Background

With the advancement of next-generation sequencing technology and the reduction in the cost of human genome decoding, large-scale genomic variant data are now widely available. Among these, missense variants, which alter a single amino acid sequence, are considered key targets in precision medicine due to their high potential to cause disease. However, the majority of genetic variants encountered in clinical settings remain classified as variants of uncertain significance (VUS) because their disease-causing potential is ambiguous.

To address this issue, various computational models have been developed to assess pathogenicity. However, most existing prediction tools rely heavily on clinical reports and population frequency data from resources such as ClinVar. This training method can lead to circularity bias, where the performance of the model is overestimated when the new variants analyzed overlap with the training data. As a result, the predictive accuracy significantly drops when diagnosing rare variants or newly identified disease genes that lack clinical records. This highlights the need for a new evaluation standard that objectively assesses variant harm based on biological mechanisms, rather than clinical bias.

Key Findings

An international research team from Bilkent University in Turkey and Rockefeller University in the United States introduced FuncVEP, an artificial intelligence model trained solely on biological functional evidence, rather than clinical outcomes. To eliminate circularity bias at its source, the researchers constructed a dataset of 18,556 missense variant functional data using large-scale functional variant analysis techniques (MAVE), expert-reviewed literature, and text mining. This training dataset includes 9,357 functionally impaired variants, 2,558 neutral variants, and 6,641 alternative benign variants selected from the gnomAD genome database.

The research team applied a lightweight gradient boosting (LightGBM) classifier structure to this function-centered dataset, incorporating a total of 571 biological features. In real-world evaluations, FuncVEP outperformed 47 existing publicly available variant prediction tools. While the average classification accuracy of existing tools was around 82%, FuncVEP achieved an impressive accuracy of 93%. Additionally, it successfully reduced the proportion of variants classified as VUS from 11% to 2%.

To demonstrate the practical utility of FuncVEP, the research team focused on 490 genes associated with inborn errors of immunity. Analysis of the UK Biobank and Mount Sinai Million Health Discoveries Program cohort data using this model revealed 50 previously unknown gene–phenotype associations. This finding demonstrates that disease-causing variants can be accurately traced using only predicted protein function loss, without prior clinical interpretation of the gene.

Significance and Outlook

FuncVEP provides a foundation for shifting the current variant interpretation paradigm from clinical data-centric to function-based evidence. Clinicians can use this model to make faster diagnostic decisions, as the prediction uncertainty is significantly reduced. Furthermore, it is expected to contribute to the development of a high-precision variant classification system that integrates computer-based prediction criteria (PP3 and BP4) and function-based evidence (PS3 and BS3) as outlined in the ACMG classification guidelines.

However, challenges remain in making this tool universally applicable. Since the model heavily relies on protein structure information and large-scale functional analysis data, its performance may be less stable when predicting rare protein variants that lack MAVE studies or functional validation data. For future clinical applications of AI-based predictions, it will be essential to complement this model with a clinical decision-making system that integrates individual patient clinical profiles and pedigree analysis results in a coordinated manner.