Comprehensive comparison of homologous recombination deficiency predictors in early-stage triple-negative breast...

Homologous recombination (HR) deficiency (HRD) is prevalent in ovarian, prostate, and specific subgroups of breast cancer, particularly triple-negative breast cancer (TNBC). Tumor HRD status can be inferred through DNA-based, RNA-based, functional, or image-based approaches. A comprehensive evaluation of the concordance and discordance among HRD prediction methods derived from these different data modalities has been lacking. In the present study, we systematically compared HRD classifications generated by seven distinct methods within a population-representative early-stage TNBC multi-omics cohort and contrasted them to an FDA-approved assay.

A total of 235 patients from a reported population-based TNBC cohort from southern Sweden profiled by RNA-sequencing, whole genome sequencing, and with available RAD51 foci staining on tissue microarrays and digital whole slide H&E images were included. Seven different HRD classification methods were applied to available data, including sequencing-based (HRDetect and Classifier of HOmologous Recombination Deficiency, CHORD), copy number-based (scarHRD, and copy number signature 17), functional HR (RAD51-FFPE), mRNA-based, and image-based (DeepHRD) methods. Eighteen selected tumors were analyzed with the Myriad myChoice CDx assay for exploratory comparison. Survival analysis was performed using invasive disease-free survival as clinical endpoint in patients treated with adjuvant standard-of-care chemotherapy.

Overall, our results revealed substantial concordance across HRD assessment methods, alongside method-specific discordances attributable to differences in data preprocessing and, importantly, to training strategies that insufficiently account for the well-established clinical and molecular heterogeneity within breast cancer. Sequencing-based methods and scarHRD showed the greatest classification agreement, with discordance to some extent explained by aspects of inadequate tumor cell content, sequencing depth, and fundamental data processing steps (like segmentation). Discordance in mRNA- and image-based classifications appeared associated with molecular subtype features, suggesting that training cohort context may impact performance by incorporating signals (e.g., mRNA expression patterns) that are not specific to HRD status. Despite variation in HRD classification agreement, all seven methods displayed approximately similar prognostic performance in the subset of patients treated with adjuvant chemotherapy.

Collectively, our findings underscore the necessity for rigorous optimization of data processing workflows and threshold definitions to ensure consistency, comparability, and reproducibility across HRD classification platforms.