Neural Network Learns to Predict How Metals Bind to Molecules
Researchers at Lomonosov Moscow State University (MSU) have developed a unified model that estimates the stability of complexes formed by different metals, including metals not represented in the training data.
The study was published in The Journal of Chemical Physics. The researchers compiled a database of stability constants from NIST46 and scientific papers, accounting for temperature and the ionic strength of the solution. They trained a graph neural network on this dataset, representing each molecule as a graph of atoms and bonds. When a metal was entirely excluded from training, prediction accuracy declined but remained sufficient for qualitative assessment. The model also estimates the uncertainty of its predictions.
Artem Mitrofanov, a senior researcher at MSU’s Artificial Intelligence Center, explained that experimental data are scarce for some actinides. The researchers therefore combined data on different metals in a single model to capture shared patterns in how they interact with organic molecules.
The study is limited to aqueous solutions.
