Artificial intelligence-assisted Mendelian randomization: a methodological framework for causal inference with applications in oncology

Mendelian randomization (MR) has become an increasingly important approach for causal inference in genomics and epidemiology, enabling investigation of causal relationships between exposures and health outcomes using genetic instrumental variables. In parallel, recent advances in artificial intelligence (AI), including machine learning (ML), deep learning (DL), and large language models (LLMs), have transformed biomedical research by facilitating the analysis of large-scale and high-dimensional datasets. To critically examine the emerging integration of AI and MR, highlighting current applications, methodological opportunities, limitations, and future directions. This narrative review was conducted according to the principles of the Scale for the Assessment of Narrative Review Articles (SANRA). Relevant studies were identified through literature searches in PubMed, Scopus, and Web of Science and were selected to illustrate methodological developments, practical applications, and emerging AI-assisted MR frameworks. AI contributes to distinct stages of the MR workflow. Before MR, AI may support phenotype construction, exposure discovery, and biological target prioritization. Within MR, AI may assist biologically informed instrument prioritization and nonlinear causal estimation while remaining fully dependent on the core instrumental-variable assumptions. Around MR, AI enhances workflow automation, data retrieval, reporting, and evidence synthesis without modifying causal inference itself. Applications are particularly relevant in oncology, where AI-assisted MR may facilitate biomarker discovery and therapeutic target prioritization. Emerging approaches, including Deep Mendelian Randomization (DeepMR), foundation models, and LLM-based analytical agents, represent promising developments but require rigorous methodological validation. Nevertheless, AI cannot overcome the fundamental assumptions of MR, and robust causal inference continues to depend on biological plausibility, instrument validity, and transparent analytical design. AI has the potential to substantially augment MR by improving efficiency, scalability, and data integration. Nevertheless, the validity of causal inference continues to depend on rigorous methodological design, biological plausibility, and adherence to established MR principles. Future progress will likely rely on transparent, explainable, and human-supervised AI frameworks integrated within robust causal inference methodologies.