An international team of scientists has unveiled a groundbreaking AI system that can identify harmful mutations in human DNA, even if those mutations have never been documented before. Known as popEVE, the technology builds on years of evolutionary data and human genetic variation, offering a powerful new tool for decision-making in the diagnosis of rare diseases.
Approximately half (50%) of patients with a rare disease never receive a definitive diagnosis, often due to limited genetic reference material or missing DNA samples from parents. popEVE could significantly reduce that blind spot. By ranking which mutations in a patient's genome are most harmful, clinicians can investigate the strongest clues first, saving precious time, especially for infants and children, where early intervention can be life-saving.
The technology is described in Nature Genetics by researchers at Harvard Medical School and the Centre for Genomic Regulation (CRG) in Barcelona. Their ambition: to make genetic diagnoses faster, more accessible and less dependent on rare cases. For healthcare systems under increasing pressure, this represents an important step towards precision medicine on a large scale.
Learning from billions of years of evolution
Every human genome contains millions of genetic variations, most harmless, some devastating. Distinguishing between the two remains one of the greatest challenges in the care of rare diseases. Although current AI tools can indicate whether a mutation may be harmful, they rarely provide a severity score. For the many patients in whom no genetic change has ever been observed, this limitation often leads to diagnostic impasses.
The popEVE team approached the challenge differently. Instead of searching patient databases for mutation patterns, they turned to evolution as the ultimate laboratory.
Over billions of years, organisms have tested protein variations through natural selection. Mutations that are incompatible with survival disappear; mutations that are tolerated persist. By analysing genetic sequences from hundreds of thousands of species, popEVE learns which regions of the approximately 20,000 human proteins are essential for life and which can vary safely. As a result, it can detect with high specificity when a single amino acid change is likely to disrupt biological function.
This approach builds on the earlier EVE model from 2021, which already helped classify variants as harmful or benign. However, popEVE's breakthrough lies in its ability to compare the mutation risk of all genes, something clinicians urgently need when faced with thousands of possible variants in a single patient.
Improving clinical decision-making
To rank mutations, the system integrates evolutionary data with population data from gnomAD and the UK Biobank, which include millions of known genetic variants. This allows popEVE to assess whether a mutation is harmful, regardless of how common or rare it is, avoiding the biases found in traditional tools, which over-mark mutations simply because they are less documented in non-European populations.
The model was tested on 31,000 families affected by severe developmental disorders. In 98% of cases, popEVE correctly identified the known pathogenic mutation as the most harmful genetic change, outperforming advanced systems such as DeepMind's AlphaMissense.
Even more strikingly, 123 previously unknown disease-related genes were identified, many of which are involved in brain development and have only been observed once or twice worldwide. For clinicians working with ‘patients of one,’ this offers an unprecedented diagnostic direction.
Equality built into the algorithm
Human genetic databases still underrepresent a large part of the world. Many AI tools flag unknown genetic variants as dangerous simply because they have not been observed before. This is a structural bias that disproportionately affects African, Asian, Middle Eastern and indigenous communities.
popEVE addresses this gap by treating every variant equally, regardless of ancestry distribution. This reduces the number of false positives and supports fairer access to genomic medicine. As co-author Dr Jonathan Frazer notes, ‘No patient should receive a frightening result because their ancestry is missing from genetic databases. popEVE helps to restore balance.’
From development to practice
popEVE is an important step forward, but not a complete genetic solution. The tool only evaluates protein-altering mutations, which means that regulatory and structural changes in DNA still require separate analysis. And despite its predictive power, the final diagnosis will always be based on clinical expertise, symptoms and the patient's history.
Nevertheless, the implications are far-reaching, because when popEVE is integrated into diagnostic workflows, the tool is capable of:
- accelerating the time to diagnosis for patients with rare diseases
- reducing healthcare costs associated with long diagnostic odysseys
- prioritise precision medicine, even in regions with limited resources
- provide guidance in the design of therapies and targeted research
Where sequencing used to raise questions, AI can now provide answers.
By learning from the entire family tree of life, popEVE shows how artificial intelligence, guided by evolution, can bring clarity to the most complex aspects of human genetics. For healthcare, this represents a decisive step towards future-proof diagnostics for rare diseases: faster, fairer and more accessible for every patient.
Better diagnostics thanks to AI
Over the past two years, there have been many examples of AI-driven tools and solutions proving their ability to accelerate and improve medical diagnostics. A great deal of time and energy is being invested in AI diagnostic solutions for the early detection of (various forms of) cancer, such as melanoma, prostate cancer and lung cancer. But also for the earlier detection of other conditions, such as heart failure.
