Predicting Protein Cascade Expression from H&E Images.

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Biomedical data analysis

Background:

• Protein expression in oncogenic pathways is key to cancer development.
• Predicting downstream protein signals is essential for understanding cancer progression.
• Current AI models often predict single proteins, lacking insight into signal propagation.

Purpose of the Study:

• To develop and evaluate an AI model for predicting downstream protein expression in breast cancer.
• To compare the performance of cellular-level versus patch-level Vision Transformers (ViT) for this task.
• To assess the utility of apoptosis and DNA damage/repair (DDR) cascades for predicting protein expression.

Main Methods:

• Utilized Reverse Phase Protein Array (RPPA) and whole-slide images (WSIs) from the TCGA-BRCA dataset.
• Developed a cellular-level ViT model (CellViT) and compared it with patch-level ViT models.
• Focused on predicting five key proteins in the apoptosis cascade, using DDR cascades as a control.

Main Results:

• Patch-level ViT models failed to achieve statistically significant predictive results (R-squared values < 0.1).
• CellViT demonstrated predictive capabilities, achieving R-squared values > 0.1 across five test folds.
• The apoptosis cascade, being morphologically indicative, yielded significantly higher predictive performance than the DDR cascade.

Conclusions:

• Cellular-level AI models, like CellViT, are more effective than patch-level models for predicting downstream protein expression from WSIs.
• Morphologically relevant biological pathways, such as apoptosis, are better targets for AI-driven protein expression prediction.
• This approach offers a novel way to infer functional protein signaling in cancer.