Attribution Prior

An attribution prior is implemented as a penalty term added to the model's loss function during training. This penalty is a function of the feature attributions themselves, making the model's interpretability a direct optimization target.

• Mechanism: The total loss becomes Loss = Prediction_Error + λ * Attribution_Penalty
• Gradient Flow: The penalty must be differentiable with respect to the model's parameters to enable backpropagation
• Strength Control: The hyperparameter λ controls the trade-off between predictive accuracy and explanation quality

The prior mathematically formalizes expert expectations about what a valid explanation should look like in a specific biological context. This transforms qualitative domain expertise into a quantitative training signal.

• Smoothness Priors: Penalize large differences in attribution between adjacent nucleotides, reflecting the biological reality that regulatory elements span contiguous regions
• Sparsity Priors: Encourage only a small fraction of input positions to receive non-zero attribution, matching the expectation that only specific motifs drive predictions
• Conservation Priors: Penalize attributions on nucleotides with low evolutionary conservation scores, focusing explanations on functionally constrained regions

A common implementation uses the expected gradients method as the attribution estimator within the prior. This satisfies key axioms like completeness and implementation invariance.

• Completeness: The sum of all feature attributions equals the difference between the model's output and a baseline, ensuring no importance is unaccounted for
• Path Integral: Attributions are computed by integrating gradients along a straight-line path from a reference input to the actual sequence
• Robustness: Unlike raw gradients, expected gradients are less susceptible to saturation artifacts in deep genomic networks

Attribution priors explicitly manage the tension between model accuracy and explanation quality. Strong priors can degrade predictive performance if they conflict with the true data-generating process.

• Pareto Frontier: Training with varying λ values reveals the optimal accuracy-interpretability trade-off curve
• Debugging Signal: A large drop in performance when applying a biologically sound prior may indicate the model is exploiting spurious correlations
• Validation: Attribution priors should be validated against orthogonal experimental data, such as known transcription factor binding sites from ChIP-seq

For genomic sequences, smoothness priors can be implemented in the frequency domain using Fourier transforms. This penalizes high-frequency variation in the attribution map that is biologically implausible.

• Spectral Penalty: The prior computes the total energy in high-frequency components of the attribution signal
• Implementation: Penalty = || HighPassFilter(Attribution_Map) ||²
• Biological Basis: Regulatory elements like enhancers and promoters span 100-1000 base pairs, so attributions should vary smoothly at the single-nucleotide level

Attribution priors share a deep mathematical connection with adversarial robustness. Both techniques regularize the model's input-gradient behavior, leading to more stable and interpretable predictions.

• Gradient Regularization: Both methods penalize large or noisy input gradients
• Robust Explanations: Models trained with attribution priors produce explanations that are less sensitive to small, semantically meaningless input perturbations
• Certified Interpretability: This connection enables the adaptation of adversarial robustness certificates to provide formal guarantees on attribution stability

Penalizes large differences in attribution scores between adjacent nucleotides. In genomics, this encodes the biological prior that functional regulatory elements (e.g., transcription factor binding sites) span contiguous stretches of bases, not isolated positions.

• Mechanism: Adds a penalty proportional to the sum of squared differences between attributions at neighboring positions
• Effect: Eliminates salt-and-pepper noise in saliency maps, producing coherent, block-like attribution patterns
• Use case: Improving the interpretability of variant effect predictors by ensuring nearby silent mutations don't receive wildly different importance scores

Encourages the model to explain predictions using the smallest possible set of nucleotides. This reflects the biological reality that most phenotypic effects are driven by a limited number of causal variants or regulatory elements.

• Mechanism: Applies an L1 penalty directly to the attribution values during training
• Effect: Drives most nucleotide-level attributions to exactly zero, highlighting only the truly salient positions
• Trade-off: Must be carefully tuned; excessive sparsity can suppress genuinely distributed regulatory signals like homopolymeric runs

Aligns attribution scores with evolutionary conservation tracks (e.g., phyloP or GERP scores). The prior encodes that nucleotides conserved across species are more likely to be functionally important.

• Mechanism: Penalizes the Kullback-Leibler divergence between the attribution distribution and a normalized conservation score distribution
• Effect: Model learns to focus on phylogenetically constrained regions, reducing false positive attributions in neutrally evolving sequence
• Limitation: Cannot discover novel functional elements not captured by comparative genomics

Encourages attribution maps to form patterns matching known transcription factor binding motifs from databases like JASPAR or HOCOMOCO. This injects decades of biochemical knowledge into the training process.

• Mechanism: Penalizes the distance between high-attribution subsequences and the closest position weight matrix (PWM) in a motif library
• Effect: Attribution peaks naturally align with experimentally validated binding sites
• Implementation: Often combined with TF-MoDISco for post-hoc validation of the learned attributions

Enforces sparsity at the level of predefined genomic regions (e.g., exons, promoters, enhancers) rather than individual nucleotides. Reflects the prior that functional effects localize to discrete regulatory units.

• Mechanism: Uses a group lasso penalty where each annotated region forms a group; the model is penalized for using too many distinct regions
• Effect: Produces explanations that are interpretable at the level of biological annotation tracks
• Data requirement: Requires high-quality genome annotation files (GTF/GFF) to define group boundaries

Trains the model to produce attribution maps that are invariant to biologically irrelevant perturbations. For example, shuffling non-coding sequence far from any regulatory element should not change attributions at the gene body.

• Mechanism: An adversarial network attempts to distinguish attributions from original vs. perturbed sequences; the primary model is penalized when the discriminator succeeds
• Effect: Robust attributions that ignore spurious correlations in the input
• Inspiration: Derived from domain-adversarial training in transfer learning