Skip to content

Latest commit

 

History

History
150 lines (117 loc) · 6.18 KB

README.md

File metadata and controls

150 lines (117 loc) · 6.18 KB

EnzymeCAGE: A Geometric Foundation Model for Enzyme Retrieval with Evolutionary Insights

EnzymeCAGE

Environment

conda create -n enzymecage python=3.8

pip install pyyaml==6.0
pip install tqdm==4.66.2
pip install numpy==1.24.3
pip install pandas==1.4.2
pip install ase==3.22.1
pip install rdkit-pypi==2022.9.5
pip install fair-esm==2.0.0
pip install biopython==1.83
pip install drfp==0.3.6
pip install mlcrate==0.2.0
pip install rxn-chem-utils==1.5.0
pip install torch==2.1.0 torch-geometric==2.5.3 torch-scatter==2.1.2 torch-cluster==1.6.3 -f https://download.pytorch.org/whl/cu122/torch_stable.html
pip install transformers==4.14.1

Running pipeline

  • Download csv format dataset and alphafold2 structures of enzymes
  • Run AlphaFill to get the catalytic pocket
  • Prepare the feature of enzymes and reactions
  • Train the model, retrieve the candidate enzymes for reactions, do inference and ranking
  • Evaluate the performance of the model

Dataset

First, download the dataset and extract it to the current directory.

Option 1: Directly use the extracted pockets

We have run AlphaFill and pre-extracted the enzyme pockets from the dataset, and you can directly use this part of the data to reproduce the experimental results. The pockets are located in ./dataset/pocket/alphafill_8A

Note: To improve the running speed of AlphaFill, we set the number of homologous proteins to query to 5 (the default is 200). The corresponding parameter is --blast-report-limit=5

Option 2: Rerun AlphaFill to extract the pockets

If you need to calculate features for model training or directly perform inference on your own dataset, you will need to run AlphaFill to extract the pockets yourself.

Step 1: Get the CIF format structure of enzymes

If your dataset only contains protein sequences, you will first need to obtain structural information (preferably in CIF format). You can either deploy colabfold locally or choose to run it online，

Step 2: Deploy AlphaFill

First, download the PDB-REDO database (this may take some time):

cd dataset/PDB-REDO
rsync -av --exclude=attic rsync://rsync.pdb-redo.eu/pdb-redo/ pdb-redo/

Next, download the code and deploy AlphaFill:

cd feature/pkgs
git clone https://github.com/PDB-REDO/alphafill.git
cp ../../scripts/run_alphafill ./alphafill
cd alphafill

Then, follow the tutorial to set up the environment.

Step 3: Run AlphaFill

After setup is complete, run the script run_alphafill.py to execute AlphaFill and extract the pockets:

python scripts/run_alphafill.py \
    --input_dir {CIF_DIR} \ # Folder containing CIF files
    --output_dir {OUT_DIR} \ # Output folder for extracted pockets
    --pdb_fasta {FASTA_PATH} \ # Located in dataset/PDB-RED/pdbredo_seqdb.txt
    --pdb_redo_dir {PDB_REDO_DIR} \ # Located in dataset/PDB-REDO/pdb-redo/

Feature

After extracting pockets, we need to generate features for model training. Since calculating molecular conformations for the full dataset is time-consuming, it is recommended to directly use our precomputed conformation files. For feature computation on test data, calculations can be performed directly.

cd feature
python main.py \
    --data_path {DATA_PATH} \
    --alphafill_result_dir {TRANS_DIR} \
    --pocket_dir {POCKET_DIR} \
    --skip_calc_mol_conformation # If you already have conformation data
  • --data_path: Required, specifies the location of the enzyme-reaction pairs dataset.
  • --alphafill_result_dir: If you re-extracted the pockets using AlphaFill, this should be the path to the AlphaFill output. Note that the CIF files in the result must end with the suffix '_transplant.cif'.
  • --pocket_dir: If you are using our pre-extracted pocket data, set this parameter to the folder containing the precomputed pocket data.
  • --skip_calc_mol_conformation: Add this parameter to skip molecular conformation calculations if you have already computed the conformations or are using our precomputed conformation data.

Feature computation for test data (200 pairs):

cd feature
python main.py \
    --data_path ../dataset/testing/test.csv \
    --alphafill_result_dir ../dataset/testing/alphafill_results \

Feature computation for full data(~1M pairs):

cd feature
python main.py \
    --data_path ../dataset/all_train_test.csv \
    --pocket_dir ../dataset/pocket/alphafill_8A \
    --skip_calc_mol_conformation

Retrieve & Inference

The evaluation processes for Loyal-1968 and Orphan-194 differ slightly. For the Loyal-1968 test set, candidate enzymes have already been assigned to each reaction, allowing us to directly use the trained model to predict the catalytic scores. In contrast, for the Orphan-194 test set, we must first retrieve candidate enzymes for each reaction before using the trained model to predict the catalytic scores. Pretrained models are here.

Retrieve candidate enzymes to orphan reactions:

python retrieve.py --data_path dataset/orphan-rxns/orphan_reactions.csv

Do inference on the retrieved candidate enzymes:

python infer.py --config config/infer/infer_orphan-rxns.yaml

Evaluation

Specify the score file after inference and evaluate the results:

# For test set Orphan-194
python evaluate.py --result_path checkpoints/orphan-rxns/seed_42/orphan_reactions_retrievel_cands_best_model.csv

# For test set Loyal-1968
python evaluate.py --result_path checkpoints/unseen-enzymes/seed_42/test_result.csv

License

No commercial use of the model or data; see LICENSE for details.

Citation

Please cite the following preprint when referencing EnzymeCAGE:

@article{liu2024enzymecage,
  title={EnzymeCAGE: A Geometric Foundation Model for Enzyme Retrieval with Evolutionary Insights},
  author={Liu, Yong and Hua, Chenqing and Zeng, Tao and Rao, Jiahua and Zhang, Zhongyue and Wu, Ruibo and Coley, Connor W and Zheng, Shuangjia},
  journal={bioRxiv},
  year={2024},
  publisher={Cold Spring Harbor Laboratory}
}