---
tags: ['graph-neural-networks', 'biological-networks', 'napistu', 'pytorch', 'graph_conv', 'relation_attention_mlp', 'edge_prediction', 'relation-aware']
library_name: napistu-torch
license: mit
metrics:
- auc
- average_precision
---

# 128-dim GraphConv model trained for relation-stratified edge prediction with a relation-attention MLP head

This model was trained using [Napistu-Torch](https://www.shackett.org/napistu_torch/), a PyTorch framework for training graph neural networks on biological pathway networks.

The dataset used for training is the 8-source ["Octopus" human consensus network](https://www.shackett.org/octopus_network/), which integrates pathway data from STRING, OmniPath, Reactome, and others. The network encompasses ~50K genes, metabolites, and complexes connected by ~8M interactions.

## Task

This model performs **edge prediction** on biological pathway networks. Given node embeddings, 
the model predicts the likelihood of edges (interactions) between biological entities such as 
genes, proteins, and metabolites. This is useful for:

- Discovering novel biological interactions
- Validating experimentally observed interactions
- Completing incomplete pathway databases
- Predicting functional relationships between genes/proteins

The model learns to score potential edges based on learned embeddings of source and target nodes, 
optionally incorporating relation types for relation-aware prediction.

## Model Description

- **Encoder**
  - Type: `graph_conv`
  - Hidden Channels: `128`
  - Number of Layers: `3`
  - Dropout: `0.2`
  - Edge Encoder: ✓ (dim=32)
- **Head**
  - Type: `relation_attention_mlp`
  - Relation-Aware: ✓

**Training Date**: 2025-12-14

For detailed experiment and training settings see this repository's `config.json` file.

## Performance

| Metric | Value |
|--------|-------|
| Validation AUC | 0.8187 |
| Test AUC | 0.8186 |
| Validation AP | 0.8205 |
| Test AP | 0.8204 |


## Links

- 📊 [W&B Run](https://wandb.ai/napistu/napistu-experiments/runs/63xqckzk)
- 💻 [GitHub Repository](https://github.com/napistu/Napistu-Torch)
- 📖 [Read the Docs](https://napistu-torch.readthedocs.io/en/latest)
- 📚 [Napistu Wiki](https://github.com/napistu/napistu/wiki)

## Usage

### 1. Setup Environment

To reproduce the environment used for training, run the following commands:

```bash
pip install torch==2.8.0
pip install torch-scatter torch-sparse -f https://data.pyg.org/whl/2.8.0+cpu.html
pip install 'napistu==0.8.2'
pip install 'napistu-torch[pyg,lightning]==0.2.14'
```

### 2. Setup Data Store

First, download the Octopus consensus network data to create a local `NapistuDataStore`:
```python
from napistu_torch.load.gcs import gcs_model_to_store

# Download data and create store
napistu_data_store = gcs_model_to_store(
    napistu_data_dir="path/to/napistu_data",
    store_dir="path/to/store",
    asset_name="human_consensus",
    # Pin to stable version for reproducibility
    asset_version="20250923"
)
```

### 3. Load Pretrained Model from HuggingFace Hub
```python
from napistu_torch.ml.hugging_face import HuggingFaceLoader

# Load checkpoint
loader = HuggingFaceLoader("seanhacks/relation_attention_mlp")
checkpoint = loader.load_checkpoint()

# Load config to reproduce experiment
experiment_config = loader.load_config()
```

### 4. Use Pretrained Model for Training

You can use this pretrained model as initialization for training via the CLI:
```bash
# Create a training config that uses the pretrained model
cat > my_config.yaml << EOF
name: my_finetuned_model

model:
  use_pretrained_model: true
  pretrained_model_source: huggingface
  pretrained_model_path: seanhacks/relation_attention_mlp
  pretrained_model_freeze_encoder_weights: false  # Allow fine-tuning

data:
  sbml_dfs_path: path/to/sbml_dfs.pkl
  napistu_graph_path: path/to/graph.pkl
  napistu_data_name: edge_prediction

training:
  epochs: 100
  lr: 0.001
EOF

# Train with pretrained weights
napistu-torch train my_config.yaml
```

## Citation

If you use this model, please cite:
```bibtex
@software{napistu_torch,
  title = {Napistu-Torch: Graph Neural Networks for Biological Pathway Analysis},
  author = {Hackett, Sean R.},
  url = {https://github.com/napistu/Napistu-Torch},
  year = {2025},
  note = {Model: graph_conv-relation_attention_mlp_h128_l3_edge_prediction}
}
```

## License

MIT License - See [LICENSE](https://github.com/napistu/Napistu-Torch/blob/main/LICENSE) for details.