In-silico Mutagenesis (ISM)

ISM performs an exhaustive computational scan by substituting every possible nucleotide at every position in a sequence. For a sequence of length L, this generates L × 3 mutated sequences (replacing the reference base with the three alternative bases). Each mutated sequence is fed through the trained model to record the change in prediction score. This creates a complete mutation map that reveals which positions are functionally critical and which are robust to change.

The core quantitative output of ISM is the delta score. This is calculated as:

• ΔScore = Prediction(Alternate Allele) - Prediction(Reference Allele) A large negative delta for a regulatory model indicates a disruptive mutation that destroys a binding site. A positive delta might indicate a gain-of-function mutation. These scores provide a direct, nucleotide-resolution measure of variant effect prediction without requiring separate models.

By aggregating delta scores across all three possible mutations at each position, ISM generates a nucleotide-level attribution map. Common aggregation strategies include:

• L2-norm: The Euclidean norm of the three delta scores, capturing the total mutational sensitivity.
• Max absolute delta: The single most impactful mutation at that position. This map visually highlights functionally constrained regions such as transcription factor binding sites or splice junctions.

ISM predictions are frequently validated against Deep Mutational Scans (DMS), which are high-throughput experimental assays measuring the functional impact of thousands of variants. Because ISM is a computational method, its predictions can be directly correlated with DMS measurements to quantify faithfulness. High correlation between ISM delta scores and experimental DMS scores validates both the predictive model and the interpretability method.

A naive ISM implementation requires 3L forward passes, which is computationally prohibitive for long sequences or large datasets. Optimization strategies include:

• Strided or windowed ISM: Only mutating positions within a sliding window of interest.
• Predictive approximations: Using Taylor expansion or gradient-based methods to estimate delta scores without full forward passes.
• Batch processing: Leveraging GPU parallelism to process thousands of mutated sequences simultaneously. Despite the cost, ISM is considered a gold-standard for its model-agnostic, assumption-free nature.

ISM attribution maps are often processed by tools like TF-MoDISco to identify recurring, high-contribution subsequences. These subsequences are clustered and aligned into sequence logos representing conserved biological motifs. This directly links a model's internal decision logic to known (or novel) protein-binding consensus sequences, bridging the gap between a black-box neural network and established molecular biology.

The general class of methods that assign a relevance score to each input nucleotide or genomic region for a specific neural network prediction.

• ISM is a perturbation-based attribution method
• Contrasts with backpropagation-based methods like Integrated Gradients
• Outputs a saliency map indicating which bases drove the prediction
• Essential for moving beyond 'black box' predictions to mechanistic understanding

An axiomatic feature attribution method that computes the path integral of gradients from a baseline input to the actual input.

• Satisfies the completeness axiom: attributions sum to the prediction difference
• Requires a non-informational baseline (e.g., all zeros or reference sequence)
• Unlike ISM, does not require individual forward passes for every mutation
• Provides nucleotide-level attribution with theoretical guarantees

A backpropagation-based attribution algorithm that compares neuron activations to a reference state.

• Uses rescale and revealcancel rules to propagate relevance scores
• Computationally efficient compared to exhaustive ISM
• Explains variant effects by comparing alternate to reference alleles
• Often combined with SHAP to create DeepSHAP for high-speed approximation

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

• KernelSHAP: model-agnostic approximation using weighted linear regression
• DeepSHAP: combines DeepLIFT rules with Shapley calculations for speed
• Guarantees local accuracy and consistency properties
• Computationally expensive for full genomic sequences without approximation

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 model after removing top-attributed features
• Infidelity Measure: quantifies error between input perturbation and attribution perturbation
• ISM itself serves as a ground-truth benchmark for validating faster attribution methods

A high-throughput experimental method that assays the functional impact of thousands of genomic variants.

• Serves as the gold-standard ground truth for validating ISM predictions
• Measures variant effect on molecular phenotypes like binding or expression
• ISM is the computational analog of DMS, enabling virtual saturation mutagenesis
• Concordance between ISM and DMS validates model interpretability