Variant Effect Predictors

Computational approaches can leverage variant effect data for functional interpretation

Are we missing a predictor? Click here to suggest a coding or non-coding VEP for inclusion in the tables.

What is a variant effect predictor?

A variant effect predictor (VEP) is a computational tool that estimates how a genetic variant may affect molecular properties, biological function, or disease risk.

These tools have often been called by different names, including variant impact predictors, pathogenicity predictors, deleteriousness predictors, functional effect predictors, and variant scoring methods. Here, we use VEP as a broad umbrella term that includes methods for predicting diverse molecular or clinical consequences of genetic variants. Note that this usage is distinct from Ensembl VEP, which annotates variants with information such as affected genes, transcripts, and predicted molecular consequences, and can also report scores from some external variant effect prediction methods. In contrast, here we use VEP to refer more generally to computational methods that assign scores or predictions about variant effects.

Coding VEPs

Predict the effects of variants in protein-coding genes, especially missense variants that alter amino-acid sequence.
Some methods produce general scores intended to rank variants from tolerated/benign to deleterious/pathogenic, while others focus on a specific molecular property, such as the effect of a substitution on protein stability.
May use evolutionary conservation, protein structure, protein language models, clinical labels, population variation, experimental variant effect maps, or combinations of these.
Many are human-specific but some can be run for variants from different organisms.

Non-coding VEPs

Some non-coding VEPs provide general scores of constraint, deleteriousness, or likely functional impact, analogous to many coding VEPs.
Other non-coding VEPs belong to the class of sequence-to-function (S2F) models, which predict specific molecular consequences. These may include effects on transcription factor binding, enhancer or promoter activity, chromatin accessibility, chromatin contacts, RNA processing, splicing, gene expression, or other functional readouts.
Because many non-coding effects are context-dependent, predictions often depend on the relevant cell type, tissue, developmental stage, assay, genomic window, or molecular readout.
Most are human-specific.

Choosing a VEP and interpreting its outputs

When choosing a VEP, first check what question it was designed to answer. A predictor trained to recognise clinically labelled missense variants is not equivalent to a model predicting enhancer activity in a specific cell type, even if both produce variant-level scores.

Useful checks before using a VEP:

What variant classes, genes, tissues, cell types, or genomic regions does it cover?

What molecular consequence does the score represent?

Are predictions precomputed, available through a web interface, or generated by running software locally?

What training data were used, and is there a risk of circularity with the benchmark or downstream use case?

Has the method been benchmarked independently?

Does the licence allow the intended use, especially for redistribution, clinical use, or training other models?

VEPs as clinical diagnostic evidence

VEPs have traditionally been used to prioritise candidate variants and, under the ACMG/AMP framework, as supporting computational evidence for variant classification. This cautious use reflected differences in predictor performance, limited validation in specific clinical contexts, and the risk that multiple predictors may not provide independent evidence.

More recently, ClinGen SVI recommendations have provided a framework for calibrating computational evidence using likelihood ratios, allowing some missense VEPs to contribute stronger evidence. Some predictors can achieve moderate or strong evidence levels (PP3) for pathogenicity and correspondingly for benignity (BP4) when their scores exceed predefined thresholds (Pejaver et al, 2022; Bergquist et al, 2025) derived from large, independent validation datasets.

Obtaining VEP predictions

The predictor tables should indicate, where possible, how predictions can be obtained:

Web interface or API

Precomputed score downloads

Locally installed software, code, or model weights

For coding VEPs, precomputed scores are often available for many or all possible human missense variants. These downloads may be indexed by genomic coordinate, transcript, RefSeq, UniProt, or other identifiers, and may use different genome builds.

For non-coding VEPs, precomputed results may be less common or may cover only specific variant classes, tissues, cell types, genomic regions, or molecular readouts. Users should check the input format, genome build, reference annotation, required sequence context, co0mputational requirements, and licensing before using predictions downstream.

Releasing a VEP

Our ‘Guidelines for releasing a variant predictor’ paper provides recommendations for developers of new VEPs, including how to share methods, code, training data, precomputed scores, documentation, licences, and benchmarking information. The aim is to make VEPs easier to evaluate independently,

Helpful resources

ProteinGym: A benchmark for protein fitness prediction and design, including functional assay components.
dbNSFP: A database of precomputed predictions and conservation metrics for many non-synonymous variants.
VIPdb: A broad catalogue of variant interpretation and prediction tools.
CAGI: A community experiment for assessing progress in genome interpretation.

This resource was compiled by Ben Livesey and Tesni Walsh from the lab of Joe Marsh (University of Edinburgh), working with the AVE Analysis, Modelling and Prediction (AMP) Workstream.