Abstract LB391: Deep learning AI predicts HRD and platinum response from histologic slides

Abstract

Abstract Background: Cancers with homologous recombination deficiency (HRD) can benefit from platinum salts and PARP inhibitors. Standard diagnostic tests, including FDA-approved companion diagnostics, for detecting HRD require molecular profiling, which is not universally available with global testing rates lowest among minority, rural, and other underserved populations. Methods: We trained DeepHRD, a deep-learning platform for predicting HRD from hematoxylin and eosin (H&E)-stained histopathological slides, using primary breast (n=1,008) and ovarian (n=459) cancers from The Cancer Genome Atlas (TCGA). DeepHRD was compared to four standard HRD molecular tests using breast (n=349) and ovarian (n=141) cancers from multiple external and independent datasets, including clinical cohorts with platinum complete response, progression-free survival (PFS) and overall survival (OS) endpoints. Results: DeepHRD detected HRD from held-out H&E-stained breast cancer slides in TCGA with an AUC of 0.81 ([0.77-0.85]; 95% confidence interval). This performance was confirmed in two independent primary breast cancer cohorts (AUC=0.76; [0.71-0.82]). In an external platinum-treated metastatic breast cancer cohort, samples detected as HRD had a higher complete response (AUC=0.76; [0.54-0.93]) with 3.7-fold increase in median PFS (14.4 versus 3.9 months; p-value=0.0019) and hazard ratio (HR) of 0.45 (p=0.0047) after correcting for PAM50 molecular subtype and age at diagnosis. This deep-learning classifier outperformed four genomic HRD tests used in the clinic, including standard HRD score, BRCA1/2, 26-HR gene panel and single-base substitution signature 3 (SBS3). Multiresolution spatial mapping identified morphological features utilized by DeepHRD for detecting HRD, notably enriched for neoplastic and necrotic tissues, and a higher macrophage density. Through transfer learning to high-grade serous-ovarian cancer, DeepHRD-positive samples exhibited better overall survival after TCGA first-line (HR=0.46; p=0.030) and an external neoadjuvant (HR=0.49; p=0.015) platinum-treated cohorts. Conclusion: In summary, DeepHRD exhibits consistent hazard ratios ranging from 0.45 to 0.49 across the three clinical cohorts and captures 1.8- to 3.1-fold more HRD-positive breast and ovarian cancer patients. DeepHRD-positive breast cancer patients that received platinum exhibited better complete response and PFS. Similarly, DeepHRD-positive platinum-treated ovarian cancer patients had a better OS. DeepHRD’s ability to detect HRD from digital H&E slides provides an important precision oncology tool that can be utilized in resource-constrained and underserved areas where genomic testing is generally not existent. Citation Format: Erik N. Bergstrom, Ammal Abbasi, Marcos Diaz-Gay, Loïck Galland, Sylvain Ladoire, Scott M. Lippman, Ludmil B. Alexandrov. Deep learning AI predicts HRD and platinum response from histologic slides [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 2 (Late-Breaking, Clinical Trial, and Invited Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(7_Suppl):Abstract nr LB391.