Data, code package, and ,models for "Interpretable Molecular Property Prediction and Fragments Discovery via Multi-scale Graph Attention Networks with Hierarchical Fragmentation"

This deposit is shared for editorial assessment and peer review. A public release will follow upon publication. This deposit contains the manuscript-ready data and code package accompanying the study “Interpretable Molecular Property Prediction and Fragments Discovery via Multi-scale Graph Attention Networks with Hierarchical Fragmentation”. The package is organized into two top-level folders that should remain side by side under `Manuscript_release`: - `Data availability`: 20 property datasets used in this study, spanning thermophysical, pharmacokinetic, toxicological, and industrial safety properties, together with the source data underlying the analyses in Fig. 1 and Fig. 3.- `Code availability`: scripts for dataset construction, hierarchical fragmentation, model training, hyperparameter optimization, prediction-result analysis, matched molecular pair analysis, fragment contribution ranking, top-20 fragment visualization, and GC/BRICS-based classification. The model folder includes implementations of HierAttnGNN and the comparison baselines Attention FP, BRICS-GNN, GC-GNN, HiGNN, and SPE-GNN. The package supports reproduction of the manuscript figures and tables, rebuilding of `.pt` datasets, retraining of models, interpretability analyses, and visualization of fragmentation examples. Main software environment: Python 3.8.20, PyTorch 2.4.0+cu118, DGL 1.1.2+cu118, RDKit 2023.9.5, torch-geometric 2.6.1, torch_scatter 2.1.2+pt24cu118, NumPy 1.24.4, pandas 2.0.3, scikit-learn 1.3.0, matplotlib 3.7.1, openpyxl 3.1.5, joblib 1.4.2, tqdm 4.66.5, fastprogress 1.0.5, and Optuna 4.5.0. Most subdirectories include README files with the required commands and run order. Some scripts depend on prebuilt `.pt`, `.pth`, or `.gz` files included in the package.