Research map / Graph learning / Knowledge and temporal graphs
GeSubNet: Gene Interaction Inference for Disease Subtype Network Generation
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.
Proposes GeSubNet, a multi-step framework combining a VQ-VAE over patient gene expression with a Neo-GNN over prior gene networks to infer subtype-specific gene networks for four TCGA cancers.
Problem
Generic knowledge-database gene networks (e.g., STRING, KEGG) lack specificity for disease subtypes, and existing gene network inference methods do not explicitly learn subtype-specific gene interactions from expression data.
Approach
Learn a VQ-VAE subtype representation from patient expression, a Neo-GNN link-prediction representation from prior gene networks, and then fine-tune the graph encoder with a reconstruction loss that reconstructs patient expression from the product of patient codes and graph embeddings, so inferred links become subtype-aware.
Key results
- Relative to the second-best baseline CSGNN, GeSubNet improves GED by 35.8%, 32.4%, 20.2%, and 34.1% across BRCA, GBM, LGG, and OV.
- In the simulated gene knockout on BRCA, high-ranking genes reach an 83% shift rate versus 20-30% for other baselines.
- GeSubNet attains the highest #EBF on all four cancers (8, 6, 6, 5 in BRCA, GBM, LGG, OV).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GeSubNet instead of CSGNN for subtype network inference | Subtype-specific gene network inference · BRCA, GBM, LGG, and OV | GED: 35.8%/32.4%/20.2%/34.1% | Helped |
| Use GeSubNet instead of CSGNN for subtype network inference | Subtype-specific gene network inference · BRCA, GBM, LGG, and OV | DCS: 29.8%/13.5%/21.6%/19.3% | Helped |
| Use GeSubNet instead of CSGNN for subtype network inference | Subtype-specific gene network inference · BRCA, GBM, LGG, and OV | CDV: 33.4%/13.7%/17.9%/15.3% | Helped |
| Use GeSubNet to select high-ranking rather than low-ranking genes for simulated knockout | BRCA gene selection · BRCA | Shift Rate (ΔSR): 12% -> 83% | Helped |
| Set latent dimension to 16 instead of 32 (ablation) | GeSubNet · Hyperparameter sensitivity experiment | GED: 0.78 -> 0.76 | Hurt |
| Set latent dimension to 64 instead of 32 (ablation) | GeSubNet · Hyperparameter sensitivity experiment | GED: 0.78 -> 0.79 | Helped |
| Set code book size to 16 instead of 32 (ablation) | GeSubNet · Hyperparameter sensitivity experiment | GED: 0.78 -> 0.72 | Hurt |
| Set code book size to 64 instead of 32 (ablation) | GeSubNet · Hyperparameter sensitivity experiment | GED: 0.78 -> 0.75 | Hurt |
| Set batch size to 16 instead of 32 (ablation) | GeSubNet · Hyperparameter sensitivity experiment | GED: 0.78 -> 0.76 | Hurt |
| Set batch size to 64 instead of 32 (ablation) | GeSubNet · Hyperparameter sensitivity experiment | GED: 0.78 -> 0.77 | Hurt |
Surprising findings
- Performance on OV is comparatively poor, which the authors attribute to limited samples (291 patients) and sparse pathogenic knowledge in databases.
Limitations
- Small patient cohorts and small prior networks (about 100 to 150 nodes) limit generality.
- Evaluation relies on graph statistics and enrichment proxies rather than experimentally validated subtype-specific interactions.
- Results on OV are weak, indicating sensitivity to sample size and database coverage.
Experimental setup
Datasets: TCGA BRCA TCGA GBM TCGA LGG TCGA OV KEGG STRING InterPro Monarch
Benchmarks: TCGA BRCA TCGA GBM TCGA LGG TCGA OV
Metrics: GED DCS CDV #EBF Shift Rate (SR)
Baselines: WGCNA wTO ARACNe LEAP GAERF LR-GNN CSGNN
Builds on: VQ-VAE Neo-GNN WGCNA ARACNe
Closest papers by meaning
- InfoSEM: A Deep Generative Model with Informative Priors for Gene Regulatory Network Inference (ICML 2025 · similarity 0.86)
- KINDLE: Knowledge-Guided Distillation for Prior-Free Gene Regulatory Network Inference (NeurIPS 2025 · similarity 0.86)
- OgBench: A Framework for Evaluating Graph Neural Networks on Omics Data (NeurIPS 2026 · similarity 0.85)
- Rethinking Cancer Gene Identification Through Graph Anomaly Analysis (AAAI 2025 · similarity 0.85)
- Improving Cancer Gene Prediction by Enhancing Common Information Between the PPI Network and Gene Functional Association (AAAI 2025 · similarity 0.84)
- A Generic Family of Graphical Models: Diversity, Efficiency, and Heterogeneity (ICML 2025 · similarity 0.83)
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.