🧬 drGT

Official implementation of drGT: interpretable drug response prediction with attention-guided gene attribution on a drug-cell-gene heterogeneous graph, published in BMC Bioinformatics (2026).　　

drGT predicts how cancer cell lines respond to drugs and highlights genes that may help explain those responses.

It brings together drug response data, gene expression, chemical structures, and known drug targets using a graph neural network. Researchers can use its predictions and attention scores to identify genes for further study and explore possible mechanisms of drug sensitivity or resistance.

Attention scores suggest research directions; they do not prove causal relationships.

---

🚀 Quick Start

Requires: Python 3.10 or 3.11 Uses uv for lightweight execution

1. Clone the repository:

git clone https://github.com/inoue0426/drGT.git
cd drGT

2. Run a short training demo (CPU or GPU):

uv run --script run_drGT.py --task test1 --data nci --method GATv2

✅ If uv is not installed:

pip install uv

This demo trains for 3 epochs per fold using a 5-fold random split of observed drug-cell pairs. It checks the training workflow; it does not run pretrained inference or reproduce the paper's full-training results. For inference without training, use the pretrained notebook below. For leave-cell-out or leave-drug-out evaluation, use Test2_leave_X_out/run_drGAT.py; the demo runner only supports test1.

3. Example output from the short demo (values vary between runs):

Using device: cpu
Best model found at epoch 2
ACC           : 0.511 (±0.009)
Precision     : 0.215 (±0.296)
Recall        : 0.221 (±0.438)
F1            : 0.170 (±0.289)
AUROC         : 0.535 (±0.024)
AUPR          : 0.532 (±0.029)

---

🧠 Using Pretrained Models

To evaluate without retraining:

from drGT import drGT
from drGT.metrics import evaluate_predictions

probs, true_labels, attention = drGT.predict('best_model_nci.pt', sampler, params)
evaluate_predictions(true_labels, probs)

ℹ️ If you only want to generate predictions (inference) without reporting performance metrics, you can use the same pretrained model and skip evaluation.

✅ Ensure that params match the pretrained model's configuration (e.g., GNN layer, hidden sizes, etc.).

📓 For a full example, see predict_with_pretrained_model.ipynb.

You can use this pretrained model for existing drugs and cell lines. See cell_drug_availability_across_datasets.ipynb to check which cell lines and drugs are available across datasets.

---

⚙️ Interactive Use with Jupyter

To analyze results or explore predictions interactively:

# Create and activate a project environment
uv venv --python 3.10
source .venv/bin/activate

# Install libraries
uv pip install -e . jupyter

# Register Jupyter kernel
python -m ipykernel install --user --name=drGT --display-name "Python (drGT)"

# Launch notebook
jupyter notebook

---

🔄 Retraining

To retrain drGT from scratch, please use the scripts provided in the Test1_random_split and Test2_leave_X_out directories (e.g., run_drGAT.py). These scripts allow you to retrain the model under various experimental setups and data-splitting strategies.

If you wish to use your own dataset, please prepare the following:

• A drug response matrix (drugs × cell lines)
• A gene expression matrix (cell lines × genes)
• Drug chemical structures as SMILES strings (column 1: Drug, column 2: SMILES)

You can modify the drGT/load_data module to support your custom file formats or alternative data sources as needed.

---

🔄 When is Retraining Required?

You do NOT need retraining if:

• You use the same dataset and split as the pretrained model
• Model architecture and hyperparameters are unchanged

Retraining IS REQUIRED if:

• You use a different dataset or add new drugs/cell lines
• You change data splitting strategy (e.g., leave-drug-out)
• You modify model architecture or hyperparameters
• You aim to interpret attention scores on your own data

---

⚡️ GPU Acceleration

All experiments were conducted on Linux with NVIDIA A100. drGT benefits significantly from GPU acceleration via PyTorch and PyTorch Geometric.

✅ Ensure you install a CUDA-compatible PyTorch version (e.g., torch==2.x with CUDA 11.8 for A100)

👉 PyTorch Installation Guide

---

📁 Directory Overview

drGT/                 # Core model implementation
configs/              # YAML configs for experiments
Test1_random_split/   # Scripts for random masking experiments
Test2_leave_X_out/    # Scripts for leave-X-out experiments
preprocess/           # Data preprocessing notebooks
data/                 # Preprocessed datasets

---

❓ Questions or Issues

Please feel free to:

• Open a GitHub Issue and mention @inoue0426
• Or email inoue019@umn.edu

We're happy to help and collaborate!

---

📖 Citation

If you use drGT in your research, please cite the published article in BMC Bioinformatics.

@article{inoue2026drgt,
  title={drGT: Interpretable Drug Response Prediction with Attention-Guided Gene Attribution on a Drug-Cell-Gene Heterogeneous Graph},
  author={Inoue, Yoshitaka and Lee, Hunmin and Fu, Tianfan and Kuang, Rui and Luna, Augustin},
  journal={BMC Bioinformatics},
  volume={27},
  number={1},
  pages={204},
  year={2026},
  doi={10.1186/s12859-026-06417-z},
  url={https://doi.org/10.1186/s12859-026-06417-z}
}