Graph neural network explanations reveal a topological signature of disease-associated hubs in biological networks

Technical Summary: Graph Neural Network Explanations Reveal a Topological Signature of Disease-Associated Hubs

Problem Statement
Graph Neural Networks (GNNs) are increasingly utilized to model biological systems by integrating omics data (e.g., RNA-seq) with prior network structures like Protein-Protein Interaction (PPI) networks. However, the reliability of post-hoc explanation methods in recovering meaningful molecular mechanisms from these "black box" models remains uncertain. It is unclear whether existing explainers recover biologically valid structures or merely reflect method-specific biases, graph topology artifacts, or model idiosyncrasies. Specifically, there is a lack of systematic understanding regarding how different explanation algorithms (Saliency Attribution, Integrated Gradients, GNNExplainer, and Layer-wise Relevance Propagation) handle distinct biological signal geometries, such as sparse drivers versus distributed pathways.

Methodology
The authors employed a two-pronged evaluation strategy combining controlled synthetic benchmarks with real-world biological data.

1. Synthetic Benchmarks: Four distinct signal archetypes were constructed to mimic biological signal organizations:

   • Needle: A single decisive node (sparse signal).
   • Hub: A central node with elevated expression in its immediate 1-hop neighborhood (mimicking transcription factors).
   • Pathway/Module: A coordinated subnetwork of interacting nodes.
   • Cascade: A hierarchical, directional chain of increasing expression.
     A standard GNN architecture (three graph convolution layers with global mean pooling) was trained to achieve perfect accuracy on these tasks. Four explanation methods were evaluated: Saliency Attribution (SA), Integrated Gradients (IG), GNNExplainer, and Layer-wise Relevance Propagation (LRP, specifically the $\epsilon$-rule). Performance was measured using top-k hit rates and the fraction of attribution mass assigned to ground-truth signal nodes.

2. Real-World Application (TCGA BRCA): The study applied these explainers to a GNN trained on 1,222 breast cancer RNA-seq samples from The Cancer Genome Atlas (TCGA), projected onto the STRING PPI network. The model classified cancer vs. normal adjacent tissue with 89.7% balanced accuracy.

   • Hub Analysis: The authors analyzed attribution patterns around four canonical cancer driver genes (TP53, BRCA1, ESR1, MYC) and their 1-hop neighborhoods.
   • Topological Scoring: A "shell-based hub score" was developed to quantify the "eye-of-the-storm" pattern (low attribution at the center, peak at 1-2 hops, decay thereafter).
   • Consensus Framework: A consensus ranking approach was introduced, combining weighted percentile ranks from IG, SA, and LRP. This was further optimized via grid search.
   • Degree Confounding: Linear regression and residualization were used to determine if hub scores were driven by node degree rather than biological signal.
   • Pathway Enrichment: Reactome pathway enrichment analysis was performed on the top 100 ranked genes for each method.

Key Results

• Method-Specific Signal Recovery:

  • Saliency (SA): Performed best for the "Needle" (single-node) signal but struggled with distributed structures. In real data, SA assigned high scores to central hub nodes but showed diffuse attribution across the graph, resulting in high global ranking performance (AUC) but lower local enrichment.
  • Integrated Gradients (IG) and LRP: Consistently outperformed other methods in recovering distributed signals (Hubs, Pathways, Cascades). In synthetic "Hub" settings, IG and LRP achieved near-perfect recovery of the full neighborhood structure. In real data, they exhibited a distinct topological signature: attribution was minimized at the central seed node (0-hop), peaked in the immediate 1-hop neighborhood, and decayed sharply across subsequent shells (2–4 hops).
  • GNNExplainer: Often failed to localize signals to the correct topological structures, diffusing attribution to distant nodes and showing poor enrichment ratios for known disease hubs.

• The "Eye-of-the-Storm" Signature:
  In TCGA BRCA data, IG and LRP revealed that disease-associated signal is not concentrated at the hub center but is enriched in the immediate neighborhood. This pattern was most pronounced for IG, which showed strong enrichment ratios (e.g., ~11.2 for MYC) and high overlap with known hubs in the top-50 ranked genes.

• Consensus and Tuning:
  A consensus approach combining IG, SA, and LRP significantly improved the prioritization of canonical cancer genes.

  • TP53: Ranked 79th by IG alone; improved to 5th (100th percentile) by the tuned consensus.
  • BRCA1: Ranked 900th by IG; improved to 72nd (99.4th percentile) by the tuned consensus.
  • The tuned consensus (weights: IG=0.69, SA=0.31, LRP=0) outperformed individual methods, balancing local enrichment sensitivity with global discrimination.

• Degree Confounding:
  While IG showed a moderate dependence on node degree ($R^2 \approx 0.196$), the consensus hub scores substantially reduced this bias ($R^2 \approx 0.023$). Crucially, the prioritization of canonical cancer genes remained stable even after residualizing hub scores against node degree, indicating that the signal captured is not solely an artifact of network centrality.

• Biological Coherence:
  Pathway enrichment analysis revealed that individual methods often highlighted generic or biologically implausible programs (e.g., IG highlighted translation/ribosomal processes; SA highlighted neuronal pathways). In contrast, the consensus approach recovered biologically coherent cancer programs, including ERBB2, RTK, MAPK signaling, and immune/cytokine signaling, aligning better with known breast cancer biology.

Significance and Claims
The paper claims that graph explanations in biological systems should not be interpreted as single-method outputs but as structured signals shaped by both network topology and method-specific biases. The authors demonstrate that:

1. Different explainers recover distinct signal geometries; IG and LRP are superior for distributed, pathway-like signals, while SA is better for sparse, single-node drivers.
2. A consistent topological signature exists for disease-associated hubs in real data, characterized by attribution peaks in the immediate neighborhood rather than the central node.
3. Integrating multiple explainers via a topology-aware consensus framework improves the recovery of biologically relevant molecular mechanisms and reduces dependence on node degree.
4. This framework provides a practical approach for hypothesis generation from graph-based omics models, emphasizing the need for model-sensitivity analyses and independent biological validation before drawing mechanistic conclusions.

The study concludes that topology-aware integration of graph explanations enhances biological interpretability and the recovery of disease-relevant molecular signatures.