Genomic Ablation

Unlike correlation-based feature attribution methods, genomic ablation establishes causal relationships between sequence elements and model outputs. By systematically removing a genomic region and observing the resulting change in prediction, the technique answers: 'Is this region necessary for the model's decision?' This moves beyond passive observation to active intervention, providing a ground-truth signal for validating the importance scores generated by methods like Integrated Gradients or SHAP.

The most common implementation involves a sliding window that traverses the input sequence, replacing nucleotides with a neutral baseline value. Key parameters include:

• Window size: Typically 10-100 base pairs, balancing resolution with computational cost.
• Stride: Step size between windows, often set to 1 for nucleotide-level resolution.
• Baseline value: Usually zeros, uniform distribution, or dinucleotide-shuffled sequences to preserve local composition. The resulting delta score—the difference between the original and masked prediction—is plotted as a perturbation map across the sequence.

Genomic ablation serves as the empirical backbone for faithfulness metrics that evaluate attribution methods. By comparing the regions an attribution map identifies as 'important' against the regions that actually cause prediction drops when ablated, researchers compute metrics like:

• Area Over the Perturbation Curve (AOPC): Measures the rate of prediction decline as top-ranked features are removed.
• ROAR (RemOve And Retrain): Retrains the model after removing the most salient features to test if the attribution method captured the model's true dependency structure. A faithful attribution map will show a steep, monotonic drop in performance.

While related, genomic ablation and in-silico mutagenesis (ISM) are distinct perturbation strategies:

• Ablation: Removes or masks entire regions, testing for necessity (is this region required?).
• ISM: Substitutes individual nucleotides with alternatives, testing for sufficiency (which specific base changes alter the prediction?). Ablation identifies critical regulatory modules like enhancers or promoters, while ISM pinpoints the exact causal single nucleotide variants (SNVs) within those modules. Both are often used together for a complete causal picture.

A primary application is the de novo discovery of regulatory elements without prior biological annotation. By ablating across a long sequence and measuring the impact on a gene expression prediction model, researchers can identify:

• Novel enhancers: Distal regions that, when removed, cause a significant drop in predicted expression.
• Silencers: Regions whose removal increases predicted expression.
• Insulator boundaries: Segments that, when ablated, cause ectopic regulatory interactions. This provides a hypothesis-generation engine for experimental validation via CRISPR interference or massively parallel reporter assays (MPRAs).

Exhaustive ablation is computationally expensive. For a sequence of length L with a window of size W and stride S, the number of forward passes is approximately (L - W) / S + 1. For a 10kb sequence with a 50bp window and 1bp stride, this requires ~9,951 model inferences. Optimization strategies include:

• Larger strides for coarse mapping, followed by fine-mapping.
• Batch processing of all masked sequences.
• Gradient-based approximations that estimate ablation effects without full forward passes. This computational burden is the primary trade-off against the method's causal rigor.

Ablation identifies cis-regulatory elements by measuring prediction drops when specific non-coding regions are masked. By sliding an ablation window across intergenic space, researchers pinpoint enhancers and silencers that drive cell-type-specific expression.

• Causal validation: Unlike correlation-based methods, ablation proves necessity
• Resolution: Can isolate functional elements down to 10-50bp windows
• Benchmark: Validated against MPRA and STARR-seq experimental data

Systematic in-silico ablation of single nucleotides quantifies the functional impact of genetic variants. By comparing prediction scores between reference and alternate alleles, ablation generates delta scores that rank variants by deleteriousness.

• Non-coding variants: Prioritizes pathogenic variants in regulatory regions
• Saturation mutagenesis: Computationally tests all possible single-nucleotide changes
• Clinical relevance: Outperforms CADD and other conservation-based scores

Ablation maps transcription factor binding sites by measuring the sensitivity of binding predictions to localized sequence perturbations. When ablation of a 6-20bp motif causes a sharp prediction drop, it reveals a causal binding footprint.

• Motif discovery: Identifies novel binding motifs without prior knowledge
• Cooperativity: Detects composite elements requiring multiple factors
• Cell-type specificity: Reveals context-dependent binding logic

Ablation serves as a diagnostic tool for identifying spurious correlations learned by genomic models. If masking a biologically irrelevant region (e.g., a poly-A stretch) significantly alters predictions, the model has learned a shortcut feature.

• Artifact detection: Flags models relying on sequencing bias
• Robustness testing: Validates that models use causal biology, not confounders
• Regulatory compliance: Provides evidence for FDA and EMA submission

Ablation of nucleotides flanking splice junctions reveals the sequence determinants of splicing decisions. By masking donor and acceptor sites, branch points, and polypyrimidine tracts, models quantify each element's contribution to spliceosome assembly.

• Cryptic splice sites: Identifies latent splice sites activated by mutations
• Spliceopathy research: Maps causal logic of splicing diseases
• Therapeutic design: Informs antisense oligonucleotide target selection

Ablation across syntenic regions from multiple species tests whether models learn evolutionarily conserved functional logic. If ablation effects are consistent across orthologous sequences, the model has captured phylogenetically stable regulatory grammar.

• Evolutionary biology: Distinguishes conserved from lineage-specific elements
• Model generalizability: Validates transfer learning across species
• Deep homology: Reveals ancient regulatory circuits preserved across taxa

A systematic perturbation technique that computationally mutates every nucleotide in a sequence to quantify its impact on model predictions. ISM creates a complete map of variant effects by introducing single-nucleotide substitutions at each position and measuring the change in output.

• Produces delta scores for every possible variant
• Serves as a ground-truth benchmark for attribution methods
• Computationally expensive for long sequences
• Often used to validate Deep Mutational Scan experiments

An axiomatic feature attribution method that computes the path integral of gradients from a baseline input to the actual input. This technique satisfies the completeness axiom, meaning the sum of all feature attributions equals the difference between the model's output for the input and the baseline.

• Uses a neutral reference sequence as the baseline
• Provides nucleotide-level attribution scores
• Avoids gradient saturation issues common in other methods
• Computationally efficient for large genomic models

A backpropagation-based attribution algorithm that compares neuron activations to a reference state using rescale and revealcancel rules. DeepLIFT explains the difference in output from a reference sequence by propagating activation differences backward through the network.

• Handles non-linearities through custom gradient rules
• Avoids zero-gradient artifacts in saturated neurons
• Forms the computational backbone of DeepSHAP
• Effective for identifying regulatory variants

A method that clusters high-contribution genomic subsequences identified by attribution maps into recurring, biologically meaningful motif patterns. TF-MoDISco transforms raw per-nucleotide importance scores into interpretable sequence logos.

• Identifies transcription factor binding sites
• Groups similar patterns across thousands of sequences
• Outputs sequence logos for each discovered motif
• Bridges the gap between attribution scores and biological insight

A unified framework based on Shapley values from cooperative game theory that assigns each genomic feature an importance score for a particular prediction. SHAP guarantees a fair distribution of credit among all input nucleotides.

• Satisfies consistency, local accuracy, and missingness axioms
• KernelSHAP provides model-agnostic estimation
• DeepSHAP combines DeepLIFT with Shapley calculations
• Computationally intensive but theoretically rigorous

Quantitative measures that evaluate how accurately an attribution map reflects the true decision-making logic of a genomic model through perturbation experiments. These metrics test whether removing highly attributed nucleotides actually changes predictions.

• ROAR iteratively retrains models after feature removal
• AOPC measures prediction drop as salient bases are perturbed
• Infidelity quantifies expected perturbation error
• Essential for validating any interpretability method