Graph Neural Networks for Computational Chemistry

Abstract

Graph Neural Networks are behind many pharmacological breakthroughs due to their innate ability to learn structural properties of molecules and accelerate high-throughput screening for favorable characteristics that could serve as a treatment or cure to a disease. Much of the world’s natural data, such as social networks and molecules, can be represented in the form of graphs. However, advancements in graph-based problems like chemistry have been lacking because graphs are a form of non-Euclidean data, and encoding them into a format that is compatible with deep learning is considerably more challenging. This thesis seeks to understand and benchmark the techniques used to preserve the structure and properties of a graph in the encoded form. Specifically, characteristics of a graph that are distinct in the graph form should be distinct in the encoded form. Preserving both the expressiveness and the distinctness of the encoded graph is a challenging task that has received a lot of attention in geometric deep learning. This work evaluates and compares various graph neural network methods on a large public dataset to quantify the expressive power of more detailed graph neural networks that consider dihedrals and bond information, for example. It becomes evident that simply constructing homogeneous graphs of nodes and edges is insufficient.