A 39 person Phase 1 readout of a machine learning selected sarbecovirus antigen shows what 'future proof' vaccine design actually means, and what it does not.
For three years, the COVID vaccine story has been a chase. Researchers pick a spike protein from the variant dominant this season, manufacture a shot against it, and the virus mutates again. A Cambridge-led team is testing a different theory: that the right target is not what the virus is doing right now, but what the entire sarbecovirus family cannot easily change.
The trial — registered ISRCTN87813400 and published in the Journal of Infection00084-8/fulltext) (Munro et al., 2026;92(6):106759) — is a Phase 1 safety and immunogenicity study of a DNA vaccine whose antigen was selected entirely by machine learning. The vaccine is called pEVAC-PS, delivered via needle-free PharmaJet Tropis device, and the construct is a DNA platform, not mRNA. The team is DIOSynVax, a Cambridge spin-out led by Professor Jonathan Heeney of the Laboratory of Viral Zoonotics, and the sponsor is University Hospital Southampton. The chief investigator is Professor Saul Faust. In 39 healthy adults aged 18 to 50 who had received two or three prior COVID-19 vaccine doses, the vaccine was administered in a dose-escalation regimen (0.2 mg, 0.4 mg, 0.8 mg, and 1.2 mg at day 0 and day 28), and triggered measurable cross-reactive antibodies against SARS-CoV-2, SARS-CoV-1, and related bat sarbecoviruses at day 56. The responses were characterized as modest in the context of substantial pre-existing immunity. No efficacy data exists, and none could.
That last clause is the part the press release does not lead with, and the part the story turns on.
What the machine-learning model actually did is narrower than "AI designed a vaccine." The DIOSynVax pipeline takes the family's genetic sequences, looks for the regions evolution has left largely untouched across SARS-CoV-2, SARS-CoV-1, and their animal relatives, and selects those conserved features as the antigen. As immunologist Neil Mabbott of the University of Edinburgh explained in The Conversation, this is target selection, not end-to-end protein design in the AlphaFold sense. The immunologists, structural biologists, and trial team still do the bench work, the formulation, the clinical operations, and the readout. The model is a sieve.
That distinction matters because the public-health pitch is structural. Current COVID shots train the immune system against the spike, which is also the part of the virus that mutates fastest. A vaccine aimed at conserved features instead bets on a target the pathogen cannot easily shed. The Phase 1 result is the first human evidence that this bet can produce cross-reactive antibodies at all. The magnitude of those antibodies, the durability, and whether they translate into protection against any real outbreak are questions the trial was not designed to answer.
The interpretation of the immunogenicity data was complicated by high baseline antibody levels in participants and heterogeneous prior exposure histories, including ongoing Omicron waves during recruitment, which introduced unavoidable immune bias across dose-escalation cohorts. The trial was not designed to disentangle vaccine effect from background immunity.
The construct is also a deliberate departure from the mRNA platform that defined the COVID response. This is a DNA vaccine, which is more thermally stable and cheaper to transport. In this trial it is delivered needle-free, via a high-pressure microfluidic jet through the skin. The needle-free delivery is a property of the PharmaJet Tropis device, not the AI. The DNA platform is a property of the construct, not the AI either. Only the antigen sequence is the model's contribution.
DIOSynVax is positioning the same architecture for Ebola and influenza, but only the sarbecovirus candidate has human data, and the "future-proof" framing outruns the Phase 1 result. The company is now planning a larger Phase 2 to assess immune response in a broader, more diverse population, and to test how well the breadth of coverage holds up.
The trial also has company. At least two other pan-coronavirus efforts sit further along in some respects: a CalTech RBD-conjugate candidate and Walter Reed's SpFN nanoparticle, both designed around the same conserved-target hypothesis through different chemistry. The Cambridge work is one entry in a crowded race, and the first to take a machine-learning-selected antigen into humans. That is a real milestone, and a small one. The 39-person readout is a proof of concept for a pipeline, not a vaccine the world can use next year, and not a guarantee that the conserved-target theory will survive a real outbreak.
What to watch next is whether the cross-reactive readout replicates in Phase 2 at meaningful titers, whether the breadth narrows when the construct is tested against sarbecoviruses outside the small panel used here, and whether the DIOSynVax platform can carry the same conserved-target logic into Ebola and influenza constructs without losing potency. Until those questions are answered, the honest read is that machine learning has produced its first human-tested antigen, and the architecture question is still open.