DeepO-GlyThr: an interpretable deep learning framework for predicting O-linked threonine glycosites in human proteins

O-linked glycosylation is a widespread protein post-translational modification with site selection strongly shaped by local sequence context. Threonine is a key acceptor residue, yet computational methods for predicting human O-linked threonine glycosites remain limited and often lack interpretability. Here, we present DeepO-GlyThr, an interpretable deep learning framework for O-linked threonine glycosite prediction in human proteins. DeepO-GlyThr integrates sequence features and employs CNN, BiGRU, and attention modules to learn discriminative patterns. On the independent test set, DeepO-GlyThr achieved a sensitivity of 0.885, an accuracy of 0.900, an F1-score of 0.898, and an MCC of 0.800, outperforming existing methods under the default classification threshold. Interpretability analyses showed that predictions were mainly driven by the local physicochemical environment and positional context around the central threonine. Volume, hydrophobicity, polarity, and positional features contributed most strongly. Attention, Integrated Gradients, and mutagenesis analyses further revealed context-dependent sequence patterns in local flanking regions. In summary, DeepO-GlyThr provides an accurate and interpretable framework for predicting O-linked threonine glycosites. Additionally, a user-friendly web server has been developed to facilitate its use, available at http://i-health.info/DeepO-GlyThr.