The development of the COVID-19 vaccine, a remarkable scientific achievement, was accelerated by unprecedented collaboration between governments, researchers, and manufacturers, dramatically shortening a process that typically spans decades. Despite this speed, the vaccine rollout, beginning in December 2020, felt slow to many, particularly as January 2021 saw a devastating surge in COVID-19 cases, claiming an estimated 445,000 lives globally and marking the pandemic’s deadliest month.
However, the landscape for future viral outbreaks is poised for significant change, thanks to the burgeoning power of artificial intelligence (AI). According to Lbachir BenMohamed, PhD, an immunologist at the University of California, Irvine, and vice president of research at TechImmune, the development timeline for a vaccine in a potential future coronavirus pandemic could be dramatically reduced.
BenMohamed, utilizing AI, has spearheaded the creation of a novel broad-spectrum coronavirus vaccine. This vaccine is designed to activate T cells, crucial components of the immune system, to effectively clear the virus. He theorizes that this targeted immune response could act as a universal key, preparing the body to combat all viruses within the coronavirus family, including SARS-CoV-2 (responsible for COVID-19), MERS-CoV, and SARS-CoV.
Clinical trials are anticipated to commence early next year for a long-COVID therapy derived from this research. These trials will be pivotal in assessing the treatment’s safety and efficacy in humans, and crucially, whether it can indeed stimulate the desired broad immune response. Should these trials prove successful, BenMohamed believes the therapy can be adapted into a vaccine ready for deployment against future coronavirus epidemics or pandemics. "We are much better prepared today than we were the last time around," he stated, emphasizing the inevitability of future pandemics.
BenMohamed is among many researchers who recognize the transformative potential of AI, particularly machine learning, in the field of vaccinology. This technology excels at analyzing vast and complex datasets, identifying intricate patterns within the immune system that would be nearly impossible for human researchers to discern. Such patterns could relate to how antibodies interact with specific receptors or the speed at which B cells produce antibodies.
The economic implications of AI in vaccine development are also significant. The U.S. Department of Health and Human Services estimates the average cost to bring a new vaccine to U.S. markets is nearly $887 million. Duxin Sun, PhD, founding director of the Institute of AI-Driven Therapeutics Discovery at the University of Michigan College of Pharmacy, expressed hope that machine learning will reduce both the cost and time associated with drug and vaccine discovery.
Major pharmaceutical corporations are increasingly investing in AI for drug discovery. Pfizer, for instance, has entered into a licensing agreement with AI startup Chai Discovery for antibody design software. Eli Lilly is collaborating with Chai, NVIDIA, and Insilico Medicine, with Insilico Medicine reporting the development of at least 28 drugs using generative AI, nearly half of which are already in clinical trials. Moderna is also integrating AI into its vaccine development efforts with the Coalition for Epidemic Preparedness Innovations (CEPI).
Despite these advancements, Sun points out that no AI-discovered drugs have yet progressed beyond clinical trials. He remains cautious about whether machine learning will fundamentally alter drug discovery by accurately predicting adverse side effects or uncovering entirely novel therapeutic approaches that human scientists might overlook.
Peter McCaffrey, MD, director of AI initiatives at the University of Texas Medical Branch (UTMB), observed that current AI often "regurgitates the scope of known things." He believes the true art of scientific discovery lies in identifying the most promising and impactful avenues for research. Both Sun and McCaffrey see AI’s primary current value in enabling researchers to "fail faster," allowing them to quickly move on to more promising lines of inquiry. "The future is not here yet," Sun remarked.
Nonetheless, BenMohamed affirmed that AI significantly accelerated the development timeline for his broad-spectrum coronavirus vaccine. Machine learning played a crucial role from the outset of the preclinical phase. By analyzing 15 million coronavirus strains and variants, the AI model identified 10 proteins that are consistently present across all coronaviruses, including those found in bats. These conserved proteins served as the foundation for vaccine candidates. After AI narrowed down the options, animal studies further refined the selection, leading to the identification of three key proteins upon which the vaccine is based. This process, which BenMohamed estimates would have taken three times longer without machine learning, highlights AI’s capacity to dramatically shorten development cycles, akin to "driving 90 miles an hour" compared to "driving 10 miles an hour."
The potential for AI to enhance existing vaccines, such as the annual flu shot, is also being explored. Influenza poses a significant challenge due to its rapid mutation rate, requiring vaccine developers at least six months to produce sufficient doses. This necessitates predictions of dominant strains from organizations like the World Health Organization (WHO), which are often imperfect, resulting in flu vaccines with typically less than 50% efficacy against infection, though they do reduce severe illness.
A decade-long retrospective study introduced VaxSeer, a machine learning algorithm that demonstrated superior accuracy in predicting dominant flu strains and identifying corresponding antigenic matches compared to annual WHO recommendations. Wenxian Shi, the study’s lead author and a PhD student at MIT, attributed VaxSeer’s success to the extensive data from the WHO’s Global Influenza Surveillance and Response System (GISRS), spanning 20 years and involving 166 institutions globally. Shi also utilized data from the Francis Crick Institute to train VaxSeer to understand how antibodies from different vaccine candidates interact with various viral strains.
The effective utilization of AI in vaccine development hinges on access to robust and well-organized data. Nikolaos Vasilakis, PhD, an arbovirologist at UTMB who collaborated on an alphavirus vaccine using AI, emphasized that "Data is hard currency." He and McCaffrey noted that researchers could benefit from improved access to data currently held as intellectual property, as some funding entities are reluctant to share experimental findings.
Another significant hurdle is the ability of algorithms to efficiently integrate datasets from disparate sources. Oliver He, DVM, PhD, an AI vaccine researcher at the University of Michigan, explained that variations in data collection and measurement methods across studies can lead to algorithm confusion and flawed results. He advocates for standardized data annotation to ensure clarity and consistency for AI models. Even with these challenges addressed, the availability of sufficient data to refine models to produce a small number of highly specific vaccine candidates, rather than several promising options, remains a concern. McCaffrey foresees a "huge amplification of the bottleneck" in hands-on experimentation as AI research proliferates, stating, "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."
Breakthroughs in AI for bacterial vaccines are anticipated to be further off due to the complex and varied mechanisms bacteria employ to evade the immune system. However, Sir Andrew Pollard, MBBS, PhD, director of the Oxford Vaccine Group, believes AI’s pattern recognition capabilities, albeit on a larger scale, can still assist in this area. The Oxford Vaccine Group is working on a vaccine for Staphylococcus aureus, a bacterium responsible for severe skin and bloodstream infections, a major contributor to antibiotic resistance, and a leading cause of bacterial infection deaths. Despite approximately 30 clinical trials, an effective S. aureus vaccine has remained elusive. Pollard suggests AI could help overcome this challenge, provided larger datasets are available, derived from studies exposing healthy individuals to the bacterium and measuring their immune responses. He also noted that as algorithms become more sophisticated, they may require less data and could potentially predict immune responses to pathogens, much like generative AI models create text.
While some express concern about growing distrust in AI mirroring skepticism towards technologies like mRNA, Vasilakis stresses the importance of scientists returning to fundamental communication principles. The researchers CIDRAP News spoke with do not foresee a future where AI solely generates ready-to-produce vaccine candidates. Human oversight and experimental validation will remain critical for ensuring safety and efficacy. "You have to still use your brain," BenMohamed stated, acknowledging that AI, while powerful, can miss crucial biological nuances. Shi expressed willingness for the WHO to consult VaxSeer for future flu vaccines, provided scientists remain in control.
Tarik Jašarević, a WHO spokesman, confirmed the organization’s exploration of emerging AI applications. He highlighted that virus selection involves not only diverse datasets but also expert interpretation to bridge knowledge gaps in viral evolution. He anticipates AI’s application in influenza may initially focus on short- to mid-term predictions. Officials from the U.S. Department of Health and Human Services did not respond to requests for comment regarding U.S. efforts in AI-driven vaccine production.
Derek Fleming, PhD, a researcher specializing in coronavirus vaccine development at the University of Minnesota, suggested AI could be applied to other aspects of vaccine research, such as designing clinical trials. He speculated that if AI models surpass animal trials in predicting vaccine safety and efficacy, animal testing could potentially become obsolete, although he acknowledges this view is not universally shared and that AI is still in its nascent stages. Ultimately, Fleming concluded, "AI will likely never completely replace humans," emphasizing the uncertainty of future developments.