mfold

An attempt to reproduce MiniFold in a compact, script-first form.

This project is a lightweight reimplementation inspired by MiniFold: Simple, Fast, and Accurate Protein Structure Prediction. It uses a pretrained language model and a MiniFold checkpoint to predict protein structures from amino-acid sequences and export them as PDB files.

Note: this repository is a reproduction attempt, not the official MiniFold codebase.

---

Features

• Single-sequence structure prediction from a FASTA input file
• Automatic checkpoint download from Hugging Face on first run
• CPU-friendly inference with PyTorch
• PDB export with atom coordinates and residue-level confidence scores
• Simple workflow: drop in a prot.fasta, run the script, get output.pdb

---

System Requirements

Runtime

• Python 3
• PyTorch
• huggingface_hub
• esm

Recommended environment

• A machine with atleast 24 RAM to load the model checkpoint
• Internet access on first run so the checkpoint can be downloaded

Optional tuning

The script respects these environment variables:

• MFOLD_NUM_THREADS
• MFOLD_NUM_INTEROP_THREADS

Example:

set MFOLD_NUM_THREADS=8
set MFOLD_NUM_INTEROP_THREADS=1

---

Installation

Install the required Python packages in your environment:

uv sync

---

Usage

1. Create a file named prot.fasta in the project directory.
2. Put your protein sequence in standard FASTA format.
3. Run the predictor:

python ownmfold.py

On the first run, the model checkpoint is downloaded automatically.

---

Input

The script expects a file named prot.fasta in the working directory.

Required format

• Line 1: FASTA header, for example >protein_name
• Line 2: the amino-acid sequence

Example:

>example_protein
MKTAYIAKQRQISFVKSHFSRQDILD

Important notes

• The sequence is read from the second line of prot.fasta
• Keep the sequence on a single line
• Use the standard 20 amino-acid letters

---

Output

The script writes output.pdb in the current directory.

Output details

• Atomic coordinates are written in PDB format
• Residue confidence is stored in the B-factor column as a pLDDT-style score
• The output currently uses a single chain A

---

Project Notes

• The implementation is intentionally minimal and focused on inference.
• The checkpoint is loaded from the Hugging Face repository jwohlwend/minifold.
• The code is designed to run without gradients and with CPU execution by default.

---

Citation

If you use this project or compare against it, please cite the MiniFold paper:

@article{wohlwend2025minifold,
  title={MiniFold: Simple, Fast, and Accurate Protein Structure Prediction},
  author={Jeremy Wohlwend and Mateo Reveiz and Matt McPartlon and Axel Feldmann and Wengong Jin and Regina Barzilay},
  journal={Transactions on Machine Learning Research},
  issn={2835-8856},
  year={2025},
  url={https://openreview.net/forum?id=1p9hQTbjgo},
  note={Featured Certification}
}

---

License

This project is licensed under the MIT License. See LICENSE for the full text.

---

Acknowledgements

This project is inspired by the MiniFold paper and its released checkpoint. Thanks to the authors for making the model available.