Crystal-pair benchmark tests scaffold hopping without docking
- Builds a benchmark from PDB ligand pairs sharing a pocket but differing in scaffold, scoring hops by how well reference protein contacts are conserved after crystal superposition, with no docking.
- SV-Hop combines a large core-replacement rule library, property and synthesizability filters, and a learned contact ranker.
- Evaluated retrospectively; it outperforms property-, 2D-similarity-based and random baselines, but the advantage shrinks markedly when whole targets are held out.
- A prospective compound panel is released. Preprint, not peer reviewed.
💡 Points Based on the abstract
The held-out-target drop is the number to watch: PDB pairs sharing a pocket invite leakage, since a ranker can memorize contact patterns for familiar families. Contact conservation is also only a proxy; it says nothing about whether a hop keeps potency. Before trusting it, run it on our own series and compare synthesized hops against measured activity.