This page condenses the manuscript into a web-first format, covering benchmark construction, feature engineering, model architecture, and evaluation strategy.
DeepO-GlyThr follows the benchmark used in the original O-GlyThr study. Positive and negative samples are 41-residue windows centered on threonine, with 20 residues on each side. The final benchmark contains 1,078 positive and 1,078 negative samples, split into training and independent test sets.
Each residue is represented by 27 channels: 20-dimensional amino-acid one-hot encoding, six z-score normalized physicochemical descriptors, and one center-prior channel that emphasizes the candidate Thr neighborhood.
The architecture combines a CNN stem, two residual convolution blocks, a two-layer BiGRU, additive attention, and fully connected classification layers. This design captures local motifs, contextual dependencies, and position-specific importance.
Performance is reported with sensitivity, specificity, accuracy, F1-score, MCC, ROC-AUC, and PR-AUC. The model is intended to rank candidate Thr-centered windows while keeping interpretation accessible for downstream biological inspection.
DeepO-GlyThr: an interpretable sequence-based deep learning framework for human O-linked threonine glycosites prediction.
Source code and data are availability in the github.