Abstrait

3D Equivariant Graph Neural Networks for Drug Discovery

Max Welling

In this talk I will first review a very useful machine learning tool for pharmaceutical applications: the Graph neural Network (GNN). A GNN analyses a signal on graph structured data, such as molecules or knowledge graphs. We then define the concept of equivariance, an organizing principle that embues the neural network with the right transformation properties under symmetry transformations. For instance, a molecule has the same properties whether we rotate it in space or not. We will finally apply the new 3D equivariant graph neural network to molecular data and show that it can successfully predict properties that are relevant for drug screening. If time permits we will briefly describe further relevant generalizations of this technnology.