Artificial intelligence is becoming an essential tool in healthcare, particularly for disease prediction. This is the case for Delphi-2MA generative AI capable of calculating the individual risk of contracting more than 1,000 diseases in the next decade. This predictive capability is now also being used for diagnosis, particularly to predict the deleterious effect of a given mutation. Two new tools presented in November 2025 demonstrate the potential of these approaches to identify the genetic causes of very rare diseases, facilitating the diagnosis of patients whose conditions were previously difficult to explain due to their rarity.
An AI that identifies mutations responsible for rare diseases
Indeed, more traditional methods use knowledge from multiple cases of the same disease to understand its causes and thus be able to detect them in a new patient. But when it comes to extremely rare, or even unique, cases, there simply isn't enough data to perform these analyses. One way to solve this problem is to reverse the logic, and instead of looking for what type of mutation might be associated with a disease, to identify the potentially pathogenic mutations in an individual in order to pinpoint the one or ones that could be causing the disease.
This is what the first of these AIs, named popEVE, does. This tool, designed by researchers from Harvard University and the Barcelona Institute of Science and Technology and presented in the journal Nature Genetics, evaluates the pathogenic potential of any genetic variant. To do this, it has been trained with data on protein changes during evolution and within human populations.
This dual analysis gives it an idea of how much a given protein has changed during its evolution, to determine which mutations are permitted or not by natural selection (based on their deleterious effects), and the distribution of a type of mutation in healthy humans or those affected by diseases. Thus, popEVE evaluates all the genetic mutations in an individual suffering from a rare disease and classifies them according to their potential pathogenicity, highlighting the one(s) that most likely cause the disease.
Read alsoGeneral AI, a poorly defined fantasy
Associating a mutation with a potential phenotype
The second AI was presented just four days later, on November 28, 2025, in the journal Nature Communications by researchers at the Icahn School of Medicine at Mount Sinai (in New York). It has been named V2P (for "variant-to-phenotype"), and as its name suggests, it links a genetic variant to the phenotype it could cause. In doing so, it goes further than popEVE: it doesn't simply assess the pathogenicity of an individual's mutations one by one, but studies them together. V2P evaluates the impact of a mutation within its genetic context, because most of the time diseases are not caused by a single mutation but by the combined effect of several.
A potentially deleterious mutation will only be so in certain individuals, depending on their genomes, and this AI makes it possible to estimate these combined effects and associate them with a potential phenotype. Our approach allows us to pinpoint the genetic changes most critical to a patient's disease, instead of having to juggle thousands of potentially important mutations., explains in a press release David Stein, author of the study. Knowing not only whether a variant is pathogenic but also what type of disease it might cause makes diagnosis easier.
In addition to helping patients better understand their condition, this technology could also advance medical research, shedding light on the links between genetics and disease: »V2P allows us to understand more clearly how genetic changes cause diseases, adds Yuval Itan, director of the study. Its ability to link variants to the types of diseases they could cause will also help us identify the genes and biological pathways that we will need to target to design therapeutic approaches based onu genomic profile of each patient. Another step towards personalized medicine, boosted by AI.
English 
French