Abstract: The Surprising Effectiveness of Topology in the Chemical Sciences

Screening large spaces of billions of molecules to identify candidates for applications requires computationally efficient molecular property prediction. In this talk, we show that graph-topological descriptors coupled with linear models are fast, accurate, and interpretable. Specifically, such models are capable of exceeding DFT accuracy relative to experiment in energy prediction tasks. Further, such models are competitive with more complex models fit on more expensive features such as DFT geometries, electronic structures, and experimentally measured properties. Finally, we show that by exploiting the hierarchical structure of molecular graph fragments we are able to explain model predictions and show that such explanations agree with physical theory and experiment. LA-UR-22-28451