A research consortium at KAIST has developed a protein-structure prediction model that, in internal tests, matched the accuracy of Google DeepMind’s AlphaFold3 and ran up to 25 times faster than comparable systems.
The model, called K-Fold, was built by Team KAIST, the group leading the Ministry of Science and ICT’s AI-specialized foundation model development project. The team is led by Professor Kim Woo-youn of the Department of Chemistry, with AI model development handled by professors Hwang Sung-ju and Ahn Sung-su of the Kim Jaechul Graduate School of AI, and protein data construction and validation carried out by professors Oh Byung-ha, Kim Ho-min and Lee Gyu-ri of the Department of Biological Sciences.
What K-Fold Predicts
K-Fold predicts how proteins bind to drug candidates — a step in drug development that researchers use to identify promising compounds. Rather than only modeling a single protein’s three-dimensional shape, the system predicts the structures formed when multiple biomolecules come together, including protein-protein, protein-drug, protein-DNA and protein-RNA complexes.
Accuracy and Benchmark Performance
In a project milestone evaluation in March, K-Fold’s accuracy in predicting molecular complex structures was assessed as comparable to AlphaFold3. In an internal performance evaluation the research team conducted in August, K-Fold exceeded existing global models on several benchmarks. The model performed particularly well on G protein-coupled receptors (GPCRs) and kinases — both major drug targets in cancer and other diseases — and on targeted protein degradation (TPD), an approach that eliminates disease-causing proteins directly.
How the Speed Gain Works
The speed gain comes from removing a step common to conventional protein-structure models. Those systems typically search and compare large volumes of sequence data from similar proteins before calculating structure. K-Fold skips that preprocessing stage, which the team attributes to a new approach that does not depend on it.
Access Through HyperLab
K-Fold is already accessible to outside researchers. The team has integrated it into HyperLab, a multi-agent platform run by HITS, a faculty startup at KAIST. Researchers can run drug-discovery work through a conversational web interface without setting up their own high-performance computing infrastructure or managing AI software directly.
Sovereign Biotech AI
Biotech AI has drawn heavy investment from large technology companies and research institutions in the United States, the United Kingdom and China, in part because of the time and cost savings it offers in drug development. KAIST framed K-Fold as a step toward sovereign biotech AI — a system built on domestic technology rather than foreign models.
KAIST President Bae Choong-sik said national competitiveness in the AI era depends on sovereign AI, defined as the ability to develop and deploy core technologies independently. He added that K-Fold combines homegrown AI technology with biotech to reach world-class performance and links that capability to a service usable in real drug-discovery research.
Professor Kim Woo-youn said the model was built with a new AI architecture intended to move past the limitations of conventional approaches, with the goal of turning it into a scientific AI platform that makes biotech AI accessible to any researcher.
Team KAIST plans to release K-Fold at no cost. HyperLab will offer a beta service to domestic and international researchers before rolling out commercial services in stages by the end of the year.