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

Area of Science:

• Oncology
• Genomics
• Bioinformatics

Background:

• Homologous recombination deficiency (HRD) is a key biomarker for targeted cancer therapy.
• The full spectrum of genetic alterations causing HRD across various cancers is not well understood.

Purpose of the Study:

• To develop a tumor-agnostic machine learning framework to identify somatic alterations driving HRD.
• To uncover novel HRD regulators and understand their role across different cancer types.

Main Methods:

• Integrated multi-omics data (copy-number variations, single-nucleotide variants, DNA methylation, gene expression) from over 8,000 cancer patients.
• Employed a genome-wide mutational signature-based HRD score as ground truth for model training.
• Utilized SHAP-based explainability to identify key HRD-associated molecular drivers.

Main Results:

• Achieved high predictive performance in identifying HRD.
• Discovered novel HRD regulators beyond BRCA1/2.
• Identified both shared and cancer-specific molecular determinants of HRD across tumors.
• Highlighted key molecular and cellular processes involved in HRD.

Conclusions:

• The AI framework expands the known genetic alterations linked to HRD.
• Provides a valuable resource for further mechanistic research into HRD.
• Demonstrates the potential of AI in improving patient stratification for HR-targeted therapies in diverse cancers.