DR-GEM enables self-supervised machine learning for single-cell embeddings and annotations

Abstract Dimensionality reduction and clustering are instrumental to single-cell and spatial genomics data analysis. Here we show that existing methods tend to fit oversampled classes and miss more unique signals and patterns, such as rare cell types and states. We find that the problem in such cases is not primarily data scarcity per se, but the model equating abundance with importance. Addressing this, we present “Distributionally Robust and latent Group-AwarE consensus Machine learning” or DR-GEM, a self-supervised meta-algorithm that uses the reconstruction error to reorient the model attention to patterns it initially missed and applies balanced consensus learning to increase robustness and filter low-quality data. We apply DR-GEM to synthetic and real-world single-cell, spatial transcriptomics, and Perturb-seq datasets and demonstrate that it outperforms existing methods in obtaining reliable embeddings, recovering rare cell types, filtering noise, mitigating cell-level and patient-level underrepresentation, and uncovering held-out genetic and spatial features. We thus surface and address an underappreciated problem and present a self-supervised framework to help support AI-based and data-driven discoveries.