K-Fold

K-Fold is currently in the preview stage. Preprint will be released soon.

K-Fold predicts biomolecular complex structures and binding-induced conformational changes. Through an apo-to-holo diffusion bridge, K-Fold aims to capture conformational changes in systems such as G protein-coupled receptors (GPCRs).

K-Fold supports proteins, DNA, RNA, small molecules, and chemical modifications without multiple sequence alignments (MSAs). Protein component structures are generated with AtlasFold by default or supplied from other predictors or experiments. This repository provides pretrained models, inference and training code, and data preprocessing workflows.

Installation

K-Fold requires Python 3.11 or later.

Install from PyPI:

pip install 'kfold[cuequiv]'

Or install from source:

git clone https://github.com/SeonghwanSeo/kfold.git
cd kfold
pip install '.[cuequiv]'

The cuequiv extra installs cuEquivariance kernels for faster inference on NVIDIA GPUs. Model weights and the chemical component dictionary (CCD) are downloaded automatically on first use.

Inference

Run prediction on a YAML or JSON query file from the repository root:

kfold predict --input examples/8and.yaml --out-dir predictions/ --seed 42

--input accepts a YAML or JSON file, or a directory of query files. The default run generates missing protein apo structures and predicts five samples, saving structures and confidence scores under --out-dir.

Run kfold predict --help for all options. See the inference guide for input formats, sampling, multiple GPUs, and outputs, or the Python API guide for use in Python.

Training

See the training guide for data preparation, training commands, and configuration.

License

Copyright © 2026 Korea Advanced Institute of Science and Technology (KAIST).

K-Fold is licensed under the Apache License 2.0.