Writing in Nature Biotechnology, Jinbi Tian, Khanh T. M. Tran, and colleagues describe AGENT, short for Algorithm-Guided Experimental design for lipid Nanoparticle Thermostabilization. The platform couples high-throughput experimentation with Bayesian optimization, a machine-learning strategy that is specifically built for situations where every experiment is costly and data are scarce. Rather than screening thousands of candidate formulations blindly, AGENT extracts the maximum amount of information from each small batch of results, using Gaussian process models to predict how untested excipient combinations will behave and to decide which experiments to run next.

The formulation problem the team tackled is genuinely multidimensional. To convert liquid mRNA–lipid nanoparticles into a solid-state, water-free product, the researchers used vacuum drying and then had to identify a cocktail of excipients—sugars, buffers, and other stabilizing agents—that would preserve the nanoparticles’ structure and the mRNA’s bioactivity through drying and subsequent storage. The search space of possible excipient types, ratios, and concentrations is enormous, and the relationships between composition and stability are nonlinear and poorly understood. Traditional one-factor-at-a-time approaches, the authors note, have been constrained by narrow formulation scope and low-throughput screening, which is precisely why prior stabilization efforts have progressed slowly.

Bayesian optimization changes the economics of that search. The algorithm maintains a probabilistic model of the entire design space and, at each iteration, balances exploration of uncertain regions against exploitation of promising ones, typically using an expected-improvement acquisition function to select the next experiment. Implemented within the AutODEx platform developed at MIT’s Computer Science and Artificial Intelligence Laboratory, AGENT completed its optimization in just six iterative rounds within a single month—a pace that would be unattainable with conventional design-of-experiments workflows, which often require many more cycles and far more material to converge on an optimum.

The results are striking. The team worked with two clinically relevant lipid nanoparticle systems, one based on SM-102, the ionizable lipid in Moderna’s vaccine, and the other on ALC-0315, used by Pfizer-BioNTech. After optimization, both were converted into solid-state formulations that retained 100 percent of their bioactivity after more than two months of storage at 37 degrees Celsius—roughly human body temperature. For a class of products that normally degrades within hours or days at room temperature, that level of thermal tolerance represents a qualitative leap, and it was achieved without altering the core lipid compositions already validated in clinics.

Stability on the shelf means little if the vaccine no longer works in the body, so the researchers subjected their thermostable formulations to a demanding battery of animal tests. In rodents, solid-state SARS-CoV-2 vaccine formulations delivered by standard injection induced antigen-specific immune responses that were non-inferior to those elicited by freshly prepared liquid vaccines given intramuscularly. The team then pushed the concept further, incorporating the stabilized nanoparticles into microneedle patches—tiny arrays of dissolving needles that painlessly deliver vaccine into the skin and could, in principle, be mailed to patients or administered by minimally trained personnel.

The microneedle experiments extended into non-human primates, a critical translational step. Patch delivery of the thermostable solid-state vaccine in cynomolgus monkeys again produced antigen-specific immune responses comparable to conventional intramuscular injection of freshly prepared vaccine. Because microneedle patches eliminate the need for needles, syringes, and highly trained injectors, and because the solid-state formulation removes the need for refrigeration, the combination addresses two of the most persistent bottlenecks in global immunization campaigns simultaneously.

The significance of the work extends beyond any single vaccine. The mRNA–LNP platform is now being pursued for cancer immunotherapies, gene editing, and treatments for rare genetic diseases, and all of these applications inherit the same storage fragility. Solid-state, water-free formulations open the door to integrating mRNA payloads into emerging delivery modalities, from microneedle patches to other advanced systems, that would be incompatible with a liquid product requiring ultra-cold handling. The authors also point out that the cold chain carries substantial economic and environmental costs, with estimates suggesting that vaccine distribution inefficiencies contribute significantly to global inequity in immunization coverage.

Methodologically, the study adds to a growing body of evidence that Bayesian optimization can compress discovery timelines in materials science and pharmaceutical development, where experiments are expensive and the parameter space is combinatorially large. Similar approaches have accelerated the identification of 3D printing materials and complex heterostructures, and the MIT team has made both its datasets, deposited in a public Zenodo repository, and its source code openly available, lowering the barrier for other laboratories to adopt the workflow. The optimization framework treats excipient selection as a multi-objective problem, allowing researchers to specify what matters—potency retention, storage stability, manufacturability—and let the algorithm navigate trade-offs automatically.

Caveats remain before thermostable mRNA vaccines reach clinics. The reported storage data cover months rather than years, and regulatory agencies will require extensive documentation of long-term stability, manufacturing consistency, and safety under Good Manufacturing Practice conditions. Human clinical trials of solid-state formulations and microneedle delivery lie ahead. Yet the demonstration that AI-guided experimentation can find stabilizing excipient compositions in six iterations and one month suggests that the development bottleneck may no longer be the search itself. If the approach generalizes across payloads and lipid systems, the freezer—and the costly, fragile supply chain it demands—could finally become optional for mRNA medicine.

Subject of Research: AI-guided optimization of thermostable solid-state mRNA–lipid nanoparticle vaccine formulations

Article Title: Accelerated discovery of thermostable mRNA–lipid nanoparticle vaccines using data-efficient AI

Article References: Tian, J., Tran, K. T. M., Pogostin, B. H., Sheridan, O., Mursalova, S., Lee, A. H., Liu, S., Hamkins, J., Antov, D., Power, A. L., Dash, Z. S., Yun, D., Konaković Luković, M., Langer, R. S., & Jaklenec, A. (2026). Accelerated discovery of thermostable mRNA–lipid nanoparticle vaccines using data-efficient AI. Nature Biotechnology. https://doi.org/10.1038/s41587-026-03331-w

Image Credits: AI Generated

DOI: 10.1038/s41587-026-03331-w

Keywords: mRNA vaccines, lipid nanoparticles, Bayesian optimization, thermostability, cold chain, microneedle patches, vacuum drying, excipients, SARS-CoV-2, drug delivery, machine learning, Nature Biotechnology