Explainable AI with Single-Residue Resolution

In the life sciences, AI and machine learning have significantly advanced protein analysis, yet their full potential is held back by the black-box nature of traditional models, which obscures how a prediction is made. Explainable AI (eXAI) closes this gap: it not only forecasts an outcome but also surfaces the reasons for it, giving detailed insight into protein function at the amino acid level. In AAanalysis this matters twice over, because CPP features are already interpretable by construction, so an explanation resolves to a concrete Part × Split × Scale signal rather than an anonymous input.

Provided by

In AAanalysis this is TreeModel (predictions and global feature importance), ShapModel (SHAP-based local, per-sequence feature impact), and CPPPlot (the visualisations). See the API reference for signatures and the tutorials for hands-on use.

What is explainable AI?

Explainable AI (eXAI) transforms opaque AI models into transparent systems, letting human experts grasp the rationale behind specific predictions. This is particularly pivotal in the life sciences, where understanding each amino acid’s role can inform drug discovery and disease treatment. Explanations come at two levels: global importance ranks which features matter across a whole dataset (TreeModel), while local attribution explains one specific sequence, residue by residue (ShapModel).

Combining CPP with SHAP

To explain machine-learning predictions for individual proteins at single-residue resolution, AAanalysis combines CPP with the explainable-AI framework SHAP. Because every CPP feature maps back to a specific sequence part and physicochemical scale, a SHAP value can be projected onto the exact residues it describes. AAanalysis exposes this through CPPPlot, which offers four complementary views: a group-level feature map, a ranked feature profile, and per-sequence CPP-SHAP plots that colour the sequence by each residue’s contribution.

Overview of explainable AI with single-residue resolution by combining CPP with SHAP, from [Breimann25].