Transcriptional profiling of microglia; current state of the art and future perspectives

Abstract Microglia are the tissue macrophages of the central nervous system (CNS) and the first to respond to CNS dysfunction and disease. Gene expression profiling of microglia during development, under homeostatic conditions, and in the diseased CNS provided insight in microglia functions and changes thereof. Single‐cell sequencing studies further contributed to our understanding of microglia heterogeneity in relation to age, sex, and CNS disease. Recently, single nucleus gene expression profiling was performed on (frozen) CNS tissue. Transcriptomic profiling of CNS tissues by (single) nucleus RNA‐sequencing has the advantage that it can be applied to archived and well‐stratified frozen specimens. Here, we give an overview of the significant advances recently made in microglia transcriptional profiling. In addition, we present matched cellular and nuclear microglia RNA‐seq datasets we generated from mouse and human CNS tissue to compare cellular versus nuclear transcriptomes from fresh and frozen samples. We demonstrate that microglia can be similarly profiled with cell and nucleus profiling, and importantly also with nuclei isolated from frozen tissue. Nuclear microglia transcriptomes are a reliable proxy for cellular transcriptomes. Importantly, lipopolysaccharide‐induced changes in gene expression were conserved in the nuclear transcriptome. In addition, heterogeneity in microglia observed in fresh samples was similarly detected in frozen nuclei of the same donor. Together, these results show that microglia nuclear RNAs obtained from frozen CNS tissue are a reliable proxy for microglia gene expression and cellular heterogeneity and may prove an effective strategy to study of the role of microglia in neuropathology.

microglia can adopt a range of phenotypes, depending on the local neural microenvironment and stimulation type (De Biase et al., 2017;Gosselin et al., 2017). Over the last decade, gene expression profiling of purified microglia has greatly contributed to our understanding and characterization of these cells, both under normal and disease condi- tions. An overview of these studies will be provided in this review, which will conclude with a detailed description of recent technological developments on nuclear sequencing of microglia.

| Mouse and human microglia transcriptomes by bulk population sequencing
The gene expression profile of mouse and human microglia was first identified using bulk population samples, for example, a large number of purified microglia in one sample. For mouse microglia, the first core microglia signature was generated in 2012 using microarrays in the Immunological Genome (ImmGen) project (Gautier et al., 2012). Based on this study, Chiu et al. compared microglia microarray data with data from 22 other myeloid cell types collected by the ImmGen project. Ninety-nine genes were identified that were fivefold or more enriched in microglia relative to other myeloid immune cells (Chiu et al., 2013). Furthermore, they also compared spinal cord microglia RNA-seq data with RNA-seq data obtained from astroglia, motor neurons, and whole spinal cord, yielding 288 genes enriched in microglia.
Using microglia marker Tmem119, a mouse microglia gene expression profile during development and after an LPS challenge was generated (Bennett et al., 2016). During development, 37 of 100 top microglia-enriched genes are consistently upregulated from E17 to P60. Again, a homeostatic microglia core signature was identified, and LPS induced a typical inflammatory gene profile.
All these studies led to the identification of a homeostatic microglia core gene expression signature, which includes Sall1, Hexb, Fcrls, Gpr43, Cx3cr1, Tmem119, Trem2, P2ry12, Mertk, Pros1, and Siglech, genes that are abundantly expressed in microglia compared to other brain or myeloid cells.
In two studies published in 2017 (Galatro,Holtman, et al., 2017;Gosselin et al., 2017), the human microglia transcriptome was reported. Gosselin et al. expression profiled microglia isolated from surgically resected brain tissue of epilepsy, brain tumor, or acute ischemia patients. Microglia were isolated by Percoll gradient centrifugation and fluorescence-activated cell sorting (FACS) on live-CD11b + CD45 Low CD64 + CX3CR1 High cells, while excluding most activated cells with moderate to high levels of CD45. The 30 most abundant transcripts in microglia across different patients are related to known microglia properties and functions like ramification and motility (P2RY12 and CX3CR1), synaptic remodeling (C3 and C1QA-C), and immune response (HLA-DRA and HLA-B) (Gosselin et al., 2017).
Comparison of the microglia transcriptome to the transcriptome of cortical brain tissue used to isolate microglia, 881 microglia-enriched (10-fold increased) genes were detected (Gosselin et al., 2017).
In both studies, the human and mouse microglia transcriptomes were compared. Gosselin et al. showed that human and mouse microglia are very similar, 13,253 of 15,768 orthologous genes pairs expressed within a fourfold range. At a cutoff of 10-fold difference, they identified 400 human microglia enriched orthologous genes and 293 mouse microglia-enriched orthologous genes. Human microglia are characterized by a higher expression of regulators of the complement system (C2, C3, VSIG4, SERPING1) and genes involved in brain structure development (SYNDIG1, GLDN, CTTNBP2, and ROBO3).
In the study of Galatro et al., the human microglia transcriptome was also compared with mouse microglia transcriptomes (Grabert et al., 2016;Matcovitch-Natan et al., 2016). An extensive overlap with human microglial data were observed, however, noteworthy differences were also observed. Human microglia-specific genes were shown to involve in immune pathways, example genes are GNLY, CD58, APOBEC3C, CLECL1, CD89, and CARD8 (Galatro, Holtman, et al., 2017). Most notably, when comparing age-associated changes in microglia gene expression between humans and mice (Grabert et al., 2016), a suprisingly low overlap was detected. Most genes with an aging-associated change in expression in humans were associated with the actin cytoskeleton (Galatro, Holtman, et al., 2017). Details on isolation methods, tissues used and which kind of comparison were used to identify microglia signature genes in these studies are summarized in Table 1.

| Mouse and human microglia heterogeneity revealed by bulk population sequencing
Microglia are plastic cells and their morphology, phenotype, and immune response display region-dependent heterogeneity (Lawson, Perry, Dri, & Gordon, 1990;Yang et al., 2013). A detailed study of the basal ganglia region revealed region-specific phenotypes of mouse microglia and this microglial diversity was partly determined by the local microenvironment (De Biase et al., 2017). At the transcriptional level, Grabert et al. was the first to demonstrate transcriptional differences between mouse microglia from the cerebral cortex, hippocampus, cerebellum, and striatum. Cerebellar and hippocampal microglia exhibited a more immune-vigilant state, with higher expression of the genes Camp and H2-Ab1 (Grabert et al., 2016). Using a microgliaspecific translating ribosome affinity purification approach, it was determined that cerebellar mouse microglia displayed a more pronounced cell clearance phenotype (Ayata et al., 2018). In contrast, using bulk population RNA-seq, Li et al. detected very limited transcriptomic heterogeneity between Tmem119 + FACS sorted microglia isolated from the above mentioned brain regions (Li et al., 2019).
Besides regional differences, gender-dependent heterogeneity in microglia gene expression was also reported. Transcriptomic profiles were generated of microglia from male and female mouse hippocampus and cortex. Male mouse microglia displayed a higher capacity to present antigens and increased responsiveness to purinergic stimuli (Guneykaya et al., 2018). Expression profiling of microglia isolated from male and female mice revealed that the gene expression program in male microglia was delayed (Hanamsagar et al., 2017).
Microglia maturation during development is shaped by microbiome-derived short chain fatty acids (Erny et al., 2015) and the effect of the microbiome on mouse microglia differentiation is sexually dimorphic (Thion et al., 2018). Perturbation of the microbiome had more profound effects in male embryos and female adults. These studies show that the microbiome is important for microglia development and maturation.
Regional heterogeneity was also demonstrated in the human CNS. Human microglia were isolated from grey matter (GM; occipital cortex) and white matter (WM; corpus callosum) from postmortem control and multiple sclerosis (MS) donor CNS. Between WM and GM microglia, 453 differentially expressed genes (logFC >2) were detected in samples from control donors, and 124 genes in MS donor-derived samples. Genes highly expressed in control GM microglia were related to "cytokine-mediated signaling," such as TNFRSF25 and CCL2; WM microglia were enriched for genes involved in "chemotaxis" and 1.3 | Single-cell RNA-sequencing (scRNA-seq) of mouse and human microglia Expression profiling of bulk population human microglia revealed changes associated with age, neurodegenerative diseases and psychiatric disorders (Galatro, Holtman, et al., 2017;Gosselin et al., 2017), and regional and gender-dependent mouse microglia heterogeneity (Ayata et al., 2018;Grabert et al., 2016;Guneykaya et al., 2018). The first disease-associated single-cell mouse microglia study was published by Keren-Shaul et al. (Keren-Shaul et al., 2017). In 5XFAD mice, an amyloid AD mouse model, a cluster of disease-associated microglia (DAM) was identified, characterized by the upregulation of genes such as Apoe, Trem2, and Tyrobp. These genes are associated with lipid metabolism and phagocytosis and were already previously identified in a meta-analysis of microglia gene expression changes in relation to aging and CNS disease . These disease-associated microglia subtypes might be promising targets for treatment of neurological diseases (Deczkowska et al., 2018). In these DAMs, genes associated with homeostatic microglia, such as P2ry12 and Tmem119, were downregulated. While the scRNA-seq study was performed in a mouse model for AD, the expression of some DAM signature genes was confirmed by immunostaining in human AD brain tissue, where these DAMs spatially associated with sites of AD pathology (Keren-Shaul et al., 2017

| Databases for transcriptome information
There are several online databases that can be used to obtain information on quantitative expression analysis in microglia. The most recent applications are: GOAD , Brain RNA-Seq (Zhang et al., 2014), and Neuroexpresso (Mancarci et al., 2017).
The aim of the GOAD database was to generate a platform for glia transcriptomics. Information from the most important glia transcriptome studies was aligned, quantified, and stored in a database. 2.2 | Microglia and nuclei isolation 2.2.1 | Microglia isolation from mouse and human brain tissue Microglia were isolated from adult mouse brain using the protocol as described before . Briefly, the brains were isolated and triturated using a tissue homogenizer. The homogenized brain samples were passed through a 70 μM cell strainer to obtain a single cell suspension. The cells were centrifuged at 220 rcf for 10 min at 4 C and the pellet was resuspended in 24% Percoll gra- Post-mortem human brain tissue of the superior frontal gyrus of two donors was obtained from the Dutch Brain Bank. Microglia were FACS sorted as DAPI neg DRAG5 pos CD45 pos CD11B pos , as previously described (Galatro, Holtman, et al., 2017). For bulk sequencing, three mice were used per condition and sequenced separately. For single cell/nucleus sequencing, sorted microglia cells/nuclei from three mice/condition were pooled and loaded on a 10× Genomics Chromium chip according to the manufacturer's instructions.

| Nuclei isolation from sorted microglia
The nuclei isolation protocol was adopted from (Krishnaswami et al., 2016

| Nuclei isolation from frozen brain tissue
Nuclei were isolated as described previously with a few adaptations

| Library preparation
For bulk sequencing, total RNA was isolated from either sorted microglia or nuclei using the RNeasy Plus Micro Kit (Qiagen,74034).
RNA quantity and quality were analyzed using an Experion electro-    Table S1).
Among these genes, in response to LPS, 111 genes were  (Table S2).
Only the top eight most significant terms are depicted, as these were most representative for the overall outcome. As expected, many significantly enriched terms were associated with the inflammatory response of microglia, and showed extensive overlap between cells and nuclei ( Figure 1e).
To determine the similarity between cellular and nuclear microglia expression profiles, we performed differential gene expression analysis between cells and nuclei in both the PBS and LPS condition. Only 22 genes were differentially expressed between cells and nuclei after either PBS (12 DE genes) or LPS treatment (16 DE genes; Figure 1f).
In summary, these data indicate that mouse microglia nuclear transcriptomes are a close approximation of cellular transcriptomes when analyzed using bulk sequencing and that the transcriptional response to LPS observed in nuclei and microglia was highly similar.  Table S4. Next, we performed differential gene expression analysis between cells and nuclei from PBS and LPS mice in the single cell/nucleus dataset. In the PBS samples, seven genes were enriched in cells compared to nuclei (logFC >1.5, FDR < 0.05), and except for Tmsb4x, these genes were all mitochondrial-related (mt-Atp6, mt-Cytb, mt-Co3, mt-Nd1, mt-Nd4, mt-Co2) (Table S5). In the LPS samples, apart from the previous seven genes, five additional cell-enriched genes were detected (mt-Co1, Rps23, Rps14, Fth1, Rpl32; Table S5). Nucleus-enriched genes were only detected in the PBS condition (Acaca, Spag5, and Gm17660) and with a less stringent cut off (logFC <−1 and FDR < 0.05), six genes were enriched in LPS nuclei (Acaca, Kazn, Gm17660, Gm26916, Mylip, and Vps13a) (Table S5) Table S5), in agreement with earlier findings (Bahar Halpern et al., 2015;Bakken et al., 2018). Some representative genes, mentioned above, are depicted in Figure 2f,g. Taken together, single-microglia nucleus gene expression profiles are a reliable proxy for single microglia transcriptomes in mice.

| Comparison of nuclear and whole cell transcriptome by single cell/nucleus sequencing in human microglia
To investigate whether a similar overlap between the cellular and nuclear transcriptomes was present in human microglia, we isolated microglia and microglia nuclei from fresh post-mortem brain tissue of two human donors. In addition, we froze adjacent tissue blocks of the same donors and then isolated nuclei from these samples to evaluate whether nuclei isolated from frozen tissue can be used to determine  (Table S6).
Overall, microglia nuclear transcriptomes from both fresh and frozen CNS tissue are a good proxy for freshly isolated microglia and potential subpopulations present in the tissue.

| DISCUSSION
Single nucleus RNA sequencing is considered to have several advantages over single cell RNA sequencing. First, nuclei are more resistant to mechanical stress and cryopreservation, which would make large collections of well-characterized (frozen) tissues in biorepositories amenable for single nucleus profiling (Krishnaswami et al., 2016). Second, single-nucleus RNA sequencing is less cell type-biased than single cell RNA sequencing. Some cell (sub)types are more vulnerable to tissue dissociation than other cell types, resulting in potential underand over-representation of cell types or subsets thereof in the data (Bakken et al., 2018). Several studies using brain tissue have shown that nuclei reflect transcriptional changes at the tissue level, single cell level and also recapitulate subtypes and diversity in neurons (Habib et al., 2016;Lake et al., 2016;Lake et al., 2017). However, for microglia, it is yet unknown whether nuclei can serve as an alternative for cellular transcriptomes, whether nuclear transcriptomes contain enough information to identify microglia subtypes, and whether this is amenable to frozen CNS tissues. Hence, we performed a systematic comparison of mouse and human fresh microglia and nuclei, and additionally microglia nuclei isolated from frozen human brain tissue.
The number of microglia in the brain is relatively low (Keller, Ero, & Markram, 2018) and, as a consequence the number of microglia in total CNS tissue single cell sequencing data is also relatively low (Darmanis et al., 2015;Mathys et al., 2019). In order to enrich for microglia nuclei from frozen tissue, nuclei from neurons and oligodendrocytes were labeled with antibodies against NeuN and OLIG2 and selected against during FACS isolations. After sequencing, first the NeuN/OLIG2 double negative nuclei population was clustered, and To investigate the differences and similarities between single cell and single nucleus RNA sequencing (Selewa et al., 2019), we compared these technologies in LPS-challenged mice, as that induces a strong and well characterized transcriptional response in microglia (Holtman, Raj, et al., 2015). First, we used bulk RNA sequencing to determine to what extent the nuclear and cellular gene expression profiles in mice overlap. As expected, we obtained less RNA from nuclei, approximately 20% of the amount of total RNA typically iso-  (Figures 1d and 2e).
Next, we investigated whether human microglia nuclei, including nuclei isolated from frozen CNS tissues, could reliably recapitulate cellular transcriptomes generated with microglia isolated from fresh CNS tissue. Clustering analysis of cells and nuclei from both donors combined showed that cells and fresh nuclei clustered very similarly.
The distribution of clusters in frozen nuclei was slightly altered but all the clusters detected in fresh microglia were recapitulated in frozen nuclei (Figure 3e). The fresh nuclei were isolated from the same microglia sample used for cellular profiling, and hence should be extremely similar. The frozen nuclei were isolated from an adjacent tissue block of the same donor which may had a slightly different cellular composition. Differences in the relative amounts of WM and GM between the fresh and frozen tissue samples would already result in changes in gene expression and cluster sizes. In addition, the isolation methods used for fresh and frozen nuclei were different, possibly contributing to the observed differences by preferential enrichment or loss of nuclear subtypes (Cluster 2 in Donor-1 and Cluster 1 in Donor-2), due to different sensitivities to freeze-thaw attrition. Importantly, donor variation, a reported parameter in single-cell microglia data (Olah et al., 2018), was equally detected in microglia cells and both fresh and frozen nuclei in the two donors analyzed. This indicates that donor-associated changes in gene expression were preserved in frozen microglia nuclei and that they hence reliably recapitulate the gene expression profile of fresh tissue microglia (Figure 3e).
By comparing nuclear and whole cell microglia transcriptomes by bulk sequencing and single nucleus/cell sequencing in human and mouse, we confirm that microglia nuclei are a reliable proxy for the single cell microglia transcriptome. This enables the use of banked human specimens to investigate microglia in neurodegenerative disease and neurological disorders.

| FUTURE PERSPECTIVES
Gene expression profiling of isolated cells, either in bulk or at the single cell levels, is inherently associated with loss of spatial, contextual information of the used tissue. Relatively recently, two technologies were reported where gene expression profiles were generated while retaining spatial information of the analyzed tissue section.
The first technology, spatial transcriptomics (Stahl et al., 2016), makes use of glass slides on which oligo d(T) primers with positional barcodes are spotted to capture the mRNA present in overlaid tissues. These positional barcodes enable the maintenance of positional information throughout the process of cDNA synthesis, library preparation, and sequencing. With decreasing spot diameters and spot distance, resolution will further increase, which is required for single cell analysis. A second approach to sequencing RNAs in the context of cells and tissues is fluorescent in situ sequencing (FISSEQ) (Lee et al., 2014). FISSEQ is a technology combining RNA-FISH and next generation sequencing, allowing for detection of multiple RNAs at subcellular resolution. Where FISSEQ provides a higher resolution than spatial transcriptomics, it requires a panel of oligonucleotides to detect mRNAs of interest where spatial transcriptomics is unbiased and does not required a priori knowledge about genes of interest.

ACKNOWLEDGMENTS
The authors thank N. Brouwer for technical support and G. Mesander and J. Teunis from the central flow cytometry unit at the UMCG. This work was supported by a China Scholarship Council fellowship to YH.