Researchers at New York University have developed an AI model that predicts the positions of hydrogen atoms in drug-like molecules. This innovation, published in Chemical Science, tackles a key challenge in molecular design and drug discovery.
Many drug-like molecules can exist in multiple forms called tautomers. These forms differ by the position of a single hydrogen atom, which can significantly alter how the molecule interacts with protein targets. Accurate identification of the correct tautomer is key for effective molecular modeling and structure-based drug discovery.
A Needle in a Moving Haystack
Determining the correct tautomer has been likened to finding a needle in a moving haystack. Experimental data on tautomer structures is scarce, particularly in resources like the Protein Data Bank (PDB), where the location of hydrogen atoms is often unavailable.
Traditional methods, such as using quantum mechanics or machine learning, have limitations. Quantum mechanics is computationally expensive, while machine learning suffers from small datasets of experimentally characterized tautomers.
Untapped Data, Powerful AI
The NYU team, led by Professor Yingkai Zhang, turned to the Cambridge Structural Database (CSD), the world’s largest repository of experimental crystal structures. They extracted a dataset of over 1.1 million tautomeric states, significantly larger than existing datasets.
Read Also: FFS Campaign Launches to Address Dialysis Catheter Fibrin Sheath
This vast dataset was used to train a graph neural network, a type of AI adept at finding patterns in connected data. The model learned to predict stable tautomers directly from 2D molecular forms, eliminating the need for 3D structures or quantum-mechanical calculations.
The AI-predicted tautomers showed improved hydrogen bonding patterns, suggesting potential revisions to existing chemical representations.
For instance, in one molecule, the model predicted a tautomer with a differently positioned hydrogen atom. This change led to additional hydrogen bonding with nearby protein residues, highlighting the impact of seemingly small molecular alterations.
Faster, More Accurate Drug Discovery
The researchers have made Tautomer-Predictor publicly available as an open-source tool. Its ability to rapidly analyze large molecular libraries, processing 4.6 million compounds in just 3.2 hours on a single GPU-enabled node, promises to accelerate drug discovery efforts significantly.
