Multimodal learning enables chat-based exploration of single-cell data

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Single-cell sequencing generates vast amounts of detailed biological data.
• Interpreting this complex gene expression data remains a significant challenge for researchers.

Purpose of the Study:

• To develop an AI-powered tool, CellWhisperer, for intuitive, chat-based interrogation of single-cell RNA sequencing data.
• To enhance the accessibility and interpretability of single-cell transcriptomic information.

Main Methods:

• Constructed a multimodal embedding of transcriptomes and textual annotations using contrastive learning on 1 million RNA sequencing profiles.
• Integrated this embedding with a large language model to enable natural-language querying of gene expression data.
• Developed a chat interface integrated with the CELLxGENE browser for interactive data exploration.

Main Results:

• CellWhisperer demonstrates strong performance in zero-shot prediction of cell types and biological annotations.
• The tool facilitates biological discovery, as shown in a meta-analysis of human embryonic development.
• Users can interactively explore gene expression through a combined graphical and chat interface.

Conclusions:

• CellWhisperer effectively leverages large-scale data repositories to bridge transcriptomes and text.
• This AI model significantly enhances the interactive exploration of single-cell RNA sequencing data using natural-language chats.