NYU develops Tautomer-Predictor AI to identify stable drug-like molecules

Mining Crystal Databases to Train Graph Neural Networks

Traditional molecular modeling often stalls because many compounds share a single molecular formula while constantly shifting between related structures where hydrogen atoms migrate. These variations change local bonding patterns, yet identifying the correct tautomer has historically relied on scarce experimental data. As noted by Dr. Yingkai Zhang, professor of chemistry at NYU and the study’s senior author:

“Although this may seem like a small change, different tautomers of the same molecule can alter how a molecule interacts with a protein target.”

To overcome the limitations of low-resolution Protein Data Bank structures and computationally expensive quantum mechanics, Dr. Xiaolin Pan mined the Cambridge Structural Database. By extracting experimentally resolved hydrogen positions from high-resolution small-molecule crystal structures, the NYU team built a training dataset of over 1.1 million tautomeric states. This vast repository allowed their graph neural network to learn stability patterns directly from 2D structures.

Reassessing Existing Biomolecular Complexes

When the research team tested their model on 5,075 biomolecular complexes from the PDBbind database, it exposed hidden discrepancies in previously cataloged data. The AI flagged 126 instances—roughly 2.5 percent of the total tested ligands—where the originally assigned tautomer in the Protein Data Bank was likely incorrect. Zhang clarified that this does not mean the underlying protein structures are flawed:

“Rather [that] our results suggest that the previously assigned chemical representation may warrant revision.”

In these identified cases, the model offered alternative tautomers featuring superior hydrogen bonding networks with adjacent protein residues. Ensuring accurate hydrogen placement is vital for molecular dynamics simulations, which track how drug candidates and protein targets move and interact over time.

High-Throughput Screening of Massive Compound Libraries

Built for speed and accessibility, the open-source Tautomer-Predictor tool can rapidly analyze millions of compounds without needing three-dimensional observations. During testing, the software processed approximately 4.6 million drug-like compounds in just 3.2 hours on a single GPU-enabled computing node. By rapidly sorting through vast libraries, the system helps researchers prevent costly errors in virtual screening caused by incorrect tautomer assignments affecting solubility, safety, or biological activity.