Explainability

The sequence model is designed to be interpretable at the residue level.

Methods

Saliency

• Gradient-based importance
• Identifies residues influencing predictions

Integrated Gradients

• Path-integrated attribution
• More stable and robust than raw gradients

Outputs

• Per-residue importance scores
• Sequence-level importance maps
• Global amino acid importance summaries

Interpretation

Important residues often align with:

• catalytic sites
• cofactor-binding regions
• redox-sensitive motifs
• structural hotspots

Practical use

• Generate hypotheses for mutagenesis
• Identify potential binding regions
• Validate model behavior against known biology

Explainability bridges prediction and mechanism.