Feature Importance — Integrated Gradients

CAPTUM

Per-residue attribution analysis using Axiomatic IG. Implementation: Captum with manual Riemann fallback.

Integrated Gradients

IG_i(x) = (x_i − x′_i) · ∫₀¹ ∂F(x′ + α(x−x′))/∂x_i dα

Approximated via Riemann sum with n_steps=50. Baseline x′ = zero vector (uninformative). Axiom completeness: Σ_i IG_i = F(x) − F(x′).

Implementation Details

LibraryCaptum 0.6+ (PyTorch-native)

FallbackManual Riemann sum (n_steps=50)

BaselineZero vector x′ = 0₁₂₈₀

TargetEach predicted GO term independently

Attribution Shape[L, 1280] → aggregated to [L] by L2 norm

Proteins analyzed—

GO terms per protein—

Axioms Satisfied

✓ Completeness

Σ_i IG_i(x) = F(x) − F(x′). The attributions sum to the difference in model output between the input and baseline.

✓ Sensitivity

If F(x) ≠ F(x′) for some feature i where x_i ≠ x′_i, then IG_i ≠ 0.

✓ Implementation Invariance

IG is independent of internal network architecture; computed on the input embedding, not activations.

✓ Linearity

IG is linear in the network output: IG(αF₁ + βF₂) = αIG(F₁) + βIG(F₂).

Full residue heatmap visualization

Interactive per-residue attribution heatmaps are available in the XAI Dashboard — select any protein and GO term to view its full attribution map and top-20 influential residues.