MORPHIOUS: A Machine Learning Workflow to Naively Detect the Activation of Microglia and Astrocytes

Abstract

In cases of brain injury, degeneration and repair, defining microglia and astrocytic activation using cellular markers alone remains a challenging task. We developed MORPHIOUS, an unsupervised machine learning workflow that utilizes a one-class support vector machine to segment clusters of activated glia by only referencing examples of non-activated glia. Here, glial activation was triggered using focused ultrasound to permeabilize the hippocampal blood-brain barrier. Analyzing the hippocampal sections seven days later, MORPHIOUS identified two classes of microglia which showed characteristic activation features, including increases in ionized calcium-binding adapter molecule 1 expression, soma size, and de-ramification. MORPHIOUS was further used to identify clusters of activated astrocytes, which showed increased expression of glial fibrillary acidic protein and branching. Thus, by only referencing untreated glia morphologies, MORPHIOUS can identify diverse and novel manifestations of glial activation. This provides significant improvements for characterizing glial activation in cases of injury, neurodegeneration, and regeneration.

Transgene distribution and immune response after ultrasound delivery of rAAV9 and PHP.B to the brain in a mouse model of amyloidosis

“…To further quantify the activation of microglia, we used a recently developed machine learning method (morphological identification of outlier clusters, MORPHIOUS; manuscript deposited in bioRxiv). 34 The contralateral hemisphere, not treated with MRIgFUS, from mice injected with rAAV9 was used to teach MORPHIOUS the definition of microglial morphologies found in conditions without MRIgFUS exposure and AAV transduction ( Figure 4 P). Hereafter, areas of microglial cells that are responding to MRIgFUS and/or rAAVs were determined and defined as highly active (focal, red) or active (proximal, yellow) clusters.…”

“…We identified activated microglia using MORPHIOUS, a recently developed machine-learning method (manuscript deposited in bioRxiv). 34 In brief, MORPHIOUS learns the definition of a microglial morphology from control tissues not activated by MRIgFUS or rAAV transduction and from this definition, infers regions of interest corresponding to non-normal or activates microglia in treated tissues. Microglial clusters identified as activated are divided into focal and proximal clusters depending on the degree of activation, with focal clusters harboring the highest degree of activation.…”

“…In sections immunostained with IBA1, a marker of microglia and macrophages, morphological features related to IBA1 fluorescence, soma characteristics, and skeleton complexity were quantified using custom ImageJ scripts, as described in detail in a manuscript deposited in bioRxiv. 34 In brief, images were first post-processed using the subtract background command (radius = 50) and the local contrast-enhanced command (blocksize = 127, histogram = 256, maximum = 2.5) and despeckled. A local threshold was applied to the images (auto-local threshold: method = Phansalkar, parameter 1 = 0, parameter 2 = 0) to delineate IBA1 foreground pixels, and the IBA1 intensity was measured.…”

“…Microglia cell bodies were delineated and counted using a custom ImageJ script, which we are describing in detail in a manuscript deposited in bioRxiv, 34 based on the ImageJ MorphoLibJ package. 66 In a slight modification to the original method, images were first local contrast enhanced (slope = 2.0).…”

“…The NND was calculated using a custom Python script (see detailed description in manuscript deposited in bioRxiv). 34 …”