To run:
-
Install requirements: pip install -r requirements.txt
-
Run: python championship_disease_specific.py
That's it! The code will use the data in the dataset/ folder and output the results.
-
Final Score: 0.053167 (Micro Average Precision)# Install dependencies (if not already installed)
-
Ranking: 4th Place / ~150 teamspip install -r requirements.txt
-
Improvement: +34.6% over baseline```
-
Academic Grade: A- (92/100)
## 🎯 Overview# Open Jupyter notebook in VS Code or browser
jupyter notebook gene_disease_prediction.ipynb
This project tackles the **gene-disease association prediction** problem using graph-based semi-supervised learning. We predict which genes are associated with 305 different diseases using:```
- **Protein-Protein Interaction (PPI) Networks** (19,765 genes, 2.3M edges)Or simply open `gene_disease_prediction.ipynb` in VS Code.
- **Graph Augmentation** via k-NN (k=20, cosine similarity)
- **Label Propagation** with optimized hyperparameters (α=0.80)### 3. Run the Notebook
- **10-Fold Cross-Validation** ensemble for robust generalization
The notebook is organized into 10 main sections:
### Problem Characteristics
- **Highly Sparse Labels:** 98.9% of gene-disease pairs are unlabeled1. **Setup & Configuration** - Import libraries, set parameters
- **Imbalanced Classes:** Diseases range from 3 to 412 associated genes2. **Data Loading** - Load BioGRID network data
- **Graph-Structured Data:** Disease proteins cluster in PPI networks3. **EDA** - Analyze network properties and visualizations
4. **Utility Functions** - Helper functions for training
---5. **Baseline (MLP)** - Simple neural network baseline
6. **GNN Models** - GCN, GAT, GraphSAGE implementations
## 🚀 Quick Start7. **Training Functions** - Training and evaluation pipeline
8. **Class Imbalance** - Focal loss and weighted BCE
### Prerequisites9. **Model Training** - Train all models and compare
10. **Predictions & Submission** - Generate Kaggle submission files
```bash
# Python 3.10+ required**⚠️ Important**: For initial testing, modify the configuration:
pip install -r requirements.txt```python
```CONFIG['EPOCHS'] = 50 # Reduce from 200 for faster testing
cd FINAL_SUBMISSION/code## 📊 Project Structure
python winning_strategy_alpha_sweep.py
Complex_network_FP/
Output: Best submission with validation AP ~0.0878├── dataset/ # Competition data files
│ ├── train_idx.pt
📖 Detailed Instructions: See FINAL_SUBMISSION/QUICKSTART.md│ ├── test_idx.pt
│ ├── node_features.pt
---│ ├── y.pt
│ ├── edge_index.pt
├── gene_disease_prediction.ipynb # MAIN NOTEBOOK - Start here!
gene-disease-prediction/├── models/ # Saved model checkpoints
│├── outputs/ # Submission files and results
├── README.md # This file├── requirements.txt # Python dependencies
├── requirements.txt # Python dependencies├── eda.py # Standalone EDA script (optional)
│└── README.md # This file
└── FINAL_SUBMISSION/ # Complete solution package```
│
├── QUICKSTART.md # 5-minute guide---
├── README.md # Detailed package overview
│## 🎯 Competition Overview
├── code/ # Python scripts
│ ├── winning_strategy_alpha_sweep.py # Main solution ⭐### Dataset
│ ├── build_augmented_graph.py # Graph preprocessing- **Nodes**: 19,765 genes
│ └── generate_visualizations.py # Plot generator- **Edges**: 1,554,790 protein-protein interactions
│- **Features**: 37 per node (gene type, chromosome, strand, length)
├── results/ # Competition outputs- **Labels**: 305 diseases (multi-label binary classification)
│ ├── submission_alpha0.80_10fold_FIXED.csv # Best submission- **Challenge**: Highly imbalanced (98.9% sparsity)
│ └── RESULTS_SUMMARY.md # Performance analysis
│### Evaluation Metric
├── reports/ # Documentation**Average Precision (micro)** - Kaggle leaderboard metric
│ ├── STEP_BY_STEP_SOLUTION_REPORT.md # Technical methodology
│ └── FINAL_ACADEMIC_REPORT.md # Academic analysis### Task
│Predict which genes are associated with which diseases using:
└── visualizations/ # Plots & figures (9 plots)- Node features (gene properties)
├── alpha_sweep_results.png- Graph structure (protein interactions)
├── kfold_validation_results.png- Semi-supervised learning (only ~25% nodes labeled)
├── score_progression_timeline.png
└── ... (6 more plots)---
- Run EDA and understand data
- Get first Kaggle submission
| Method | Test Score | Improvement |
|--------|------------|-------------|### Phase 2: GNN Models (Days 3-4)
| Baseline (α=0.45) | 0.039507 | - |- [ ] Train GCN, GAT, GraphSAGE
| + K-Fold CV | 0.046358 | +17.3% |- [ ] Compare performance
| + Higher α (0.60) | 0.052736 | +13.8% |- [ ] Submit best single model
| Final (α=0.80, 10-fold) | 0.053167 | +34.6% |
- Feature engineering (add degree, centrality)
| Model | Train AP | Test AP | Issue |- [ ] Try different loss functions
|-------|----------|---------|-------|- [ ] Ensemble methods
| Label Propagation | 0.9234 | 0.0532 | ✅ Balanced |- [ ] Submit best ensemble
| GCN (Deep Learning) | 0.9532 | 0.0382 | ❌ Overfitting |
| GAT (Deep Learning) | 0.9687 | 0.0368 | ❌ Worse overfitting |### Phase 4: Refinement (Days 8-10)
| XGBoost | 0.8500 | 0.0340 | ❌ Ignores graph |- [ ] Advanced architectures (GIN, Transformer)
- Better handling of imbalance
Lesson: 98.9% label sparsity requires robust, graph-aware methods.- [ ] Final submissions
- Start report writing
- Create presentation slides
1. Graph Augmentation (k-NN, k=20)- [ ] Final submission preparation
└─ 1.5M edges → 2.3M edges (+50%)
---
2. Base Predictions (Logistic Regression, per-disease)
└─ C=0.1, SAGA solver## 💡 Key Insights from EDA
3. Label Propagation (α=0.80, 100 iterations)1. **Network is sparse** but well-connected (avg degree ~157)
└─ Y_new = 0.80 × (Adj @ Y) + 0.20 × Y_02. **Severe class imbalance**: 98.9% of labels are negative
3. **Power-law degree distribution**: Few hub genes, many peripheral genes
4. K-Fold Ensemble (10 folds)4. **Positive correlation** (r~0.3) between degree and disease count
└─ Final = mean(fold_predictions)5. **Hub genes** (high degree) tend to be associated with more diseases
Implication: Graph structure is crucial! GNNs should outperform MLP.
📖 Full Details: See FINAL_SUBMISSION/reports/STEP_BY_STEP_SOLUTION_REPORT.md
```python
# Add degree as feature
degrees = degree(data.edge_index[0], num_nodes=data.num_nodes)
# Add PageRank
from torch_geometric.utils import to_scipy_sparse_matrix
# ... compute PageRank and add to features
📈 All 9 Visualizations: See FINAL_SUBMISSION/visualizations/### 2. Hyperparameter Tuning
- Learning rate: Try 0.01, 0.001, 0.0001
---- Hidden dimensions: 128, 256, 512
- Number of layers: 2, 3, 4
- Loss function: BCE, Weighted BCE, Focal Loss
Our approach is validated by peer-reviewed research:
📄 DIAMOnD Algorithm (Ghiassian, Menche, Barabási 2015) - PLoS Computational Biology- GIN (Graph Isomorphism Network)
- PNA (Principal Neighbourhood Aggregation)
Key Finding: "Disease proteins localize in specific neighborhoods of the Interactome"- GraphTransformer
- Hybrid architectures
Validation:
-
✅ Disease proteins cluster in PPI networks (confirmed by research)### 4. Ensemble Strategies
-
✅ Label propagation superior to community detection for sparse labels- Average predictions from top 3 models
-
✅ Graph structure contains critical disease information- Weighted average (weight by validation AP)
-
Stack different model types (GCN + GAT + SAGE)
Citations: 500+ (highly influential paper by Albert-László Barabási, founder of network science)
📖 Full Academic Analysis: See FINAL_SUBMISSION/reports/FINAL_ACADEMIC_REPORT.md
Your report must include:
- Introduction: Problem description, motivation
-
Graph augmentation: k-NN adds informative edges- [ ] Methodology: Model descriptions, loss functions, training procedure
-
Higher α=0.80: Trust graph structure more than features- [ ] Experiments: Hyperparameters, ablation studies, model comparison
-
K-fold ensemble: Prevents overfitting (validation→test gap reduced)- [ ] Results: Performance metrics, tables, plots
-
Simplicity: Label propagation outperforms deep learning- [ ] Discussion: What worked, insights, challenges
-
Conclusion: Summary and future work
-
Deep learning (GCN/GAT): Severe overfitting on sparse labels
-
XGBoost: Ignores graph structureAll visualizations are automatically saved in
figures/folder! -
Lower α (0.45): Insufficient graph trust
-
Complex features: Added noise, not signal---
- Results & Insights (3-4 min)
📖 Professor Evaluation: See FINAL_SUBMISSION/reports/FINAL_ACADEMIC_REPORT.md4. Conclusion (1 min)
---Visual Elements:
- Network visualization
- Label distribution plot
All results are fully reproducible:- Model architecture diagrams
-
✅ Fixed random seeds (
seed=42)- Training curves -
✅ Documented dependencies (
requirements.txt)- Model comparison bar chart -
✅ Clear instructions (QUICKSTART.md)
-
✅ Hardware: MacBook M4 (CPU only, no GPU needed)Key Messages:
-
Why graph structure matters
Expected Runtime:- How GNNs work for this problem
-
Full solution: ~15 minutes- Your best insights from experiments
-
Alpha sweep: ~75 minutes- Final leaderboard position
| Document | Reading Time | Audience |### Papers
|----------|--------------|----------|- GCN: Kipf & Welling, 2017
| QUICKSTART.md | 5 min | First-time users |- GAT: Veličković et al., 2018
| STEP_BY_STEP_SOLUTION_REPORT.md | 20 min | Technical readers |- GraphSAGE: Hamilton et al., 2017
| FINAL_ACADEMIC_REPORT.md | 30 min | Academic reviewers |
| RESULTS_SUMMARY.md | 15 min | Performance analysis |### Documentation
---- BioGRID Database
- Bochra Chemam
- Yvonne Heiser
- Ronah Nakonde
This project is licensed under the MIT License.
- Kaggle Community for organizing the competition
- Course Instructors for guidance and feedback
- Barabási Lab for foundational research (DIAMOnD algorithm)
- Open-Source Tools: PyTorch, scikit-learn, NetworkX, matplotlib
---- Share Kaggle scores immediately
- Update notebook regularly
If you use this work, please cite:---
@misc{gene_disease_prediction_2025,
author = {Bochrache},### Issue 1: Out of Memory
title = {Gene-Disease Prediction via Label Propagation on PPI Networks},**Solution**: Reduce batch size, use gradient checkpointing, or train on CPU
year = {2025},
publisher = {GitHub},```python
url = {https://github.com/Bochrache/gene-disease-prediction}CONFIG['BATCH_SIZE'] = 256 # Reduce from 512
}```
---Solution: Use GPU, reduce epochs for testing, or train overnight
-
Weighted graph edges (PPI vs k-NN)print(device) # Should show 'cuda' if available
-
Disease-specific alpha optimization```
-
Meta-learning ensemble
-
Node2Vec embeddings### Issue 3: Poor Performance
-
Biological validation of predictionsSolution:
-
Check for bugs in data loading
---- Ensure proper train/val/test split
- Verify loss function handles imbalance
⭐ If you find this work useful, please star the repository!- Try different learning rates
Last Updated: December 13, 2025 ### Issue 4: Overfitting
Status: 4th Place (Public Leaderboard) - Awaiting Private ResultsSolution: Increase dropout, add weight decay, use early stopping
CONFIG['DROPOUT'] = 0.6
CONFIG['WEIGHT_DECAY'] = 1e-3
CONFIG['PATIENCE'] = 15- Check the notebook comments first
- Review PyTorch Geometric documentation
- Ask your team members
- Consult course materials
Remember:
- Start early - Don't wait until the deadline
- Iterate quickly - Submit often to Kaggle
- Learn from failures - Every experiment teaches something
- Work together - Teamwork makes the dream work
Kaggle Deadline: December 12, 2025
Final Submission: December 19, 2025
Last Updated: December 13, 2025
Team: Bochra Chemam, Yvonne Heiser, Ronah Nakonde