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

Since every biological system requires capillaries to support its oxygenation, design of engineered preclinical models of such systems, for example, vascularized microphysiological systems (vMPS) have gained attention enhancing the physiological relevance of human biology and therapies. But the physiology and function of formed vessels in the vMPS is currently assessed by non-standardized, user-dependent, and simple morphological metrics that poorly relate to the fundamental function of oxygenation of organs. Here, a chained neural network is engineered and trained using morphological metrics derived from a diverse set of vMPS representing random combinations of factors that influence the vascular network architecture of a tissue. This machine-learned algorithm outputs a singular measure, termed as vascular network quality index (VNQI). Cross-correlation of morphological metrics and VNQI against measured oxygen levels within vMPS revealed that VNQI correlated the most with oxygen measurements. VNQI is sensitive to the determinants of vascular networks and it consistently correlates better to the measured oxygen than morphological metrics alone. Finally, the VNQI is positively associated with the functional outcomes of cell transplantation therapies, shown in the vascularized islet-chip challenged with hypoxia. Therefore, adoption of this tool will amplify the predictions and enable standardization of organ-chips, transplant models, and other cell biosystems.

Evaluation of the Morphological and Biological Functions of Vascularized Microphysiological Systems with Supervised Machine Learning

Vascularized microphysiological systems and organoids are contemporary preclinical experimental platforms representing human tissue or organ function in health and disease. While vascularization is emerging as a necessary physiological organ-level feature required in most such systems, there is no standard tool or morphological metric to measure the performance or biological function of vascularized networks within these models. Further, the commonly reported morphological metrics may not correlate to the network's biological function-oxygen transport. Here, a large library of vascular network images was analyzed by the measure of each sample's morphology and oxygen transport potential. The oxygen transport quantification is computationally expensive and user-dependent, so machine learning techniques were examined to generate regression models relating morphology to function. Principal component and factor analyses were applied to reduce dimensionality of the multivariate dataset, followed by multiple linear regression and tree-based regression analyses. These examinations reveal that while several morphological data relate poorly to the biological function, some machine learning models possess a relatively improved, but still moderate predictive potential. Overall, random forest regression model correlates to the biological function of vascular networks with relatively higher accuracy than other regression models.

Evaluation of the morphological and biological functions of vascularized microphysiological systems with supervised machine learning

Vascularized microphysiological systems and organoids are contemporary preclinical experimental platforms representing human tissue or organ function in health and disease. While vascularization is emerging as a necessary physiological organ-level feature required in most such systems, there is no standard tool or morphological metric to measure the performance or biological function of vascularized networks within these models. Further, the commonly reported morphological metrics may not correlate to the network's biological function - oxygen transport. Here, a large library of vascular network images was analyzed by the measure of each sample's morphology and oxygen transport potential. The oxygen transport quantification is computationally expensive and user-dependent, so machine learning techniques were examined to generate regression models relating morphology to function. Principal component and factor analyses were applied to reduce dimensionality of the multivariate dataset, followed by multiple linear regression and tree-based regression analyses. These examinations reveal that while several morphological data relate poorly to the biological function, some machine learning models possess a relatively improved, but still moderate predictive potential. Overall, random forest regression model correlates to the biological function of vascular networks with relatively higher accuracy than other regression models.