2026-09-21
USC MGVS uses three SBDD generators as probes, then GED-searches catalog libraries. On 30 targets, docking 2k candidates beats 50k random ZINC, a 25x efficiency gain.
Structure-based generative models can place atoms inside a protein pocket. The molecules they emit are often not makeable. Gao and Coley already showed in 2020 that goal-directed generators routinely propose structures a synthesis planner cannot route. Early discovery needs compounds that can be bought and assayed in days; custom synthesis does not move at that speed.
Two existing exits both cost something. Screening only catalog libraries (Enamine REAL, ZINC, WuXi GalaXi) has scaled to billions of make-on-demand entries, and exhaustive docking does not keep up. Constraining generators to synthesizable subspaces cuts diversity and control. The USC Rohs lab's bet is different: treat the generator as a probe that compresses chemical space, then retrieve neighbors that already sit in a buyable catalog.
The pipeline is model-guided virtual screening (MGVS), in five steps:
The test set is 30 PDBbind targets, one per cluster after 30% sequence-identity clustering. The VLS baseline is 50k random compounds from the ZINC20 drug-like subset. Raw Vina tracks molecular size, so the headline metric is Vina efficiency (score divided by heavy-atom count). ECFP4 and Daylight distances are reported as weaker controls; later numbers show GED tracks pocket shape better.
Catalog hits are far more synthesizable by SA score than the generated queries. Docking scores usually hold: 98.7% of top-1 hits land inside the query's Vina error band of ±1.5 kcal/mol, and the mean is slightly better. Versus the PDB co-crystal ligand, almost every top-1 hit matches or beats the crystal score, and 38.8% do so by more than 1.5 kcal/mol.
Worse similarity predicts worse scores. Spearman ρ between GED and a worse binding score is 0.44, versus 0.31 for ECFP4 and 0.14 for Daylight. The generated query is acting as a real template.
Against random ZINC:
| Source | Docking budget | vs random ZINC |
| Pocket2Mol queries | 1k generated | beats 10k |
| DrugHIVE queries | 1k generated | beats 30k |
| DiffSBDD queries | 1k generated | beats 50k |
| Search-hits, all three models | 2k including queries | all beat 50k |
That 2k-versus-50k comparison is the claimed 25x screening efficiency. Even one nearest neighbor per query keeps the upper 95% CI of ΔVina at +0.3 kcal/mol. Mean ΔVina crosses zero around 40 docked hits per query.
Poses are not coincidental. On non-hydrophobic contacts from PLIP (H-bonds, salt bridges, π interactions), 80.7% of queries have a top-1 hit that matches every specific interaction at residue level, and 99.2% share at least one. The weakest target, PDB 5G1P, still yields a residue-level full match for 20.0% of queries. For π-stacking, 0.92 of top-100 hits share at least one residue-level contact; for H-bonds the figure is 0.69.
This is an engineering claim, not a new generator. Existing 3D models can already enter early discovery if they are not asked to emit orderable molecules. Generation compresses space; catalogs land the hit. All three architectures worked, so the result is not tied to one generator family.
For screening teams with a known pocket and a catalog, 2k dockings for 50k-quality top-10s is immediately testable. For model builders, the optimization budget belongs on occupying the pocket; synthesizability can sit with retrieval. The authors read that as a reason not to lock generators inside reaction-template subspaces, because those analogs often docked better than the generated queries themselves.
Every number is a docking number. QuickVina2 itself quotes ±1.5 kcal/mol uncertainty. There is no wet assay, no ADMET, no purchase-and-make. "Synthesizable" here means present in Enamine REAL, ZINC, or WuXi GalaXi, not already made in a flask.
The baseline is random ZINC, not the synthon-based VLS methods that actually scale to billion-compound spaces. The 25x figure is versus random enumeration, not versus V-SYNTHES. Filtering out non-five/six-membered rings and complex fused systems also trims the generated space before retrieval success is measured.
The authors note the method is hostage to approximate search over growing catalogs; if retrieval lags, MGVS idles. Their case against synthesizable-subspace generators rests on the same docking-only evidence. Whether a constrained model would convert better in a wet lab is unanswered.
The JCIM version of record sits behind ACS Cloudflare. This write-up follows the public-domain bioRxiv preprint of the same paper; journal-only additions, if any, are not covered.