AI variant interpretation uses machine learning to predict whether a genetic change is likely to cause disease, helping labs work through the millions of variants that sequencing finds. Tools such as AlphaMissense can score almost every possible protein-changing variant. But their predictions are one line of evidence among several, not a diagnosis, and they are least tested where uncertainty is highest.
A woman gets her genetic test results back after a family history of heart disease. The report does not say she carries a disease-causing variant. It does not say she is clear either. It says she has a variant of uncertain significance in a cardiac gene: a spelling change in her DNA that nobody can yet classify.
She is not unusual. In a 2023 study in JAMA Network Open, Chen and colleagues analyzed hereditary disease genetic testing in nearly 1.7 million people. Forty-one percent had at least one variant of uncertain significance, often called a VUS. People who were not of European white ancestry had more of them per gene sequenced. When VUS results were later reclassified, about 80 percent turned out to be benign or likely benign, and reclassification took, on average, more than two years.
Twenty years ago, the hard and expensive part of genomic medicine was reading DNA. Today sequencing is fast and relatively cheap. The hard part is knowing what the sequence means. That is the thesis of this piece. The bottleneck in genomic medicine has moved from sequencing to interpretation, and AI is aimed squarely at it. Whether AI helps depends less on how clever the models are than on how their predictions are weighed, calibrated and checked in the people who were left out of the data they learned from.
Why is variant interpretation so hard?
Every person carries millions of genetic differences from the reference genome. The overwhelming majority are harmless. A clinical test narrows the search to specific genes, but even there, labs routinely find rare variants that have never been seen before, or seen only a handful of times.
To classify a variant, labs in many countries follow a framework published in 2015 by the American College of Medical Genetics and Genomics and the Association for Molecular Pathology, usually called the ACMG/AMP guidelines. It combines different types of evidence: how common the variant is in the population, whether it appears in affected family members, what laboratory experiments show about its effect, whether it has been seen in other patients, and what computer predictions suggest. Each type is weighted, and the combination determines whether a variant is called pathogenic, likely pathogenic, uncertain, likely benign or benign.
The difficulty is that for rare variants most of that evidence simply does not exist. There are no affected relatives to test, no laboratory study of that exact change and no previous patients. The computational prediction is often the only evidence available. For years it was also the weakest, allowed to count only as supporting evidence.
What can AI tools like AlphaMissense actually do?
They can make predictions at a scale no laboratory can.
In September 2023, Google DeepMind described AlphaMissense in Science. It considered all 71 million possible missense variants in the human genome, the single-letter changes that swap one amino acid in a protein for another. It classified 89 percent of them: 32 percent as likely pathogenic and 57 percent as likely benign. By DeepMind’s count, only about 0.1 percent of those possible variants had been confirmed by human experts. The predictions were made freely available, and DeepMind stated that they are not designed to be used in the clinic directly and should be interpreted with other evidence.
AlphaMissense is part of a wider wave. Other models predict how variants affect splicing, gene regulation or protein stability. Deep learning has also transformed earlier steps, such as calling variants accurately from raw sequencing data. The overall effect is that almost any variant a lab encounters now comes with a machine prediction attached.
Can AI predictions be used to classify variants in the clinic?
Increasingly, with care. The key development is calibration.
A prediction is only clinically useful if you know how often it is right at a given score. In 2022, a ClinGen working group led by Vikas Pejaver published in the American Journal of Human Genetics a calibration of several computational tools against variants of known classification. It showed that, at well-chosen score thresholds, some tools provide evidence stronger than the “supporting” level the 2015 guidelines assumed, in some cases up to strong evidence toward pathogenicity or benignity. That allows labs to give well-calibrated predictions more weight, but only within defined thresholds, and still in combination with other evidence.
This is the right model for AI in variant interpretation. The prediction is not the answer. It is a measured amount of evidence, whose weight has been checked against known cases.
The weakness is the phrase “known cases.” Calibration depends on variants that have already been classified, and those skew toward well-studied genes and toward people of European ancestry. That is the same population that is overrepresented in large genomic databases. Where the data are thin, both the predictions and their calibration are less certain. Our piece on bias in clinical algorithms argued that the most important fix is to test performance in the groups a tool will serve. That applies with particular force here, because the people with the most unresolved variants are the ones least represented in the data.
A worked example: one variant, three outcomes
Consider the woman from the opening, in a hypothetical case with illustrative details.
Her variant is a missense change in a gene linked to an inherited heart muscle disease. It has never been reported in population databases or in other patients. There is no laboratory study of it. Under the original 2015 approach, the only evidence is a computational prediction, counted as supporting. The variant stays a VUS. She is told to wait.
Now the lab applies calibrated AI predictions. The variant scores well into the range that the lab’s calibration treats as moderate evidence of pathogenicity. That alone is still not enough to call it likely pathogenic. But the lab also notes that the variant sits in a part of the protein where other disease-causing variants cluster, which is another recognized line of evidence. Those two lines together still fall just short under the guideline’s combining rules. But if one more piece arrives, such as the variant turning up in an affected relative, the combination can reach likely pathogenic. At that point, she can be offered cardiac screening, and her relatives can be tested for the same variant.
The third outcome is the one people forget. If the AI score falls in the benign range, and other evidence is consistent, the variant might be reclassified toward benign. Given that most reclassified VUS end up benign, this may be where AI helps most: resolving uncertainty toward reassurance, and sparing families anxiety and unnecessary surveillance.
In each version, the AI prediction changes the conclusion only when combined with other evidence and applied within calibrated limits. The danger is a shortcut: a clinician, a direct-to-consumer company or a patient reading a raw AI score of “likely pathogenic” as a diagnosis.
The strongest case for relying more heavily on AI
The counterargument is serious, and its core is about people waiting.
Families live with VUS results for years. The JAMA Network Open data suggest reclassification takes more than two years on average. Meanwhile, some people with truly pathogenic variants miss preventive care, and many with benign ones live with unnecessary worry. If AI predictions are accurate enough, strict evidence frameworks that give them limited weight may cause more harm through delay than they prevent through caution.
There is also a practical argument. The volume of sequencing is growing faster than the number of clinical geneticists and variant scientists. Without greater reliance on computation, the backlog of uncertain results will grow.
This argument is right that the status quo has costs. But the answer the field has chosen, giving calibrated predictions more weight within an evidence framework, captures much of the benefit while keeping checks in place. Going further, to AI-only classification, would be most risky for the variants where predictions are least validated, in the populations already most poorly served.
What this means for labs, clinicians and health systems
For clinical laboratories, the priority is disciplined use of calibrated tools, documented thresholds and systematic reanalysis. AI predictions change over time as models improve, so labs should plan for periodic reinterpretation of old VUS results and a way to recontact patients when classifications change.
For clinicians, a VUS is not a positive result, and a raw AI score is not a classification. Management decisions should rest on the lab’s final interpretation and family history, and clinicians should explain uncertainty to patients honestly. As we argued in a test without a treatment is just bad news, a result’s value depends on what can be done with it.
For health systems and funders, the biggest lever is data. Classifications improve when labs share variant interpretations in public databases such as ClinVar and when underrepresented populations are included in genomic research, with their consent and on terms they trust. AI will reproduce the gaps in its training data unless those gaps are closed.
For patients and families, a VUS result deserves patience and follow-up rather than alarm. It is reasonable to ask your clinician or genetic counselor whether the lab reanalyzes results and how you will be told if a classification changes.
Frequently asked questions
What is a variant of uncertain significance?
A variant of uncertain significance, or VUS, is a genetic change for which there is not enough evidence to say whether it causes disease. It is common: a 2023 JAMA Network Open study of nearly 1.7 million people undergoing hereditary disease testing found 41 percent had at least one. Most VUS that are later reclassified turn out to be benign.
Is AlphaMissense accurate enough for diagnosis?
Not on its own. AlphaMissense predicts the likely effect of missense variants and classified 89 percent of 71 million possible variants, but Google DeepMind has stated its predictions are not designed for direct clinical use. In practice, labs use calibrated computational scores as one line of evidence, combined with population, family, functional and clinical data.
Will AI reduce uncertain genetic test results?
It is likely to help, especially where calibrated predictions can be combined with other evidence to move variants toward benign or pathogenic classifications. Reclassification toward benign may be the biggest benefit. But AI predictions are less validated for rare genes and for populations underrepresented in genomic data, where uncertain results are already most common.
Why do people of non-European ancestry get more uncertain results?
Because genomic reference databases and research studies have historically included far more people of European ancestry. Variants common and harmless in other populations may look rare and suspicious when there are few comparison samples. The 2023 JAMA Network Open study found more VUS per sequenced gene in people who were not of European white ancestry.
The takeaway
Sequencing a genome is no longer the achievement. Understanding it is. AI is the most promising tool yet for that job, and its predictions deserve more weight than the old guidelines gave them, as calibration work has shown. But genomic medicine runs on evidence, not scores. The gains will go to labs and health systems that treat AI as one measured witness, keep checking it against reality and invest in the diverse data that make every witness more reliable.
Have you or your patients received a variant of uncertain significance, and did it ever get resolved? We’d like to hear how long it took.
