Researchers at New York University have created an artificial intelligence (AI) model that can learn chemical patterns related to stability in drug-like molecules and accurately predict the positions of hydrogen atoms. This advancement, published in the journal Chemical Science, aims to solve the problem of identifying the correct stable form of molecules that share the same molecular formula but can exist in different forms called tautomers. Tautomers are structural isomers that differ only in the position of a hydrogen atom, and these differences can significantly affect how a molecule interacts with proteins. Accurately assigning the correct tautomer is crucial for drug development and molecular modeling. Determining the correct tautomer is difficult because of a lack of experimental data. The Protein Data Bank, a major source of 3D structures of biological molecules, typically does not include the positions of hydrogen atoms, which are essential for distinguishing between tautomers. Experimental techniques like X-ray crystallography, commonly used to determine molecular structures, often lack the precision needed to locate hydrogen atoms reliably. Other approaches also have limitations. Quantum mechanical calculations, while accurate, are computationally intensive and not practical for analyzing large molecular libraries. Machine learning models, on the other hand, are limited by the relatively small number of experimentally characterized tautomers available in solution—often just a few hundred molecules. To overcome these challenges, the researchers used the Cambridge Structural Database, which contains high-resolution X-ray crystal structures of small molecules and includes hydrogen atom positions. They mined this database to create a dataset with over 1.1 million tautomeric states. Using this dataset, they trained a graph neural network to predict stable tautomers directly from 2D molecular structures, without needing 3D models or quantum calculations. When the model was tested on 5,075 ligands from the Protein Data Bank—molecules that bind to proteins—it identified 126 cases (about 2.5%) where the previously assigned tautomer was likely incorrect. In each of these cases, the model suggested a more stable tautomer that improved hydrogen bonding with the surrounding protein. These changes highlight how even minor adjustments in molecular structure can significantly influence interactions with proteins. Accurately identifying the correct tautomer can improve drug discovery by helping scientists understand how a molecule fits into a protein's binding site. It can also enhance computer simulations, which rely on correct hydrogen positions and bonding patterns. Using the wrong tautomer could lead to inaccurate predictions of molecular behavior. The researchers have made their method available as an open-source tool called Tautomer-Predictor, which can quickly analyze large molecular libraries. In a test, it processed over 4.6 million compounds in just 3.2 hours using a single GPU-enabled computer.