Table 1.

Selecting an architecture for a model predicting protein–glycan interactions. For the task of predicting protein–glycan interactions, we trained deep learning models with varying architectures to identify a suitable model for this study. In this table, we note the differences in the various models in three modules, the arm analyzing protein sequences, the arm analyzing glycan sequences, and the downstream module combining protein and glycan information for binding prediction. Mean values from five independent training runs (Table S3, Supporting Information, n = 5) are provided for the mean squared error (MSE) and mean absolute error (MAE) on a separate test set. For each metric, the best value is shown in bold

Protein arm	Glycan arm	Combined head	MSE	MAE
RNN	SweetNet	Fully connected + Multi‐sample dropout + Sigmoid	0.9925	0.501
ESM‐1b	SweetNet	Fully connected + Multi‐sample dropout + Sigmoid	0.7475	0.4276
ESM‐1b + fine‐tune	SweetNet	Fully connected + Sigmoid	0.7375	0.4238
ESM‐1b + fine‐tune	SweetNet	Fully connected + Multi‐sample dropout	0.7415	0.4301
ESM‐1b + fine‐tune	SweetNet	Fully connected + Multi‐sample dropout + Sigmoid	0.7283	0.4137