AI found a vulnerability shared by all coronaviruses — and turned it into a vaccine candidate
In the spring of 2026, a research team at the University of Cambridge announced results from a Phase I clinical trial of an experimental pan-coronavirus vaccine — a candidate designed not to target one strain of SARS-CoV-2, but to produce an immune response against a structural feature shared by every known coronavirus. The trial enrolled 96 healthy adults in the UK and was primarily designed to assess safety and optimal dosing. It was not designed to measure protection against infection. That distinction matters, and it is where most early reporting has been imprecise.
The candidate passed its primary endpoint: no serious adverse events were recorded, and all participants generated measurable antibody responses to the target antigen. Phase I data do not tell us whether those antibodies will prevent disease in the real world. What they do establish is that the vaccine is safe enough to proceed to larger trials — and that the AI-driven design process produced an antigen the human immune system can recognise and respond to.
The problem with strain-specific vaccines
Every approved COVID-19 vaccine currently in use targets the spike protein of SARS-CoV-2 — specifically the receptor-binding domain (RBD), the region the virus uses to enter human cells. This was a sensible design choice: the spike protein is highly immunogenic and the RBD is essential to infection. The problem is that the spike protein mutates rapidly. The variants that drove successive waves — Alpha, Delta, Omicron, and its descendants — all carried mutations in the RBD that partially evaded immunity from earlier vaccines or prior infection. Manufacturers have updated vaccine sequences repeatedly to track the dominant circulating strain, in a process that resembles the annual influenza vaccine update.
The deeper problem is that the coronavirus family is large. SARS-CoV-2 is one of seven coronaviruses known to infect humans. MERS-CoV, which causes Middle East Respiratory Syndrome, has a case fatality rate of roughly 35 percent and circulates in camel populations across the Arabian Peninsula and parts of Africa. There are hundreds of bat coronaviruses, several of which have already demonstrated the ability to infect human cells in laboratory conditions. A strain-specific vaccine against SARS-CoV-2 provides no cross-protection against any of these.
How AI was used in the design process
The Cambridge team's approach started with a question that previous vaccine research had not been able to answer computationally at scale: across the full diversity of coronaviruses, are there structural regions that do not change — and if so, can an antigen be designed to make the immune system attack exactly those regions?
Three AI methods were applied in sequence.
AlphaFold for structural prediction. The team used AlphaFold 3 to predict the three-dimensional structure of proteins from dozens of coronavirus strains — including strains for which no experimental crystal structure existed. AlphaFold's ability to predict protein folding from amino acid sequence alone meant that structurally conserved regions could be identified without the years of wet-lab work previously required to determine each structure experimentally. The researchers identified a segment of the fusion peptide — a region the virus uses to merge its membrane with a host cell's — that maintained nearly identical geometry across all coronaviruses analysed.
ML-based epitope mapping. Machine learning models trained on immunogenicity databases were used to identify which sub-regions of the conserved structural segment were most likely to be recognised by human T cells and B cells — the components of adaptive immunity responsible for long-lasting protection. This step filtered out conserved regions that, while structurally stable, were unlikely to generate a robust immune response, and flagged those most likely to be genuine epitopes: specific molecular targets the immune system can learn to attack.
Generative antigen design. Rather than using the naturally occurring peptide sequence as the vaccine antigen, the team applied a generative ML model to redesign the antigen's surface while preserving its three-dimensional shape and the key epitope residues. The goal was to make the antigen more immunogenic — more visible to the immune system — than the native viral sequence, which coronaviruses have partly evolved to conceal. This process produced several candidate antigen sequences that were tested first in mouse models before the lead candidate was selected for human trials.
What "pan-coronavirus" means in practice
A pan-coronavirus vaccine targets conserved regions — parts of viral proteins that cannot mutate without destroying the virus's ability to function. The fusion peptide region the Cambridge team focused on is conserved because it plays a mechanically essential role in infection: it inserts into the host cell membrane to initiate membrane fusion. A virus that mutated this region would lose the ability to infect cells and would not survive to replicate. Evolution has locked this sequence in place across the coronavirus family, which is exactly what makes it an attractive vaccine target.
"Pan-coronavirus" does not mean the vaccine is guaranteed to work against all coronaviruses in all settings. It means the immune response it generates is directed at a target that is structurally present in all known coronaviruses — so if that immune response is effective, it would in principle provide cross-protection across the family. The Phase I trial confirmed the immune system mounts a response against the target. Whether that response translates to clinical protection against SARS-CoV-2, MERS-CoV, or a hypothetical future zoonotic coronavirus is what Phase II and III trials are designed to determine.
Phase I trial: what was measured and what was found
The UK trial enrolled 96 healthy adults aged 18 to 65 across three dosing cohorts. Participants received two intramuscular injections 28 days apart. The primary endpoints were safety — adverse events, laboratory abnormalities, any serious adverse event within 90 days — and reactogenicity: local injection-site reactions and systemic symptoms such as fever, fatigue, and headache. Secondary endpoints measured immunogenicity: serum antibody titres against the target antigen and neutralising activity against pseudovirus models of several coronavirus strains.
The results: no serious adverse events were recorded in any cohort. Reactogenicity was in line with other protein subunit vaccines — injection site pain and mild fatigue were the most common reactions, typically resolving within 48 hours. All participants seroconverted, meaning all developed detectable antibody responses. In the highest-dose cohort, researchers observed cross-reactive neutralising antibodies against pseudovirus models of SARS-CoV-2 (Wuhan strain and the most recent Omicron subvariant), SARS-CoV-1, and MERS-CoV. The magnitude of neutralising responses was lower than those seen after natural SARS-CoV-2 infection or current mRNA vaccines, which the researchers attribute to the immune system having had less prior exposure to this novel target antigen.
These are promising immunogenicity results for a first-in-human study, but they are not efficacy data. A neutralising antibody response in a lab assay does not automatically translate to clinical protection. That connection has to be demonstrated in larger trials with actual infection endpoints.
The path to approval
Phase II trials, expected to begin in late 2026 pending regulatory review by the UK's Medicines and Healthcare products Regulatory Agency (MHRA), will enroll several hundred participants across broader age groups — including older adults, in whom immune responses typically differ — and begin to assess immunogenicity against circulating coronavirus strains in a larger and more diverse population. Phase II will also extend safety monitoring over a longer follow-up period.
Phase III — the pivotal efficacy trial — would require tens of thousands of participants and enough circulating coronavirus to measure protection against infection or severe disease. The statistical requirements for a Phase III trial mean it is unlikely to complete before 2029 under an optimistic timeline. Regulatory approval by the FDA, EMA, or WHO prequalification for global deployment would follow, adding further time. The manufacturing scale-up challenge is substantial: the adjuvanted protein subunit format the Cambridge team used is more complex to produce at scale than mRNA vaccines, though it has the advantage of not requiring ultra-cold storage.
AI as infrastructure for vaccine development
The Cambridge pan-coronavirus project is one of several that illustrate a structural shift in how vaccines are designed. AlphaFold's public release in 2021 removed a major bottleneck — protein structure determination — from the early stages of vaccine antigen design. What previously required years of X-ray crystallography or cryo-electron microscopy can now be computed in hours. That compression does not accelerate clinical trials, which are governed by biology and regulatory requirements that cannot be shortened. But it dramatically reduces the time between "identify a target" and "have a testable antigen candidate."
The generative antigen design methods applied in this work are still maturing. The models were trained primarily on data from well-studied protein families; their reliability for novel viral targets depends on how well those targets resemble the training distribution. The field is also grappling with how to validate AI-generated antigens — the lab assays for immunogenicity are the same as before, and they remain the bottleneck after the computational design phase.
What has changed is the hypothesis space researchers can explore. A team that previously might evaluate five to ten antigen candidates before entering preclinical testing can now computationally screen thousands of design variants and carry the best-performing handful into the lab. If even a fraction of the resulting candidates succeed in human trials at the rate the Cambridge pan-coronavirus work suggests is possible, the pipeline from viral discovery to Phase I trial — which took roughly 18 months for the first COVID-19 vaccines under emergency conditions — could become achievable in under a year for future pandemic threats. That is not a guarantee the science delivers on that timeline. It is a reason to watch Phase II results closely.