AI-Derived Biomarker From Routine Pathology Images Predicts Individual Benefit From PARP Inhibitors in Ovarian Cancer
Key Insights
A deep learning model using routine H&E pathology slides from the phase III PAOLA-1 trial predicts individual patient benefit from maintenance olaparib plus bevacizumab versus bevacizumab alone.
The model's Estimated Treatment Improvement (ETI) score showed a strong treatment–biomarker interaction (HR 0.36, 95% CI 0.22–0.59), comparable to HRD (search) status.
ETI remained independently associated with outcomes after adjustment for clinical variables and HRD (search) status, and further stratified benefit within both HRD-positive and HRD-negative tumors.
A new study published in NPJ Digital Medicine, part of the Nature Portfolio, demonstrates that artificial intelligence can extract clinically meaningful predictive information from routine diagnostic pathology slides to refine treatment selection in ovarian cancer (search). Using data from the phase III PAOLA-1 randomized trial, researchers from DiaDeep (search) and collaborating clinical institutions developed a treatment-aware deep learning model that estimates each patient's individual benefit from maintenance olaparib plus bevacizumab versus bevacizumab alone.
The model analyzes treatment-naive primary tumor whole-slide images from hematoxylin and eosin (H&E) stains—the standard diagnostic pathology preparation available in routine clinical practice worldwide.
Addressing Heterogeneity Beyond HRD (search) Status
Homologous recombination deficiency (HRD (search)) status is currently a key biomarker for selecting ovarian cancer (search) patients for PARP inhibitor therapy. However, response heterogeneity remains substantial even among patients sharing the same HRD classification. As Jean-Sébastien Frenel, Professor of Medical Oncology at the University of Nantes and lead author, noted: "HRD does not tell the whole story, and patients with the same HRD status can experience very different outcomes."
The research team sought to determine whether routine diagnostic H&E slides could provide additional predictive information beyond what molecular testing currently offers.
Model Design and Key Findings
The deep learning model generates a continuous Estimated Treatment Improvement (ETI) score by combining image-derived features with randomized treatment assignment and progression-free survival data from the PAOLA-1 trial. This counterfactual approach allows the model to estimate what each patient's outcome would be with and without the addition of olaparib.
The study's key findings include a strong treatment–biomarker interaction for the ETI score, with a hazard ratio of 0.36 (95% CI 0.22–0.59)—a magnitude comparable to that observed for HRD (search) status itself. ETI remained independently associated with outcomes after adjustment for clinical variables and HRD status, confirming that the image-derived biomarker captures predictive information beyond established molecular markers.
Notably, the model further stratified predicted treatment benefit within both HRD (search)-positive and HRD-negative tumor subgroups, suggesting that AI-derived histology biomarkers could refine patient selection even among populations already defined by molecular testing.
Biological Correlates
Attention maps generated by the model suggested that multifocal tumor-infiltrating lymphocytes may be associated with greater benefit from olaparib, providing a potential biological rationale for the model's predictions and linking the image-based biomarker to the tumor immune microenvironment.
Clinical and Economic Implications
Beyond improving precision medicine, the researchers emphasize that better treatment selection carries broader implications. "Better treatment selection means fewer unnecessary treatments, reduced toxicity, lower healthcare costs, and more sustainable cancer care," Frenel stated. The ability to derive predictive biomarkers from routine pathology slides—without requiring additional tissue, assays, or cost—positions this approach as a potentially scalable complement to molecular testing.
Giuseppe Caruso, Gynecology Oncology Specialist at the European Institute of Oncology and study contributor, described the work as "a great example of computational pathology moving closer to clinical impact."
The study authors—including Jean-Sébastien Frenel, Pierre-Etienne Heudel, Emmanuelle Guinaudeau, Céline Bossard, Carmela Pisano, Giuseppe Caruso, Joseph Rynkiewicz, Sanae Salhi, Yahia Salhi, Jérôme Chetritt, Eric Pujade-Lauraine, and Isabelle Ray-Coquard—acknowledged the patients of the PAOLA-1 trial whose participation made the research possible.