EGAT: A Graph Attention Network for Your Chemical Property Prediction Needs

Determining the optimal design of molecules for industrial applications hinges upon accurately elucidating their physiochemical, spectral, and reactive characteristics. Historically, such properties have been found either using expansive experimental studies or exhaustive quantum chemistry tools based on the highest level of theory afforded. However recently, several groups have popularized the idea of directly predicting chemical properties based only on graphical representations of molecules and reactions along with a broad enough dataset. In this talk, we revisit this task using a set of 1) several newly developed edge-featured graph attention architectures colloquially known as EGAT and 2) large datasets such as the Reaction Graph Depth 1 (RGD1) transition state dataset. Such architectures achieve state-of-the-art withheld test set predictions on a variety of molecular and reaction property prediction tasks. Further sets of case-studies on other contemporary architectures indicated that such poor out-of-sample performance is a common trait. Thus, we conclude that standard graph architectures can achieve results comparable to the irreducible error of current reaction datasets, but out-of-sample performance remains poor.