Skip to main content
. Author manuscript; available in PMC: 2022 Jun 20.
Published in final edited form as: Cell Rep. 2021 Jun 15;35(11):109251. doi: 10.1016/j.celrep.2021.109251

Table 3.

Developing a model predicting viral glycan-binding behavior

Model Train MSE Test MSE
Fully connected 0.8508 0.8753
Language model (SweetTalk) 0.8253 0.8726
Graph model (SweetNet) 0.7455a 0.7352a

Models consisted of a recurrent neural network analyzing the protein sequences of viral hemagglutinin as well as either a fully connected neural network using the counts of mono-, di-, and trisaccharides as input (“Fully connected”); a SweetTalk-based glycan language model; or a SweetNet-based GCNN. All models were trained to predict Z score transformed glycan binding of hemagglutinin from various influenza virus strains. Average MSEs from five independent training runs (N = 5), from both the training and independent test set, are shown here.

a

The superior value for each metric.