Structure-guided prediction of catalytic residues in enzymes by deep learning

Background

We introduced SCREEN (Structure-based Catalytic Residue Recognition in Enzymes) for precise and high-throughput prediction of catalytic residues across a wide range of taxonomic enzyme groups. This framework integrates heterogeneous data sources regarding enzyme structures, functions and evolutionary profiles.

Note: A streamlined version named SCREEN-Light, designed for faster training and minimal software dependencies, is available at https://github.com/BioColLab/SCREEN-Light.

Installation

SCREEN is implemented using Python 3.6.13 and various scientific computing packages (numpy, biopython, scikit-learn, torch). We recommend creating a dedicated virtual environment and installing the same package versions, see the full list in requirements.txt. Command lines that can be helpful in setting up the environment (tested on linux with python 3.6), this can be done in few lines using Anaconda (https://www.anaconda.com).

conda create --name SCREEN_env python=3.6
conda activate SCREEN_env
conda install pytorch==1.7.0 torchvision==0.4.0 torchaudio==0.7.0 -c pytorch 
pip install -r requirements.txt

Additionally, SCREEN utilizes a pre-trained language model protT5, BLAST, and HMMER for protein residual embedding and evolutionary information extraction. The protT5, PSIBLAST and HHM must be installed and appropriately linked.

(1) Install protT5

The language model protT5 (referring to https://github.com/agemagician/ProtTrans) can be installed via

pip install transformers
pip install sentencepiece

(2) Install PSIBLAST, HHblits, DSSP

** Install BLAST+ for extracting PSSM (position-specific scoring matrix) profiles!

To download and install the BLAST+ package (https://ftp.ncbi.nlm.nih.gov/blast/executables/LATEST/) and BLAST database, Please refer to BLAST® Help (https://www.ncbi.nlm.nih.gov/books/NBK52640/). Set the absolute paths of PSIBLAST and databases in the script "feature_extract.py".

** Install HHblits for extracting HMM profiles!

To install HHblits and download uniclust30_2018_08(http://wwwuser.gwdg.de/~compbiol/uniclust/2018_08/uniclust30_2018_08_hhsuite.tar.gz) for HHblits, please refer to https://github.com/soedinglab/hh-suite. Set the absolute paths of HHBLITS and uniclust30_2018_08 databases in the script "feature_extract.py".

** Install DSSP for extracting SS (Secondary structure) profiles! Set the absolute path of dssp in the script "feature_extract.py".

Note: Difference versions of blast+, HHblits and their databases may result in slightly different PSSM and HMM profiles, leading to slight different predictions. Typical download databases and install bioinformatics tools time on a "normal" desktop computer is 10 hours.

Enzyme feature extraction pipline

To extract the handcrafted enzyme features, please use:

python feature_extract.py

The main stages of enzyme feature extraction are:

1. Writing the enzyme sequence fasta file
2. Downloading the pdb files
3. Preprocessing of the pdb files
4. Contact map construction according to the residue 3D coordination
5. MSA construction and PWM calculation
6. Atom feature extraction
7. Protein residual level embedding using language model ProtT5

Note: Paths to the PSIBLAST and HHBLITS are defined in feature_extract.py.

Predicting catalytic residues using evolutionary information

The script for predicting catalytic residues with experimentally determined enzyme structure is predict.py, see some examples:

python predict_SCREEN.py --PDBfile ./Example/PDB_id.txt

The first argument is:

• A .txt file containing PDB ID (e.g. Example/PDB_id.txt);

This file can contain multiple PDB_ids. A specific chain or list of chain can be added to the ID with an underscore (e.g. 1brs_A, 2p6b_B).

Visualizing the outputs

Here we provide an example of outputs obtained by predicting catalytic residues on the 1aui_A protein

• A Visualization/1aui-A.py script for visualizing the results with Pymol.
• The PDB file with catalytic residue probabilities inserted in the B-factor field

Below is a visualization of predicrted outputs for 1aui_A protein.

Retraining SCREEN

Catalytic residue data sets are provided in the Dataset/training_data folder, which included:

• the PDB_ids used for training in text format (training_id_withEC.txt).
• the label files in text format (training_label.txt).

Training from scratch on the enzyme catalytic residue data set

This can be done via:

python train_SCREEN.py

The main stages of train.py are:

1. Parsing the enzyme catalytic residue label files
2. Loading the extracted features of enzymes
3. Training of the network
4. Prediction on the validation sets.

Testing on five test data sets

This can be done via:

python test_SCREEN.py

All source data in the paper can be found at https://huggingface.co/datasets/Biocollab/SCREEN/tree/main.

Contact

For any question and comment regarding the code, please reach out to tong.pan@monash.edu