Faithfulness Metrics

A core metric that evaluates attribution faithfulness by sequentially perturbing the most salient nucleotides and measuring the drop in prediction probability. A steeper, earlier drop indicates higher attribution quality.

• Process: Rank nucleotides by attribution score, then iteratively replace top-k positions with a baseline token (e.g., N or reference allele)
• Interpretation: A higher AOPC score means the attribution map correctly identified the sequence elements most critical to the model's prediction
• Application: Commonly used to benchmark Integrated Gradients, DeepLIFT, and attention-based methods against each other on genomic variant effect prediction tasks
• Example: When evaluating a splice site predictor, a faithful attribution map will cause a 90% probability drop after mutating only the canonical GT-AG dinucleotides

A rigorous benchmarking framework that iteratively retrains a genomic model after removing the most relevant features according to an attribution map. This tests whether the attributions capture causal dependencies rather than correlations.

• Methodology: At each iteration, replace the top-k most salient nucleotides with a baseline, then retrain the model from scratch on the modified dataset
• Key Insight: If an attribution method is faithful, retrained model performance should degrade faster than when removing random nucleotides
• Computational Cost: Requires multiple full training cycles, making it expensive for large genomic foundation models but providing the gold standard for faithfulness evaluation
• Variant: ROAR can be applied at the motif level rather than nucleotide level to test whether attribution maps correctly identify whole regulatory elements

A metric that quantifies the expected error between a significant perturbation of the input sequence and the corresponding perturbation of the attribution map. It formalizes the notion that faithful attributions should predict the effect of input changes.

• Definition: Computed as the expected squared difference between the model's output change and the dot product of the attribution map with the perturbation vector
• Mathematical Basis: Derived from the idea that an attribution map should function as a local linear approximation of the model in the neighborhood of the input
• Advantage: Does not require retraining and can be computed efficiently for any differentiable genomic model
• Perturbation Types: Can use Gaussian noise, random mutagenesis, or biologically-informed substitution matrices (e.g., transition-transversion ratios)

A correlation-based metric that measures how consistently an attribution method identifies the same genomic regions as important when the model is trained with different random initializations. High sensitivity-n indicates attribution robustness.

• Procedure: Train n identical architecture models with different random seeds, compute attributions for the same input, then measure the correlation (e.g., Spearman rank) between attribution maps
• Interpretation: Low sensitivity-n suggests the attribution method is capturing noise or artifacts rather than stable model logic
• Genomic Context: Particularly important for regulatory genomics where multiple degenerate motifs can achieve similar function
• Threshold: A sensitivity-n above 0.8 across 5 independent training runs is generally considered acceptable for publication-quality interpretability claims

The gold-standard experimental validation that compares computational attribution scores against high-throughput in vitro or in vivo mutagenesis data measuring the functional impact of thousands of genomic variants.

• Benchmark Datasets: Includes massively parallel reporter assays (MPRAs), saturation mutagenesis of regulatory elements, and deep mutational scans of protein-coding regions
• Metric: Spearman or Pearson correlation between the predicted effect (delta score) from the attribution map and the experimentally measured functional consequence
• Regulatory Relevance: For non-coding variants, DMS correlation validates whether attribution maps correctly identify pathogenic single-nucleotide polymorphisms in enhancers and promoters
• Limitation: DMS data is expensive to generate and only available for a limited set of genomic loci, restricting its use as a universal benchmark

A suite of diagnostic tests designed to verify that an attribution method is genuinely sensitive to the learned parameters of the genomic model rather than producing edge-case artifacts or input-dependent noise.

• Model Parameter Randomization Test: Randomize the model's weights layer by layer from top to bottom; attributions should degrade progressively, not remain unchanged
• Data Randomization Test: Train the model on randomly permuted labels; attributions should differ substantially from those of a model trained on real data
• Edge Case Detection: Verify that attributions change appropriately when the input sequence is reversed, complemented, or padded with random nucleotides
• Implementation: These checks should be run before any biological interpretation of attribution maps to ensure the method is functioning correctly