Research map / Label-efficient and robust learning / Recognition and long-tail
Improving Cancer Gene Prediction by Enhancing Common Information Between the PPI Network and Gene Functional Association
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 DISFusion, which fuses a Chebyshev GCN over the PPI network and a hypergraph residual network over annotated gene sets, with a cross-view decorrelation loss to enhance shared information for cancer gene identification.
Problem
Identifying cancer genes by combining PPI network and gene functional association information, where the shared (common) information between these two views is overlooked by existing methods.
Approach
Adds a cross-view decorrelation loss (Barlow-Twins-style) that pushes the cross-correlation matrix between graph and hypergraph gene embeddings toward identity, enhancing shared information and reducing redundancy before fusing both embeddings for prediction.
Key results
- Pan-cancer AUROC 0.948 and AUPRC 0.897, versus DISHyper at 0.930 and 0.870.
- On independent test sets, DISFusion shows a relative AUPRC improvement of 25% and 13% over DISHyper.
- The cross-view decorrelation loss improves AUROC by 1.1% and AUPRC by 1.5% over the variant without it (ablation).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DISFusion instead of DISHyper | Cancer gene identification · Pan-cancer gene identification | AUROC: 1.8% | Helped |
| Use DISFusion instead of DISHyper | Cancer gene identification · Pan-cancer gene identification | AUPRC: 2.7% | Helped |
| Use DISFusion instead of DORGE | Cancer gene identification · CESC | AUPRC: 3.4% | Helped |
| Use DISFusion instead of DISHyper | Cancer gene identification · HNSC | AUPRC: 6.8% | Helped |
| Use DISFusion instead of DISHyper | Cancer gene identification · LUAD | AUPRC: 9.8% | Helped |
| Add cross-view decorrelation loss (ablation) | DISFusion · Pan-cancer gene identification | AUROC: 1.1% | Helped |
| Add cross-view decorrelation loss (ablation) | DISFusion · Pan-cancer gene identification | AUPRC: 1.5% | Helped |
| Fuse PPI and functional-association information (ablation) | DISFusion without cross-view decorrelation loss · Pan-cancer gene identification | AUROC: 0.9% | Helped |
| Fuse PPI and functional-association information (ablation) | DISFusion without cross-view decorrelation loss · Pan-cancer gene identification | AUPRC: 1.2% | Helped |
| Increase α from 0 to 0.005 (ablation) | DISFusion · Pan-cancer gene identification | Model performance | Helped |
| Increase α beyond 0.01 (ablation) | DISFusion · Pan-cancer gene identification | Model performance | Hurt |
Surprising findings
- Genes such as ANXA1 and TLR2 ranked low by both network-based and functional-based baselines rank highly in DISFusion, attributed to the decorrelation loss.
Limitations
- Performance drops sharply when the loss weight alpha exceeds 0.01, so the trade-off is sensitive to hyperparameters.
- Novel cancer gene validation relies on CPTAC proteomic differential expression, which only partly confirms predictions (21 of 30 genes).
Experimental setup
Datasets: CPDB PPI network TCGA pan-cancer multi-omics features (via EMOGI) MSigDB C2 and C5 gene sets NCG v6.0 COSMIC CGC v91 DigSEE DriverDB v3 OncoKB ONGene CPTAC proteomics CancerMine
Benchmarks: pan-cancer cancer gene identification (AUROC/AUPRC) OncoKB independent test set ONGene independent test set six single cancer types (AUPRC) CPTAC proteomics validation
Metrics: AUROC AUPRC
Baselines: 20/20+ DORGE EMOGI MTGCN MODIG DISHyper
Builds on: EMOGI DISHyper Chebyshev GCN Barlow Twins
Closest papers by meaning
- Rethinking Cancer Gene Identification Through Graph Anomaly Analysis (AAAI 2025 · similarity 0.85)
- GeSubNet: Gene Interaction Inference for Disease Subtype Network Generation (ICLR 2025 · similarity 0.84)
- PRING: Rethinking Protein-Protein Interaction Prediction from Pairs to Graphs (NeurIPS 2025 · similarity 0.82)
- Tumor Micro-environment Interactions Guided Graph Learning for Survival Analysis of Human Cancers from Whole-slide Pathological Images (CVPR 2024 · similarity 0.82)
- Modeling Dense Multimodal Interactions Between Biological Pathways and Histology for Survival Prediction (CVPR 2024 · similarity 0.81)
- Gene-Gene Relationship Modeling Based on Genetic Evidence for Single-Cell RNA-Seq Data Imputation (NeurIPS 2024 · similarity 0.81)
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.