SkipGNN: predicting molecular interactions with skip-graph networks

Molecular interaction networks are powerful resources increasingly used with machine learning to predict biologically meaningful interactions. Existing graph neural networks (GNNs) mainly aggregate information from immediate neighbors and thus optimize for direct similarity between interacting nodes, overlooking indirect similarities that are highly informative in biological networks. We present SkipGNN, a GNN that aggregates messages from both direct neighbors and two-hop neighbors (skip similarity) by constructing a second-order skip graph and iteratively fusing representations from the original and skip graphs. Evaluations on four interaction networks—drug-drug, drug-target, protein-protein, and gene-disease—show that SkipGNN achieves superior and robust performance, learns biologically meaningful embeddings, and performs especially well on noisy, incomplete networks.