Graph neural networks for biochemistry that incorporate substructure

Abstract

The success of neural networks has been considerable interest in applying neural networks to biological and chemical data. Often, this data is graph-like in nature: chemical structures, biological signalling networks, and cytoskeletons all take the form of graphs. Most attempts to apply neural networks to this data have used generic graph neural network architectures such as message-passing neural networks. However, these networks are strictly limited in their expressivity. Moreover, there is no obvious way to include our knowledge of important biochemical structures into the network.