Imagine that your DNA is like a vast library full of letters, hiding many small changes—so-called variants. Every person has tens of thousands of them, and most are harmless, like a minor typo in a book that changes nothing. But every now and then, one appears that causes real chaos, such as a rare disease. How can these troublesome variants be found in such an enormous stack? This is where a new AI model called popEVE, developed by scientists at Harvard Medical School, comes in. It is like a super-detective that searches the genome and says: "This variant is innocent, but this one could cause serious problems!"
The popEVE model can calculate a score for every variant in a patient’s genome—a probability that this particular change will cause disease. It is not simply a black-and-white decision about whether a variant is bad or good, but rather a continuous scale ranking variants from harmless ones to those that can lead to serious illnesses or even death in childhood or adulthood. The scientists described it in the journal Nature Genetics and showed that popEVE found more than 100 new alterations responsible for undiagnosed rare genetic diseases. That is like finding lost treasures in a genetic ocean!
How was popEVE created, and what can it do?
It all began with the EVE model, developed in Debora Marks’s laboratory several years ago. EVE learns from deep evolutionary information across different species—it draws lessons from how mutations behave in nature across species. But EVE had a weakness: it could not easily compare variants across different human genes. So the scientists, including Rose Orenbuch, added two new things to the mix. The first is a large language model for proteins (large-language protein model), which learns from the sequences of amino acids that make up proteins. The second is human population data that captures natural genetic variability.
As a result, popEVE not only sees how a variant affects protein function, but also how important it is to a person’s overall health. Debora Marks, a professor of systems biology at the Blavatnik Institute at HMS, says the goal was to create a model that ranks variants by disease severity and provides a clinically meaningful view of a patient’s genome. And guess what? It works! When tested on known variants and case studies, popEVE successfully distinguished between pathogenic (disease-causing) and benign (harmless) variants. It was able to distinguish healthy people from those with severe developmental disorders, determine whether a variant would cause death in childhood or adulthood, and even recognize whether a variant was inherited or occurred randomly—all without any information about the parents.
What is particularly impressive is that the model shows no bias against people from underrepresented genetic groups and does not overestimate the number of pathogenic variants. Rose Orenbuch explains that it combines information across species and within the human population, providing a clear picture of a variant’s impact on physiology.
Discovering new genes with popEVE
The scientists took popEVE and applied it to a group of about 30,000 patients with severe developmental disorders who had not yet received a diagnosis. These diseases appeared to be genetic and caused by a single variant, but no one had found that variant. PopEVE provided a diagnosis in about one-third of the cases! And best of all, it identified variants in 123 genes associated with developmental disorders that had not previously been known. Of these, 25 genes have since been confirmed by other laboratories as causes of these disorders.
The team is now working to bring popEVE into clinical practice. They are collaborating with organizations such as the Children’s Rare Disease Collaborative at Boston Children’s Hospital. A clinic at the Centro Nacional de Análisis Genómico in Barcelona is already using popEVE to interpret patient variants, and it has helped produce several rare disease diagnoses.
What is next for popEVE?
The scientists are now integrating scores from popEVE into databases such as ProtVar and UniProt so that scientists around the world can use them. The model is available online through a portal where variants can be viewed in graphs. Rose Orenbuch is excited about its potential for patients who have not received a diagnosis through standard methods—popEVE has already found many candidate genes beyond the genes already known to cause disease.
This model could not only speed up diagnoses, but also open the door to new drugs by showing where to look for therapeutic targets in the genome.
