Astrocyte Regional Heterogeneity Revealed Through Machine Learning‐based Glial Neuroanatomical Assays

Abstract

Astrocytes undergo activation during brain injury through an ill‐defined process known as gliosis. Recent work focused on elucidating the molecular underpinnings of glial response to injury. However, astrocytes exist in a unique three‐dimensional space within the brain, and this location is highly relevant to the astrocyte’s function. Therefore, generating neuroanatomical assays, which maintain CNS positional information of astrocytes relative to their neural networks, represent optimal modalities to assess astrocyte function and reactivity. The majority of astrocytic neuroanatomical assays in the scientific literature measure mean fluorescence intensity of GFAP or ALDH1L1. Modern advances in computer vision and machine learning algorithms have now given us the tools with which we can extract a multitude of features from microscopic imaging data. By segmenting astrocytes from an image into numerous GFAP or ALDH1L1 positive “objects”, extracted numerical values indicate object location, object intensities, object shapes, as well as object textures from the original image. We tested the hypothesis that morphological changes in astrocyte reaction to injury represents a heterogeneous response in different brain regions. To test this hypothesis, we developed an automated segmentation algorithm for both GFAP and ALDH1L1 immunostained mouse brain sections treated with vehicle or systemic lipopolysaccharide to induce neuroinflammation. These images were then captured on a laser scanning confocal microscope and features characterizing shape, fluorescent intensity, and fluorescent texture. We then utilized a covariance matrix, random forest‐based boruta algorithm, cluster analysis and principal component analysis to evaluate the differences between reactive astrocytes and baseline astrocytes. By implementing a support vector machine algorithm, we show classification identification methods of reactive versus baseline astrocytes. Machine learning models developed to classify neuroanatomical images provides a highly insightful methodology to discover new processes by which astrocytes react to injury.Support or Funding Information3R01HL132355