Optimized detection of homologous recombination deficiency improves the prediction of clinical outcomes in cancer

Abstract

Homologous recombination DNA-repair deficiency (HRD) is a common driver of genomic instability and confers a therapeutic vulnerability in cancer. The accurate detection of somatic allelic imbalances (AIs) has been limited by methods focused on BRCA1/2 mutations and using mixtures of cancer types. Using pan-cancer data, we revealed distinct patterns of AIs in high-grade serous ovarian cancer (HGSC). We used machine learning and statistics to generate improved criteria to identify HRD in HGSC (ovaHRDscar). ovaHRDscar significantly predicted clinical outcomes in three independent patient cohorts with higher precision than previous methods. Characterization of 98 spatiotemporally distinct metastatic samples revealed low intra-patient variation and indicated the primary tumor as the preferred site for clinical sampling in HGSC. Further, our approach improved the prediction of clinical outcomes in triple-negative breast cancer (tnbcHRDscar), validated in two independent patient cohorts. In conclusion, our tumor-specific, systematic approach has the potential to improve patient selection for HR-targeted therapies.

Introduction

As a part of the Fanconi Anemia (FA) pathway, homologous recombination (HR) is an evolutionarily conserved, tightly regulated mechanism for high-fidelity repair of DNA double-strand breaks (DSBs)¹. Deficiency in homologous recombination (HRD) has profound consequences for replicating cells driving genomic instability and oncogenic transformation. In cancer, HRD results in a fundamental vulnerability, and tumors with HRD are markedly sensitive to DSB-inducing agents such as platinum-based chemotherapy and Poly-ADP Ribose Polymerase (PARP) inhibitors².

High-grade serous ovarian cancer (HGSC), the most common and most lethal subtype of ovarian cancers³, is characterized by profound genomic instability. Around half of the HGSC cases harbor genomic alterations leading to HRD⁴, and these patients have been shown to benefit from treatment with PARP inhibitors^5,6. The HRD test previously used in PARP inhibitor clinical trials (MyriadMyChoise®CDx)^5,6 works by quantifying specific allelic imbalances (AIs): (1) Large scale transitions (LSTs)⁷, (2) Loss of heterozygosity (LOH)⁸ and (3) Telomeric allelic imbalances (TAIs)⁹. However, the decision criteria for these HRD-specific AIs (HRD-AIs) and the HRD status classification were originally designed using a mixture of breast and ovarian cancer samples^7,8,9,10. Further, other algorithms for HRD detection have primarily focused on BRCA1/2 mutation prediction^11,12. As the genomic drivers and mutational processes differ across the cancer types, the details of the genomic instability occurring due to HRD in HGSC remain unclear.

Herein, via pan-cancer analysis, we show that HGSC harbors unique patterns of AIs, which are also distinct from triple-negative breast cancers (TNBC). Using a systematic approach based on machine learning and statistics on The Cancer Genome Atlas ovarian cancer (OVA-TCGA) multi-omics dataset, we optimized the criteria for HRD-AIs on HGSC. We implemented these criteria as an open-source algorithm (ovaHRDscar) to reliably define HRD status beyond the prediction of BRCA1/2 mutations. We show that ovaHRDscar improves the prediction of clinical outcomes in three independent clinical datasets compared to previous algorithms. Further, we show that our approach improves the prediction of clinical outcomes also in TNBC (tnbcHRDscar). Thus, our machine learning-aided disease-specific approach (HRDscar) shows promise as a biomarker that can improve outcome prediction and patient selection for HR-targeted therapies in cancer.

Results

Systematic pan-cancer characterization reveals unique features of allelic imbalances in HGSC

To elucidate the potential differences in the patterns of AIs across human cancers, we first characterized the quantity and the length distributions of AIs in the 18 most common cancer types from the TCGA (Fig. 1a). Interestingly, HGSC had the highest number of AIs (Fig. 1b) and the lowest median length (Fig. 1c). Concordantly, HGSC showed the highest levels of LOH events (Supplementary Fig. 1a) with one of the lowest median length (Supplementary Fig. 1b)….