Improving Cancer Driver Gene Prediction using Biological knowledge-guided Prompts for LLM

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Accurate identification of cancer driver genes is essential for cancer research and therapeutic development.
• Current methods face challenges due to limited sample sizes and poor gene characteristic differentiation.
• Large language models (LLMs) offer potential for enhanced gene representation due to their extensive biomedical knowledge.

Purpose of the Study:

• To develop a novel framework, Bioprompt, for improved cancer driver gene identification.
• To leverage biological knowledge-guided prompts for LLMs to enhance gene semantic representations.
• To integrate LLM-derived features with existing gene features for robust driver gene prediction.

Main Methods:

• Utilized Gene Ontology (GO) information to guide LLM prompt design for generating gene semantic representations.
• Employed a text encoder and variational autoencoder to encode semantic representations.
• Integrated LLM-derived features with existing gene features using contrastive learning.
• Used a logistic regression model for final driver gene identification.

Main Results:

• Bioprompt demonstrated improved performance on pancancer and 15 individual cancer datasets compared to existing models.
• Ablation studies confirmed the significance of GO-guided prompts in generating valuable semantic information.
• The method showed strong generalization ability on independent test sets.
• Analysis of LLM-generated gene functional vocabulary provided insights into cancer-related gene functions and mechanisms.

Conclusions:

• Bioprompt effectively enhances cancer driver gene identification by integrating LLM-derived semantic features guided by biological knowledge.
• The framework offers a promising approach for advancing cancer research and therapeutic strategies.
• The study highlights the potential of LLMs in bioinformatics and the interpretability of generated gene functions.