Beyond Spike: What GSK’s Two-Antigen Flu Vaccine Tells Us About Programmable Vaccine Design

The next generation of mRNA vaccines may be defined less by the novelty of mRNA itself than by a harder question — how do we decide what the mRNA should encode, how it should be engineered, how it should be delivered, and which combination of those choices is worth testing? That question has become particularly relevant following new results from GSK’s seasonal influenza programme.

On 1 September 2026, GSK announced that its investigational mRNA seasonal influenza vaccine would advance into Phase III development, with the trial expected to begin that same month, following positive Phase II results. The Flu-028 study enrolled 971 adults aged 18 years and older and compared several mRNA vaccine candidates with licensed age-appropriate influenza vaccines; according to GSK, its selected candidate generated higher immune responses against all influenza strains tested than licensed standard-dose comparators in younger adults and high-dose comparators in older adults, while showing an acceptable safety and reactogenicity profile. [1]

The development is noteworthy not simply because another mRNA influenza vaccine is progressing through clinical development, but because of what GSK has chosen to encode: its candidate targets both haemagglutinin (HA) and neuraminidase (NA), making it an instructive example of a broader transition in vaccine development, from treating a vaccine as a relatively fixed biological product towards thinking about it as a multivariable design problem with real implications for computational biology.

Influenza vaccines have traditionally concentrated on HA

Influenza virions display two major surface glycoproteins, haemagglutinin and neuraminidase. HA facilitates viral attachment and entry into host cells, while NA plays a different role, promoting the release of newly formed virions and facilitating viral spread — and both can be targets of protective immunity.

Yet conventional seasonal influenza vaccine development has historically concentrated heavily on HA: licensed influenza vaccines are principally standardised according to HA content, while the amount and immunogenicity of NA can be considerably less controlled. [2] There are good historical and biological reasons for that emphasis, but it does not necessarily mean that HA represents the complete solution.

A growing literature suggests that antibodies directed against NA can contribute independently to protection, including through inhibition of viral replication and reductions in disease severity. A 2025 systematic review of human responses to influenza vaccination concluded that NA immunity represents an increasingly important independent correlate of protection, while highlighting considerable variability in NA content and responses among existing vaccine approaches, [3] and a recent 2026 review similarly argues that NA-specific immunity may be particularly useful where HA-directed immunity is poorly matched to circulating strains. [2]

GSK’s approach therefore asks a subtly different question from the traditional influenza-vaccine model: rather than which HA should the vaccine contain?, the design problem begins to resemble which combination of viral targets should be encoded, in what form, and with what delivery architecture? — and that distinction matters.

mRNA changes the size of the design space

One attraction of mRNA is its programmability: changing an encoded antigen does not necessarily require rebuilding the entire manufacturing platform around a new recombinant protein, and multiple antigens can also potentially be encoded within a single vaccine formulation. But programmability introduces a paradox — the easier it becomes to construct different biological designs, the more difficult it becomes to decide which designs should actually be constructed.

Consider even a simplified influenza mRNA vaccine. A development programme might need to make decisions about antigen selection, strain selection, antigen combinations, antigen sequence engineering, coding-sequence optimisation, 5′ and 3′ untranslated regions, RNA secondary structure, nucleoside chemistry, poly(A) architecture, relative RNA quantities, ionisable lipid chemistry, helper lipid composition, cholesterol composition, PEG-lipid composition, lipid ratios, particle size and physicochemical properties, dose, tissue distribution, cellular uptake, endosomal escape, innate immune activation, antigen expression, reactogenicity and manufacturing conditions — and these variables do not necessarily act independently: changing an untranslated region can alter translation, changing RNA structure can alter stability, changing an ionisable lipid can affect cellular uptake, endosomal escape, biodistribution and tolerability, and changing antigen dose can alter both immunogenicity and reactogenicity.

The vaccine therefore cannot always be reduced to antigen + carrier; it is more accurately considered a coupled biological system. Recent research illustrates how large these individual design spaces have already become — in 2026, work integrating ionisable-lipid engineering with RNA optimisation screened a 96-member biodegradable lipid library while also using massively parallel experiments to rank more than 630,000 3′ UTR variants and more than 68,000 5′ UTR sequences. [4] Those are only portions of the total system, and once multiple design dimensions are considered simultaneously, exhaustive experimental search rapidly becomes impractical.

The LNP is part of the programme

This problem becomes especially important when lipid nanoparticles are considered. LNPs are sometimes described simply as delivery vehicles, but that description increasingly understates their biological importance: an LNP typically contains an ionisable lipid, phospholipid, cholesterol and PEG-lipid, and variations in the chemistry, proportions and physical properties of these components can alter encapsulation, biodistribution, cellular uptake, intracellular trafficking, endosomal escape, immune interactions and ultimately therapeutic performance. [5,6] The delivery system is therefore not merely packaging around an mRNA programme — it is part of the programme.

Recent work increasingly describes LNPs themselves as programmable. A 2026 review in Materials Today Bio, for example, defines programmable LNPs as systems in which composition and interfacial chemistry can be manipulated to influence organ tropism, cellular specificity, intracellular trafficking and immune interactions, and identifies data-guided formulation, machine learning, in-vivo barcoding and physics-based prediction as increasingly important approaches to LNP development. [5]

The scale of the search problem is already prompting a change in experimental methodology. Li and colleagues demonstrated this in Nature Materials by synthesising 584 ionisable lipids, experimentally measuring mRNA transfection, using those results to train machine-learning models, and subsequently using the model to interrogate a virtual library of approximately 40,000 candidate lipids — only a small number of model-selected candidates then needed to be synthesised and tested experimentally. [7] Other investigators have gone further computationally: an AI-assisted study reported screening a virtual chemical space containing almost 20 million candidate ionisable lipids, using predictive models for properties including apparent pKa and mRNA delivery performance before experimental validation of selected molecules. [8]

This represents an important inversion of the traditional experimental workflow. Instead of make → test → make → test → make → test, the process increasingly becomes represent → predict → prioritise → make → test → learn → redesign — a fundamentally different model of biological engineering.

The sequence is becoming computational too

The same transformation is occurring on the RNA side. An mRNA molecule is not simply an antigen translated into nucleotides — its behaviour is influenced by coding sequence, codon usage, 5′ and 3′ UTRs, RNA secondary structure, stability, translation kinetics and interactions with cellular machinery. Jin and colleagues describe four major optimisation problems for mRNA vaccine design: UTR optimisation for translation, codon optimisation, global sequence optimisation for structural stability, and local structural optimisation affecting intracellular persistence and expression. [9]

These create enormous combinatorial spaces: a 5′ UTR only 200 nucleotides long theoretically occupies a sequence space on the order of 4²⁰⁰ possibilities, and experimental enumeration is obviously impossible. Machine learning and generative models therefore become useful not because they somehow replace molecular biology, but because molecular biology has created a search problem too large to navigate experimentally.

A 2026 review of AI-assisted mRNA design describes the field as moving towards coordinated optimisation of UTRs and coding sequences rather than optimising individual components independently. [10] This distinction may prove important, since optimising each component separately does not guarantee an optimal complete system.

From vaccine development to design-space navigation

This is where GSK’s HA-plus-NA influenza vaccine becomes interesting beyond influenza. Adding another antigen does not merely add another ingredient — it creates another design dimension: if a vaccine can encode HA and NA, it could theoretically encode different HA variants, different NA variants, engineered antigens, conserved epitopes or combinations thereof, each interacting with sequence design, expression level, dose and delivery, so the number of plausible candidate vaccines expands rapidly.

This creates a problem familiar across engineering disciplines: when the space of things we can build grows faster than our ability to test them, experimentation alone is no longer an efficient search strategy, and the challenge becomes deciding which experiment contains the most information — a different computational objective from simply predicting whether a vaccine will work.

The goal becomes iterative learning: design → build → measure → learn → redesign. Experimental data update the model, the model identifies promising or informative regions of the design space, those candidates are tested, the resulting observations become new training data, and the cycle repeats.

Where Noviota fits

This is closely related to the problem Noviota is being built to explore. Noviota is developing computational approaches for representing, reasoning over and navigating programmable biological systems, particularly where sequence, formulation, delivery and biological response interact — not to replace experimental biology with an algorithm, nor to claim that biological performance can be predicted perfectly from sequence or formulation (biological systems are too context-dependent, heterogeneous and incompletely understood for that), but to make increasingly complex design spaces structured, computable and experimentally navigable.

For an mRNA-LNP programme, a computational representation might connect therapeutic objective, biological target, antigen architecture, RNA construct, sequence features, LNP composition, manufacturing parameters, physicochemical characteristics, cellular delivery, expression, immune response and experimental outcome, with each layer producing information that can constrain or inform another. This is one reason Molecular Precision, part of the Noviota group, is developing the concept of a Nanomedicine Design Stack: rather than treating sequence design, formulation, experimental results and manufacturing information as disconnected datasets, the aim is to represent them as parts of the same design system.

The distinction is subtle but important. A database tells us what was tested; a computational design system should increasingly help us reason about what should be tested next — and why.

Negative results matter too

There is another consequence. Most biological development programmes generate far more failed candidates than successful ones, yet those failures frequently disappear into laboratory notebooks, internal reports or poorly structured datasets — even though computational design makes those negative results valuable: if formulation A failed while formulation B succeeded, the distinction can constrain subsequent search; if an RNA sequence produced excellent expression but unacceptable innate immune activation, that result contains information; and if an LNP performs well in hepatocytes but poorly in antigen-presenting cells, that difference contains information too.

The value lies not only in discovering the successful candidate, but in progressively constructing a map of the design landscape. This also means provenance becomes important — models require knowledge of exactly which sequence, formulation, manufacturing process, assay and experimental conditions produced each observation, since without that context, apparently comparable biological measurements may not actually be comparable. Programmable biology therefore creates both a computational-design problem and an information-architecture problem.

The laboratory remains inside the loop

There is a temptation when discussing AI-driven biological design to imagine autonomous systems predicting optimal medicines directly, but that is not the near-term reality: biological prediction remains uncertain, in-vitro performance may not translate in vivo, animal models may not translate into humans, LNP biodistribution is influenced by biological interactions that are incompletely understood, manufacturing changes can affect critical quality attributes, and immune responses vary between individuals. The experimental loop therefore remains essential.

The more realistic model is model-guided experimentation, in which computational systems reduce the number of experiments required to explore a design space effectively while experiments continually correct the computational representation — a productive division of labour in which computation explores possibility and biology supplies reality.

Why influenza is a useful demonstration

Seasonal influenza provides an unusually good illustration of this transition. It is not a static target: influenza evolves continuously, vaccine composition is repeatedly updated, different populations respond differently, antigenic match varies between seasons, and manufacturing timelines matter. A programmable vaccine platform potentially allows more variables to be changed more rapidly, but flexibility alone does not solve the problem — it increases the number of decisions.

The strategic advantage therefore may not ultimately belong simply to whoever can manufacture mRNA fastest; it may belong to whoever can most effectively integrate surveillance data, antigen selection, sequence engineering, delivery design, experimental evidence and manufacturing constraints into a coherent design cycle. GSK’s decision to advance a vaccine targeting both HA and NA is only one development programme, and its Phase II immunogenicity results do not establish clinical superiority — Phase III efficacy data will be required before conclusions about protection can be drawn. [1]

But scientifically, the programme illustrates something larger. Vaccines are becoming increasingly configurable, RNA is becoming increasingly engineerable, delivery systems are becoming increasingly programmable, and experimental datasets are becoming large enough for computational methods to influence what gets built next. The resulting bottleneck is no longer simply our ability to construct biological systems — increasingly, it is our ability to navigate the space of possible biological systems intelligently, and that is the problem Noviota intends to work on.