Abstract
Graph neural networks were used to model protein isoelectric points. Predictions contained
markedly fewer outliers (predicted with errors > 0.5 pH units) compared to tools published in the
literature, despite slightly higher root-mean-squared errors. This result was reproduced for graph
convolutional and graph isomorphism networks when node features used only one-hot encoding
of amino acid sequences. Graph isomorphism networks could also produce similar predictive
powers when employing physical descriptors of the amino acids, either alone or in addition to the
one-hot encoded features.
Supplementary weblinks
Title
Code for GNN calculations
Description
The skeleton of the (Spektral/Keras/Tensorflow) code used to produce this work.
Actions
View