AI-driven HRD biomarker prediction using OncoPredikt® deep learning inference engine trained on 1,892 whole slide images (WSIs) from TCGA datasets, split into 80/10/10 train-validation-test sets
ResNet50 deep learning model processes H&E-stained diagnostic slides via 256×256 tile segmentation and quality control, eliminating need for expensive multi-technique HRD testing (~USD 4,000)
HRD aggregate score derived from loss of heterozygosity (LOH), large-scale genomic instability, and telomeric allelic imbalance — validated against BRCA1/2 and HRR gene status
99.3% accuracy, 100% sensitivity, and 99% specificity achieved on 120 external Ambry Genetics WSIs, with ROC/AUC = 1.00 demonstrated on receiver operating characteristic curve
Sample-level prediction table shows consistent concordance between standard HRD results and AI predictions across 10 patient samples with turnaround times of 2.5–6.4 minutes
Clinical advantage over conventional HRD testing: faster, affordable, operable on limited tissue samples, directly applicable to routine diagnostic H&E slides for PARP inhibitor patient selection Read Abstract