New Artificial Intelligence Tool Predicts Hydrogen Atom Positions to Accelerate Drug Discovery

Researchers at New York University have developed a novel artificial intelligence tool designed to predict the precise locations of hydrogen atoms within drug-like molecules. Published in the journal Chemical Science, the research tackles a long-standing obstacle in molecular design: rapidly determining the stable configuration of chemical structures that share identical molecular formulas yet shift into different forms known as tautomers.

Many molecules relevant to pharmacology can exist as multiple tautomers, where a hydrogen atom shifts location alongside a corresponding alteration in bonding patterns. Although this structural shift may appear minor, different tautomers can significantly change how a molecule interacts with a specific protein target. Consequently, accurate tautomer assignment remains vital for effective molecular modeling and structure-based drug discovery. However, identifying the correct tautomer has historically resembled finding a needle in a moving haystack.

The difficulty largely stems from a scarcity of experimental data characterizing tautomer structures. Widely utilized repositories, such as the Protein Data Bank, typically lack the precise location of hydrogen atoms because macromolecular X-ray crystallography resolutions are generally insufficient for dependable hydrogen mapping. As a result, scientists frequently have to infer these crucial positions. Alternative methods also present limitations. While quantum mechanics can assess tautomers, the approach is computationally demanding and expensive for screening extensive compound libraries. Meanwhile, traditional machine learning techniques have been restricted by small datasets containing only a few hundred experimentally characterized tautomers in solution.

To overcome these data bottlenecks, the research team turned to the Cambridge Structural Database, the world’s largest repository of experimental crystal structures, which frequently displays resolved hydrogen positions in high-resolution small-molecule crystals. Lead author Xiaolin Pan, a postdoctoral researcher in the lab of senior author Yingkai Zhang, systematically mined this database. This process generated a dataset containing more than 1.1 million tautomeric states—orders of magnitude larger than previous experimental collections.

Using this expansive dataset, the investigators trained a graph neural network, a deep learning artificial intelligence model capable of identifying patterns across connected data points. This trained system predicts stable tautomers directly from two-dimensional molecular forms, entirely bypassing the need for complex three-dimensional structures or intensive quantum-mechanical calculations.

When the NYU investigators tested their model on 5,075 biomolecular complexes from the Protein Data Bank featuring multiple potential tautomeric states, the tool identified approximately 2.5 percent—totaling 126 instances—where the originally assigned ligand tautomer was likely incorrect. In each of these instances, the artificial intelligence model reassigned an alternative stable tautomer that demonstrated superior hydrogen bonding patterns. Senior author Yingkai Zhang noted that these reassignments generally yielded more chemically reasonable interactions, suggesting that earlier chemical representations may warrant revision rather than implying that the underlying protein structures themselves are flawed.

Improving tautomer identification holds significant promise for subsequent computer simulations. Molecular dynamics simulations track how drug candidates and their protein targets interact and move over time. These complex calculations rely heavily on correct hydrogen positions and chemical bonding assignments, meaning that utilizing an inaccurate tautomer could alter predicted cellular interactions and dynamics.

The research team has made their methodology publicly available as an open-source utility named Tautomer-Predictor. Designed to evaluate massive molecular libraries with high speed, the tool successfully processed approximately 4.6 million compounds in just over three hours using a single graphics processing unit node. Funding for the study was provided by the National Institutes of Health.

Key Features of the Tautomer-Predictor Tool

• Trained on a dataset of over 1.1 million tautomeric states mined from high-resolution crystal structures.
• Employs a graph neural network to predict stable tautomers directly from 2D molecular forms.
• Operates without requiring resource-intensive 3D structures or quantum-mechanical calculations.
• Capable of rapidly screening massive molecular libraries, processing millions of compounds in hours.

Source: News-Medical.net