gc-gnn-papers

Trained checkpoints and results for reproducing benchmarks from two papers that introduced the GroupGAT architecture.

Models are implemented in gc-gnn — a PyTorch Geometric library for hierarchical multi-level graph neural networks that combines the concept of group-contribution and GNNs. This repo pins v0.1.0 of that library.

Results may vary slightly due to the initialization seed. All checkpoints were trained on Linux.

---

Paper 1 — GroupGAT (JCIM 2023)

Aouichaoui et al., Journal of Chemical Information and Modeling, 63(3), 725–744 (2023)

Property	Checkpoint	Test MAE	Test R²	Overall MAE	Overall R²
ESOL	groupgat_esol	0.387	0.939	0.343	0.950
Bradley Melting Point	groupgat_nmp	26.61	0.843	18.13	0.934

All checkpoints in checkpoints/paper_1/.

---

Paper 2 — GroupGAT (CACE 2023)

Aouichaoui et al., Computers & Chemical Engineering, 176, 108291 (2023)

Property	Checkpoint	Test MAE	Test R²	Overall MAE	Overall R²
BCF	groupgat_bcf	0.335	0.864	0.290	0.912
BioD	groupgat_biod	0.115	0.787	0.124	0.770
LC50	groupgat_lc50	0.537	0.790	0.443	0.844
LD50	groupgat_ld50	0.338	0.627	0.239	0.813
LogKow	groupgat_logkow	0.294	0.943	0.160	0.983
PCO	groupgat_pco	0.120	0.788	0.102	0.922
PEI	groupgat_pei	0.446	0.841	0.595	0.644
pKa	groupgat_pka	1.019	0.818	0.732	0.905

All checkpoints in checkpoints/paper_2/.

---

Usage

Requires Python ≥ 3.9.

Install dependencies and the model library (pinned to v0.1.0):

pip install -r requirements.txt
pip install git+https://github.com/arnaou/gc-gnn@v0.1.0

Evaluate a checkpoint against its original split:

# Paper 1
python scripts/evaluate.py checkpoints/paper_1/groupgat_esol/

# Paper 2
python scripts/evaluate.py checkpoints/paper_2/groupgat_ld50/

Predict on new molecules interactively using the included notebook:

notebooks/inference.ipynb

The notebook walks through loading a checkpoint, featurizing SMILES, running inference, and visualizing fragment attention weights. Both requirements.txt and the gc-gnn pip install above are prerequisites.

---

Citation

If you use these checkpoints or results, please cite the relevant paper(s). If you use the underlying datasets, please also cite the original data sources referenced in each paper.

Paper 1 (JCIM 2023)

@article{aouichaoui2023combining,
  title     = {Combining Group-Contribution Concept and Graph Neural Networks Toward Interpretable Molecular Property Models},
  author    = {Aouichaoui, Adem R. N. and Fan, Fan and Abildskov, Jens and Sin, G{\"u}rkan},
  journal   = {Journal of Chemical Information and Modeling},
  volume    = {63},
  number    = {3},
  pages     = {725--744},
  year      = {2023},
  publisher = {ACS Publications}
}

Paper 2 (CACE 2023)

@article{aouichaoui2023application,
  title     = {Application of Interpretable Group-Embedded Graph Neural Networks for Pure Compound Properties},
  author    = {Aouichaoui, Adem R. N. and Fan, Fan and Mansouri, Seyed S. and Abildskov, Jens and Sin, G{\"u}rkan},
  journal   = {Computers \& Chemical Engineering},
  volume    = {176},
  pages     = {108291},
  year      = {2023},
  publisher = {Elsevier}
}