(474l) Multi-Input E(n)-Graph Neural Networks for Learning Molecular Interaction Properties

Here we have developed a multi-input equivariant graph convolution-based deep learning model, named the MEGNN, for the prediction of chemical properties that result from the interaction of heterogeneous molecular structures. The model is an equivariant model which allows for the efficient learning of molecular properties by utilizing atomic spatial coordinates to make predictions that are agnostic to translations, rotations, and reflections in input molecules. We applied this model to predict the frictional properties between polymer brush monolayers with different functional end-groups attached to the ends of the brushes and achieved improvements to the prediction of these tribological properties. The MEGNN enables conditioned property prediction, meaning properties can be predicted qualified on their interaction with another molecule. In this talk I will discuss some of the technical details of this model and their relevance to chemistry.