Deep learning has shown promise in mass spectrometry-based proteomics and is emerging in glycoproteomics. While deep learning models effectively predict peptide fragment mass spectra, they struggle with the non-linearity of glycan structures in intact glycopeptides. This paper introduces DeepGlyco, a deep learning approach for predicting fragment spectra of intact glycopeptides. DeepGlyco utilizes tree-structured long-short term memory networks (LSTMs) for glycan processing and graph neural networks to model glycan fragmentation pathways, enhancing model explainability and isomer differentiation. The study demonstrates the use of predicted spectral libraries for data-independent acquisition (DIA) glycoproteomics, supplementing library completeness. DeepGlyco offers a valuable resource for glycoproteomics research.

Yi Yang, Qun Fang