K-Fold

K-Fold is currently in the preview stage and under active development. The first official release is expected around September 20. 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). This repository provides pretrained models, inference and training code, and data preprocessing workflows.

Model parameters

K-Fold uses pretrained AtlasLM for protein sequence representations and TriProRep for protein structure representations. The parameters for these models and K-Fold are downloaded automatically on first use from Hugging Face.

Installation

K-Fold requires Python 3.11 or later.

Install from PyPI:

pip install kfold

Or install from source:

git clone https://github.com/SeonghwanSeo/kfold.git
cd kfold
pip install -e .

K-Fold uses custom Triton kernels by default when Triton is installed and CUDA is available.

Inference

Run predictions from a YAML or JSON query file, or a directory of query files for bulk runs:

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

By default, K-Fold prepares apo structures with AtlasFold, then runs K-Fold predictions. You can also provide apo structures from experiments or other prediction tools (e.g., AlphaFold2).

Use --kernel {triton,cuequiv,torch} to select the backend for K-Fold complex prediction. Automatic selection prefers Triton, then cuEquivariance, then PyTorch, depending on availability.

For multi-seed inference, use --share-apo-seeds 1 2 3 to generate one apo ensemble per target and reuse it across all inference seeds specified by --seeds. This is particularly useful for relatively rigid apo structures or runs with many inference seeds; see automatic apo generation.

Use --stage apo to prepare apo structures only, or --stage complex to predict complexes from prepared apos:

kfold --stage apo --input examples/ --out-dir predictions/ --seeds 42
kfold --stage complex --input examples/ --out-dir predictions/ --seeds 42

See kfold --help, the inference guide, or the Python API guide for details.

Training

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

Acknowledgements

K-Fold was developed at KAIST as part of the K-Fold initiative supported by the Ministry of Science and ICT (MSIT), Republic of Korea.

Members of Team KAIST are listed below (alphabetical order):

• Project management: Hyeongwoo Kim^3,†
• K-Fold architecture: Seokhyun Moon^3,†, Jun Hyeong Kim³, Shinwoo Kim³, Minha Park³, Jisu Seo³, Mingyeong Shin³, Wonho Zhung³
• Protein structure encoder: Hyosoon Jang^1,†, Taewon Kim^1,†, Hyunjin Seo¹
• RNA sequence encoder: Dongki Kim^1,†, Jun Hyeong Kim^1,†, Jinheon Baek¹, Jaehyeong Jo¹
• Training data preparation: Yeongnam Bae^2,†, Woosung Jeon^2,†, Joongwon Lee^3,†, Junyup Lee^2,†, Yunsu Shin^2,†, Eugene Choi², Jeong Hun Choi², Hyeongyu Han², Calvin Samuel²
• Kernel optimization: Youngchan Kim⁴
• Engineering lead: Seonghwan Seo^3,†
• Supervision: Sungsoo Ahn¹, Dongsu Han^4,1, Sung Ju Hwang¹, Ho Min Kim², Woo Youn Kim³, Gyu Rie Lee², Byung-Ha Oh²

^† Core contributor; ¹ KAIST AI; ² KAIST Biological Sciences; ³ KAIST Chemistry; ⁴ KAIST Electrical Engineering.

We thank our collaborators at HITS for their contributions to K-Fold.

License

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

K-Fold source code and model weights are licensed under the Apache License 2.0.