Saturation Mutagenesis Scoring

The core scoring mechanism computes the log-ratio between the probability of the reference allele and the alternate allele under a pretrained genomic language model. A negative score indicates the variant disrupts the learned regulatory grammar, while a positive score suggests a neutral or benign substitution. This approach leverages autoregressive sequence likelihood or masked token prediction to quantify the model's surprise at the mutated sequence without requiring any labeled variant effect training data.

Unlike experimental deep mutational scanning, which is constrained by library complexity and transformation efficiency, computational saturation mutagenesis systematically iterates through all three possible alternate nucleotides at every position in a target locus. For a 1,000-base-pair regulatory element, this generates 3,000 variant effect predictions in a single inference pass. Genomic language models with subquadratic scaling, such as those using the Hyena operator or Mamba state space models, can extend this exhaustive scanning to megabase-length loci without prohibitive compute costs.

The scoring output reveals position-specific constraint profiles across the locus. Critical regulatory positions—such as transcription factor binding motifs or splice donor sites—exhibit high sensitivity, where nearly all substitutions produce strongly negative scores. In contrast, spacer regions or positions with degenerate consensus sequences tolerate a broader range of nucleotides. This fine-grained mapping enables researchers to identify cryptic functional elements that are invisible to conservation-based methods like PhyloP or GERP++.

Saturation scoring naturally produces a quantitative allelic series for every position, ranking the three alternate nucleotides by their predicted functional severity. This capability distinguishes between hypomorphic variants (partial loss of function) and null variants (complete disruption), providing mechanistic hypotheses for variant interpretation. For clinical genomics applications, this stratification aids in classifying variants of uncertain significance by predicting whether a substitution is likely to be pathogenic, benign, or intermediate based on its position-specific score distribution.

Genomic language models perform saturation mutagenesis scoring in a zero-shot regime, meaning the model was never explicitly trained on labeled variant effect data. The functional impact emerges purely from the model's learned representation of evolutionary sequence constraints. This property enables cross-locus generalization: a single pretrained model can score variants at any genomic region—promoters, enhancers, splice sites, or untranslated regions—without task-specific fine-tuning. The approach is limited only by the model's pretraining corpus diversity and the evolutionary depth captured during self-supervised learning.

Computational saturation scores are benchmarked against Massively Parallel Reporter Assays (MPRAs) and deep mutational scanning datasets. High-performing genomic language models achieve Spearman correlations exceeding 0.7 with experimental measurements of regulatory activity. Discrepancies between computational predictions and experimental data often reveal trans-acting factors or chromatin context dependencies not captured by sequence-alone models, providing actionable hypotheses for follow-up investigation. Concordance is highest for loci where cis-regulatory grammar dominates functional output.

The foundational capability that enables saturation mutagenesis scoring without labeled training data. A genomic language model computes the log-likelihood ratio between reference and alternate alleles—the difference in probability reflects functional impact. This approach leverages evolutionary sequence constraints learned during pretraining to identify deleterious mutations across an entire locus in a single pass.

The computational workhorse that systematically introduces every possible single-nucleotide substitution into a reference sequence and records the resulting change in model predictions. Key steps include:

• Reference anchoring: establishing baseline predictions for the wild-type sequence
• Exhaustive perturbation: iterating through all 3 alternate nucleotides at every position
• Delta quantification: measuring the magnitude of prediction shift for each variant This technique identifies critical regulatory nucleotides without wet-lab experimentation.

A numerical prediction of a genetic variant's functional consequence, typically computed as the log-likelihood ratio (LLR) between reference and alternate alleles under a pretrained genomic language model. Scores can be:

• Continuous: raw LLR values reflecting the magnitude of functional disruption
• Binarized: thresholded classifications distinguishing pathogenic from benign variants
• Ranked: positional prioritization across a locus to identify mutation-sensitive regions These scores form the output matrix of a saturation mutagenesis scan.

The probability assigned to a genomic sequence by an autoregressive model, serving as the mathematical foundation for variant scoring. The log-likelihood quantifies how well a sequence conforms to learned patterns of natural DNA. In saturation mutagenesis, the delta log-likelihood between reference and alternate alleles isolates the variant's specific contribution to functional disruption, controlling for background sequence context.

A metric derived from a language model's cross-entropy loss that quantifies how surprised the model is by a given sequence. In genomics, low perplexity indicates strong evolutionary constraint and functional importance. When applied to saturation mutagenesis:

• Positional perplexity shifts highlight nucleotides where substitutions cause disproportionate model uncertainty
• Constraint landscapes reveal functional domains within a locus
• Comparative perplexity across species identifies conserved regulatory elements

Standardized collections of curated datasets and evaluation protocols that rigorously measure variant effect prediction performance. These benchmarks provide ground-truth labels from deep mutational scanning experiments and clinical variant databases, enabling:

• Calibration assessment: measuring how well predicted scores align with empirical functional measurements
• Model comparison: head-to-head evaluation of different genomic language models on identical variant sets
• Stratified analysis: performance breakdowns by variant type, genomic context, and evolutionary conservation level