Loading…
Equivariant heterogeneous graph network that rescores docked protein-ligand poses for virtual screening and ranks structural analogs by activity.
No providers recorded yet. Browse all providers
A deep-learning scoring function can post an excellent number on DUD-E and still be useless on a target it has never seen, because the benchmark's proteins are usually in the training set too. Deduplicating by PDB ID does not fix it: all 67 proteins in CASF-2016 also appear in PDBbind2020, and over 70% of its ligands have an analog bound to the same protein there. A model can then rank actives by recognizing a familiar scaffold rather than by reading the interaction, and stripping the overlapping targets out makes the scores collapse — RTMScore's BEDROC on DEKOIS2.0 falls from 0.541 to 0.352, TANKBind's DUD-E AUROC from 0.778 to 0.583, first place to last.
EquiScore, from the Drug Discovery and Design Center of the Shanghai Institute of Materia Medica with collaborators at Zhejiang University, attacks that failure from both directions. On the model side it represents a docked pose as a heterogeneous graph in which chemistry is written into the topology rather than inferred from distances: a virtual node stands in for each aromatic ring so pi-systems act as units, and empirical interaction fingerprints computed by ProLIF become their own class of edge, making a hydrogen bond or hydrophobic contact an explicit relation the network can attend to. Attention between atom pairs is gated by geometric distance, and node features are updated by equivariant message passing, so rotating or translating the complex cannot change the score.
On the data side the authors built a training corpus, PDBscreen, and then deliberately shrank it. Positives are re-docked crystal complexes, keeping near-native poses within 2 Å RMSD so the model learns from the imperfect poses it will meet. Negatives come from cross-docking, so the same ligand appears as both positive and negative against different proteins and scaffold memorization cannot pay off, plus property-matched DeepCoy decoys filtered by shape overlap. Proteins sharing a UniProt ID with any test-set target were then removed.
EquiScore is a binary classifier over a graph built from residues within 8 Å of the ligand, with geometric edges between atoms closer than 5.5 Å, covalent edges, and ProLIF interaction edges. Stacked EquiScore layers combine info-aware attention, an EGNN-style node update and an edge update that folds attention back onto structural edges; ligand atoms are pooled into a three-layer MLP. PDBscreen was assembled from the 180,207 PDB entries available in July 2021, filtered to better than 2.5 Å resolution and stripped of cofactors and covalent binders, with cross-docked negatives outnumbering positives roughly ten to one. Against 21 competing methods EquiScore reached the top AUROC of 0.821 on DEKOIS2.0 (81 targets, 40 actives and 1,200 decoys each) versus 0.756 for the runner-up, and on DUD-E (102 targets, 8 protein families) it led on unseen targets with BEDROC 0.432 and 1% enrichment factor 17.675. Rescoring Glide SP poses raised 1% EF to 16.83, and lifted weaker docking programs' 1% EF two- to three-fold. On the LeadOpt analog set its weighted Spearman correlation of 0.57 trailed only Schrödinger's FEP+ at 0.73, at a fraction of the compute.
The practical use is as a rescoring stage: dock a library with whatever program a group already licenses, extract the pocket, and re-rank the poses before ordering compounds. That helps most where hit rate matters and the target is novel — orphan receptors, newly solved structures, targets with no congeneric series. The analog-ranking result extends it into lead optimization, where a chemist choosing among close analogs would otherwise wait on a free-energy perturbation run. Released scripts cover pocket extraction, library screening and benchmark reproduction, and the attention visualizations give a per-atom rationale rather than a bare number.
EquiScore's most durable contribution may be methodological rather than architectural: its analysis of soft overlap between PDBbind and the standard screening benchmarks, and its UniProt-level deduplication scheme, showed concretely that published scoring-function comparisons were inflated. The thread runs into later work on resolving data bias in affinity prediction, while catalog entries such as BioScore pursue the same goal of one checkpoint spanning many target classes. The limits are worth stating plainly: EquiScore ranks rather than predicting a calibrated affinity; it needs an external docking program for poses; and PDBscreen was built with commercial Schrödinger tooling, so rebuilding the training corpus requires licenses the released code does not include.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.