Contrastive learning enables rapid mapping to multimodal single-cell atlas of multimillion scale

Abstract

Single-cell datasets continue to grow in size, posing computational challenges for dealing with expanded scale, extended modality and inevitable batch effects. Deep learning-based approaches have recently emerged to address these points by deriving nonlinear cell embeddings. Here we present contrastive learning of cell representations, Concerto, which leverages a self-supervised distillation framework to model multimodal single-cell atlases. Simply by discriminating each cell from the others, Concerto can be adapted to various downstream tasks such as automatic cell type classification, data integration and especially reference mapping. Unlike current mainstream packages, Concerto’s contrastive setting well supports operating on all genes to preserve biological variations. Concerto can flexibly generalize to multiomics to obtain unified cell representations. Benchmarking on both simulated and real datasets, Concerto substantially outperforms competing methods. By mapping to a comprehensive reference, Concerto recapitulates differential immune responses and discovers disease-specific cell states in patients with COVID-19. Concerto is easily parallelizable and efficiently scalable to build a 10-million-cell reference within 1.5 h and query 10,000 cells within 8 s. Overall, Concerto will facilitate biomedical research by enabling iteratively constructing single-cell reference atlases and rapidly mapping novel dataset against them to transfer relevant cell annotations.