PiDeeL: metabolic pathway-informed deep learning model for survival analysis and pathological classification of gliomas

Official implementation of PiDeeL (Pathway-Informed Deep Learning Model), a deep learning framework that incorporates metabolic pathway information for improved survival prediction in glioma patients using HRMAS NMR spectroscopy data.

Paper

PiDeeL: metabolic pathway-informed deep learning model for survival analysis and pathological classification of gliomas

Published in Bioinformatics, Volume 39, Issue 11, November 2023

📄 Read the paper

Overview

PiDeeL integrates biological metabolic pathway knowledge into the neural network architecture to predict patient survival outcomes. By constraining the network's weights according to metabolite-pathway relationships, PiDeeL achieves better interpretability and performance.

Installation

1. Clone the repo

   git clone https://github.com/ciceklab/PiDeeL/
   cd PiDeeL

2. Create conda environment

   conda env create --name PiDeeL --file PiDeeL.yml
   conda activate PiDeeL

Key Dependencies

• PyTorch 2.0+
• pycox (for survival analysis)
• scikit-survival
• scikit-learn
• pandas, numpy, matplotlib
• shap

Repository Structure

PiDeeL/
├── README.md
├── PiDeeL.yml                    # Conda environment file
├── system_fig.png
│
├── run/                          # Inference with pretrained models
│   ├── predict.py                # Main prediction script
│   ├── model.py                  # Model architecture
│   ├── model_utils.py            # Utility functions
│   ├── PiDeeL_2layer.pth         # Pretrained 2-layer model
│   ├── PiDeeL_3layer.pth         # Pretrained 3-layer model
│   ├── PiDeeL_4layer.pth         # Pretrained 4-layer model
│   └── sample_quant.pickle       # Sample input data
│
└── reproduction/                 # Full experiment reproduction
    ├── config.py                 # Central configuration
    ├── scripts/                  # Training scripts
    │   ├── load_targeted_data.py
    │   ├── model_utils.py
    │   ├── 1layer/, 2layer/, 3layer/, 4layer/
    │   └── baseline/             # Cox-PH, CWGB, RF
    ├── figures/                  # Paper figure generation
    │   ├── run_fig2.py, run_fig4.py, ...
    │   └── run_all_figures.py
    ├── models/                   # Saved model weights
    ├── logs/                     # C-Index results
    ├── plots/                    # Training loss plots
    └── pideel_data/              # Data directory

---

Prediction Using Pretrained PiDeeL

Use pretrained models to predict survival risk scores for new samples.

1. Navigate to the run directory:

   cd run/

2. Prepare your input data:

   • Use the provided sample_quant.pickle as a template
   • Or use the automated metabolite quantification pipeline from Cakmakci et al. to quantify your HRMAS NMR spectroscopy data

3. Run prediction:

   python predict.py --layer 4 --dev gpu

   Options:

   • --layer: Select model architecture (2, 3, or 4)
   • --dev: Device to use (gpu or cpu)

4. Output: Risk scores printed to terminal

---

Reproducing Paper Results

Quick Start: Interactive Notebook

The easiest way to reproduce the main comparison figure is using the Jupyter notebook at the repo root:

jupyter notebook reproduce_main_comparison.ipynb

By default, the notebook uses pretrained model logs included in the repository to generate Figure 2 immediately.

To retrain models from scratch, set the RETRAIN flag in the notebook:

RETRAIN = True  # Set to True to retrain all models instead of using pretrained logs

Data Setup

The preprocessed samples are already included in the repository under reproduction/pideel_data/targeted/. No additional download is required to reproduce the results.

If you want access to the raw HRMAS NMR spectroscopy data, you can download it from Zenodo:

https://zenodo.org/record/7228791

Training Models

Navigate to the reproduction directory:

cd reproduction/

PiDeeL and DeepSurv models:

# 2-layer
python scripts/2layer/no_pathway/main.py  # DeepSurv
python scripts/2layer/pathway/main.py     # PiDeeL

# 3-layer
python scripts/3layer/no_pathway/main.py
python scripts/3layer/pathway/main.py

# 4-layer
python scripts/4layer/no_pathway/main.py
python scripts/4layer/pathway/main.py

Baseline models:

python scripts/baseline/coxph/main.py  # Cox Proportional Hazards
python scripts/baseline/cwgb/main.py   # Component-wise Gradient Boosting
python scripts/baseline/rf/main.py     # Random Survival Forest

Generating Figures

cd reproduction/figures/

# Generate all figures
python run_all_figures.py

# Or generate individual figures
python run_fig2.py      # Main comparison (Fig. 2)
python run_fig4.py      # Random connections ablation
python run_fig5.py      # Dropout analysis
python run_fig6.py      # External validation

Results

• C-Index results: reproduction/logs/*/c_indices.txt
• Generated figures: reproduction/figures/
• Training plots: reproduction/plots/

---

GPU/CPU Support

The code automatically detects GPU availability:

• GPU available: Uses CUDA for training (recommended)
• CPU only: Falls back to CPU (slower but functional)

No code changes needed - device selection is automatic via config.py.

---

Citation

If you use this code in your research, please cite:

@article{kaynar2023pideel,
  title={PiDeeL: metabolic pathway-informed deep learning model for survival analysis and pathological classification of gliomas},
  author={Kaynar, Gun and Cakmakci, Doruk and Bund, Caroline and Todeschi, Julien and Namer, Izzie Jacques and Cicek, A Ercument},
  journal={Bioinformatics},
  volume={39},
  number={11},
  pages={btad684},
  year={2023},
  publisher={Oxford University Press}
}

License

Distributed under the MIT License.

Contact

Gun Kaynar - gunkaynar.com

A. Ercument Cicek - http://ciceklab.cs.bilkent.edu.tr/ercument

Acknowledgements

This work was supported by grants from BPI France (ExtempoRMN Project), Hôpitaux Universitaires de Strasbourg, Bruker BioSpin, Univ. de Strasbourg and the Centre National de la Recherche Scientifique; also by TUBA GEBIP, Bilim Akademisi BAGEP and TUSEB Research Incentive awards to AEC.