Blog

AI algorithms may keep us up with the mutation rate of the novel coronavirus

In a paper, researchers at the Massachusetts Institute of Technology describe a machine learning algorithm that can predict which mutant strains pose the greatest threat to the world’s newly developed vaccines.
As new variants of novel coronavirus continue to emerge like wildfire around the globe, researchers have been working around the clock to determine which new strains might defeat our vaccine.
The good news is that artificial intelligence (AI) may help with this work.
In a paper published Friday in the journal Science, MIT researchers describe a machine-learning algorithm that can predict which mutant strains pose the greatest threat to the world’s newly developed vaccines.
The tool can be used to rapidly shrink the mutant strains most likely to escape the immune system of vaccinated or previously infected people, allowing researchers to test for these suspected variants in the laboratory and update the vaccine accordingly.
“This is a real-time partner in vaccine development,” says Bryan Bryson, a bioengineer at MIT and co-author of the paper. “Now, what we can do with models is much faster than we can do in the lab.”
The tool comes at a critical moment in the Covid-19 pandemic.
Millions of doses of a vaccine against SARS-Cov-2 (the coronavirus that causes CoviD-19) are finally being made available to the public, and more than three percent of Americans have already been vaccinated.
These vaccines are designed to train our immune system to recognize the particular strain of the coronavirus.
But the more the virus mutates, the greater the likelihood that those who have been vaccinated and previously infected will have reduced immunity to the new strain, a so-called “viral escape.”
A variant of the coronavirus that could escape would have vaccine manufacturers scrambling to update their vaccines in a high-stakes game of catch-up.
In recent weeks, new strains of the virus from Britain, South Africa, California and other areas have begun to spread around the world.
These unwieldy variants seem to be more infectious than the original, though happily less lethal.
Several experts have said publicly that current vaccines are still able to fight the new strains.
“Of course, there will be more mutations.”
Bryson and his colleagues claim that their algorithm could help vaccine manufacturers keep up with the race, “which would reduce the amount of effort currently spent in testing for such mutations.”
“This is a tool that tells you when to investigate,” says co-author Bonnie Berger, a computer scientist at MIT. “As new strains emerge, we can flag which ones have the potential to escape and are worth studying.”
A number of AI-based tools were instrumental in the early development of the Covid-19 vaccine. For example, AI helped researchers determine which parts of the virus’s genetic code were most likely to mutate and how certain mutations affected its physical structure.
MIT’s new machine learning algorithm extends AI’s capabilities by applying it to viruses escaping.
The team’s model was originally developed for machine language understanding, aiming to find both syntax and semantics.
Identifying an escape mutation as one that preserves the virus’s infectivity but causes the virus to look different from the immune system is akin to a word change that preserves the grammar of a sentence but changes its meaning.
Using the similarities between the two, researchers have creatively modified it to sense changes in the virus’s genetic code.
They call this process Constrained Semantic Change Search (CSCS).
As the model learns about the coronavirus genome, it also begins to understand how that genome might change.
It then generates a short list of suspected strains for testing in the lab.
To test the strain, the researchers will first produce a dummy virus that carries a suspicious mutation identified by the computational model.
They then subjected the fake virus to antibodies collected from people who had previously been vaccinated with or infected with Covid-19.
If the antibodies do not neutralize the virus, it indicates that the new strain is able to evade the immune system, so a newer vaccine is needed.
Then it goes back into the algorithm, looking for more suspected mutants.
“It’s like a loop” between computers and wet LABS, Bryson said. “You’re just going back and forth and trying to keep up with the epidemic.”
The researchers added nearly 1,000 sequences of spikes from SARS-CoV-2 to another 3,000 sequences from other types of coronaviruses, such as those that cause the common cold, and then trained the models.
The virus enters human cells through spikes, which are also used by our immune system to recognize the virus.
Thousands of examples have enabled the model to understand how the sequence of amino acids is controlled in coronaviruses.
“The advantage of language models is that they can learn rules directly from a large training set,” said Brian Hie, a doctoral student in Berger’s group and co-author of the paper. “That’s why we use this model in a biological environment, because we don’t know the rules for which amino acids can coexist.”
The MIT researchers tested some of the new variants into their algorithm and found that the British and South African strains both scored “very high” on their likelihood of escaping.
However, Berger said they didn’t score as high as the escape mutants created in the lab.
“The ability to predict when high scores will translate into actual escape from the human immune system is beyond the capacity of the model,” Hie said.
In the long term, he hopes to continue using the model to predict viral mutations that have yet to occur.
“This is a bold and innovative goal for the field of research: to vaccinate against the virus of the future.”

Leave a Reply

Your email address will not be published. Required fields are marked *

X