Genomic Scar Score: A robust model predicting homologous recombination deficiency based on genomic instability.

OBJECTIVE: To develop a novel machine learning-based algorithm called the Genomic Scar Score (GSS) for predicting homologous recombination deficiency (HRD) events. DESIGN: Method development study. SETTING: AmoyDx Medical Laboratory and Jiangsu Cancer Hospital. POPULATION OR SAMPLE: A cohort of individuals with ovarian or breast cancer (n = 377) were collected from the AmoyDx Medical Laboratory. Another cohort of patients with ovarian cancer treated with PARP inhibitors (n = 58) was enrolled in the Jiangsu Cancer Hospital. METHODS: We used linear support vector machines to build a Genomic Scar (GS) model to predict HRD events, and Kaplan-Meier analyses were performed by comparing the progression-free survival (PFS) of patients in different groups using a two-sided log-rank test. MAIN OUTCOME MEASURES: The performance of the GS model and the result of clinical validation. RESULTS: The GS model displayed more than 97.0% sensitivity to detect BRCA-deficient events, and the GS model identified patients that could benefit from poly(ADP-ribose) polymerase inhibitors (PARPi), as the GS score (GSS)-positive group had a longer progression-free survival (PFS) (9.4 versus 4.4 months; hazard ratio [HR] = 0.54, P < 0.001) than the GSS-negative group after PARPi treatment. Meanwhile, the GSS showed high concordance among different NGS panels, which implied the robustness of the GS model. CONCLUSIONS: The GS was a robust model to predict HRD and had broad clinical applications in predicting which patients will respond favourably to PARPi treatment.