A network trained on crystal hydrogens disputes the tautomers of 126 PDBbind ligands

The percentage needs its denominator. The 5,075 PDBbind ligands in the test were selected because each admits more than one tautomer, so the 126 disagreements sit inside the population where disagreement was possible at all ^[13]. That is about one in 40 of the ambiguous cases ^[17].

In those 126 cases the disagreement is about chemical plausibility. "Although this may seem like a small change, different tautomers of the same molecule can alter how a molecule interacts with a protein target," said Yingkai Zhang, the NYU chemistry professor and senior author of the study, which is published in Chemical Science ^[4][2]. The model moved a hydrogen and the resulting hydrogen-bonding pattern improved; in one ligand the reassignment let the molecule form additional hydrogen bonds with nearby protein residues ^[13][15]. "Reassigning these tautomers generally produced more chemically reasonable interactions. This does not mean that the experimentally determined protein structures themselves are incorrect; rather, our results suggest that the previously assigned chemical representation may warrant revision," Zhang said ^[14].

The training labels come out of crystals ^[11]. Molecules in the Cambridge Structural Database sit packed in a lattice; a drug candidate binds in a solvated pocket. The other source of experimental truth is solution-phase measurement, and those datasets hold a few hundred molecules ^[8]. Take 300 as a working size: the crystal-derived set is then roughly 3,700 times larger ^[18].

phys.org's account puts the throughput at 4.6 million compounds in hours ^[16]. The evidence offered for the corrections is the hydrogen-bonding pattern, and the account does not report a docking score or an affinity estimate recomputed before and after the reassignment ^[19]. For a screening campaign, the test worth running is whether any of the 126 corrections changes a rank order.

On the evidence reported, the sensible use is a bulk pass whose disagreements go into a review queue. Two conditions attach to that. The labels are solid-state, and the check on the reassignments is internal chemical reasonableness rather than an independent measurement of where the hydrogen sits in the bound complex ^[13]. Calling library curation finished would take a solution-phase test set far larger than the few hundred molecules now available ^[8], or a downstream result showing that corrected tautomers predict binding affinities better than the deposited ones.