EDKG-DL Model Assets

Model asset repository for EDKG-DL — a Deep Learning framework with a causality-integrated Endocrine Disruption Knowledge Graph. It predicts endocrine-disrupting chemicals (EDCs), applicability domain (ADs), biological entities, and sensitive pathways from compound SMILES using machine learning and graph deep learning algorithmes.

Pipeline: PaDEL 2D descriptors/fingerprints → biological entity-specific qualitative/quantitative models → causality-integrated EDC classification → sensitive adverse outcome (AO) and adverse outcome pathway (AOP) inference, producing JSON / Excel reports.

Repository contents (~930 MB)

├── settings.json           # global settings
├── manifest.json           # SHA-256 digests for integrity verification
├── qualitative_models/     # 74 biological entity-specific qualitative models
├── quantitative_models/    # 52 biological entity-specific quantitative models
├── edc_models/             # EDC classification model
├── gnn_models/             # graph neural network weights
└── padel_runs_cache/       # cached PaDEL descriptors

Usage

The recommended way is the edkg-dl Python package (Python ≥ 3.10), which downloads these assets automatically on first use — no manual setup required:

pip install edkg-dl
edkg-dl-predict "CCO"

Or download the assets directly with huggingface_hub:

from huggingface_hub import snapshot_download

snapshot_download(repo_id="HaoyueTan/edkg-dl-models", local_dir="./models")

Once downloaded, assets can be reused offline by setting HF_HUB_OFFLINE=1.

Integrity & security

Every file digest is verified against manifest.json before the models load. Note that PKL/Joblib files execute code when deserialized; only load assets from trusted sources.

License

Released under the Custom Academic and Research Use License — free for non-commercial educational and research purposes only; commercial use requires a separate license from Nanjing University.

Disclaimer

Results are for research support only and must not replace experimental or regulatory conclusions.