Cancer Research · uk
Reading DNA Repair Weaknesses in Ovarian Cancer Tissue Slides: HRDPath Explores New Clues for Treatment Stratification
Routine pathology images may hold clues to drug selection. HRDPath aims to identify homologous recombination repair deficiency in tumors, with performance improving when multiple slides are combined. Whether it can support clinical decisions still requires larger-scale validation.
Clues to ovarian cancer treatment sometimes lie in cancer cells’ ability to repair DNA. Identifying this weakness in tissue slides already obtained at diagnosis could allow more patients to receive further testing and treatment assessment earlier. HRDPath, an artificial intelligence model developed by researchers at King’s College London, is attempting to read this information from microscopic images. The study was published in *Nature Communications* on October 21.
The study focused on highly aggressive high-grade serous tubo-ovarian cancer. Some of these tumors have “homologous recombination deficiency” (HRD), meaning that an important pathway for repairing DNA damage has failed, potentially in association with abnormalities in genes such as BRCA1 and BRCA2. This deficiency may also provide an opening for treatment, making cancer cells more sensitive to platinum-based chemotherapy or PARP inhibitors. However, determining HRD status usually relies on molecular testing, with cost and accessibility posing barriers.
HRDPath used cases with existing genomic analysis results to learn associations between routinely stained tissue images and HRD status. The study covered three datasets: the main UWOV dataset included 152 patients and 720 slides, while independent datasets of 81 and 93 patients were used for evaluation. Slides from the same patient were assigned to the same group to avoid inflating performance through images from the same individual appearing in both the training and test data.
Among the 29 patients in the main test set, the model had an area under the curve (AUC) of 0.846, a sensitivity of 69.2%, and a specificity of 93.8%. Sensitivity reflects the model’s ability to identify HRD-positive cases, while specificity reflects its ability to correctly rule out negative cases. These results indicate that although the model relatively rarely misclassified negative cases as positive, it still missed some tumors that truly had the repair deficiency.
Performance improved when more slides were combined. The “approximately 91%” mentioned in the university announcement and the report republished by Medical Xpress refers to the proportion of HRD-positive cases identified through analysis of multiple images. The paper further explains that this sensitivity was achieved in an analysis using more than six slides per patient and cannot be treated as overall accuracy across all settings. The number of slides and sampling conditions are therefore also part of interpreting the model’s performance.
The researchers also tried to understand what the model was seeing. According to the university, certain regions of HRD-positive tumors had rounder cell nuclei and more densely packed cancer cells. These associations suggest that abnormalities in DNA repair may be accompanied by recognizable differences in tissue morphology, but appearance alone cannot yet establish an individual patient’s molecular status.
A more practical potential application is to help prioritize molecular testing and provide supplementary information for treatment stratification. Before entering clinical practice, HRDPath still needs validation in larger and more diverse patient populations to determine whether it can maintain its performance under slide preparation and imaging conditions at different hospitals. The impact of missed cases must also be assessed. The current results support further research but are not yet sufficient for image-based predictions to replace molecular testing or independently determine drug selection.