Revealing a coherent cell state landscape across single cell datasets with CONCORD

Resolving the intricate structure of the cellular state landscape from single-cell RNA sequencing (scRNAseq) experiments remains an outstanding challenge, compounded by technical noise and systematic discrepancies—often referred to as batch effects—across experimental systems and replicate. To address this, we introduce CONCORD (COntrastive learNing for Cross-dOmain Reconciliation and Discovery), a self-supervised contrastive learning framework designed for robust dimensionality reduction and data integration in single-cell analysis. The core innovation of CONCORD lies in its probabilistic, dataset– and neighborhood-aware sampling strategy, which enhances contrastive learning by simultaneously improving the resolution of cell states and mitigating batch artifacts. Operated in a fully unsupervised manner, CONCORD generates denoised cell encodings that faithfully preserve key biological structures, from fine-grained distinctions among closely related cell states to large-scale topological organizations. The resulting high-resolution cell atlas seamlessly integrates data across experimental batches, technologies, and species. Additionally, CONCORD’s latent space capture biologically meaningful gene programs, enabling the exploration of regulatory mechanisms underlying cell state transitions and subpopulation heterogeneity. We demonstrate the utility of CONCORD on a range of topological structures and biological contexts, underscoring its potential to extract meaningful insights from both existing and future single-cell datasets.