Machine Learning-Accelerated Discovery of Stable Solid-State Lipid Nanoparticle-Formulated RNA Vaccines

Introduction: : Lipid-nanoparticle (LNP)-formulated RNA therapeutics continue to attract attention across various clinical areas, particularly following the success of SARS-CoV-2 mRNA vaccines. RNA therapeutics and prophylactics offer advantages in safety, manufacturing scalability, and design flexibility, yet enhancing the stability of RNA-LNPs remains a critical challenge. Current FDA-approved mRNA vaccines (developed by Moderna and Pfizer/BioNTech) in liquid formulations require cold storage at 4C for short-term or -20C for long-term stability, posing significant challenges for worldwide distribution, especially in resource-limited and rural areas. Solid-state RNA-LNPs, obtained through lyophilization or freeze-drying, show improved stability and extended shelf-life compared to liquid forms. However, these methods require specialized equipment, extensive labor, and complex procedures. While excipients are commonly employed to improve the long-term stability of solid-state RNA-LNP formulations, the methods to identify optimal excipients in a high throughput manner remains inadequately explored. The complex physicochemical properties of LNPs, combined with the extensive range of potential excipients, create a multivariable challenge for optimization. Here, we present a ML-driven, closed-loop experimental framework to guide the selection of excipients for stabilizing RNA-LNP formulations in solid states using vacuum-drying at ambient temperature. We further demonstrate the broad utility of solid-state RNA-LNPs by integrating them into microneedle (MN) patches for SARS-CoV-2 mRNA vaccine delivery, a next-generation technology that enables patient self-administration, eliminates sharp waste, and broadens global access to vaccines.

Materials and

Methods: : LNPs were synthesized using an organic-aqueous precipitation method in an automated microfluidic mixer. The organic phase consisted of ionizable lipids (SM-102 or D-Lin-MC3-DMA or ALC-0315), helper lipids (DOPE or DSPC or DOTAP), PEGylated lipids (DMG-PEG 2000 or ALC-0159), and cholesterol dissolved in ethanol at specified molar ratios. The aqueous phase contained mRNA (encoding for firefly luciferase (Fluc) protein or the spike receptor-binding domain (RBD) of SARS-CoV-2 spike protein) in citrate buffer. LNPs were dialyzed against 1x phosphate-buffered saline (PBS) followed by water, concentrated, and stored at 4°C. Excipients (polyvinyl alcohol (PVA), polyvinyl pyrrolidone (PVP), sucrose, glucose, L-arginine hydrochloride) were dissolved in Tris or sodium acetate buffer. mRNA-LNPs were combined with excipients and vacuum-dried for three days. A549 cells were treated with reconstituted solid-state and soluble FLuc mRNA-LNPs. Luminescence was measured 24 hours post-transfection and used as the input for a Bayesian optimization algorithm. Size, encapsulation efficiency and morphology of reconstituted solid-state mRNA-LNPs were characterized using dynamic light scattering measurements, RiboGreen assays, and cryogenic electron microscopy, respectively. For in vivo validations, BALB/c mice received intramuscular injections of reconstituted solid-state or soluble FLuc mRNA-LNPs. Bioluminescence was assessed using an IVIS system 6 hours post-injection. To fabricate vaccine-loaded MN patches, mixture of algorithm-optimized excipient formulation and RBD mRNA-LNPs were dispensed onto PDMS molds and vacuum-dried. BALB/c mice were vaccinated with soluble or reconstituted solid-state RBD mRNA-LNPs, or RBD mRNA-LNP loaded MN patches, all at the same dose. Serum samples were collected weekly for ELISA to assess antigen-specific IgG antibody titers.

Results, Conclusions, and Discussions:: Results and Discussions:
We successfully developed a new method of formulating solid-state mRNA-LNPs via vacuum drying (Fig. 1a). To evaluate the effects of excipients on LNP stability during this process, we synthesized six mRNA-LNP formulations, including clinically relevant ones from Moderna (mRNA-1273) and Pfizer/BioNTech (BNT162b2) (Fig. 1b, c). In vitro screening of solid-state mRNA-LNPs formulated with L-arginine HCl, PVA, PVP, sucrose, and two buffers (Tris, sodium acetate) demonstrated that excipients are essential for maintaining solid-state mRNA-LNP stability, no single excipient is sufficient, and excipient performance is formulation-specific (Fig. 1d). To systematically optimize excipient combinations, we developed a ML-guided Bayesian optimization framework. Solid-state FLuc mRNA-LNPs were formulated with various excipient combinations, reconstituted, and tested for in vitro transfection efficiency, guiding successive optimization cycles (Fig. 2a). Using the ML-guided framework, optimized excipient formulations were identified for LNP1 (Moderna lipid composition) after six and three iterations. Compared to manually optimized or initial formulations, the ML-optimized solid-state LNPs showed 30-40% greater in vitro transfection efficiency and ~90% improvement over standard excipient combinations (Fig. 2b). In vivo IVIS imaging further confirmed that reconstituted optimized formulations achieved protein expression levels comparable to solution-state LNPs (Fig. 2c). Encapsulation efficiency and particle size were largely maintained, with Tris buffer reducing particle size increase compared to PBS (Fig. 2d, e). To assess the efficacy and stability of algorithm-optimized solid-state mRNA-LNPs, we formulated SARS-CoV-2 spike RBD mRNA-LNPs using LNP1 lipid compositions and evaluated their performance in both intramuscular injections and MN patches (Fig. 3a). Antibody titers in mice showed a strong dependence on excipient composition, with the optimized formulation achieving responses statistically equivalent to fresh soluble LNPs. MN patches fabricated with optimized excipients enabled robust skin penetration and generated immune responses comparable to intramuscular delivery (Fig. 3b). Furthermore, the optimized LNPs demonstrated excellent thermostability, retaining efficacy after one month storage at room temperature and 37°C (Fig. 3c).

Conclusions:
The ML-guided stabilization approach introduced here provides a powerful means to extend the shelf life of LNP-based RNA biologics and enable their integration into advanced controlled-release delivery systems, paving the way for broader clinical impact in vaccines, oncology, and other RNA-based treatments.

Acknowledgements and/or References (Optional):: The authors would like to thank the support from the Bill & Melinda Gates Foundation, the Sieble Scholars Foundation, and the MIT-Takeda Fellowship.