Regulators of homologous recombination deficiency identified by machine learning using somatic multi-omics data

Homologous recombination deficiency (HRD) is a critical biomarker for guiding targeted therapies, yet the full range of somatic alterations driving HRD across cancers remains incompletely characterized. Here, we present a tumor-agnostic machine learning framework that integrates somatic multi-omics data, including copy-number variations, single-nucleotide variants, DNA methylation, and gene expression from over 8,000 patients in The Cancer Genome Atlas. Using a genome-wide mutational signature-based HRD score as ground truth, our model achieved high predictive performance and leveraged SHAP-based explainability to uncover HRD regulators beyond BRCA1/2 Cross-tumor analysis revealed both shared and cancer type-specific molecular determinants, whereas functional enrichment highlighted key molecular and cellular processes. These findings expand the known repertoire of HRD-associated alterations, provide a resource for mechanistic investigation, and demonstrate the potential of integrative AI approaches to improve patient stratification for HR-targeted therapies across diverse malignancies.