Interpretation dashboard

Ablation and Integrated Gradients views of DeepO-GlyThr

Attention is available per submitted window. This page complements it with global evidence from feature ablation and manuscript-level Integrated Gradients analysis.

Feature ablation and importance analysis

Performance degradation after removing feature groups

The main result is that removing physicochemical channels causes the most severe collapse of positive-site recognition.

Ablation figure
Attention patterns and position-specific residue analysis

DeepO-GlyThr captures a position-dependent pattern centered on the target threonine residue

The model appears to rely primarily on local sequence context in the central regions, whereas negative samples exhibit more dispersed and distal dependency patterns. Residue-level analysis further suggests distinct position-residue preferences between positive and negative samples.

Integrated Gradients figure
Ablation

Feature ablation

To evaluate the contribution of different input features, we conducted feature ablation experiments on the independent test set. Compare the full model against variants without center prior, one-hot encoding, or physicochemical channels.

Absolute performance metrics
Performance drop relative to the full model
Model AUC AUPRC Acc F1 MCC Sn Sp
Full model 0.921 0.93 0.9 0.898 0.8 0.885 0.915
w/o center prior 0.839 0.873 0.57 0.252 0.266 0.145 0.995
w/o one-hot 0.755 0.74 0.51 0.039 0.101 0.02 1.0
w/o physicochemical 0.472 0.474 0.5 0.0 0.0 0.0 1.0
Position-wise IG analysis

DeepO-GlyThr captures a position-dependent pattern centered on the target threonine residue.

The model appears to rely primarily on local sequence context in the central regions, whereas negative samples exhibit more dispersed and distal dependency patterns.

Relative feature-importance tiers
Position hotspots around the central Thr

Highest-priority IG signals

  • Tier 5
    Volume
    Strongest contribution in the IG analysis.
  • Tier 5
    Center prior
    Positional context around the candidate Thr remains essential.
  • Tier 4
    Hydrophobicity
    Consistently emphasized among physicochemical descriptors.
  • Tier 4
    Flexibility
    Supports local context sensitivity in the window.

Interpretation notes

  • Physicochemical context is consistently more important than most identity channels.
  • Upstream and downstream local neighborhoods around the central Thr dominate attribution.
  • The overlap between attention and IG strengthens the biological plausibility of the learned context.