HRDPath

Homologous recombination deficiency (HRD) is one of the few actionable biomarkers in high-grade serous tubo-ovarian carcinoma (HGSOC), guiding the use of platinum chemotherapy and PARP inhibitors. Establishing HRD status currently requires genomic assays that are costly, slow, and inconsistent across vendors, which limits testing in exactly the settings where treatment decisions are most time-pressured. HRDPath predicts genomically derived HRD status directly from routine H&E-stained whole-slide images, converting a slide that pathology departments already produce into a candidate digital biomarker.

Developed by the Sailem lab at King's College London with collaborators at the University of Washington, the University of Pittsburgh, University College London Hospital, and Cambridge, HRDPath differs from earlier weakly supervised HRD classifiers in two ways. It reasons at the level of the patient rather than the individual slide, aggregating evidence across every slide available for a case, and it fuses three complementary views of the tissue instead of relying on a single attention-pooling mechanism. Both choices target the central difficulty of the task: the morphological signature of HRD is diffuse, spread across scales, and not confined to a handful of diagnostic regions.

The model was posted as a bioRxiv preprint in September 2025 under a CC BY license and has not yet been peer reviewed.

#Key Features

• Patient-level prediction: A transformer encoder aggregates slide embeddings across all slides belonging to a patient, and accuracy increases significantly when multiple slides per patient are available — matching how ovarian cancer cases are actually sampled in practice.
• Three-stream tissue representation: A DSMIL branch locates coarse regions of interest, a global vision transformer summarizes the whole slide, and a local window transformer captures fine-grained structure within the regions DSMIL selects.
• Few-shot transfer to new cohorts: Starting from a fixed checkpoint trained on the internal University of Washington ovarian cohort, fine-tuning on one to three example images per class transfers the model to TCGA-OV and PTRC-HGSOC, reaching an AUC of 0.82 and specificity up to 0.98.
• High specificity: The reported specificity of 0.938 is the operating characteristic that matters clinically, since a false HRD-positive call routes a patient toward a therapy unlikely to benefit them.
• Interpretability pipeline: Attention maps are paired with cell-level morphometrics extracted using HoVer-NeXt, surfacing the cellular and tissue-level changes associated with HRD-positive tumours rather than reporting a bare score.

#Technical Details

Whole-slide images are tiled through the CLAM preprocessing pipeline at 15x magnification, colour-normalized, and embedded with RetCCL, a self-supervised pathology encoder trained with clustering-guided contrastive learning. The resulting patch features feed the three streams, whose outputs are combined and passed to the patient-level aggregator. The transformer encoders use depth 2, 8 attention heads, and dimension 1024, with k = 8 local windows of width w = 32, dropout 0.1, learning rate 1e-4, weight decay 1e-5, batch size 1, and Kaiming uniform initialization under fixed seeds. Colour normalization measurably improved cross-dataset generalizability.

Across the three cohorts, HRDPath reports an AUC of 0.846 and a specificity of 0.938, exceeding previously published H&E-based HRD prediction for HGSOC by 0.09 AUC. It was benchmarked against DeepHRD's ResNet-18 ensemble, ZoomMIL, and both CLAM-SB and CLAM-MB. Substituting UNI as the patch feature extractor performed significantly worse than RetCCL, a useful counterexample to the assumption that the largest available pathology foundation model is always the right encoder for a downstream biomarker task.

#Applications

The immediate use case is triage: flagging likely HRD-positive HGSOC cases from slides already scanned at diagnosis, either to prioritize confirmatory genomic testing or to inform therapy planning where genomic testing is unavailable or delayed. The few-shot transfer results matter here, because a laboratory adopting the model can adapt it with a handful of locally stained examples rather than a labelled institutional cohort. A variant trained only on tubo-ovarian slides also predicts HRD at other anatomical sites, suggesting the learned morphology is not purely site-specific.

#Impact

HRDPath is an argument that patient-level architecture and encoder choice, not encoder scale, drive performance on genomic-biomarker prediction from histology. Its limitations are concrete: it addresses a single tumour type, the internal University of Washington cohort is not publicly available (though TCGA-OV and PTRC-HGSOC are), and no code or trained weights have been released, so independent reproduction currently depends on reimplementation from the preprint.