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

Area of Science:

• Oncology
• Computational Biology
• Pathology

Background:

• Poly (ADP-ribose) polymerase (PARP) inhibitor therapy is increasingly affordable and crucial for breast and ovarian cancers.
• Homologous recombination deficiency (HRD) is a key biomarker for PARP inhibitor response, but its identification is challenging.
• Current next-generation sequencing (NGS) for HRD testing is tissue-dependent, has high failure rates, and long turnaround times.

Purpose of the Study:

• To develop a non-invasive method for predicting HRD status.
• To overcome the limitations of current tissue-based HRD testing methods.
• To create a rapid, cost-effective, and tissue-sparing alternative for guiding PARP inhibitor therapy.

Main Methods:

• Development of a multimodal AI model named TRINITY.
• TRINITY integrates imaging, image-based transcriptome, and clinico-molecular data.
• Whole-slide images (WSIs) from H&E-stained samples were analyzed to predict HRD status.

Main Results:

• TRINITY achieved high performance metrics (e.g., AUC-ROC of 0.91 and 0.72) in TCGA breast and ovarian cancer samples.
• The model demonstrated promising results in an external blind study (AUC-ROC of 0.89).
• TRINITY showed potential for predicting HRD status across different cohorts.

Conclusions:

• TRINITY shows potential as a rapid, cost-effective, and tissue-sparing alternative to conventional NGS testing for HRD.
• The AI model may aid in identifying patients who will benefit from PARP inhibitor therapy.
• Further validation is required to confirm TRINITY's generalizability across diverse cancer types.