AlphaFold-Research

Using AlphaFold and GNN to Predict Catalytic Efficiency of Enzymes

Model Scratch

Pipeline

0. Preparation

1. Create a conda environment with Python>3.7

conda create --name afgnn python=3.8
conda activate afgnn

2. Install required packages

pip install -r requirements.txt

1. Dimensionality Reduction

Use PCA to reduce the dimensionality of the logits in the output file XXXXX.pkl of modified AlphaFold.

XXXXX.pkl is a Python dict object denoting a AlphaFold output of a protein sequence with length of e_dim, whose dict_keys is given by

• distogram,
  • bin_edges
  • logits
• experimentally_resolved,
• masked_msa,
• predicted_lddt,
• representations,
  • msa
  • msa_first_row
  • pair
  • single
  • structure_module
• structure_module,
• plddt

We only need the logits and single for the enzyme features.

1. Modify the ./pca/pca.py to change the AF2_OUTPUT_DIR to the corresponding path where the actual AlphaFold outputs are saved.

2. You may need to change the padding size if the maximum length of the e_dim is greater than 1600 in ./pca/pca.py.

3. Conduct PCA on the pkl in AF2_OUTPUT_DIR

cd pca
nohup python3 pca.py >> log.txt

2. Feature Engineering

Construct the reaction graph as well as the node feature from the raw data.

1. Modify the model_name in the second block of the ./datasets/preprocess.ipynb to match the target dataset (i.e. iMM904, iYO844 or iML1515)

2. Run through the code blocks in ./datasets/preprocess.ipynb

3. [Optional] Modify the train_test_split ratio in the 19th code block to get different size of test set.

The output graph and feature will be saved to ./dataset/<dataset_name>/.

• The node information with enzymes' single representation and molecules' feature are saved to node.pkl as a pandas dataframe.
• The link information is saved to link.dat.
• The training label is saved to label.dat.
• The test label is saved to label.dat.test.
• The logits feature of enzyme nodes are saved to logits/X.npy with X denotes the node index.

3. Network Training

1. Train the modified GCN

cd gnn
nohup python3 train.py --dataset iYO844 --num-layers 2 >> logs/iYO844.txt

Acknowledgements

Code derived and reshaped from HGB.