Nature Biotechnology, Published online: 05 January 2026; doi:10.1038/s41587-025-02978-1

Comprehensively resolving the cell state landscape requires integrating single-cell omics data from diverse studies. We developed CONCORD, a contrastive learning framework that leverages principled mini-batch sampling to learn denoised, batch-integrated and high-resolution representations of cells, capturing intricate structures such as differentiation trajectories and cell-cycle loops across numerous biological contexts.