4.6 million molecules passed in 3.2 hours: AI replaces forgotten hydrogen atoms

Researchers at NYU publish an open-source tool that determines the stable shape of tautomers; out of 5,075 reference complexes, it reports 126 structures that are probably poorly recorded.

Two molecules can share exactly the same chemical formula and behave very differently depending on the position of a single hydrogen atom. These interconvertible shapes, called tautomers, are a stubborn obstacle to computer-aided drug design: to get the wrong tautomer is to model an interaction that doesn't exist. A team at New York University trained a model to recognize the chemical patterns associated with stability and then predict where the hydrogens should go.

The work, published in the journal Chemical Science, is accompanied by an open-source tool called Tautomer-Predictor. Its speed is its main argument: about 4.6 million compounds processed in 3.2 hours on a single GPU-equipped node, a rate that allows entire molecular libraries to be screened rather than isolated candidates.

Applied to 5,075 ligand-protein complexes from the PDBbind database, the model identified 126 cases—about 2.5% of the sample—in which the recorded tautomer likely differs from the thermodynamically stable form. In other words, a non-zero fraction of the structures that serve as a reference for pharmaceutical research relies on questionable hydrogen placement.