This repository consists of following two parts:
- Running active learning pipeline for ultra-large-scale docking
- Anlyzing why active learning was so effective
Technical details and thorough analysis can be found in our paper, Understanding active learning of molecular docking and its applications, written by Jeonghyeon Kim, Juno Nam and Seogok Ryu. If you have any question, feel free to open an issue or reach out at [email protected].
conda create -n activelearning python=3.10 numpy scipy matplotlib pandas scikit-learn pytorch pytorch-cuda=11.7 cuda=11.7 dgl parmap openbabel rdkit -c pytorch -c dglteam/label/cu117 -c nvidia -c conda-forge --override-channels
conda install conda-forge::plip
conda activate activelearningedit_dataset.py is to acquire the molecules to dock following acquisition function. train_map.py is to train your model on the acquired molecules before. inference_map.py is to inference on remaining dataset.
If you use slurm, submit_active_learning.py would write down whole active learning pipeline script. Following is the example.
python submit_active_learning.py --title Test --csv_path Enamine_HTS.csv --num_iter 10 Before you start training, you need to dock acquired molecules in to prepared receptor. All the scripts need for docking is in scripts/prepare and scripts/docking.
All the analysis performed in our paper is in scripts/analysis. They are seperated into the figure number wrote down in our paper.
-
fig2-6: Analysis about model's RMSE,$R^2$ , Success rate, and ordering. -
fig7: Analysis about 3D pose similarity of docked molecules into each receptor. -
fig9: Analysis about interaction pattern of top scored compounds. -
tab3: Linear factor analysis between number of functional group and docking score. -
fig11: Calculating AUROC of DUD-E active and decoy set when using surrogate model for virtual screening -
fig12: Compare ability to acquire higher docking score between fingerprint screening and our model's inference. Chemical space visualization using t-sne also.