Pan-cancer landscape of homologous recombination deficiency

Abstract

Homologous recombination deficiency (HRD) results in impaired double strand break repair and is a frequent driver of tumorigenesis. Here, we develop a genome-wide mutational scar-based pan-cancer Classifier of HOmologous Recombination Deficiency (CHORD) that can discriminate BRCA1- and BRCA2-subtypes. Analysis of a metastatic (n = 3,504) and primary (n = 1,854) pan-cancer cohort reveals that HRD is most frequent in ovarian and breast cancer, followed by pancreatic and prostate cancer. We identify biallelic inactivation of BRCA1, BRCA2, RAD51C or PALB2 as the most common genetic cause of HRD, with RAD51C and PALB2 inactivation resulting in BRCA2-type HRD. We find that while the specific genetic cause of HRD is cancer type specific, biallelic inactivation is predominantly associated with loss-of-heterozygosity (LOH), with increased contribution of deep deletions in prostate cancer. Our results demonstrate the value of pan-cancer genomics-based HRD testing and its potential diagnostic value for patient stratification towards treatment with e.g. poly ADP-ribose polymerase inhibitors (PARPi).

Fig. 1. CHORD is a random forest Classifier of HOmologous Recombination Deficiency able to distinguish between BRCA1- and BRCA2-type HRD phenotypes in a pan-cancer context.

a The features used for training CHORD are relative counts of different mutation contexts, which fall into one of three groups based on mutation type. (i) Single nucleotide variants (SNV): six possible base substitutions (C > A, C > G, C > T, T > A, T > C, T > G). (ii) Indels: indels with flanking microhomology (del.mh, ins.mh), within repeat regions (del.rep, del.none), or not falling into either of these 2 categories (del.none, ins.none). (iii) Structural variants (SV): SVs stratified by type and length. Relative counts were calculated separately for each of the 3 mutation types. b Training and application of CHORD. From a total of 3,824 metastatic tumor samples, 2026 samples were selected for training CHORD. The model outputs the probability of BRCA1-type HRD and BRCA2-type HRD, with the probability of HRD being the sum of these 2 probabilities. The performance of CHORD was assessed via a 10-fold nested cross-validation (CV) procedure on the training samples, as well as by applying the model to the BRCA-EU dataset (543 primary breast tumors) and PCAWG dataset (1,854 primary tumors). Lastly, CHORD was applied to all samples in the HMF and PCAWG dataset in order to gain insights into the pan-cancer landscape of HRD. c The features used by CHORD to predict HRD as well as BRCA1-type HRD and BRCA2-type HRD, with their importance indicated by mean decrease in accuracy. Deletions with 2 to ≥5 bp (i.e. ≥2 bp) of flanking microhomology (del.mh.bimh.2.5) was the most important feature for predicting HRD as a whole, with 1–100 kb structural duplications (DUP_1e03_1e04_bp, DUP_1e04_1e05_bp) differentiating BRCA1-type HRD from BRCA2-type HRD. Boxplot and dots (n = 10) show the feature importance over 10-folds of nested CV on the training set, with the red line showing the feature importance in the final CHORD model. Boxes show the interquartile range (IQR) and whiskers show the largest/smallest values within 1.5 times the IQR.

Fig. 2. Performance of CHORD.

Performance was determined by 10-fold cross-validation (CV) on the HMF training data, as well as prediction on two independent datasets: BRCA-EU (primary breast cancer dataset) and PCAWG (primary pan-cancer dataset). BRCA-EU and PCAWG samples shown here all passed CHORD’s QC criteria (i.e. MSI absent, ≥50 indels, ≥30 SVs if a sample was predicted HRD). a, d, g The probability of HRD for each sample (total bar height) with each bar being divided into segments indicating the probability of BRCA1- (orange) and BRCA2-type HRD (purple). Stripes below the bar plot indicate biallelic loss of BRCA1 or BRCA2. In a, probabilities have been aggregated from the 10 CV folds. b, e, h Receiver operating characteristic (ROC) and c, f, i precision-recall curves (PR) and respective area under the curve (AUC) values showing the performance of CHORD when predicting HRD as a whole (gray), BRCA1-type HRD (orange), or BRCA2-type HRD (purple).

Fig. 3. The genetic causes of HRD in patients from the HMF and PCAWG datasets.

a The bar plot shows the probability of HRD for each patient (total bar height) with each bar being divided into segments indicating the probability of BRCA1-type HRD (orange) and BRCA2-type HRD (purple). 310 patients were predicted HRD while 4812 were predicted HRP by CHORD. b A one-tailed Fisher’s exact test identified enrichment of BRCA1 (q = 9.4e-51), BRCA2 (q = 4.8e-101), RAD51C (q = 5.6e-5) and PALB2 (q = 0.02) biallelic inactivation in CHORD-HRD vs. CHORD-HRP patients (from a list of 781 cancer and HR related genes). Each point represents a gene with its size/color corresponding to the statistical significance as determined by the Fisher’s exact test, with axes indicating the percentage of patients (within either the CHORD-HRD or CHORD-HRP group) in which biallelic inactivation was detected. Multiple testing correction was performed using the Hochberg procedure. c Biallelic inactivation of BRCA2, RAD51C and PALB2 was associated with BRCA2-type HRD, whereas only BRCA1 inactivation was associated with BRCA1-type HRD. Top: BRCA1- and BRCA2-type HRD probabilities from CHORD. Middle: SV contexts (duplications 1–10 kb and 10–100 kb) used by CHORD to distinguish BRCA1- from BRCA2-type HRD. Bottom: The biallelic status of each gene. Samples were clustered according to HRD subtype, and by the impact of a biallelic/monoallelic event (based on ‘P-scores’ as detailed in the methods). Clusters 1, 2, 3, and 5 correspond to patients with identified inactivation of BRCA2, RAD51C, PALB2 and BRCA1, while clusters 4 and 6 correspond to patients without clear biallelic inactivation of these 4 genes. Tiles marked as “Known pathogenic” refer to variants having a “pathogenic” or “likely pathogenic” annotation in ClinVar. “Other” variants include various low impact variants such as splice region variants or intron variants (these are fully specified in Supplementary Data 4). LOH: loss-of-heterozygosity. Only data from samples that passed CHORD’s QC criteria are shown in this figure (MSI absent, ≥50 indels, and ≥30 SVs if a sample was predicted HRD).

Fig. 4. Percentage breakdown of the incidence and genetic causes of HRD in CHORD-HRD patients pan-cancer and by cancer type.

Left and right bars represent the HMF and PCAWG datasets respectively. The vertical split in the figure separates cancer types with (left side) and without (right side) ≥10 CHORD-HRD patients in at least one of the datasets. a Frequency of HRD. Cancer types where no frequency of HRD is displayed contain no data in either the HMF or PCAWG datasets. b The gene deficiency associated with HRD. Bar segments are grouped into BRCA2-type HRD genes (BRCA2, RAD51C, PALB2) and BRCA1-type HRD genes (BRCA1 only). c The likely combination of biallelic events in BRCA1/2, RAD51C or PALB2 causing HRD. d Whether the genetic cause of HRD was purely due to somatic events, due to germline predisposition, or unknown. In c, d, “Unknown” and/or “LOH + unknown” bar segments refer to patients where no clear biallelic loss of the aforementioned BRCA1/2, RAD51C, or PALB2 was identified (i.e. clusters 4 and 6 of Fig. 3c). LOH: loss-of-heterozygosity. Only data from samples that passed CHORD’s QC criteria are shown in this figure (MSI absent, ≥50 indels, and ≥30 SVs if a sample was predicted HRD).