Large language models in personalized oncology

Background . Large language models are considered a promising tool for personalized oncology; however, their role in the analysis of molecular data and treatment selection should be clearly distinguished from the functions of specialized predictive models. Objective . To systematize the potential applications of large language models in personalized oncology, primarily for the interpretation of clinical, genomic, and transcriptomic data and of predictions of tumor chemo- and radiosensitivity. Methods . We performed a narrative review of publications identified in PubMed and Google Scholar, with the search updated through July 1, 2026. Original studies and reviews on medical LLM applications, foundation models for genomic and transcriptomic data, clinical decision support, and prediction of anticancer treatment sensitivity were considered. Results . The most justified approach is to use a language model as an interface layer: a specialized machine-learning or deep-learning model calculates the prediction, while the language model generates an explanation, compares the result with clinical guidelines, and indicates limitations. The main risks include unreliable responses, insufficient reproducibility, weak external validation, and legal constraints. Conclusion . A clinically promising architecture separates computation from explanation: the language model does not prescribe treatment autonomously but links molecular data, a validated predictive model, and the physician.