Accurate, database-free protein folding from single sequences using structure-aware synthetic MSAs

GhostFold is a next-generation protein folding framework that predicts 3D structures directly from single sequences — without relying on large evolutionary databases. By generating synthetic, structure-aware multiple sequence alignments (MSAs), GhostFold achieves high accuracy while remaining lightweight and portable.

This repository provides scripts and configurations to set up GhostFold locally using Mamba and ColabFold environments, along with integration support for Hugging Face Transformers.

GhostFold uses Mamba for virtual environment management.

You can install Mamba using one of the following methods:

conda install -n base -c conda-forge mamba

# For Linux or macOS (ARM/x86)
wget https://github.com/conda-forge/miniforge/releases/latest/download/Mambaforge-Linux-x86_64.sh
bash Mambaforge-Linux-x86_64.sh

Follow the on-screen prompts to complete installation, then restart your terminal or run:

source ~/.bashrc

git clone https://github.com/brineylab/ghostfold.git
cd ghostfold

mamba create -n ghostfold python=3.10
mamba activate ghostfold

If you receive an error indicating Mamba isn’t initialized, activate it manually:

source ~/.bashrc

GhostFold relies on the following core libraries:

torch
transformers
sentencepiece

Install them using:

mamba install pytorch torchvision torchaudio -c pytorch
mamba install transformers sentencepiece -c conda-forge

Install the appropriate CUDA drivers for PyTorch.

Refer to the Transformers Installation Guide for platform-specific setup details.

We use Rich for printing log messages. Install it within the environment with:

pip install rich

When using from_pretrained() to load models, GhostFold automatically fetches pretrained weights from the Hugging Face Hub. You may need to create and configure a Hugging Face access token. See the instructions here.

huggingface-cli login

To enable local structure prediction, first ensure the setup scripts are executable:

chmod +x install_localcolabfold.sh ghostfold.sh

Then run the installation script:

./install_localcolabfold.sh

This script will:

• Configure a compatible ColabFold environment
• Install all required dependencies
• Download model weights automatically

If you prefer to run predictions via Google Colab, you can use the generated pseudoMSAs directly in ColabFold by selecting “custom_msa” under MSA settings.

Once setup is complete, ensure all scripts are executable:

chmod +x ghostfold.sh

Then, launch a prediction using:

./ghostfold.sh --project_name <your_project_name> --fasta_file <path/to/your/fasta_file.fasta>

GhostFold includes a sample FASTA file for demonstration:

./ghostfold.sh --project_name 7JJV --fasta_file query.fasta

GhostFold can operate in two primary modes:

1. MSA Generation Mode – Generates structure-aware synthetic MSAs only.
2. Full Mode – Runs both synthetic MSA generation and structure prediction.

You may run folding separately after generating MSAs. To explore the available options:

./ghostfold.sh --help

• ProstT5: Protein Language Modeling
• ColabFold: AlphaFold Simplified