Artificial intelligence in cancer transcriptomics: methods, applications in oncology, and limitations

Abstract

Background. Transcriptomic data show which genes are active in cells and tissues. Machine-learning and deep-learning methods use these data to classify samples, integrate different datasets, and predict biological properties and treatment response. In oncology, a major challenge is transferring patterns identified in experimental models to tumor samples from patients. Objective. To critically synthesize current artificial-intelligence methods for transcriptomic analysis, with emphasis on oncology applications, transfer of models to patient data, and requirements for clinical validation. Materials and Methods. We performed a critical narrative review of primary studies and methodological papers and additionally described the authors' publication sample for 2021–2026. The literature review was updated through August 13, 2026. Results. The most developed areas include single-cell data analysis, spatially resolved tumor diagnostics, and drug-response prediction. Independent comparative evaluations show that increasing model size and complexity does not guarantee better performance than simpler methods. Clinical translation requires accounting for cellular composition and technical differences between measurement batches, aligning experimental and clinical response measures, preventing information leakage, and performing external validation. Conclusion. Artificial intelligence should be viewed as a data-analysis tool rather than a substitute for experimental or clinical validation. Its value depends on data quality, correct task definition, and the ability to maintain performance in independent cohorts.