AI Model Predicts Stable Molecular Forms to Sharpen Drug Discovery

Tautomerism occupies a curious position in chemistry: it is conceptually simple, involving nothing more than the relocation of a single proton and a corresponding rearrangement of bonds, yet its consequences ripple through nearly every stage of drug development. The relative populations of tautomeric forms are governed by subtle thermodynamic differences, often amounting to only a few kilocalories per mole, which is precisely the energy scale at which computational chemistry becomes both most valuable and most fragile. A model that misplaces a proton by one heteroatom can therefore produce binding free energy estimates, docking poses, and molecular dynamics trajectories that look plausible while being quietly wrong. This is why the problem has persisted for decades despite the maturity of the surrounding computational infrastructure.

The historical reliance on inference rather than observation for hydrogen positions deserves some elaboration. X-ray crystallography detects electrons, not nuclei, and hydrogen atoms, with their single electron participating in the bond, scatter X-rays weakly relative to carbon, nitrogen, and oxygen. In small-molecule crystallography, where crystals are typically highly ordered and data can be collected to very high resolution, hydrogen positions can often be refined directly or placed with confidence using geometric constraints. In macromolecular crystallography, by contrast, protein crystals diffract to lower resolution, contain enormous amounts of solvent-disordered content, and rarely permit direct localization of hydrogens on ligands. Neutron crystallography can locate hydrogen nuclei explicitly because neutrons scatter from nuclei rather than electron clouds, but neutron sources are scarce, beamtime is limited, and growing large protein crystals suitable for neutron work is difficult. As a result, the tautomeric state of a ligand bound in a protein structure is usually assigned by software or by the crystallographer’s judgment, using rules of thumb about which heteroatoms are likely protonated at physiological pH. These assignments are then inherited by downstream users of the structure, often without re-examination.

The scale of the Cambridge Structural Database mining effort described in the study is worth appreciating in context. The CSD has accumulated small-molecule crystal structures for over a century, and its high-resolution entries preserve the geometric detail needed to distinguish, for example, a hydroxyl oxygen from a carbonyl oxygen, or an imidazole nitrogen bearing a proton from one that does not. By systematically extracting more than 1.1 million tautomeric states from these structures, the researchers effectively converted a repository built for crystal engineering and structural chemistry into a training corpus for machine learning. This kind of data repurposing has become an increasingly productive strategy in computational chemistry, where experimental archives originally collected for one purpose turn out to contain the labels needed to train models for another. The orders-of-magnitude increase over prior tautomer datasets, which typically contained a few hundred experimentally characterized examples in solution, changes what is statistically learnable: rare tautomeric motifs, substituent effects, and heteroaromatic patterns that would be absent or underrepresented in smaller collections can now inform predictions.

The choice of a graph neural network operating on two-dimensional molecular input is also consequential. Two-dimensional representations encode the connectivity of atoms and bonds but not the three-dimensional conformation, meaning the model must learn tautomer preferences from topology and chemical context alone. This is a deliberate trade-off. Requiring a 3D structure would mean generating conformations first, introducing an additional source of error and computational cost, and would tie the model’s accuracy to the quality of conformer generation. By working directly from the 2D form, the tool can be applied at the earliest stages of library triage, before any geometry has been computed, which is exactly where tautomer decisions have the greatest leverage since they propagate into every subsequent calculation.

The finding that roughly 2.5 percent of ligands in the PDBbind set may carry incorrect tautomer assignments warrants careful interpretation. PDBbind is a curated collection of protein-ligand complexes drawn from the Protein Data Bank, widely used as a benchmark for docking and binding affinity prediction. If a few percent of its ligands have misassigned protonation or tautomeric states, then benchmark results that appear to reflect method performance may in part reflect data noise. The researchers are careful to note that the experimental structures themselves are not necessarily wrong; rather, the chemical annotation layered onto them may need revision. This distinction matters because the electron density in a crystal structure constrains the heavy-atom framework, and a reassigned tautomer often fits the same density equally well while offering more chemically reasonable hydrogen bonding. The example highlighted in the release, where reassignment enables additional hydrogen bonds with nearby protein residues, illustrates the practical payoff: a corrected tautomer can transform an apparently suboptimal binding pose into a coherent one.

Hydrogen bonding is particularly sensitive to tautomeric state because it depends on which atoms carry hydrogens and which act as acceptors. A pyridone and a hydroxypyridine, for instance, present entirely different donor-acceptor patterns to a binding site despite sharing the same atomic composition. In structure-based drug design, where hydrogen bonds are often the dominant specificity-determining interactions, an incorrect tautomer can lead medicinal chemists to draw the wrong conclusions about why a ligand binds or how to improve it. Free energy perturbation calculations and other rigorous simulation methods are similarly affected, since they require a fully specified molecular topology including every hydrogen. An error introduced at the tautomer assignment stage cannot be corrected later by more sampling or better force fields; it is baked into the simulation from the start.

The throughput figures reported for the released tool give a sense of its practical reach. Processing roughly 4.6 million compounds in 3.2 hours on a single GPU-enabled node corresponds to on the order of hundreds of molecules per second, fast enough to annotate entire corporate or public screening libraries overnight. This matters because tautomer enumeration is combinatorial: a molecule with several heteroatoms and exchangeable protons can generate many candidate tautomers, and exhaustively evaluating each with quantum mechanics is infeasible at library scale. A learned model that ranks candidate tautomers by predicted stability provides a practical filter, allowing expensive quantum or experimental follow-up to be reserved for the most plausible candidates.

It is also worth situating this work within the broader movement toward physics-aware machine learning in chemistry. Rather than attempting to predict binding affinities or biological activity end-to-end, the model targets a discrete, well-defined chemical inference problem that sits upstream of many downstream tasks. Problems of this kind, where the answer is a structured chemical choice rather than a continuous property, can be validated against experimental ground truth and integrated into existing workflows as drop-in improvements. The open-source release of the tool as Tautomer-Predictor follows a pattern increasingly common in computational chemistry, where reproducibility and community adoption are treated as integral to the research contribution rather than afterthoughts.

Certain limitations naturally accompany this approach. The training data derive from small-molecule crystals, which reflect tautomeric preferences in the solid state, where packing forces and crystal environments can shift equilibria relative to solution or to the protein binding pocket. The study’s authors and the field more generally recognize that tautomer populations are environment-dependent: a form favored in a crystal may not dominate in water, and the protein microenvironment, with its own hydrogen bond donors, acceptors, and dielectric properties, can stabilize forms that are minor in bulk solution. Models trained on crystallographic data therefore capture one informative slice of tautomer stability, and their predictions are best understood as strong priors to be refined in context rather than final answers. Nevertheless, the availability of a fast, data-driven prior represents a meaningful advance over rule-based tautomer perception, which has historically relied on hardcoded chemical heuristics that vary between software packages and can disagree with one another.

The funding context, an NIH grant supporting methods development for molecular modeling, reflects the recognized importance of infrastructure-level improvements in drug discovery research. Individual breakthroughs in target identification or clinical validation attract considerable attention, but cumulative progress often depends on quietly fixing foundational layers of the computational pipeline. Tautomer assignment is one such layer. As machine learning methods continue to permeate molecular design, the quality of the chemical representations fed into them becomes a limiting factor, and efforts to clean and standardize those representations, whether through better data, better models, or better experimental characterization, compound across the entire discovery enterprise. The work from NYU exemplifies this genre of contribution: not a new drug or a new target, but a repair to the substrate on which countless downstream calculations rest.

Subject of Research: Scientists use AI to find stable forms of molecules for drug discovery

Article Title: Scientists use AI to find stable forms of molecules for drug discovery

Article References: Scientists use AI to find stable forms of molecules for drug discovery. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: tautomers, deep learning, drug discovery, chemical science, hydrogen positions, graph neural network, Cambridge Structural Database, Protein Data Bank, molecular modeling, NYU, Scientists, find