Simple and effective embedding model for single-cell biology built from ChatGPT

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Large-scale gene-expression data is crucial for pretraining models to understand gene and cellular functions.
• Current models require extensive data curation and training, posing significant challenges.

Purpose of the Study:

• To explore a simpler alternative to traditional gene-expression-based models using literature-derived embeddings.
• To develop a novel method, GenePT, for encoding single-cell biological knowledge.

Main Methods:

• Utilized GPT-3.5 to generate gene embeddings from literature text descriptions.
• Created single-cell embeddings by averaging weighted gene embeddings based on expression levels.
• Generated sentence embeddings for cells using ordered gene names by expression level.

Main Results:

• GenePT achieved comparable or superior performance to traditional models on gene-property and cell-type classification tasks.
• Demonstrated the effectiveness of large language model embeddings in capturing single-cell biological information.
• Showcased a simplified approach to single-cell data analysis.

Conclusions:

• Leveraging large language model embeddings from scientific literature offers a computationally efficient and effective method for single-cell analysis.
• GenePT provides a powerful new tool for researchers in genomics and computational biology.
• This approach bypasses the need for extensive gene-expression data curation and model training.