DeepSHAP

DeepSHAP's core innovation is the linear composition rule, which propagates SHAP values through a network by recursively applying DeepLIFT's rules. This avoids the exponential complexity of naive Shapley calculations.

• Mechanism: Treats each layer as a function and chains attributions via the chain rule
• Efficiency: Reduces computation from O(2^n) to O(n) for deep networks
• Benefit: Makes exact Shapley values tractable for models with millions of genomic features

DeepSHAP requires a reference baseline against which feature contributions are measured. In genomics, this is typically a dinucleotide-shuffled sequence or a uniform background distribution.

• Dinucleotide-shuffled baselines preserve local sequence composition while destroying functional motifs
• Multiple references can be averaged to reduce baseline-dependent variance
• Impact: The choice of reference directly shapes which genomic features are highlighted as important

DeepSHAP inherits the Shapley axioms from cooperative game theory, providing mathematical guarantees that other attribution methods lack.

• Completeness: Attributions sum exactly to the difference between the model's output and the baseline output
• Symmetry: Features with identical effects receive identical attributions
• Dummy: A feature that contributes nothing receives zero attribution
• Additivity: Attributions for ensemble models equal the sum of individual model attributions

DeepSHAP quantifies the functional impact of single-nucleotide variants by computing the SHAP value difference between reference and alternate alleles at each position.

• Delta SHAP scores isolate the marginal contribution of a specific base change
• Positional aggregation sums nucleotide-level scores across regulatory elements
• Application: Prioritizing non-coding variants in whole-genome sequencing studies by their predicted regulatory disruption

Unlike KernelSHAP, which requires sampling perturbations to the input, DeepSHAP propagates attributions directly through the network graph in a single backward pass.

• Single-pass efficiency: One forward and one backward pass per reference
• Scalability: Handles convolutional, recurrent, and attention layers common in genomic architectures
• Comparison: Orders of magnitude faster than perturbation-based Shapley approximations for deep models

DeepSHAP's nucleotide-resolution scores can be aggregated to identify transcription factor binding motifs that drive model predictions.

• Sliding window summation reveals contiguous high-importance regions
• TF-MoDISco integration: High-scoring subsequences are clustered into motif patterns
• Biological validation: Recovered motifs often match known binding sites from databases like JASPAR and ENCODE

The foundational backpropagation algorithm that DeepSHAP builds upon. DeepLIFT compares neuron activations to a reference state to compute contribution scores.

• Uses rescale and revealcancel rules to handle non-linearities
• Avoids saturation problems that plague pure gradient methods
• Essential for explaining variant effects in regulatory genomics models

The unified game-theoretic framework that DeepSHAP approximates. SHAP assigns each genomic feature an importance score based on Shapley values from cooperative game theory.

• Satisfies completeness: attributions sum to the model output difference
• Guarantees consistency: if a feature's contribution increases, its attribution never decreases
• Provides a theoretically rigorous foundation for genomic variant effect prediction

A model-agnostic alternative to DeepSHAP that estimates Shapley values by solving a weighted linear regression over a sampled coalition space.

• Works with any model type, not just deep networks
• Computationally expensive for high-dimensional genomic inputs
• Often used as a ground-truth benchmark to validate DeepSHAP's speed-accuracy tradeoff

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

• Requires selecting a meaningful baseline sequence (e.g., all zeros or reference genome)
• Produces nucleotide-level attributions for genomic sequence models
• Often compared against DeepSHAP in faithfulness benchmarks

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

• Produces mutation effect maps by measuring prediction change per substitution
• Computationally exhaustive but model-agnostic
• Serves as a ground-truth reference for evaluating attribution method accuracy

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

• AOPC: Measures prediction drop as salient nucleotides are sequentially perturbed
• ROAR: Iteratively retrains the model after removing top-attributed features
• Infidelity Measure: Quantifies expected error between input perturbation and attribution perturbation
• Critical for validating DeepSHAP outputs in regulatory genomics applications