variantfold 2.0.0

Classify variants of uncertain significance using AlphaFold-predicted protein structures and graph neural networks.

VariantFold

Classify Variants of Uncertain Significance (VUS) using AlphaFold-predicted protein structures and Graph Neural Networks.

VariantFold leverages protein structure predictions from ColabFold/AlphaFold and a Graph Convolutional Network (GCN) to classify VUS based on the standardised ACMG-AMP variant classification system.

Workflow

ClinVar data → Parse variants → Mutate sequences → ColabFold 3-D prediction
    → PDB-to-graph conversion → Train GCN (benign vs pathogenic) → Classify VUS

1. Parse — Extract missense variants from ClinVar downloads (benign, pathogenic, VUS).
2. Mutate — Apply each variant to the reference protein sequence.
3. Predict — Run ColabFold to generate 3-D structure models for every variant.
4. Convert — Transform PDB files into PyTorch Geometric residue-level graphs with rich node features (one-hot amino acid, 3-D coordinates, pLDDT).
5. Train — Train a multi-layer GCN on the benign vs pathogenic graph dataset.
6. Classify — Run the trained model on VUS structures to predict likely benign / likely pathogenic with probabilities.

Installation

# Core package (graph conversion + GCN training/inference)
pip install .

# With ColabFold for structure prediction (GPU recommended)
pip install ".[structure]"

# With visualisation tools
pip install ".[viz]"

# Everything
pip install ".[all]"

Quick start — Python API

from variantfold import VariantFoldConfig, VariantFoldPipeline

cfg = VariantFoldConfig(
    gene_symbol="VHL",
    entrez_email="[e-mail]",
)

pipe = VariantFoldPipeline(cfg)
pipe.step1_parse_variants()       # Parse ClinVar files + fetch sequence
# pipe.step2_predict_structures() # Run ColabFold (long — needs GPU)
pipe.step3_collect_models()       # Gather best PDB models
metrics = pipe.step4_train()      # Train GCN
print(f"Test accuracy: {metrics['accuracy']:.2%}")

vus_df = pipe.step5_classify_vus()
print(vus_df)

Quick start — CLI

# Run steps 1, 3, 4, 5 (assumes PDB libraries are already populated)
variantfold run --gene VHL --email [e-mail] --steps 1,3,4,5

# Standalone inference on new PDB files
variantfold predict --model variantfold_VHL/variantfold_model.pt \
                    --pdb-dir ./new_vus_pdbs/

Input data

Place these files in the working directory (./variantfold_<gene>/):

File	Description
clinvar_result_bng.txt	ClinVar download filtered to benign variants
clinvar_result_ptg.txt	ClinVar download filtered to pathogenic variants
clinvar_result_vus.txt	ClinVar download filtered to VUS (optional)

Download from ClinVar using the tab-delimited download with default settings.

Configuration

All parameters are set via VariantFoldConfig:

cfg = VariantFoldConfig(
    gene_symbol="TP53",
    entrez_email="[e-mail]",
    distance_threshold=6.5,   # Å, residue contact cutoff
    gcn_hidden_dim=64,        # GCN layer width
    gcn_num_layers=3,         # depth
    epochs=200,
    learning_rate=0.01,
    train_fraction=0.8,
    use_residue_features=True, # 24-dim features (set False for legacy 1-dim)
)

Development

pip install -e ".[dev]"
pytest

Licence

MIT