Machine learning for predicting and overcoming tumor radioresistance: current approaches

Introduction . Intertumoral differences in biological response to radiation remain a major obstacle to the full personalization of radiotherapy. Radioresistance is driven by multiple factors, including DNA repair, hypoxia, metabolic reprogramming, and cancer stem cells, which renders the search for a single biomarker ineffective. Objective . To analyze current machine learning (ML) approaches for predicting radiosensitivity and overcoming tumor radioresistance, and to assess their clinical potential and methodological limitations. Methods . A narrative review of recent peer-reviewed publications in PubMed / MEDLINE focusing on ML applications in radiation oncology was conducted. Studies on radiomics, radiogenomics, multimodal data integration, adaptive planning, and radiosensitizer discovery were analyzed. Results . Examples of transcriptomic (gradient boosting models), metabolic (integration with flux modeling), and radiomic models for predicting treatment response are summarized. Models with external validation and a biologically interpretable link to resistance mechanisms demonstrate the highest value. Key challenges include small sample sizes, heterogeneous data acquisition protocols, and the predominance of retrospective designs. Conclusion . The most feasible application scenarios for ML are narrow assistive tasks: risk stratification, localization of resistant subvolumes, support for plan adaptation, and prioritization of candidates for radiosensitization. Clinical deployment of models influencing dose or fractionation requires mandatory prospective evaluation and assessment of impact on outcomes (local control and toxicity).