RAGCell: Retrieval-Augmented Generation as Supervision for Versatile Single-cell Analysis.

RAGCell achieves high-performance single-cell analysis at less than one-tenth the cost of pretraining-based models by integrating retrieval-augmented generation techniques.

• Why it matters: Efficient and versatile single-cell analysis tools are essential for advancing biological research, but current models face high resource demands and struggle with the heterogeneity between cellular and textual data.
• What they did: The approach involves constructing cell- and feature-level knowledge databases using large language models, which serve as supervision signals, and aligning cellular representations with text embeddings to transfer knowledge across spaces.
• The result: Extensive experiments on six tasks show RAGCell outperforms state-of-the-art models, enabling cost-effective, accurate analysis and bridging the gap between textual and cellular data representations.