README.md

metadata

title: MXfold2 RNA Structure Predictor (mxfold2-rna-structure)
emoji: 🧬
colorFrom: blue
colorTo: green
sdk: docker
pinned: false
license: mit
tags:
  - rna
  - secondary-structure
  - bioinformatics
  - deep-learning
  - pytorch

MXfold2: RNA Secondary Structure Prediction

This Hugging Face Space provides an interactive interface for MXfold2, a state-of-the-art RNA secondary structure prediction tool that uses deep learning with thermodynamic integration.

Features

• Deep Learning + Thermodynamics: Combines neural networks with Turner thermodynamic parameters
• High Accuracy: Trained on comprehensive RNA structure datasets
• Fast Predictions: Optimized for quick structure prediction
• Multiple Models: Choose between hybrid model or pure thermodynamic model
• API Access: Can be called programmatically via Gradio API

Usage

Web Interface

1. Enter your RNA sequence (using nucleotides: A, U, G, C)
2. Select a prediction model (default TrainSetAB recommended)
3. Click "Predict Structure"
4. View the predicted dot-bracket structure and minimum free energy (MFE)

API Access

You can call this Space programmatically using the Gradio Python client:

from gradio_client import Client

client = Client("joeyisgoed/mxfold2-rna-structure")
result = client.predict(
    sequence="CGCGAAUUUGCG",
    model_choice="Default (TrainSetAB)",
    api_name="/predict"
)

structure, mfe, visualization, status = result
print(f"Structure: {structure}")
print(f"MFE: {mfe} kcal/mol")

Using cURL

curl -X POST https://joeyisgoed-mxfold2-rna-structure.hf.space/api/predict \
  -H "Content-Type: application/json" \
  -d '{"data": ["CGCGAAUUUGCG", "Default (TrainSetAB)"]}'

Structure Notation

The output uses dot-bracket notation:

• ( and ) represent base pairs forming helices
• . represents unpaired nucleotides (loops, bulges)

Example:

Sequence:  GGAUGGAUGUCUGAGCGG...
Structure: (((((((........((((
           ↑       ↑        ↑
         Helix    Loop    Helix

Models

Default (TrainSetAB) - Recommended

• Hybrid model combining deep learning and thermodynamics
• Trained on TrainSetA and TrainSetB datasets
• Best overall accuracy for diverse RNA sequences

Turner2004 (Thermodynamic only)

• Pure thermodynamic model using Turner parameters
• No machine learning component
• Fast and interpretable

Technical Details

Method: MXfold2 integrates:

• Deep neural networks (CNN/LSTM/Transformer)
• Turner 2004 thermodynamic parameters
• Zuker dynamic programming algorithm

Input Constraints:

• Sequence length: 3-5000 nucleotides
• Valid nucleotides: A, U, G, C (case insensitive)
• FASTA format not required (plain sequence accepted)

Output:

• Dot-bracket structure notation
• Minimum Free Energy (MFE) in kcal/mol
• Lower MFE = more stable structure

Citation

If you use MXfold2 in your research, please cite:

@article{sato2021mxfold2,
  title={RNA secondary structure prediction using deep learning with thermodynamic integration},
  author={Sato, Kengo and Akiyama, Manato and Sakakibara, Yasubumi},
  journal={Nature Communications},
  volume={12},
  number={1},
  pages={941},
  year={2021},
  publisher={Nature Publishing Group},
  doi={10.1038/s41467-021-21194-4}
}

Links

• Paper: Nature Communications (2021)
• GitHub: mxfold/mxfold2
• Original Web Server: http://www.dna.bio.keio.ac.jp/mxfold2/

License

This Space uses MXfold2, which is licensed under the MIT License.

Acknowledgments

Developed by Kengo Sato, Manato Akiyama, and Yasubumi Sakakibara at Keio University.

Hugging Face Space deployed by joeyisgoed.