AgGRU-Check

Identify and visualize amyloidogenic domains in proteins using gated recurrent units. Try it live at aggru-check.fly.dev!

Read more about this model in details.pdf.

Overview

AgGRU-Check uses an ensemble of bidirectional GRUs to predict and analyze amyloidogenic regions in protein sequences. The model combines sliding window analysis with ensemble predictions to provide:

• Position-specific amyloid propensity scores
• Confidence measures based on ensemble agreement
• Interactive visualization of amyloidogenic domains
• Detailed window-based structural analysis
• Exportable results in CSV format

Features

Advanced Analysis

• Multi-scale window analysis (15-27 residues)
• Ensemble prediction with confidence scoring
• Position-specific saliency mapping

Interactive Visualization

• Color-coded propensity scores
• Detailed residue-level tooltips
• Window analysis overlays
• Confidence indicators
• Sequence position markers

Input Support

• Raw amino acid sequences
• FASTA format with headers
• Multiple sequence analysis
• Batch processing with streaming results

Project Structure

.
├── deployment/                  # Deployment-ready code
│   ├── backend/                # FastAPI server
│   │   └── main.py            # Core analysis engine
│   └── frontend/              # SvelteKit UI
│       └── src/
│           ├── routes/        # Main pages
│           └── lib/           # Shared components
├── details.pdf                # Technical paper
└── amyloidClassification.ipynb # Training notebook

Quick Start

Using the Web Interface

1. Visit aggru-check.fly.dev
2. Paste your protein sequence(s)
3. Click "Predict" to see interactive results
4. Hover over residues for detailed predictions
5. Download results as CSV if needed

Example input:

>TDP-43 (amyloid domain) [318-343]
INPAMMAAAQAALKSSWGMMGMLASQ

Local Development

Backend

cd deployment/backend
python -m venv venv
source venv/bin/activate  # or `venv\Scripts\activate` on Windows
pip install -r requirements.txt
uvicorn main:app --reload

Frontend

cd deployment/frontend
npm install
npm run dev

API Usage

The API accepts POST requests and streams real-time analysis results:

curl -X POST https://aggru-check.fly.dev/api/predict \
  -H "Content-Type: application/json" \
  -d '{"sequenceList": [">TDP-43\nINPAMMAAA"]}'

The API streams results as Server-Sent Events (SSE) including:

• Real-time position-specific scores
• Per-model analysis progress
• Ensemble agreement metrics
• Position-specific confidence scores
• Detailed window analysis results

Example response events:

{"type": "sequence_start", "sequence_index": 0, "total_models": 5}
{"type": "position_result", "position": 10, "saliency": 0.85, "confidence": 0.92}
{"type": "model_complete", "sequence_index": 0, "model_index": 1}

Training Data

Before running the training notebooks, download these datasets:

• GPD_proteome.faa
• GPD_proteome_orthology_assignment.txt

Available at: Gut Phage Database

Technical Details

• Model Architecture:

  • Bidirectional GRU ensemble
  • Multi-scale window analysis

• Analysis Methodology:

  • Adaptive Window Analysis: Uses multiple window sizes (15, 21, 27 residues) to capture both local and broader structural patterns
  • Context-Aware Scoring:
    • For amyloid sequences: Identifies critical regions by masking windows and measuring prediction impact
    • For non-amyloid sequences: Requires stronger evidence with higher thresholds for positive predictions
  • Position Scoring:
    • Gaussian weighting around window centers
    • Nonlinear impact scaling to emphasize strong signals
    • Conservative score aggregation using weighted max/mean combinations
  • Confidence Calculation:
    • Based on ensemble agreement and prediction variance
    • Scales with number of models completed
    • Incorporates standard deviation of predictions

• Stack:

  • Backend: PyTorch, FastAPI
  • Frontend: SvelteKit, TypeScript
  • Deployment: Docker + Fly.io

Citation

If you use AgGRU-Check in your research, please cite:

@software{aggru_check_2024,
  author = {Mockler, Evan},
  doi = {10.5281/zenodo.13147167},
  title = {AgGRU-Check: Interactive Analysis of Amyloidogenic Domains},
  url = {https://github.com/ejmockler/AgGRU-Check}
}

License

MIT