Effective Resistance and Graph Neural Network Reliability in Tissue-Specific Interactomes
Effective resistance added a small but reproducible signal for where tissue-specific GNN predictions fail.
The paper tests 24 tissue-specific protein interaction networks and finds effective resistance is mostly driven by inverse degree, with that collapse stronger in larger filtered networks. After controlling for entropy, degree, annotation count, local structure, and feature-only difficulty, its residual still explains extra per-node loss in 19 of 24 held-out networks. The effect grows with GNN depth, but the added variance explained is modest, peaking at 0.37% of what the controls leave unexplained. Selective prediction barely improves. ArXiv · AI/CL/LG's note
The paper tests 24 tissue-specific protein interaction networks and finds effective resistance is mostly driven by inverse degree, with that collapse stronger in larger filtered networks. After controlling for entropy, degree, annotation count, local structure, and feature-only difficulty, its residual still explains extra per-node loss in 19 of 24 held-out networks. The effect grows with GNN depth, but the added variance explained is modest, peaking at 0.37% of what the controls leave unexplained. Selective prediction barely improves. ArXiv · AI/CL/LG's note
score 3