A machine‐learned microvasculature optimizes physiological insulin secretion in a vascularized pancreas‐chip

Abstract

Microphysiological systems have gained increasing attention for their applications in drug discovery and testing, toxicity screening, and more recently disease modeling. However, microvascular networks are rarely incorporated into these devices despite being a major component of almost all tissues of the body. Of the reports integrating these networks, many use different quantification tools and fail to relate these quantifications to the biological function of the networks. Here, we standardized vasculature engineering by first generating 500 samples of microvascular networks‐on‐chips composed of different endothelial cell densities and sources, supporting cells, growth factors, and extracellular matrices. For each sample, 10 individual morphological metrics were generated using a peer‐reviewed, open‐source network quantification tool. Simultaneously, the mass transport of each sample was quantified using numerical simulation, resulting in a continuous value representing the biological function of each sample. Using machine learning, we then developed models using neural network regression methods that related the morphological metrics with the simulated biological function. Further, we identified which metrics most significantly impact the mass transport, as well as the integrable chip components that optimized mass transport and its related morphological metrics. Finally, we demonstrated how these optimizations lead to the vascularization of a pancreas microenvironment featuring insulin‐secreting islet cells. Further, vascularization rescues insulin secretion when pancreas‐chips are placed in hypoxia, demonstrating that the integration of microvascular networks plays a crucial role in the oxygenation and maintenance of a tissue‐specific microphysiological system.