Migrating the SNP array-based homologous recombination deficiency measures to next generation sequencing data of breast cancer

1Reliable quantification of homologous recombination deficiency of human tumor biopsies, especially in the case of ovarian and breast cancer, is expected to identify patients that are particularly sensitive to platinum or PARP inhibitor-based therapy. Before the widespread introduction of next generation sequencing (NGS) to characterize tumor biopsies, SNP arrays were used to identify large-scale genomic aberrations associated with homologous recombination deficiency, often induced by the loss of BRCA1 or BRCA2 function. Three such measures were identified: telomeric allelic imbalance (HRD-TAI score), loss of heterozygosity profiles (HRD-LOH score), and large-scale state transitions (HRD-LST score). These three measures have also been combined into a single summary measure of HR deficiency. The HRD-LOH score has also become an integral part of a recently published, whole-genome sequencing-based measure of homologous recombination deficiency, HRDetect. These measures, along with functional assays, showed promise to identify HR-deficient cases and thus predict response to platinum or PARP inhibitor therapy.,, Since NGS has become the main genomic characterization method of cancer biopsies, it has become essential to migrate the SNP array-based methodology to NGS-based platforms.

2TCGA breast cancer biopsies have been both SNP array profiled and subjected to NGS allowing a direct comparison.

3We found good correlation between the SNP array-based and NGS-based HRD scores (Fig. 1). When comparing the results of the scarHRD R package to SNP array-based measurements, we found the following Pearson correlation coefficients: number of telomeric allelic imbalances (NtAI): r = 0.84 (R2 = 0.70, adjusted R2 = 0.70, p < 2.2e–16), large-scale transition (LST) r = 0.79 (R2 = 0.62, adjusted R2 = 0.62, p < 2.2e–16) loss of heterozygosity (HRD−LOH) r = 0.73 (R2 = 0.53, adjusted R2 = 0.52, p < 2.2e–16). These three measures are often combined for diagnostic purposes and in HRDetect. Therefore, we also compared the sum of the three scores across the two platforms (HRD sum): r = 0.87 (R2 = 0.75, adjusted R2 = 0.75, p < 2.2e–16) (Fig. 1). The artificial reduction of coverage to 30× did not affect this correlation (Supplementary Material, Figure S7-S8). The BRCA1/2-mutated samples showed significantly higher NGS-based HRD-sum values (Fig. 2, Supplementary Figure S6). The predictive value of HRD-sum, measured as AUC value of the corresponding ROC curve, was 80.8% (Supplementary Figure S2).

4There was no significant difference in SNP versus WXS-based estimation of tAI, LST, and HRD-sum, but the number of LOH events were significantly lower in the WXS-based estimation (p = 0.012, Kolmogorov–Smirnov test). This could be attributed to differences in segmentation algorithm (the more segmented the WXS data is the lower number of LOHs that are called) or to low sample quality, coverage. However, when comparing the ROC curves for BRCA1/2 status of the SNP-based and WXS-based HRD-score, there was no significant difference between the SNP array-based and NGS-based methods. (Supplementary Figure S3).

5According to our expectations and previous results the BRCA1/2-deficient cases showed higher values for each of the four scores (Supplementary Figure S4-S5).

6The sum of the three HRD scores showed good correlation across the two platforms. Thus in more advanced NGS-based HR deficiency measures such as HRDetect, the SNP array-based step could be replaced by an NGS-based estimate of the HR deficiency scores.

7Based on receptor status determined by immunohistochemistry, 139 paired tumor and normal samples of the TCGA breast cancer cohort could be classified as triple-negative breast cancer. From these patients 95 had Affymetrix SNP 6.0 array-based HRD estimates (LOH, TAI, LST), previously published by our group. In this publication we present the scarHRD R package (https://github.com/sztup/scarHRD) which estimates the level of the three HR deficiency measures using NGS data.

8A sample’s LOH score is the total number of LOH regions across the entire genome that are larger than 15 Mb but do not cover whole chromosomes. In the original publication this 15 Mb lower limit for LOH was determined by comparing SNP array profiles between BRCA mutant and BRCA wild-type cases. We performed a similar analysis using NGS data and found that the original 15 Mb cutoff performed best in this case as well (Supplementary Figure S1).

9The LST is defined as a chromosomal break between adjacent regions of at least 10 Mb, with a distance between them not larger than 3 Mb.

10The number of telomeric allelic imbalances is the number of AIs (the unequal contribution of parental allele sequences with or without changes in the overall copy number of the region) that extend to the telomeric end of a chromosome.

11Allele-specific copy number estimation is a crucial part of estimating HR deficiency. As previously shown, allele-specific copy number estimation from NGS data performed using the Sequenza R package show high agreement with SNP array-based copy number profiles. The scarHRD package is, therefore, able to use Sequenza preprocessed files as well as other allele-specific segmentation files in the same format.

12As it has been previously shown that in ovarian cancer the sum of the genomic scar scores is elevated in BRCA-deficient cancers, an additional aim of our study was to compare the unweighted numeric sum of LOH, tAI, and LST, called here HRD-sum, to the BRCA1/2 status of the patients. A sample was classified as BRCA-deficient if (1) there was a deep deletion of BRCA1/2, (2) a germline and a somatic mutation in BRCA1/2 with LOH, or (3) if LOH had co-occurred with promoter methylation in one of the BRCA1/2 genes. The somatic mutation status (mutations with likely pathogenic function) and methylation data was acquired from the TCGA data portal. The germline mutation status was determined using HaplotypeCaller, and was annotated with Intervar, likely pathogenic mutations and frameshift insertion/deletion with unknown significance were used in our analysis. LOH was determined using Sequenza’s allele-specific segmentation results (Supplementary Table S1).

13The data sets generated during the current study are available from the corresponding author on reasonable request.

14The code/algorithm for performing the experiments is available for download at https://github.com/sztup/scarHRD.