PPI-Net connects molecular protein interactions to functional processes in disease

Researchers introduce PPI-Net, a hierarchical graph neural network that integrates protein-protein interaction networks with biological pathway data to predict cancer outcomes and mechanisms. Demonstrating over 90% balanced accuracy across ten cancer types, the model reveals how molecular changes propagate through biological systems to drive disease, offering both predictive power and mechanistic interpretability.

PPI-Net addresses a fundamental limitation in current cancer prediction models: the inability to connect molecular-level observations to functional biological processes in interpretable ways. By embedding patient molecular profiles within protein interaction networks and propagating signals through a hierarchical pathway structure, the model bridges multiple scales of biological organization simultaneously. This approach proves particularly valuable because cancer involves complex cascades where initial mutations trigger downstream effects across interconnected cellular systems.

The research builds on decades of systems biology work demonstrating that protein interactions and signaling pathways follow organized hierarchies. PPI-Net's innovation lies in leveraging modern graph neural networks and publicly available databases—STRING for protein interactions and Reactome for pathway hierarchies—to automatically learn how molecular signals aggregate into functional programs. The 6.7% accuracy improvement from incorporating pathway hierarchy over protein interactions alone validates this multi-scale approach.

For biomedical AI development, this work demonstrates that incorporating structured biological knowledge directly into model architecture outperforms purely data-driven approaches. The model's ability to recover known oncogenic modules like TP53-AKT signaling while discovering convergent pathways like ion signaling suggests genuine biological insight rather than statistical artifacts. This interpretability matters tremendously for clinical translation, where physicians need to understand predicted outcomes.

The multi-omics extension combining RNA-seq and methylation data indicates the framework's flexibility for integrating diverse data types. Future applications could extend beyond cancer to other complex diseases, establishing PPI-Net as a foundational approach for systems-level disease modeling. The work's reliance on open databases ensures reproducibility and accessibility across research institutions.

• →PPI-Net achieves over 90% balanced accuracy predicting cancer outcomes by integrating protein interactions with pathway hierarchies
• →Hierarchical pathway information improves model accuracy by 6.7% compared to protein-interaction-only approaches
• →The model identifies known cancer mechanisms while discovering novel convergent biological programs driving disease
• →Multi-scale graph neural networks incorporating structured biological knowledge outperform data-driven methods alone
• →Framework demonstrates path toward clinically interpretable AI by connecting molecular predictions to functional biological processes