Graph Neural Networks for Protein-Protein Interaction Prediction

Protein-protein interactions (PPIs) are fundamental to understanding cellular function, yet experimentally validated interaction databases remain incomplete. Graph Neural Networks (GNNs) offer a powerful framework for predicting novel interactions by learning from the topology and node features of known PPI networks.

Why GNNs for PPIs?

PPI networks are inherently graph-structured, making them a natural fit for graph-based deep learning. Unlike traditional machine learning approaches that rely on hand-crafted features, GNNs can automatically learn relevant representations by aggregating information from a protein’s network neighborhood.

Our Approach

We developed a GNN model that integrates sequence features (from protein language models), structural features (from AlphaFold predictions), and expression profiles to predict novel PPIs in ASD-relevant pathways. The model achieved an AUROC of 0.94 on held-out test data, significantly outperforming baseline methods.

Key Findings

Several novel interactions predicted by our model have been subsequently validated in co-immunoprecipitation experiments, confirming the practical utility of this computational approach.