New AI tool could make future vaccines ‘variant-proof,’ researchers say

A new AI tool that predicts viral mutations could guide treatments for COVID-19 — and the next pandemic.

The system, named EVEscape, was developed at Harvard Medical School and Oxford University. In tests, the tool accurately predicted the most concerning and frequent variants of the SARS-CoV-2 virus that emerged during the pandemic.

A study published last week in Nature revealed an array of promising results. EVEscape’s forecasts proved more accurate than experimental approaches, while faster and more efficient than lab-based tests. The tool also successfully pinpointed therapies that would struggle to subdue new variants.

The predictions are already informing pandemic monitoring efforts. For over a year, the researchers have been releasing biweekly rankings of the most concerning new SARS-CoV-2 strains. The findings are shared with groups including the World Health Organization (WHO). 

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol' founder Boris, and some questionable AI art. It's free, every week, in your inbox. Sign up now!

“There are still thousands of new strains emerging each month — too many to experimentally test,” Pascal Notin, an Oxford University researcher who co-authored the study, told TNW. “EVEscape allows us to rapidly determine the threat level of the new strains.”

Notin and his colleagues have also used EVEscape to successfully predict mutations of HIV and influenza. They’re now testing the tool on lesser-known that could also cause pandemics, such as Nipah and Lassa.

In the future, the researchers envision EVEscape informing vaccine design. At present, vaccines and therapeutics are tested retrospectively against previous pandemic mutations.

EVEscape could add evaluations on where the virus might go next. This offers hope for a powerful new treatment: variant-proof vaccines.

How EVEscape predicts virus mutations 

The new tool is based on a generative model called EVE (Evolutionary model of Variant Effect).

Initially, EVE was developed to predict the risks of genetic mutations causing human diseases, such as cancers. When COVID-19 proved alarmingly adept at mutating beyond the constraints of treatments, the researchers adapted their model to SARS-CoV-2.

Generative models have unique strengths for this job. A key aspect of predicting which mutations will evade immunity is whether they will preserve the so-called “fitness” of the viral protein. This fitness leads to a functional protein that expresses, folds, and binds to the host cell receptor. 

“Generative models trained on evolutionary sequences are critically helpful in supporting that prediction,” said Nodin.

To make the predictions, EVE is trained to learn a compact representation of certain protein sequences. These sequences have maintained their fitness over thousands to millions of years of evolution.

“By doing so, it has to implicitly learn the biochemical constraints that underpin functional sequences,” Nodin explained. “We can subsequently leverage that understanding to predict whether new mutated forms for the protein — whether they are human or viral proteins — will be functional themselves.”

AI’s future fight against viruses

EVEscape’s adaptability stems from its simplicity. The tool learns from a dataset comprised of only viral protein sequences and their structure. As a result, it can be applied to any virus and at the very start of a pandemic.

“This is a great improvement over methods that typically need to wait for relevant pandemic antibodies to be broadly available for testing,” Nodin said.

Another string to EVEscape’s bow is its modular design. When more powerful generative models emerge, the current fitness prediction models can be swiftly replaced.

COVID-19 has also expanded the potential uses of EVEscape. The pandemic provided important lessons about AI predictions — and an immense pool of data.  With that wealth of information, EVEscape can make surprisingly accurate predictions about understudied viruses for which we have scarce data. That could prove immensely valuable in future outbreaks.