New AI Model Revolutionizes Ranking of Genetic Disease Variants

A groundbreaking study published in Nature Genetics has introduced an innovative AI model called popEVE, designed to rank genetic variants based on their potential to cause severe or mild disease. By combining deep evolutionary signals with data from human populations, researchers are providing a new tool for clinicians to better prioritize genetic variants in cases that remain unresolved.

The study addresses a critical challenge faced by many families with rare diseases. Approximately one in four individuals diagnosed with rare conditions receives a genetic diagnosis, even after undergoing whole-exome sequencing (WES). This leaves many families searching for answers while clinicians navigate through millions of variants in each genome. Current computational tools typically assess changes within a single gene, making it difficult to gauge the overall severity of a variant across different proteins.

Enhancing Variant Effect Prediction

The popEVE model integrates evolutionary evidence with human population constraints to facilitate the ranking of missense variants. This approach allows clinicians to understand the organism-level impact of previously unseen variants, which can significantly aid in diagnosing rare diseases. The model was developed using two advanced protein models: the Evolutionary Model of Variant Effect (EVE), a Bayesian variational autoencoder trained on multiple sequence alignments, and the Evolutionary Scale Modeling 1 variant (ESM-1v), a large language model trained on protein sequences.

Researchers introduced a population constraint through a latent Gaussian process that correlates evolutionary scores with missense intolerance, utilizing data from the United Kingdom Biobank (UKBB) and the Genome Aggregation Database (gnomAD). To address potential ancestry bias, the model focused on presence or absence indicators rather than allele frequencies.

The performance of popEVE was rigorously benchmarked against leading predictors, including AlphaMissense, Bayesian Deleteriousness (BayesDel), and Rare Exome Variant Ensemble Learner (REVEL). Using ClinVar labels and deep mutational scans, the researchers validated popEVE’s effectiveness in identifying pathogenic variants among large cohorts, including approximately 31,000 trios from a severe developmental disorder (SDD) metacohort.

Implications for Rare Disease Diagnosis

The results from the study indicate that popEVE outperforms existing tools in capturing disease severity. For instance, pathogenic variants linked to childhood mortality exhibited more deleterious scores compared to those associated with adult mortality. In the SDD metacohort, de novo missense (DNM) scores indicated a significant shift towards higher deleteriousness when compared to unaffected siblings. A Gaussian mixture defined severity thresholds, with a severe cut-off established at −5.056, representing a 99.99% likelihood of being deleterious.

The model also demonstrated its ability to accurately prioritize likely causal variants without requiring parental genomes. Among the 513 individuals with severe DNMs, 98% had their variant ranked as the most deleterious in their exome. This prioritization method successfully identified 410 candidate genes within the SDD cohort, recovering a substantial 94% of previously reported missense-identified genes while also uncovering 123 novel candidates.

The novel variants identified by popEVE had not been recorded in the UK Biobank or gnomAD, and functional analyses indicated that these genes interact with known genes associated with developmental disorders. Structure mapping further supported the findings, revealing that 91% of severe substitutions were in close proximity to interaction partners.

Overall, the popEVE model signifies a substantial advancement in the field of clinical genetics. By integrating deep evolutionary insights with human population data, it provides a calibrated, proteome-wide assessment of missense variant severity. This innovative approach not only distinguishes between childhood-lethal and adult-onset pathogenicity but also enriches the identification of truly damaging variants in severe developmental disorder cohorts.

As genetic sequencing technologies continue to expand globally, the application of severity-aware, minimally biased scoring systems like popEVE has the potential to enhance diagnostic processes, improve counseling, and facilitate research efforts. This could ultimately lead to faster and clearer answers for families affected by rare diseases, making a significant impact on the landscape of genetic medicine.

The full study can be accessed in the March 2025 edition of Nature Genetics, authored by Orenbuch, R., Shearer, C. A., Kollasch, A. W., and their colleagues.