Abstract 2149: A whole genome sequencing classifier of homologous recombination deficiency

Abstract

Homologous recombination deficiency (HRd) is a DNA repair defect prevalent in but not exclusive to breast and ovarian cancer most commonly associated with BRCA1 or BRCA2 alterations. HRd results in accumulation of small and large scale genetic alterations across the genome, including allele specific copy number alterations (aCNAs), small nucleotide variants (SNVs), deletions, and structural variants (SVs). Detection of HRd in tumors predicts response to genotoxic drugs such as PARP inhibitors and platinum. Genome wide aCNAs such as large state transitions (LST), loss of heterozygosity (LOH), and telomeric allelic imbalances (TAI) in conjunction with BRCA1/2 mutation detection have been implemented in routine diagnostic testing to identify HRd in tumors. However, these features represent a subset of the genetic signatures predictive of HRd, and we hypothesize that a significant portion of tumors with HRd are missed using these existing assays. Whole genome sequencing (WGS) enables the detection of the full spectrum of genetic lesions that arise in an HRd tumor in a single assay. To demonstrate the added value of WGS to identify HRd, we trained and validated a pan-cancer classifier of HRd. A tumor/normal matched cohort of 321 cancer patients sequenced by WGS was assembled and analyzed as part of a retrospective study, representing 62 tumor types. An unbiased analysis of HRd associated SV signatures revealed the top quartile of samples harboring tandem duplications (Dups) and deletions (Dels) in the size range of 1-10kbp were enriched with BRCA1, BRCA2, and RAD51C/D alterations. Through curating Dels, Dups, HRd SNV/InDel signatures, and alteration of HRd associated genes, 37 unique patients were found to have high confidence HRd, out of which 13% had no alterations in BRCA1, BRCA2, or other HRd genes. We then trained a random forest classifier to identify HRd tumors. The most important predictive features were WGS-specific, namely small deletions with microhomology, SV Dels, and SV Dups. The HRd classifier was validated using an independent cohort of 556 samples from the Pan-Cancer Analysis of Whole Genomes (PCAWG) study. Of 46 samples with biallelic BRCA1/2 alterations, the classifier achieved high areas under receiver-operator characteristic (AUROC, 0.99) and precision recall curves (AUPRC, 0.96). The aCNA score, the number of segments harboring LST, LOH, and TAI, had similar AUROC (0.96) but lower AUPRC (0.87). There were 11 BRCA1/2 non-altered cases predicted to be HRd with the classifier which were not identified by CNA scores, in which 10 had at least 1 alteration in an HRd gene, including RAD51C, CHEK2 biallelic alterations and SVs in PALB2, Fanconi pathway genes, and ATM/ATR. We conclude that a classifier incorporating the additional mutational features which can only be detected using WGS can achieve superior precision in identifying HRd tumors and, in the future, uncover additional patients for therapeutic options. Citation Format: Kevin Hadi, Gunes Gundem, Max F. Levine, Aditya Deshpande, Minal Patel, Stan Skzrypczak, Majd Al Assaad, Juan Miguel Mosquera, Olivier Elemento, Andrew L. Kung, Juan S. Medina-Martínez, Elli Papaemmanuil. A whole genome sequencing classifier of homologous recombination deficiency [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 2149.