An mRNA vaccine is a message molecule inside a fat bubble. Both parts are fragile. That is why most RNA vaccines travel at minus 20 to minus 80 degrees Celsius, a cold chain large parts of the world do not have. A team at MIT has now found a way to attack the problem not with better refrigeration but with a better bubble — and the way they found it was an algorithm.

The work, published in Nature Biotechnology as "Accelerated discovery of thermostable mRNA–lipid nanoparticle vaccines using data-efficient AI," comes from the labs of Ana Jaklenec and Robert Langer at MIT's Koch Institute, with Jinbi Tian and Khanh Tran as lead authors and collaborators at MIT's Computer Science and Artificial Intelligence Laboratory. It was supported in part by the Gates Foundation.

The problem: lipid nanoparticles (LNPs) protect the mRNA and carry it into cells. Which excipients, in which ratios, make those particles heat-tolerant is a search space with effectively endless combinations. The team had tried sugars, salts and polymers before, with partial success but not in the formulations used by approved vaccines. Running thousands of formulation experiments by hand was not an option.

The answer was a model built for small datasets. It began by analysing nearly 50 FDA-approved excipients. For each candidate, the researchers measured how well it stabilised an LNP carrying mRNA for firefly luciferase, a protein that emits light — more light after transfection meant the mRNA had survived better. From that screen the team picked five promising excipients and asked the algorithm to predict the ratios most likely to produce particles resembling the ones Moderna uses in its Covid-19 vaccine. They tested two candidates at a time, fed the results back in, and let the model generate the next predictions. After several rounds, one formulation went to animal testing. The guided process took weeks rather than the months earlier screening rounds had consumed.

Then the actual test. The particles packaged Covid-19 mRNA antigens, the preparation was vacuum-dried, and the result was stored either at 37 degrees Celsius — 98 Fahrenheit — for two months, or at room temperature for a year. Mice given the stored vaccine mounted immune responses equivalent to those from a reference formulation similar to Moderna's original.

The same heat-resistant formulation also worked in solid microneedle patches: small needles that dissolve in the skin, carrying a SARS-CoV-2 antigen, producing responses in mice similar to injectable RNA vaccines. The team additionally applied the algorithm to another LNP composition resembling the one Pfizer uses — same excipients, different proportions.

What this does not mean:

- There are no human data. Nothing here establishes human shelf life, safety, efficacy or readiness for commercial distribution. - The findings cover specific formulations, antigens and laboratory procedures, not every commercial RNA product. - Vacuum drying and large-scale manufacturing were not tested in this study. - The authors propose that once a heat-resistant combination is found for one LNP, it could be adapted to carry other mRNA payloads. That is a development prospect, not a demonstrated result.

Why it still matters: the bottleneck for mRNA was never only biology, it was logistics. A vaccine that holds for a year at room temperature changes what is possible in a clinic without a freezer, on a boat up a river, or in a campaign run off a generator. The Gates Foundation does not fund this kind of lipid chemistry out of curiosity; it funds it because of that arithmetic. And the study's real contribution is methodological — a model that learns from a handful of measurements can turn a formulation search from months into weeks, a pattern that reaches well past vaccines.

Until independent groups rebuild the formulation and larger studies exist, it remains what the authors say it is: early evidence, not a product.

Sources