< (From left) KAIST Professors Byung-Ha Oh, Gyuri Lee, Sung Ju Hwang, Woo Youn Kim, Sungsoo Ahn, and Ho Min Kim >
"Design a drug candidate that binds effectively to this protein."
In response to such a request, AI predicts the protein's three-dimensional structure, analyzes which compounds are most likely to bind to it, and designs promising drug candidates. KAIST researchers have developed K-Fold, the world's fastest Bio-AI model for protein structure prediction, which also supports drug candidate design.
KAIST (President Choongsik Bae) announced on August 28 that it had formed "Team KAIST" after being selected as the lead institution for the Ministry of Science and ICT's "AI Specialized Foundation Model Project" and unveiled K-Fold, a next-generation Bio-AI model developed by Team KAIST.
Team KAIST is led by Professor Woo Youn Kim from the Department of Chemistry. His research group, together with Professors Sung Ju Hwang and Sungsoo Ahn's groups at the Kim Jaechul Graduate School of AI, developed the AI model. Professors Byung-Ha Oh, Ho Min Kim, and Gyuri Lee from the Department of Biological Sciences oversaw protein data construction and validation. HITS, a KAIST faculty startup, integrated K-Fold into HyperLab, its web-based AI research platform, enabling researchers to use the model in real-world research workflows. In addition, the Korea Pharmaceutical and Bio-Pharma Manufacturers Association (KPBMA) and the Korea Biotechnology Industry Organization (KoreaBIO) will lead efforts to raise awareness of K-Fold's achievements and promote its use across the industry.
K-Fold's defining capability is its ability to predict the binding between proteins and drug candidates—a critical step in drug discovery.
Drug development begins with determining the structure of a disease-related protein and identifying, among numerous compounds, those most likely to bind to the protein and produce the desired effect. K-Fold not only predicts a protein's three-dimensional structure, but also calculates where and how a drug candidate is likely to bind, helping researchers identify promising candidates more quickly.
K-Fold goes beyond predicting the structure of a single protein. It can also predict the structures formed when different biomolecules interact, including protein–protein and protein–drug candidate complexes, as well as complexes involving DNA and RNA.
In the project's stage evaluation conducted in March, its accuracy in predicting molecular complex structures was assessed as approaching that of AlphaFold3, developed by Google DeepMind. In an in-house performance evaluation conducted by the research team in August, K-Fold also outperformed existing global models in selected evaluation categories.
K-Fold demonstrated particularly strong performance in predicting how drug candidates bind to and act on key therapeutic targets, including G protein-coupled receptors (GPCRs) and kinases, which are major drug targets for cancer and other diseases. It also performed strongly in targeted protein degradation (TPD), an emerging drug discovery approach designed to directly eliminate disease-causing proteins.
K-Fold also significantly increased the speed of structure prediction. Conventional protein structure prediction models often require a complex preprocessing step that searches for and compares large amounts of data on similar proteins before calculating a structure. K-Fold applies a new approach that does not depend on this process, eliminating the need for preprocessing calculations and increasing structure prediction speeds by up to 25 times compared with existing models.
This means that researchers can evaluate more drug candidates within the same amount of time. By reducing the time and computing resources required for structure prediction, K-Fold can help rapidly identify the most promising compounds from a vast pool of candidates and narrow the selection for experimental validation.
"National competitiveness in the AI era depends on sovereign AI capabilities, which is the crucial ability to develop and deploy core technologies independently," said KAIST President Choongsik Bae. "K-Fold is significant because it combines homegrown AI technology with biotechnology to challenge the world's leading technologies and translates that capability into a service applicable to real-world drug discovery. KAIST will continue to strengthen Korea's technological sovereignty in AI and its future competitiveness by advancing the convergence of foundational AI technologies with science and technology."
The research team went beyond developing K-Fold as a standalone model, turning it into an AI research service that researchers can use through a conversational web-based interface.
K-Fold has been integrated into HyperLab, a multi-agent platform developed by HITS, a KAIST faculty startup. This allows researchers to use the model without having to build their own high-performance computing infrastructure or operate complex AI software.
For example, if a researcher asks the AI to "design an antibody that binds strongly to this protein," it provides step-by-step support for predicting the protein's structure, designing candidates with a high likelihood of binding, and computationally evaluating the results. Researchers can also ask it to "find a suitable peptide candidate for this cancer target protein." In practical terms, instead of moving between multiple software tools to calculate structures and analyze results, researchers can simply state their research objective and have the AI carry out the necessary analyses and design tasks in sequence.
To support these capabilities, HyperLab incorporates approximately 120 computational tools and 160 specialized functions for structure prediction, drug design, and the analysis of life science data, including genomic and proteomic data. It also connects more than 100 specialized databases with a large-scale knowledge graph, enabling the platform to retrieve relevant scientific information and apply it to its analyses.
HyperLab aims to serve as an AI Co-Scientist that assists researchers throughout the research process by supporting the full workflow, from understanding a research question and selecting the appropriate tools to predicting structures, analyzing results, and iteratively improving designs.
Bio AI is emerging as a critical technology capable of reducing the time and cost required for drug discovery, driving intense competition among global technology companies and major research institutions in the United States, the United Kingdom, and China. The development of K-Fold is significant because it lays the foundation for sovereign bio AI by securing core bio AI technology domestically, rather than relying solely on overseas models, and making it available for real-world research.
The achievement was first presented at the 2026 Annual Meeting of the Korean Federation of Biomolecular Science, held on June 23, where Professor Woo Youn Kim from the KAIST Department of Chemistry delivered a keynote lecture titled "Generative Drug Design Powered by Agentic AI."
"K-Fold was developed not to follow existing models, but to overcome the limitations of conventional approaches through a new AI architecture," explained Professor Kim. "We will develop it into an AI-for-Science platform that makes world-class bio AI technology accessible to researchers everywhere."
The Team KAIST consortium plans to release K-Fold free of charge. HyperLab will provide beta access to researchers in Korea and abroad, and gradually expand its commercial services by the end of this year.
Meanwhile, industry training on K-Fold is gaining momentum. An online session hosted by KoreaBIO on August 27 attracted 85 participants, while 118 have registered for KPBMA's hybrid session on September 1. Designed primarily for researchers and practitioners at pharmaceutical and biotech companies, the training covers how to use K-Fold and presents case studies of its application to drug design. The initiative is intended to accelerate the adoption of sovereign bio-AI across the industry.
