Grad-CAM

An axiomatic attribution method that satisfies the completeness axiom—the sum of all feature attributions equals the difference between the model's output and a baseline. For genomic models, it computes the path integral of gradients from a neutral reference sequence (e.g., all zeros or shuffled dinucleotides) to the actual input.

• Key property: Sensitivity and implementation invariance are mathematically guaranteed
• Genomic use: Identifying regulatory variants by attributing prediction differences between reference and alternate alleles
• Limitation: Requires careful baseline selection; a poor baseline can produce misleading attributions

A backpropagation-based attribution algorithm that compares each neuron's activation to a reference activation from a neutral input. It uses rescale and revealcancel rules to handle non-linearities, avoiding the saturation problem that plagues pure gradient methods.

• Rescale rule: Distributes contribution in proportion to activation differences
• Revealcancel rule: Handles cases where positive and negative contributions cancel
• Genomic advantage: Efficiently scores variant effects by comparing mutant vs. reference sequences in a single backward pass

A unified framework grounded in Shapley values from cooperative game theory. Each nucleotide or genomic region is treated as a player in a coalition; SHAP computes the marginal contribution of each feature averaged over all possible feature subsets.

• Additive property: Prediction = base value + sum of SHAP values
• KernelSHAP: Model-agnostic approximation using weighted linear regression
• DeepSHAP: High-speed variant combining DeepLIFT rules with Shapley calculations
• Genomic context: Provides theoretically justified importance scores for regulatory element discovery

A systematic perturbation technique that computationally mutates every nucleotide position to all possible alternate bases and measures the resulting change in model prediction. Unlike gradient-based methods, ISM provides exact effect measurements without linear approximations.

• Exhaustive approach: Tests all 3 alternate bases at every position
• Saturation mutagenesis: Analogous to experimental deep mutational scans
• Output: Produces a mutation effect map showing which positions are most sensitive to change
• Computational cost: Requires L × 3 forward passes for a sequence of length L

A motif discovery algorithm that clusters high-contribution genomic subsequences from attribution maps into recurring, biologically meaningful patterns. It bridges the gap between raw nucleotide-level scores and interpretable transcription factor binding motifs.

• Input: Attribution scores from any method (Grad-CAM, DeepLIFT, ISM)
• Process: Extracts high-scoring seqlets, aligns them, and clusters by similarity
• Output: Position weight matrices representing discovered motifs
• Validation: Recovers known motifs (e.g., GATA, AP-1) and discovers novel binding patterns

The learned coefficients in transformer-based genomic models (e.g., Enformer, DNABERT) that indicate the relative importance of different nucleotide positions when computing sequence representations. Unlike Grad-CAM, attention weights are intrinsic to the architecture rather than post-hoc.

• Multi-head attention: Different heads capture distinct biological relationships (e.g., proximal vs. distal enhancer-promoter interactions)
• Caveat: Raw attention weights do not always correlate with feature importance; they reflect token mixing, not necessarily prediction relevance
• Best practice: Combine with gradient-based methods for robust interpretation