pepTrans

Embedding-Based Transformer Framework for Multi-Level Peptide–Protein Interaction Prediction

---

Overview

pepTrans is a transformer-based deep learning framework designed for comprehensive peptide–protein interaction (PepPI) analysis using only amino acid sequences.

The framework integrates large-scale pretrained protein language model (PLM) embeddings with task-specific convolutional neural networks (CNNs) to perform multiple peptide–protein interaction prediction tasks without requiring structural information, molecular docking, or handcrafted features.

Unlike traditional structure-dependent approaches, pepTrans learns interaction-relevant representations directly from protein and peptide sequences, enabling scalable and high-throughput prediction across diverse biological applications.

---

Key Features

✅ Sequence-only prediction framework

✅ No requirement for 3D structures

✅ No handcrafted biochemical features

✅ Transformer-based protein language model embeddings

✅ Multi-task peptide–protein interaction prediction

✅ Strong generalization to unseen proteins and peptides

✅ Competitive performance against AlphaFold3-associated evaluation pipelines

✅ Superior performance compared with several structure-based docking methods

✅ Suitable for large-scale peptide therapeutic discovery

---

Supported Tasks

The released repository contains pretrained models for:

Task	Description
Binary PepPI Prediction	Predict whether a peptide interacts with a protein
Peptide Binding Residue Prediction	Identify interaction-responsible residues within peptides
Peptide–Protein Binding Affinity Prediction	Estimate interaction strength
Peptide Virtual Screening	High-throughput candidate ranking
Peptide–PBD Prediction	Predict peptide interactions with protein binding domains
Virtual Alanine Scanning	Assess residue contributions to binding

---

Architecture

pepTrans combines:

1. Protein Language Models

• ProtT5-XL-U50
• Transformer encoder representations
• Context-aware residue embeddings

2. Convolutional Neural Networks

Task-specific CNN modules are used to capture:

• Local residue motifs
• Interaction signatures
• Spatial sequence patterns
• Long-range contextual information

3. Multi-Level Prediction Heads

The learned representations are used for:

• Binary interaction prediction
• Residue-level binding prediction
• Affinity estimation
• Virtual screening

---

Model Workflow

Protein Sequence
        │
        ▼
 ProtT5 Embedding
        │
        ▼
 Protein CNN Module
        │
        ├────────────┐
        │            │
        ▼            ▼

Peptide Sequence
        │
        ▼
 ProtT5 Embedding
        │
        ▼
 Peptide CNN Module

        ▼
 Feature Fusion
        ▼
 Fully Connected Layers
        ▼

 ┌─────────────────────────┐
 │ Binary PepPI Prediction │
 └─────────────────────────┘

 ┌─────────────────────────┐
 │ Binding Residue Mapping │
 └─────────────────────────┘

---

Performance Highlights

pepTrans was evaluated using benchmark datasets and independent external test sets.

Binary Interaction Prediction

• Consistently outperformed:

  • CAMP
  • DeepDTA
  • PIPR
  • NRLMF

• Demonstrated strong performance under:

  • Novel proteins
  • Novel peptides
  • Novel peptide–protein pairs

Binding Residue Prediction

pepTrans achieved:

• Average MCC ≈ 0.55
• Average AUC ≈ 0.77

on independent peptide–protein complexes.

Comparison Against AlphaFold3-Based Evaluation

pepTrans demonstrated:

• Higher average MCC
• Competitive AUC
• More stable prediction distributions

while requiring only sequence information and significantly lower computational cost.

Virtual Screening

pepTrans outperformed:

• GalaxyPepDock
• AutoDock CrankPep
• CABS-Dock
• MDockPep
• CAMP

on benchmark virtual screening datasets.

---

Repository Structure

pepTrans/
│
├── Binary pepPIs prediction/
├── Binding Affinity/
├── Generalizability/
├── Peptide Binding Residues/
├── Peptide PBD Prediction/
└── Peptide Virtual Screening/

Each directory contains task-specific pretrained weights and checkpoints.

---

Model Weights

Important Notice

GitHub imposes storage limitations for large deep learning model files.

To ensure long-term availability and reproducibility, all pretrained pepTrans weights are hosted on Hugging Face.

Official Model Repository

👉 https://github.com/SyedKumailHussainNaqvi/pepTrans/tree/main

Researchers should download all model checkpoints directly from this repository.

---

Scientific Impact

pepTrans advances peptide–protein interaction modeling by:

• Eliminating dependence on experimental structures
• Enabling scalable peptide screening
• Improving residue-level interpretability
• Supporting peptide therapeutic discovery
• Facilitating large-scale interaction prediction

The framework provides a practical alternative to computationally intensive structure-based pipelines while maintaining competitive predictive performance.

---

Citation

If you use pepTrans in your research, please cite:

@article{Naqvi2026pepTrans,
  title={pepTrans: Embedding-Based Transformer Framework for Multi-Level Peptide–Protein Interaction Prediction},
  author={Naqvi, Syed Kumail Hussain and Cho, Hwangeui and Chong, Kil To and Tayara, Hilal},
  journal={Under Review},
  year={2026}
}

---

Authors

Syed Kumail Hussain Naqvi
Department of Physical-AI Convergence
Jeonbuk National University, Republic of Korea

Hwangeui Cho
School of Pharmacy
Jeonbuk National University, Republic of Korea

Kil To Chong
Jeonbuk National University, Republic of Korea

Hilal Tayara
School of International Engineering and Science
Jeonbuk National University, Republic of Korea

---

Contact

For questions regarding:

• pretrained weights
• model reproduction
• datasets
• benchmarking
• collaborations

please contact the Syed Kumail Hussain Naqvi .

---

pepTrans

Advancing sequence-based peptide–protein interaction modeling through transformer-powered representation learning.