Variant Effect Prediction

The foundational capability of genomic and protein language models to predict mutation effects without any supervised fine-tuning on labeled variant data. The model computes the likelihood of the wild-type sequence versus the mutated sequence using its pre-trained probability distribution. A significant drop in likelihood indicates a potentially deleterious variant. This approach, exemplified by ESM-1v and EVE, leverages the model's deep understanding of evolutionary sequence constraints learned during self-supervised pre-training on millions of natural sequences.

The mathematical core of zero-shot variant effect prediction. The model calculates the log-likelihood ratio between the alternate allele and the reference allele given the surrounding sequence context. Formally, this is often expressed as: score = log P(mutant | context) - log P(wild-type | context). A negative score indicates the mutation is less probable under the model's learned distribution of functional sequences. Advanced methods use pseudo-log-likelihoods or masked marginal probabilities to score mutations while conditioning on the rest of the sequence.

The underlying mechanism that makes zero-shot prediction possible. During pre-training on diverse natural sequences, the model implicitly learns sequence conservation patterns and co-evolutionary couplings between residues. Positions critical for structure or function exhibit low tolerance for substitution, a signal captured in the model's probability distribution. This allows the model to recognize mutations that violate fundamental biochemical constraints without ever being shown a labeled pathogenic variant.

A systematic application of variant effect prediction where the model computationally introduces every possible single amino acid or nucleotide substitution at each position in a sequence and scores the predicted impact. This generates a comprehensive mutational landscape or heatmap of functional tolerance across the entire protein or regulatory element. The resulting scores correlate strongly with experimental deep mutational scanning data, enabling prioritization of variants for functional validation.

The translational application of variant effect prediction to human genetics. Models like PrimateAI-3D and AlphaMissense leverage protein language model embeddings and structural context to classify missense variants as likely benign or likely pathogenic. These predictions serve as computational evidence under ACMG/AMP guidelines for variant interpretation, helping clinical geneticists prioritize variants of unknown significance (VUS) in rare disease diagnosis and population screening programs.

The process of mapping raw model likelihood scores to clinically interpretable pathogenicity scales. Raw scores are often calibrated against curated databases like ClinVar or gnomAD to produce well-calibrated probabilities. Techniques include:

The foundational paradigm where a pre-trained language model scores mutations without any supervised fine-tuning on labeled variant data. The model computes the log-likelihood ratio between the reference and alternate allele given its surrounding sequence context.

• Relies entirely on evolutionary signals learned during self-supervised pre-training
• A lower likelihood for the mutant sequence implies a deleterious effect
• Eliminates the need for scarce, biased clinical variant databases

A computational technique that systematically introduces every possible single-nucleotide change at each position in a sequence and measures the predicted impact using a pre-trained model.

• Generates a comprehensive effect map for all possible variants in a region
• Used to identify constrained positions where any change is predicted to be damaging
• Enables saturation mutagenesis experiments that are infeasible in a wet lab

A measure of the degree to which a nucleotide or amino acid position remains unchanged across evolutionary time. This is the primary signal that language models learn during self-supervised pre-training.

• Highly conserved positions are typically functionally critical
• Mutations at conserved sites are far more likely to be pathogenic
• Transformer attention patterns implicitly capture conservation without explicit alignment

A visualization of the self-attention weights from a transformer model's layers, used to interpret which specific nucleotides or residues the model focuses on when scoring a variant.

• Reveals potential binding sites or structural contacts
• High attention at a mutated position often correlates with functional importance
• Provides a mechanistic hypothesis for why a variant is predicted to be damaging

A family of large-scale protein language models from Meta AI that uses masked language modeling on millions of evolutionarily diverse sequences. ESM-2 variant scoring is a leading zero-shot method.

• Captures structural, functional, and evolutionary information in learned representations
• The log-likelihood ratio from ESM-2 correlates strongly with experimental deep mutational scanning data
• Scales to 15 billion parameters, improving variant effect resolution

A standardized suite of curated datasets and evaluation protocols for rigorously comparing genomic language model performance on nucleotide-level classification tasks.

• Includes benchmarks for enhancer identification, promoter prediction, and splice site detection
• Provides a controlled framework to assess how well pre-training objectives transfer to variant effect prediction
• Enables reproducible comparison across architectures like DNABERT, Enformer, and the Nucleotide Transformer