GNN Explanations Reveal Topological Signature of Disease Hubs in Biological Networks
Recent evaluations of four Graph Neural Network (GNN) explanation methods—SA, IG, GNNExplainer, and LRP—highlighted their effectiveness in analyzing biological networks. Utilizing synthetic benchmarks with established motifs, the study found that SA excels at identifying sparse single-node drivers, while IG and LRP are adept at retrieving distributed pathway-like and cascade-like signals. The analysis, grounded in TCGA BRCA breast cancer RNA-seq data, projected findings onto a protein-protein interaction network, revealing disease-associated hubs that exhibited attribution peaks within a one-hop neighborhood, with a gradual decay in attribution across subsequent hops. The results were shared on arXiv.
Key facts
- Four GNN explanation methods were evaluated: SA, IG, GNNExplainer, LRP.
- Synthetic benchmarks with known ground-truth motifs were used.
- SA best recovers sparse single-node drivers.
- IG and LRP recover distributed pathway-like and cascade-like signals.
- Analysis used TCGA BRCA breast cancer RNA-seq data.
- Data projected onto a protein-protein interaction network.
- Disease-associated hubs show attribution peaks in 1-hop neighborhood.
- Attribution decays across successive hops from hubs.
Entities
Institutions
- arXiv