DrugCLIP for Drug-The-Whole-Genome

Model weights and encoded embeddings

link: https://huggingface.co/datasets/bgao95/DrugCLIP_data

download model_weights.zip, encoded_mol_embs.zip, targets.zip, unzip them and put them inside ./data dir

Set environment

you can set the environment with the Dockerfile in docker dir, or use the requirements.txt file.

Do virtual screening

bash retrieval.sh

You need to set pocket path to ./data/targets/{target}/pocket.lmdb

target is one of the name in ./data/targets

you need to set num_folds to 8 for 5HT2A

The molecule library for the virtual screening is 1648137 molecules inside ChemDIV.

each line in result file look like this:

smiles,score

For data won't fit in mem

To perform screening on chucked mol embedding files, run:

python utils/screening_chucks.py --gpu_num 8 --mol_embs <path_to_multiple_chucks> --zscore_embs <uniform_small_set_for_approx_zscore> --pocket_reps <path_to_pocket_reps_pkl> --batch_size <batch_size> --output_dir <path_to_output_dir> --rm_intermediate

The --mol_embs argument allows multiple args. Each arg should be a path to a chunk of molecule embeddings. The --zscore_embs argument should be a path to a small set of molecule embeddings for approximating the zscore; if not specified, the first mol_embs file is used. The resulting files will be saved as merge{mol_embs_file_id}_{gpu_id}.pkl in output_dir . The --rm_intermediate flag will remove the intermediate files after the next chunk for saving disk space.

To retrieve SMILES strings and original ids for the output files, run:

python utils/retrieve_chunk.py --input_files output/merge*.pkl --mol_lmdb <path_to_mol_lmdb> --output_dir retrieval_results --num_threads <num_threads>

The --mol_lmdb should be extractly the same order as the --mol_embs argument in the last step. The resulting files will be saved as retrieval_results/{pocket_name}.csv.

Benchmarking

link: https://huggingface.co/datasets/bgao95/DrugCLIP_data

download DUD-E.zip, LIT-PCBA.zip, unzip them and put inside ./data dir

bash test.sh

select TASK to DUDE or PCBA in test.sh

Other tools

Pocket Pretraining: https://github.com/THU-ATOM/ProFSA

virtual screening post-processing: https://github.com/THU-ATOM/DrugCLIP_screen_pipeline

Pocket detection: https://github.com/THU-ATOM/Pocket-Detection-of-DTWG