Identifying genes causally linked to cancer from a multi-omics perspective is essential for understanding the mechanisms of cancer and improving therapeutic strategies. Traditional statistical and machine-learning methods that rely on generalized correlation approaches to identify cancer genes often produce redundant, biased predictions with limited interpretability, largely due to overlooking confounding factors, selection biases, and the nonlinear activation function in neural networks. In this study, we introduce a novel framework for identifying cancer genes across multiple omics domains, named ICGI (Integrative Causal Gene Identification), which leverages a large language model (LLM) prompted with causality contextual cues and prompts, in conjunction with data-driven causal feature selection. This approach demonstrates the effectiveness and potential of LLMs in uncovering cancer genes and comprehending disease mechanisms, particularly at the genomic level. However, our findings also highlight that current LLMs may not capture comprehensive information across all omics levels. By applying the proposed causal feature selection module to transcriptomic datasets from six cancer types in The Cancer Genome Atlas and comparing its performance with state-of-the-art methods, it demonstrates superior capability in identifying cancer genes that distinguish between cancerous and normal samples. Additionally, we have developed an online service platform that allows users to input a gene of interest and a specific cancer type. The platform provides automated results indicating whether the gene plays a significant role in cancer, along with clear and accessible explanations. Moreover, the platform summarizes the inference outcomes obtained from data-driven causal learning methods.
