DISCODE

DISCODE (Deep learning-based Iterative pipline to analyze Specificity of COfactors and to Design Enzyme) is a transformer-based NAD/NADP classification model. The model uses ESM-2 language model for amino acid embedding, finally represents the probability of NAD and NADP. This code can also be run on Google Colaboratory.

Repository Overview

This structure helps users easily navigate and understand the DISCODE repository.

• discode.yaml : A YAML file that lists all the necessary dependencies for creating a Conda environment.
• discode/models.py : Contains the model architecture and loading functions.
• discode/utils.py : Utility functions such as sequence processing, model inference, visualization, and mutant design tools.
• example/example.ipynb : A Jupyter notebook for identifying and designing cofactor-switching mutants.
• weights/weights.pt : The weight file that needs to be loaded for using the model.

Installation

Manual install

Note : This code was developed in Linux, and has been tested in Ubuntu 18.04 and 20.04 with Python 3.8.
While the code requires a conda environment for easier use, it can also be used without conda by installing the dependencies listed in the yaml file via pip.

1. Clone github repository

git clone https://github.com/SBML-Kimlab/DISCODE.git

2. Create and activate virtual environment

cd DISCODE
conda env create -f discode.yaml
conda activate discode

Running DISCODE on Google Colaboratory workspace

https://colab.research.google.com/drive/1Gm9QrmYHqLfUqZY0xz6jqfIOZSoqkoiP?usp=sharing

Usage

Preparation

from discode import models, utils

model_path = "weights/weights.pt" # please specify the model weight path
model = models.load(model_path) # if gpu available, it will automatically load on gpu
model.eval() # Model must be specified "eval"

# Q9K3J3 is Streptomyces coelicolor malate dehydrogenase
name, sequence = "Q9K3J3", "MTRTPVNVTVTGAAGQIGYALLFRIASGQLLGADVPVKLRLLEITPALKAAEGTAMELDDCAFPLLQGIEITDDPNVAFDGANVALLVGARPRTKGMERGDLLEANGGIFKPQGKAINDHAADDIKVLVVGNPANTNALIAQAAAPDVPAERFTAMTRLDHNRALTQLAKKTGSTVADIKRLTIWGNHSATQYPDIFHATVAGKNAAETVNDEKWLADEFIPTVAKRGAAIIEARGASSAASAANAAIDHVYTWVNGTAEGDWTSMGIPSDGSYGVPEGIISSFPVTTKDGSYEIVQGLDINEFSRARIDASVKELSEEREAVRGLGLI"

Example of classification

# Predict label of wildtype sequence
# The sequence data will be preprocessed with ESM2-t12 model
data = utils.tokenize_and_dataloader(name, sequence)

# The processed data will be transferred into model, and predict the probability, attention weights, and outlier residues
# The default threshold for selecting outliers is set to 2S
outlier_idx, probability, predicted_label, _name, attention_weights = utils.model_prediction(data, model, threshold="2S")
# The first column of probability is NAD probability, and the second column is NADP probability
print(f"The label probability of NAD is {probability.detach().numpy()[0]:.3f}, NADP is {probability.detach().numpy()[1]:.3f}")
# The label probability of NAD is 0.999, NADP is 0.001

Plot the attention sum and outlier residues

# The default threshold for selecting outliers is set to 2S
utils.plot_attention_sum(attention_weights, sequence, threshold="2S")

Designing Cofactor-Switching Mutants:

Ipynb example of a designing pipeline for cofactor switching mutants is in example/example.ipynb.
We provide various thresholds for outlier selection to accommodate user's purposes and convenience.
[Options]

• max_num_mutation: set the maximum number of mutations to yield cofactor switching mutant (default=3)
• max_num_solution: set the maximum number of solutions to return (default=50)
• prob_thres: set a probability threshold for cofactor specificity reversal (default=0.5)
• pickle_path: directory where a pickle file is saved (default='.')
• sequence: a protein sequence aimed at changing cofactor specificity
• name: sequence id (default='unknown')
• threshold (default="2S") : Specifies the method for selecting outliers
  • Standard deviation-based thresholds : 1S (1 sigma), 2S (2 sigma), 3S (3 sigma)
  • Percentile-based thresholds : IQR (Interquartile Range), P90 (percentile 90), P95 (percentile 95), P99 (percentile 99)
• mode (default=iter_num):
  • iter_prob : scan all combinations of mutations guided by attention analysis and return those for optimal probabilities (exhaustively calculate most probable designs).
  • iter_num : scan all combinations of residues guided by attention analysis. However, if a cofactor switching design is obtained, scan only to combinations from the same number of mutations (exhaustively calculate a minimal requirement of designs).
  • shortest : scan minimum number of combinations by selecting mutations that show optimal probability changes (fastest search).

The results generated by each mutation step will be saved on pickle_path as {name_mode_mutation_step}.pkl

utils.scan_switch_mutation(model = model,
                           max_num_mutation = 3,
                           name = name,
                           sequence = sequence,
                           mode = "iter_num",
                           threshold = "2S",)

Contact

If you have any questions, problems or suggestions, please contact us.

Citation

Kim J., Woo J., Park J.Y., Kim K.J., Kim D. Deep learning for NAD/NADP cofactor prediction and engineering using transformer attention analysis in enzymes. Metab. Eng. 2025; 87:86-94. https://doi.org/10.1016/j.ymben.2024.11.007.

Reference

1. A. Vaswani et al., Attention Is All You Need. Adv Neur In 30 (2017).
2. Z. M. Lin et al., Evolutionary-scale prediction of atomic-level protein structure with a language model. Science 379, 1123-1130 (2023).