NYU's Tautomer-Predictor Skips Quantum Calculations and Flags 126 PDBbind Ligands: What the September 2026 Chemical Science Paper by Pan, Han, Han and Zhang Shows
Quentir Medicine Monitor
Evidence-based insights for quantum medicine. Published by Quentir Systems LLC · September 27, 2026.
A single hydrogen atom can sit in two places on the same drug molecule. When it moves, the pattern of single and double bonds shifts with it, and the molecule presents a different face to the protein it is meant to bind. Chemists call these interchangeable forms tautomers, and choosing the wrong one at the start of a computer-aided drug design project can distort the docking scores and free-energy estimates that follow.
A paper in the Royal Society of Chemistry journal Chemical Science, "Deep learning of tautomer stability from crystallographic proton positions" by Xiaolin Pan, Chao Han, Fengyang Han and Yingkai Zhang of New York University, published online in September 2026, attacks this problem with an unusual source of training data. The team mined hydrogen positions resolved in small-molecule crystal structures from the Cambridge Structural Database, built more than 1.1 million tautomeric states from them, and trained a graph neural network to predict tautomer stability directly from a molecule's two-dimensional structure. The authors state that the model needs no 3D conformers and no quantum-mechanical calculations, which is what makes it fast enough for structure-based drug discovery at library scale.
The quantum-medicine interest lies in what the network replaces. Quantum chemistry has long been the accurate but slow reference for deciding which tautomer a molecule prefers, and it is also one of the problems quantum computing companies cite when they pitch molecular simulation to pharmaceutical buyers. Here a classical model trained on experimental crystal data takes over a large part of that job on a single GPU node.
The results come from retrospective tests on public structure collections. The study screened no new compounds in the laboratory and involved no patients, and its training labels describe molecules in the crystal, which is a different environment from a protein pocket or blood plasma.
What Xiaolin Pan and Yingkai Zhang built at NYU, and where the training data came from
Zhang, a professor of chemistry at NYU, leads the group; Pan, a postdoctoral researcher in the lab, is first author. Their starting point was a data shortage. According to the account of the study published by Genetic Engineering & Biotechnology News on 23 September 2026, experimentally characterized tautomer sets in solution usually hold only a few hundred molecules, far too few to train a model that generalizes. The Protein Data Bank, the main public archive of protein structures, rarely helps either, because macromolecular X-ray structures are generally not resolved finely enough to locate hydrogen atoms.
Small-molecule crystallography is different. Many high-resolution structures in the Cambridge Structural Database, maintained by the CCDC in Cambridge, England, carry explicitly assigned hydrogen atoms. The NYU team treated those assigned positions as experimental labels. From them Pan assembled the training set of more than 1.1 million tautomeric states, which the authors describe as orders of magnitude larger than earlier experimental sets. The model itself is a graph neural network: it reads a molecule as atoms connected by bonds, enumerates the candidate tautomers, scores each one, and ranks the likely stable states.
Why quantum-mechanical calculations have been the slow reference for tautomer assignment
Deciding which tautomer is most stable is at heart a question about electronic energies, and quantum chemistry answers it from first principles. The difficulty is cost. The paper's abstract states that quantum-mechanical approaches are often too computationally demanding for library-scale use, while rule-based tools and 3D-dependent machine-learning methods trade away either transferability or throughput. A screening campaign that runs to millions of purchasable compounds cannot afford a careful quantum calculation for every candidate form of every molecule.
This is the same bottleneck that quantum computing vendors describe when they offer molecular simulation to drug developers. Quentir has followed several of those efforts, including the 3,000 variational quantum eigensolver emulations that Phasecraft and NVIDIA computed on GPUs for a molecular dataset. The NYU paper is a useful reference point for that market. In Quentir's reading, the paper sets a new baseline for one well-defined, high-volume task: a classical network trained on experimental data now offers a fast answer, and any quantum method aimed at the same task will have to beat it on accuracy, cost or both. The paper itself makes no comparison with quantum computing.
Quantum pillar: not applicable. Technology readiness: TRL 4 of 9. The tool runs no quantum technology of its own; it has been checked retrospectively against public crystal and protein-ligand structure collections and released as open-source software, with no prospective drug-design study or laboratory confirmation of its reassignments reported yet.
What the model found in 5,075 PDBbind ligands and the 4.6-million-compound Enamine collection
The most concrete test used PDBbind, a curated set of protein-ligand complexes drawn from the Protein Data Bank. The authors applied the model to 5,075 ligands that can exist in more than one tautomeric state. It flagged 126 ligands, about 2.5 percent, whose assigned tautomer was likely incorrect. In each of those cases, according to the abstract, the reassigned stable tautomer showed improved hydrogen-bonding patterns with the surrounding protein and fewer unsatisfied polar atoms. Zhang was careful about what this means. "This does not mean that the experimentally determined protein structures themselves are incorrect," he said in the GEN account; the chemical representation of the ligand "may warrant revision."
The second test measured speed. The released workflow, Tautomer-Predictor on GitHub, processed the 4.6-million-compound Enamine collection in about 3.2 hours on one GPU-enabled node. The group also offers a free web version on Hugging Face. For a medicinal chemistry team, that turns tautomer checking from a specialist step applied to a shortlist into a routine clean-up pass over an entire screening library.
Where crystal-trained predictions can mislead a drug-discovery team
The training labels carry a built-in limitation that the authors themselves name: the hydrogen positions describe tautomeric states "in the crystalline environment." A molecule packed into a crystal lattice, a molecule dissolved in water and a molecule bound inside a protein pocket can prefer different forms, because neighboring molecules, solvent and charged residues all shift the energy balance. A crystal-trained model inherits the preferences of the crystal.
The PDBbind result should be read with the same care. The improved hydrogen bonding of the 126 reassigned ligands is a plausibility check computed from existing structures. This study does not report a prospective test showing that the reassignments change which compounds a team would synthesize, or that they improve hit rates or binding predictions for a real program. Those are the measurements that would move the tool up the readiness ladder, and they belong to the next round of work.
What the Chemical Science paper means for medicinal chemists and quantum chemistry vendors in 2026
For drug discovery groups, the practical gain is immediate and modest: an open, fast way to catch a class of input errors that can distort docking, free-energy calculations and molecular dynamics before any compound is made. Better starting structures do not produce a medicine, and the path from a corrected ligand file to a patient still runs through synthesis, assays, animal studies and clinical trials.
For the quantum side of the field, the paper sharpens a question buyers are already asking. Quantum chemistry remains the reference for tautomer energetics, and quantum computers are promoted partly on their promise to make such calculations cheaper at scale. When experimental data can train a classical model that handles millions of molecules in an afternoon, a quantum approach has to show its advantage on the harder cases that remain: unusual scaffolds, metal-containing drugs, and tautomers whose preference depends on the protein environment. That is where a hospital pharmacy researcher or a biotech buyer should expect the next credible quantum claim to be tested.
Sources
Primary source: Xiaolin Pan, Chao Han, Fengyang Han and Yingkai Zhang (New York University), Chemical Science, published online September 2026. Also drawn on: Genetic Engineering & Biotechnology News for the study account and Yingkai Zhang's quotations (23 September 2026); the authors' Tautomer-Predictor repository; and the CCDC and RCSB Protein Data Bank for the structure archives; the readiness assessment is this Monitor's own.