DISCODE

DISCODE (Deep learning-based Iterative pipeline 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 and represents the probability of NAD and NADP specificity.

Installation

Note: Developed and tested on Linux (Ubuntu 18.04+), Python 3.8+.

Option 1: pip install (recommended)

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

Option 2: conda environment

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

Classification

from discode import models, utils

model = models.load("weights/weights.pt")  # automatically loads on GPU if available
model.eval()

name, sequence = "3M6I", "MASSASKTNIG..."
dataloader = utils.tokenize_and_dataloader(name, sequence)

outlier_idx, probability, predicted_label, _name, attention_weights = utils.model_prediction(dataloader, model)
# outlier_idx: zero-indexed positions of salient residues
# probability: [NAD_prob, NADP_prob]

print(f"NAD: {probability.numpy()[0]:.3f}, NADP: {probability.numpy()[1]:.3f}")

Visualization

utils.make_max_attention_map(attention_weights)
# Plots maximum attention map across all layers/heads [8 x 20]

utils.plot_attention_sum(attention_weights, sequence, threshold="2S")
# Plots per-residue attention sum with outlier threshold line

Supported thresholds: "1S", "2S", "3S", "IQR", "P90", "P95", "P99"

Mutation Design

Results are saved at each step as {pickle_path}/{name}_{mode}_mutation_{step}.pkl.

utils.scan_switch_mutation(
    model=model,
    name=name,
    sequence=sequence,
    mode="shortest",       # "shortest", "iter_num", or "iter_prob"
    max_num_mutation=3,
    max_num_solution=20,
    prob_thres=0.5,
    threshold="2S",
    batch_size=32,
    pickle_path=".",
)

Arguments:

Argument	Default	Description
model	—	Loaded TransformerClassifier model
name	"unknown"	Protein accession or identifier
sequence	—	Wildtype amino acid sequence (single-letter)
mode	"iter_num"	Mutation search mode (see below)
max_num_mutation	3	Maximum number of mutations per design (search depth)
max_num_solution	50	Maximum number of results returned in the final DataFrame
prob_thres	0.5	Probability threshold for cofactor switching (used in iter_prob mode)
threshold	"2S"	Outlier detection method for selecting salient residues ("1S", "2S", "3S", "IQR", "P90", "P95", "P99")
batch_size	32	Number of mutation candidates processed per GPU batch
pickle_path	"."	Directory to save intermediate results as .pkl files

Modes:

• shortest — greedily selects the single highest-probability mutation at each step. Fastest; finds the minimal mutation path.
• iter_num — exhaustively scans all candidate combinations and stops as soon as at least one converting mutation is found.
• iter_prob — exhaustively scans all combinations across all steps and returns all results sorted by switching probability.

A complete example is provided in example/example.ipynb.

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).