Zero-Shot Variant Effect Prediction

The core mechanism computes a variant effect score as the log-ratio of probabilities assigned to the alternate versus reference allele by an autoregressive model. A negative score indicates the variant disrupts learned regulatory grammar, suggesting pathogenicity. This approach requires no labeled training data—only the pretrained model's sequence likelihoods.

• Formula: Score = log P(alt_allele) - log P(ref_allele)
• Captures disruption to splice sites, promoters, and enhancers
• Used by models like EVE and ESM-1v for protein variants

By training on diverse genomes, DNA language models implicitly learn evolutionary conservation patterns. Variants in highly constrained regions produce large negative log-likelihood shifts, serving as a zero-shot signal for purifying selection without requiring multiple sequence alignments.

• Detects ultra-conserved elements and non-coding constraint
• Outperforms phyloP and GERP scores on some benchmarks
• Captures lineage-specific constraints missed by comparative genomics

Unlike position weight matrices, genomic language models generate contextualized embeddings where the representation of a nucleotide depends on surrounding sequence. A variant's effect is assessed by how it alters the attention patterns and hidden states across potentially megabase-scale contexts.

• Captures long-range enhancer-promoter interactions
• Models epistatic effects between distal variants
• Enabled by architectures like Enformer and HyenaDNA

Zero-shot prediction transfers across species boundaries. A model pretrained on human genomes can score variants in mouse or zebrafish by exploiting shared regulatory syntax. This enables functional annotation in non-model organisms where labeled training data is scarce or nonexistent.

• Leverages deep homology of regulatory grammar
• Validated for primate, rodent, and plant genomes
• Reduces need for species-specific fine-tuning

Robust zero-shot scoring requires models to produce strand-agnostic predictions. Techniques like reverse complement augmentation during pretraining ensure that a variant and its reverse complement yield identical effect scores, reflecting the double-helical nature of DNA and preventing strand-biased artifacts.

• Enforces Watson-Crick parity in predictions
• Critical for accurate indel scoring near palindromic sequences
• Improves calibration of variant effect scores

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

• Calculation: Score = log P(alt | context) - log P(ref | context)
• Negative scores indicate the alternate allele is less likely, suggesting a deleterious effect
• Positive scores suggest the variant is more probable than the reference
• Used to prioritize variants in clinical sequencing pipelines without requiring labeled training data

The probability assigned to a genomic sequence by an autoregressive model, quantifying how well the sequence conforms to learned patterns of natural DNA.

• Serves as the foundation for zero-shot variant scoring
• A low likelihood on the alternate allele relative to the reference signals potential pathogenicity
• Captures evolutionary constraint without explicit conservation analysis
• Enables ranking of variants by their deviation from the model's learned genomic grammar

A computational technique that systematically introduces virtual mutations into a DNA sequence and measures the resulting change in model predictions.

• Saturation mutagenesis: testing all possible single-nucleotide substitutions at every position
• Identifies nucleotides critical for regulatory function without wet-lab experiments
• Genomic language models accelerate this by scoring thousands of variants in parallel
• Used to map transcription factor binding sites and splice regulatory elements

A metric derived from a language model's cross-entropy loss that quantifies how surprised the model is by a given sequence.

• Lower perplexity indicates the sequence fits the model's learned distribution
• A variant that increases perplexity suggests disruption of functional sequence patterns
• Provides a normalized, interpretable measure of sequence constraint
• Enables comparison of variant effects across different genomic contexts and model architectures

Standardized collections of curated datasets and evaluation protocols designed to rigorously measure the performance of genomic language models on zero-shot variant effect prediction.

• Includes datasets like ClinVar, gnomAD constraint metrics, and deep mutational scans
• Provides ground-truth labels for benchmarking without requiring model fine-tuning
• Enables fair comparison across architectures: DNABERT, HyenaDNA, Enformer, and Mamba-based models
• Critical for validating that zero-shot predictions correlate with experimental functional assays

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 computing likelihoods for all 3 alternate alleles per position in a single forward pass
• Produces variant effect maps that reveal regulatory grammar at nucleotide resolution
• Used to interpret non-coding variants in promoters, enhancers, and splice sites
• Bridges the gap between statistical association and mechanistic understanding of disease variants