Do protein representations encode variant function? Diagnosing residue-level protein attributions via embedding-space gradients 1. The paper introduces IGLOO, a benchmarking framework to test whether residue-level attributions (via Integrated Gradients on protein language model embeddings) actually recover functionally important residues, rather than merely producing plausible-looking heatmaps. 2. IGLOO’s core idea is to create synthetic fitness landscapes where per-residue “ground-truth importance” is known, enabling direct scoring of attribution fidelity. It builds 9 protein–landscape settings using 3 proteins (GB1, T4 lysozyme, p53) × 3 landscape types (additive, epistatic, nonlinear “cliff”), and systematically sweeps label noise to decouple accuracy from interpretability. 3. A central finding: predictive accuracy (held-out Spearman correlation on fitness) weakly tracks attribution quality. Models can predict variant fitness well while producing residue attributions that are near-random relative to known important sites—an explicit demonstration of the prediction–explanation fallacy in pLM-based variant effect prediction. 4. The work diagnoses why this happens using cascading parameter randomization: in common “pool-first” architectures (attention pooling on per-residue embeddings before the MLP head), attributions barely change when randomizing most of the predictor head, but collapse when randomizing the pooling layer—showing the pooling module largely determines residue-level attribution patterns. 5. A practical architectural intervention improves attribution: moving learned computation before pooling (“pool-later” via per-residue processing + pooling) substantially strengthens the link between model learning and residue attributions. In one example (GB1 additive), the correlation between attribution quality and predictive accuracy increases dramatically (Pearson r from 0.23 to 0.83). 6. Backbone choice matters for interpretability: across seven pLMs from ESM-C, Profluent E1, and Ankh families, Ankh models yield the most consistent attribution quality across landscapes, and bigger is not always better (Ankh-base can outperform Ankh-large). 7. Intermediate layers can predict “well” but attribute poorly: for Ankh-base, attribution quality collapses when using intermediate hidden states (even when prediction Spearman can reach ~0.6), while final-layer representations give the best attribution. The paper argues this may relate to scaling/normalization properties of intermediate residual streams. 8. Ensembling improves stability of interpretability: averaging attributions across independently seeded predictor heads reduces attribution variance roughly at the expected 1/sqrt(K) rate, making residue-level interpretations more reliable without major added cost. 9. Real-protein case studies mirror synthetic conclusions. On FLIP2 α-amylase splits, the split with best predictive performance does not necessarily yield best recovery of annotated functional categories; pool-later heads outperform pool-first heads for attribution across splits. On TEM-1 β-lactamase datasets, top-attributed sites show partial agreement with known functional residues, and in silico checks (stability and binding proxies) match attribution sign in ~2/3 of tested mutant–ligand combinations. 💻Code: github.com/chemcognition-lab… 📜Paper: biorxiv.org/content/10.64898… #ProteinEngineering #ProteinLanguageModels #Interpretability #IntegratedGradients #DeepLearning #ComputationalBiology #MachineLearning #VariantEffectPrediction