Histopathology Based AI Model Predicts Anti-Angiogenic Therapy Response in Renal Cancer Clinical Trial

Area of Science:

• Oncology
• Computational Pathology
• Biomarker Discovery

Background:

• Metastatic clear-cell renal cell carcinoma (ccRCC) lacks predictive biomarkers for treatment response.
• Angiogenesis is a key target, with the RNA-based Angioscore predicting anti-angiogenic (AA) therapy response.
• Current transcriptomic assays for Angioscore are limited by cost, time, and heterogeneity challenges.

Purpose of the Study:

• To develop a deep learning (DL) model for predicting the Angioscore directly from histopathology slides.
• To overcome the interpretability limitations of traditional DL models by generating a visual vascular network.
• To validate the model's reliability across multiple independent cohorts, including a clinical trial.

Main Methods:

• A novel deep learning approach was employed to analyze histopathology images.
• The model was trained to predict the RNA-based Angioscore.
• Interpretability was enhanced by generating a visual vascular network as the basis for predictions.
• The model was validated on diverse patient cohorts, including a clinical trial dataset.

Main Results:

• The DL model accurately predicted the RNA-based Angioscore in independent cohorts (Spearman correlations of 0.77 and 0.73).
• Model predictions revealed biological associations between angiogenesis, tumor grade, stage, and driver mutations.
• The model successfully predicted response to AA therapy in both real-world and clinical trial data.
• Its predictive performance nearly rivaled the ground truth RNA-based Angioscore at a significantly lower cost.

Conclusions:

• This interpretable DL approach provides robust Angioscore prediction from histopathology slides.
• The method offers valuable insights into angiogenesis and anti-angiogenic treatment response in ccRCC.
• This technique presents a practical and cost-effective alternative to current genomic assays for clinical application.