PreFold-dG

Official repository for PreFold-dG.

Requirements

This code is tested using the following environment:

python==3.12
pytorch==2.6.0
numpy==2.4.2
pandas==2.2.3
scipy==1.13.1
scikit-learn==1.6.1
biopython==1.84
tqdm==4.67.1

Preprocessed Data

Raw data can be preprocessed with the provided notebook process_data.ipynb.

We share preprocessed files for the datasets used in the paper. Due to file size, they are hosted externally:

File	Link
processed_data/skempi_mut.pkl	Download
processed_data/skempi_wt.pkl	Download
processed_data/her2.pkl	Download

Data Sources

Split and group files under data/ are derived from the following sources:

File	Source
data/split/rdenet.csv	RDE-PPI
data/split/ppiformer.csv	PPIRef
data/split/gearbind.csv	GearBind
data/group/prot2cplx.csv	PPIformer

Usage

python main.py --config config/skempi.yaml

Any config option can be overridden via CLI arguments:

python main.py --config config/skempi.yaml --epochs 100

See Configuration for the full list of options.

Configuration

YAML config files define model architecture, training hyperparameters, and data/split settings.

Parameter	Description
train_data	Training dataset name
test_data	Optional external test dataset
split	Path to fold assignment CSV (PDB,fold). If omitted, a random 3-fold split is generated
train_folds	List of folds to use as hold-out pool. If omitted, all folds are used
test_fold	Fixed test fold index
group_map	Optional CSV for grouping (PDB,group)
group_min_size	Minimum group size for group-level metrics

See config/ for examples: skempi.yaml (cross-validation), ppiformer.yaml (fixed test fold), her2.yaml (external test set).

Citation

TBA

License

This project is licensed under the BSD-3-Clause-LG AI Research License.