(K-434) Machine Learning-assisted Syringe Design for Programmed mRNA-LNP Vaccine Release

Many drugs and vaccines require multiple doses to provide sufficient treatment and protection. Similarly, lipid nanoparticle (LNP)-based mRNA vaccines necessitate multiple dose administrations at predetermined time intervals. Most mRNA vaccines for SARS-CoV-2 involve two to three doses [1], and recent clinical trials for mRNA cancer vaccines for rabies (CureVac) [2], prostate cancer (CureVac) [3], and melanoma (Moderna) [4] required 3, 5, and 9 dose administrations, respectively. However, requiring multiple doses poses challenges of missed or mistimed doses. In fact, challenges in administering such a dosing regimen lead to a large untreated population, especially in the developing world where healthcare resources and patient access are both limited. Considering nearly half of the unimmunized children have received at least one dose of vaccine, but not the complete vaccination schedule needed for full protection [5], a delivery system that can mimic the current multiple bolus administrations in a single injection, could significantly improve the globe immunization coverage. Core-shell StampEd Assembly of polymer Layers (SEAL) microparticles exhibit discrete, pulsatile release of therapeutics at different times by changing the material properties of the shell [6]. These particles can be injected as a mixed population with different release kinetics to achieve the desired pulsatile cargo released over long periods of time to mimic current administration regimens. To provide safe and complete injections of SEAL microparticles, and ultimately, sufficient treatment or full vaccination, a syringe design must be optimized for maximum injectability of intact particles and minimum dead volume [7].

To build a syringe framework that offers the optimal SEAL microparticle injections, the fluid velocity profile and particle trajectory with different design parameters were first simulated using COMSOL Multiphysics computational fluid dynamics (CFD) Module (Stockholm, Sweden) (Figure 1). Assuming injections of 1,000 particles with a particle density of 1,340 kg/m³ in a liquid solution (density of 1,000 kg/m³) with a 2.88 mm/sec inlet velocity at 1 atm, syringe injections were simulated for different syringe design parameters (D, d, l, ll, l2, dn) (Figure 1), particle diameters (5, 10, 25, 50, 100, and 150 µm), and solution viscosities (0.37, 0.01, 0.001 Pa∙S). A total of 576 injection conditions, which were the design of experiments selected via Taguchi method [8], were simulated for the best injectability, which was calculated by dividing the ‘number of particles reached the needle end’ by the ‘total number of particles injected’. Secondly, the simulation input and output data were used to create a continuous deep learning-assisted model for predicting the chance and range of injectability using MATLAB Neural Net Fitting (MathWorks, Natick, MA). Levenberg–Marquardt Algorithm and 10 hidden layers were used with data partition of 70% for training, 15% for validation, and 15% for testing to generate a prediction model with the lowest mean square error (MSE) and R value closest to 1. Lastly, two syringe designs predicted for the highest injectabilities were illustrated via computer aided design (CAD) modeling (SolidWorks, Dassault Systemes, France) and 3D printed with FormLabs stereolithography (SLA) printers (Somerville, MA).

Smaller particles (5-25 µm) exhibited higher injectability at a lower solution viscosity (0.001 Pa∙S), while bigger particles (50-150 µm) exhibited higher injectability at a higher solution viscosity (0.37 Pa∙S) (Figure 2a). This is because the fluid injection drag force affects small particles more than the gravity force, while gravity affects big particles more than the drag force in solution with a lower viscosity. For all solution viscosities, larger D and smaller d, l, ll, and l2 values resulted in higher particle injectabilities (Figure 2b). For all solution viscosities, design parameter d affected the injectability the most and l2 affected the injectability the least (Figure 2c). As for the different needle diameters (dn), there was no significant difference in terms of injectability for all solution viscosities. After feeding these simulation data into Neural Net Fitting, prediction models with testing Rs equal to 0.99826, 0.99882, and 0.999 were generated for solution viscosities 0.37, 0.01, and 0.001 Pa∙S, respectively (Figure 3a). These machine learning-generated models confirmed with the CFD results and predicted the two best syringe framework designs for the highest injectabilities: 1) 10mm D, 2.5mm d, 5mm l, 5mm ll, 5mm l2, and 500µm dn for 75.2% injectability, and 2) 15mm D, 3.75mm d, 10mm l, 15mm ll, 20mm l2, and 500µm dn for 71.6% injectability (Figure 3b). Syringe frameworks with these dimensions were CAD modeled and prototyped using 3D printing (Figure 3c). As the next step, the prototypes will be empirically tested for microparticle injectability in vitro. The syringe framework designs optimized for the best SEAL microparticle injectability using machine learning can potentially provide safe and reliable programmed vaccine releases that mimic current multiple-dose administration regimens, reduce the waste of costly drugs, and improve global healthcare equity.