KAIST Unveils K-Fold Bio-AI Foundation Model for Protein and Drug Design

KAIST unveils K-Fold for biological structure prediction

KAIST announced K-Fold on August 28, 2026, describing it as a next-generation biological foundation model for protein-structure prediction and computational drug design.

The system is designed to predict protein structures as well as interactions involving protein-protein, protein-drug, DNA and RNA complexes. KAIST says a key goal is to support researchers in identifying how candidate compounds may bind to disease-related proteins and in narrowing candidates for experimental validation.

K-Fold was developed by Team KAIST under South Korea's Ministry of Science and ICT AI Specialized Foundation Model Project. The work brings together KAIST chemistry, AI and biological-sciences groups, with HITS integrating the model into its HyperLab research platform.

Up to 25x faster is a team-reported result

KAIST says K-Fold avoids a conventional preprocessing stage that searches and compares large amounts of similar protein information before structure calculation. According to the university, removing that dependency can make structure prediction up to 25 times faster than existing models in the team's testing.

KAIST also says a March project-stage evaluation found K-Fold's molecular-complex structure accuracy approaching AlphaFold 3, while an August in-house evaluation showed stronger results than existing global models in selected categories.

Those are KAIST-reported evaluation claims, not an independent benchmark demonstrating that K-Fold is generally 25x faster or more accurate than AlphaFold 3 across all tasks. The exact datasets, hardware, task mix and evaluation protocol matter when interpreting headline comparisons.

Drug-binding and complex prediction are central targets

KAIST highlights performance on major drug-discovery targets including GPCRs and kinases, as well as targeted protein degradation workflows.

The model is not presented as a clinical diagnostic system or as a substitute for laboratory testing. Its role is computational: predict structures, rank or design candidates and help researchers decide what deserves further experimental validation.

That distinction is important in AI-for-science coverage. Faster in-silico screening can reduce search costs, but candidate binding predictions still need physical experiments and, for therapeutics, the full preclinical and clinical development process.

HyperLab turns K-Fold into a multi-agent research workflow

K-Fold has been integrated into HyperLab, a multi-agent platform from HITS, a KAIST faculty startup. KAIST says HyperLab connects roughly 120 computational tools, 160 specialized functions and more than 100 databases through a large knowledge graph.

The platform is designed so a researcher can state an objective—such as finding a peptide candidate for a target protein—and have the system coordinate structure prediction, candidate design, database retrieval and computational evaluation.

This makes the announcement broader than a single structure model. It is also an example of AI-for-science moving toward agentic orchestration around specialized scientific tools rather than relying on one general-purpose language model to perform every step.

Release status: unveiled now, free distribution planned

K-Fold has been unveiled and integrated into HyperLab, but KAIST says the Team KAIST consortium plans to distribute K-Fold free of charge. HyperLab is expected to provide beta access to researchers in Korea and abroad and then expand commercial services gradually by the end of 2026.

That means the model should be described as an announced research system with a planned free release, not as a universally downloadable open model today.

KAIST says the work was supported by South Korea's Ministry of Science and ICT. The team had also presented the underlying achievement at a biomolecular-science meeting in June, while the university's detailed public announcement and rollout information arrived on August 28.

Why K-Fold matters

K-Fold is notable because it combines a specialized biological foundation model with an agentic research environment that can connect structure prediction to downstream drug-design tools and scientific databases.

The strongest evidence will come from transparent evaluation data, external replication and real research use after broader access expands. For now, the verified development is that KAIST has unveiled the model, reported promising speed and complex-prediction results, integrated it into HyperLab and outlined a free-release/beta-access path.