Whole genome sequencing approach to assess homologous recombination deficiency in a pan-cancer cohort

Abstract

Background

Homologous recombination deficiency (HRD) impacts cancer treatment strategies, particularly effective utilization of PARP inhibitors. However, the variability of different HRD assays has hampered the selection of oncology patients who may benefit from these therapies. Our study aims to use the whole genome landscape to better define HRD in a pan-cancer cohort.

Methods

We employed a whole genome sequencing HRD classifier that includes genome-wide signatures associated with HRD to analyze 580 tumor/normal paired samples. The HRD phenotype was correlated with genomic variants in BRCA1/2 and other homologous recombination repair genes.

Results

In this paper we show that the HRD phenotype is identified in various cancers including breast (21%), pancreaticobiliary (20%), gynecological (17%), prostate (9%), upper gastrointestinal (GI) (2%), and other cancers (1%). HRD cases are not confined to BRCA1/2 mutations; 24% of HRD cases are BRCA1/2 wild-type. A diverse range of gene alterations involved in HRD are elucidated, including biallelic mutations in FANCF, XRCC2, and FANCC, and deleterious structural variants. In a subset of cases, the whole genome sequencing-based classifier offers more insights and a better correlation to treatment response when compared to other assays.

Conclusions

Although HRD is a biomarker used to determine which cancer patients would benefit from PARP inhibitors, a lack of harmonization of tests to determine HRD status makes it challenging to interpret their results. Our study highlights the use of comprehensive whole genome sequencing analysis to better predict HRD and elucidates genomic mechanisms associated with this phenotype.

Plain language summary

Homologous recombination deficiency is a condition in which a cancer cell cannot repair certain types of DNA damage. It causes genetic instability and is often due to changes in parts of the DNA called genes, such as BRCA1 and BRCA2. Cancers with this deficiency can be more readily killed by certain drugs that prevent DNA repair. Some of these drugs are approved for the treatment of several types of cancer, including ovarian, breast, pancreatic, and prostate cancers. To better identify tumors with this deficiency, we characterize the whole genome of cancer samples. We find that a comprehensive analysis of the entire genome improves the detection of homologous recombination deficiency. This type of analysis may provide a more accurate way to guide treatment decisions for people with cancer.

Assaad, Hadi and Levine et al. develop a whole-genome sequencing classifier to improve the detection of homologous recombination deficiency (HRD) across a pan cancer cohort. The classifier detects HRD beyond BRCA1/2 mutations, reveals HRD-related genomic events, and correlates with treatment response in a subset of patients.

Introduction

BRCA1 and BRCA2 (BRCA1/2) play a significant role in an error-free DNA damage repair pathway known as homologous recombination repair (HRR)¹. This pathway corrects DNA double-strand breaks (DSBs) and interstrand cross-links^1,2. Somatic or germline mutations in BRCA1/2 (BRCAmut) lead to homologous recombination deficiency (HRD) and have long been associated with breast, ovarian, pancreatic, and prostate cancers^3,4. Poly (ADP-ribose) polymerase inhibitors (PARPi) have been developed to treat cancers associated with mutations in BRCA1/2 and other HRR genes based on the synthetically lethal relationship with HRD⁵. Furthermore, the use of alkylating-like agents, especially platinum-based chemotherapy (PBCT), has shown increased effectiveness for cancers with HRD^6,7.

Loss of function in HRR genes, such as RAD51B, ATM, FANC genes, CHEK2, PALB2, among others, can result from small mutations, structural variants (SV), or epigenetic changes^8–10. Therefore, evaluation of different deleterious mechanisms would require several assays and advanced analysis. Due to the broad range of HRR-inactivating genes, it is necessary to not only survey the full genetic landscape of HRR genes, but also detect DNA damage signatures associated with HRD for selective treatment with PARPi and PBCT¹⁰.

Current clinical methods to assess HRD status (see below) are either genetic (detecting mutations of relevant HRR genes) or scar-based (assessing the tumoral mutational landscape). Other methods, such as the estimation of HRD status by functional assays can provide more accurate evidence of a defect in the HR pathway in real-time. However, their routine clinical use is limited, and they remain as experimental approaches¹¹.

HRD is associated with DNA damage signatures, including single base substitution (SBS), SV signatures, in addition to loss of heterozygosity (LOH), large scale transition (LST), and telomeric allelic imbalance (TAI)^12–14. Assays that report HRD, such as MyChoice® CDx and FoundationOne CDx, among others, are based on targeted next-generation sequencing (NGS) and only employ a subset (LOH, LST, and TAI) of the available signals to detect HRD. Their use in PARPi treatment of ovarian cancer has been approved based on the results of multiple clinical trials^15–18. It is worth noting that in some trials several patients with BRCAness (HRD phenotype) have experienced shorter survival under first-line maintenance PARPi and numerous cases lacking BRCAness have shown extended survival^16,19. In the absence of reversion BRCA2 mutations^20–23, the underlying reason for the lack of response remains unknown but may be attributed to a potential false-positive HRD phenotype, either due to the testing method or a disconnect between the presence of genomic mutations and BRCAness^24,25. In contrast to the current commercially available companion diagnostics (CDx) for HRD testing, algorithms using whole genome sequencing (WGS), such as HRDetect and CHORD, employ all mutation classes^3,26. While not yet widely adopted in clinical practice, HRDetect, for example, has been used in the RIO trial to identify HRD tumors in triple-negative breast cancer patients treated with PARPi inhibitors²⁷.

The clinical and research assays have shown a variable association between BRCAness and the presence of HRR genes mutations^3,26,28. This variability indicates that not every deleterious mutation in these genes results in BRCAness. Conversely, BRCAness can also be associated with BRCA1/2 wild type (BRCAwt) and variants of unknown significance (VUS)²⁸. This highlights the necessity of focusing on the BRCAness phenotype rather than solely relying on the known pathogenicity of mutations.

WGS offers a comprehensive assessment of the genome, covering simple and complex SVs, copy number alterations (CNA), and mutational patterns associated with HRD^19,29. Because of that, it has the potential to achieve superior precision and sensitivity in identifying BRCAness when compared to other methods.

In this study, we analyzed WGS data from pan-cancer samples and investigated the prevalence and characteristics of HRD employing a WGS-based classifier. We compared our HRD results with scores from other WGS-based algorithms and, in a subset, with commercial HRD tests. In this subset, WGS-based HRD scores show a stronger correlation with treatment responses. We find a strong association of the HRD classifier with BRCA1/2 mutations, including SVs, and identify high-confidence HRD in BRCA1/2 wt cases.

Methods

Patient enrollment, tissue samples, and clinical data acquisition

Participants were prospectively enrolled at Weill Cornell Medicine (WCM) in the Institutional Review Board (IRB)-approved protocol WCM IRB #1305013903 (Research for Precision Medicine) that includes an informed consent for their data to be used. The WCM IRB protocol #1007011157 (Comprehensive Cancer Characterization by Genomic and Transcriptomic Profiling) allows for utilization of retrospective biospecimens. Tumor DNA for WGS was extracted from frozen tumor samples, formalin-fixed, paraffin-embedded (FFPE) archival tissue, or fluid from malignant ascites. Histopathology review was performed before DNA extraction. Germline DNA was primarily extracted from blood. When unavailable, saliva or benign tissue (frozen or FFPE) was used. We collected comprehensive clinical data, which encompassed age, sex, ethnicity, treatment-related information, radiologic findings, and pathologic data.

DNA Extraction and WGS

For DNA extraction from FFPE blocks, we used 5-micrometer-thick unstained slides macro-dissected for at least 80% tumor content. For DNA extraction from frozen specimens, we employed 3 mm core punches from the frozen OCT-embedded tissue. The Maxwell® 16 FFPE Plus DNA kit (Promega, Cat# AS1135) was employed, in combination with the Maxwell® 16 instrument (Promega, Madison, WI). DNA quality and quantity was assessed by using the Agilent Tapestation 4200 (Agilent Technologies) and the Qubit Fluorometer (ThermoFisher), respectively. WGS was carried out at the New York Genome Center on an Illumina Novaseq6000 sequencer using 2 × 150 bp cycles. Libraries were generated using the KAPA Hyper Library Preparation Kit (KAPABiosystems KK8502, KK8504), targeting 500 bp fragments, in compliance with the manufacturer’s instructions. DNA fragments underwent a series of preparation steps including shearing, end-repair, adenylation, and ligation to Illumina sequencing adapters. The prepared DNA fragments were size-selected using bead-based methods and amplified³⁰. Quality and quantity of the final libraries were assessed prior to sequencing.

WGS data processing pipeline and HRD curation

We employed Isabl GxT analytics³¹, a comprehensive data management, processing, and visualization platform, to analyze cancer whole genome and transcriptome sequencing data. This platform was used to process all WGS and RNA sequencing data and generate comprehensive reports^32,33. Within the Isabl pipeline deployed in Amazon Web Services (AWS) cloud high performance computing (HPC) environments, DNA and RNA alignment with Burrows-Wheeler Aligner (BWAMem) and Spliced Transcripts Alignment to a Reference (STAR), quality control, somatic and germline variant calling, and annotation were performed as previously described^32,34–37. Briefly, ensemble calling was done for both somatic and germline single nucleotide variants (SNV), insertions deletions (InDels), and somatic SVs. These variant classes were then annotated and those in which at least 2 of 3 callers of each class were included for reporting. Purity, ploidy, and genome-wide copy number states were estimated with Battenberg, followed by annotation of CNA events at the gene level. Driver alterations for all variant classes and their potential treatment targets were assessed by cross-referencing protein-coding variants with Catalog Of Somatic Mutations In Cancer (COSMIC)³⁸ and Oncology Knowledge Base (OncoKB)³⁹.

The WCM WGS pan-cancer cohort

The initial number of tumor/normal pairs that underwent WGS was 593, obtained from 469 patients. Thirteen samples from the dataset were excluded due to restrictions in the agreement with a collaborating investigator. The total number of genomes included in the final analyses was 580, that comprised two groups allocated for training (n = 381) and non-training (n = 212) cohorts (see below).

HRD classification

To determine HRD, we used the Isabl HRD classifier, a proprietary algorithm developed by Isabl Inc. As detailed below, data from a cohort at WCM was used for training. For validation the Isabl HRD classifier was applied across a dataset from the International Cancer Genome Consortium (ICGC) and across additional WCM samples.

Training cohort

WGS was performed on a WCM cohort of 381 matched tumor/normal (T/N) pairs from 321 patients, representing 62 tumor types, with a median sequencing depth of 92X for tumor and 48X for normal samples. Of these, 305 samples had tumor purity greater than 20% and compiled the training cohort. HRD-associated features were evaluated, focusing on COSMIC SBS3, deletions with microhomology, and small SV duplications and deletions. Cases were selected if they were top-quartile outliers for each feature, followed by further curation of cases exhibiting at least two outlier signatures. This approach identified 37 high-confidence HRD tumors and 268 HRR-proficient (HRP) cases.

A random forest classifier (Isabl HRD) was trained using these 37 cases with outlier HRD features and the remaining 268 HRP cases. The training model incorporated features including 96 trinucleotide single-nucleotide variant (SNV) contexts, 45 insertion/deletion (InDel) types categorized by type, length, microhomology versus repeat-mediated status, and fraction, as well as 38 SV features classified by type, size, and clustered versus dispersed patterns²³. Additionally, LST, LOH, and TAI scores were included in the classifier. Each sample was assigned an HRD probability score ranging from 0 to 1, with a score ≥0.5 indicating HRD and <0.5 classified as HRP. Based on histopathology review prior to model training, a tumor purity greater than 20% was required for accurate HRD assessment, as mentioned above.

External validation cohort

The performance of the classifier to predict biallelic BRCA1/2 status was evaluated using an external validation cohort of 556 breast, ovarian, prostate, and pancreatic cancer cases, with mutation calls obtained from the International Cancer Genome Consortium (ICGC). Performance was assessed using area under the receiver operating characteristic (AUROC = 0.96) and precision-recall (AUPRC = 0.60) curves, demonstrating high predictive accuracy of biallelic BRCA1/2 mutations (Figure S5).

Non-training cohort

WGS was then performed on 212 additional tumor samples with matched normal tissues, obtained from 148 patients, which we designated as the test cohort.

Fluorescence in situ Hybridization

Four-µm-thick formalin-fixed paraffin-embedded tissue sections were used for fluorescence in situ hybridization (FISH) analysis, as described in our protocols^36,37,40. Bacterial artificial chromosomes were designed against loci of interest to prepare break-part dual-color FISH probes⁴¹. For BRCA2 RP11-110O22 BAC clone was labeled red and RP11-11K16 clone was labeled green, and for ATM RP11- 144G7 BAC clone was labeled red and RP11-589O5 clone was labeled green. All clones were validated on normal metaphase spreads before any application on FFPE tissue. A positive break-apart was determined by one red, one green, and one yellow signal (combination of red and green signal indicating the normal chromosome homolog). At least 200 nuclei were analyzed per case using a fluorescent microscope (Olympus BX51; Olympus Optical). Cytovision 7.3.1 software was used for imaging and analysis.

Statistics and reproducibility

All statistical analyses and WGS HRD classifier training were conducted using R and Python. The classifier performance was evaluated using AUROC and AUPRC, calculated separately in the training and test cohorts. The primary training and internal validation cohort consisted of samples from the WCM dataset. An external validation was performed on an independent cohort from the ICGC. The same WGS processing pipeline (Isabl) was applied across all cohorts. Replicates refer to either biologically distinct samples from the same patient or samples from different tumor regions. No statistical method was used to predetermine sample size.

Ethics approval and informed consent

The use of the clinical data and samples for this study was approved by WCM institutional review board (IRB) protocols # 1305013903 (Research for Precision Medicine) and # 1007011157 (Comprehensive Cancer Characterization by Genomic and Transcriptomic Profiling).

Results

Sample characteristics and frequency of HRD

We performed WGS analysis on 580 tumor samples and the matching germline samples from 453 patients. These encompass 77 unique histology types (Fig. S1), classified according to the MSKCC Oncotree⁴², and obtained from 33 unique primary sites. Sites were divided into six cancer subgroups: prostate (29%), gynecological (20%), pancreaticobiliary (15%), breast (11%), upper gastrointestinal (GI) (10%), and “others” (12%), which included smaller cancer cohorts and rare tumors (Figs. 1a, b and S1).

Fig. 1. Cohort Characteristics.

A Left panel, overall breakdown of cohort by specimen type (left) and whether biopsy was from a primary or metastatic tumor (right). B Breakdown of the Isabl HRD score and HRD status (cutoff > 0.5) across the cohort. FFPE formalin-fixed, paraffin-embedded, GI gastrointestinal, HRD homologous recombination deficiency.

HRD analysis (see HRD Classification paragraph in Methods) was conducted on 521 of the 580 tumor samples, with 59 cases excluded due to low purity ( < 20%). The median purity was 71.5%, and the median coverage was 90x. Sixty-two samples across 53 patients exhibited HRD by WGS (Supplementary Data 1). The highest percentage of HRD cases was found in breast cancer (21%, including three of seven triple negative cases), followed by pancreaticobiliary (20%), gynecological (17%), and prostate cancers (9%). Upper GI cancers had the lowest percentage of HRD, with only one case harboring this phenotype. Additionally, one case of carcinoma of unknown primary in the “others” cohort demonstrated HRD.

Mutational landscape of HRD

While BRCAness is most commonly explained and clinically tested through BRCA1/2 alterations, dysfunction in other HRR pathway genes can also result in phenocopy. To investigate the genetic basis of HRD, we examined alterations in BRCA1, BRCA2, and other HRR pathway genes, considering LOH and compound hits indicative of biallelic loss of function. We focused particularly on SVs and somatic mutations with unknown effects on HRD (Fig. 2a, b), as follows.

Fig. 2. Genomic alterations and features of HRD.

A Landscape of BRCA1/2 and other HR-associated genes in HRD cases. B Top, BRCA1/2 alteration status in the HRD cohort by variant class. Bottom, further delineating alterations in genes outside of BRCA1/2 in cases that have BRCA1/2 events not considered in standard clinical practice (SSV or VUS) or are BRCAwt. C Genomic feature enrichment (Z-score calculated across the full HRD & HRP cohort) in HRD. Notably, the features of HRD are not distinguished by gene target or variant type. Del deletion, Dup duplication, GI gastrointestinal, Hom homozygous, HRD homologous recombination deficiency, HRR homologous recombination repair, MH microhomology, Mut mutation, SBS single base substitution, SV structural variant, VUS variant of unknown significance, WGS whole genome sequencing, WT wild-type.

BRCA1/2-Mutated HRD Cases

Among the 62 HRD samples from 53 patients, 47 (76%) harbored BRCA1 and/or BRCA2 alterations (BRCAmut-HRD). Specifically, 34 samples (55%) had biallelic pathogenic SNVs and InDels, and 4 (6%) had homozygous deletions in BRCA1/2 (Fig. 2a, b, Fig.e S2). Additionally, two samples (3%) harbored SVs with LOH in BRCA1/2, with a predicted impact on the coding sequence, and notably, neither of these cases had deleterious mutations in other HRR genes (Fig. 2b). Other 7 cases (11%) harbored BRCA1/2 VUS, either as small mutations or SVs. A notable example of a BRCA2 SV was observed in a case of advanced prostatic carcinoma with neuroendocrine differentiation, where WGS revealed a BRCA2::TMPRSS2 fusion associated with HRD with a predicted deleterious effect on BRCA2. The rearrangement was validated using a BRCA2 FISH break-apart assay (Fig. 3a).

Fig. 3. Validation of alterations in two WGS HRD positive cases.

A Prostate neuroendocrine carcinoma, (H&E image) with high-confidence HRD score. Genome circos and the number of Indels, SNVs of molecular signatures and rearrangements illustrate HRD features including a high rate of microhomology and SV deletions. The tumor harbors a BRCA2::TMPRSS2 fusion with deleterious effect on BRCA2, which rearrangement was validated by FISH. B High grade serous ovarian carcinoma (H&E image) harboring HRD signals as shown in the circos plot with high rate of microhomology deletions, SBS3, SV tandem duplications, and SV deletions. WGS detected an in-frame HIF1A::UIMC1 fusion that leads to a truncated, non-functional chimeric transcript that is missing exon 6, necessary for binding to BRCA1. Scale bars represent 100 µm for H&E images and 10 µm for the FISH image. H&E hematoxylin and eosin, HRD homologous recombination deficiency, Indel insertion deletion, SNV single-nucleotide variant, LOH loss of heterozygosity, SBS single base substitution, SV structural variant, FISH fluorescence in situ hybridization.

BRCA1/2-Wildtype HRD Cases

The remaining 15 HRD samples (24%) were BRCA1/2-wildtype (BRCAwt-HRD) (Fig. 2b). Among these, three had known pathogenic mutations in other HRR genes: one prostate cancer harbored a FANCF pathogenic mutation (p.W193), another prostate cancer had a biallelic FANCC deletion, and one endometrial carcinoma had a biallelic XRCC2 deletion. The remaining 12 BRCAwt-HRD cases fell into two categories: five (33%) had no detectable mutations in any HRR genes, while seven (47%) harbored a VUS in at least one HRR pathway gene, including RAD51B, PALB2, ATM, CHEK2, FANCD2, and RAD50 (Fig. 2a, b). Examples of BRCAwt-HRD cases include a high-grade serous ovarian carcinoma (HGSOC) with a HIF1A::UIMC1 fusion involving intron 13 of UIMC1, which interacts with BRCA1 in recognizing and repairing DNA lesions. The gene fusion was validated by RNA sequencing (Fig. 3b). Another example is a high-grade serous carcinoma of the fallopian tube (HGSFC) with a biallelic ATM VUS (p.P2842S). Both cases were classified as HRD-positive by MyChoice CDx, and both showed clinical responses: the HGSOC case responded to PBCT (RECIST 1.1, Supplementary Data 2), while the HGSFC case remained progression-free on PARPi for 20 months at the time of writing.

Genomic Features Enriched in HRD Tumors

HRD tumors exhibited distinct genomic signatures, including genome-wide small microhomology deletions (MH-dels), COSMIC SBS3 or SBS40 mutational signatures, SV deletions (1-10 kb), and SV duplications (1-10 kb) (Fig. 2c, Fig. S3). SV duplications were predominantly enriched in BRCA1-mutated tumors. While MH-dels were the most commonly enriched feature among HRD samples, three BRCAwt HRD cases with RAD51B mutations exhibited intermediate levels of MH-dels but demonstrated enrichment of SV deletions, SV duplications, and SBS3/40 signatures, suggesting a potentially distinct HRD feature profile.

(HRR)-proficient (HRP) cases with BRCA1/2mut (BRCA1/2mut-HRD)

In the HRP tumor cohort, 12.4% (57 of 459) samples had at least one mutated BRCA1/2 allele. Among these, 10% (5 of 57) had biallelic pathogenic mutations and 23% (13 of 57) had monoallelic pathogenic mutations. The remaining monoallelic and biallelic mutations were VUS small mutations or SV with germline pathogenic mutations. For the five samples with biallelic pathogenic BRCA1/2 mutations and HRP, one was from a patient with mixed ovarian carcinoma with a POLE pathogenic mutation (c.857 C > G, p.P286R) that led to a hypermutator state with a TMB of 897.9/mb and no HRD signatures. This suggests the possibility of the BRCA2 mutation being a result of the hypermutator phenotype. The remaining four samples were from one patient with prostate adenocarcinoma, where no HRD signatures were found despite the presence of two biallelic pathogenic BRCA2 mutations. The reason for this discordance remains unclear.

HRD concordance in cases with multiple samples

Our cohort included 54 patients with two or more samples, totaling 169 samples. Among these patients, 85% (46 of 54) did not have any samples with HRD, and 13% (7 of 54) had all their samples consistently showing HRD. Only 1 of 54 (2%) had a discrepant HRD status between samples. The patient had histologically and molecularly different synchronous ovarian and endometrial serous carcinomas. The sample from the brain metastatic lesion originated from the endometrial cancer, which displayed HRD associated with BRCA1, RAD51B, and RAD54B deleterious rearrangements. This sample also had drivers including TP53 (p.G244R), RB1 (p.R579fs29), and SMARCA2 (p.A7fs15), among others. In contrast, the ovarian cancer sample displayed only a FANCM rearrangement, with wild-type BRCA1, RAD51B, and RAD54B. The drivers for this sample included TP53 (p.G244R) and RRAS2 (p.G23V), with wild-type RB1 and SMARCA2.

Comparison of Isabl WGS HRD with other assays

Current commercial tests for HRD employ only GIS or LOH scores based on allele-specific copy number segmentation, a subset of genome-wide features predictive of HRD and captured by WGS. To explore the potential clinical significance of WGS-based HRD testing, we retrospectively assessed cases with WGS, commercial HRD scores, and clinical history that included treatment response. From the 39 serous carcinoma cases in the gynecological cancer cohort, where an HRD score can inform treatment decisions, we identified 15 cases that had undergone a commercial assay that included an HRD score. This group comprised 9 cases evaluated with MyChoice CDx GIS and six cases with FoundationOne HRD score. The results of both commercial assays and WGS testing are summarized in Table 1. Of these cases, three out of five identified as HRD-positive by WGS were deemed HRP by commercial assays; these included one case tested by FoundationOne (negative) and two cases by MyChoice CDx, one negative and the other inconclusive for HRD. All three cases had confident HRD phenotype features, including enrichment for MH-dels, SV deletions, or duplications. Two of the three patients were maintained on PARPi after showing a response to PBCT (RECIST 1.1) and experienced prolonged periods of non-progression, lasting 10 months and 5 years, respectively. The third patient did not receive PARPi.

Table 1.

Comparison of WGS HRD with commercially available assays

ID	Isabl	Commercial Assay	Result	Clinical response to Maintenance PARPi
WCM209	HRD	MyChoice CDx	HRD	No available PARPi response data
WCM414	HRD	MyChoice CDx	HRD	No disease recurrence to date (PFS = 20 months)
WCM295	HRP	MyChoice CDx	HRP	Did not receive PARPi
WCM435	HRP	MyChoice CDx	HRP	Did not receive PARPi
WCM442	HRP	FoundationOne CDx	HRP	Did not receive PARPi
WCM284	HRP	MyChoice CDx	HRP	Did not receive PARPi
WCM331	HRP	MyChoice CDx	HRP	Did not receive PARPi
WCM419	HRP	MyChoice CDx	HRP	Did not receive PARPi
WCM231	HRD	FoundationOne	HRP	Progression 10 months after maintenance therapy (PFS = 10months)
WCM205	HRD	MyChoice CDx	HRP	Did not receive PARPi
WCM228	HRD	MyChoice CDx	Inconclusive	No disease recurrence to date (PFS = 60 months)
WCM425	HRP	FoundationOne CDx	HRD	Progression 2 months after maintenance therapy (PFS = 2 months)
WCM436	HRP	FoundationOne CDx	HRD	Did not receive PARPi
WCM236	HRP	FoundationOne CDx	HRD	Progression 3 months after maintenance therapy (PFS = 3 months)
WCM272	HRP	FoundationOne CDx	Inconclusive	Progression 3 months after maintenance therapy (PFS = 3 months)

Concordant cases are in gray rows. HRD homologous recombination deficiency, HRP homologous recombination proficiency, PARPi PARP inhibitor, PFS progression free survival.

Among cases determined as HRP by WGS (N = 10), one was found to be inconclusive and three were found to be positive according to FoundationOne. All four cases had an enrichment in MH-dels, SV deletions, duplications, and SBS3/40. Our assessment using RECIST 1.1 (Supplementary Data 2) confirmed disease progression on PARPi after 3 months for the case deemed inconclusive by FoundationOne, and after 2 and 3 months, respectively, for two of the three cases with HRD by FoundationOne but HRP by WGS. The third case was not treated with PARPi.

We also investigated HRD positivity using two other HRD algorithms, CHORD and HRDetect (Fig. 2a, Supplementary Data 1)^3,26. Overall, 15% (9 of 62) of our WGS HRD samples were negative by CHORD, which included three cases that are simultaneously negative by HRDetect. Interestingly, two of those cases were BRCAwt with RAD51B SVs, and one with BRCA2 monoallelic VUS somatic small mutation and BRCA1 biallelic pathogenic somatic small mutation. Among the samples classified as HRD-negative by the Isabl HRD classifier (n = 508), 98.1% (498) were also negative by CHORD, and 94.6% were negative by HRDetect. Separately, in non-training cohort (n = 212), the three HRD algorithms showed concordance in 92.0% (195) of the samples. Among the 17 discordant cases, two were positive by Isabl HRD, negative by CHORD, but positive by HRDetect, while 15 were negative by Isabl HRD but positive by CHORD and/or HRDetect. Overall, in the non-training cohort, the agreement between Isabl HRD and CHORD was 96.2% (Cohen’s kappa = 0.79), while the agreement between Isabl and HRDetect was 95.3% (Cohen’s kappa = 0.77) (Fig. S4). All three WGS-based assays (CHORD, HRDetect, Isabl HRD) and the CNA score were applied to an external validation cohort, demonstrating that all three algorithms identified additional HRD cases compared to the CNA score, using biallelic BRCA1/2 mutations as the reference for true positives (Fig. S5).

Discussion

The significance of HRD in precision oncology has grown because of the development of PARPi and the correlation with response to PBCT in solid tumors^19,43–46. The routine clinical methods to determine HRD, either genetic (detecting associated mutations) or scar-based (mutational landscape), are restricted. Targeted NGS panels are limited in detecting the full range of BRCA1/2 mutational events and mainly focus on exonic regions^47–49. Targeted panels may also overlook alterations in other genes of the HRR pathway that may cause HRD^47–49. Consequently, recent research has focused on scoring the genomic signatures of the disruption in the HRR pathway like LOH, TAI, LST, SBS3, small deletions, and tandem duplications^1,29,49. However, these HRD ‘footprints’ are not fully revealed by common targeted panels and exome sequencing^1,50. Functional tests offer an alternative or a complementary approach to these DNA-based methods. They measure the actual current activity and dynamic changes of the HR pathway and may be more relevant for predicting response to therapy. However, they require living cells and assessing HRD status may be impracticable for tumors with a low proliferation rate (e.g., via RAD51 foci formation assay). Published studies on functional methods demonstrate analytical validity but not necessarily clinical validity or utility^51,52. In the end, the real gold standard for validating any HRD test is the patient’s response to therapies like PARPi and PBCT. To address these limitations, we employed a WGS approach to better determine the HRD status in a pan-cancer cohort.

Epithelial carcinomas of the ovary and fallopian tube are the only cancer types approved for PARPi treatment based on clinical HRD signature testing, according to the most recent NCCN guidelines⁵³. In our cohort, we compared WGS HRD phenotype testing results to MyChoice CDx and FoundationOne CDx when available. We identified several discrepant cases: patients with HRD on WGS but HRP by clinical panels showed treatment response to PARPi and/or PBCT, whereas those patients with HRP tumors by our WGS classifier but HRD on other panels exhibited treatment resistance to PARPi or PBCT. Although there was a limited number of cases for comparison, we noticed that the discrepancy in HRD status, as determined by WGS analysis, was higher in FoundationOne CDx than Myriad CDx cases. This is likely because FoundationOne CDx only reports the LOH metric, whereas Myriad CDx provides a score that incorporates three metrics, LOH, LST, and TAI, allowing for a more accurate assessment of HRD status^18,47. While preliminary and limited by the number of cases, our study provides evidence that WGS would improve prediction of HRD with subsequent treatment response, beyond current FDA approved CDx for HRD. As reported in other studies, a WGS classifier of HRD might reduce false negatives and increase the inclusion of appropriate patients in relevant clinical trials⁵⁴.

In an effort to assess concordance across commercial and research tests that employ GIS or LOH scores, recent analysis of 13 GIS-based assays revealed challenges in unifying a definition of HRD and reported a wide range of percent positive HRD cases ( > 50% differences) in the cohorts studied⁵⁵. More fundamentally, comparative studies that continue to be designed around performance metrics other than treatment or other clinical response limit objective assessment of the use of such assays.

In addition, multiple algorithms have been developed that rely on CNA features, with or without additional features derived from whole-exome sequencing (WES) data, to generate a GIS. These include HRDsum and HRProfiler, among others^56,57. A key distinction between these approaches and WGS-based assays is the absence of SV-based features, which are critical signatures of HRD-related genomic damage and are not reliably captured in WES due to its limited coverage of non-coding regions and lack of genome-wide SV information, which are essential to the performance of classifiers like Isabl HRD, CHORD, and HRDetect⁵⁸. To evaluate the impact of this difference, we compared the performance of three WGS-based classifiers—Isabl HRD, CHORD, and HRDetect—against the GIS score, representing CNA-based algorithms as a whole. Our findings showed that WGS-based classifiers have superior performance in identifying additional HRD cases, underscoring the added value of incorporating SV-based features in HRD detection.

PARPi are progressively becoming a part of the therapeutic arsenal against tumors with HRD. However, treatment of non-ovarian cancers with PARPi depends on specific conditions, according to NCCN guidelines. These include the presence of a pathogenic BRCA1/2 mutation for breast and pancreatic cancers, or the presence of pathogenic mutations in BRCA1/2 and/or other HRR genes in prostate cancer⁵³. Our results reveal that the HRD phenotype was not only identified across a variety of cancer types, but also associated with pathogenic variants, VUS, and wild-type status of BRCA1/2 or other HRR genes.

Our study further expands the spectrum of molecular events that cause HRD and includes biallelic SVs in BRCA1/2 and other HRR pathway genes. Our results in this area confirm the data reported by Ewing et al., on BRCA1/2 SVs in ovarian serous carcinomas, prostatic carcinoma, and breast carcinoma and expand our knowledge of SVs impacting other HRR genes like RAD51B, FANCC, and CHEK2 among others⁵⁹. We also report that cases with structural or small mutation VUS in BRCA1/2 or other HRR genes were the most probable drivers in cases with HRD phenotype, not coinciding with other known pathogenic mutations in the HRR pathway. This suggests that those VUS are possibly related to HRD and might be considered pathogenic, supporting a previous study that reported a correlation between BRCA1/2 VUS and the HRD phenotype⁶⁰.

We also interrogated those BRCAwt cases with HRD phenotype. These patients, identified as having HRD tumors through clinically approved testing assays like MyChoice CDx and Foundation CDx, as well as other WGS-based assays, have been reported in the literature to constitute approximately 20–30% of cases^19,61,62. This range aligns with the data observed in our study. Based on WGS, some of our patients had biallelic mutations in other HRD genes, like PALB2 and RAD51B, known to be associated with HRD. In other cases, we identified variants in genes reported to be linked to HRR pathway dysfunction without previously known clinical impact. One example is a case with a UIMC1 fusion in which the deletion of the AIR domain, essential for binding to the BRCA1-A complex, could have caused HRD⁶³. In some BRCAwt HRD cases, the genomic cause could not be uncovered by WGS, which may be due to epigenetic changes like BRCA1/2 promoter methylation^64–66.

One of the main conclusions of our study is that, compared to commercial assays, WGS-based HRD assessment could clarify inconclusive or misleading results from NGS-based assays that rely on fewer signatures. On a routine basis, we discuss such cases during a research molecular tumor board (MTB) format, a Continuing Medical Education (CME) accredited conference^34,67–71. Examples include two patients (detailed above), both of which showed a WGS HRD phenotype but were HRP and inconclusive by MyChoice CDx respectively. For the first patient, following a recent disease recurrence of HGSOC, the consensus was to consider PARPi as a valid treatment option, especially in light of a previous response to PBCT. For the second patient, who had HGSOC with a WGS-determined HRD phenotype, the commercial HRD assay results were inconclusive. However, WGS identified a biallelic SV in RAD51B as the likely cause of the HRD phenotype, rather than the monoallelic BRCA2 VUS found through clinical targeted sequencing. The consensus was to continue the patient on maintenance PARPi treatment started earlier, and to which the patient was responding.

In summary, we illustrate how a WGS HRD classifier could be universally applicable in precision oncology, unearthing the HRD phenotype that may be invisible by other methods. Future functional studies will help elucidate the role of SVs in causing HRD. We acknowledge that larger clinical response data is required to definitively assess accuracy of HRD predictions against treatment response. This study provides major impetus to evaluate WGS as a potential clinical diagnostic tool to target PARPi and PBCT responses and paves the way for both clinical validity studies and drug trial designs that include WGS methodology in their endpoints.

Author contributions

M.A., K.H.a., and J.M.M. performed study concept and design and wrote the manuscript. J.M.M., M.A., M.S., J.M., and J.O. generated the patient cohort. K.H.a. performed statistical analysis and interpretation of data. K.H.a., M.L., G.G., J.S.M., and E.P. performed genomic analyses. JM.M and M.A. performed the histopathological assessment. A.Se. performed the FISH experiments. M.A, A.S.e., M.T., K.H.a., and M.L. prepared the figures. A.K. and D.C.W. performed specimen processing. M.P., A.Sb., and H.-H.K. helped with data organization. D.G., E.A., E.C.-D., S.T.T., A.J.O., M.A.S., C.N.S., K.Ho., and D.M.N., participated in patient consenting. O.E. provided institutional support. All authors read and approved the final paper.

Peer review

Peer review information

Communications Medicine thanks Gianluca Tedaldi, Eva Romanovsky and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Data availability

The analyzed data is available in the Supplementary Data files and illustrates the cohort tumor subtypes, the BRCA1/2 alterations in HRD positive cases, the top features contributing to the HRD score, the agreement between HRD classifiers, and the ROC and PRC curves. The raw WGS data that support the findings of this study are available in the database of Genotypes and Phenotypes (dbGaP) under the accession number: phs004351.v1.p1.

Competing interests

Kevin Hadi, Max F. Levine, Gunes Gundem, Juan S. Medina-Martinez, Minal Patel and Elli Papaemmanuil are Isabl, Inc. employees. All other authors declare no competing interest.

Supplementary information

The online version contains supplementary material available at 10.1038/s43856-025-01308-5.