FuncVEP: functional training yields best-performing missense variant predictor, uncovers new disease links

A new family of variant effect predictors, FuncVEP, trained exclusively on experimental functional data rather than clinical labels, outperforms 48 existing tools across multiple benchmarks, scientists report today in Nature Genetics. The work, led by Tayfun Özçelik at Bilkent University, with first author Barış Kayaalp, shows that learning from direct measures of protein function produces more generalizable and accurate predictions—improving accuracy on functional benchmarks from 78.8% to 84.6% and on clinical benchmarks from 90.1% to 92.4%.

Missense variants, which change a single amino acid in a protein, account for nearly half of protein-coding variation but most are classified as variants of uncertain significance (VUS). Existing predictors often rely on clinical outcomes or population allele frequencies, introducing data circularity and limited generalizability. FuncVEP sidesteps these issues by training on balanced, high-quality functional data from multiplexed assays of variant effect (MAVEs), curated literature, and proxy-benign variants from population databases.

The models integrate 580 features spanning conservation, structural, protein language model (PLM) outputs, and existing VEP scores. Three versions were created: CTI (includes all available VEPs), CTE (excludes clinically trained predictors and AlphaMissense), and SP (no VEP-derived features). Across independent benchmark categories—including clinical, functional, deep mutational scanning, de novo mutations in developmental disorders, and cancer hotspot mutations—all three FuncVEP models ranked at the top overall. FuncVEP-CTI achieved the highest overall rank percentile (mean 1.4), followed by CTE (2.6) and SP (6.5).

Importantly, FuncVEP also demonstrated translational utility. In a phenome-wide association study (PheWAS) of 494 genes linked to inborn errors of immunity in the UK Biobank, the study identified 210 new gene–phenotype associations across all tests, with FuncVEP-CTE identifying 17.7% more new associations than the best competitor (CPT-1). Of these, 101 reached genome-wide significance or were supported by independent evidence. The findings also revealed the genetic epidemiology of these disorders, estimating that 58% of individuals carry at least one pathogenic or likely pathogenic variant across IEI genes.

Limitations: FuncVEP currently classifies variants only as damaging or neutral, without specifying the direction of effect (e.g., loss- vs. gain-of-function) or the magnitude of impact. It is a binary classifier, not a regression model.

The authors suggest that functionally trained predictors like FuncVEP could become a central paradigm for missense variant interpretation, improving both diagnostic precision and the discovery of genotype–phenotype relationships.

Reference:
Kayaalp, B., Çil, K., Conil, C. et al. Prediction of human missense variant effects from functional evidence. Nat Genet (2026). https://doi.org/10.1038/s41588-026-02727-3