NYU Tautomer-Predictor Challenges a Hidden Weakness in AI Drug Discovery

NYU researchers trained the NYU Tautomer-Predictor on more than 1.1 million molecular states, targeting a small chemical ambiguity with consequences across drug discovery. The model predicts where hydrogen atoms sit in drug-like molecules, without first running expensive quantum calculations. That matters because moving one hydrogen can change how a candidate appears to bind with a protein.

The open-source system addresses tautomerism, where a molecule switches between closely related forms as a hydrogen moves and chemical bonds rearrange. Researchers often must choose one form before docking, virtual screening, or molecular dynamics begins. A wrong choice can distort every calculation that follows.

The deeper story is not that AI has learned another chemistry task. It is that many sophisticated drug-design pipelines still depend on a basic molecular assignment that their experimental structures do not reliably reveal. AlphaFold-style structure models, docking software, and generative systems can produce detailed predictions, yet those predictions remain vulnerable to incorrect chemical inputs.

NYU Tautomer-Predictor Turns Crystal Data Into a Screening Tool

The NYU team converted an overlooked archive of experimentally located hydrogen atoms into training data for large-scale AI tautomer prediction.

The study, published in Chemical Science on September 23, 2026, was led by first author Xiaolin Pan and senior author Yingkai Zhang. Chao Han and Fengyang Han also contributed. All four researchers are affiliated with New York University’s Department of Chemistry, while Zhang is also part of the Simons Center for Computational Physical Chemistry.

Their target was a persistent problem in computational chemistry. Many drug-like molecules exist as two or more tautomers, which have the same molecular formula but different hydrogen positions and bonding patterns. A tautomer is therefore not a different compound in the usual sense. It is an alternative chemical arrangement that the same compound can adopt.

That distinction can affect molecular recognition. A moved hydrogen changes which atoms can donate or accept hydrogen bonds. It can also change charge distribution, local geometry, and the energetic cost of binding. A protein pocket that appears compatible with one representation may interact differently with another.

Drug-design software cannot ignore that choice. A docking calculation needs a defined ligand structure. Molecular dynamics requires assigned atoms and bonds. Free-energy calculations also begin from a particular chemical representation. When the selected tautomer is wrong, computational precision cannot repair the initial mistake.

The data needed to resolve the issue has historically been limited. Solution experiments that distinguish tautomers tend to cover only hundreds of molecules, according to the NYU research announcement. That scale constrains machine-learning systems, especially when researchers want them to generalize across broad chemical libraries.

The Protein Data Bank presents another difficulty. Its protein and protein-ligand structures are often determined with X-ray crystallography, but macromolecular structures generally lack enough resolution to locate hydrogen atoms reliably. Those atoms are especially difficult to observe because hydrogen contains only one electron and scatters X-rays weakly.

Pan and colleagues looked instead to the Cambridge Structural Database. This repository contains experimentally determined small-molecule crystal structures, many with hydrogen positions resolved at higher precision. The researchers extracted more than 1.1 million tautomeric states from that material.

They used those states to train a two-dimensional graph neural network. A graph neural network represents atoms as connected points and bonds as their links, allowing the model to learn chemical patterns from molecular topology. The model does not require a three-dimensional conformer or a quantum-mechanical energy calculation before making its prediction.

The peer-reviewed tautomer study reports recall of 0.97 and precision of 0.88 across a test set covering crystal structures, aqueous-solution benchmarks, and drug-like molecules. Recall measures how often the system retrieves relevant stable states. Precision measures how often its positive predictions are correct under the study’s labels.

Those results make the system promising as a filtering and assignment tool. They do not mean it delivers a universally correct answer in every solvent, protein pocket, or biological environment. The model learned primarily from proton positions observed in crystals, while real drug discovery involves changing chemical conditions.

Still, the scale changes what is operationally possible. Earlier experimental datasets were too small to support the same breadth. By treating crystallographic hydrogen positions as supervision, the team built a training set large enough for library-scale screening.

That data strategy is the central advance. The architecture matters, but the study’s sharper insight is that an existing scientific archive contained useful labels hiding in plain sight. The team did not need to generate more than one million new laboratory measurements. It reorganized available structural evidence around a neglected prediction task.

One Hydrogen Can Reshape a Drug-Protein Interaction

Tautomer assignment is an upstream decision, so an error can spread through docking, scoring, simulation, and lead selection.

A drug candidate must interact with its target through a combination of shape, electrostatics, hydrogen bonding, and movement. Tautomers can alter several of those factors without changing the compound’s molecular formula. That makes the problem easy to underestimate and difficult to catch later.

Consider a ligand with a hydrogen that can occupy either of two nitrogen atoms. In one form, the first nitrogen donates a hydrogen bond while the second accepts one. Moving the hydrogen can reverse those roles. The protein pocket has not changed, but the predicted interaction network has.

Docking programs typically place candidate ligands into a binding site and score the resulting poses. If the input tautomer offers the wrong donors and acceptors, the software can prefer an unrealistic orientation. It can also rank a weaker candidate above a stronger one because their chemical representations were prepared inconsistently.

Molecular dynamics adds time and motion, but it does not automatically correct every chemical assignment. The simulation follows the supplied atoms, bonds, and force-field parameters. As Zhang explained in the research announcement, the wrong tautomer can alter predicted interactions and dynamics because those calculations require correctly assigned hydrogen positions and bonding.

This is why tautomer prediction in drug discovery is more than a cleanup task. It sits near the beginning of a dependency chain. Errors introduced there can survive into virtual screening results, free-energy estimates, and the selection of molecules for synthesis.

The team tested this concern on 5,075 PDBbind ligands that could occupy multiple tautomeric states. PDBbind is a curated collection connecting protein-ligand structures with binding data. The model flagged 126 ligand assignments as likely incorrect, equal to about 2.5 percent of the evaluated set.

For each flagged case, the model’s alternative assignment produced improved hydrogen-bonding patterns and fewer unsatisfied polar atoms. An unsatisfied polar atom is capable of a favorable polar interaction but lacks a suitable partner in its modeled environment. Reducing those mismatches can make a binding pose more chemically plausible.

That result should be interpreted carefully. The researchers did not claim that the experimentally determined protein structures were wrong. They argued that the chemical representation assigned to the ligand may warrant revision. Atom coordinates and the formal description attached to those coordinates are related, but they are not identical.

The distinction matters because structural databases influence later models. Researchers use deposited complexes to benchmark docking methods, train machine-learning systems, and evaluate protein-ligand predictions. A modest rate of questionable tautomer assignments can therefore affect more than the original records.

The reported 126 cases remain model-generated reassignments, not 126 newly confirmed experimental corrections. Improved hydrogen-bonding patterns support their plausibility, but prospective experiments would provide stronger evidence. Neutron diffraction, high-resolution spectroscopy, or other targeted techniques can sometimes locate hydrogens more directly.

The pressure therefore falls on computational chemistry teams, database curators, and developers of AI drug-design systems. They need to treat proton placement as an explicit uncertainty rather than a hidden preprocessing choice.

A single global accuracy number cannot settle that issue. Teams must know whether the model performs consistently across their chemical series, charge states, functional groups, and relevant environments. They also need workflows that retain alternative tautomers when the evidence is ambiguous.

Fast AI Tautomer Prediction Pressures Quantum-First Workflows

The primary contest is between fast, learned assignment and slower physics-based evaluation, not between two drug-discovery companies.

Quantum-mechanical methods estimate molecular energy from electronic structure. They provide a principled way to compare tautomers, especially when researchers account for solvent and conformational effects. Their limitation is computational cost, particularly across libraries containing millions of candidates.

Rule-based systems occupy the faster end of the spectrum. They enumerate or rank tautomers using expert-written transformations, empirical scores, and known chemical patterns. These tools are practical, but fixed rules can struggle outside the chemistry represented by their design assumptions.

Earlier machine-learning approaches created a middle path. Some used three-dimensional conformers or calculated molecular properties, which introduced additional preparation costs. Others learned from small solution-phase datasets that limited chemical coverage.

The NYU Tautomer-Predictor changes that tradeoff by scoring candidates directly from two-dimensional topology. The authors say it needs neither conformer generation nor quantum calculations. That does not make physics obsolete, but it gives screening teams a faster first pass.

The throughput test illustrates the difference. The open-source workflow processed about 4.6 million compounds from an Enamine collection in roughly 3.2 hours on one GPU-enabled node. That is close to 400 compounds per second when averaged across the reported run.

At that scale, AI tautomer prediction can sit near the entrance of a screening pipeline. A team could enumerate plausible forms, rank them, and pass a smaller set into docking or more expensive energy calculations. Physics-based methods could then focus on uncertain or high-value compounds.

This tiered strategy is more credible than replacing quantum chemistry outright. A learned model offers speed and pattern recognition, while quantum methods provide detailed energetic treatment. Experiments remain necessary when a tautomer assignment directly influences a costly medicinal chemistry decision.

The study’s open-source code also lowers the barrier to testing. Researchers can inspect the workflow, run it on internal molecules, and compare its rankings against established preparation procedures. Open availability matters because reproducibility is a recurring concern in AI-assisted chemistry.

Tautomer-Predictor uses an AttentiveFP-style architecture, according to the released implementation. This model family learns molecular representations by repeatedly combining information from neighboring atoms and bonds. Attention mechanisms help the network weight local chemical environments during prediction.

That architecture differs from the large generative models receiving much of the drug-discovery attention. It does not propose an entirely new molecule, predict a protein’s full structure, or generate a clinical candidate. It performs a narrower classification task that can improve the inputs to those larger systems.

Narrowness is an advantage here. The output maps to a recognizable chemical decision, and the team can evaluate it against experimental structural labels. The system also exposes uncertainty more clearly than an end-to-end claim that a model has “designed a drug.”

Related work from the same group shows why environment remains important. A 2025 study introduced a Siamese neural network for predicting tautomer ratios in aqueous solution. That solution-phase model focused on relative tautomer populations rather than classifying stability from crystal-derived data.

Together, the projects suggest a layered direction for computational chemistry. Crystal data can supply broad structural supervision. Solution measurements can refine environmental behavior. Quantum calculations and experiments can resolve cases where context changes the preferred form.

The immediate competitive pressure lands on software pipelines that treat tautomer preparation as a fixed rules engine or an unexamined default. A fast learned alternative gives pharmaceutical teams a reason to benchmark those defaults again.

It also raises expectations for other chemistry models. If a system predicts binding affinities or generates ligands, developers should disclose how they chose protonation and tautomeric states. Otherwise, an impressive benchmark may partly reflect inconsistent molecular preparation.

Crystal Labels Are Valuable, but They Are Not Biology

The model’s strongest asset, its large crystallographic dataset, also defines the main limit on what its predictions establish.

A crystal lattice is not an aqueous solution, a protein binding pocket, or a living cell. Molecules in crystals interact with neighboring copies, counterions, and packing forces. Those conditions can favor a tautomer that differs from the dominant state elsewhere.

The paper’s test set included aqueous-solution benchmarks and drug-like molecules, which supports some transfer beyond crystals. Its reported precision and recall are encouraging across those categories. Still, aggregate performance does not guarantee correct behavior for every scaffold or chemical environment.

Protein pockets create particularly complex conditions. Local electrostatics, buried water molecules, metal ions, nearby residues, and conformational changes can shift tautomer preferences. Binding can even stabilize a less abundant solution-state tautomer when that form makes favorable interactions.

The model also works from two-dimensional topology. This design enables speed, but it omits some three-dimensional context during its initial ranking. The absence of a protein environment is especially relevant when two tautomeric forms have similar intrinsic stability.

For that reason, the 126 PDBbind reassignments should be treated as high-value review candidates. Better hydrogen bonds and fewer unsatisfied polar atoms strengthen the alternative representations. They do not independently establish which tautomer occupied each experimental complex.

Dataset construction introduces further questions. Crystallographic records are not free from curation errors, disorder, uncertain hydrogen placement, or biases toward molecules that crystallize successfully. The model can learn useful chemical patterns while also inheriting systematic features of its source database.

Precision of 0.88 also means false positive predictions remain part of the expected output under the reported evaluation. That level can be acceptable for screening, where the model helps prioritize possibilities. It is less suitable for treating every top-ranked tautomer as settled fact.

Drug developers should therefore preserve provenance. A downstream result should record which tautomers were enumerated, how they were scored, what cutoff was used, and whether protein context changed the ranking. Without that trail, the tool risks turning an explicit chemical uncertainty into another invisible default.

External validation will be important. Independent groups need to benchmark the model on proprietary medicinal chemistry series, prospective experiments, and targets with unusual binding-site environments. Comparisons should include common commercial preparation tools, open-source rules, quantum calculations, and expert review.

The model’s release also arrives during broader scrutiny of AI drug discovery. Systems can excel at retrospective benchmarks while losing accuracy on new chemistry or real experimental endpoints. A 2026 evaluation of molecular generation methods found that different architectures offered different advantages and no method optimized every criterion consistently.

That wider pattern reinforces the appropriate role for the NYU system. It can strengthen molecular preparation and catch suspicious assignments. It cannot determine efficacy, toxicity, selectivity, pharmacokinetics, or clinical success.

Even the term “stable” needs context. A molecule can occupy several tautomeric states in an equilibrium, with their relative populations changing by solvent, temperature, pH, or binding partner. A classifier that selects the most plausible state simplifies a distribution that may remain chemically meaningful.

The right response is not to dismiss the model because biology is complicated. It is to use the model where its evidence is strongest and escalate uncertain cases. Library-scale speed is most valuable when paired with a clear route to higher-fidelity validation.

This approach also helps avoid a common AI failure mode. Automation can make an answer appear final simply because it arrives quickly. In chemistry, rapid output should increase the number of hypotheses tested, not reduce the care used to verify important ones.

The Next Tests Will Determine Its Role in Drug Discovery

Three signals will show whether AI tautomer prediction becomes standard infrastructure or remains a promising research utility.

The first signal is independent prospective validation. Researchers should select molecules before experiments, predict their preferred tautomers, and then measure the relevant states under controlled conditions. Success on new compounds would provide stronger evidence than retrospective agreement with existing databases.

Such tests should span crystals, aqueous solutions, and protein-bound environments. Agreement in one setting does not guarantee agreement in another. Reporting confidence and alternative states would also reveal whether the model knows when a case is ambiguous.

A prospective study that confirms top-ranked predictions across diverse chemistry would strengthen the case for routine use. Repeated failures tied to particular functional groups or environments would narrow the model’s practical scope. Either result would improve deployment decisions.

The second signal is adoption in preparation and screening pipelines. The reported ability to process 4.6 million compounds in 3.2 hours makes integration technically plausible. The more important question is whether the model changes decisions.

Teams should measure how often it alters docking poses, binding rankings, free-energy results, or candidate selection. A preprocessing tool delivers value only when its corrected assignments improve downstream outcomes or prevent wasted calculations.

Integration with established chemistry platforms would also indicate maturity. Useful implementations need dependable tautomer enumeration, molecule standardization, batch processing, error handling, and reproducible output. A model checkpoint alone does not create production infrastructure.

The public prediction workflow gives researchers a starting point. Its impact will depend on independent testing, maintenance, documentation, and compatibility with the software already used by medicinal chemists.

The third signal is database correction and curation. The model flagged approximately 2.5 percent of the evaluated PDBbind ligands, creating a concrete set for expert review. If curators and structural chemists confirm a meaningful portion, the study will have identified a correctable weakness in widely reused research data.

Confirmed revisions would matter beyond those 126 complexes. Protein-ligand databases support benchmarking and model training across computational drug discovery. Better chemical assignments can improve the quality of future systems built on the same records.

If reviewers reject most proposed changes, that would weaken claims about the model’s ability to transfer from crystal-derived labels into protein environments. The test is therefore unusually clear. The authors have produced specific predictions that specialists can investigate.

The broader lesson already extends past tautomerism. AI systems depend on how scientists encode molecules, proteins, assays, and experimental conditions. More capable architectures cannot compensate for every ambiguity in those representations.

The NYU Tautomer-Predictor brings one such ambiguity into view. It shows that a hydrogen atom absent from an experimental map can influence a long chain of computational conclusions. It also shows how curated scientific data can make that problem tractable at scale.

For developers, the practical question is whether their models treat tautomer choice as data, uncertainty, or an unnoticed default. For pharmaceutical teams, the question is whether inexpensive AI triage can reduce the number of costly calculations launched from questionable structures.

For researchers evaluating AI tautomer prediction, the next step is concrete: run the system beside an existing molecular-preparation workflow, retain competing states, and inspect every disagreement that reaches a high-value candidate. The NYU model offers speed and a substantial experimental training base. Prospective measurements must now establish where that advantage holds, where environmental context changes the answer, and whether better tautomer assignments produce better drug-discovery decisions.