Technical Summary: IGLOO – Diagnosing Residue-Level Protein Attributions via Embedding-Space Gradients

Problem Statement

Protein language models (pLMs) are increasingly used to predict the fitness of protein variants. A common workflow involves embedding variant sequences with a pLM and training a downstream predictor to estimate fitness. While these predictors often achieve high accuracy on benchmarks, a critical gap remains: high predictive accuracy does not guarantee that the model has learned biologically meaningful residue-level features.

Attribution methods, such as Integrated Gradients (IG), are frequently used to interpret these models by assigning contribution scores to individual residues. However, without ground truth for residue importance, it is unclear whether these attributions reflect the true biological drivers of function or are merely artifacts of the model architecture. This "prediction–explanation fallacy" suggests that models optimized for prediction may produce conflicting or untrustworthy explanations. Currently, there is no standardized framework to quantify whether pLM-based fitness predictors recover the residues that actually determine fitness, nor is there a clear guide on which modeling choices (e.g., architecture, backbone, pooling strategy) yield trustworthy attributions.

Methodology: The IGLOO Framework

The authors introduce IGLOO (Integrated Gradients for LOcational sensitivity in prOteins), a quantitative framework designed to benchmark and diagnose the fidelity of residue-level attributions.

1. Synthetic Benchmark with Ground Truth

To evaluate attributions, the authors constructed synthetic fitness landscapes with known per-residue importance:

• Proteins: Three proteins with distinct topologies were selected: GB1, T4 Lysozyme, and p53.
• Ground Truth: Residues were assigned importance weights based on literature-derived functional annotations (e.g., active sites, binding sites, structural roles).
• Fitness Landscapes: Three types of fitness functions were implemented to vary complexity:
  • Additive: Fitness decreases linearly with weighted mutation penalties.
  • Epistatic: Includes pairwise interactions between mutations based on structural contacts.
  • Nonlinear (Stability Cliff): Fitness collapses if critical residues are mutated beyond a threshold.
• Noise Sweep: Label noise was introduced to create predictors with a wide range of test accuracies, allowing the authors to decouple predictive performance from attribution quality.

2. Experimental Variables

The framework systematically varied several axes to determine their impact on attribution quality:

• pLM Backbones: Seven models from three families (ESM-C, Profluent E1, Ankh) of varying sizes.
• Predictor Architectures:
  • Pool-First: Attention pooling is applied to pLM embeddings before the predictor head (standard approach).
  • Pool-Later: Per-residue processing (MLP) is applied before pooling, allowing the head to learn residue-specific features.
• Representation Depth: Testing intermediate layers vs. final layers of the pLM.
• Ensembling: Averaging attributions across multiple independently seeded models.

3. Evaluation Metrics

Attribution quality was assessed using five metrics comparing predicted residue scores against ground-truth weights:

1. Spearman correlation between attribution magnitude and ground truth.
2. Precision@k (fraction of top-ranked positions that are annotated).
3. Top-category recall.
4. AUROC for separating annotated from background positions.
5. Sign accuracy.
   These metrics were aggregated into a single TOPSIS score (Technique for Order of Preference by Similarity to Ideal Solution) to facilitate cross-comparison.

Key Results

1. Predictive Accuracy $\neq$ Interpretability

In standard "pool-first" architectures, attribution quality showed little to no correlation with predictive accuracy. Models with high Spearman correlation ($\rho$) on test sets often produced attributions no better than random. This confirms that a model can learn to predict fitness without learning the correct residue-level drivers.

2. The Critical Role of Architecture (Pool-Later)

The most significant finding was that model architecture dictates attribution quality.

• Cascading Randomization: Randomizing predictor weights revealed that attributions in "pool-first" models are largely determined by the pooling layer weights, not the learned predictor head.
• Pool-Later Intervention: Moving learned per-residue computation before the pooling layer (e.g., using a per-residue MLP before attention pooling) dramatically improved the link between predictive accuracy and attribution quality. In the GB1 additive landscape, the correlation between attribution quality and predictive performance rose from $r=0.23$ to $r=0.83$.
• Performance Gains: The "pool-later" architecture nearly doubled per-metric skill scores (e.g., AUROC skill increased from 0.42 to 0.88) and raised the composite TOPSIS score by ~29% compared to the standard baseline.

3. Backbone and Layer Selection

• Backbone Choice: The Ankh family of models produced the most consistent and high-quality attributions across landscapes. Interestingly, larger models (Ankh-large) did not necessarily outperform smaller ones (Ankh-base), suggesting scale is not the primary driver of interpretability.
• Layer Depth: For Ankh-base, intermediate layers (12–36) produced poor attributions despite sometimes yielding good predictive performance. The final layer consistently provided the best attribution quality. Intermediate states were found to be dominated by a few massive-activation dimensions, degrading alignment with ground truth.

4. Ensembling Reduces Variability

Model attribution quality can vary drastically between random seeds. Ensembling 5–10 independently seeded models reduced the standard deviation of TOPSIS scores, following the theoretical $1/\sqrt{K}$ reduction, thereby producing more stable and trustworthy attributions.

5. Validation on Real Proteins

The framework was applied to real datasets for $\alpha$-amylase and $\beta$-lactamase:

• $\alpha$-amylase: Confirmed that the split with the highest predictive accuracy did not yield the best attribution performance. "Pool-later" models outperformed "pool-first" models in identifying functional residues even when predictive performance was comparable.
• $\beta$-lactamase: Attributions guided the selection of mutant candidates. Across 35 substituted position and ligand combinations, 63% of combinations led to the expected effect when compared against binding affinity, and 69% led to the expected effect when compared against stability, demonstrating the utility of IGLOO in guiding protein engineering.

Significance and Claims

The paper claims that high predictive accuracy is insufficient for trusting residue-level attributions in protein engineering. The authors argue that practitioners must move beyond relying solely on regression performance to guide sequence edits.

The primary significance of IGLOO lies in providing:

1. A Diagnostic Protocol: A method to probe what pLMs actually learn about protein function, distinguishing between models that predict well and those that explain well.
2. Actionable Guidelines: A set of defaults to improve trustworthiness:
   • Use "pool-later" architectures (learn per-residue features before global pooling).
   • Prioritize Ankh backbones for feature extraction.
   • Utilize ensembles (5–10 seeds) to reduce variance.
   • Benchmark against absolute baselines (random chance) rather than relative rankings.
3. A Community Standard: An open, extensible framework to pre-screen new pLMs and predictor architectures for their suitability in mechanistic interpretation and attribution-guided design.

The authors conclude that since attribution analysis is computationally cheap relative to training, it should be a mandatory step whenever pLM-based predictors are deployed for sequence-editing decisions.