GhostFold

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

---

Overview

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.

---

Installation

1. Install PyTorch with CUDA

GhostFold requires PyTorch with CUDA support. Install the appropriate version for your system before installing GhostFold:

# Example for CUDA 12.1 (adjust for your CUDA version)
pip install torch --index-url https://download.pytorch.org/whl/cu121

Refer to the PyTorch installation guide for platform-specific instructions.

2. Install localcolabfold

Operations that involve protein structure prediction (ghostfold run and ghostfold fold) require a working local ColabFold runtime. From the GhostFold repository root:

chmod +x scripts/install_localcolabfold.sh
./scripts/install_localcolabfold.sh

If you prefer cloud-based structure prediction, you can use the generated pseudoMSAs directly in ColabFold by selecting "custom_msa" under MSA settings and uploading the pseudoMSA generated by GhostFold.

3. Install GhostFold

pip install ghostfold

For development:

git clone https://github.com/brineylab/ghostfold.git
cd ghostfold
pip install -e ".[dev]"

---

Hugging Face Authentication

GhostFold uses ProstT5 from the Hugging Face Hub. You may need to configure a Hugging Face access token:

huggingface-cli login

See the Hugging Face documentation for details.

---

CLI Usage

GhostFold provides a single command-line tool with five subcommands:

Generate pseudoMSAs

ghostfold msa --project-name my_project --fasta-path query.fasta

Options:

• --config PATH — Custom YAML config (overrides bundled defaults)
• --recursive — Recursively search directories for FASTA files
• --coverage FLOAT — Coverage values (repeatable, default: 1.0)
• --num-runs INT — Independent runs per sequence (default: 1)
• --evolve-msa — Enable MSA evolution with substitution matrices
• --mutation-rates JSON — Mutation rates per matrix
• --sample-percentage FLOAT — Fraction of sequences to evolve (default: 1.0)
• --plot-msa-coverage — Generate MSA coverage heatmaps
• --no-coevolution-maps — Skip coevolution map generation

Run structure prediction

ghostfold fold --project-name my_project

Options:

• --subsample — Enable MSA subsampling (multiple depth levels)
• --mask-fraction FLOAT — Mask a fraction of MSA residues (0.0-1.0)
• --num-gpus INT — Override auto-detected GPU count
• --localcolabfold-dir PATH — Path to localcolabfold pixi checkout (default: ./localcolabfold)
• --colabfold-env TEXT — Legacy mamba env name for ColabFold fallback (default: colabfold)

Full pipeline (MSA + folding)

ghostfold run --project-name my_project --fasta-path query.fasta

Combines all options from msa and fold commands.

Mask MSA files

ghostfold mask --input-path input.a3m --output-path masked.a3m --mask-fraction 0.15

Calculate Neff scores

ghostfold neff my_project/

Version

ghostfold --version

---

Python API

GhostFold can also be used as a Python library:

from ghostfold import run_pipeline, mask_a3m_file, calculate_neff, MSA_Mutator
from ghostfold.core.config import load_config

# Load config with optional overrides
config = load_config("my_config.yaml")

# Run MSA generation pipeline
run_pipeline(
    project="my_project",
    fasta_path="query.fasta",
    config=config,
    coverage_list=[1.0],
    evolve_msa=True,
    mutation_rates_str='{"MEGABLAST": 5, "PAM250": 20, "BLOSUM62": 10}',
    sample_percentage=1.0,
    plot_msa=False,
    plot_coevolution=False,
)

---

References

• ProstT5: Protein Language Modeling
• ColabFold: AlphaFold Simplified