Positive unlabeled learning-based enzyme promiscuity prediction

This repository contains the positive unlabeled learning-based enzyme promiscuity prediction (PU-EPP) model as described in the paper Deep Learning Enables Rapid Identification of Mycotoxin-degrading Enzymes.

Requirements

A Linux workstation with GPUs is essential for deploying PU-EPP. The final PU-EPP model was trained on 5 NVIDIA Tesla V100 GPUs, which took about 2 weeks.

Installation

Dependency

The code has been tested in the following environment:

Package	Version
Python	3.9.12
PyTorch	1.12.0
CUDA	11.6.1
RDKit	2022.3.5
Gensim	4.1.2
Scikit-learn	1.1.3

Install dependencies

Install anaconda first, then install the conda environment by:

conda env create -f PU_EPP_environment.yml
conda activate PU_EPP
pip install --upgrade pip
pip install jupyter
jupyter notebook

Training

Run train.ipynb and specify --class CFG to your own config.

Testing

Run test.ipynb and specify --class CFG to your own config.

Predicting

Screening functional enzymes for a substrate from a .faste file

1. To load the PU-EPP model and make predictions from a .faste file, put the example1.fasta (.fasta file of candidate enzymes) in the data folder,
2. Run predict.ipynb and specify:

• --PreCFG.useFasteFile = True
• --PreCFG.fasteFile Path to the .faste file of candidate enzymes
• --PreCFG.compound The molecular structure of the substrate in simplified molecular input line entry system (SMILES) format

After a few minutes of calculation, you will find the result in the result folder with the name example1_result.csv.

Predicting the probes of enzyme-substrate pairs

1. To load PU-EPP and make predictions for enzyme-substrate pairs, put the example2.csv file (data of enzyme-substrate pairs) in the data folder.

   example2.csv

   Substrate	Enzyme
   SMILES1	SEQ1
   SMILES2	SEQ2

2. Run predict.ipynb and specify:

• --PreCFG.useFasteFile = False
• --PreCFG.csvFile Path to the .csv file of enzyme-substrate pairs

You will find the result in the result folder with the name example2_result.csv.

Fine-tuning

To fine-tune PU-EPP on a new dataset:

1. put the example3_train.csv file (data of enzyme-substrate pairs) and example3_test.csv file in the data folder.

   example3_test.train or example3_test.csv

   Substrate	Enzyme	Label
   SMILES1	SEQ1	1
   SMILES2	SEQ2	0

   Label 1 stands for positive, and 0 stands for negative or unlabeled.

2. Run finetuning.ipynb and specify:

• --CFG.traindata_path Path to the .csv file of the training set
• --CFG.testdata_path Path to the .csv file of the test set
• --CFG.modelsave_file_suffix The suffix of the model name to save
• --CFG.result_file_suffix The suffix of the log file name to save

You will find the fine-tuned model named with --CFG.modelsave_file_suffix as a suffix in the model/model_funetuning folder.

Assistance

For researchers who do not have the hardware to deploy PU-EPP, please send your data in one of the following formats to us ([email protected] or [email protected]). We will then do the calculation and get the results back to you.

1. a substrate and a list of candidate enzymes

2. an enzyme and a list of candidate substrates

3. a list of enzyme-substrate pairs

Link to other repositories

Zenodo, https://doi.org/10.5281/zenodo.7813738