Precision oncology aims to customize cancer treatment by identifying and targeting molecular defects in a patient's tumor. In breast and ovarian cancers, defects in homologous recombination repair (HRR) pathways involving genes such as BRCA1/2, PALB2, and RAD51 make tumors more responsive to treatments like platinum salts and PARP inhibitors. Traditional detection methods for homologous recombination deficiency (HRD) involve expensive and time-consuming genomic testing, such as Myriad myChoice CDx and FoundationOne CDx, which analyze genomic instability and BRCA1/2 status.
Despite their effectiveness, these methods are costly (around $3,000 per patient) and have long processing times (4-6 weeks), limiting their widespread use, especially in resource-constrained settings. Routine tumor biopsies, processed for histopathological examination, are more accessible and timely but typically lack the genomic information necessary for precision oncology.
Recent AI advancements have shown promise in predicting genomic changes from digital images of H&E-stained slides. DeepHRD, a new AI-based classifier, addresses these limitations by detecting HRD directly from digitized biopsy slides. Trained and validated on data from The Cancer Genome Atlas (TCGA), DeepHRD outperforms current molecular tests in detecting HRD and predicting treatment outcomes, offering a faster, more cost-effective, and broadly applicable solution for precision oncology.
DeepHRD employs soft labeling during training to avoid overconfidence and uses weakly supervised multiple-instance learning. It mimics the diagnostic process of pathologists by performing initial predictions at low magnification (5×) and refining them at higher magnification (20×) within regions of interest (ROIs). Using ResNet18 convolutional neural networks for feature extraction and techniques like principal component analysis and k-means clustering for ROI selection, the model ensures accuracy and prevents overfitting through random dropout and multiple inference passes. The final prediction score, along with confidence intervals, aids in making treatment recommendations based on a single diagnostic whole-slide image (WSI), showing improved detection of HRD and better clinical outcomes compared to traditional methods.
The development of DeepHRD demonstrates the feasibility of using AI to detect HRD in breast and ovarian cancers from digital H&E slides. The tool has shown better prediction of complete response (CR) and progression-free survival (PFS) in an external cohort of platinum-treated metastatic breast cancer patients compared to existing molecular tests, capturing more BRCA1/2 wild-type tumors responsive to platinum therapy. By incorporating transfer learning, DeepHRD classified more high-grade serous ovarian cancers (HGSOCs) as HRD-positive, correlating with significantly higher overall survival (OS). DeepHRD’s rapid, cost-effective HRD detection supports treatment decisions and clinical trial enrollment, particularly in resource-constrained settings. While further validation in specific trials is needed, DeepHRD could complement existing molecular HRD assays, enhancing precision oncology.
Ref : Deep Learning Artificial Intelligence Predicts Homologous Recombination Deficiency and Platinum Response From Histologic Slides | Journal of Clinical Oncology (ascopubs.org)
