Cell2Text: Multimodal LLM for Generating Textual Descriptions from Single-Cell RNA-Seq Profiles

Abstract

Single-cell RNA sequencing has revolutionized biological research by enabling gene expression measurement at cellular resolution, revealing diverse cell types, states, and disease contexts. Recent single-cell foundation models can learn generalizable representations directly from expression data, improving downstream classification and clustering tasks. However, such models typically rely on fixed label spaces that limit their ability to express cellular diversity.

This thesis presents Cell2Text, a multimodal generative framework that transforms single-cell transcriptomic profiles into structured natural language descriptions. By integrating pretrained single-cell encoders with large language models through learnable projection modules, Cell2Text generates coherent summaries describing cellular identity, tissue of origin, disease relevance, and biological pathway activity. Experimental results show that Cell2Text achieves higher accuracy than baseline models, maintains strong ontological consistency through PageRank-based similarity metrics, and produces semantically faithful text outputs. Overall, the proposed approach highlights the potential of combining biological and linguistic representations for scalable and informative single-cell characterization.