AI Says It Can Beat the WHO at Picking Flu Vaccines. Can It?

Twice a year, a small group of virologists convenes on behalf of the World Health Organization to make an educated guess that affects the health of billions of people: which influenza strains will dominate the coming season, and therefore which strains belong in that year's vaccine (WHO Global Influenza Programme). Most years, the guess holds up well enough. Some years it doesn't: during the 2014-2015 season, the H3N2 strain in the vaccine had already been overtaken by a mutated version of the virus, and protection suffered as a result (Chambers et al., Cell Reports, 2015).

On 18 September 2026, researchers from Northwestern Polytechnical University, the Harbin Veterinary Research Institute and the Chinese Academy of Sciences published a new tool built to sharpen that guess. Called VirPLM, it repurposes an AI system originally designed to understand the shape and behaviour of proteins, and points it at one of virology's oldest problems: forecasting how flu will look to the human immune system next winter (Li et al., Bioinformatics, 2026). It's a careful piece of work. It also arrives in the middle of an unresolved fight, playing out this year in one of medicine's most-read journals, over whether this entire category of AI tool does what it claims.

What happened#

VirPLM is a two-stage system built around ESM-2, a protein language model that learns general patterns in protein sequences by training on millions of them, in the same way a large language model learns patterns in text (Li et al., 2026). In stage one, the team fine-tuned ESM-2 on the HA1 subunit, the part of the flu virus's surface protein that antibodies recognise, across the historical record of human H3N2 sequences. In stage two, they trained the adapted model to predict antigenic distance: how different two virus samples look to the immune system, a property normally measured in the laboratory using haemagglutination inhibition (HI) assays.

Those assays remain the gold standard, but they are slow, and every one requires live virus, ferret or human sera, and a trained technician. VirPLM instead works directly from a genetic sequence, which is far cheaper to generate and now accumulates by the thousands each season. Tested against other sequence-based methods, under both standard cross-validation and a stricter retrospective test that only ever looks backward in time (mimicking how the tool would actually be used), VirPLM performed consistently better, and it flagged mutation sites that independently line up with regions already known to drive antigenic drift. In a season-by-season comparison, strains that VirPLM ranked highest achieved better estimated coverage of that season's circulating viruses than the strains the WHO actually recommended, in most of the seasons tested (Li et al., 2026). The code and trained models are freely available on GitHub, with a fixed version archived on Zenodo, and the paper itself is open access.

Why flu keeps outrunning its vaccine#

Influenza's surface is covered in haemagglutinin, a protein shaped like a small spike that lets the virus latch onto and enter human cells. It's also the main target of the antibodies a flu vaccine trains the body to produce. The trouble is that haemagglutinin, and specifically its HA1 subunit, mutates constantly under pressure from those same antibodies, a process called antigenic drift. Enough drift, accumulated over a year or two, and antibodies raised against last year's strain stop recognising this year's, which is why the vaccine has to be reformulated on a rolling basis rather than fixed once and left alone.

Deciding what to put in that reformulation is WHO's job, carried out through its Global Influenza Surveillance and Response System, with recommendations issued every February for the Northern Hemisphere season and every September for the Southern Hemisphere one (WHO). The recommendation has to be made roughly six to eight months before the season it covers, because manufacturing an egg- or cell-based vaccine at global scale takes that long, and the virus keeps evolving for the entire gap. That timing problem is exactly what has drawn AI researchers to the field: a protein language model like ESM-2 doesn't get tired, doesn't need a live virus sample, and can in principle score thousands of candidate mutations overnight. Whether it scores them accurately enough to matter is the open question.

Why this matters#

If sequence-based tools like VirPLM prove reliable, they offer something WHO's expert panel cannot generate on its own: a fast, cheap, standing estimate of antigenic drift that can be recomputed the moment new sequences are deposited in public databases, rather than only twice a year. That would not replace laboratory testing or expert judgement, but it could give both an earlier warning when a new variant is heading somewhere unwelcome, in between the formal recommendation meetings.

There is a broader story here too, about what protein language models are turning out to be good for. ESM-2 was built to help predict protein structure and function, not to forecast evolution. VirPLM is part of a small but growing body of work, including tools such as PREDAC-FluB for influenza B and FluEvoFormer for prospective strain ranking, asking whether the same underlying representations can be repointed at genuinely predictive, forward-looking tasks in virology. That is a meaningfully harder test of these models than the structure-prediction benchmarks they are usually judged on.

Critical analysis#

Here is the part of the story that VirPLM's own paper does not mention, but that anyone reading it should know. VirPLM is not the first tool to claim it can out-predict WHO's flu vaccine strain choices, and its predecessor's experience is a caution against taking that claim at face value.

In August 2025, a team at MIT published VaxSeer, an AI system that models both how flu strains are likely to spread and how well antibodies will recognise them, then combines the two into a single "coverage" score (Shi et al., Nature Medicine, 2025). Using a decade of retrospective data, the authors reported that VaxSeer's picks beat WHO's actual recommendations for H3N2 in nine of the ten seasons tested, and matched or beat them for H1N1 in six of ten, according to MIT's own announcement of the work (MIT News, 2025). The paper also reported that its coverage score correlated with real-world vaccine effectiveness and averted disease burden, which is what turned a methods paper into a public health story.

Ten months later, virologists Simon de Jong and Colin Russell published a formal challenge to that result in the same journal (de Jong & Russell, Nature Medicine, 2026). Their critique made two specific technical points. First, they argued that the headline correlation with vaccine effectiveness only appears when data from different flu subtypes are pooled together, and disappears once each subtype is examined on its own, which is the only comparison that is biologically meaningful, since H3N2 and H1N1 behave differently and are not interchangeable. That pattern, where an association appears in combined data but vanishes or reverses within each subgroup, is a textbook statistical trap known as Simpson's paradox. Second, they pointed out that the disease-burden figures VaxSeer was validated against are themselves calculated by public health agencies using vaccine effectiveness as an input, so using them to independently confirm a vaccine-effectiveness prediction is somewhat circular. In their published reply, Shi and colleagues defended the broader framework but acknowledged that VaxSeer's edge over a simple, untrained baseline model, when tested the way the critics specified, did not reach statistical significance (Shi et al., reply, Nature Medicine, 2026).

None of this means VirPLM inherits VaxSeer's specific flaws. VirPLM is built differently, restricts itself to one subtype (H3N2) rather than pooling across strains, and is not, in its published form, validated against real-world vaccine effectiveness or disease burden at all, only against retrospective antigenic coverage, a narrower and more defensible claim. But that narrower claim is also less than the "beats the WHO" framing suggests: outperforming a historical vaccine choice on estimated antigenic coverage, in hindsight, is not the same as prospectively guiding a real recommendation, and VirPLM has not yet been tested against a simple, non-AI baseline model in the way that exposed VaxSeer's weakest point. Given how recently the paper was published, that scrutiny simply has not happened yet. That's worth flagging as an open question, not a settled shortcoming.

Expert perspective#

VirPLM and VaxSeer sit at opposite ends of how one might build an AI flu-forecasting tool. VaxSeer explicitly models two biological processes, viral fitness (which strains will spread) and antigenicity (which strains antibodies will recognise), each with its own machine-learning component, then multiplies the two together into a single score. VirPLM instead fine-tunes a general-purpose protein language model to predict antigenic distance directly from sequence, without an explicit fitness model, and leans on the fact that ESM-2 has already learned a great deal about protein biophysics from far more data than any flu-specific dataset alone could provide.

That distinction mirrors a wider pattern in computational biology over the past few years: general-purpose foundation models, fine-tuned for a narrow task, are increasingly competitive with bespoke systems built from scratch for that task, provided the underlying model has learned something transferable in the first place. Earlier work such as "Language models learn to represent antigenic properties of human influenza A(H3) virus" established that protein language model embeddings do carry usable antigenic signal; VirPLM's contribution is a more thorough fine-tuning and evaluation pipeline built on that foundation, rather than a wholly new idea. Given the pace of publication in this niche, and the unresolved argument still hanging over VaxSeer, this will not be the last word on whether any of these tools are ready to inform an actual WHO recommendation. The more useful marker of progress will be a prospective test, a model whose predictions are locked in before a season starts, not measured against it afterwards, and no such trial exists yet for VirPLM or its predecessors.

Key takeaways#

VirPLM, published 18 September 2026 in Bioinformatics, fine-tunes the ESM-2 protein language model to predict how flu's HA1 protein will drift, and reports better estimated antigenic coverage than WHO's historical vaccine choices in most seasons tested (Li et al., 2026).

It is not the first AI tool to make this kind of claim. VaxSeer, from MIT, reported similar results in Nature Medicine in 2025 and was widely covered in the press (Shi et al., 2025; MIT News, 2025).

VaxSeer's central evidence was formally challenged in the same journal in 2026, on the grounds of a statistical artefact (Simpson's paradox) and a circular validation metric (de Jong & Russell, 2026).

VirPLM's narrower, single-subtype design avoids that specific flaw, but has not yet been tested against a simple non-AI baseline or validated against real-world vaccine effectiveness.

The underlying problem, flu mutating faster than a twice-yearly committee process can fully track, is real and well documented, which is why cheap, sequence-based tools remain worth pursuing even amid the current uncertainty over whether any of them yet outperform expert judgement.

Frequently asked questions#

What is VirPLM? A computational tool that fine-tunes ESM-2, a general-purpose protein language model, to predict how the flu virus's HA1 surface protein will change in ways that let it evade existing antibodies (Li et al., 2026).

What is a protein language model? An AI system trained on large numbers of protein sequences to learn general patterns in how amino acids combine to form working proteins, similar in structure to how language models like ChatGPT learn patterns in text. ESM-2 is one such model, originally developed for protein structure and function prediction.

Does VirPLM actually beat the WHO's flu vaccine choices? Its authors report higher estimated antigenic coverage than WHO's historical picks in most of the seasons they tested, in a retrospective analysis. That is different from a prospective trial made before a season starts, which has not yet been done for VirPLM (Li et al., 2026).

What was wrong with the earlier VaxSeer claim? Critics Simon de Jong and Colin Russell argued its headline result depended on pooling data across flu subtypes in a way that created a statistical illusion (Simpson's paradox), and that a second validation metric was circular because it used vaccine effectiveness data to "confirm" a vaccine effectiveness prediction (de Jong & Russell, 2026).

Did VaxSeer's authors accept the criticism? Partly. In their published reply, they defended the overall approach but acknowledged that VaxSeer's advantage over a simple untrained baseline was not statistically significant under the comparison the critics specified (Shi et al., reply, 2026).

How does the WHO currently choose flu vaccine strains? Through its Global Influenza Surveillance and Response System, which issues recommendations every February for the following Northern Hemisphere season and every September for the Southern Hemisphere one, based on laboratory antigenic testing and genetic surveillance data pooled from national influenza centres worldwide (WHO).

Could tools like VirPLM replace that process? Nothing published so far supports that. The realistic near-term role for these tools, if they hold up to further scrutiny, is as an additional, cheaper signal that complements laboratory testing and expert review, not a replacement for either.

Is this the same as the AI-designed-virus stories from earlier in 2026? No. Those stories involved AI generating entirely new viral genomes. VirPLM does not design anything; it analyses naturally occurring flu sequences to predict how the virus is likely to change.

Glossary#

Antigenic drift: The gradual accumulation of mutations in a virus's surface proteins that lets it evade antibodies raised against earlier versions of the virus.

Haemagglutinin (HA): The spike-shaped protein on the surface of the influenza virus that lets it bind to human cells; it is also the main target of vaccine-induced antibodies. HA1 is its outer, most variable subunit.

Protein language model: An AI model trained on large numbers of protein sequences to learn general patterns in protein structure and function, which can then be adapted (fine-tuned) for specific tasks.

ESM-2: A widely used protein language model, originally developed to help predict protein structure, that VirPLM adapts for flu antigenic prediction.

Haemagglutination inhibition (HI) assay: A laboratory test that measures how well antibodies block a virus from binding to red blood cells, used as the traditional gold standard for assessing antigenic similarity between virus strains.

Fine-tuning: Further training a general-purpose AI model on a smaller, specific dataset so that it performs better on a narrower task.

Simpson's paradox: A statistical pattern in which a trend appears in a combined dataset but disappears, or reverses, when the data are separated into their proper subgroups.

Retrospective versus prospective evaluation: A retrospective evaluation tests a model's predictions against outcomes that are already known; a prospective evaluation locks in predictions before the outcome occurs, which is a stricter and more convincing test.

References#

Li, X., Xiao, K., Zhou, C., Jia, X., Zhao, D., Xu, J., Zeng, X., Shi, J., Shang, X., Zhu, J., Kong, H. "VirPLM: Antigenic prediction of influenza A/H3N2 viruses with a fine-tuned protein language model." Bioinformatics, 18 September 2026. DOI: 10.1093/bioinformatics/btag692. Peer-reviewed, open access. Code: GitHub.
Shi, W., Wohlwend, J., Wu, M., Barzilay, R. "Influenza vaccine strain selection with an AI-based evolutionary and antigenicity model." Nature Medicine, 28 August 2025. DOI: 10.1038/s41591-025-03917-y. Peer-reviewed.
MIT News. "MIT researchers develop AI tool to improve flu vaccine strain selection." 28 August 2025. Link. Official institutional source.
de Jong, S. P. J., Russell, C. A. "Limited evidence of AI superiority in seasonal influenza vaccine strain selection." Nature Medicine, 2026. DOI: 10.1038/s41591-026-04461-z. Peer-reviewed.
Shi, W., Wohlwend, J., Wu, M., Barzilay, R. "Reply to: Limited evidence of AI superiority in seasonal influenza vaccine strain selection." Nature Medicine, 2026. DOI: 10.1038/s41591-026-04460-0. Peer-reviewed.
World Health Organization. "Recommendations for influenza vaccine composition." Global Influenza Programme. Link. Official organisation source.
Chambers, B. S., Parkhouse, K., Ross, T. M., Alby, K., Hensley, S. E. "Identification of Hemagglutinin Residues Responsible for H3N2 Antigenic Drift during the 2014-2015 Influenza Season." Cell Reports, 25 June 2015. DOI: 10.1016/j.celrep.2015.06.005. Peer-reviewed.
Yang, F. et al. "PREDAC-FluB: predicting antigenic clusters of seasonal influenza B viruses with protein language model embedding based convolutional neural network." PMC12264208. Peer-reviewed.
"FluEvoFormer: A Structure-Guided Generative Foundation Model for Prospective Influenza Antigenic Evolution and Vaccine Strain Selection." PMC13517461. Peer-reviewed.
"Language models learn to represent antigenic properties of human influenza A(H3) virus." Scientific Reports. Link. Peer-reviewed.