Fig. 1: Overview of the deep learning model for glycopeptide fragment spectrum prediction.

a The input glycopeptide comprises a peptide sequence and a glycan tree, where monosaccharides are one-hot encoded. b The peptide sequence is processed by a linear LSTM network. c The glycan tree is traversed by a tree LSTM network. d, e The peptide features extracted by the linear and the glycan features by the tree LSTM are fused with each other. Then peptide features are processed by another linear LSTM network to predict the relative intensities of peptide b/y fragments. The glycan features are traversed by another tree LSTM network, updating the feature of each monosaccharide node in the glycan tree. f Features of potential cleavage sites are aggregated from the monosaccharide nodes that are lost or retained after the cleavage. Features of structure-specific glycan fragments are aggregated from the corresponding cleavages to predict the relative intensities of Y ions, where structural isomeric fragments are combined. g The peptide and glycan fragment ions are finally merged to form the output glycopeptide spectrum. The monosaccharide symbols are defined in Supplementary Table 1.