AI Systems Already Improving Vaccine Development and Public Health

The cleanest test of AI in vaccine development is not whether a model can generate a plausible antigen or produce a more impressive heat map. It is whether the model changes a decision that someone actually has to make. VaxSeer, an influenza strain-selection system reported in Nature Medicine in 2025, is unusually useful because it was tested against a decision everyone in seasonal flu vaccination understands: which viral strain should go into the vaccine before the next season arrives.[1]

In a 10-year retrospective analysis of H3N2 influenza, VaxSeer outperformed World Health Organization recommendations in 9 of 10 years. More specifically, the WHO-recommended vaccine strain matched the best available strain in 0 of 10 years for H3N2, while VaxSeer matched the best available strain in 5 of 10 years. The system’s predicted coverage scores also correlated with CDC-estimated real-world vaccine effectiveness, with r=0.861 and p=0.0014.[1]

That is the kind of result that deserves attention because the comparator is not vague. The model was not merely ranking sequences in isolation; it was evaluated against historical strain choices and against an external estimate of vaccine effectiveness. The practical question was whether a system looking at viral evolution and immune escape signals could have selected strains with broader expected coverage than the recommendation process used at the time.

The phrase “outperformed WHO recommendations,” however, needs to stay attached to the design of the study. VaxSeer’s result was retrospective. It did not have to sit in a global meeting balancing manufacturing feasibility, regulatory timing, egg- or cell-based production constraints, supply agreements, national program logistics, and the practical limits of changing a recommendation once the vaccine-production clock has started. A retrospective benchmark can show that the historical recommendation left measurable antigenic coverage on the table. It cannot, by itself, prove that the same model should replace the recommendation process.

What VaxSeer actually accelerated

For influenza, the accelerated step is antigen selection under uncertainty. Seasonal vaccines have to be matched against viruses that continue to evolve after candidate strains are chosen. A model that can better estimate which candidate strain is likely to cover circulating variants is intervening at a real bottleneck: the point where surveillance data becomes a vaccine composition decision.

The important part of the VaxSeer study is not simply that machine learning was involved. It is that the authors could define the task, replay past decision windows, compare against actual recommendations, and ask whether the model’s coverage estimates lined up with observed vaccine effectiveness. That combination is still rarer than it should be in AI vaccine papers. Many systems predict something biologically interesting. Far fewer can be tied to a decision point where a different output would have implied a different vaccine choice.

That last row is not a small footnote. The public health value of better strain prediction depends on whether committees can use the output early enough, whether manufacturers can produce the selected strain reliably, and whether regulators can evaluate the resulting product within the necessary seasonal timeline. Still, VaxSeer gives the field something solid: a named system, a defined comparator, and a result that would have changed some historical choices if it had been trusted and deployable at the time.

DIOSynVax moves the question into humans

DIOSynVax answers a different question. VaxSeer is about selecting among candidate strains with historical validation. DIOSynVax is about whether an AI-designed vaccine candidate can make it through first-in-human testing with an acceptable early safety and immunogenicity profile.

The University of Cambridge spinout’s universal Sarbeco coronavirus vaccine was described as the first AI-designed vaccine to enter human trials. In June 2026, DIOSynVax completed a Phase 1 trial with 39 participants, demonstrating safety and broad immune responses against SARS-CoV-2, SARS, and bat coronaviruses.[2]

That is a milestone, but it is a narrow one. Phase 1 does not establish clinical effectiveness. It does not show population-level protection, durability, performance in older or immunocompromised groups, or superiority over existing coronavirus vaccine strategies. It shows that an AI-designed candidate was safe enough in a small first-in-human study and generated immune responses broad enough to justify further testing.

The distinction matters because “AI-designed vaccine passes Phase 1” is already strong enough. It does not need to be inflated into “AI has solved vaccine development.” In the vaccine pipeline, Phase 1 is a gate, not the destination. The candidate still has to face larger studies, clearer clinical endpoints, and the ordinary discipline of adverse-event monitoring, manufacturing scale-up, and regulatory review.

The less visible gains are upstream

Some of the most plausible uses of AI in vaccine development sit earlier than strain selection or first-in-human testing. They do not yet produce a licensed vaccine or a changed public recommendation, but they can reduce the number of biological possibilities that researchers have to carry forward.

PATH announced agentic AI workflows in April 2026 for identifying immune biomarkers, or correlates of protection, for rotavirus and RSV vaccines.[3] This is not the same claim as “AI found a vaccine.” A useful correlate of protection can help trial designers and regulators interpret immune responses without waiting for every possible clinical endpoint to mature. If validated, that kind of biomarker work can shorten or clarify development pathways because it gives teams a better signal for whether a candidate is likely to protect.

The evidence stage is still early. The PATH example is best read as an AI-assisted scientific workflow aimed at biomarker discovery, not as proof that an AI-identified correlate has already changed licensure decisions for rotavirus or RSV vaccines. Its value will depend on whether the candidate biomarkers survive biological validation and whether regulators accept them for specific development decisions.

UC Irvine’s broad-spectrum coronavirus work is another upstream example with a concrete computational narrowing step. Machine learning screened 15 million viral strains to identify 10 conserved proteins, and the reported preclinical timeline fell from an estimated roughly 12 years to roughly 4 years.[4] The acceleration here is not a completed clinical program. It is the reduction of an enormous search space into a smaller set of conserved targets that can be tested experimentally.

That is still meaningful. Preclinical vaccine research often loses time not because scientists lack imagination, but because the number of plausible antigens, epitopes, constructs, adjuvant combinations, and animal-study designs can expand faster than budgets and laboratory capacity. A system that narrows candidates intelligently can save years before a human trial is even designed. But the claim should remain attached to the stage: preclinical acceleration, not demonstrated protection in people.

Public health communication is a separate endpoint, not an afterthought

AI in vaccine development and public health communication should not be treated as one seamless pipeline. Designing or selecting a vaccine candidate is a different problem from getting a public agency to post a message that its audience might actually see. The data are different, the risks are different, and the outcome measures should be different.

The Penn LDI HIV social-media study is useful because it measured an institutional behavior rather than asking whether AI-generated language sounded persuasive in a lab. Across 42 U.S. counties, AI-driven campaign messages were six times more likely to be posted by health agencies.[5]

That is evidence of improved uptake by agencies, not evidence that HIV outcomes changed. It says that AI-generated or AI-selected messaging fit agency needs well enough to be used more often. For public health operations, that is not trivial. A message that never clears a local agency’s threshold for posting cannot influence anyone. But the next evidentiary step would be different: whether those posts changed awareness, testing, vaccination behavior where relevant, linkage to care, or other health outcomes.

This distinction is especially important for vaccine communication, where institutional trust, misinformation, and platform dynamics can distort even carefully designed campaigns. A system that helps agencies choose more postable messages could become valuable infrastructure. It could also produce shallow optimization if the endpoint is only “posted” and not “understood,” “trusted,” or “acted on.”

Why partial improvements still matter

The economic argument should not be the main evidence for AI in vaccines, but it explains why narrow improvements attract serious attention. A 2025 HHS-cited estimate placed the average cost of bringing a novel vaccine to U.S. markets at $886.8 million.[6] Against that background, shaving uncertainty from antigen selection, reducing failed preclinical candidates, or clarifying immune endpoints can matter even if no single model transforms the whole enterprise.

The most credible cases so far have a common shape. They do not ask AI to “discover vaccines” as a single grand task. They give it a constrained job: predict the best influenza strain from surveillance data, design a candidate antigen for a defined viral group, identify immune markers that may stand in for protection, narrow conserved coronavirus targets, or help agencies select messages they are more likely to use.

• There is a defined prediction or selection task.
• The input data are structured enough for the model to learn from them.
• The output can be compared with a meaningful benchmark or endpoint.
• A real decision point exists where the model output could change what humans do next.
• The claim is kept within the deployment stage actually tested.

VaxSeer fits this pattern best because the modeled decision is familiar and the comparator is measurable. DIOSynVax fits it differently because the model’s design work has now been tested in humans, albeit only at Phase 1. PATH and UC Irvine sit earlier, where the output is a more manageable set of biological hypotheses. Penn LDI sits downstream, where the decision is whether public agencies use a message.

The regulatory frame is still being built

Regulators are not ignoring AI, but the governance structure is still catching up to the variety of uses now appearing in drug and vaccine development. FDA published draft guidance on AI for drug development in January 2025, CDER established an AI Council in 2024, and EMA-FDA guiding principles were released in January 2026.[7]

Those steps matter because the regulatory question is not simply whether a model is accurate. It is whether the model is being used to generate evidence, select candidates, enrich a trial population, define an endpoint, support manufacturing decisions, or communicate with the public. Each use creates a different burden of transparency, validation, monitoring, and accountability.

Data access remains a practical constraint. Influenza strain selection depends on surveillance quality and timely sharing. Biomarker discovery depends on well-characterized clinical and immunologic datasets. Communication tools depend on local context, language, platform behavior, and agency capacity. Bias can enter at any of these points: in which sequences are available, which populations are represented in immune datasets, which counties or agencies are included in communication studies, and which outcomes are optimized.

The field is therefore in an uneven but important position. AI systems are already improving specific parts of vaccine development and public health communication: influenza strain selection, AI-designed candidate generation, immune-marker discovery, preclinical target narrowing, and agency-facing message selection. The strongest evidence comes where the task is bounded, the comparator is visible, and the output can be tied to a decision. That is not yet a broadly transformed vaccine ecosystem. It is a set of concrete advances that deserve to be judged case by case, with the same attention to endpoints and constraints that vaccine science has always required.

References

1. VaxSeer: an AI platform for forecasting influenza vaccine effectiveness and guiding strain selection. Nature Medicine, 2025.
2. First AI-designed vaccine completes Phase 1 clinical trial. ScienceDaily, June 5, 2026.
3. Using artificial intelligence to accelerate vaccine development. PATH, April 2026.
4. Artificial intelligence could usher in new era for vaccine development. CIDRAP, 2025.
5. AI Breakthrough May Transform Public Health Campaigns. Penn LDI, 2025.
6. New paradigms for vaccine development. The Lancet Infectious Diseases, 2025.
7. Artificial Intelligence for Drug Development. U.S. Food and Drug Administration.