In Silico Mutagenesis

Every nucleotide position in an input DNA sequence is individually mutated to all three alternative bases, creating a comprehensive map of causal regulatory effects. Unlike experimental methods limited to specific loci, this approach achieves complete saturation across sequences spanning up to 200 kilobases in models like Enformer.

• Generates 3 × L predictions for a sequence of length L
• Reveals position-specific sensitivity to perturbation
• Identifies both activating and repressive mutations
• Maps epistatic interactions when combined with multi-nucleotide perturbations

The quantitative difference between the model's prediction on the reference allele and the mutated allele, computed for every possible substitution. These ISM scores form a mutation effect matrix that directly quantifies the regulatory impact of each nucleotide.

• Positive scores indicate predicted activation upon mutation
• Negative scores indicate predicted repression
• Magnitude reflects functional importance of the position
• Aggregated scores reveal transcription factor binding logic

By analyzing ISM score patterns, models automatically recover known transcription factor binding motifs and discover novel regulatory elements without prior biological annotation. The spatial pattern of high-impact mutations often traces the consensus sequence of DNA-binding proteins.

• High-impact positions cluster at motif cores
• Flanking low-impact regions define motif boundaries
• Enables de novo motif detection from sequence alone
• Validates model's internal representation of cis-regulatory grammar

In silico mutagenesis directly predicts the functional consequences of single-nucleotide variants (SNVs) and rare mutations by measuring their predicted impact on molecular phenotypes like chromatin accessibility or gene expression. This bridges the gap between population genetics and regulatory genomics.

• Prioritizes non-coding variants in genome-wide association study (GWAS) loci
• Estimates effect sizes for rare and private mutations
• Distinguishes deleterious from benign regulatory variants
• Complements tools like SpliceAI for splicing variant interpretation

The choice of reference sequence against which mutations are compared critically influences interpretation. Common baselines include the actual genomic sequence, a dinucleotide-shuffled control preserving k-mer frequencies, or a uniform background. Each baseline answers a different biological question.

• Genomic reference: Measures deviation from natural sequence
• Shuffled control: Isolates motif-specific effects from background
• Uniform background: Reveals absolute nucleotide preferences
• Baseline selection must align with the causal hypothesis being tested

Full saturation mutagenesis requires 3 × L forward passes through the neural network per sequence, making it computationally intensive for long sequences and large cohorts. Optimization strategies include gradient-based approximations and batched inference on specialized hardware.

• Integrated Gradients provides a first-order Taylor approximation
• GPU-accelerated inference pipelines reduce wall-clock time
• Surrogate models can approximate ISM scores for rapid screening
• Trade-off between exhaustiveness and throughput guides experimental design