AI Reads DNA-Repair Clues From Ovarian Cancer Slides. What HRDPath Shows, and What It Does Not Yet Prove
A King’s College London model flags HRD in ovarian cancer from routine slides. Promising, but validation and better sensitivity are needed.
What the study did
Some ovarian cancers cannot repair a certain kind of DNA damage properly. This is called homologous recombination deficiency, or HRD. It matters in treatment: HRD is linked to sensitivity to platinum chemotherapy and to benefit from PARP inhibitors, a class of drugs that work on cells with weak DNA repair. Today HRD is generally found through molecular or genomic testing, which can be costly and is not available everywhere.
Get OncoFrontier’s stories in your inbox
Join Medium for free to get updates from this writer.
Researchers at King’s College London, led by Dr Heba Sailem, built an AI model called HRDPath to look for HRD another way. Instead of analysing DNA, it reads high-resolution digital images of tissue samples that hospitals already prepare for diagnosis. It breaks each slide into small patches and combines two AI approaches: one picks out informative patches and how they relate to one another, and a vision-transformer model weighs the relevant regions across the image. The model was trained on slides from patients whose HRD status was already known from genomic testing. The study was published in Nature Communications and focuses on high-grade serous tubo-ovarian cancer.
What it found
The university reports that HRDPath correctly identified about 91% of HRD-positive cases when it considered several images from the same patient. The paper’s abstract reports a different analysis: an AUC of 0.846, with a sensitivity of 0.692 and a specificity of 0.938, when considering tubo-ovarian and omental sites. The team also found that HRD-positive tumours tended to have rounder cell nuclei and more densely packed cancer cells in certain tumour regions and that the model’s predictions correlated with patient outcomes.
How to read the numbers
These figures come from different ways of counting, so they should not be read side by side as one scoreboard. The AUC of 0.846 summarises how well the model separates HRD-positive from HRD-negative tumours, where 0.5 is chance and 1.0 is perfect. The sensitivity of 0.692 means that in that analysis roughly three in ten HRD-positive tumours were missed, while the specificity of 0.938 means few HRD-negative tumours were wrongly flagged. That profile fits a screening or triage role, where a positive result is fairly reliable, but a negative result does not rule HRD out, better than a replacement for genomic testing.
What it could be used for
If validated, an image-based test could widen access to precision oncology, especially where genomic testing is expensive or unavailable. It could also point clinicians toward patients who might benefit from platinum chemotherapy or PARP inhibitors, using slides already made for diagnosis. Treatment decisions would still rest with the oncology team and standard testing.
What is not yet known
The researchers say validation in larger patient populations is still needed. Questions include whether HRDPath performs the same across slide scanners, staining methods, laboratories, and patient groups, and whether it matches genomic HRD testing closely enough to guide treatment. It is also unknown whether using it would improve treatment choices or outcomes. The work is a research model, not an approved diagnostic.
The throughline
Pathology slides already exist for nearly every patient. The appeal of HRDPath is that it tries to read molecular information out of them. The test of that idea is not one good number but consistent performance on other hospitals’ slides and evidence that it helps patients get the right treatment.
This article is for general informational and educational purposes only and does not constitute medical advice. The technologies discussed here are in various stages of research, development, or early clinical evaluation, and none replace the judgement of a qualified oncologist. Always consult a licensed clinician regarding diagnosis or treatment decisions.
Related Reading
This article is part of OncoFrontier’s ongoing coverage of AI & Technology in Oncology, an initiative by CancerFax.
PARP inhibitors are one class of targeted therapy that depends on biomarker testing. CancerFax’s targeted therapy guide explains how biomarker testing is used to match treatments:
