Hyper Binding Co-folding

A docking program treats the receptor as furniture: one deposited structure, a box around a pocket someone already knows about, and a search for the best-scoring ligand conformation inside it. Each of those assumptions fails routinely in early discovery. Proteins are not rigid — a few angstroms of rearrangement on binding moves both the pose and the score. The pocket has to be identified before the calculation starts. And when a target has several deposited structures, the choice of which to dock into is itself a large source of error: HITS reports that for one derivative series, swapping the input structure moves the correlation between predicted and measured activity from 0.86 down to 0.28. For a first-in-class target with no holo structure, there is nothing to dock into at all.

Hyper Binding Co-folding drops the receptor structure from the input. Given a protein sequence and a ligand drawn as a 2D structure, it predicts the folded protein and the bound ligand conformation simultaneously, so induced fit is an output rather than a property of whichever crystal structure was picked; the technical report states that the method "does not utilize any 3D structural information of the target protein during inference." HITS Inc., a Seoul drug discovery company spun out of KAIST, shipped it on 10 April 2025 as the headline feature of HyperLab 2.0.

It sits in the co-folding lineage opened by AlphaFold 3, which HITS says publicly it extended; the technical report cites Chai-1 as the co-folding approach it implements while benchmarking Chai as a separate baseline. It is one of two modes of HyperLab's Hyper Binding function — the other, a distinct model, holds the receptor rigid and scores poses inside a user-defined site with a physics-informed network.

#Key Features

• Sequence-in, complex-out: A protein name or raw sequence plus a ligand SMILES string is the whole input; HyperLab resolves the name against UniProt and pulls in cofactors, ions and post-translational modifications.
• Induced-fit prediction: Protein conformation and ligand pose are predicted jointly. On E3 ligase–glue–target ternary complexes, removing the molecular glue changes the predicted binding mode, not merely the score.
• No pocket required: Valid poses are produced without a predefined binding site. Supplying site information when it is known raises accuracy rather than being a precondition.
• Affinity alongside geometry: The same run returns a predicted binding score, so a derivative series can be ranked for synthesis priority.
• Covalent co-folding: A module added in February 2026 co-folds warhead-bearing ligands together with their target residue, with warhead and product form selected explicitly.

#Technical Details

On the PoseBusters v2 PB-valid benchmark, HITS reports 68% pose prediction accuracy from sequence and ligand alone, rising to 77% with binding-site information. The same figure places DiffDock at 13%, AutoDock Vina at 58%, Chai Discovery at 66%, AlphaFold 3 at 73% without a site and 84% with one, and Boltz-2 at 78% with a site. Inference is quoted at roughly three minutes per complex on cloud hardware, against the fifteen minutes the report cites for AlphaFold 3 on an RTX 3060. For affinity, the Hyper Binding function reaches Pearson correlations of 0.70 and 0.53 on two free-energy perturbation sets, ahead of Glide SP, Vina, Genscore and Luminet on both. On a covalent set curated from PDBbind, covalent co-folding reports 88.7% pose accuracy against 61.3% for covalent docking, 48.4% for COV SMINA and 46.8% for GNINA. Neither the parameter count nor the training corpus has been disclosed, and every number here is self-reported in a technical report that has not been peer reviewed or independently reproduced.

#Applications

The intended use is hit triage and lead optimization where the structural starting point is weak or absent — novel mechanisms with no holo structure, targets whose deposited structures disagree, and interactions that exist only as a ternary assembly. On a JAK2 derivative set, HITS reports that co-folded structures lifted the activity correlation from about 0.38 under conventional docking to 0.71, the difference between a ranking that cannot guide synthesis and one that can. Delivered inside HyperLab rather than as a local package, it runs without a GPU cluster or a computational chemistry background, and feeds the platform's screening, SAR and design modules.

#Impact

Hyper Binding Co-folding is a clear example of AlphaFold 3-style co-folding productized for bench chemists rather than distributed as a checkpoint — which is also its main limitation. No code, weights or training data have been released, the model is reachable only through a commercial platform, and every benchmark figure is the vendor's own, so none of it can be independently verified. The developers are candid about failure modes: degraded accuracy on targets with few structural neighbors, occasional chirality errors, and difficulty with large conformational transitions. HITS has since integrated K-Fold, the alignment-free complex predictor from Team KAIST, into the same platform, so HyperLab now offers two structure engines side by side.