AI Model Converts 3D Molecular Structures into Chemical Formulas with Record Accuracy
Researchers from the AIRI Institute, HSE University, Lomonosov Moscow State University, Tomsk State University, and Ural Federal University have developed Uni-Bond, an AI algorithm that converts the three-dimensional coordinates of atoms into a two-dimensional structural formula of a molecule. The team presented the technology at the International Conference on Machine Learning (ICML) in Seoul.
Molecular simulations typically rely on three-dimensional representations of molecules, while property prediction methods more often use molecular graphs that focus on the types of chemical bonds between atoms rather than their precise spatial positions. Uni-Bond is designed to bridge these two approaches more reliably. Unlike conventional software, which infers chemical bonds primarily from the distances between atoms, Uni-Bond analyzes the geometry of the entire molecule. As a result, it achieves 99.53% accuracy, reducing errors by a factor of 20 compared with previous methods.
The researchers evaluated the algorithm using water clusters, in which molecules are packed closely together but do not form chemical bonds with one another. Uni-Bond correctly identified bonds within individual water molecules while avoiding false intermolecular connections – a scenario in which earlier methods frequently produced incorrect results.
The neural network can distinguish five types of interactions: no bond, single, double, triple, and aromatic bonds, such as those found in a benzene ring. The researchers plan to extend the model to recognize charged particles and free radicals in future versions.
