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

Accurately identifying cancer driver genes is crucial for understanding cancer and developing therapies. However, challenges such as limited sample sizes and insufficient differentiation of gene characteristics hinder current approaches. Large language models (LLMs) possess a wealth of biomedical knowledge offering the potential to enhance gene representations. However, effectively integrating LLMs into cancer driver gene prediction requires careful prompt engineering and multiomics fusion strategies. This study introduces a novel framework, called Bioprompt, that leverages biological knowledge-guided prompts for LLM to improve cancer driver gene identification. We utilize Gene Ontology(GO) information to guide the design of prompts for LLMs, enabling them to generate relevant gene semantic representations. Semantic representations are then encoded through a text encoder and a variational autoencoder. Subsequently, we integrate these LLM-derived gene semantic features with existing model-generated gene features via contrastive learning. Finally, a logistic regression model is employed to combine these two feature types for the identification of cancer driver genes. Validation on pancancer and 15 individual cancer datasets shows improvement over existing models. Ablation studies confirm the critical role of GO-guided prompts in generating valuable gene semantic information. Independent test set tests have demonstrated the generalization ability of our method. Furthermore, by analyzing the gene functional vocabulary generated by the LLM, we gain valuable insights into cancer related gene functions and new perspectives on cancer mechanisms.