Cancer gene identification through integrating causal prompting large language model with omics data-driven causal...

Area of Science:

• Genomics and Computational Biology
• Cancer Research
• Artificial Intelligence in Medicine

Background:

• Identifying cancer-related genes is crucial for understanding cancer mechanisms and developing targeted therapies.
• Traditional methods for gene identification often yield biased and poorly interpretable results due to confounding factors and methodological limitations.

Purpose of the Study:

• To develop a novel framework, Integrative Causal Gene Identification (ICGI), for identifying cancer genes across multiple omics domains.
• To leverage large language models (LLMs) and causal feature selection for improved accuracy and interpretability in cancer gene identification.

Main Methods:

• Development of the ICGI framework integrating LLMs with causality contextual cues and data-driven causal feature selection.
• Application of the causal feature selection module to transcriptomic data from six cancer types in The Cancer Genome Atlas (TCGA).
• Comparison of ICGI's performance against state-of-the-art methods for identifying cancer-distinguishing genes.

Main Results:

• The ICGI framework demonstrates superior capability in identifying cancer genes that effectively distinguish between cancerous and normal samples compared to existing methods.
• LLMs show potential in uncovering cancer genes and understanding disease mechanisms, particularly at the genomic level.
• Current LLMs may not fully capture information across all omics levels, indicating areas for future improvement.

Conclusions:

• The proposed ICGI framework offers a powerful and interpretable approach for multi-omics cancer gene identification.
• The study highlights the potential of integrating LLMs with causal inference for advancing cancer research.
• An online service platform has been developed to provide accessible gene-cancer association insights based on causal learning methods.