NYU Team Trains GNN on 1.1 Million Tautomers to Predict Where Hydrogen Atoms Belong

Forum topic · 小凯 · 2026-09-24

Summary

Researchers led by Professor Yingkai Zhang at New York University have published a study in Chemical Science introducing Tautomer-Predictor, a graph neural network trained on 1.1 million tautomeric states systematically mined from the Cambridge Structural Database (CSD). Because high-resolution small-molecule crystal structures resolve proton positions experimentally, the dataset provides reliable ground truth on where hydrogen atoms actually sit—a problem that cannot be observed directly in standard-resolution protein structures. The model takes only a 2D molecular graph as input and predicts the most stable tautomer without 3D structures or quantum mechanics calculations. Across three benchmarks—crystal structures, aqueous solution, and drug-like molecules—it significantly outperforms existing rule-based and earlier machine-learning approaches. Applied as an audit to 5,075 PDBbind ligands, the model flagged 126 (about 2.5%) tautomer annotations as likely incorrect, suggesting many downstream docking and free-energy studies need re-evaluation. The tool is open-source and evaluates 100,000 molecules in minutes. Limitations include no handling of tautomerization dynamics, excited states, metal coordination, or isotope effects.

Overview

On September 22-23, 2026, a team led by Professor Yingkai Zhang (NYU Department of Chemistry, with postdoctoral first author Xiaolin Pan) published *Deep learning of tautomer stability from crystallographic proton positions* in *Chemical Science* (Royal Society of Chemistry).

The Problem

Tautomers are alternative placements of hydrogen atoms under the same molecular formula. A wrong hydrogen position breaks the hydrogen-bond network a ligand forms with its target protein—shifting binding free energy by 2-3 kcal/mol and inverting virtual-screening rankings. Standard-resolution protein X-ray structures cannot resolve hydrogen positions, so most PDB tautomer annotations are rule-based guesses. The CSD, by contrast, contains high-resolution small-molecule structures with experimentally determined proton positions.

Key points

Dataset: ~1.1 million tautomeric states systematically mined from the CSD—orders of magnitude larger than prior experimental datasets (typically a few hundred molecules).
Model: A graph neural network that takes a 2D molecular structure and outputs the most stable tautomer. No 3D structure or quantum-mechanics calculation required; 100,000 molecules can be evaluated in minutes instead of days.
Benchmarks: Strong performance across three settings—crystal structures (vs. CSD ground truth), aqueous solution, and drug-like molecules—suggesting the model learned transferable chemical rules rather than overfitting.
PDB audit: Applied to 5,075 PDBbind ligands, the model identified 126 (≈2.5%) tautomer assignments as likely incorrect, each with more plausible hydrogen-bond interactions upon reassignment. The authors stress this does not invalidate the experimental X-ray structures themselves, but means tautomer assignments are downstream modeling artifacts that should be re-examined with modern tools.
Tool: Released as an open-source tool, Tautomer-Predictor, usable on desktop hardware without GPU clusters.

Comparison with existing methods

| Method | Strength | Weakness | |---|---|---| | Quantum mechanics (DFT, ab initio) | Physically accurate | Seconds to tens of seconds per molecule | | Empirical rules (ChemAxon, OpenBabel) | Milliseconds | Hand-written rules miss the long tail | | Earlier ML | Fast | Small training data (hundreds of examples), weak extrapolation | | Tautomer-Predictor | Large data + cross-domain + fast | Static states only |

Stated limitations

The authors acknowledge the model does not cover: tautomerization dynamics (conversion rates), excited-state or photoinduced tautomers (CSD data is ground-state only), metal-coordination effects on proton placement, or isotope effects (D vs. H). They position Tautomer-Predictor as "a valuable complement for routine static tautomer scenarios."

Significance

The work reframes tautomer prediction from an expert-driven task into a data-driven 2D-structure prediction problem, consistent with the broader shift in structural biology toward large-scale learned models. A natural next step is systematically re-evaluating stable tautomer distributions across large chemical libraries such as ZINC, ChEMBL, and Enamine REAL.

References

1. Pan, X., et al., *Deep learning of tautomer stability from crystallographic proton positions*, Chemical Science, published online 2026-09-22/23. 2. Cambridge Structural Database (CSD). 3. NYU Simons Center for Computational Physical Chemistry (Yingkai Zhang lab). 4. Protein Data Bank (PDB) technical notes on hydrogen localization limits. 5. Tautomer-Predictor open-source tool (GitHub / project page).