Attribution Uncertainty

Instead of a single importance score per nucleotide, Bayesian methods produce a posterior distribution over attributions. This is achieved by placing priors over model weights (e.g., via Monte Carlo Dropout or Stochastic Weight Averaging-Gaussian) and performing multiple stochastic forward passes. The resulting variance across passes quantifies epistemic uncertainty—the model's lack of knowledge due to limited data. A wide credible interval for a transcription factor binding site attribution signals that the model has not seen enough similar motifs to be confident, directly informing regulatory risk assessment.

Deep ensembles—training multiple identical architectures with different random initializations—provide a frequentist approach to uncertainty. Attribution maps are generated independently for each model in the ensemble. The coefficient of variation or interquartile range at each nucleotide position is then computed. High disagreement among ensemble members on the importance of a specific intronic variant indicates model uncertainty. This method is computationally expensive but is considered a gold standard for detecting brittle explanations that disappear with minor changes to the training trajectory.

A faithful attribution should be stable under trivial input noise. This characteristic measures the local Lipschitz continuity of the explanation function. By adding small, biologically irrelevant perturbations (e.g., synonymous codon changes) to the input sequence and recalculating the attribution map, one can quantify aleatoric uncertainty stemming from irreducible data noise. A high max-sensitivity score indicates that the explanation is fragile. Formally, it bounds the maximum change in attribution relative to the change in input, ensuring the explanation does not rely on adversarial, brittle features.

Rather than providing a confidence interval for the importance score itself, conformal attribution wraps the feature selection process in a rigorous statistical guarantee. It produces a prediction set of genomic regions that contains the truly important features with a user-specified probability (e.g., 95%). This is achieved by calibrating a heuristic notion of uncertainty (like a model's softmax score) on a held-out calibration set. The result is a distribution-free, finite-sample guarantee that the most salient motif is not excluded from the explanation, which is crucial for clinical reporting where false negatives are unacceptable.

Many attribution methods like Integrated Gradients and DeepLIFT require a reference or baseline input. The choice of baseline (e.g., all-zero vectors, shuffled sequences) is arbitrary and introduces a hidden source of uncertainty. This characteristic is addressed by sampling from a distribution of plausible baselines (e.g., a set of dinucleotide-shuffled sequences that preserve local structure) and aggregating the resulting attributions. The variance introduced by baseline sampling reveals how much the explanation depends on the chosen reference point, preventing over-interpretation of artifacts caused by a single, poor baseline choice.

Adapted from model calibration, this concept evaluates whether the magnitude of an attribution score corresponds to the true probability of a biological effect. Attributions are binned by score magnitude, and the empirical frequency of a functional effect (validated via Deep Mutational Scans or MPRA) is plotted against the mean score. A perfectly calibrated attribution map would lie on the diagonal. The Expected Calibration Error (ECE) quantifies the deviation. This transforms the abstract importance score into a physically meaningful probability, allowing a CTO to set a threshold based on acceptable false discovery rates.

Quantitative measures that evaluate how accurately an attribution map reflects the true decision-making logic of a genomic model. These metrics operate by introducing perturbations to the input sequence and measuring the correlation between the change in prediction and the change in the attribution score.

• Comprehensive deletion: Iteratively removes the most salient features.
• Sufficiency: Tests if the most salient features alone can recover the prediction.
• Correlation: Measures the monotonic relationship between attribution magnitude and prediction impact.

A specific metric that quantifies the expected error between a significant perturbation of the input sequence and the corresponding perturbation of the attribution map. It formally defines the local fidelity of an explanation by measuring how well the attribution function predicts the model's output change under a perturbation distribution.

• Mathematical basis: Rooted in the completeness axiom.
• Perturbation types: Gaussian noise, mutagenesis, or masking.
• Use case: Comparing the reliability of different attribution methods like Integrated Gradients vs. DeepLIFT.

The study of how the uncertainty in the output of a genomic model can be apportioned to different sources of uncertainty in its nucleotide inputs. Unlike feature attribution, which explains a single prediction, sensitivity analysis maps the global variance of the model's output space.

• Global methods: Sobol indices, Morris method.
• Local methods: Partial derivatives at a specific input point.
• Application: Identifying which regulatory regions most influence the variance of a gene expression prediction across a population.

A high-throughput experimental method that assays the functional impact of thousands of genomic variants, often used as a ground-truth benchmark for validating attribution methods. DMS data provides a direct empirical measurement of the effect of every possible single amino acid substitution in a protein.

• Validation: Correlating in-silico attribution scores with in-vitro DMS fitness scores.
• Scale: Measures ~10^4 variants in a single experiment.
• Relevance: The definitive standard for proving that a model's attribution map identifies functionally critical nucleotides.

A suite of tests designed to verify that an attribution method is sensitive to the learned parameters of the genomic model and not just the input data. The most critical check is model parameter randomization, where the weights of the trained model are progressively destroyed.

• Cascading randomization: Attributions must degrade as layers are randomized from top to bottom.
• Data randomization: The method must produce different attributions for a model trained on shuffled labels.
• Purpose: Filters out edge-detectors and other non-explanatory methods that produce visually pleasing but model-agnostic heatmaps.

A benchmarking framework that iteratively retrains a genomic model after removing the most relevant features according to an attribution map. If the attribution is faithful, retraining on the modified data should cause a sharp, predictable drop in model performance.

• Process: Rank features by attribution, remove top fraction, retrain, measure accuracy.
• Comparison: A better attribution method causes a faster degradation in performance.
• Contrast with AOPC: Unlike Area Over the Perturbation Curve (AOPC), ROAR accounts for the model's ability to adapt to the missing features.