KAIST发布开源生物AI模型K-Fold，称结构预测速度最高快于既有模型25倍

The Korea Advanced Institute of Science and Technology (KAIST) has developed a homegrown bio artificial intelligence model to compete with Google DeepMind's AlphaFold3. The AI predicts structural changes that occur when proteins bind with drugs, helping researchers search for new drug candidates. The model is an outcome of the government's "AI-specialized foundation model" program, which aims to secure independent AI technology in specialized fields such as biotechnology and medical science, beyond general-purpose large language models. It is expected to reduce reliance on foreign technology and broaden the base for AI use in domestic pharmaceutical and biotech research.
KAIST said on the 28th that it had developed K-Fold, a next-generation bio AI model, through the Ministry of Science and ICT's AI-Specialized Foundation Model Project. The research was overseen by Kim Woo-youn, a professor in KAIST's Department of Chemistry. Teams led by professors Hwang Sung-ju and Ahn Sung-soo of the Kim Jaechul Graduate School of AI developed the AI model, while teams led by professors Oh Byung-ha, Kim Ho-min and Lee Gyu-ri of the Department of Biological Sciences handled protein data construction and validation. HITS, a faculty startup spun out of KAIST, developed the service that allows researchers to use K-Fold in practice.
A Korean AlphaFold, 25 Times Faster
K-Fold predicts not only the three-dimensional structure of proteins but also what structures form when proteins bind with drug candidates. In drug development, this allows researchers to use AI to screen which substances are more likely to bind well to a particular protein, before testing numerous candidates one by one in the lab. The model can also predict complex structures formed when various biomolecules bind, including protein-protein pairs, antibodies and antigens, and DNA and RNA.
In a government project stage evaluation in March, K-Fold's accuracy in predicting molecular complex structures was assessed as approaching that of AlphaFold3. In an internal performance evaluation the research team conducted this month, the model outperformed global models on some assessment items. It showed particularly strong performance on drug targets where structural changes are difficult to predict, such as targeted protein degradation (TPD) drugs, G protein-coupled receptors (GPCR) and kinases.
Prediction speed also improved. Existing structure-prediction AI must go through multiple sequence alignment (MSA), a process of searching and comparing vast sequence data from similar proteins. K-Fold reduced its dependence on that process through large-scale pretraining and adopted its own approach, using generative AI to learn the structural changes that occur when a protein binds with another substance. By eliminating the precomputation step required for structure prediction, the model raised prediction speed by up to 25 times compared with existing models, the researchers said.
Open Platform Lets Any Researcher Work With an AI Scientist
The researchers also connected K-Fold to a service that can be used in actual drug research. K-Fold has been installed on HyperLab, HITS' AI platform, so that when a researcher types a natural-language request such as "design an antibody that binds strongly to this protein," the AI predicts the protein structure, designs candidates with a high likelihood of binding, and then evaluates the results, carrying out the process step by step.
HyperLab is linked to more than 120 computational tools, more than 160 specialized functions, and more than 100 specialized databases and knowledge graphs that can be used for structure prediction, drug design and life-science data analysis. The plan is to develop it into an "AI co-scientist" that understands a researcher's question, selects the necessary tools, predicts structures and then continues on to result analysis and design improvement.
K-Fold consists of a 7B-class main model and a 2B-class lightweight model, and will be released as open source under the Apache 2.0 license. The researchers plan to make K-Fold available not only to domestic universities and research institutes but also to pharmaceutical and biotech companies for their own research and service development. The Korea Pharmaceutical and Bio-Pharma Manufacturers Association and the Korea Biotechnology Industry Organization will also help spread the model across the industry, including training company staff on how to use K-Fold and how it applies to drug design.
Beyond General-Purpose AI, Technological Sovereignty in Biotech
The development of K-Fold is also significant because the government is widening the scope of its push for AI technological sovereignty from general-purpose AI to specialized fields such as biotechnology and medical science. The Ministry of Science and ICT is running the AI-Specialized Foundation Model program separately from its general-purpose independent AI foundation model effort. The aim is to secure independent AI models in specialized areas that general-purpose AI alone cannot address, and to spread them across domestic industry and research sites.
Biotechnology and medical science are the first strategic fields. A total of 18 consortiums applied in a public call held last year, and two consortiums, led by KAIST and medical AI company Lunit, were selected. KAIST is developing bio AI for protein complex structure prediction and drug design, while Lunit is developing full-cycle medical science AI that spans molecules and drugs through clinical trials and actual clinical settings. Since November last year, the government has provided each consortium with 256 Nvidia B200 GPUs.
"Securing independent AI in the biotechnology field means more than simply adding one more homegrown model," the researchers said in explaining the significance of the achievement. Foreign models that currently lead the field of protein structure prediction, including AlphaFold3, may carry restrictions on usage terms or commercial application. As AI becomes established as a core research tool in drug development, domestic research institutes and pharmaceutical and biotech companies face a growing likelihood of being affected by the policies governing foreign models and services. The KAIST researchers said the ultimate goal of developing K-Fold is "to reduce dependence on foreign technology and build a sovereign bio AI based on domestic independent technology."