LingoCell interprets gene perturbations through natural language and tissue niches

Abstract

Single-cell foundation models are general tools for resolving cell identities and disease states, but struggle to interpret changes in expression magnitude in biological language, to transfer to sparse spatial measurements, and to explain cellular niches. We introduce LingoCell, a foundation model connecting single-cell and spatial transcriptomics. It aligns cell representations with language while keeping gene expression editable, supporting cell annotation, cross-batch integration, description generation, and disease-language readouts of single-gene and combinatorial in silico perturbations. Extended to tissues, it annotates cells without single-cell references, learns identity-conditioned niche embeddings, and performs spatial in silico perturbation by replacing and editing neighbor expression. In lung adenocarcinoma, LingoCell revealed synergistic disease responses to jointly weakening alveolar differentiation programs and strengthening malignant programs. In colorectal cancer, it revealed different programs conveyed by the same fibroblast state to T cells, macrophages, and tumor cells. LingoCell provides a unified framework for prioritizing target combinations and stromal intervention programs.