Signatera HRD评分实现对同源重组缺陷的高准确度分类

Homologous recombination deficiency (HRD) is known to predict patient response to poly (ADP-ribose) polymerase (PARP) inhibitors and platinum agents, yet reliable detection remains difficult in patients who lack pathogenic variants in canonical homologous recombination repair genes. For these cases, HRD status can be assessed by the genomic “scars” caused by genomic instability, such as loss of heterozygosity (LOH), telomeric allelic imbalance (TAI), and large-scale state transitions (LST). However, current approaches have limited sensitivity, especially in specimens with low tumor purity. Here, we describe the development and performance of Signatera HRD score, an algorithm that uses tissue-based whole-exome and -genome sequencing (WES and WGS) data to determine HRD status. The algorithm integrates somatic single nucleotide variant (SNV) and copy number variant (CNV) features within a probabilistic framework to infer genomic scarring with high sensitivity and specificity. The underlying model was trained on WES and WGS data from 1,600 tumors of multiple cancer types. From genome-wide copy-number profiles of these tumors, non-negative matrix factorization identified 22 distinct and recurrent CNV signatures. Of these 22 signatures, 9 correlated with LOH and broader genomic instability and were retained as HRD features. The CNV signatures, combined with established single-base substitution (SBS) and insertion and deletion (ID) mutational signatures adjusted for tumor purity, were then fed into a machine learning classification model to predict HRD status. We retrospectively evaluated the performance of Signatera HRD score model in a cohort of 206 patients who underwent bespoke, mPCR-NGS tumor testing (Signatera TM ) with orthogonally determined HRD status. Initial performance (measured as the area under the curve [AUC]) was 0.8, but increased to 0.92 by adjusting the CNV signatures for the ploidy of each sample. Incorporating established SBS and ID mutational signatures further improved the discriminative power of the model to an AUC of 0.97. Using an HRD-status threshold derived from the training set of patients, the model achieved 94.4% sensitivity at 94.2% specificity in the evaluation cohort. Feature-importance analyses indicated that mutational signatures were the dominant contributors to model performance. In conclusion, Signatera HRD score is a tissue-based model that combines ploidy-adjusted CNV signatures with SBS/ID mutational signatures to reliably predict tumor HRD status, achieving performance comparable to or better than established genomic-scar metrics such as LOH, TAI, and LST. By leveraging widely available NGS data, this approach can expand access to HRD assessment and help identify patients more likely to benefit from PARP inhibitors without requiring specialized scar assays.