Variant Effect Prediction

Leverages multiple sequence alignments (MSAs) to measure how conserved a residue position is across homologous proteins. The core principle: positions critical for function or stability tolerate very few substitutions over evolutionary time.

• Position Weight Matrices (PWMs) quantify amino acid frequencies at each site
• Tools like SIFT predict a substitution is deleterious if the variant amino acid is rarely observed in the alignment column
• PROVEAN extends this by scoring the impact of the variant on the alignment score of the surrounding sequence neighborhood
• Strength: computationally efficient and interpretable
• Limitation: struggles with proteins that have few known homologs

Models trained on labeled datasets of known pathogenic and benign variants, typically sourced from databases like ClinVar and HGMD. These classifiers integrate diverse features to make predictions.

• PolyPhen-2 uses a Naive Bayes classifier with features including sequence conservation, structural parameters (solvent accessibility, crystallographic B-factors), and Pfam domain annotations
• CADD (Combined Annotation Dependent Depletion) trains a support vector machine (SVM) on a diverse feature set, contrasting fixed alleles in the human lineage against simulated de novo mutations to generate a C-score
• REVEL uses a random forest trained on rare missense variants, combining scores from multiple individual tools (SIFT, PolyPhen, MutationTaster, etc.) to produce an ensemble prediction
• These methods excel at integrating heterogeneous data but are limited by training label quality and potential circularity

Self-supervised models like ESM-1v and EVE learn the distribution of amino acids at each position from millions of unaligned protein sequences, without requiring explicit MSAs.

• ESM-1v (a variant of ESM-2) uses masked language modeling: it predicts the probability of an amino acid at a given position given its sequence context. A low probability for the variant allele indicates a deleterious effect
• EVE (Evolutionary Model of Variant Effect) uses a deep variational autoencoder (VAE) to learn the distribution of functional sequences across an entire protein family, outputting an evolutionary index that quantifies how likely a variant sequence belongs to the functional family
• Tranception combines a protein language model with a retrieval mechanism over MSAs for state-of-the-art performance
• Key advantage: captures higher-order epistatic dependencies beyond pairwise coevolution

Calculates the change in Gibbs free energy of folding (ΔΔG) caused by a point mutation. A significant destabilization (ΔΔG > 2 kcal/mol) is a strong indicator of pathogenicity.

• FoldX uses an empirical force field to compute the energy difference between wild-type and mutant structures, accounting for van der Waals clashes, solvation, hydrogen bonding, and entropy changes
• Rosetta ddg_monomer applies the Rosetta all-atom energy function with conformational sampling of the mutated side chain and local backbone relaxation
• ThermoMPNN leverages the inverse folding model ProteinMPNN to predict stability effects by assessing how well the variant sequence fits its structural environment
• These methods require a high-quality 3D structure (experimental or predicted via AlphaFold) and are computationally more expensive than sequence-only approaches

High-throughput experimental assays that measure the functional effect of thousands of variants in parallel provide the gold-standard training and benchmarking data for predictors.

• A DMS experiment generates a variant effect map: a quantitative score (e.g., growth rate, fluorescence) for nearly every possible single amino acid substitution in a protein
• Datasets like ProteinGym aggregate DMS results across hundreds of proteins, enabling rigorous benchmarking of computational methods
• Models like GEMME and DeepSequence are unsupervised generative models trained directly on natural sequence variation, but validated against DMS data to confirm they recapitulate experimental fitness landscapes
• The correlation between predicted scores and DMS measurements (Spearman's ρ) is the primary evaluation metric for variant effect predictors

Combine the outputs of multiple orthogonal methods to improve accuracy and robustness, mitigating the biases of any single approach.

• REVEL is a random forest meta-predictor that integrates 13 individual tools (including SIFT, PolyPhen, MutationTaster, and FATHMM) with allele frequency data
• EVEscape combines the EVE evolutionary index with structural features (antibody accessibility, binding interface proximity) to predict viral immune escape mutations, as demonstrated during the COVID-19 pandemic
• AlphaMissense fine-tunes AlphaFold's internal representations on population frequency data from gnomAD and ClinVar labels to directly predict pathogenicity, achieving state-of-the-art performance without relying on MSA conservation
• Meta-predictors are the preferred approach for clinical variant interpretation guidelines from ACMG/AMP

A high-throughput experimental technique that measures the functional effect of thousands of single amino acid substitutions in parallel. DMS combines saturation mutagenesis, a selection pressure for function, and deep sequencing to quantify variant enrichment or depletion. The resulting genotype-phenotype maps serve as the gold-standard training and benchmarking datasets for supervised variant effect predictors, enabling models to learn the sequence-function landscape directly from empirical data.

A foundational approach to variant effect prediction based on the principle that functionally critical residues are conserved across evolution. Methods like GERP++ and phastCons quantify constraint at each genomic position by comparing the observed substitution rate to a neutral expectation. A high conservation score at a variant position strongly suggests the reference amino acid is functionally important and a substitution is likely deleterious. These scores are often used as input features for ensemble predictors.

Ensemble classifiers that integrate the outputs of multiple individual variant effect tools into a single, more robust score. Tools like REVEL, CADD, and EVE combine conservation, biochemical, structural, and population frequency data. By training on diverse features, meta-predictors mitigate the biases of any single method and consistently outperform individual tools in clinical variant interpretation benchmarks, making them essential for ACMG/AMP guideline-based classification.

A modern paradigm where large-scale models like ESM-2 and ProtTrans, trained on hundreds of millions of protein sequences, generate dense vector representations that implicitly capture structure and function. The log-likelihood ratio of the reference vs. alternate amino acid at a given position, as scored by the pLM, serves as a zero-shot variant effect predictor. This approach requires no experimental training data and has shown competitive performance with supervised methods.

A biophysical approach that predicts the change in Gibbs free energy of folding upon mutation. Tools like FoldX and Rosetta ddg_monomer use physics-based force fields and statistical potentials to model the energetic impact of a substitution on the folded state. A large positive ΔΔG indicates strong destabilization, which is a common mechanism of pathogenicity. This method is particularly powerful for interpreting variants in proteins with known 3D structures.

A critical population genetics filter that uses large-scale databases like gnomAD to identify variants that are too common in the general population to cause a rare Mendelian disease. The underlying assumption is that highly penetrant pathogenic variants are purged by purifying selection. A variant with a minor allele frequency (MAF) above a disease-specific threshold is strong evidence for a benign classification, making this a foundational step in any variant interpretation pipeline.