YuelDesign

In unbound cyclin-dependent kinase 2, a salt bridge holds lysine-33 and aspartate-145 together. In the inhibitor-bound structures that contact is broken, freeing D145 to accept a hydrogen bond or stack against the ligand's aromatic ring. A generative model handed a rigid crystal snapshot inherits whichever conformation it was given: with the salt bridge intact, it is a steric and electrostatic wall, and the sampler designs molecules that avoid D145 rather than the ones that engage it.

YuelDesign removes the wall by generating the pocket and the ligand at the same time. It runs a diffusion process over a full-atom representation of the complex — protein side chains and ligand atoms concatenated into a single coordinate matrix, distinguished only by binary masks — so the binding site rearranges as the molecule takes shape. Jian Wang, Dong Yan Zhang and Nikolay Dokholyan at the University of Virginia School of Medicine, with Shreshty Budakoti at Pennsylvania State University, posted it to bioRxiv in May 2025 and published it in Science Advances in April 2026.

It is the generative member of the group's Yuel line, alongside Yuel 2 for binding affinity and YuelPocket for locating the site in the first place. Its backbone, E3former, adapts the Evoformer stack of AlphaFold 2: triangle attention and triangle multiplication over pair features, but with multiple sequence alignments and positional encodings dropped — pockets and ligands have no canonical residue ordering — and an equivariant coordinate head that emits 3D displacements instead of a distogram.

#Key Features

• Joint pocket and ligand generation: Side-chain coordinates denoise alongside ligand coordinates, so induced fit emerges from sampling rather than being ruled out by a frozen receptor. Generated pockets sit at a median 1.8 Å RMSD from their native counterparts.
• Dual diffusion over geometry and chemistry: An elucidated diffusion model handles continuous coordinates while a discrete denoising diffusion probabilistic model handles atom types, both conditioned on shared features and time steps so chemical identity and geometry stay consistent.
• Equivariant Evoformer backbone: E3former preserves rotational and translational equivariance across full-atom complexes of up to roughly 400 atoms, against the 15 to 50 ligand-only atoms earlier generators handled.
• Leakage-controlled data split: Complexes above 30% sequence identity, pockets within 2 Å RMSD, and ligands above 0.85 Tanimoto similarity are forced into the same split, so test pockets are not near-copies of training ones.
• Chemically well-behaved output: Connectivity stays high and the rate of rings larger than six atoms stays lowest across every molecule size tested, with validity near 1.0.

#Technical Details

Training uses Binding MOAD, a curated set of roughly 41,400 protein-ligand crystal structures drawn from the PDB, split 8:2 under the similarity constraints above. A pocket is every residue with an atom within 6 Å of any ligand atom; each atom is a token carrying its one-hot type and a backbone/side-chain/ligand indicator, with interatomic distance and a same-residue flag as pair features. Sampling starts from Gaussian noise in coordinate space and a random categorical draw in atom-type space, then denoises over 100 steps; coordinates largely settle by step 75 while atom types stabilize in the final 20. Against DiffSBDD and PMDM, two rigid-pocket diffusion generators, YuelDesign holds higher QED and lower large-ring rates across all size bins at comparable synthetic accessibility, achieves better AutoDock Vina and MedusaScore energies, and shows lower RMSD when its molecules are redocked. On CDK2 it produces a far larger share of ligands within 3 Å of D145 and a broad distribution of K33-D145 distances that includes the native ligand-bound geometry, a distance the rigid-pocket baselines cannot move at all.

#Applications

The intended use is hit generation for structure-based drug discovery, and the flexibility argument matters most for targets whose sites reorganize on binding — kinases, reductases, and other proteins where a single crystal snapshot misrepresents the druggable shape. Inference takes a pocket PDB file and a target heavy-atom count and writes complexes holding both the generated ligand and its accompanying pocket conformation, giving medicinal chemists a starting scaffold together with the receptor geometry that scaffold implies.

#Impact

YuelDesign makes a specific and testable claim about a common shortcut: that freezing the receptor is not a harmless simplification but the thing keeping generators away from a large class of real binders, and the CDK2 salt-bridge analysis demonstrates it cleanly. The authors are candid about what remains unsolved. Generated molecules retain only modest fingerprint similarity to native ligands; three- and four-membered rings are overproduced because nothing penalizes ring strain during sampling from noise; and quality degrades for larger molecules as the dimensionality of the joint diffusion grows, with latent diffusion offered as the likely fix. All evaluation is computational — docking scores, chemical property distributions and pocket RMSD — with no experimental synthesis or assay.