Effective Resistance and Graph Neural Network Reliability in Tissue-Specific Interactomes

Effective resistance in 24 tissue interactomes mostly tracks inverse degree, limiting GNN reliability signals.

The authors test whether effective resistance in tissue-specific protein interaction networks can flag unreliable graph neural network predictions for protein function. Across 24 interactomes the signal is dominated by inverse degree, with Spearman correlation -0.955, and that degeneration deepens as co-expression filtering grows the network. Residual departure from the inverse-degree limit exceeds degree-preserving null graphs in every network and, after controls for entropy, degree, annotation size, local structure, and feature difficulty, explains additional per-node loss in 19 of 24 held-out networks. The added variance is only 0.37% of what the controls leave unexplained, and selective prediction improves negligibly.