Artificial Intelligence for Multi-omics Breast Cancer Subtype Prediction and Discovery: A Systematic Review

Artificial intelligence models integrate molecular omics to reproduce breast cancer subtype labels or discover patient groups, but heterogeneous endpoints, cohorts, and validation hinder interpretation. We searched PubMed/MEDLINE, Europe PMC, and OpenAlex through 2026 September 8, supplemented by citation searching and targeted retrieval through Google Scholar and IEEE Xplore. Eligible studies combined at least 2 molecular omics in one machine-learning model for breast cancer subtype prediction or discovery. Findings were synthesized by endpoint without meta-analysis. Of 705 exported records, 301 duplicates were removed, 404 unique records were screened, and 70 studies were included. Fifty-seven addressed intrinsic subtype or Prediction Analysis of Microarray 50 (PAM50) prediction, and 11, subtype discovery or recovery. Sixty-two publications used The Cancer Genome Atlas Breast Invasive Carcinoma, and 9 used the Molecular Taxonomy of Breast Cancer International Consortium, demonstrating extensive reuse of limited patient populations. Among 59 supervised prediction studies, only 1 independently evaluated a complete multi-omics model in another cohort, and none reported prospective or nested full-pipeline validation. Calibration, uncertainty reporting, and external validation were uncommon, although 34 studies provided code or software. Among 11 discovery studies, 4 assessed resampling stability and 3 attempted independent subtype recovery. All 57 PAM50 studies used expression, and none reported excluding PAM50 classifier genes or close proxies. Thirty-nine included a matched expression-only baseline, but only one tested the complete model independently. Risk of bias was high in 48/59 prediction studies. Graph, hypergraph, attention, and uncertainty-aware models are increasingly used, but evidence of transportability and calibrated clinical value remains limited.