A Clinically Translatable Multimodal Deep Learning Model for HRD Detection from Histopathology Images

Background: With extensive research and development in the past decade, the affordability of Poly (ADP-ribose) polymerase (PARP) inhibitor therapy has drastically improved. Homologous recombination deficiency (HRD), a key biomarker, has been identified as an important guiding factor for PARP inhibitor therapeutic decisions in breast and ovarian cancer. However, identification of patients who will respond to Poly (ADP-ribose) polymerase (PARP) inhibitor therapy is challenging due to the lack of a unifying morphological phenotype. Current HRD testing via next-generation sequencing (NGS) is tissue-dependent, has high failure rates, misses relevant HRD genes, and involves longer turn-around times. Methods: To overcome these limitations, we developed a multimodal AI model, TRINITY, combining imaging, image-based transcriptome data, and clinico-molecular data, to examine whole-slide images (WSIs) obtained from hematoxylin and eosin (H&E)-stained samples to non-invasively predict HRD status. Results: The TRINITY model, tested on 316 TCGA breast and OV samples, presented a sensitivity of 0.77 and 0.91, NPV of 0.94 and 0.86, PPV of 0.63 and 0.58, specificity of 0.89 and 0.47, and AUC-ROC of 0.91 and 0.72, respectively. The model also yielded a similar outcome in a blind study of 74 samples, with a sensitivity of 81.2, NPV of 0.85, PPV of 0.77, specificity of 0.81, and high AUC-ROC value of 0.89, showing its promising preliminary evidence of predicting HRD status on external cohorts. Conclusions: These findings demonstrate TRINITY's potential as a rapid, cost-effective, and tissue-sparing alternative to conventional NGS testing. While promising, further validation is needed to establish its generalizability across broader cancer types.