druglikeness 0.2.2

AI-based drug-likeness prediction

Drug-likeness scoring based on unsupervised learning

This repository contains API for Drug-likeness scoring based on unsupervised learning. Original code is available at SeonghwanSeo/DeepDL.

If you want to train the model with your own dataset, please see the section #train-model.

If you have any problems or need help with the code, please add an issue or contact [e-mail].

TL;DR

# evaluate a molecule with DeepDL
>>> python scoring.py 'CC(=O)Oc1ccccc1C(=O)O'
score: 88.856

# evaluate a molecule with naive setting using another model
>>> python scoring.py 'CC12C(O)CNC13C1CC(N1)C23' --naive -m chemsci-2021
score: 41.399

# screening
>>> python screening.py data/examples/chembl_1k.smi -o out.csv --naive --cuda

• 88.856 is the predicted score. The higher the score, the higher the drug-likeness.
• For fast screening, consider using naive setting, which evaluates a single stereoisomer.
• Multiple models are providen; see #model-list for details.

Installation

# use pip
pip install druglikeness

# from official github (python 3.9-3.13)
git clone https://github.com/SeonghwanSeo/drug-likeness.git
cd drug-likeness
pip install -e .

Python API

from druglikeness.deepdl import DeepDL

# Enter the name of model (see `# Model List`) or the path of your own model.
pretrained_model_name_or_path = 'extended'

# This will download the model weights if you provide the model name.
model = DeepDL.from_pretrained(pretrained_model_name_or_path, device="cpu")

# Evaluate the molecule.
score = model.scoring(smiles='CC(=O)Oc1ccccc1C(=O)O', naive=False)

# Screen the molecules.
score_list = model.screening(smiles_list=['c1ccccc1', 'CCN'], naive=True, batch_size=64)

Model List

Model Name	Description
extended (default)	New model trained on an updated drug database. (excluding test set: FDA-approved drugs)
chemsci-2025	Retrained model of chemsci-2021 with hyperparameter tuning.
chemsci-2021	Finetuned model from the paper (PubChem pretrained, World Drug finetuned).
chemsci-2021-pretrain	Pretrained model from the paper (trained on PubChem)

If your environment is offline, you can manually download the models from Google Drive.

Model Performance

Following shows the scoring performance (AUROC) of the models on the various test datasets.

Model	Mode	FDA vs ChEMBL	FDA vs ZINC15	FDA vs GDB17
extended	strict	0.862	0.961	0.991
extended	naive	0.861	0.961	0.989
chemsci-2025	strict	0.817	0.941	0.984
chemsci-2025	naive	0.817	0.941	0.982

Train Model

You can finetune the model with your own dataset using the pretrained model on PubChem dataset.

pip install -e '.[train]'

# train with the 2.8k training set from the paper
bash ./scripts/download_data.sh
python ./scripts/finetune.py --data_path ./data/train/worlddrug_not_fda.smi

python ./scripts/finetune.py --data_path <smi_file>

Evaluation

# download train/test datasets
>>> bash ./scripts/download_data.sh

# evaluate the model
>>> python ./scripts/evaluate.py --cuda

# Output
Test 1489 molecules in data/test/fda.smi
Average score: 79.40051813170444

Test 1792 molecules in data/test/investigation.smi
Average score: 68.71052120625973

Test 10000 molecules in data/test/chembl.smi
Average score: 64.11581128692627

Test 10000 molecules in data/test/zinc15.smi
Average score: 52.7382759815216

Test 10000 molecules in data/test/gdb17.smi
Average score: 39.37572152862549

AUROC
FDA vs Investigation   : 0.789
FDA vs ChEMBL          : 0.862
FDA vs ZINC15          : 0.961
FDA vs GDB17           : 0.991

Citation

@article{lee2022drug,
  title={Drug-likeness scoring based on unsupervised learning},
  author={Lee, Kyunghoon and Jang, Jinho and Seo, Seonghwan and Lim, Jaechang and Kim, Woo Youn},
  journal={Chemical science},
  volume={13},
  number={2},
  pages={554--565},
  year={2022},
  publisher={Royal Society of Chemistry}
}