AI Designed a Vaccine to Fight Every Coronavirus. In 39 Humans, the Immune Response Was ‘Modest.’

One out of twelve.

That is the immunogenicity hit rate for pEVAC-PS, the first vaccine whose active ingredient was designed entirely by artificial intelligence, in a Phase I clinical trial published June 5 in the Journal of Infection. Four dose groups of healthy volunteers aged 18 to 50 received two intradermal injections 28 days apart, delivered by a needle-free PharmaJet Tropis device that fires a micro-fluid jet through the skin. Thirty-nine people enrolled. Zero experienced a serious adverse event. Safety was, by every measure the trial tracked, flawless.

Immunogenicity was not, and the gap between the safety profile and the immune response tells most of the story about where AI-designed vaccines stand right now.

The researchers tested binding antibody responses across four dose groups (0.2 mg, 0.4 mg, 0.8 mg, 1.2 mg) against three antigens: the vaccine’s own computationally designed receptor-binding domain, the SARS-CoV-2 Wuhan wild-type RBD, and the SARS-CoV-1 RBD. Twelve group-antigen combinations, and exactly one of them — Group 4, the highest dose, against the vaccine’s own construct — showed a statistically significant increase in binding antibodies (p < 0.05 by Friedman test with Dunn’s correction). The other eleven did not reach significance. Cross-neutralization against SARS-CoV-1, the metric that would most directly validate the “pan-sarbecovirus” claim, showed no change in any group.

The Press Release and the Paper

Read the coverage and you get one story; read the data tables and the numbers tell something different entirely.

The Scottish Sun described a “pandemic-proof vaccine that conquers viruses we don’t even know about yet.” The article quoted Professor Marian Knight of the National Institute for Health and Care Research calling the trial results a “remarkable success” and a “pivotal leap forward.” It reported 49 healthy volunteers, a number that appears nowhere in the published paper, which enrolled 39 and analyzed 30 for immunogenicity after excluding participants who tested positive for COVID during the trial.

The paper itself uses different language entirely. The authors, led by Professor Saul Faust at University Hospital Southampton, write that findings “do not yet substantiate broad or robust neutralizing activity” but “support the underlying design concept.” Where the Sun says “pandemic-proof,” the Journal of Infection says “modest and variable.”

This is not a contradiction, because both statements can be true simultaneously — Phase I trials are designed to establish safety and find the right dose, not to demonstrate population-level protection. But the distance between “pandemic-proof” and “modest and variable” is the kind of gap that erodes public trust in vaccine science when the Phase II results inevitably introduce more complexity and the public remembers only the headline.

What the AI Actually Did

DIOSynVax — short for Digital Immune-Optimised Synthetic Vaccines — is a Cambridge spin-out founded by Professor Jonathan Heeney, a veterinary immunologist who spent two decades studying cross-species coronaviruses before asking the question that launched the company: what if you stopped chasing individual variants and designed an antigen that targets what all sarbecoviruses share? The company’s computational platform does something genuinely novel: it analyzes the three-dimensional surface topology of the receptor-binding domains across the entire sarbecovirus family and designs a synthetic antigen that presents conserved epitopes, the structural features most likely to elicit antibodies effective against multiple related viruses, not just the strain circulating today.

Think of it as computational averaging across a protein family, a kind of molecular consensus-building. Instead of training the immune system to recognize one specific lock, you train it to recognize the shape all the locks share.

The peptide microarray data from the trial supports this design intent. Antibodies in vaccinated subjects targeted the S309 epitope, a conserved region known to enable broadly neutralizing activity across sarbecoviruses. Increased reactivity appeared in both the N-terminal and ACE2-binding regions of the spike protein. The antigen design worked at the molecular level. The immune system made antibodies that bound where they were supposed to bind.

Binding and neutralization are different things, and this distinction explains why the immunogenicity data can look encouraging and disappointing at the same time. You can have antibodies that stick to the right spot on a virus but lack the structural precision or concentration to actually prevent infection. The neutralization data — tested in Groups 3 and 4 only, against four pseudoviruses — showed significant increases against Omicron BA.1 in Group 3 and Delta in Group 4, meaning two of eight pseudovirus-group combinations reached significance: twenty-five percent.

Metric	Result
Volunteers enrolled	39
Serious adverse events	0
Binding significance (of 12 combos)	1 (8.3%)
Neutralization significance (of 8 combos)	2 (25%)
SARS-CoV-1 cross-neutralization	No change
Delivery platform	DNA (needle-free intradermal)

The Strongest Case for Optimism

The counterargument deserves its full weight, and it is genuinely strong enough to reframe the entire trial narrative.

Phase I trials are designed to test safety, not efficacy. Judging pEVAC-PS on immunogenicity at this stage is like grading a pilot on the pre-flight checklist before they have left the gate. The modest immune response is likely a function of two compounding factors that have nothing to do with the AI-designed antigen itself.

First, the delivery platform. pEVAC-PS is a DNA vaccine delivered intradermally, and DNA vaccines have historically produced weaker immune responses than mRNA vaccines in humans — a known limitation of the platform, not a failure of the antigen design. DIOSynVax is already developing an mRNA version backed by a $42 million investment from the Coalition for Epidemic Preparedness Innovations. Swap the delivery vehicle, keep the antigen, and the immunogenicity picture could change dramatically.

Second, the trial population. Recruitment ran from December 2021 through September 2023, squarely through the Omicron BA.1, BA.2, BA.5, and XBB waves. Every participant had prior COVID vaccination, and many had breakthrough infections on top of that. Their immune systems were already primed with antibodies against the very epitopes pEVAC-PS targets. Detecting a statistically significant boost above that high, heterogeneous baseline is extraordinarily difficult in a 39-person cohort, like trying to hear a whisper in a room where everyone is already shouting. A future trial in immunologically naïve participants — or infants without prior exposure — would be a cleaner test of the antigen’s intrinsic potency.

The real innovation is the computational design platform, not this specific formulation on this specific delivery vehicle in this specific population at this specific moment in the pandemic’s immunological history. If the platform can design one conserved-epitope antigen, it can design others for entirely different virus families. DIOSynVax is already pivoting toward influenza: supra-seasonal flu, pre-pandemic bird flu, and ultimately a universal influenza vaccine, according to COO Dr. Rebecca Kinsley. Platform can be upgraded; antigen design is the hard part, and the hard part appears to work.

The Math Nobody Ran

Here is the original calculation that nobody else ran. The Innovate UK grant that funded pEVAC-PS was £1.93 million — roughly $2.4 million at current exchange rates — and that bought a computationally designed antigen, a Phase I trial in 39 volunteers, published safety data, and proof-of-concept immunogenicity data showing the designed epitope actually elicits targeted antibodies.

Compare that to the traditional vaccine development pipeline. A single conventional Phase I trial typically costs between $1 million and $5 million for a simple design, and pEVAC-PS sits comfortably within that range. But the antigen design cycle — the part where molecular biologists and crystallographers spend months or years identifying conserved epitopes, synthesizing candidate antigens, and testing them in animal models — was compressed by computation. Heeney’s team used AI to do in silico what would otherwise require hundreds of wet-lab iterations.

The per-trial cost savings are modest at best. The time savings are where the platform earns its keep. If a pandemic-capable sarbecovirus emerged tomorrow and the goal was a broadly protective vaccine candidate rather than a variant-chasing reformulation, the computational platform could generate a new antigen design in weeks rather than the months it takes to go from sequence publication to candidate nomination through traditional structural biology. Speed is the product.

Limitations

This analysis relies on the published Phase I data and secondary reporting. We have not independently accessed the raw immunogenicity datasets. The trial used an in-house ELISA assay rather than a standardized commercial kit, making it difficult to benchmark antibody titers against established correlates of protection from other vaccine trials. No correlate of protection has been established for pan-sarbecovirus vaccines; we do not know what titer level would actually prevent infection. The 39-person sample size is standard for Phase I but insufficient for drawing conclusions about efficacy. GBP511, a nanoparticle-platform candidate from the University of Washington and SK Biosciences, also claims the “first pan-sarbecovirus vaccine in human trials” designation, and the priority claim may be contested.

The Bottom Line

The first vaccine designed by AI has passed the test that matters most at this stage: it does not hurt people. It also does not yet protect them, at least not at the doses and on the platform tested. The distance between those two facts is where the next five years of computational vaccinology will be won or lost.

If you are a public health official, the actionable finding is not the immunogenicity — it is the platform. DIOSynVax’s computational design produced a synthetic antigen that elicited antibodies targeting the intended conserved epitope in human subjects. That is proof-of-concept for a design approach, not a finished vaccine. Watch the mRNA reformulation funded by CEPI’s $42 million and the Phase II trial enrolling over 200 participants. If the same antigen on a stronger delivery platform produces robust cross-neutralization, the platform becomes a rapid-response tool for the next pandemic — not a cure for this one.

If you are a biotech investor, the signal worth tracking is the pivot to influenza. A universal flu vaccine is a larger market than a pan-coronavirus vaccine, and the same computational approach applies: conserve the epitope, design the antigen, let the delivery platform catch up. The $42 million from CEPI is validation funding. The influenza program is the commercial play.

If you are a science journalist, compare the paper to the press coverage before you write. “Pandemic-proof” and “modest and variable” cannot both be the headline. One of them is the data. Choosing the wrong one does more damage to vaccine confidence than any Phase I result ever could.