Search

3019 Background: Homologous recombination deficient (HRD) tumors are highly responsive to platinum-based chemotherapy and poly (ADP-ribose) polymerase inhibitor (PARPi) therapy. Pathogenic BRCA-1 and BRCA-2 (BRCA1/2) alterations are key members of the HR DNA repair pathway but genomic instability status, including loss of heterozygosity, telomeric allelic imbalance and large scale state transitions across the genome are also predictive of HRD. HRD testing is currently performed by next generation sequencing which can take 2-4 weeks for results, has a high failure rate, requires significant tissue and is costly. We developed and tested the ability of an AI enabled platform to predict HRD status from the analysis of whole slide imaging of the diagnostic H&E slide. This platform, iPREDICT-HRD is rapid, precise, and cost effective. Methods: The AI engine was trained on 120 H&E slides that were used to identify tumor prior to manual microdisseection for HRD assessment by NGS. Histopathological features were extracted, followed by feature mapping to predict HRD status based on the results of NGS testing. ResNet AI algorithm was trained to segment, annotate and predict HRD status. 10 lac tiles of 256x256 size at 40x magnification were generated per pathological class. 70% of the data set was used for training and 30% for validation of the AI model. Results: Using single blinded clinical samples, iPREDICT-HRD tool detected HRD + ve samples with 99.3% accuracy with 100% sensitivity and 99% specificity in the test set. Patch-level predictions of HRD status demonstrated intra-tumor heterogeneity within the H&E slides. Visual inspection of the heatmap suggested the presence of patches with high predictive ability of HRD status and this outperformed an average HRD score for slides with heterogeneity. Conclusions: AI-enabled prediction of HRD status can be accurately performed on diagnostic H&E slides potentially yielding results quickly and afforadably, even when limited tissue is available for testing.