Scientists build AI that predicts viral mutations before they hit

For the past few years, the entire world has felt like it was playing a losing game of tag with SARS-CoV-2. Every time a new vaccine rolled out, the virus seemed to debut a fresh mutation that slipped past our defenses.

Now, researchers at Harvard Medical School and the University of Oxford have built an AI designed to put humans on the offensive. Detailed in a new study published in Nature, their tool—called EVEscape—can forecast how a virus will mutate before those dangerous variants even exist in the real world.

A time machine for pandemics

To see if the system actually works, the team effectively rewound the clock to January 2020, feeding EVEscape only the data available at the very start of the COVID-19 pandemic.

Working with that early snapshot, the AI correctly predicted the most frequent mutations and flagged the most concerning variants that would emerge years later. In predicting which mutations would become dominant, EVEscape actually outperformed standard lab testing. It also accurately forecasted which antibody-based treatments would eventually lose their punch as the virus evolved.

Crucially, EVEscape delivers these answers in silico. Traditional experimental methods require researchers to wait for humans to catch a variant, produce antibodies, and send samples to a lab. EVEscape bypasses that multi-month lag entirely by computing probabilities straight from viral code.

How it decodes viral evolution

EVEscape is built on a two-part framework. The first is a generative biological model adapted from EVE, a previous tool the team developed to determine whether human genetic mutations cause disease.

That generative engine learns the evolutionary rules of viral proteins—essentially mapping out what a virus can change without breaking itself. The team then layered on structural details about the virus, specifically flagging the anatomical regions that the human immune system attacks most frequently.

By marrying structural vulnerability with evolutionary viability, EVEscape can forecast the exact escape routes a pathogen is most likely to take. The approach is viral-agnostic: the researchers demonstrated strong predictive accuracy on both HIV and influenza, and they are already pointing the tool at understudied threats with pandemic potential, including the Lassa and Nipah viruses.

Every two weeks, the team now publishes an updated ranking of SARS-CoV-2 variants on its website and shares findings with the World Health Organization. The full codebase has also been released as open-source software.

Why it matters

We think EVEscape represents a fundamental shift in how medicine handles viral threats, turning vaccine development from a reactive scramble into an anticipatory science.

Historically, designing vaccines has been strictly retrospective. You wait for an outbreak, sequence the culprit, and build a countermeasure—often arriving just in time for a new variant to render it obsolete. If predictive tools like this hold up, biotech developers can stress-test drug targets against tomorrow's mutations before entering clinical trials, clearing the path for truly future-proof vaccines.

For global health authorities, an automated variant tracker that runs computationally rather than through physical antibody draws creates an early warning radar for dangerous lineages before they dominate hospital wards.

That said, EVEscape is not a standalone magic bullet. Viral spread in the wild relies on human behavior, local immunity, and random superspreading events that no protein model can forecast. Pharma companies will still need to run wet-lab validations, but letting an algorithm filter out 99% of dead-end variants could shave months off life-saving drug pipelines.