Cell2Sentence: Teaching Large Language Models the Language of Biology.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Single-cell transcriptomics generates high-dimensional gene expression data.
• Integrating advanced computational methods like natural language processing (NLP) can unlock new insights.
• Existing NLP models require adaptation for biological data.

Purpose of the Study:

• To introduce Cell2Sentence (C2S), a novel method for adapting large language models (LLMs) to single-cell transcriptomics.
• To demonstrate the utility of C2S in enabling LLMs to perform biological tasks.
• To bridge the gap between NLP and single-cell biology.

Main Methods:

• Transforming gene expression data into "cell sentences."
• Fine-tuning pre-trained LLMs (e.g., GPT-2) using cell sentences.
• Evaluating the fine-tuned models on tasks like cell generation and cell type annotation.

Main Results:

• Fine-tuned GPT-2 models can generate biologically valid cells based on cell type inputs.
• The models accurately predict cell types from cell sentences.
• LLMs fine-tuned with C2S acquire a significant understanding of single-cell biology.

Conclusions:

• Cell2Sentence (C2S) provides a flexible framework for integrating NLP with transcriptomics.
• C2S enables LLMs to perform complex biological tasks, including cell generation and annotation.
• This approach leverages existing NLP models and libraries for broad biological applications.