DeepOTG: An effective deep learning framework for identifying human protein O-linked threonine glycosylation sites...

GalNAc transferase (GalNAc-T)-initiated O-linked N-acetylgalactosamine (GalNAc) glycosylation at threonine residues, referred to as O-linked threonine glycosylation (OTG), is a prevalent protein post-translational modification involved in critical biological processes and associated with diverse diseases. However, the distinct yet overlapping substrate specificities of GalNAc-T isoenzymes pose a significant challenge in experimentally validating and computationally predicting OTG sites. Here, we introduce DeepOTG, a dual-branch deep learning framework that combines global semantic context from the protein language model ESM-2 with local evolutionary patterns derived from the position-specific scoring matrix (PSSM) using a multi-scale CNN-BiGRU architecture. To effectively fuse complementary information while suppressing redundancy, we design an attention-based fusion module incorporating multi-head self-attention and a variational information bottleneck. On benchmark datasets, DeepOTG achieves an MCC of 0.807 on the balanced test set and 0.737 on the imbalanced test set, outperforming state-of-the-art predictors. Furthermore, interpretability analyses demonstrate that the learned residue importance is consistent with reported substrate preferences and structural observations. Thus, the framework provides a valuable tool for the discovery of OTG sites and further biological research.