A generalizable machine learning framework for classifying DNA repair defects using ctDNA exomes.

Area of Science:

• Oncology
• Genomics
• Bioinformatics

Background:

• DNA damage repair (DDR) defects influence treatment sensitivity in metastatic prostate cancer.
• Current DDR gene sequencing biomarkers lack predictive accuracy for treatment benefit.
• Somatic alteration signatures require whole-genome sequencing, limiting clinical utility.

Purpose of the Study:

• To develop and validate a machine learning model for identifying DDR deficiencies using accessible liquid biopsy data.
• To improve the prediction of treatment response in metastatic prostate and bladder cancers.
• To introduce DARC Sign, a tool for classifying multiple DDR deficiencies from plasma cell-free DNA.

Main Methods:

• Whole-exome sequencing data from 155 plasma and leukocyte samples were analyzed.
• XGBoost ensemble models were trained using mutation and copy number features.
• Deep targeted sequencing established ground truth labels for DDR gene alterations.

Main Results:

• XGBoost models achieved high performance (AUCs of 0.99-1.00) for identifying BRCA2, CDK12, and mismatch repair deficiencies.
• The model re-classified samples with discordant genomic features and identified a novel BRCA2 copy loss.
• The developed approach outperformed existing exome-based BRCA2 deficiency classifiers.

Conclusions:

• DARC Sign, a machine learning tool, utilizes liquid biopsies for simultaneous identification of multiple DDR deficiencies in metastatic prostate cancer.
• This method offers a clinically practical approach to identifying DDR deficiencies.
• The tool has the potential to guide patient selection for DDR-directed therapies in clinical trials.