general-purpose features directly from tomographic data using a Siamese self-supervised objective, then transfers those features to multiple tasks without fine-tuning. A central contribution is that the network is trained entirely on synthetic tomograms generated by a physics-aware simulator that models defocus variation, sample thickness, and molecular crowding, sidestepping the need for large annotated experimental datasets while still transferring to real data.

The framework learns at two complementary scales: a dense, voxel-level representation that captures local structural context across the full tomogram, and a subtomogram-level representation that encodes individual macromolecular complexes. Together these enable a single pretrained backbone to support denoising, semantic and instance segmentation, and macromolecular detection across both prokaryotic and eukaryotic systems.

#Key Features

• Self-supervised Siamese learning: Built on a SimSiam-style objective that learns representations by maximizing agreement between augmented views, removing the dependence on labeled tomograms during pretraining.
• Synthetic-to-real transfer: Trained on the CryoETSim synthetic dataset modeling defocus, sample thickness, and molecular crowding, the learned features transfer zero-shot to experimental tomograms without fine-tuning.
• Hierarchical representations: Produces both dense voxel-level features for whole-tomogram tasks and subtomogram-level embeddings for individual macromolecular complexes.
• Multi-task from one backbone: A shared pretrained encoder supports denoising, semantic and instance segmentation, and particle detection rather than requiring a bespoke model per task.
• Released pretrained checkpoints: Ships ready-to-use models including denoising (with an AreTomo3-aware variant), semantic and instance segmentation, ribosome segmentation, lamella detection, particle candidate prediction, and subtomogram embedding models.

#Technical Details

CryoSiam couples a convolutional backbone with a Siamese self-supervised framework operating on two levels. The dense pathway uses a SimSiam-pretrained backbone to produce voxel-wise features that initialize semantic segmentation heads, while the subtomogram pathway learns per-particle embeddings (offered in convex-hull masking, no-masking, and strict-masking variants for different downstream uses). Pretraining relies exclusively on synthetic weighted-back-projection tomograms from the CryoETSim simulator, with a denoising variant additionally trained on reconstructions from both IMOD and AreTomo3 pipelines to improve robustness across reconstruction software. The authors demonstrate denoising, multi-class subcellular segmentation, ribosome and lamella detection, and instance separation of macromolecular complexes on experimental data from diverse prokaryotic and eukaryotic samples, with all evaluation performed without task-specific fine-tuning of the pretrained representations.

#Applications

CryoSiam targets structural and cell biologists working with cryo-ET who need to process tomograms at scale without building bespoke annotation pipelines. Practical uses include denoising raw tomograms to improve interpretability, segmenting organelles and subcellular structures, restricting analysis to lamella regions in FIB-milled samples, and detecting and embedding macromolecular complexes such as ribosomes to support particle picking and subtomogram averaging workflows. Because the released checkpoints run on experimental data out of the box, the framework lowers the barrier for labs that lack large annotated training sets.

#Impact

By showing that a single self-supervised model trained purely on synthetic data can transfer zero-shot across denoising, segmentation, and detection tasks, CryoSiam advances cryo-ET toward general-purpose, label-efficient analysis and reduces reliance on the manual annotation that has bottlenecked the field. The code is openly released under GPL-3.0 with documentation and pretrained weights distributed on Hugging Face; notably, the weights repository carries an MIT license that differs from the GPL-3.0 code license, a discrepancy adopters should verify before redistribution. As a 2025 preprint the results await peer review, and performance on samples outside the simulated training distribution remains to be characterized, but the multi-task, no-fine-tuning design represents a meaningful step for accessible tomogram analysis.