ROAR

ROAR operates by repeatedly retraining a model from scratch after removing the most relevant features identified by an attribution map. In each iteration, a fraction of top-ranked nucleotides is masked or zeroed out, and the model is retrained on the degraded dataset. The rate at which model performance degrades serves as a direct measure of attribution faithfulness. A high-quality attribution map causes a steep, monotonic drop in accuracy, confirming that the identified features were genuinely critical to the model's decision logic.

Unlike perturbation-based metrics that evaluate a frozen model, ROAR measures causal importance by allowing the model to adapt to missing features. This exposes attribution methods that highlight spurious correlations or brittle patterns the model can easily replace. Key metrics derived from ROAR curves include:

• Area Under the Curve (AUC): Summarizes overall degradation
• Top-k degradation: Performance drop after removing the most salient k% of features
• Rank correlation: Alignment between attribution ranking and true feature necessity

Standard perturbation methods like AOPC or infidelity measurement evaluate attributions on a static, pre-trained model. This can produce misleading results when the model relies on redundant features. If an attribution map highlights one redundant feature while ignoring its equally predictive counterpart, a single-pass perturbation shows no performance drop—falsely suggesting poor attribution. ROAR corrects this by retraining, forcing the model to rely on remaining features and revealing whether the highlighted features were truly indispensable.

In genomic sequence models, ROAR validates whether attribution maps correctly identify transcription factor binding motifs. The process:

• Train a model to predict chromatin accessibility from DNA sequence
• Generate attribution maps using methods like Integrated Gradients or DeepSHAP
• Iteratively mask top-attributed nucleotides and retrain
• Monitor whether predictive power collapses rapidly True motif-driven models degrade sharply when motif cores are removed. Models exploiting flanking sequence context or dinucleotide frequencies degrade more gradually, exposing interpretability gaps.

ROAR has several implementation variants tailored to different genomic tasks:

• ROAR-KL: Replaces masked features with a baseline distribution and measures KL divergence from original predictions
• ROAR-CAUSAL: Incorporates causal inference by conditioning on confounders during retraining
• Progressive ROAR: Removes features in small increments (e.g., 1% per iteration) for high-resolution faithfulness curves
• Cross-model ROAR: Evaluates whether attributions transfer across architectures, testing if highlighted features are universally important or model-specific artifacts

ROAR's primary limitation is its computational expense. Each iteration requires a full model retraining cycle, making it impractical for large foundation models without significant resources. For a typical genomic model with 10 retraining steps, ROAR demands 10x the compute of single-pass evaluation. Mitigation strategies include:

• Using smaller proxy models for rapid iteration
• Early stopping when degradation plateaus
• Parallelizing retraining runs across independent compute nodes Despite the cost, ROAR remains the gold standard for attribution fidelity when regulatory or clinical decisions depend on interpretability.