K-Fold

Every predictor in the AlphaFold lineage begins by looking backwards: before a single coordinate is computed, the pipeline searches large sequence databases for homologs and assembles a multiple sequence alignment. That search frequently dominates wall-clock time for one prediction, and it is why folding thousands of protein-ligand pairs is an infrastructure project rather than an afternoon. K-Fold is built to remove that step. Team KAIST reports that large-scale pretraining reduced the model's dependence on alignments enough to run inference from sequence alone, and attributes the speedup it announces — up to 25 times faster than comparable models — to skipping that precomputation rather than to a cheaper folding trunk.

K-Fold predicts complexes, not just single chains: protein-protein pairs, protein-ligand binding poses for drug candidates, and assemblies involving DNA and RNA. It was developed by Team KAIST, the consortium KAIST leads under the Korean Ministry of Science and ICT's "AI Specialized Foundation Model Project," and unveiled on 28 August 2026. Woo Youn Kim of the Department of Chemistry leads the project; Sung Ju Hwang and Sungsoo Ahn of the Kim Jaechul Graduate School of AI built the model, while Byung-Ha Oh, Ho Min Kim and Gyuri Lee of the Department of Biological Sciences handled protein data construction and validation. HITS Inc., a KAIST faculty spin-off, serves the model inside its HyperLab platform.

The model is positioned directly against AlphaFold 3 and Boltz-2, and the framing is as much about sovereignty as accuracy: the consortium's stated goal is a domestically controlled structure predictor whose terms of use are not set abroad. As of the announcement no preprint, code or weights have been published, and every performance figure below comes from the developers or from the funding program's own milestone review.

#Key Features

• Alignment-free inference: Structure prediction begins from raw sequence, with no homology search or MSA construction stage in front of it.
• Multi-molecule complexes: One model covers protein-protein, protein-drug candidate, and protein-DNA/RNA assemblies rather than single-chain folding alone.
• Binding-mode emphasis: Development targeted where a candidate compound attaches and in what pose, the question that decides whether a virtual screen is worth acting on.
• Two release scales: A 7B-class main model and a 2B-class lightweight variant are announced under one name, for full-accuracy and throughput-oriented use respectively.
• Delivered as a service: HITS integrated K-Fold into HyperLab, so the model is used through a conversational web interface instead of a local GPU cluster.

#Technical Details

Team KAIST describes an architecture that learns the physical and chemical principles of intramolecular and intermolecular interaction, generatively modeling the conformational change a protein undergoes on binding, rather than relying solely on statistical patterns in deposited structures. The consortium built a 2B-parameter model in the project's first phase and scaled it to 7B in the second, training on the 256 NVIDIA B200 GPUs the ministry allocated to it from November 2025. In the program's March 2026 stage evaluation, complex structure accuracy was assessed as approaching AlphaFold 3. An in-house benchmark run by the team in August 2026 reported that K-Fold exceeded existing models on selected evaluation items, with the strongest results on G protein-coupled receptors, kinases, and targeted protein degradation — target classes where the relevant conformational change is hard to predict. Neither evaluation has been peer-reviewed or independently reproduced, and no benchmark protocol or per-target scores have been published.

#Applications

The intended use is early-stage drug discovery, where a team must fix a target protein's structure and then rank many compounds by how well they bind it. Removing the alignment search changes the arithmetic of that screen: more candidates can be co-folded in the same compute budget, narrowing the set that goes to assay. HITS delivers the model through HyperLab, its multi-agent platform of roughly 120 computational tools, so a researcher can state an objective — design a binder for this target, find a peptide candidate for this oncoprotein — and have structure prediction, design and scoring run in sequence. Atolab, a KAIST alumni startup, is to provide on-premise HyperLab installations for institutions with data-residency constraints, and KPBMA and KoreaBIO are running usage training for Korean pharmaceutical and biotech staff.

#Impact

K-Fold is the biotechnology half of Korea's push to extend "sovereign AI" from general-purpose language models into specialized scientific domains; a parallel consortium led by Lunit covers medical AI. The consortium's practical argument is about dependence: leading structure predictors carry usage or commercial restrictions set by their owners, and a domestic model removes that exposure. KAIST states that the 7B and 2B models will be released free of charge under Apache 2.0 and that HyperLab will move from beta to staged commercial service. Until that release lands the model cannot be evaluated independently — there is no paper, no repository, no checkpoint and no public benchmark submission, so the accuracy and speed claims rest on the developers' own reporting.