Sorting of viable unlabeled cells based on deep representations links morphology to multiomics.

Abstract

Phenotyping of single cells has dramatically lagged advances in molecular characterization and remains a manual, subjective, and destructive process. We introduce Computational Sorting and Mapping of Single cells (COSMOS), a platform for phenotyping and enrichment of cells based on deep learning interpretation of high-content morphology data in realtime. By training models on an atlas of >1.5 billion images, we demonstrate enrichment of unlabeled cells up to 33,000-fold. By applying COSMOS to tissue samples we identify malignant cells with similar accuracy to molecular approaches while enriching cells for functional evaluation. We show high-dimensional embedding vectors of morphology without any need for complex sample pre-processing, gating, or bioinformatics capabilities, which enables discovery of cellular phenotypes and integration of morphology into single-cell multimodal analyses as a new “omic”. Sorting intact cells based on deep morphological representations could further provide a powerful tool for tying together data across different modalities.