AI predicts DNA repair defects from ovarian cancer tissue images in new study

An AI model could help identify a treatment-relevant weakness in ovarian cancer without additional genetic testing. But the promise rests partly on a held-out test of just 29 patients. Published October 21 in Nature Communications, the HRDPath study connects routine tissue images with faulty DNA repair—a finding researchers say needs validation in larger patient populations.

A repair weakness that helps guide treatment

The research focuses on high-grade serous tubo-ovarian carcinoma, an aggressive cancer arising in the ovaries or fallopian tubes. Some of these tumors have homologous recombination deficiency, or HRD: an impaired ability to repair DNA damage. That weakness can make them more vulnerable to platinum-based chemotherapy and drugs called PARP inhibitors.

Identifying HRD can therefore help doctors choose treatments. Existing tests examine genetic changes, but the study’s authors describe them as expensive, time-consuming and variable. Their alternative starts with hematoxylin and eosin-stained tissue slides, a standard preparation already used in clinical practice. The goal is to extract a molecular clue from an image rather than order another molecular test.

Learning from several slides, not one isolated view

HRDPath breaks large digital slide images into smaller patches and converts them into features for analysis. It combines two AI approaches: one identifies informative patches and their relationships; the other uses a vision transformer to weigh relevant regions across the image. The system does not require researchers to mark out tumor regions precisely before making its prediction.

Its central design choice is patient-level analysis. HRDPath combines information from multiple slides belonging to one person and learns to predict both an HRD-positive or negative result and a continuous HRD score. Training used patients whose HRD status had already been established through genomic analysis; HRDetect supplied the reference labels for the study’s UWOV cohort.

The in-house UWOV collection contained 152 patients. Model development used a subset of 146: 117 for training and 29 for held-out evaluation. All slides from any one patient stayed on the same side of that split, helping prevent the model from learning a person’s distinctive patterns and then encountering them again during testing.

69.2%Sensitivity

The model correctly identified 69.2% of HRD-positive patients in the 29-patient test set.

93.8%Specificity

The model correctly identified 93.8% of HRD-negative patients in that test set.

Why the detection figures need context

The benchmark shows an uneven result: the model was better at recognizing HRD-negative patients than finding HRD-positive ones. The authors also report that one training strategy improved specificity and overall performance at the expense of sensitivity. Those measures capture different errors, rather than a single interchangeable score for how well the model works.

A King’s College London account published by Medical Xpress highlights approximately 91% detection of HRD-positive cases when several images from the same patient were considered. That figure should not be substituted for the held-out test’s sensitivity. The paper separately examines how predictions change across combinations of one to 11 slides per patient, using its broader 152-patient collection.

Testing also extended beyond UWOV to two independent public cohorts: TCGAOV, with 81 patients, and PTRC, with 93. The authors report comparable performance there. Their image analyses also linked HRD-positive tumors to distinct cellular and tissue features, including changes in cell nuclei. Those findings support the study’s premise that a DNA repair defect can leave visible clues in tissue.