First universal vaccine designed entirely by AI: Cambridge human trials confirm safety

First universal vaccine designed entirely by AI: Cambridge human trials confirm safety

🔎 Never before had a vaccine component been entirely conceived by a machine — until now

In June 2026, the University of Cambridge published the results of the first human clinical trial of a vaccine whose active component was entirely designed by artificial intelligence. Not assisted. Not accelerated. Designed. The vaccine candidate pEVAC-PS, developed by biotech DIOSynVax, was administered to 39 healthy volunteers at two National Institute for Health Research sites in the United Kingdom, and the results are unequivocal: safety confirmed, excellent tolerance, cross-immune responses detected against several coronaviruses in the Sarbecovirus subgroup.

What changes the game is the very nature of what the AI produced. According to Medical Xpress, the algorithm did not simply sift through existing protein libraries faster. It designed an antigenic structure — a "super-antigen" — that human science had not envisioned. An unprecedented molecular architecture, conceived to simultaneously target thousands of coronavirus variants, including those that do not yet exist.

The implication is staggering. If an AI can design a universal vaccine in a few days of computing, the next pandemic could be countered before it even spreads. We are not there yet — Phase II still lies ahead — but the signal sent by Cambridge is clear: AI in biology is no longer an assistance tool, it is an autonomous designer.

The Essentials

• The pEVAC-PS vaccine is the world's first whose active component (a "super-antigen") was entirely designed by AI, with no human intervention on the molecular structure.
• The Phase I trial, conducted on 39 healthy volunteers in the United Kingdom, confirms the safety and good tolerability of the candidate, with no significant adverse reactions according to EMJ Reviews.
• Cross-immune responses were observed against several Sarbecovirus coronaviruses, including SARS-CoV-1, SARS-CoV-2, and other future threats from the same family.
• Delivery is needle-free, via microfluidic jet — a technical innovation coupled with the biological innovation.
• The next steps include a Phase II trial to evaluate the actual protective efficacy, as Science & Vie points out.

AI tools and models involved in the design

The design of pEVAC-PS relies on proprietary computational pipelines developed by DIOSynVax, derived from recent advances in protein modeling. AI analyzed the genetic data of Sarbeco viruses collected by global surveillance programs, according to Université de Cambridge.

The DIOSynVax pipeline follows on from the revolution sparked by AlphaFold, but goes further. AlphaFold predicts the structure of an existing protein. The DIOSynVax system, on the other hand, generates new antigenic structures that do not exist in any living organism — this is de novo design.

How AI designed this "super-antigen" that no one had imagined

The underlying problem: cross-immunity is a headache

Coronaviruses of the Sarbecovirus subgroup share structural similarities, but their variable regions mutate rapidly. A classic vaccine targets a specific surface protein (like the SARS-CoV-2 spike protein). The result: as soon as a new variant modifies this protein, the vaccine loses efficacy.

The universal approach requires targeting conserved regions — those that do not mutate from one coronavirus to another. The problem? These regions are often poorly immunogenic. The immune system ignores them in favor of the variable regions, which are more visible.

What AI did differently

DIOSynVax's AI analyzed the genetic data from thousands of Sarbeco coronaviruses collected by global surveillance programs. It identified the conserved regions shared by this entire viral family.

But the crucial step is not the identification — it's the design. The algorithm generated an artificial antigenic structure that presents these conserved regions in a way that maximizes their visibility to the immune system. This is the "super-antigen" that does not exist in any natural virus.

As Medical Xpress points out, AI didn't simply accelerate discovery — it designed a structure that human science had not envisaged. This is a fundamental distinction. We have moved from AI as a research tool to AI as a molecular designer.

The parallel with AlphaFold

In 2020, DeepMind's AlphaFold solved the protein folding problem: predicting the 3D structure of a protein from its amino acid sequence. This was already a revolution. But AlphaFold works on what already exists.

The Cambridge project goes a step further: it involves generating a protein that does not exist, optimized for a specific functional objective (triggering cross-immunity). This is the shift from prediction to creation. Current agentic LLM models like GPT-5.5 (agentic score 98.2) or Gemini 3 Pro Deep Think (score 95.4) are paving the way for this type of autonomous design in other fields.

Trial results: safety, tolerability, immune responses

Protocol and participants

The Phase I trial was conducted on 39 healthy volunteers at two National Institute for Health Research (NIHR) sites in the United Kingdom, according to Pharmaphorum. The goal of a Phase I is not to demonstrate efficacy, but to verify safety and tolerability.

Safety: the verdict is positive

The pEVAC-PS vaccine proved to be safe and well tolerated. No significant side effects were reported according to EMJ Reviews. This is an essential result: an AI-designed artificial antigen could have triggered unpredictable immune responses. This was not the case.

Cross-immune responses: the encouraging signal

Beyond safety, the trial showed that pEVAC-PS is capable of triggering cross-immune responses against multiple coronaviruses, as reported by ScienceDaily. Specifically, the volunteers' immune systems produced antibodies and cellular responses that recognize not only SARS-CoV-2, but also SARS-CoV-1 and other coronaviruses in the Sarbeco family.

This result is preliminary — the sample size (39 people) does not allow for robust statistical conclusions on efficacy. But the biological signal is there. Open Access Government confirms that the vaccine candidate is "capable of triggering immune responses against coronaviruses," which fully justifies moving on to Phase II.

Needle-free delivery: a detail that matters

An often underestimated aspect: the vaccine is delivered via microfluidic jet, a needle-free technology. The BBC mentions this in its reporting, and it is strategic. The acceptability of a universal vaccine also depends on how easy it is to administer. No needle means less training required for healthcare workers, less medical waste, and better patient acceptance.

From design to testing: the exact role of AI in the pipeline

Step 1 — Genetic data collection and analysis

AI ingested the genetic sequences of thousands of Sarbeco coronaviruses collected by global surveillance programs. This step falls under massive data analysis — a field where generalist LLMs like DeepSeek V4 Pro (score 88) already excel.

Step 2 — Identification of conserved regions

The algorithm identified stable protein regions across the entire Sarbecovirus family. These regions are ideal targets for a universal vaccine, since they do not mutate.

Step 3 — De novo design of the super-antigen

This is the step where AI goes beyond the role of assistant. It generated an artificial protein structure that presents the conserved regions in a way optimized for the immune system. This step involves 3D modeling, protein-antibody interaction simulation, and multi-objective optimization.

Step 4 — In silico validation before synthesis

Even before manufacturing the vaccine, AI simulated the interactions between the super-antigen and immune cells. This is where the advanced reasoning capabilities of models like Claude Opus 4.7 (score 90) or Gemini 3 Pro Deep Think (score 90 in reasoning) become relevant for similar pipelines.

Step 5 — Synthesis, preclinical trials, and then human trials

Once the structure has been digitally validated, the protein is synthesized in the lab, tested on animal models, and then on humans. The Cambridge Phase I trial is the culmination of this pipeline.

What this changes for the management of future pandemics

Response time is the key factor

During the emergence of SARS-CoV-2, the development of the first vaccine took about 11 months — a historic record. But 11 months is still too long when a virus spreads at the speed of global air travel.

With an AI de novo design pipeline, the time to design a vaccine candidate can be reduced to a few days. The rest of the process (synthesis, preclinical trials, clinical trials) remains long, but the main bottleneck — discovering the right antigenic target — disappears.

From reaction to anticipation

The Cambridge vaccine is not designed for a specific virus. It targets an entire family of coronaviruses, including those that have not yet infected humans. This is a paradigm shift: we move from manufacturing vaccines after the appearance of a threat to preparing defenses before emergence.

This proactive approach is made possible by AI. The algorithm identified vulnerabilities shared by all known Sarbecoviruses, and designed an antigen that exploits them. If a new coronavirus from this family emerges, the vaccine is theoretically already ready.

Current limitations

We must remain honest: Phase I does not prove efficacy. As Science & Vie reminds us, pEVAC-PS still needs to prove that it effectively protects against coronaviruses. Phase II, which will evaluate efficacy on a larger sample, is the real trial by fire.

Furthermore, this vaccine only targets the Sarbecovirus family. A virus from another family (influenza, hantavirus, etc.) would require a separate design pipeline. Universality is relative — it is universal within a viral family, not against all existing viruses.

AI in biology: beyond vaccines

From AlphaFold to drug design

Protein structure prediction by AlphaFold (DeepMind) paved the way. But the Cambridge project represents the next step: the generation of functionally optimized proteins. This is no longer analysis; it is engineering.

This evolution follows the same trajectory as LLMs: first capable of completing text, then of reasoning, then of acting autonomously via agents. In computational biology, we are moving from prediction to creation.

LLMs as scientific research tools

Current models are playing an increasing role in research pipelines. OpenAI's GPT-5.5, with its agentic score of 98.2, can orchestrate complex research workflows. Anthropic's Claude Opus 4.7 excels at analyzing scientific literature and synthesizing clinical data. Google's Gemini 3.1 Pro is particularly effective for genetic sequence analysis.

These models do not replace specialized pipelines like DIOSynVax's. But they significantly accelerate the upstream and downstream steps: literature review, trial data analysis, protocol writing, and identification of therapeutic targets.

The emergence of autonomous AI agents in research

The concept of an autonomous AI agent — a system that plans, executes, and iterates without constant human supervision — finds one of its most promising use cases in biology. The DIOSynVax pipeline is an example of this: the AI designed the super-antigen relatively autonomously, with the researchers validating it a posteriori.

This dynamic raises fascinating questions. How far can we let an algorithm go in designing molecules intended to be injected into the human body? Current regulations are not designed for this scenario. Agencies like the MHRA in the UK or the FDA in the United States will need to adapt their frameworks.

The DIOSynVax ecosystem and the UK context

Who is behind pEVAC-PS?

DIOSynVax is a British biotech company spun out of the University of Cambridge. The start-up specializes in the computational design of vaccines, using AI algorithms to generate optimized antigens. pEVAC-PS is its most advanced candidate.

The trial took place at two NIHR sites, the UK NHS's clinical research network, as detailed by Pharmaphorum. This choice is not insignificant: the UK has invested heavily in its clinical trial infrastructure post-Brexit, and the Cambridge results are also an institutional victory.

Funding and strategy

Universal vaccines are a complex area of funding. They are expensive to develop, but their market is uncertain — a vaccine against a future pandemic has no immediate customer. DIOSynVax relies on public funding (UK Research and Innovation) and biotech-focused investors.

The Phase I success strengthens the company's position to raise the funds necessary for Phase II, which will require a much larger sample and multiple clinical sites.

Comparison with other universal vaccine approaches

Several teams worldwide are working on universal vaccines against coronaviruses. The Cambridge approach stands out due to the central role of AI in designing the antigen itself.

The advantage of the DIOSynVax approach is its ability to generate novel structures. The drawback is the regulatory risk: no framework existed to evaluate an antigen that does not exist in nature. Phase I has partially lifted this uncertainty.

Next steps: Phase II and beyond

What Phase II will need to demonstrate

Phase II will evaluate the real-world efficacy of pEVAC-PS. The key questions: are the cross-immune responses observed in Phase I sufficient to protect against infection? Does the vaccine protect against coronaviruses that the volunteers have never encountered? What is the duration of immunity?

The trial will require a larger sample (likely several hundred participants) and could include challenge tests or studies in endemic areas.

Logistical and regulatory challenges

An AI-designed vaccine raises unprecedented questions for regulators. How do you evaluate the long-term safety of a molecule that does not exist in nature? The Phase I data are reassuring, but regulatory agencies will want longer-term follow-up data.

The jet microfluid delivery technology, while promising, also adds a layer of regulatory complexity. It is a medical device combined with a biological product — two evaluation pathways that must align.

The market availability horizon

Even if Phase II is successful, a Phase III will be required before market launch. In the most optimistic scenario, pEVAC-PS would not be available before 2028-2029. But in the event of a new coronavirus pandemic, regulators could accelerate the process via emergency procedures — and the vaccine would already be designed.

❌ Common errors in understanding this result

Error 1: Confusing "AI-designed" and "AI-accelerated"

Many commentators talk about AI "accelerating" vaccine discovery. In the case of pEVAC-PS, this is inaccurate. AI did not accelerate an existing human process — it designed an antigenic structure that human science had not envisioned. The distinction is fundamental. Accelerating is doing the same thing faster. Designing is doing something new.

Error 2: Believing the vaccine is already ready

Phase I proves safety, not efficacy. Headlining "AI created a vaccine that works" is misleading. The vaccine is safe and triggers immune responses — this is encouraging, but insufficient. Phase II is the real test. Science & Vie is right to headline "it's time for the verdict": the trial is beginning, not ending.

Error 3: Thinking that "universal" means "anti-everything"

The vaccine is universal within the Sarbecovirus family. It does not protect against the flu, dengue, or even against other coronavirus families (like those of the common cold, which belong to other subgenera). The universality is targeted, not global.

Error 4: Downplaying the role of needle-free delivery

Jet microfluidics technology is not a gadget. It determines the capacity for large-scale deployment. A universal vaccine that requires an intramuscular injection by trained personnel will always be slower to deploy than a vaccine that can be administered by jet. This is a strategic element of the project.

❓ Frequently asked questions

What is an AI-designed "super-antigen"?

It is an artificial protein structure, which does not exist in any natural organism, designed by algorithm to present the conserved regions of several coronaviruses to the immune system simultaneously. AI optimized its geometry to maximize the cross-immune response.

Why are 39 volunteers enough?

For a Phase I, yes. The goal is solely to assess safety and tolerability, not efficacy. 39 participants is a standard sample size for this type of trial. Phase II, on the other hand, will require a much larger sample.

Does this vaccine protect against future Covid variants?

Theoretically yes, if they belong to the Sarbecovirus family. The super-antigen targets the conserved regions of this family, which do not mutate. But this remains to be confirmed by Phase II with tests against specific variants.

What exact AI was used?

The computational pipeline is proprietary to DIOSynVax and has not been made public. It relies on protein modeling techniques similar in principle to AlphaFold, but extended to de novo structure generation.

When will this vaccine be available?

Not before 2028-2029 in a normal scenario, after Phases II and III. In the event of a pandemic emergency, this timeline could be reduced via accelerated authorization procedures, the vaccine already being designed.

Could AI design vaccines against other diseases?

That is the ambition. The same de novo design pipeline could theoretically be applied to other viral families (flu, HIV, dengue). But each family requires high-quality genetic data and an understanding of the specific immune mechanisms.

✅ Conclusion

The Phase I trial of the pEVAC-PS vaccine marks a turning point: for the first time, a vaccine component entirely designed by AI has been injected into humans and proven safe. AI didn't just accelerate science — it expanded it into structures no one had imagined. Phase II will tell whether this promise turns into real protection. Until then, one thing is certain: AI in biology is no longer experimental. It is producing clinical results.