A new artificial-intelligence-based approach could make one of genetics’ most difficult problems substantially easier: determining which single-letter changes in DNA are harmless and which can disrupt human health. The method, called FuncVEP, predicts the functional effects of missense variants—genetic alterations that change one amino acid in a protein—and has outperformed 48 existing prediction tools across a broad series of tests. In a study published in Nature Genetics, researchers report that FuncVEP raised accuracy on functional benchmarks from 78.8 percent to 84.6 percent, while improving performance on clinical benchmarks from 90.1 percent to 92.4 percent. The system also helped identify 210 previously unrecognized gene–phenotype associations involving 494 genes connected to inborn errors of immunity, suggesting that functional prediction tools could accelerate both diagnosis and the discovery of disease biology.
Missense variants are among the most common forms of genetic variation. A change in a DNA letter can replace one amino acid with another in a protein, potentially altering the protein’s shape, stability, location or interaction with other molecules. Yet the majority of missense variants found in human genomes have uncertain consequences. Some have no measurable effect, while others interfere with essential biological processes and contribute to inherited disorders, cancer susceptibility or immune dysfunction. For clinicians interpreting genome-sequencing results, this uncertainty can stand between a patient and a molecular diagnosis. A variant may appear suspicious because it is rare or occurs in a biologically important gene, but those clues alone do not establish that it causes disease. FuncVEP was designed to address this bottleneck by focusing on direct evidence of what variants do to cellular or molecular function rather than relying primarily on whether they have previously been observed in patients.
Many existing variant-effect predictors draw heavily on clinical records, disease databases, evolutionary conservation or patterns of variation in human populations. These sources are valuable, but they can introduce limitations. Clinical databases may contain incomplete or unevenly classified information, while population-based methods often assume that variants depleted from healthy populations are harmful. Predictors trained on clinical labels can also suffer from data circularity: the same evidence, or closely related evidence, may influence both the training data and the benchmark used to evaluate performance. In that situation, a tool can appear highly accurate partly because it has learned features of existing annotations rather than general biological principles. The researchers instead trained the FuncVEP family on diverse functional data, aiming to connect predictions more directly to experimental measurements of variant impact and improve the method’s ability to generalize to datasets it had not encountered before.
Functional evidence can take many forms, depending on the gene and the biological system being studied. Experiments may measure whether a mutated protein remains stable, reaches the correct cellular compartment, binds its molecular partners or performs a biochemical reaction. Other assays can examine how a variant affects the survival or behavior of cells carrying it. These measurements are not identical: an amino-acid substitution that destabilizes a protein may produce a different signal from one that leaves the protein intact but blocks an interaction. By bringing together varied functional datasets, a predictor can learn recurring relationships between amino-acid changes and biological consequences. The challenge is that functional experiments are often generated in different laboratories, using different technologies, cell types and scoring systems. FuncVEP’s reported performance indicates that the approach was able to extract useful signals across this heterogeneity rather than becoming narrowly tuned to one experimental platform.
The strongest gains appeared on functional benchmarks, where accuracy increased from 78.8 percent for existing state-of-the-art approaches to 84.6 percent with FuncVEP. That improvement is important because functional benchmarks test a question close to the core of variant interpretation: does the altered sequence measurably change biological activity? The method also improved results on clinical benchmarks, rising from 90.1 percent to 92.4 percent. Clinical prediction remains essential because the ultimate goal of variant interpretation is not simply to classify laboratory measurements, but to help determine whether a genetic change contributes to a patient’s condition. The two types of benchmarks answer related but different questions. Functional tests assess molecular or cellular consequences, whereas clinical tests compare predictions with disease-associated evidence. Better performance in both categories suggests that FuncVEP’s functional grounding did not come at the expense of clinical relevance, although no computational prediction can replace patient history, family data and direct clinical assessment.
The researchers then used FuncVEP in a discovery setting involving inborn errors of immunity, a broad group of inherited disorders that impair immune development or immune responses. These conditions can produce recurrent infections, inflammatory disease, unusual susceptibility to particular pathogens or problems affecting multiple organs. Many remain difficult to diagnose because their genetic causes are diverse and because the same gene can produce different clinical presentations. By applying the predictor to human genetic and health data from the UK Biobank and the Mount Sinai Million Health Discoveries Program, the team identified 210 new gene–phenotype associations involving 494 genes linked to these immune disorders. A gene–phenotype association does not automatically prove that a variant causes disease; rather, it points to a statistically or biologically supported relationship that can be investigated further. Even so, the scale of the reported discoveries suggests that functional prediction can help researchers prioritize genes and variants that might otherwise remain hidden in large genomic datasets.
This discovery potential is particularly relevant as sequencing projects expand. Biobanks now contain genetic information from hundreds of thousands of people, often paired with electronic health records, laboratory measurements and diagnoses. Those resources can reveal relationships between rare variants and traits, but the signal is difficult to detect when most variants have unknown functional effects. A computational tool that ranks variants according to their likely biological impact can reduce the number of candidates requiring experimental follow-up. It may also help expose connections between genes and diseases that are too uncommon, variable or newly recognized to have accumulated extensive clinical documentation. According to the study, FuncVEP substantially improved the discovery rate compared with leading existing predictors. That result positions the method not only as a diagnostic aid, but also as a way to generate hypotheses about disease mechanisms and identify potential targets for future laboratory studies.
The findings do not mean that every high-scoring variant is pathogenic or that every low-scoring variant is benign. Biological systems are context-dependent, and a variant’s effect can differ among tissues, developmental stages or environmental conditions. Functional assays themselves may not perfectly reproduce what happens in a human body, while clinical datasets can contain misdiagnoses, incomplete records and ancestry-related biases. Predictions should therefore be integrated with segregation analysis, population frequency, evolutionary information, laboratory experiments and the patient’s phenotype. The principal advance reported for FuncVEP is its use of diverse functional evidence to make predictions that generalize across datasets while avoiding some of the circularity associated with purely clinical training. If its performance continues to hold in independent populations and prospective diagnostic studies, the approach could help turn a growing flood of genomic variation into more actionable biological knowledge—and bring many patients with unexplained inherited disease closer to a precise diagnosis.
Subject of Research: Prediction of the functional effects of human missense genetic variants and discovery of gene–phenotype associations linked to inborn errors of immunity
Article Title: Prediction of human missense variant effects from functional evidence
Article References: Kayaalp, B., Çil, K., Conil, C. et al. “Prediction of human missense variant effects from functional evidence.” Nature Genetics (2026). https://doi.org/10.1038/s41588-026-02727-3
Image Credits: AI Generated
DOI: 10.1038/s41588-026-02727-3
Keywords: FuncVEP, missense variants, variant interpretation, functional genomics, genetic diagnosis, inborn errors of immunity, gene–phenotype associations, artificial intelligence
Tags: advances in genomics and personalized medicineAI-based genetic variant predictionclinical genetic diagnosisdisease genetics and mutation impactfunctional impact of missense variantsgene–phenotype association discoverygenetic variation and human healthgenetic variation effect assessmentgenetic variation predictioninborn errors of immunityperformance of variant prediction toolsprotein structure and stability prediction
