(M1530-09-52) Mass Transport Modeling and In Silico Formulation of Osmotic Controlled Release Tablets with Microscopy-Based Digital Twins

Purpose: Osmotic delivery of active pharmaceutical ingredients (API’s) in extrudable core system (ECS) tablets continues to play a key role in the development of controlled release systems and are widely accepted as the most reliable source for controlled release drug delivery. Pharmacokinetic control is achieved through manipulation of the semi-permeable membrane surrounding the tablet core. A wholistic understanding of the porous membrane microstructure is essential in understanding how drug releases between different formulations. In this work, cross sectional SEM imaging is combined with artificial intelligence (AI) based structural synthesis to generate 3D membrane volumes in silico. Comparisons of the 3D membrane structures will elucidate the impact of membrane structure on the performance of osmotic pump formulations. The influence of critical quality attributes on product performance is evaluated from simulated membrane tortuosity, capillary pressure, and permeability using the imaging data. Modelling enables a rapid assessment of membrane performance variability between formulations over the course of a few days, saving significant time, money, and resources in the development of osmotic pump tablet systems.

Methods: Four single-layered osmotic controlled release MR tablet shells were investigated using high resolution 2D focused ion beam scanning electron microscopy (FIB-SEM). Four shells were selected for analysis post dissolution testing for an evaluation of the shell’s microporous structure absent the water-soluble additives. A focused ion beam was used to mill a trench roughly 60 µm wide through the coating. SEM was used to collect high resolution images at various magnifications, revealing internal structural details in the cross sections of each tablet shell. AI-based analyses of the imaging data were performed to reconstruct porosity and solid phases of the 2D images. To evaluate membrane transport properties between formulations, iterative AI models based on generative adversarial networks (GAN) were run to produce 3D membrane volumes from the 2D cross sectional images. Key performance attributes calculated for the finalized 3D in silico volumes included total porosity, connected porosity, pore size distribution, pore spatial uniformity, tortuosity, capillary pressure, and permeability. GAN was further employed for microstructure engineering of the 3D membrane in sample D4556. The volume fraction of total porosity was modulated in the digital space to meet specific performance targets. Four additional volumes were simulated with incremental increases and decreases in total membrane porosity.

Results: The cross-sectional images collected on the membrane shells are shown in Figure 1a. Pores with distinct grayscale contrast can be seen, allowing for a quantification of the pore networks. A distinct layering pattern of the internal porosity was observed in each of the partially dissolved tablet shells. AI-based segmentation of the images was utilized for a direct comparison of pore networks. Samples D3831 and D3840 showed greater porosity than D3837 and D4556 as can be seen in Table 1.  Iterative AI generative models simulated 3D porosity networks from the 2D image inputs, and the simulated membrane volumes are shown in Figure 1b. Among these simulated models, the membrane of tablet D3837 was the most tortuous and had the greatest capillary pressure at each wetting saturation as seen in Figure 1c-d. GAN was successful in simulating four additional models of D4556 with variations in total porosity while maintaining the output target condition of 1 µm median pore size. Porosity values in the five models in Figure 2 with incremental increases and decreases in total porosity from the original images were 18.5%, 20.5%, 22.4%, 24.0%, and 25.4%. Membrane transport properties were predicted from the digital models and a porosity permeability correlation was developed. Mechanical permeability (σ) was computed in each of the five digital formulations of D4556 and plotted against their respective porosity (p).  The data was fit to the power model expressed in Eq. (1) and has an R2 of 0.9991, confirming high agreement between the computed data and the correlative model developed. σ=(6E-12)p^5.64  

Conclusion: Using four single-layered osmotic controlled release, FIB-SEM imaging and AI-based iterative generative models have been shown as a promising approach in investigating the microstructures of semi permeable membranes to predict transport properties. Sample D3840 was identified to be the most tortuous and the most porous, indicating this sample will likely show improved drug release compared to the other formulations studied. The porosity-permeability correlative model developed herein provides a mechanistic understanding of how variations in the amount of pore former may affect permeation of dissolved drug through the membrane and dictate the release rate of the API.  This technique can be broadly applied in early stages of product development to predict product performance between various formulations and model the controlled release of API through the tablet shell.

Table 1. Overview of samples studied, and characterization methods performed.

Figure 1. a) FIB-SEM image data collected on four different single-layered osmotic controlled release MR tablet shell coatings b) 3-dimensional in silico models developed with GAN representing tablet shells from D3831, D3837, and D3840. Pores (black) and semi permeable membrane (white) are shown across X, Y and Z axes. c) Distribution of tortuosity values computed for each of the 3D models simulated in (a). d) Capillary pressure vs wetting saturation for each of the 3D models simulated in (a).

Figure 2. a) Pore size distributions for five in silico models engineered to represent D4556 formulation with incremental increases and decreases in total porosity volume fractions. b) Capillary pressure vs wetting saturation for each of the five digitally engineered in silico models of D4556. c) Porosity vs permeability correlative model, where each data point represents porosity and mechanical permeability computed in each of the five digitally engineered in silico models of D4556. The model is well fit by σ=(6e-12) p^5.64 with an R2 of 0.9991.