A new artificial intelligence model, popEVE, could accelerate diagnoses for rare, single-variant genetic diseases by scoring how likely each genetic change in a patient’s genome is to cause disease. Developed by Harvard Medical School researchers and collaborators, popEVE places variants on a continuous spectrum of severity and, in testing, distinguished benign from pathogenic variants, flagged those linked to childhood versus adult mortality, and helped resolve previously undiagnosed cases.

PopEVE builds on EVE, a generative model that learns evolutionary constraints to predict how mutations affect protein function. To make variant scores comparable across different genes—a key barrier for clinical triage—the team layered in a large-language protein model trained on amino acid sequences and human population variation data. This calibration enables cross-gene comparisons that reflect both functional impact and relevance to human physiology.

In a Nature Genetics study published Nov. 24, the model demonstrated several capabilities:

• Differentiated pathogenic from benign variants and separated healthy controls from patients with severe developmental disorders.
• Predicted age-of-onset lethality (childhood versus adulthood) and inferred whether variants were inherited or de novo without parental genomes.
• Avoided ancestry bias and did not inflate pathogenicity rates in underrepresented groups.

Applied to a cohort of roughly 30,000 undiagnosed patients with severe developmental disorders, popEVE supported diagnoses in about one-third of cases. Notably, it implicated variants across 123 genes not previously tied to these disorders; 25 of those gene–disease links have since been independently confirmed. The model also surfaced more than 100 novel disease-causing alterations, highlighting its potential for discovery alongside diagnosis.

Clinically, the vision is a prioritized, interpretable genome readout that guides which variants to investigate first, improving the efficiency of genetic workups that often stall on variants of uncertain significance. Co-senior author Debora Marks emphasized the goal of ranking variants by disease severity to offer a clinically meaningful genome view, while lead author Rose Orenbuch underscored value for patients who elude standard diagnostic pipelines.

PopEVE is accessible via an online portal that visualizes variant scores from most to least likely disease-causing and maps them onto protein structures. The team is collaborating with groups including the Children’s Rare Disease Collaborative at Boston Children’s Hospital, CHOP’s Division of Human Genetics, and Genomics England with the Wellcome Sanger Institute. Early clinical use at Barcelona’s Centro Nacional de Análisis Genómico has already aided several rare-disease diagnoses.

To foster broad adoption, researchers are integrating popEVE scores into resources such as ProtVar and UniProt, enabling global comparison of variants across genes. While further validation is needed before routine clinical deployment, the developers anticipate that popEVE will boost clinician confidence in computational interpretation, shorten diagnostic odysseys, and illuminate new therapeutic targets.

By prioritizing variants based on predicted disease severity, popEVE seeks to improve the odds of timely diagnosis—and open pathways for better treatments and drug discovery for patients with rare genetic conditions.