Molecular graphs, which provide a mathematical representation of a molecule’s structural formula, are a natural choice to describe molecules in AI systems. Many reports have shown that Graph Neural Networks (GNNs) can predict molecular properties at the level of accuracy used to generate the training set at greatly reduced computational cost. However, the traditional definition of molecular graphs only contains information about atoms and bonds, and ignores information about the 3D orientation of each atom relative to the rest of the system. This omission prevents the prediction of conformational dynamics and changing bond distances. In this work, we redefine the molecular graph to include edges between atoms within a predefined Euclidian distance to capture each atom’s conformational geometry within the graph structure. With this new definition, we apply a dynamic GNN to predict changes in the bonding structure and relative neighborhood, including atom-to-atom distances in the molecular graph. We employ the ISO17 dataset, composed of molecular dynamics trajectories of isomers of C7O2H10 with 1 fs resolution computed at DFT-level accuracy, to predict changes in bond length and neighborhood environment. Using this graph definition, we show average mean absolute errors in predicted atom-to-atom distances around 0.01 Å.

Geometric learning of conformational dynamics using graph neural networks

ACS Spring Meeting 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.