tb-pnca-gnn

A Graph Convolutional Network (GCN) for predicting Pyrazinamide (PZA) resistance in Mycobacterium tuberculosis from mutations in the pncA gene.

Overview

Pyrazinamide is a critical first-line antibiotic for tuberculosis treatment, but resistance prediction remains challenging due to the diverse range of mutations in the pncA gene. This project implements a GCN that models the PncA protein structure as a graph, combining structural information with amino acid features to predict resistance phenotypes.

The model was trained on phenotypically characterised isolates and achieves an F1 score of 0.816 on the held-out test set.

Installation

git clone https://github.com/fowler-lab/tb-pnca-gnn.git
cd tb-pnca-gnn
pip install -r requirements.txt

Requirements

• Python 3.8+
• PyTorch
• PyTorch Geometric
• Biopython
• pandas, numpy, scikit-learn
• Weights & Biases (optional, for experiment tracking)

Repository Structure

tb-pnca-gnn/
├── src/                    # Core source code
│   ├── gcn_model.py        # GCN architecture
│   ├── protein_graph.py    # Protein graph construction
│   ├── evaluation.py       # Model evaluation utilities
│   └── model_helpers.py    # Training helpers
├── train/                  # Training scripts
│   ├── train.py            # Main training script
│   └── configs/            # YAML configuration files
├── notebooks/              # Analysis notebooks (follow paper workflow)
│   ├── 01_create_sequences.ipynb
│   ├── 02_create_graph_dataset.ipynb
│   ├── 03_train_model.ipynb
│   ├── 04_bootstrapping.ipynb
│   ├── 05_gnn_explainer.ipynb
│   └── ...
├── data/                   # Datasets and features
├── saved_models/           # Pre-trained model weights
└── requirements.txt

Usage

Training a Model

Using the training script:

cd train
python train.py

Configuration can be modified in train/configs/config.yaml.

Using Pre-trained Models

Pre-trained models are available in saved_models/:

Model	F1 Score	Description
full_model/F1=0.816_epoch=1119.pth	0.816	Full model with all features
codon-split/F1=0.835_epoch=246.pth	0.835	Codon-based train/test split
cluster-split/F1=0.798_epoch=248.pth	0.798	Cluster-based train/test split

Notebooks

The notebooks in notebooks/ follow the analysis pipeline described in the paper:

1. 01_create_sequences.ipynb - Process mutation data and create sequences
2. 02_create_graph_dataset.ipynb - Build protein graphs with node features
3. 03_train_model.ipynb - Train and evaluate the GCN
4. 04_bootstrapping.ipynb - Bootstrap confidence intervals
5. 05_gnn_explainer.ipynb - Model interpretability analysis
6. 06_node_feature_importance.ipynb - Feature importance analysis

Data

Relevant data files are included in the data/ directory, including:

• Training and test sequences
• Protein structure features (MAPP scores, SNAP2 predictions, ΔΔG values)
• Pre-computed graph datasets in data/inputs/

Citation

If you use this code, please cite:

Dissanayake, D. et al. (2025). Predicting pyrazinamide resistance in Mycobacterium tuberculosis using a graph convolutional network. bioRxiv. https://doi.org/10.1101/2025.10.28.685176

@article{dissanayake2025pnca,
  title={Predicting pyrazinamide resistance in Mycobacterium tuberculosis using a graph convolutional network},
  author={Dissanayake, Dylan and others},
  journal={bioRxiv},
  year={2025},
  doi={10.1101/2025.10.28.685176}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.

Contact

Dylan Dissanayake - dylan.dissanayake@msdtc.ox.ac.uk

Philip W Fowler - philip.fowler@ndm.ox.ac.uk

Fowler Lab, University of Oxford