Type 1 diabetes (T1D) is a chronic autoimmune disease characterized by the destruction of insulin-producing beta cells in the pancreas. Current treatments, such as insulin injections or pumps, require constant monitoring and management, making them burdensome for patients. Pancreatic islet transplantation and transplantation of stem cell-derived insulin-producing cells (SC-βs) have shown promise in restoring glycemic control, but the limited availability of donor islets and the need for lifelong immunosuppression restrict its widespread applicability. Bioartificial pancreas (BAP) devices offer a potential solution by encapsulating islets or SC-βs within a semipermeable membrane, protecting them from immune rejection while allowing for nutrient and oxygen exchange. However, this encapsulation introduces mass transport challenges, primarily oxygen limitation, significantly affecting cell survival and insulin secretion. Beta cells, in their natural environment, are closely associated with blood vessels for efficient oxygen supply. In BAP devices, oxygen delivery relies solely on passive diffusion, which is limited by the low oxygen tension in the transplantation site, the low permeability of the encapsulation material (usually a hydrogel), and the high oxygen consumption rate (OCR) of beta cells. Hypoxia can lead to cell death and the release of danger-associated molecular patterns (DAMPs), triggering an immune response and graft failure. Consequently, accurate prediction of oxygen transport and its effect on cell viability and insulin secretion is crucial for optimizing BAP design. Existing models often simplify the system by assuming uniform islet size and distribution, neglecting the inherent stochasticity of actual BAP devices. This study addresses these limitations by developing a sophisticated computational platform that accounts for the stochastic properties of cell size, distribution, and their impact on device performance.

Previous computational models of mass transfer in BAP devices have typically simplified the system by treating islets as uniformly sized spheres or circles, distributed uniformly within the device. These simplifications, while computationally efficient, fail to capture the inherent heterogeneity and stochasticity of islet size and spatial distribution observed in real-world BAPs. Mammalian islets exhibit a wide range of sizes, from <50 µm to >500 µm in diameter, with distributions varying across species and isolation protocols. Moreover, islets are not uniformly distributed within the encapsulation matrix. The neglect of this inherent probabilistic nature of islet size and position in previous models limits their predictive power and may lead to biased estimations of device performance. Existing models, often based on deterministic approaches, have provided limited insights into the impact of this variability on overall BAP functionality. This lack of accurate representation has hampered efforts to optimize device design and predict therapeutic outcomes.

The researchers developed a computational platform called SHARP (Simulated Heterogeneity and Randomness Program) based on the stochastic finite element method. SHARP incorporates the probabilistic nature of islet size and distribution into the mass transport model. The platform begins by characterizing the size distributions of islets and SC-βs from various sources (mouse, rat, juvenile pig, human, and two sources of SC-βs). Islet sizes were measured from images and fit to lognormal or Weibull distributions. SHARP then simulates the device in silico, generating a virtual representation of the device by randomly distributing islets or SC-βs based on their empirically determined size distributions within the specified device geometry. The program accommodates various geometries, including planar slabs, cylinders, and hollow cylinders. Oxygen transport is modeled using the oxygen conservation equation, accounting for diffusion and consumption by the cells (using Michaelis-Menten kinetics). Oxygen-dependent insulin secretion is modeled using a Hill relationship. The program calculates several key outputs including volume-average pO2, necrotic fraction, and loss of insulin secretion capacity for the whole islet population and for each individual islet. This Monte Carlo simulation is repeated numerous times to obtain statistically meaningful results and quantify the impact of stochasticity. To validate the model's predictions, the researchers created BAP devices with rat islets partitioned into large and small size fractions, evaluating device performance in vitro under hypoxic conditions and in vivo by implanting the devices in diabetic mice. Finally, a machine learning model, SHARP-ML, was developed and trained on data generated by SHARP to rapidly predict device performance without the computational burden of the stochastic finite element method. SHARP-ML provides a user-friendly web application to generate device-specific islet equivalence conversion tables.

SHARP accurately captured the effect of islet size distribution and spatial distribution on BAP performance. The study found that the variance in human islet size distributions significantly affects device function, with larger islets experiencing greater hypoxia, necrosis, and loss of insulin secretion capacity. This was validated experimentally using rat islets partitioned into large and small size fractions. BAP devices containing smaller islets demonstrated significantly improved function both in vitro and in vivo. In the in vivo study, devices with smaller islets effectively lowered blood glucose levels in diabetic mice, while devices with larger islets failed to maintain glycemic control. Optimization studies using SHARP showed that planar slab devices were more effective than cylindrical or hollow cylindrical devices for delivering a therapeutic dose of islets, though the required islet volume was still quite high. Using SHARP-ML, the researchers also demonstrated that the traditional islet equivalence (IEQ) calculation overestimates the functional capacity of larger islets, particularly in high-density devices, necessitating a functional adjustment to the IEQ conversion table. The use of smaller islets or SC-βs (due to their more uniform size distributions) along with strategies to enhance oxygen transport were proposed as means to improve BAP device design and function.

The findings highlight the importance of considering islet size distribution and spatial arrangement in modeling BAP performance. The significant impact of size distribution on device efficacy underscores the need for more sophisticated islet dosing strategies that account for functional capacity, rather than simply relying on islet volume. The development of SHARP and SHARP-ML provides researchers with valuable tools for optimizing BAP design and predicting clinical outcomes. The results suggest a focus on minimizing device thickness and optimizing cell density, as well as employing strategies to enhance oxygen delivery to the encapsulated cells. The successful validation of SHARP's predictions in an in vivo model strengthens the model's credibility and its potential for guiding the design and development of next-generation BAP devices. The successful implementation of a machine learning surrogate model demonstrates the possibility of scaling up the analysis to broader applications in the field.

This study demonstrates the development and validation of a novel computational platform, SHARP, that accurately predicts the performance of bioartificial pancreas devices by incorporating the stochastic properties of islet size and spatial distribution. SHARP’s predictions were validated in vitro and in vivo experiments and shows that islet size distribution significantly impacts device function, with smaller islets performing better. The development of a surrogate machine learning model, SHARP-ML, significantly simplifies the prediction process. This work provides valuable tools for optimizing BAP design and accelerating the clinical translation of cell encapsulation therapies for type 1 diabetes. Future research could focus on incorporating additional stochastic factors, such as variations in islet OCR and insulin secretion, into the model. The integration of SHARP with other models of glucose metabolism and insulin dynamics could create a comprehensive closed-loop simulation of BAP function in vivo.

The model assumes uniform oxygen tension at the device-host boundary and a constant oxygen permeability for the encapsulation matrix, which may not accurately reflect the in vivo environment. The study used a limited set of device geometries and encapsulation materials. Although the machine learning model simplifies the analysis, its accuracy depends on the quality and quantity of training data. Further validation is needed to confirm the generalizability of the findings across different cell types, device designs, and transplantation sites. The study also did not extensively quantify the fibrotic coverage on the device surface post-transplantation, despite some visual evidence suggesting this could be a contributing factor to device failure.