Interpretable O-glycosite prediction

DeepO-GlyThr predicts human O-linked threonine glycosites from sequence context.

The server combines residue identity, physicochemical descriptors, positional priors, convolutional feature extraction, bidirectional recurrent modeling, and attention-based aggregation to score candidate Thr-centered windows and expose the positions that drive each prediction.

Launch Predictor Explore Interpretation
DeepO-GlyThr workflow
Sensitivity
0.885
Independent test set
Accuracy
0.900
Balanced default threshold
F1-score
0.898
Strong positive recovery
MCC
0.800
Robust overall performance
Why this server

An interpretable sequence-based deep learning framework for predicting human O-linked threonine glycosites

DeepO-GlyThr is designed for two complementary tasks: scanning full protein sequences for candidate O-linked threonine glycosites, and evaluating pre-extracted 41-residue windows centered on a target Thr.

Protein-level scanning

Paste full-length protein sequences and the server will automatically enumerate all valid Thr-centered windows with 20 upstream and 20 downstream residues.

Interpretable prediction

The deployed model returns probability estimates together with attention profiles, making individual predictions easier to inspect.

Paper-aligned evidence

Benchmark, ablation, attention, and attribution analyses are organized to mirror the manuscript and help readers connect results to the model design.

Benchmark

Independent-test comparison and representation learning

DeepO-GlyThr reaches the best balance of sensitivity, accuracy, F1-score, and MCC among representative O-linked threonine predictors, while progressively separating positive and negative windows in deeper layers.

Performance comparison
Interpretability

Attention, attribution, and perturbation converge on local Thr context

Ablation, Integrated Gradients, attention distributions, and mutagenesis analyses consistently indicate that the central threonine neighborhood and its physicochemical environment dominate prediction behavior.

View Interpretation Panels Read Methods
Usage

Recommended server workflow

Step 1

Submit FASTA input

Paste full-length proteins or 41-aa Thr-centered windows. The server validates amino-acid alphabet and extracts valid candidate windows automatically. Sequence length requires a minimum of 41 amino acids and has no maximum limit. For faster output of results, it is advised to submit fewer than 200 sequences at a time.

Step 2

Review candidate sites

Inspect probability scores, predicted labels, source proteins, and site positions. Export the full result table for downstream filtering or reporting.

Step 3

Inspect individual explanations

Open the site-detail page to examine attention weights across the 41-aa window and identify positions most responsible for the current score.