Machine learning chained neural network analysis of oxygen transport amplifies the physiological relevance of vascularized microfluidic systems

Introduction:: Microphysiological systems (MPSs) are recently gaining attention for their ability to predict the efficacy, toxicity, and viability of therapeutic products before in vivo studies. Relatedly, MPS vascularization and its characterization are important requirements due to blood vessels’ incidence in almost all tissues and organs of the body. Some examples of vascularized-MPSs (vMPS) exist, but most focus on the non-standardized optimization of various morphological metrics. Here, we show that these metrics have limitations, and vascular networks optimized by their basic function, the oxygenation of tissues and organs, have relatively better predictive power. To achieve this, we developed a biologically inspired vMPS optimization workflow using a chained, machine learned neural network regression aimed at increasing theoretical oxygenation, scored by a new metric – the Vascular Network Quality Index (VNQI).


Materials and Methods:: We collected 500 samples of vMPS images, analyzed each sample’s morphology with a peer-reviewed morphological analysis tool, and simultaneously quantified each vMPS’s oxygen transport using a finite element simulator we developed in-house termed AngioMT. We performed traditional data cleaning procedures before obtaining a final dataset consisting of six morphological metrics and two oxygen transport metrics. We then established five different neural network architectures using TensorFlow 2.3.0, evaluating the accuracy in predicting oxygenation data with the models’ R² and mean absolute error (MAE) scores. Finally, we generated vascularized islet-chips by injecting islet spheroids in the hydrogels during vMPS manufacturing, cultivated them for 96 hours, challenged them with hypoxia for 4 hrs, added 22 mM glucose containing media, and collected the media for insulin quantification via ELISA. Statistical analyses of all experiments following machine learning was performed by one-way and/or two-way ANOVA in GraphPad Prism.


Results, Conclusions, and Discussions:: We tested five neural network architectures, finding that a novel chained neural network regression model predicted true oxygenation with an R² value of 0.98 and a MAE of 0.02. After a SHapley Additive exPlanations (SHAP) analysis, we saw that Branchpoint Count, Vessel Length, and the intermediary neural network output significantly contributed to the oxygenation of vMPS samples. A Pearson R correlation test also showed that VNQI, the output of the neural network architecture, correlated most with the true oxygenation measurements. We also found that the independent or interdependent combination of hydrogel stiffnesses, endothelial cell sources and densities, pro-angiogenic cell densities, and/or growth factor supplements could tune a vMPS’s VNQI. We then validated this deep learning methodology by vascularizing a pancreatic islet-on-a-chip, challenged it with hypoxia, and found the chip functionality increased almost 2-fold when vascularized with high-VNQI vMPS ingredients. Therefore, this machine learning methodology has translatable potential for other vascularized organ-chips, or tissue engineering procedures such as islet replacement therapy.




Acknowledgements (Optional): :

References (Optional): : Tronolone, J. J., Mathur, T., Chaftari, C. P., & Jain, A. Evaluation of the Morphological and Biological Functions of Vascularized Microphysiological Systems with Supervised Machine Learning. Annals of Biomedical Engineering, 2023. https://doi.org/10.1007/s10439-023-03177-2