LIME Analysis
NOT IMPLEMENTEDLIME (Local Interpretable Model-agnostic Explanations) is not in the PaTz codebase. Local explanations are provided by Integrated Gradients + MC Dropout uncertainty.
LIME requires perturbation-based sampling (10–50k perturbed sequences per protein) — computationally prohibitive for variable-length protein sequences on a 6 GB GPU. The per-protein IG attributions in results/xai_results.json already provide local (per-protein, per-residue) explanations with better axiomatic guarantees.
For a single protein, generate N=1,000–5,000 perturbed versions by randomly masking residue windows.
Run each perturbation through PaTz to get GO term prediction probability changes.
Fit a sparse linear model (Ridge) on the perturbation neighbourhood → coefficients are LIME weights.
Identify residue windows with highest positive/negative weight for a target GO term.
Per-residue importance map for each protein and each GO prediction. Fast (~2s/protein). Captum + Riemann fallback.
→ OpenN=20 stochastic passes → predictive entropy H[p] per protein. Identifies unreliable predictions.
→ OpenPairformer 4-layer 8-head attention weights (residue–residue). Requires inference pass.
| Criterion | LIME | Integrated Gradients |
|---|---|---|
| Explanation type | Neighbourhood linear proxy | Path-integral attribution |
| Perturbation needed | ✓ Yes (1k–50k samples) | ✗ No (50 interpolation steps) |
| Handles variable length | ✗ Hard (masking ill-defined) | ✓ Yes (per-residue dim-mean) |
| GO-term specificity | ✓ Per-class linear fit | ✓ Per-class gradient |
| Completeness axiom | ✗ Approximate | ✓ Exact (Captum) |
| Speed per protein | ~4 min (RTX 3050) | ~2 sec |
| Used in PaTz | ✗ Not implemented | ✓ src/xai/ig.py |