Research map / Graph learning / Graph self-supervised learning
Rethinking Cancer Gene Identification Through Graph Anomaly Analysis
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Identifies "weight heterogeneity" (high edge-weight variance) in cancer gene nodes of PPI networks and proposes HIPGNN, a spectral-plus-spatial GNN that uses eigenvalue encodings and confidence/interaction context decoding to identify cancer genes.
Problem
Cancer gene identification on protein-protein interaction networks, where existing GNNs use unweighted graphs and fail to model the complex interaction patterns of cancer genes.
Approach
Treats PPI confidence scores as edge weights and observes that cancer genes show higher edge-weight variance (weight heterogeneity), which flattens spectral energy toward spectrum extremes; HIPGNN encodes eigenvalue position and proximity in a transformer-based spectral filter and adds spatial decoders for interaction and confidence context.
Key results
- On STRINGdb at 80% training, HIPGNN reaches AP 81.21 vs SMG's 77.43 (weighted setting), a gain of 3.78 points.
- On CPDB at 20% training, HIPGNN (unweighted) reaches AP 78.07 and weighted HIPGNN AP 78.13, outperforming SMG by about 7.30 points in AP.
- Ablation shows both interaction and confidence context help, with confidence context being most impactful.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HIPGNN instead of SMG | Cancer gene identification · STRINGdb, 20% train | AP: +0.44% | Helped |
| Use HIPGNN instead of SMG | Cancer gene identification · CPDB, 20% train | AP: +7.30% | Helped |
| Use HIPGNN instead of SMG | Cancer gene identification · STRINGdb, 80% train | AP: +3.78% | Helped |
| Use HIPGNN instead of SMG | Cancer gene identification · CPDB, 80% train | AP: +1.88% | Helped |
| Add spectral eigenvalue encoding and proximity-aware spectral graph representation (ablation) | HIPGNN · STRINGdb, 80% train | Model performance | Helped |
| Add interaction context decoding (ablation) | HIPGNN · STRINGdb, 20% train | Model performance | Helped |
| Add confidence context decoding (ablation) | HIPGNN · STRINGdb, 20% train | Model performance | Helped |
| Use empirical loss weights instead of uniform Average weight or Weight learner (ablation) | HIPGNN · STRINGdb | AP | Helped |
| Set eigenvalue threshold q to 3,000 rather than other tested q values (ablation) | HIPGNN · STRINGdb | Model performance | Helped |
| Increase q beyond 3,000 (ablation) | HIPGNN · STRINGdb | Model performance | Hurt |
Surprising findings
- Edge weights hurt low-pass filters like GCN on weighted graphs, while ChebNet and HIPGNN benefit from weighted edges.
- Spectral energy flattening was less pronounced on real CPDB data in the low-frequency region than in synthetic graphs.
Limitations
- Weight heterogeneity was observed only in cancer genes on two PPI networks, requiring further validation.
- Real-world edge weights are not ideally Gaussian, so synthetic-graph findings are not fully reproduced on CPDB.
- Experiments were run on a single L4 GPU in Google Colab, limiting scale.
Experimental setup
Datasets: STRINGdb CPDB TCGA
Benchmarks: STRINGdb CPDB
Metrics: AUC F1 (macro) AP
Baselines: GCN GAT GraphSAGE ChebNet EMOGI MTGCN SMG
Builds on: EMOGI SMG MTGCN ChebNet BWGNN
Closest papers by meaning
- Rayleigh Quotient Graph Neural Networks for Graph-level Anomaly Detection (ICLR 2024 · similarity 0.86)
- OgBench: A Framework for Evaluating Graph Neural Networks on Omics Data (NeurIPS 2026 · similarity 0.85)
- Improving Cancer Gene Prediction by Enhancing Common Information Between the PPI Network and Gene Functional Association (AAAI 2025 · similarity 0.85)
- GeSubNet: Gene Interaction Inference for Disease Subtype Network Generation (ICLR 2025 · similarity 0.85)
- MOTIVE: A Drug-Target Interaction Graph For Inductive Link Prediction (NeurIPS 2024 · similarity 0.84)
- Improved Expressivity of Hypergraph Neural Networks through High-Dimensional Generalized Weisfeiler-Leman Algorithms (ICML 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.