Scaffold hopping from deposited crystal pairs: docking-free contact-conservation benchmark and structure-validated method

Scaffold hopping—replacing a ligand core while retaining function—is almost always scored by two-dimensional similarity to the reference, by docking to a single pose, or by recovery of one literature ligand, while the trivial baseline of copying the reference decorations is rarely reported. We build a benchmark from the Protein Data Bank: 2,886 ligand pairs that bind the same pocket with different Bemis–Murcko scaffolds, across 27 targets and five families, with scaffold- and target-held-out splits. Superposing the two crystal structures and transferring the experimentally solved hop into the reference pocket gives a docking-free contact-conservation criterion: the true hop recovers 80.6% of reference contacts across 40 crystal pairs, and 97% of pairs exceed 0.6; the criterion tracks pharmacophore overlap (ρ = 0.51, n = 50) but not the potency gap (ρ = −0.13), so it measures pose, not affinity. SV-Hop (structure-validated scaffold hopping)—a 60,069-rule core-replacement lexicon, exact graph swap, property/SA/novelty constraints and a learned contact ranker—returns a novel, property-conserved, synthesizable hop for 62.1% of 298 scaffold-held-out references, against 50.3% (property), 44.3% (2D), 27.0% (random) and 0% (copy). The scaffold split holds out scaffolds rather than pockets, so 97.4% of test pairs share a pocket with training; under target shift the gain falls to +5.3 pp (65.3% vs 60.0%) and the ranker transfers weakly (ρ = 0.13). A 388-rule PDB-only lexicon matches coverage and reaches 81.9% on the same target-held-out split, against 65.3% for the full lexicon, and an equal-diversity control attributes part but not all of that gap to candidate diversity. We release the benchmark, lexicon, model and a 22-compound prospective panel.