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

Abstract

Revealing the underlying cell-state landscape from single-cell data requires overcoming the critical obstacles of batch integration, denoising, and dimensionality reduction. Here, we present CONCORD, a unified framework that simultaneously addresses these challenges within a single self-supervised model. At its core, CONCORD implements a unified probabilistic sampling strategy that corrects batch effects via dataset-aware sampling and enhances biological resolution through hard-negative sampling. Remarkably, using only a minimalist neural network with a single hidden layer and contrastive learning, CONCORD surpasses state-of-the-art performance without relying on deep architectures, auxiliary losses, or external supervision. It seamlessly integrates data across batches, technologies, and even species to generate high-resolution cell atlases. The resulting latent representations are denoised and biologically meaningful--capturing gene co-expression programs, revealing detailed lineage trajectories, and preserving both local geometric relationships and global topological structures. We demonstrate CONCORDs broad applicability across diverse datasets, establishing it as a general-purpose framework for learning unified, high-fidelity representations of cellular identity and dynamics.