We introduce SIMBA, a graph neural network enhanced with metabolic simulations, accurately predicting gut microbial abundances and uncovering cross-feeding interactions (Spearman ρ = 0.85).
The code and details for mechanistic simulations are available at: [https://github.com/javadamn/SIMBA-GNN/]
SIMBA is a deep learning framework that leverages heterogeneous graph neural networks to model complex interactions within microbial communities. The project implements a multi-stage training approach including self-supervised learning, supervised learning, and fine-tuning stages.
- Heterogeneous Edge-Aware Graph Neural Network architecture
- Multi-stage training pipeline:
- Self-supervised learning
- Supervised learning
- Fine-tuning
- Integration with Weights & Biases for experiment tracking
- Support for various loss functions and optimization strategies
- Flexible data handling for microbial community data
The pre-trained model checkpoints are available at
- Python 3.10+
- PyTorch
- PyTorch Lightning
- Weights & Biases
- Other dependencies (see requirements.txt)
# Clone the repository
git clone https://github.com/yourusername/MetaBiome.git
cd MetaBiome
# Install dependencies
pip install -r requirements.txtThe project can be run using the training script with Weights & Biases integration:
python train_wandb_sweep.pyIf you use SIMBA-GNN in your research or software, please cite the following preprint:
@article{Parsa2025.05.27.656377,
author = {Parsa, Mohammad S. and Aminian-Dehkordi, Javad and Mofrad, Mohammad},
title = {{SIMBA-GNN}: Simulation-augmented Microbiome Abundance Graph Neural Network},
journal = {bioRxiv},
year = {2025},
elocation-id = {2025.05.27.656377},
doi = {10.1101/2025.05.27.656377},
url = {https://www.biorxiv.org/content/early/2025/05/28/2025.05.27.656377},
publisher = {Cold Spring Harbor Laboratory}
}This project is licensed under the Creative Commons Attribution 4.0 International License - see the LICENSE file for details.