Illuminating cell states by a comprehensive and interpretable single cell foundation model

Advances in single-cell sequencing have enabled AI-driven foundation models with powerful data representation. However, their practical use is limited by real-world data sparsity, heterogeneity, and poor interpretability. To overcome these, we introduce CellVQ. To enhance generalizability, we incorporate a large-scale single-cell dataset comprising 68 million cells, model parameters totaling 500 million, and challenging pretraining tasks. Notably, we introduce a Single-Cell Discretization (SCD) module that effectively represents cell embeddings, addressing data heterogeneity. For improved interpretability, the SCD module transforms high-dimensional and sparse single-cell data into a “cell code,” facilitating recognition and analysis. Additionally, we also present CellVQ-Graph, a plug-and-play tool that integrates CellVQ’s features with multimodal data (genes, cell communication, annotations) to build a knowledge graph for biological discovery. Extensively evaluated, CellVQ outperforms strong baselines in all downstream tasks, and also uncovered intriguing biological phenomena with compelling explanations. CellVQ aspires to serve as a truly applicable and generalizable AI tool for the cell biology community. Current single-cell models face data sparsity and interpretability limits. The authors introduce CellVQ to improve generalizability with a cell-state discretisation module, and CellVQ-Graph to integrate genes, interactions, and annotations into a knowledge graph for interpretability.