For thousands of families with a child who has a severe, undiagnosed developmental disorder, the search for an answer often ends in a folder of inconclusive genetic test results. On November 24, 2025, researchers at Harvard Medical School published a study in Nature Genetics describing an AI model called popEVE that is starting to close some of those files, having flagged 123 previously unlinked genes as likely causes of disease in a cohort of roughly 30,000 undiagnosed patients with severe developmental disorders.
A Score for Every Variant in the Genome
PopEVE, built by a team led by Debora Marks, a professor of systems biology at Harvard’s Blavatnik Institute, along with lead author Rose Orenbuch and co-senior authors Mafalda Dias and Jonathan Frazer, works by assigning every possible variant in a person’s genome a score reflecting how likely it is to cause disease. Rather than analyzing one gene in isolation, the model draws on evolutionary comparisons across hundreds of thousands of species, protein language models trained on massive sequence databases, and patterns of genetic variation across human populations, then places each variant on a continuous spectrum from clearly benign to clearly pathogenic.
Why Rare Disease Diagnosis Has Been So Slow
An estimated 30 million Americans live with a rare disease, and for many families the road to a diagnosis is what geneticists have long called a diagnostic odyssey, often stretching across five to seven years and multiple specialists. Traditional variant-interpretation tools work reasonably well for genes that have already been studied extensively, but they struggle with genes where only a handful of cases have ever been documented worldwide, precisely because there is too little data for classical statistical methods to draw firm conclusions. That gap has left an estimated half of all suspected genetic disease cases unsolved even after whole genome sequencing.
What the Study Actually Found
Applying popEVE to the developmental disorder cohort, the Harvard team identified 123 genes not previously linked to any disease as strong candidates for causing the developmental disorders seen in patients carrying rare variants in them. Twenty-five of those candidate gene associations have already been independently confirmed by other research groups working with separate patient cohorts, according to the study, lending early external validation to the model’s predictions. The researchers estimate the approach could help resolve a diagnosis in roughly one-third of cases within the broader 30,000-patient cohort that had gone unexplained.
Two Views on How Much This Changes Clinical Practice
Genomics researchers who reviewed the findings describe popEVE as a meaningful advance precisely because it targets the genes classical tools ignore, the ultra-rare ones where a family might be the only documented case on record. Independent confirmation of 25 candidate genes by outside labs is being cited as evidence the model is finding real biology rather than statistical noise. Clinical geneticists are more cautious, noting that a computational prediction is not the same as a confirmed diagnosis, and that variants flagged by popEVE still require functional laboratory validation and correlation with a patient’s actual symptoms before a family can be told with confidence what is causing their child’s condition. That validation pipeline, they note, remains slow and underfunded relative to the pace at which AI models can now generate candidate genes.
Where the Research Goes From Here
The Marks Lab has made popEVE’s methodology public, and the team says it is working to extend the model beyond developmental disorders into other categories of rare genetic disease, while academic medical centers begin testing how to integrate variant scores like these into existing genetic counseling workflows. For the families still waiting on answers, the practical test will be whether hospitals and diagnostic labs can build the functional-validation infrastructure needed to turn a computational candidate gene into a confirmed diagnosis quickly enough to matter, rather than adding one more inconclusive report to an already thick file.
