DirectHRD enables sensitive scar-based classification of homologous recombination deficiency

Homologous recombination deficiency (HRD) is a predictive biomarker for efficacy of PARP (poly ADP-ribose polymerase) inhibition and platinum chemotherapy for cancer patients but remains challenging to detect. The discovery of patients without pathogenic mutations in known HR genes but exhibiting genomic scars indicative of HRD led to the FDA approval of the first scar-based HRD test. Despite advancements in whole genome sequencing (WGS) and integration of large training datasets with machine learning models, current methods lack the sensitivity required for detecting HRD scars in low tumor purity samples, especially in liquid biopsies. Here, we describe DirectHRD, a genomic scar-based HRD classifier based on WGS. Compared to other WGS-based methods, DirectHRD exclusively utilizes a highly specific type of HRD scar-small deletions with microhomology-and its associated signatures in a probabilistic framework. We applied DirectHRD to 501 tumor and 90 cell-free DNA (cfDNA) samples from 4 cancer types: breast, ovarian, prostate, and pancreas. Among all 501 tumor biopsies, DirectHRD achieved 100% detection of HRD with high specificity (>90%). Across all 90 cfDNA samples, the method achieved an area under the curve of 0.87 and demonstrated the ability to detect HRD at tumor fractions as low as 1%, making it 10 times more sensitive than state-of-the-art methods.