Development of statistical protocol to validate machine learning predictions of focal adhesion sites

Abstract

Focal Adhesion sites are important to adhesion and various other functions of human endothelial cells. A deeper understanding of these sites within the endothelium can assist in understanding different biological states. In recent studies, machine learning algorithms have been utilized to learn and predict Focal Adhesion sites within a given membrane for the development of a digital twin. This research aimed to validate these predictions when the true orientation is unknown. A protocol was developed using the programing language R that loads membrane and Focal Adhesion site coordinates of both two groups of cells and determines if the two groups are statistically equivalent using the Studentized Permutation Test. A population standard was created by comparing actual cells against a larger population of actual cells. The response to various error states was explored. The statistical protocol was utilized to compare the results of machine learning cells compared to actual cells to the population standard. This protocol was then expanded to utilize marks, or other information about the Focal Adhesion site, and the process repeated. In both scenarios, the machine learning predictions were determined to statistically fit into the population. These predictions can be used in the validation of cellular digital twins in future studies.