Neural XAI methods (deep learning)

[breimanntools]

Problem

Protein language models (PLMs) are powerful but not interpretable.

Goal

Map deep learning representations to CPP-based explanations.

Tasks

• Implement Integrated Gradients, LRP, DeepLIFT (captum)
• Extract residue-level importance from PLMs
• Map embeddings → CPP feature space (scale projection)
• Apply TCAV on embeddings
• Compare DL attribution vs CPP attribution

How this improves AAanalysis

• Makes PLMs usable in interpretable workflows
• Connects embedding space → physicochemical meaning
• Enables hybrid models (DL + CPP)

Acceptance criteria

• PLM predictions explainable via CPP features