CryoSiam: self-supervised representation learning for automated analysis of cryo-electron tomograms

Abstract

Cryo-electron tomography (cryo-ET) enables visualization of macromolecular complexes in their native cellular context, but interpretation remains challenging due to high noise levels, missing information, and lack of ground-truth data. Here, we present CryoSiam (CRYO-electron tomography SIAMese networks), an open-source framework for self-supervised representation learning in cryo-ET. CryoSiam learns hierarchical representations of tomographic data spanning both voxel-level and subtomogram-level information. To train CryoSiam, we generated CryoETSim (CRYO-Electron Tomography SIMulated), a synthetic dataset that systematically models defocus variation, sample thickness, and molecular crowding. CryoSiam trained models transfer directly to experimental data without fine-tuning and support key aspects of cryo-ET data analysis, including tomogram denoising, segmentation of subcellular structures, and macromolecular detection and identification across both prokaryotic and eukaryotic systems. Publicly available pretrained models and the CryoETSim dataset provide a foundation for scalable and automated cryo-ET analysis.