Comprehensive analysis of lectin-glycan interactions reveals determinants of lectin specificity
Fig 3
Lectin binding site features can be used to predict the identity of bound glycans.
Random forest models trained for each of the 15 glycans have strong recall performance while predicting whether interactions contain the respective glycan based on the interaction features alone. The models are predictive of glycan identity even when trained only on lectins with less than 50% sequence identity, outperforming identical classifiers trained on data with shuffled labels. Split violin plots show the recall (left-hand distribution and left y-axis) and precision (right-hand distribution and right y-axis) of ligand-specific random forest models measured during leave-one-out cross-validation. The pairs of notched boxplots for each glycan show the performance of classifiers trained on data with shuffled labels, where again the left-hand boxplots depict recall and the right-hand boxplots depict precision. Glycan symbols follow the SNFG system.