Capturing cell heterogeneity in representations of cell populations for image-based profiling using contrastive learning

Image-based cell profiling is a powerful tool that compares perturbed cell populations by measuring thousands of single-cell features and summarizing them into profiles, typically by averaging across cells. Although average profiling is commonly used, it fails to capture the heterogeneity within cell populations. We introduce CytoSummaryNet: a machine learning approach for summarizing cell populations that outperforms average profiling in predicting a compound's mechanism of action. CytoSummaryNet uses weakly supervised contrastive learning in a multiple-instance learning framework and provides an easier-to-apply method for aggregating single-cell feature data than previously published strategies. Interpretability analysis suggests that CytoSummaryNet achieves this by downweighting noisy cells (small mitotic cells or those with debris) and prioritizing less noisy cells (large uncrowded cells). Remarkably, CytoSummaryNet may also mitigate batch effects, even though this was not part of the training objective. Finally, the framework is designed to facilitate retraining, employing weak labels derived from perturbation replicates that are readily available in all cell profiling datasets. We show on a public dataset that CytoSummaryNet aggregated profiles result in a 68% increase in the mean average precision of mechanism of action retrieval compared to the commonly used average-aggregated profiles. ### Competing Interest Statement The Authors declare the following competing interests: S.S. and A.E.C. serve as scientific advisors for companies that use image-based profiling and Cell Painting (A.E.C: Recursion, SyzOnc, Quiver Bioscience; S.S.: Waypoint Bio, Dewpoint Therapeutics, Deepcell) and receive honoraria for occasional talks at pharmaceutical and biotechnology companies. R.v.D is an employee of CellVoyant. All other authors declare no competing interests.