The COVID-19 vaccine was a triumph of science. Governments, researchers, and vaccine manufacturers worked in tandem to speedrun a process that can take decades.
Still, after a year of existential terror and isolation, the rollout of these shots, starting in December 2020, felt glacially slow—especially as COVID cases surged in January 2021, taking the lives of an estimated 445,000 people globally, making it the deadliest month of the pandemic.
Credit: Steve Zylius / UC Irvine
If the next coronavirus pandemic were to begin today, the wait for a vaccine could be much shorter because of artificial intelligence (AI), said Lbachir BenMohamed, PhD, an immunologist at the University of California, Irvine, and vice president of research at California-based TechImmune.
With the aid of AI, BenMohamed led a team of researchers to create a broad-spectrum coronavirus vaccine that activates T cells to clear the virus. BenMohamed believes this antigenic response can serve as a sort of skeleton key to train the body to mount an immune response to all viruses in the coronavirus family, including SARS-CoV-2 (which causes COVID-19), MERS-CoV (Middle East respiratory syndrome coronavirus) and SARS-CoV (severe acute respiratory syndrome coronavirus).
Clinical trials are slated to begin early next year on a long-COVID therapy based on this discovery. This step determines safety and efficacy in humans and will reveal whether the vaccine can elicit the desired antigenic response. If the trials are successful, BenMohamed said the treatment can be refashioned into a vaccine for the next coronavirus epidemic or pandemic.
“If there is a pandemic that is coming in future years, which is just a matter of time, we are much more prepared today than we were the last time around,” he said.
BenMohamed is one of several researchers who told CIDRAP News that the field of vaccinology is just starting to realize what’s possible with AI, specifically a subtype of AI called machine learning.
If there is a pandemic that is coming in future years, which is just a matter of time, we are much more prepared today than we were the last time around.
The technology analyzes specific measurements—such as where antibodies bind to certain receptors or how long it takes B cells to produce those antibodies—to identify complex patterns within the immune system. In some cases, the data are so vast and complex that these patterns would be nearly impossible for researchers to recognize on their own.
Saving money by failing faster
It costs an average of $886.8 million to bring a novel vaccine to US markets, according to a 2025 report by the US Department of Health and Human Services (HHS).
The hope is that machine learning will make drug and vaccine discovery less costly and time-consuming, said Duxin Sun, PhD, the founding director of the Institute of AI-Driven Therapeutics Discovery at the University of Michigan College of Pharmacy.
Credit: Martin Janča / iStock
Major pharmaceutical companies are starting to invest in AI for general drug discovery. Pfizer recently signed a licensing agreement with the AI startup Chai Discovery to access software for antibody design. Eli Lilly has also partnered with Chai, along with AI companies NVIDIA and Insilico Medicine. The latter told CNBC in March that it had developed at least 28 drugs using generative AI tools, with nearly half already in clinical trials. Moderna is using AI in its vaccine work with the Coalition for Epidemic Preparedness Innovations (CEPI).
While there have been exciting, novel publications, Sun notes that none of these drugs discovered through AI have advanced past clinical trials. He’s not convinced machine learning will lead to a paradigm shift that enhances the success rate of drug discovery by predicting unintended and potentially dangerous side effects, or finding entirely new ways to solve problems that had never occurred to human scientists to try.
Right now, AI tends to kind of regurgitate the scope of known things, said Peter McCaffrey, MD, director of AI initiatives at the University of Texas Medical Branch [UTMB].
“The art of being the scientist is in that specific deep gap of what is worth doing,” he said.
Rather, Sun and McCaffrey both see AI as a way to help researchers fail faster and move on to the next, hopefully more promising line of inquiry sooner
“The future is not here yet,” said Sun.
The art of being the scientist is in that specific deep gap of what is worth doing.
Still, AI definitely accelerated the timeline of the broad-spectrum coronavirus vaccine, said BenMohamed.
Machine learning was used at the very start of the pre-clinical phase—when researchers evaluate vaccine candidates and filter out those that are unsuitable—by sequencing 15 million coronavirus strains and variants. Of the virus’s 29 proteins, or molecular structures, the model identified 10 proteins that were constant across all coronaviruses, including those that circulate only in bats. These 10 proteins formed the basis for vaccine candidates, or experimental formulations designed to provoke an immune response.
Once AI told BenMohamed’s team which vaccine candidates to focus on, the researchers further winnowed down the list through animal studies until they identified three proteins, on which they based their vaccine.
The process took about four years. Without the aid of machine learning, BenMohamed estimates it would have taken three times that long.
“It’s like [comparing] driving 10 miles an hour to driving 90 miles an hour,” he explained.
AI might outdo scientists at designing flu shots
Researchers are also exploring how AI might improve existing vaccines, such as the yearly flu shot or nasal spray.
Influenza is a moving target for vaccine developers. Pharmaceutical companies need at least six months to produce enough doses of the annual flu vaccine. This requires the World Health Organization (WHO) and US scientists to predict which strains of the fast-mutating flu virus will be dominant in the coming months. Guesses often miss the mark. While the vaccine reduces severe illness, it is typically less than 50% effective at protecting against infection.
But in a 10-year retrospective study published last year, the machine learning algorithm VaxSeer did a better job predicting which flu strains would become dominant and then selecting antigenic matches to those strains than the annual WHO recommendations. The antigen is a vaccine’s active ingredient that produces an immune response.
Lead author Wenxian Shi, a PhD student in the department of electrical engineering and computer science at the Massachusetts Institute of Technology (MIT), said VaxSeer’s accuracy can be attributed to the WHO’s Global Influenza Surveillance and Response System (GISRS), which provided her 20 years of data from a network of 166 institutions across 136 countries.
“They put a lot of effort in collecting all this data all over the world,” said Shi.
Shi also used data from the Francis Crick Institute, a London-based biomedical research center, to train VaxSeer to recognize how antibodies from different vaccine candidates bind to different viral strains.
Navigating diverse datasets
“Data is hard currency,” said Nikolaos Vasilakis, PhD, a UTMB arbovirologist who, in collaboration with McCaffrey, used AI to develop an alphavirus vaccine.
Vasilakis and McCaffrey said vaccine researchers who want to use AI would benefit from better access to data that are categorized as intellectual property. In some instances, entities that fund studies are unwilling to share experimental data.
Another issue: Algorithms need to combine datasets from different sources more efficiently.
Data is hard currency.
The challenge is that different studies collect and measure data in different ways, which can confuse algorithms and lead to flawed results, said Oliver He, DVM, PhD, an AI vaccine researcher at the University of Michigan. He would like researchers to agree on a standardized data annotation that’s easily understood by anyone who uses it.
Even if these issues were resolved, in many cases, there still wouldn’t be enough data to refine models to the point that algorithms could produce just one or two potential vaccine candidates instead of a half dozen promising options.
In other words, the results that a machine learning model can produce are good, but not extraordinary. And with the proliferation of AI research, McCaffrey anticipates a “huge amplification of the bottleneck” in hands-on experimentation.
“Somebody has to physically make this protein, this vaccine candidate, put it in a mouse, measure what happens to them, put it in a cell-based assay,” he said.
Bacterial vaccines are tricky, even with AI
AI breakthroughs for bacterial vaccines are likely further off, because these pathogens have much more complex and diverse methods for evading the immune system.
But how AI can help scientists with bacterial vaccines is the same concept of pattern recognition, just at a grander scale, said Sir Andrew Pollard, MBBS, PhD, director of the Oxford Vaccine Group.
One vaccine the Oxford Vaccine Group is working on is for Staphylococcus aureus, which
causes recurrent skin infections and infections of the bloodstream and internal organs. It is a major driver of antibiotic resistance and a leading cause of death from bacterial infection. After approximately 30 clinical trials, an S aureus vaccine remains elusive.
Pollard thinks AI can help break that losing streak, but, as other researchers interviewed for this story have said, his algorithm needs larger datasets. These data come from studies in which healthy people are safely exposed to the bacterium; Pollard and his collaborators then measure the research volunteers’ immune response.
With time, Pollard said algorithms might become more sensitive and therefore require less data to produce results. It’s even possible that machine learning models for vaccines might eventually predict immune responses for pathogens, similar to how generative AI models like ChatGPT write email responses.
AI will never completely replace humans
Some people’s suspicion of mRNA technology, first widely used in coronavirus vaccines, contributed to overall vaccine skepticism, which has been on the rise since the COVID-19 pandemic. Vasilakis wonders whether growing distrust of AI will have a similar effect.
“We need to come back to the basics as scientists, and we need to be able to communicate with our audience,” he said.
None of the researchers CIDRAP News spoke to envision a future in which data are simply fed to an AI algorithm that then spits out vaccine candidates ready for mass production. Human studies will always be integral to vaccine development to ensure safety and efficacy.
We need to come back to the basics as scientists, and we need to be able to communicate with our audience.
“You have to still use your brain,” said BenMohamed, explaining that although AI is a powerful tool, it can fail to grasp certain biological complexities.
Shi agrees with BenMohemed, and said she’d be fine with the WHO consulting VaxSeer for future flu vaccines, as long as scientists are in the driver’s seat.
WHO spokesman Tarik Jašarević told CIDRAP News that the international public health agency is exploring new technologies, including emerging AI applications. He noted that virus selection relies not only on diverse datasets but also on experts to navigate existing knowledge gaps, such as the biological mechanisms that drive viral evolution.
“In this context, the application of AI in influenza may most feasibly begin with short- to mid-term,” said Jašarević.
HHS officials did respond to CIDRAP News’ request to comment on US efforts to leverage AI for vaccine production.
AI could be used in other areas of vaccine research and development, such as designing clinical trials, said Derek Fleming, PhD, a researcher specializing in coronavirus vaccine research and development with the University of Minnesota’s Center for Infectious Disease Research and Policy, publisher of CIDRAP News.
Fleming said it’s even conceivable that, if AI models become more accurate than animal trials at predicting a vaccine candidate’s safety and efficacy, animal trials could become obsolete. Other researchers would disagree, but as Fleming notes, AI is still very new.
But, he said, “AI will likely never completely replace humans.”
“It’s hard to say what the future holds,” he added.
