Variant Effect Score

The core metric is typically calculated as the log-likelihood ratio (LLR) between the alternate and reference alleles. A genomic language model assigns a probability to every sequence; a pathogenic variant causes a significant drop in probability. The score is computed as LLR = log P(Sequence_alt) - log P(Sequence_ref). A highly negative score indicates the alternate allele is much less likely under the model's learned distribution of natural, functional DNA, suggesting a disruptive effect.

A defining feature is the ability to perform zero-shot prediction. Genomic language models pretrained solely on unlabeled reference genomes via self-supervision can score variant effects without any fine-tuning on labeled clinical datasets. The model's internal representation of evolutionary sequence constraints, learned during pretraining, serves as an implicit fitness landscape. This allows functional impact assessment for variants in genes or non-coding regions never seen during training.

Unlike classical conservation scores that rely on fixed evolutionary alignments, these scores are fully context-dependent. A self-attention mechanism allows the model to weigh the influence of surrounding nucleotides, capturing long-range dependencies such as enhancer-promoter interactions. A single nucleotide change in a critical transcription factor binding motif will produce a dramatically different score than the same change in a non-functional spacer region, reflecting the local regulatory grammar.

The scoring framework naturally handles heterozygous and homozygous states. For a heterozygous variant, the score can be computed as the difference between the log-likelihood of the reference sequence and the average log-likelihood of the two allelic sequences. This provides a quantitative measure of the dosage effect, distinguishing between a complete loss-of-function and a partial hypomorphic impact, which is critical for understanding dominant versus recessive inheritance patterns.

Robust models enforce reverse complement invariance. Since DNA is double-stranded, a variant's effect must be identical whether scored on the forward or reverse strand. This is achieved through reverse complement data augmentation during pretraining or by symmetrizing the LLR computation. The final score is often the average of predictions from both strands, ensuring the metric is a biophysically consistent property of the double helix.

Variant effect scores enable exhaustive in-silico saturation mutagenesis. By computationally introducing every possible single-nucleotide substitution at every position in a regulatory element or protein-coding exon and scoring each one, researchers can generate a complete functional map of a locus. This heatmap of constraint reveals critical residues and cryptic splice sites, prioritizing variants for experimental validation with a throughput impossible via wet-lab assays.

The capability of a genomic language model to predict the functional impact of a genetic variant without explicit training on labeled variant effect data. The model computes the log-likelihood ratio between the reference and alternate allele sequences using only its pretrained understanding of evolutionary sequence constraints.

• Mechanism: Compares sequence probability with reference vs. alternate allele
• Key Advantage: No need for curated training datasets of known pathogenic variants
• Example: DNABERT-2 scoring ClinVar variants without fine-tuning

The probability assigned to a genomic sequence by an autoregressive model, serving as the foundational metric from which variant effect scores are derived. A lower log-likelihood for the alternate allele indicates the variant disrupts learned patterns of natural DNA.

• Formula: Log P(sequence | model parameters)
• Application: Measures evolutionary constraint and identifies functional elements
• Relationship: Variant Effect Score = log P(alt) - log P(ref)

A computational technique that systematically introduces virtual mutations into a DNA sequence and measures the resulting change in model predictions. This enables high-throughput identification of nucleotides critical for regulatory function.

• Saturation Mutagenesis: Tests every possible single-nucleotide substitution at a locus
• Use Case: Prioritizing non-coding variants in genome-wide association study (GWAS) loci
• Output: Per-nucleotide importance scores for regulatory activity

A metric derived from a language model's cross-entropy loss that quantifies how surprised the model is by a given sequence. In genomics, lower perplexity indicates the sequence conforms to learned regulatory grammar.

• Interpretation: High perplexity at variant positions suggests functional disruption
• Relationship: Perplexity = exp(cross-entropy); inversely related to log-likelihood
• Application: Identifying constrained non-coding elements across the genome

The computational or experimental process of evaluating the functional impact of every possible single-nucleotide substitution at a given locus. Genomic language models dramatically accelerate this by predicting effects for all 3N possible variants in a single forward pass.

• Computational Advantage: Seconds vs. months for experimental assays
• Output: A complete variant effect map for a regulatory element
• Integration: Scores used to prioritize variants for functional validation

Standardized collections of curated datasets and evaluation protocols designed to rigorously measure variant effect prediction performance. These benchmarks enable fair comparison across different genomic language models.

• Key Datasets: ClinVar, gnomAD constraint metrics, deep mutational scans
• Metrics: AUROC, precision-recall curves, Spearman correlation with functional assays
• Purpose: Establish ground truth for model validation and regulatory acceptance