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).
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.