Protein O-linked threonine glycosylation (OTG) is a crucial post-translational modification in eukaryotic species, playing a vital role in diverse biological processes. In humans, dysregulation of OTG has been associated with serious diseases, including cancer and neurological disorders. However, experimental detection of OTG sites remains costly and labor-intensive, underscoring the need for effective computational approaches. In this study, we introduce DP-OTG, a feature-free deep learning model for the accurate prediction of human OTG sites. Unlike existing tools that rely heavily on handcrafted features or large-scale pretrained language models, DP-OTG employs a hybrid architecture combining multi-kernel convolutional neural networks, bidirectional long short-term memory, and a trainable embedding layer to automatically learn sequence patterns directly from raw protein sequences. This end-to-end framework captures both local and long-range sequence dependencies without the need for manual feature engineering. Extensive evaluations using 10-fold cross-validation and independent testing demonstrate that DP-OTG achieves superior predictive performance, with an accuracy of 88.8% and an Matthew's Correlation Coefficient (MCC) of 0.776 on the balanced test set, and an accuracy of 89.3% and an MCC of 0.661 on the imbalanced test set, outperforming several state-of-the-art predictors. In addition, to comprehensively assess the discriminative power and generalization ability of DP-OTG in predicting human OTG sites, we employed t-distributed stochastic neighbor embedding to visualize the feature representations before and after training. These results underscore the effectiveness of DP-OTG in extracting robust features for accurate OTG site prediction, even under challenging data distributions. Our findings highlight DP-OTG as a robust, efficient, and scalable tool for human OTG site prediction. All the code and resources related to this study have been made freely accessible at: https://github.com/nuinvtnu/DP-OTG/.
Tran et al. (Sat,) studied this question.