KAIST unveils homegrown AI model for predicting protein structures, drug binding

Researchers at the Korea Advanced Institute of Science and Technology (KAIST) said Friday they have developed a biomolecular artificial intelligence (AI) model, called K-Fold, that predicts protein structures and how drug candidates bind to them, with the aim of speeding up new drug development using homegrown technology.

The model was built by a KAIST-led group called Team KAIST as part of a project run by Korea's Ministry of Science and ICT to develop AI foundation models specialized by field. KAIST President Bae Chung-sik said the project reflects the importance of "sovereign AI" — a country's ability to develop and control its own core technology — for national competitiveness in the AI era.

K-Fold is designed to predict not only the 3D shape of a single protein, but also how it interacts with other proteins, drug candidates, or genetic material such as DNA and RNA, and where and how a candidate compound binds.

That combination of structure prediction and binding prediction is central to early-stage drug discovery, in which researchers must identify a disease-related protein's shape before screening for substances likely to bind to it effectively.

In an internal evaluation in March, the team said K-Fold's accuracy in predicting complex molecular structures approached that of Google DeepMind's AlphaFold3, a leading international model.

A follow-up benchmark test in August found K-Fold outperformed existing global models on some measures, including binding predictions for G-protein-coupled receptors and kinases — both common drug targets in cancer and other diseases — as well as for targeted protein degradation, a newer drug approach that eliminates disease-causing proteins directly rather than merely blocking them.

The researchers said K-Fold also runs up to 25 times faster than comparable existing models because it skips a computationally heavy step, used by earlier AI models, of searching for and comparing large numbers of similar protein sequences before calculating structure.

KAIST-affiliated startup HITS has integrated K-Fold into HyperLab, a web-based AI research platform, allowing researchers to design drug candidates through natural-language requests rather than operating specialized software directly.

The Korea Pharmaceutical and Bio-Pharma Manufacturers Association and the Korea Biotechnology Industry Organization are helping promote the model's use in industry; the team plans to release K-Fold free of charge, with HyperLab moving from a beta test to wider commercial service later this year.

This article was published with the assistance of generative AI and edited by The Korea Times.