NES Classifier

This repository contains a trained Graph Neural Network (GNN)-based pipeline for detecting Nuclear Export Signals (NES) from protein 3D structures in .pdb format.

---

User Manual

Input:

• A protein .pdb file containing 3D structure, with NES (chain B) and surrounding regions.

Output:

• Classification result:
  • NES POSITIVE or NES NEGATIVE
• Confidence score (value between 0 and 1)

---

Installation

Ensure you have the following installed:

• Python ≥ 3.8
• torch, torch_geometric, biopython, sklearn, matplotlib, tqdm

To install dependencies:

pip install -r requirements.txt

---

Usage

1. Place your .pdb file under any directory (e.g. examples/my_protein.pdb).

2. Run the classifier:

python user_main.py examples/my_protein.pdb

3. Example output:

Prediction: NES POSITIVE  
Confidence score: 0.843

---

How It Works

This tool uses a Graph Neural Network (GNN) to classify proteins based on the spatial proximity and amino acid types of their residues.

• Graph Construction: Each residue becomes a node, and edges are built between nearby residues (within 8Å). Only chain B (NES region) and its surrounding residues (within 15Å) are included in the graph.

• Node Features: Each node has a one-hot vector representing the amino acid type and a binary flag indicating if it belongs to the NES chain.

• Model: The default model is EGNN (Equivariant GNN), trained on labeled positive/negative NES proteins. GCN is also supported.

---

Model Training

To retrain or experiment with parameters, run:

python run.py

You can modify run.py to set:

• batch_size
• epochs
• learning_rate
• hidden_dim
• dropout, etc.

Model checkpoints are saved under Hackaton/.

---

Evaluation

After training, the model generates:

• ROC Curve: Hackaton/roc_curve.png
• Boxplot: Hackaton/boxplot.png

These help visualize model separation power between NES-positive and negative samples.

---

Acknowledgments

This model was developed as part of a protein bioinformatics Hackathon for NES signal detection using 3D structural information and deep learning.