The recognition of biallelic germline or somatic mutations in BRCA1/2 is, to date, one of the most clinically relevant and frequently used genetic biomarkers of homologous recombination repair deficiency (HRD) in the clinics (Dougherty et al., 2017; Hoppe et al., 2018). Patients harboring germline pathogenic variants (GPVs) in BRCA1/2 have a higher risk of developing breast and/or ovarian cancer (Mersch et al., 2015). Patients with germline or somatic mutations have an enhanced benefit from targeted therapies such as platinum-based chemotherapy or poly (ADP-ribose) polymerase inhibitors (PARPi) (Hennessy et al., 2010). The terms “BRCAness” or “HRD phenotype” refer to tumors with similar clinicopathological and molecular characteristics to tumors with BRCA1 and BRCA2 GPVs (Lord and Ashworth, 2016). Gene alterations occurring in other homologous recombinant associated genes, such as PALB2 (Tischkowitz et al., 2007; Thomas and Brown, 2015) and RAD51C/D (Kondrashova et al., 2017; Polak et al., 2017), have been linked to the HRD phenotype. Inactivation through promoter methylation of BRCA1 and RAD51C has also been found to result in HRD tumors (Ruscito et al., 2014; Polak et al., 2017; Staaf et al., 2019), and these tumors also demonstrate increased sensitivity to PARPi and platinum (Kondrashova et al., 2018).

Advances in tumor sequencing resulted in the development of methods to identify HRD tumors independently of identifying the cause. Cancer genomes of patients with BRCA1/2 mutations are enriched with particular mutational patterns as well as a high number of distinct LOH regions. In addition, BRCA1/2-deficient tumors include small deletions with ≥4 bp flanking homology. Several structural variations are typical of BRCA1/2-deficient cancer genomes, including deletions up to 100 kb, unclustered tandem duplications of ∼10 kb associated with BRCA1 mutations (Willis et al., 2017), and deletions up to 1-10 kb in cancers are found in patients with BRCA2 mutations (Degasperi et al., 2020). A specific single-base substitution signature (also known as single-nucleotide variants), referred to as COSMIC signature 3, is strongly associated with BRCA1/2 deficiency (Polak et al., 2017).

Whole-genome sequencing (WGS) data enable the detection of different genomic alterations such as base substitutions, indels, rearrangements, and copy number aberrations, which are the result of homologous recombination deficiency. There are two HRD classifiers that are based on features extracted from WGS data. HRDetect (Davies et al., 2017) is a weighted logistic regression model based on six input features: the proportion of small deletions with microhomology at the breakpoint junction, HRD index based on genomic scars, COSMIC signatures 3 and 8, and two rearrangement signatures 3 and 5. This model was trained on BRCA1/2-null breast cancers. The classifier of Homologous Recombination Deficiency (CHORD) (Nguyen et al., 2020) is a random forest model that uses relative counts of somatic mutation contexts from WGS data.

Both classifiers classify >90% of tumors with biallelic inactivation via DNA mutation of BRCA1/2 as HRD and have generally high accuracy as measured by AUC∼0.98 (area under the curve) (Davies et al., 2017; Nguyen et al., 2020). Mutations in PALB2, RAD51C/D, and BARD1 are associated with HRD signatures (Polak et al., 2017; Matis et al., 2021) and account for a small fraction of non-BRCA1/2-mutated HRD cases (Golan et al., 2021). Nguyen et al. (2020), in the paper that introduced CHORD, reported that a substantial proportion (∼40%) of cancer samples identified as HR-deficient did not harbor any mutation in known HR-related genes (Nguyen et al., 2020), while Davies et al. (2017) reported more than 30% of these cases. These findings indicate that conventional testing for mutations in HR genes will miss a considerable number of HRD tumors where HRD is caused by unknown reasons.

The possible source of high unexplained cases could be either technical or biological. Both HRDetect and CHORD are continuous scores, designed to determine if a tumor exhibits HRD. Both use a universal threshold that was not optimized for specific cancer types. HRDetect threshold was developed based on the breast cancer dataset but this cut-off has been used for other cancer types. The CHORD study used a 0.5 cut-off. In addition, BRCA1/RAD51C promoter methylation is not measured in most WGS studies or on only one subset of these samples.

Here, we aim to examine the range of missing proportions of HRD samples across various three tumors where HRD is frequently reported (breast, ovarian and pancreas cancer) and determined the impact of cancer-type-specific thresholds as well as of promoter methylation BRCA1/RAD51C for an available subset of cases. To do so, we used published CHORD and HRDetect scores for these three cancers (Davies et al., 2017; Degasperi et al., 2020; Nguyen et al., 2020), as well as published HRDdetect scores for pancreas cancer (Golan et al., 2021) and CHORD scores that we calculated. In the case of ovarian and breast cancers, we limited our study to the subset of patients with available data for the methylation status of the BRCA1/RAD51C promoter. We determined the proportion of unexplained cases if we use cancer-type specific thresholds (for pancreas, ovarian, and breast cancer) and promoter methylation status (for ovarian and breast cancers).

Datasets . Studies that performed homologous-recombination deficiency detection analysis on the same samples using the CHORD (Nguyen et al., 2020) and HRDetect (Davies et al., 2017; Degasperi et al., 2020) classifiers were selected. From the selected studies, we made the largest unique intersection of sample names containing prediction scores of HR deficiency for both classifiers, CHORD and HRDetect. The dataset was divided into four major groups of HR-related cancers: breast, pancreatic, and ovarian (Supplementary Table S1), while other cancer types were put into a separate category (Supplementary Table S2) due to the low number of biallelic events and samples labeled as HRD. We included only breast and ovarian cancer samples that had verified BRCA1/2 with respect to methylation (Davies et al., 2017). The methylation status of HR-related gene promotors was considered to be an important underlying cause of HRD in tumors and we wanted to include only samples with validated methylation status to perform the downstream analysis. For the pancreatic dataset, we used 391 pancreatic samples whose data alongside the HRDetect classifier results were provided by Golan et al. (2021). For pancreatic samples, we ran the CHORD classifier using the default setting as it was previously described (Nguyen et al., 2020). The final combined dataset consisted of discrete datasets of 1) 371 breast cancers, 2) 66 ovarian cancers, 3) 391 pancreatic cancers, and 4) 1 238 samples belonging to other cancer types. For each sample in selected studies, we extracted the available methylation status of BRCA1/2 genes for the breast and ovarian cancer samples alongside biallelic and monoallelic alternations in HR-related genes for all cancer types. We considered biallelic germline inactivation to be present when a germline pathogenic variant (GPV) was the first hit with the second hit being loss-of-heterozygosity (LOH) or somatic mutation. Somatic biallelic inactivation was considered where at least one hit was a somatic mutation, while promoter hypermethylation biallelic inactivation was defined as when one hit was promoter methylation and the other one was somatic or LOH. Monoallelic inactivation was considered when only one gene had any mutation other than LOH. Samples carrying biallelic inactivation in HR-related genes were considered to be true HR-deficient tumors. A detailed summary of all biallelic and monoallelic alterations in analyzed HR-related genes can be found in Supplementary Table S1 and Supplementary Table S2, alongside the source of information regarding these gene alterations.

Assessment of the accuracy of CHORD and HRDetect classifiers through ROC and precision-recall curves . To assess the accuracy of each classifier for each of the four major cancer types, we calculated receiver operating characteristics (ROCs) using the R function “roc” from package “pROC” (Robin et al., 2011) and precision-recall (PR) curves using R function “pr.curve” from package “PRROC” (Grau et al., 2015) by comparing CHORD and HRDetect probability scores against samples carrying biallelic inactivation in HR-related genes. Bootstrapping (2000 samples) was performed to estimate the 95% CI of the area under the ROC curve (AUC). Additionally, we compared the performance of these classifiers when no methylation data are available for breast and ovarian cancers to highlight the importance of promoter hypermethylation in HR-deficient tumors.

Determining the optimal threshold. We applied a tenfold nested cross-validation approach to find the optimal threshold values of HRDetect and CHORD for classifying samples as HR-deficient or -proficient within breast, pancreatic, and ovarian cancers. The inner tenfolds were used to calculate the average optimal threshold, while the outer folds in the cross-validation process containing 10% of test data were used to assess the accuracy of the classification of HR-deficient samples. The reported optimal threshold for each classifier was calculated as the mean of all the average thresholds in outer loops for each cancer type.

Statistical analysis. Probabilistic scores from CHORD and HRDetect classifiers were compared with Spearman correlation (Spearman, 1987) using R functions cor () or cor. test (). The one-sided partially overlapping samples z-test for dichotomous variables with R function “Prop.test” from package “Partiallyoverlapping” (Derrick, 2018) was used to determine the statistically significant differences in the proportion of samples classified as HRD samples with and without evidence between CHORD and HRDetect. An one-sided Fisher’s exact test using the R function “pairwise_fisher_test” from package “rstatix” (Kassambara, 2021) was used for testing the differences of explained and unexplained classifications between cancer types within each classifier. For comparison of the different ROC curves, we used the DeLong’s test (two-sided, paired-samples) for two correlated ROC curves using the R function “roc.test” from package “pROC” (Robin et al., 2011). Corrections for multiple hypothesis testing were done using Bonferroni correction, and adjusted p-values were reported. All the analyses were carried out in R statistical programming language version 4.1.0.

Our study shows an integrated overview of detecting homologous recombination deficiency in cancers using CHORD and HRDetect classifiers. Here, we have mainly focused on three cancers most commonly associated with HRD: breast, ovarian, and pancreatic cancers. We observed that biallelic inactivation of genes explains a large fraction of samples possessing HRD when using a universal default threshold, as was demonstrated in previous studies (Davies et al., 2017; Nguyen et al., 2020; Golan et al., 2021). However, around 10–28% of patients without known underlying causes were detected by these classifiers despite their high performance based on the default threshold. In this study, we found that by applying a cancer-type-specific threshold the number of unexplained cases reduced to around 8.9–12.3% without decreasing the sensitivity of 96%. We estimate that in this dataset up to ∼10% of HRD cases are caused by types of alterations that still have not been associated with HRD and therefore gene-centric testing for mutations in HR genes will likely fail to identify them. Similar results apply to the analysis of other cancer types in which the HRD cancers are rarer in comparison to the four well-known HRD cancers. The low number of HRD mutations in prostate samples and other cohorts did not allow the determination of a reliable cancer-type-specific threshold. The small fraction of unexplained cases is consistent with our previous proposal (Foulkes and Polak, 2019; Matis et al., 2021) that if alterations in novel genes lead to HRD, taken together, they will all account for only a very small proportion of all HRD cases.

The different cut-offs that we observed may be due to subtle differences across cancer in the mutational landscape even for tumors with different same gene defects, especially in mutational signatures (Degasperi et al., 2020). Furthermore, as it was highlighted by Nguyen et al. (2020), additional threshold optimization and validations are also required when applying classifiers to WGS data generated by other variant calling pipelines. Our cohort contained data generated by various pipelines for CHORD and HRDetect in each cancer type, which may affect the overall comparison of results between these classifiers. In addition to the threshold value, it is important to investigate other features affecting the mutation landscape of tumors, such as deficiency in mismatch repair (MMR), which may have a negative impact on the overall performance of classifiers in specific tumors. It was noted by Golan et al. (2021) that one pancreatic sample with biallelic inactivation in BRCA2 and PMS2 (responsible for MMR) was misclassified by HRDetect and CHORD classifier and had both scores near zero.

In addition to cancer-type-specific thresholds that reduce the number of unexplained cases, we demonstrated the importance of including the promoter methylation status of BRCA1 and RAD51C in order to evaluate the fraction of HRD cases that are explained by known causes. In breast and ovarian cancers, for which methylation analysis is most often conducted, promoter methylation of BRCA1 accounts for at least 20% of explained biallelic inactivation cases of HRD, labeled by either of the classifiers, and lack of methylation data significantly affects the performance of classifiers. The proportion of unexplained cases in other cancer types may have been reduced if methylation analysis data existed, especially in pancreatic cancer where some detected monoallelic PVs could have other events such as promoter methylation that would explain their HRD. These observations highlight the advantage of using these classifiers alongside conventional testing for patient selection and stratification in clinics, as was already suggested by several studies (Zhao et al., 2017; Staaf et al., 2019; Chopra et al., 2020). The relationship between the presence of HRD and response to therapies such as PARP inhibitors is not precise and there is currently no “ground truth” for measuring HRD. Resistance to PARP inhibitors can co-exist with HRD (Dias et al., 2021), so the presence of HRD is not by itself a direct predictor of response to PARP inhibitors and other drugs such as platinum that cause double-strand DNA breaks. Combinations of different approaches such as WGS-based, FDA-approved assays, and newer functional assays such as the RAD51 foci assay (Pellegrino et al., 2022) will ultimately lead to a better selection of HRD patients for appropriate therapies. Hence, our review re-analysis emphasizes the power of both CHORD and HRDetect in the stratification of patients possessing HRD phenotype across various cancers, as well as the importance of identification and further validation of new unrevealed oncogenic mutations.