BPS2026 – Locating macromolecules within crowded cell environments with deep representation learning

Studying macromolecules in their native cellular environment is crucial for understanding molecular function. Cryogenic electron microscopy (cryo-EM) images capture complexes directly inside frozen-hydrated cells and tissues and retain high-resolution features that are absent from tomograms with the benefit of faster data collection and processing. However, the crowded cellular environment and low signal-to-noise ratio in cryo-EM images make identifying macromolecules challenging. Supervised machine learning methods for object detection have been effectively applied to the problem of particle picking in single-particle cryo-EM. However, we observed that when trained on previously identified ribosomes in images of untilted cellular lamellae, these methods fail to detect particles of interest, likely due to the effective loss of low-resolution contrast in crowded cell environments. Scaling supervised learning methods to new data faces the additional bottleneck of the high cost of annotation. To address these limitations, we developed a self-supervised, contrastive learning method for particle localization in situ. Our approach first pretrains a deep neural encoder on large, unannotated data sets of patches of cellular cryo-EM images to learn relevant structural features. This pretrained encoder is then fine-tuned with a minimal number of labels to classify image patches and identify particles of interest. We demonstrate that our self-supervised model performs on-par with, and in some cases outperforms, supervised methods in distinguishing ribosomes from background in in situ data. Crucially, our model generalizes better to unseen data sets and offers significantly higher throughput relative to 2D template matching (2DTM). By enabling fast, template-free particle localization directly within cellular images, this approach advances our ability to explore complex biological systems.