README_APP.md

miRBind2 Gradio Web Interface

Interactive web application for predicting miRNA-mRNA binding sites with explainability visualization.

Features

Single Prediction Mode

• Simple Interface: Enter miRNA and mRNA sequences to get binding predictions
• Real-time Predictions: Uses pre-trained Pairwise CNN model (AUPRC: 84.97)
• Sequence Visualization: Color-coded nucleotide importance bars
• SHAP Explainability: Interactive heatmap showing which positions contribute to binding prediction
• PDF Reports: Download comprehensive reports with all visualizations

Batch Prediction Mode

• File Upload: Process multiple pairs from TSV/CSV files
• Interactive Table: Browse all predictions with sorting/filtering
• Click-to-View: Select any row to see detailed SHAP analysis
• Progress Tracking: Real-time progress bar during processing
• Accuracy Metrics: Automatic calculation if labels provided

General

• Local Execution: Runs entirely on your machine, no internet required
• GPU Support: Automatically uses CUDA if available, falls back to CPU
• Unified Interface: Both modes in one app with tabs

Installation

Prerequisites

• Python 3.8 or higher
• pip package manager

Quick Start

1. Install dependencies:

pip install -r requirements.txt

2. Run the application:

python app.py

3. Open your browser:

The app will automatically open at http://127.0.0.1:7860

If it doesn't open automatically, copy the URL from the terminal.

Usage

The app has two tabs: Single Prediction and Batch Predictions

Tab 1: Single Prediction

1. Enter your miRNA sequence (e.g., UAGCUUAUCAGACUGAUGUUGA)
2. Enter your mRNA target sequence
3. Click "Predict Binding"
4. View results:
   • Binding score (0-1)
   • Prediction (BINDING/NO BINDING)
   • Confidence level
   • Download PDF report

With Explainability (Single Mode)

1. Check the "Show SHAP Explainability" checkbox
2. Run prediction (will take a few seconds longer)
3. View:
   • Sequence explainability bars: Per-nucleotide importance
   • SHAP heatmap: Position-pair contributions
   • PDF report: Download everything

Tab 2: Batch Predictions

1. Upload file: Click "Upload TSV/CSV File"
   • Format: target_seq, mirna_seq, label (optional)
   • See example_batch.tsv for reference
2. Configure: Check "Compute SHAP" if you want explainability (optional)
3. Process: Click "Process File"
4. Browse: View results table with all predictions
5. Detail: Click any row to see full SHAP analysis

Batch Processing Tips

Fast screening (100+ pairs):

• Upload file
• Leave SHAP unchecked
• Process → Get results in seconds
• Identify interesting pairs from table

Detailed analysis (5-20 pairs):

• Upload file with top candidates
• Check SHAP enabled
• Process → Click each row for full analysis

Input Guidelines

• Accepted nucleotides: A, T, C, G, U, N
• Both RNA (U) and DNA (T) formats are accepted
• Sequences are automatically:
  • Converted to uppercase
  • U → T conversion
  • Padded to 28 nt (miRNA) and 50 nt (mRNA)
  • Trimmed if too long

Example Sequences

miRNA example:

UAGCUUAUCAGACUGAUGUUGA

mRNA target example (with an embedded complementary site for clearer explainability):

GGGCACUUUUUCAACAUCAGUCUGAUAAGCUAAGUGUCUUCCAGGGAAUU

Model Information

Architecture: Pairwise CNN with One-Hot Encoding

Training Data: AGO2 eCLIP (Manakov et al. 2022)

Performance:

• Test AUPRC: 84.97
• Leftout AUPRC: 83.08

Model Parameters:

• miRNA length: 28 nucleotides
• Target length: 50 nucleotides
• Embedding dimension: 8
• Filter sizes: [128, 64, 32]
• Kernel sizes: [6, 3, 3]

Technical Details

File Structure

mirbind2_web/
├── app.py                          # Main Gradio application
├── requirements.txt                # Python dependencies
├── README_APP.md                   # This file
├── miRBind_2.0-main/              # Original miRBind2 repository
│   ├── code/                       # Model and encoding code
│   │   └── pairwise_binding_site_model/
│   │       └── shared/             # Shared utilities
│   └── models/                     # Pre-trained models
│       └── pairwise_onehot_model_20260105_200141.pt

Performance Optimization

• GPU Acceleration: Automatically uses CUDA if available
• Batch Processing: Efficient tensor operations
• Caching: Model loaded once at startup

SHAP Computation

SHAP (SHapley Additive exPlanations) uses GradientShap to compute feature attributions:

• Compares input against baseline (all zeros)
• Computes gradients to determine importance
• Aggregates across nucleotide pair dimension
• Visualizes as 2D heatmap (miRNA × mRNA positions)

Troubleshooting

Model Not Found

If you see Model not found error:

# Check model exists
ls miRBind_2.0-main/models/pairwise_onehot_model_20260105_200141.pt

Download from the model repository if missing.

SHAP Not Available

If SHAP visualization doesn't work:

pip install captum

Out of Memory (GPU)

If you encounter CUDA out of memory errors:

• Close other GPU-intensive applications
• Or disable SHAP (uncheck the box)
• The model will automatically fall back to CPU

Port Already in Use

If port 7860 is busy, edit app.py:

app.launch(
    server_port=7861,  # Change to different port
    ...
)

Advanced Usage

Sharing Publicly

To create a temporary public link (for collaboration):

Edit app.py line 368:

app.launch(
    share=True,  # Creates public Gradio link
    ...
)

Warning: This shares your model publicly. Only use for trusted collaborators.

Custom Model

To use a different model:

1. Update model path in app.py (line 50):

model_path = Path(__file__).parent / "path/to/your/model.pt"

2. Ensure model type matches (pairwise_onehot or pairwise)

Batch Predictions

For multiple predictions, consider using the original inference scripts:

cd miRBind_2.0-main/code/pairwise_binding_site_model
python -m inference.predict --help

Citation

If you use miRBind2 in your research, please cite:

miRBind 2.0: Advanced miRNA Target Prediction
[Citation details to be added]

License

See LICENSE file in the miRBind_2.0-main directory.

Support

For issues or questions:

• Check the troubleshooting section above
• Review the original miRBind2 repository README
• Open an issue on the GitHub repository

---

Developed using: Gradio, PyTorch, Captum (SHAP)