Ovarian Tumor Slides Offer Clues to Faulty DNA Repair

Cells have a high-precision way to repair a dangerous kind of DNA damage, a complete break across both strands. Some ovarian tumors have a defect in it, a state called homologous recombination deficiency (HRD). Those tumors tend to be sensitive to platinum chemotherapy and to PARP inhibitors, drugs that block a second repair route so the damaged cancer cell cannot cope. ^[1] A study published in Nature Communications on October 21, 2026 asks whether a routine tissue slide can reveal HRD. ^[1]

The finding

Sequencing costs and visible tumor features

The cancer in question is high-grade serous tubo-ovarian carcinoma, an aggressive form with few reliable treatment-guiding markers. Existing HRD tests look for mutations in repair genes and for general genome instability. The authors say a well-known sequencing-based tool, HRDetect, has not entered clinics because it depends on costly whole-genome sequencing. ^[1] HRDetect combines six patterns of DNA damage, called mutational signatures, into one score. ^[3]

Hospitals worldwide routinely stain tumor sections with hematoxylin and eosin (H&E), and the stained sections can be scanned into whole-slide images, gigapixel pictures of the tissue. The hope is that HRD leaves visible traces a computer can learn. ^[1] Earlier work included DeepHRD, trained on breast and ovarian cancer slides. ^[4] An ovarian-specific model in another study reached an AUROC of 72% on its own held-out data but only 57% on an outside trial cohort. ^[5] The authors say earlier ovarian results were limited and rarely reported sensitivity or specificity. ^[1]

Prioritizing DNA-repair testing

A cheap first look at HRD could help decide who gets sequencing sooner, especially where sequencing is scarce. In one hospital's data the idea works, and using several slides helps. It does not yet show that image-based calls can guide treatment.

The evidence

Patient-level learning from ovarian slides

The data. The team used 152 patients from one hospital, with 720 slides. Labels came from HRDetect: 82 patients were HRD-positive and 70 negative. Training used 117 patients (493 slides) and validation 29 patients (138 slides). All slides from one patient stayed in the same group, so the model could not memorize a person. For outside checks they used public data from two projects, TCGA (81 patients) and PTRC (93 patients). There, HRD was defined by a different, scar-based score. ^[1]

The model. Each slide is cut into small square patches, color-corrected, and converted into numbers by a pretrained image model. The difficulty is that labels exist per patient, not per patch. Methods built for this are called weakly supervised; CLAM is a well-known example that learns from slide-level labels alone. ^[6] HRDPath combines three views: one module picks out the most informative patches, a transformer (a network that learns which parts of its input relate to each other) reads the whole slide, and a third looks at windows around the picked patches. A patient-level transformer then merges all of a patient's slides. The model predicts both a yes-or-no HRD call and a continuous HRD score. ^[1] A reasonable alternative is outlining tumor regions first, but the authors note one earlier approach that did so is harder to move into practice. ^[1]

The headline numbers. On the 29 held-out patients, HRDPath reached an AUC of 0.846, sensitivity of 0.692 and specificity of 0.938. AUC runs from 0 to 1; 0.5 is chance and 1 is perfect. Sensitivity is the share of true HRD cases caught, so the model missed about 3 in 10. Specificity is the share of non-HRD cases correctly cleared. One of its modules alone scored an AUC of 0.808. ^[1] In the authors' side-by-side comparison, HRDPath beat the other models they tried, and its AUC was 0.096 higher than CLAM-SB's. A large pathology foundation model, UNI, produced weaker results as a feature extractor, which the authors called surprising. ^[1]

More slides helped. Predictions were more consistent when three or more slides per patient were combined. With six or more slides, accuracy topped 0.95, with sensitivity 0.910 and specificity 1.000. ^[1] These are averages over patients who have that many slides, so they come from a subset. The analysis used the full cohort of 152 patients and 720 slides, combining patients used to build the model with the 29 held-out patients. A result of perfect sensitivity and specificity for patients sampled from three body sites rested on only 13 patients, which the authors flag. ^[1]

Outside data. Trained on TCGA, the model reached an AUC of 0.847, specificity of 0.714, and sensitivity of 0.786. With 1 to 2 example images per class used for fine-tuning, an AUC of up to 0.823 appeared on TCGA and PTRC, at best, with specificity of 0.700 and sensitivity of 0.981. ^[1] PTRC holds only five HRD-negative patients, so it tests little. ^[1]

What the model seems to see. HRD-positive tumors tended to have rounder, less stretched nuclei, but the effects were small. Simple morphology measures alone gave a best cross-validated AUC of 0.593, so no single visible feature explains the prediction. ^[1] The authors also found that cases where the image model disagreed with genomic labels differed in survival. These were exploratory comparisons. ^[1]

The labels in the main cohort come from HRDetect, which the authors note is not FDA-approved. The test group is small, treatments varied, and the study did not test PARP inhibitor response. We did not independently replicate any result. ^[1]

Future directions

The authors identify prediction of PARP inhibitor benefit as an open question. Does HRD classification carry information useful for that decision? Testing patient-level image scores against treatment outcomes would extend the classification evidence to an outcome this study does not establish. ^[1]

They also call for validation with multiple HRD assays. Comparing labels from different assays in the same ovarian cohort would help separate assay disagreement from image-model error. The exploratory survival signal in patients whose image and sequencing classifications disagree needs an independent cohort before it can support a prognostic claim. ^[1]