Loading…
Protein-ligand binding affinity prediction for Kd, Ki and IC50 from a pocket structure and a SMILES string, with no docked complex required.
No providers recorded yet. Browse all providers
Binding affinity predictors sit on a strange dataset. The PDBbind general set holds roughly 19,000 protein-ligand complexes, and most scoring models take the crystallographic complex as input — a structure harder to obtain than the affinity being predicted. Deep scoring functions have crept from Pearson 0.82 to 0.84 on the standard benchmark while sitting on those same labels.
Yuel 2 manufactures more. It cross-docks 500 PDBbind proteins against 500 ligands with MedusaDock and AutoDock Vina, producing 250,000 pairs annotated not with a single score but with the individual energy terms each program computes — van der Waals attraction and repulsion, solvation, hydrogen bonding. A graph encoder, Yuel-SE, is pretrained to regress those terms, then frozen; only a six-layer head, Yuel-AP, is fine-tuned on experimental affinities. Interaction geometry is learned from a force field, where labels are unlimited, and the scarce experimental data is spent only on mapping that representation onto measured affinity.
Jian Wang and Nikolay Dokholyan built Yuel 2 at Penn State College of Medicine in Hershey, posting it to bioRxiv in October 2024 and publishing it in the Journal of Chemical Information and Modeling in March 2025. It succeeds Yuel (2022), a structure-free predictor taking only sequence and a 2D ligand graph, which is not currently in this catalog. The same two authors published YuelPocket, a binding-site predictor, from the University of Virginia a year later; the two share a naming line but are architecturally distinct and trained independently on different data.
Yuel-SE represents the pocket as a graph of residue Cα nodes carrying their BLOSUM62 column, with edges encoding adjacency and Cα-Cα distance, and RDKit turns the SMILES into a ligand graph of atom features and bond orders. The two are merged by connecting every protein node to every ligand node and passed through eight graph attention layers into a latent vector; Yuel-AP is six linear layers on top of it.
Pretraining reaches a correlation of 0.97 for VDW_A, 0.94 for total MedusaScore and 0.80 for AutoDock Vina scores. Fine-tuning uses the PDBbind general set v2020 (19,443 complexes) split 80:20, with the 285 CASF-2016 test complexes withheld. On CASF-2016 scoring power Yuel 2 reaches Pearson 0.85 against 0.82 to 0.84 for the compared deep-learning scoring functions, and does so from pocket and SMILES rather than the crystal complex those methods consume. Ranking power averages Spearman 0.682 across the 57 targets. Per-metric Pearson correlations are 0.81 for Kd, 0.85 for Ki and 0.75 for IC50, and on a BindingDB activity-cliff set scored by a similarity-weighted affinity difference, predicted and true values correlate at 0.65.
The practical use is ranking compounds against a target without docking them first: a campaign builds one pocket graph and one graph per candidate, removing the step that dominates the cost of structure-based virtual screening and the dependence on a docked pose being correct. Predicting IC50 matters in lead optimization, where chemists work from cell-assay potency rather than equilibrium constants. The activity-cliff results speak to the same stage: telling apart analogs differing by one substituent is the judgment call in a congeneric series, and where scoring functions most often fail.
Yuel 2's contribution is a route around the data ceiling: a docking program is an unlimited label generator, and a frozen encoder trained on its energy terms carries structural knowledge a few thousand measurements cannot supply. Later systems such as Boltz-2 reach affinity by co-folding the complex, a heavier route to the same endpoint. Limits bound the claims. The pretraining set draws its 500 proteins from PDBbind, so its diversity is bounded by the database it is meant to escape, and its labels reproduce a force field's approximations rather than measurements. IC50 predictions inherit the assay conditions of the training data. On the activity-cliff benchmark Yuel 2 was trained per target to match ACGCN's protocol, testing the architecture rather than the transfer strategy, and the two perform comparably. All results are computational. Source code and test data sit in the lab's Bitbucket repository, which carries training and fine-tuning scripts but no trained checkpoints and no license file, leaving reuse terms unsettled; regenerating the pretraining corpus additionally requires MedusaDock executables requested from the authors.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.