EVEscape: an AI to anticipate the mutations that let viruses escape immunity

Why is it so hard to anticipate virus variants?

A virus changes all the time. Some mutations let it escape antibodies, the defence proteins made after an infection or a vaccination. They are what give rise to "variants of concern". Knowing them in advance would help design vaccines and treatments that keep working.

The problem: existing methods arrive too late. Laboratory tests need antibodies from patients, which exist only once much of the population has been infected. And computational methods lean heavily on the strains already circulating.

How does EVEscape work?

EVEscape combines two ingredients. First, a deep-learning model (an AI that learns from examples) studies historical virus sequences to estimate which mutations remain compatible with a working virus. Second, information about the protein's structure and its physical and chemical properties indicates which mutations antibodies would see most easily and which differ most from what they recognise.

This recipe has a major advantage: it can be used before surveillance data, experimental scans or 3D structures of antibodies are available.

What did the study show?

For the SARS-CoV-2 virus, the team trained EVEscape only on sequences available before 2020, then compared its estimates with what was learned afterwards about the pandemic. Result: the model is as accurate as high-throughput experimental scans at anticipating the virus's variation.

In detail, at anticipating experimental measures of antibody escape, EVEscape scores 0.53 on the area under the precision-recall curve (a score where higher is better), against 0.24 for earlier models of the same type.

The method generalises: it was validated on the Covid-19 virus, HIV and influenza, and applied to little-studied viruses with pandemic potential such as Lassa and Nipah. The researchers publish regularly updated escape scores for current Covid-19 strains.

What does this change for patients?

In time, an early-warning system like this could help teams developing vaccines and treatments to target parts of the virus less likely to mutate, and to spot worrying strains sooner.

Today it is a research tool. It says nothing about the protection of any individual patient and has not shown a direct effect on vaccines already rolled out.

What are the limits?

EVEscape cannot anticipate everything that is specific to a new pandemic, such as the particular constraints of a virus. The authors recommend using it alongside laboratory experiments and enriching it with pandemic data as soon as it exists.

The most detailed results concern Covid-19, where it was possible to compare with what really happened. For Lassa and Nipah, these are estimates that cannot yet be compared with a real outbreak.