AI found a thermostable mRNA vaccine recipe in six rounds—but the evidence is still preclinical

A peer‑reviewed study published September 28 in [Nature Biotechnology](https://www.nature.com/articles/s41587‑026‑03331‑w) reports an AI‑guided method for turning mRNA lipid nanoparticles into solid formulations that tolerate warm storage. The system, called AGENT, combines high‑throughput experiments with Bayesian optimization: each experimental round updates a model that chooses the next formulations likely to improve several stability targets at once. The researchers completed six optimization iterations in one month. That is the practical contribution of the AI component. It did not invent an mRNA vaccine or predict clinical efficacy; it selected promising combinations from a formulation design space using sparse experimental results. The distinction matters because “AI‑designed vaccine” would overstate what the paper tested. AGENT optimized the excipients and processing conditions used to preserve two clinically relevant lipid‑nanoparticle compositions after drying.

The paper reports that optimized solid‑state formulations based on the SM‑102 and ALC‑0315 lipid systems retained 100% bioactivity after storage at 37°C for more than two months. Those lipid systems are representative of the compositions used in Moderna and Pfizer‑BioNTech COVID‑19 vaccines. The result addresses a real distribution constraint: liquid mRNA‑LNP products can require cold or ultra‑cold handling because RNA and lipid particles degrade during storage. “Bioactivity” here refers to the formulations’ measured ability to deliver functional mRNA in the study’s assays. It does not mean 100% vaccine efficacy, nor does it establish shelf life under every shipping condition. A same‑day [Nature Biotechnology research briefing](https://www.nature.com/articles/s41587‑026‑03330‑x) summarizes the result as full retained bioactivity at 37°C for two months, paired with immune responses that were not inferior to fresh vaccine controls in the tested animals.

The team tested the optimized formulations in mice and non‑human primates. According to the paper's abstract, the solid‑state vaccines produced antigen‑specific immune responses that were non‑inferior to freshly prepared soluble vaccines delivered by intramuscular injection. The researchers also incorporated the formulations into microneedle patches, a delivery format that could eventually reduce dependence on trained injectors and cold‑chain equipment. Nature's [reporting summary](https://media.springernature.com/original/springer‑static/esm/art%3A10.1038%2Fs41587‑026‑03331‑w/MediaObjects/41587_2026_3331_MOESM2_ESM.pdf) records 15 cynomolgus macaques, including five females and ten males. It says exploratory screens used at least three samples, mice were randomly assigned, primates were stratified by age and sex, and in‑vivo work was replicated in at least three animals. It also says the in‑vivo experiments were not blinded because the intramuscular and microneedle routes were visibly different. The [supplement](https://media.springernature.com/original/springer‑static/esm/art%3A10.1038%2Fs41587‑026‑03331‑w/MediaObjects/41587_2026_3331_MOESM1_ESM.pdf) reports several mouse comparisons with five biologically independent animals per group. These are preclinical experiments with modest group sizes, not a human trial. Non‑inferiority in the reported animals is encouraging evidence that drying and warm storage did not erase the intended immune response. It is not evidence that a patch vaccine is ready for clinics or that the same result will hold in people.

The paper makes the generated datasets available through [Zenodo](https://zenodo.org/records/21726997) and says the Bayesian optimization was implemented in the AutODEx framework and separately links the public [AutoOED source repository](https://github.com/yunshengtian/AutoOED). The public materials do not establish that the repository is a complete snapshot of every AGENT implementation detail. That improves auditability: researchers can inspect the measurements and the general Bayesian‑optimization framework rather than treating AGENT as an unexplained model output. The system’s strength is data efficiency. Bayesian optimization is useful when each experiment is slow or expensive because it chooses measurements expected to provide the most information or improvement. The risk is that an efficient search can still optimize only the objectives and conditions encoded in the experiment. Long‑term chemical stability, manufacturing variability, transport shocks, sterility, dosing consistency and human immune response remain separate validation problems.

The next evidence threshold is reproducibility across independent laboratories and production‑scale batches, followed by longer stability windows and clinical testing. A usable thermostable vaccine also needs packaging, manufacturing and regulatory evidence that the complete product—not only a laboratory formulation—survives realistic distribution. For AI‑assisted laboratory design, the study is a concrete demonstration of a narrower promise: a model can reduce the number of wet‑lab cycles needed to find a strong formulation. The researchers report a six‑iteration, one‑month optimization and publish the resulting datasets. The public record supports that workflow and the preclinical stability result. It does not yet support claims of a cold‑chain‑free human vaccine. *Figure: Evidence milestones reported in the September 28 Nature Biotechnology paper. Source: [Nature Biotechnology](https://www.nature.com/articles/s41587‑026‑03331‑w). Figure by OpenTools Team; no third‑party expressive material used.*