Predicting BRCA mutation and stratifying targeted therapy response using multimodal learning: a multicenter study.

Area of Science:

• Oncology
• Genetics
• Computational Pathology

Background:

• BRCA1/2 gene status is critical for cancer treatment decisions, but genetic testing is often inaccessible.
• Not all patients benefit from poly (ADP-ribose) polymerase inhibitors (PARPi), necessitating improved risk stratification.
• Developing predictive models for BRCA1/2 status and PARPi response is crucial for personalized cancer therapy.

Purpose of the Study:

• To develop and validate a multimodal model for predicting BRCA1/2 gene status using histopathological images.
• To assess the model's ability to predict prognosis and response to PARPi treatment in various cancer types.
• To identify morphological features associated with BRCA1/2 mutations for model interpretability.

Main Methods:

• A multi-instance attention model (MIAM) was developed to detect BRCA1/2 status from H&E images.
• The MIAM-C model integrated tissue, cell, and clinical features for enhanced prediction.
• Model performance was evaluated using AUC and Kaplan-Meier analysis across three independent cohorts (ovarian, breast, prostate, pancreatic cancers).

Main Results:

• The MIAM-C model demonstrated superior performance in identifying BRCA1/2 genotype compared to the MIAM model.
• High-attention regions, including high-grade tumors and lymphocytic infiltration, correlated with BRCA1/2 mutations.
• MIAM-C accurately predicted PARPi therapy response and served as an independent prognostic factor for BRCA1/2-mutant ovarian cancer.

Conclusions:

• The MIAM-C model accurately detects BRCA1/2 gene status from pathological images.
• This multimodal approach effectively stratifies prognosis for patients with BRCA1/2 mutations.
• The findings support the potential of AI-driven pathological analysis for personalized cancer treatment.