Nucleotide-level Attribution

Unlike region-based methods that aggregate importance over windows, nucleotide-level attribution assigns a unique scalar score to each A, C, G, and T in the input sequence. This allows the identification of causal single-nucleotide polymorphisms (SNPs) and point mutations that drive model predictions.

• Resolves contributions at the single base-pair level
• Directly maps to variant effect prediction tasks
• Enables comparison with deep mutational scan (DMS) ground truth

Most nucleotide-level attribution methods rely on backpropagating gradients from the output prediction through the network to the input layer. The magnitude of the gradient with respect to each input nucleotide indicates its influence.

• Saliency maps use raw input gradients
• Input × Gradient multiplies gradients by input values to reduce noise
• Integrated Gradients accumulates gradients along a path from a baseline to satisfy the completeness axiom

Many attribution algorithms require defining a neutral baseline or reference sequence to compare against. The attribution score reflects the difference in prediction between the actual nucleotide and what would be predicted if that position were replaced with the reference.

• DeepLIFT compares activations to a reference state using rescale and revealcancel rules
• Common baselines include dinucleotide-shuffled sequences or uniform background frequencies
• Choice of baseline critically impacts the biological interpretability of results

The faithfulness of nucleotide-level attributions is validated through in-silico mutagenesis (ISM) — systematically mutating each position and measuring the actual change in model output. High-attribution positions should cause large prediction shifts when altered.

• ISM serves as a ground-truth approximation for attribution accuracy
• AOPC (Area Over the Perturbation Curve) quantifies prediction drop as salient bases are removed
• ROAR (Remove And Retrain) tests if attributions remain valid after model retraining

High-attribution nucleotides often cluster into contiguous sequence motifs that correspond to transcription factor binding sites or other regulatory elements. Algorithms like TF-MoDISco group these patterns into recurring, biologically meaningful motifs.

• Attribution maps are converted into sequence logos showing information content per position
• Clustering reveals known and novel regulatory grammar
• Enables discovery of cell-type-specific binding preferences

Single attribution scores can be misleading without confidence estimates. Modern methods compute attribution uncertainty by measuring variance across model ensembles, dropout masks, or Bayesian approximations.

• Bayesian neural networks provide posterior distributions over attribution values
• Deep ensembles reveal disagreement between independently trained models
• High-variance positions indicate ambiguous model logic requiring further investigation

A systematic perturbation technique that computationally mutates every nucleotide in a sequence to quantify its impact on model predictions.

• Performs a virtual saturation mutagenesis scan across the entire input
• Measures the delta score between reference and alternate allele predictions
• Computationally expensive but provides the most direct causal measure of nucleotide importance
• Often used as a ground-truth benchmark for validating faster attribution methods

An axiomatic attribution method that computes the path integral of gradients from a baseline input to the actual input, satisfying the completeness axiom.

• Guarantees that attributions sum to the prediction difference from baseline
• Requires careful selection of a neutral baseline sequence (e.g., all-zero embedding)
• Produces fine-grained nucleotide-level scores suitable for regulatory genomics
• Satisfies sensitivity and implementation invariance axioms

A backpropagation-based attribution algorithm that compares neuron activations to a reference state using rescale and revealcancel rules.

• Explains differences in output from a reference sequence rather than absolute predictions
• Handles non-linearities efficiently without requiring gradient integration
• Particularly effective for identifying variant-specific effects on transcription factor binding
• Can be combined with SHAP values via DeepSHAP for high-speed approximation

A unified framework based on Shapley values from cooperative game theory that assigns each genomic feature an importance score for a particular prediction.

• Provides a theoretically grounded distribution of credit among nucleotides
• KernelSHAP offers model-agnostic estimation via weighted linear regression
• DeepSHAP combines DeepLIFT rules with Shapley calculations for deep models
• Guarantees local accuracy, missingness, and consistency properties

A method that clusters high-contribution genomic subsequences identified by attribution maps into recurring, biologically meaningful motif patterns.

• Translates raw nucleotide-level scores into human-interpretable sequence motifs
• Uses deep clustering on attribution-weighted subsequences
• Outputs position weight matrices that can be matched against known transcription factor databases
• Bridges the gap between per-nucleotide scores and biological mechanism discovery

Quantitative measures that evaluate how accurately an attribution map reflects the true decision-making logic of a genomic model through perturbation experiments.

• ROAR iteratively retrains models after removing top-attributed features
• AOPC measures prediction drop as salient nucleotides are sequentially masked
• Infidelity measure quantifies expected error between input perturbation and attribution perturbation
• Essential for benchmarking and comparing different attribution methods objectively