Computationally efficient graph neural/convolutional network captures protein geometry to predict affinity and developability.

Abstract

There are thousands, if not hundreds of thousands, of known proteins but a significantly lower number of unique protein folds. As a result, proteins often have conserved geometric patterns, including the ways amino acids cluster together. Delaunay tessellations have been successfully used in multiple methods involving protein structures, including applications to elastic network models to predict hinges and to predict conformational heterogeneity. In this work, we use Delaunay tessellations and alpha shapes to capture these conserved geometric patterns to define edges that are better suited to model protein packing than other established ways to form the edges for Graph Neural Network (GNN) architectures, including Graph Convolutional Networks (GCN). We test the resulting GCN models on problems including antibody affinity and developability. Compared to GNN architectures based on k-nearest neighbors or a fixed cutoff radius, models employing Delaunay tessellation can capture residue-specific interactions at a lower computational cost.