PPI-Net: Graph Neural Network Models Disease from Molecular to Functional Levels

Scientists have created PPI-Net, a hierarchical graph neural network that combines protein-protein interaction networks with representations at the pathway level to simulate disease progression from molecular interactions to functional outcomes. This model incorporates patient-specific molecular profiles into a unified interaction network sourced from STRING and transmits signals through a multi-layer Reactome hierarchy utilizing graph attention, which allows for the aggregation of gene-level signals into more complex biological programs. When evaluated on RNA-seq data from ten cancer types in The Cancer Genome Atlas, PPI-Net demonstrated impressive predictive capabilities, achieving a balanced accuracy of over 90%. This research tackles the challenge of elucidating how molecular changes influence biological systems to promote disease, providing interpretability across various scales that many existing models do not offer.