GMProt: Gated Modulation Protein Language Model for Prediction of Antimicrobial Peptide Activity

GMProt is a flexible pipeline for training, prediction, and visualization of feature contributions in antimicrobial peptide activity modeling, including the estimation of Minimum Inhibitory Concentration (MIC). This framework accompanies and supports our work, “Gated Protein Language Modeling for Accurate Prediction of Antimicrobial Peptide Activity.”

For a comprehensive description of the approach, please refer to the paper: [link here]

---

📑 Data Availability

The complete dataset is publicly available on Zenodo:

• DOI: https://doi.org/10.5281/zenodo.18626920

📁 Folder Structure

├── alphafold_pdb_ecoli/
├── alphafold_pdb_stap/
├── fasta_ecoli/
├── fasta_stap/
├── five_fold_ecoli/
├── five_fold_s_aureus/
├── protT5/
└── grampa.csv

alphafold_pdb_ecoli

Predicted three-dimensional peptide structures (PDB format) targeting Escherichia coli, generated using AlphaFold2.

alphafold_pdb_stap

Predicted peptide structures (PDB format) targeting Staphylococcus aureus, generated using AlphaFold2.

fasta_ecoli and fasta_stap

FASTA-formatted peptide sequence datasets for:

• E. coli
• S. aureus

five_fold_ecoli

Preprocessed datasets stored in .pkl format for model training.

• Each file contains five cross-validation splits
• Multiple .pkl files represent alternative feature configurations
• See readme.txt for feature details
• ecoli_dataset.pkl is the final dataset used for ModProt training

five_fold_s_aureus

Five-fold cross-validation datasets for S. aureus.

protT5

Residue-level peptide embeddings computed using ProtT5 for both:

• E. coli
• S. aureus

grampa.csv

The original antimicrobial peptide dataset containing sequences across multiple bacterial species, including:

• E. coli
• S. aureus All derived datasets were constructed from this base file.

Dataset Attribution

The original grampa.csv dataset is attributed to:

Deep learning regression model for antimicrobial peptide design
https://www.biorxiv.org/content/10.1101/692681v1.full

---

Features

• Train the ModProt model on one or more datasets.
• Predict activity values and evaluate model performance on test datasets.
• Plot feature contributions to interpret model predictions.
• Flexible handling of multiple datasets via indices.
• Modular design for easy extension to other models and datasets.

---

Installation

1. Clone the repository:

git clone https://github.com/SCSE-Biomedical-Computing-Group/GMProt.git

2. Install dependencies:

pip install -r requirements.txt 

3. Usage

python train.py <mode> [--model MODEL_NAME] [--datasets DATASET_INDICES]

Argument	Description
mode	Required. Mode to run: train, predict, or plot.
--model	Optional. Model name. Default: MoPro.
--datasets	Optional. List of dataset indices (supports multiple). Default: [0].

4. Examples:

Train mode (default dataset [0]):
python train.py train

Predict mode 
python train.py predict --datasets 0 1 2 3 4

Plot feature contributions for specific datasets:
python train.py plot --datasets 0

Plots are saved to ./visualization/IG/fold_<dataset_index>.

5. Directory Structure:

├── train.py                  # Main script with argparse
├── model/                    # Pretrained or saved models
├── data/                     # Complte datasets
├── visualization/
│   └── IG/                   # Feature contribution plots
├── requirements.txt          # Python dependencies
└── README.md

Notes:

The mode argument is mandatory. Use one of train, predict, or plot. For visualization, the model must have been trained and saved beforehand. Ensure all dependencies and datasets are properly prepared before running.

Citation

If you use this code, or dataset in your research, please cite our paper:

Prem Singh Bist, et al. (2026). **. Journal/Conference Name. DOI: 10.xxxx/xxxx