Cyclic multiplex fluorescent immunohistochemistry and machine learning reveal distinct states of astrocytes and microglia in normal aging and Alzheimer’s disease

Abstract

Background: Astrocytes and microglia react to Aβ plaques, neurofibrillary tangles, and neurodegeneration in the Alzheimer’s disease (AD) brain. Single-nuclei and single-cell RNA-seq have revealed multiple states or subpopulations of these glial cells but lack spatial information. We have developed a methodology of cyclic multiplex fluorescent immunohistochemistry on human postmortem brains and image analysis that enables a comprehensive morphological quantitative characterization of astrocytes and microglia in the context of their spatial relationships with plaques and tangles. Methods: Single FFPE sections from the temporal association cortex of control and AD subjects were subjected to 8 cycles of multiplex fluorescent immunohistochemistry, including 7 astroglial, 6 microglial, 1 neuronal, Aβ, and phospho-tau markers. Our analysis pipeline consisted of: (1) image alignment across cycles; (2) background subtraction; (3) manual annotation of 5172 ALDH1L1+ astrocytic and 6226 IBA1+ microglial profiles; (4) local thresholding and segmentation of profiles; (5) machine learning on marker intensity data; and (6) deep learning on image features. Results: Spectral clustering identified three phenotypes of astrocytes and microglia, which we termed “homeostatic,” “intermediate,” and “reactive.” Reactive and, to a lesser extent, intermediate astrocytes and microglia were closely associated with AD pathology (≤ 50 µm). Compared to homeostatic, reactive astrocytes contained substantially higher GFAP and YKL-40, modestly elevated vimentin and TSPO as well as EAAT1, and reduced GS. Intermediate astrocytes had markedly increased EAAT2, moderately increased GS, and intermediate GFAP and YKL-40 levels. Relative to homeostatic, reactive microglia showed increased expression of all markers (CD68, ferritin, MHC2, TMEM119, TSPO), whereas intermediate microglia exhibited increased ferritin and TMEM119 as well as intermediate CD68 levels. Machine learning models applied on either high-plex signal intensity data (gradient boosting machines) or directly on image features (convolutional neural networks) accurately discriminated control vs. AD diagnoses at the single-cell level. Conclusions: Cyclic multiplex fluorescent immunohistochemistry combined with machine learning models holds promise to advance our understanding of the complexity and heterogeneity of glial responses as well as inform transcriptomics studies. Three distinct phenotypes emerged with our combination of markers, thus expanding the classic binary “homeostatic vs. reactive” classification to a third state, which could represent “transitional” or “resilient” glia.