NYU's AI scans 4.6 million compounds in hours: where the hydrogen atom goes, the AI decides

## NYU's AI scans 4.6 million compounds in hours A New York University team has published a paper in Chemical Science: they trained an AI to learn the stability-related chemical patterns in drug molecules and accurately predict where hydrogen atoms sit, tackling the long-standing problem of determining tautomers. **Key points** - **The problem itself**: Many drug molecules have tautomers: a hydrogen atom moves from one site to another, and the bond types change with it. Different hydrogen positions can mean entirely different ways of binding to a protein target; but hydrogen positions are usually missing from the PDB (Protein Data Bank) and have to be inferred. - **The AI's solution**: A graph neural network predicts the stable tautomer directly from the 2D molecular formula, with no need for 3D structures or quantum mechanical calculations. - **Scale and error correction**: On a single GPU node, it processed about 4.6 million compounds in 3.2 hours; among 5,075 PDB ligands it found 126 cases (about 2.5%) in which the original hydrogen positions were "very likely mislabeled", and the hydrogen-bonding patterns made more sense after correction. - **Open source**: The method has been released as an open-source tool, Tautomer-Predictor, which can quickly analyze very large molecular libraries. > Source: [read the original](https://phys.org/news/2026-09-ai-scans-million-compounds-hours.html). This text is a summary for reference only.