Self-supervised vision transformers accurately decode cellular state heterogeneity

Abstract

Abstract Characterising cellular phenotypic heterogeneity is essential to understand the relationship between the molecular and morphological determinants of cellular state. Here we report that publicly available self-supervised vision transformers (ss-ViTs) accurately elucidate phenotypic stem cell heterogeneity out-of-the-box. Moreover, we introduce scDINO, an adapted ss-ViT trained on five-channel automated microscopy data, attaining excellent performance in delineating peripheral blood immune cell identity. Thus, ss-ViTs represent a leap forward in the unsupervised analysis of phenotypic heterogeneity.