AI Model Reads Routine Tumor Slides to Flag Hidden DNA-Repair Flaws in Ovarian Cancer

Researchers at King's College London have built an artificial intelligence model that can estimate, from routine microscope images of tumor tissue, whether an ovarian cancer carries a DNA-repair defect that determines which drugs are likely to work against it. The tool, described in a paper published Wednesday in Nature Communications, is aimed at a biomarker called homologous recombination deficiency, or HRD, which currently requires genomic testing that is costly and not widely available outside major cancer centers.

The model, named HRDPath, was developed by a team led by Chun-I Wu and Dr. Heba Sailem, a reader in computational pathology at King's College London's Institute for Artificial Intelligence, working with collaborators at University College London Hospital, the University of Pittsburgh School of Medicine, the University of Washington's Division of Gynecologic Oncology and the University of Cambridge's MRC Cancer Unit. The paper's senior author list includes Serena Nik-Zainal, a Cambridge geneticist known for her work cataloguing mutational patterns in tumors, and Elizabeth Swisher, a University of Washington gynecologic oncologist who has studied HRD testing for more than a decade.

A Digital Shortcut to an Expensive Test

HRD testing matters most in high-grade serous tubo-ovarian carcinoma, the most common and lethal form of epithelial ovarian cancer, accounting for an estimated 70 to 80 percent of cases. Tumors with HRD — often but not always linked to mutations in the BRCA1 or BRCA2 genes — tend to respond well to platinum-based chemotherapy and to a class of drugs called PARP inhibitors. Tumors without the defect generally do not.

Determining HRD status today typically means sending a tissue sample out for genomic sequencing, a process the paper's authors describe as variable in quality, expensive and slow. HRDPath instead analyzes the whole-slide pathology images that are already captured as part of a standard cancer diagnosis, using deep learning to pick out patterns invisible to the naked eye. According to the Medical Xpress account of the research, the system was trained on tissue samples whose HRD status had already been confirmed through genomic testing, then checked against new samples it had not seen before.

The Numbers

According to the published abstract, HRDPath was validated across three separate patient cohorts and achieved an area under the curve of 0.846 — a standard statistical measure of diagnostic accuracy, where 1.0 would mean perfect discrimination and 0.5 would be no better than chance. The authors say this "significantly" outperformed previously reported attempts to predict HRD status from pathology images alone. When the model combined multiple whole-slide images from the same patient with additional optimization, both its sensitivity and its specificity reached 0.90, meaning it correctly flagged about nine in ten true HRD-positive cases while also correctly clearing about nine in ten HRD-negative ones. Medical Xpress reported the same result in plainer terms: the model correctly identified roughly 91 percent of HRD-positive cases when several images per patient were available.

The team also built an interpretability pipeline — a set of tools meant to show what the model is actually "looking at" — and found that HRD-positive tumors tended to have rounder cell nuclei and more densely packed cancer cells in certain tissue regions, a pattern consistent with what pathologists have separately observed about how these cancers behave.

"We are unlocking information hidden in microscopic images of cancer tissue that could transform how we match patients to treatments," Dr. Heba Sailem, a reader in computational pathology at King's College London, said of the research, according to Medical Xpress.

Why It Matters for Patients

Ovarian cancer remains one of the harder cancers to treat precisely because it is often diagnosed late and has comparatively few validated biomarkers to guide therapy. The American Cancer Society estimates that about 21,010 women in the United States will be diagnosed with ovarian cancer in 2026 and about 12,450 will die of the disease. Federal data compiled by the National Cancer Institute's SEER program put the five-year relative survival rate at 52 percent and the median age at diagnosis at 63, underscoring both the disease's lethality and the narrow window clinicians have to get treatment right the first time.

Existing genomic HRD tests, such as the one currently used as a companion diagnostic for certain PARP inhibitors, require specialized laboratory infrastructure that is concentrated in wealthier health systems. If a tool like HRDPath can eventually be validated for clinical use, the King's College London team argues it could extend HRD testing to clinics that already have a pathology lab and a scanner but lack access to genomic sequencing, without requiring additional tissue or blood samples from patients. That argument — lower cost, faster turnaround, broader geographic reach — is the central claim driving interest in the paper, though it remains a projection rather than a demonstrated outcome.

Caveats and What Comes Next

The researchers themselves are cautious about how close the tool is to the clinic. Even in a strong validation study spanning three cohorts, an AUC of 0.846 and a sensitivity/specificity ceiling of 0.90 leave a meaningful margin of error — roughly one in ten cases misclassified in each direction — in a setting where getting the answer wrong could mean a patient is denied a drug that would have helped her, or given one that will not. Medical Xpress reported that the authors themselves say larger patient studies are needed before the tool could be used in routine clinical decision-making, and the published paper frames HRDPath as "a foundation for an accessible and scalable digital biomarker" rather than a finished product.

Independent, larger-scale clinical validation — ideally including patients treated and monitored prospectively, rather than retrospectively analyzed slides — would be the next step before regulators or hospitals would consider using an image-based tool in place of, or alongside, genomic testing. The King's College London group has not announced a timeline for such a trial. For now, the paper stands as an early but methodologically serious demonstration that routine pathology images, already sitting in hospital archives worldwide, may carry more actionable genetic information than oncologists have been able to extract from them until now.