Post Syndicated from Emily Waltz original https://spectrum.ieee.org/the-human-os/biomedical/devices/ai-predicts-most-potent-covid-19-mutations
The tool comes at a crucial moment in the COVID-19 pandemic. Millions of doses of vaccines developed against SARS-CoV-2 (the coronavirus that causes COVID-19) are finally rolling out to the public. Just over three percent of the U.S. population has been vaccinated.
These vaccines were designed to train our immune systems to recognize a particular strain of the coronavirus. But the more the virus mutates, the greater the chance that those already vaccinated and those who were previously infected could enjoy less immunity to the new strains.
This harrowing process is called viral escape. Coronavirus mutations that achieve escape would then send vaccine makers scrambling to update their vaccines in a high-stakes game of catch-up.
In recent weeks, new viral variants out of the United Kingdom, South Africa, California, and other regions have begun to spread across the globe. These intractable variants seem to be more contagious than their ancestors, though gratefully not more deadly. Multiple experts have said publicly that our current vaccines should still work against the new strains.
Surely, however, there will be more mutations. Here is where Bryson and his colleagues say their algorithm could help vaccine makers keep up with the game. It would, they say, reduce the laborious experimental techniques currently used to monitor such mutations.
“This is a tool that tells you when to investigate,” says Bonnie Berger, a computer scientist at MIT and co-author of the paper. “As new strains come along, we can flag which ones are worth investigating for escape potential.”
A number of AI-based tools aided the early development of COVID-19 vaccines. For example, AI helped researchers identify which segments of a virus’s genetic code are most likely to change, and how some mutations might affect its physical structure. MIT’s new machine learning algorithm broadens AI’s repertoire by applying it to viral escape.
The group’s model was originally developed for machine language comprehension. The algorithm is designed to look for both grammar (syntax) and meaning (semantics). Using those same two principles, the researchers creatively adapted it to perceive changes to viral genetic code.
They call their process constrained semantic change search (CSCS). As the model learns about the coronavirus genome, it begins to learn what kinds of changes to that genome could be consequential.
From that, it then generates a short list of suspicious strains to test in the lab.
To test the strains, researchers would first generate a pseudovirus carrying the suspect mutations identified by the computational model. They would then subject the pseudovirus to antibodies gathered from people previously vaccinated or infected with COVID-19. If the antibodies don’t neutralize the virus, that suggests that the new strain is capable of evading the immune system—and that updated vaccines are needed.
Then it’d be back to the algorithms to look for more suspicious variants. “It’s like a loop” between the computers and the wet labs, says Bryson. “You just kind of go back and forth and try to understand the pandemic in real time.”
The researchers trained their model on just under 1000 genetic sequences of the SARS-CoV-2 spike protein, plus another 3000 spike sequences from other types of coronaviruses, such as those that cause the common cold. (The spike is what the virus uses to enter human cells, and also what our immune systems will recognize.)
Those thousands of examples taught the model the rules governing how amino acids must be sequenced in coronaviruses. “The nice thing about language models is they can learn the rules directly from a large training set,” says Brian Hie, a PhD candidate in Berger’s group, and co-author of the paper. “That’s why we wanted to use this model in the biological setting, where we don’t know the rules of which amino acids can go together.”
As an experiment, the MIT researchers fed some of the new variants into their algorithm and found that both the UK and South African strains scored “quite high” in terms of their probability of escape. However, they did not rank as high as an escape mutant generated in laboratory experiments, says Berger.
Predicting when a high score will translate into an actual escape from a human immune system is beyond the model’s capability, says Hie.
In the long term, Hie says he hopes to keep working with the model so that it can predict future mutations in viruses that haven’t yet occurred. “That’s a moonshot kind of goal for this line of research: Vaccinating against future forms of the virus,” he says.