Stem cell derived human microglia transplanted in mouse brain to study human disease

Abstract

While genetics highlight the role of microglia in Alzheimer’s disease (AD), one third of putative AD-risk genes lack adequate mouse orthologs. Here, we successfully engraft human microglia derived from embryonic stem cells in the mouse brain. The cells recapitulate transcriptionally human primary microglia ex vivo and show expression of human specific AD-risk genes. Oligomeric Amyloid-β induces a divergent response in human vs. mouse microglia. This model can be used to study the role of microglia in neurological diseases.

Introduction

Forty-one percent of human genes lack convincing 1:1 mouse orthologs, complicating modelling diseases in mice¹. We focused on 44 genome-wide significant genetic loci (p < 5x10-8) identified by different AD GWAS studies and selected the genes nearest to the lead SNP to build a list of candidate AD-risk genes^2–4 (Figure 1a, Supplementary Table 1). We found that 15 of these genes lacked a clear 1:1 mouse ortholog (Figure 1b), e.g. CR1 or APOC. Other genes, such as CD33 and the MS4A4-cluster have many-to-many orthology with low protein sequence similarity, suggesting functional divergence. Nine additional AD-risk genes are <60% identical to their mouse ortholog¹, including TREM2. Even the largest AD genetic risk factor, the APOE polymorphism, does not exist in rodents. In addition, current in vitro systems to model human microglia display artificially induced transcriptional signatures⁵, limiting their use in disease modelling.

Figure 1. Human ESC-derived microglia successfully engraft the mouse brain.

(a) Selection of 44 genes with p>5x10^-8 from 3 landmark studies in the field. See online methods. (b) From these 44 candidate AD risk genes^3,4,8, 15 (marked with a red dot) do not have a clear 1:1 mouse ortholog or display <60% identity between human and mouse at the primary amino acid sequence. Colour scale, green (high similarity) to red (low similarity). (c) Schematic representation of the area of mouse brain covered by transplanted human microglia. Microglia are represented by green dots, and the distance between anatomically consecutive sections is 500μm. (d) H9-microglia successfully engraft the mouse brain and (e) express homeostatic markers TMEM119 and P2RY12 (n=4 mice). Scale bars of 100 and 5μm, respectively. (f) Transplanted cells distribute across the parenchyma forming a mosaic with similar nearest neighbour distance (NND) and density to that of mouse cells from adjacent areas (n=4 mice per group, two-tailed t-test p=0.9, graph shows mean±SEM). H9-microglia are labelled in green (Iba1⁺ GFP⁺), whereas arrowheads highlight few mouse cells (Iba1⁺ GFP^-) co-existing with H9-microglia in the grafted areas of the parenchyma (n=4 mice). Scale bar, 100 μm. (g) Higher magnification microphotographs and 3D reconstruction by Imaris show typical morphology with high complexity branching in H9-microglia (n=4 mice). Scale bar 5 μm.

Here, we investigated survival, integration and transcriptomic features of human microglia transplanted in mouse brain.

Results

ESC-derived microglia survive and integrate in the mouse brain

We differentiated H9 embryonic stem cell (ESC) into microglia using cytokines CSF1, IL-34, TGF-β and CX3CL1 (Supplementary Figure 1)⁶, and transplanted them into the brain of Rag2^-/- Il2rγ^-/- hCSF1^KI mice (hCSF1^KI) at P4⁷. We created a permissive environment for human microglia integration by pre-treating the neonates with Colony-Stimulating Factor 1-Receptor (CSF1R) inhibitor BLZ945⁸, removing an average of 53±7% of host microglia (Supplementary Figure 2). After 8 weeks, H9-microglia, representing 9±5% of the total microglial population (Extended Data 1), showed a mosaic distribution across multiple areas of the brain (Figure 1c and d; Extended Data 2), with nearest neighbour distance⁹ and density in transplanted areas similar to host mouse cells (n=4, Figure 1e). H9-microglia showed a complex ramified morphology and expressed homeostatic markers TMEM119 and P2RY12 (Figure 1d-g).

ESC-derived microglia mimic primary human cells at the transcriptome level

We compared the single cell transcriptomic profile of 2,246 transplanted H9-microglia (in vivo) (n=3/1: 3 mice in 1 combined sequencing pool), versus 4496 H9-derived monocytes (n=2/1: 2 differentiations in 1 combined sequencing pool) and 3385 microglia in vitro (n=2/1), and 22,846 human primary microglia obtained from cortical surgical resections (n=7/7; Extended Data 3; Supplementary Table 2; online Methods and Reporting Summary). We excluded B- and NK/T-cells (316), oligodendrocytes (1159), cycling cells (60), and doublets (172) (Figure 2a-c, Extended Data 3). Using Seurat, we defined 6 main clusters named In vitro-1 Monocytes (MNC), In vitro-2 Microglia (MG), In vivo-Homeostatic Microglia (HM), Cytokine Response Microglia (CRM), CNS-Associated Macrophages (CAM)¹⁰, and Neutrophils (Nφ) (Figure 2a; Extended Data 3), based on experimental data¹¹ and meta-analysis from microglial transcriptional profiles¹². CRM represents a novel cluster and is defined by an upregulation of genes encoding cytokines/chemokines (Extended Data 3; Supplementary Table 3). More than 97% of the in vitro derived H9-monocytes and microglia were present in In vitro clusters (Figure 2a-c), whereas 79% primary microglia isolated from human brain and 60% of transplanted H9-microglia distributed into the In vivo-HM cluster (Figure 2a-c; Extended Data 3d and e). A smaller percentage (13%) of primary compared to H9 transplanted microglia (35%) were present in the In vitro clusters. In addition, some cells showed a CNS-associated macrophage (CAM) expression profile (Figure 2a and e). Immunohistochemistry and in situ hybridization confirmed that CAM cells were in proximity to blood vessels and expressed the perivascular macrophage marker MRC1 (Figure 2e, lower panels). The engrafted H9 cells expressed the microglia markers CX3CR1 and P2RY12 (Figure 2e, upper panels).

Figure 2. H9-microglia isolated 8 weeks after transplantation are similar to human primary microglia.

(a) t-SNE plot visualizing 33,144 single cells sorted based on CD11b (primary human), CD11b hCD45 and GFP (engrafted H9-microglia) staining, and in vitro derived monocytes (MNC) and microglia (MG) after quality control, and removal of peripheral cells, cycling cells and doublets. Cells are coloured according to clusters identified with Seurat’s kNN and merging: In vitro-1 MNC, In vitro-2 MG, In vivo-Homeostatic Microglia (HM) and Cytokine Response Microglia (CRM), CNS-Associated Macrophages (CAM)¹⁰, and Neutrophils (Nφ). The assignment of different clusters to distinct cell types/states is based on previous experimental data from our lab¹¹ and a recent meta-analysis describing multiple modules of microglial transcriptional profiles¹², as detailed in Extended Data 4a-c and Supplementary Table 3. (b, c) Distribution and percentage of cells from either in vitro, in vivo (engrafted) H9 or primary human microglia across the different clusters identified. (d) Most highly expressed genes in the different samples: in vitro-1 MNC; in vitro-2 MG, in vivo (engrafted) H9 and primary microglia. (e) In situ hybridization for CX3CR1 and P2RY12 (microglia) and MRC1 (perivascular macrophages) confirming the location of the two main distinct identities acquired by H9 engrafted cells (GFP) in the mouse brain (n=4 mice). Scale bar is 25 μm and 10 μm in the left and right panels, respectively. (f, g) Volcano plots showing gene expression differences between average gene expression in (f) 22,846 primary vs. 3385 in vitro MG and (g) 22,846 primary vs. 2,246 engrafted H9-microglial cells (with a logFC threshold of 0.2, Wilcoxon Rank Sum test, p-values adjusted with Bonferroni correction based on the total number of genes in the dataset). Genes associated to homeostatic or activation expression profiles are highlighted in blue and red, respectively (Supplementary Table 3).

Direct comparison between experimental groups revealed that in vitro monocytes/microglia displayed >300 differentially expressed genes (logFC>0.2) compared to microglia from surgical samples, consistent with an “activated” profile (Figure 2d and f; Extended Data 4; Supplementary Table 3). In contrast, engrafted H9-microglia displayed a comparable homeostatic signature to that of the cells isolated from the human brain, with only 41 differentially expressed genes (Figure 2d and g; Extended Data 4a and b). Therefore, the mouse CNS environment is sufficient to drive microglia from an artificial in vitro “activated” towards a more natural homeostatic brain resident phenotype.

Human ESC-derived and host mouse microglia display a divergent response to oligomeric Aβ

We tested our humanized system with an acute AD-related challenge, i.e. oligomeric Aβ₄₂ (oAβ), previously shown to induce cognitive alterations¹³ (Extended Data 5). Mice were injected in the ventricle with 5 μM oAβ (n=3) or scrambled peptide (Scr, n=3) at 8-10 weeks post transplantation. We isolated 4880 transplanted H9-microglia 6 hours after injection (n=3x2/2), excluding CNS-associated macrophages and cycling cells (Extended Data 6). Clustering analysis revealed a homeostatic (H9.HM), a “primed” (H9.PM), and a cytokine (H9.CRM) cluster (Figure 3a; Extended Data 7). The H9.HM cluster was significantly enriched with scrambled peptide treated cells (68%, Chi²-test, p-value<2.2x10^-250) and showed high expression of multiple homeostatic genes (Figure 3a; Extended Data 7; Supplementary Table 3). The “primed” H9.PM cluster was very different from the previously characterized activated response (ARM) response in wild type mouse microglia¹¹ as it expresses an unusual mixture of homeostatic and activation genes¹¹, and consisted of a larger proportion of scrambled (65%) vs. oAβ cells (35%, Chi² test, p-value<10^-250) (Figure 3a and b; Extended Data 7). Finally, the H9.CRM cluster was significantly enriched in cells from oAβ treated mice (75%, Chi² test, p-value<2.210^-250), and displayed high levels of multiple inflammatory cytokines and chemokines, such as IL1B, IL6, CCL2, CCL4, etc. (Figure 3a and b; Extended Data 7). Trajectory analysis¹⁴ revealed a phenotypical change of H9-microglia from homeostatic towards the cytokine-response state (Figure 3b; Extended Data 7) with microglia from the H9.PM cluster enriched in the initial and middle phases, indicating they might represent an early response to the injection of peptides (Figure 3b; Extended Data 7).

Figure 3. Human and host mouse microglial response to oligomeric Aβ.

(a, b) Analysis of the response of H9-microglia upon oAβ exposure. (a) t-SNE plot visualizing the 4880 H9 microglia passing quality control, and after removal of CAM, cycling cells and doublets. Cells are coloured according to clusters identified with Seurat’s kNN (upper panel; H9.HM: Homeostatic Microglia, H9. PM: Primed Microglia, H9.CRM: Cytokine Response Microglia, H9) and treatment (lower panel; Scr: scrambled peptide, oAβ: oligomeric Ab). (b) Plot of the phenotypic trajectory followed by H9-microglia upon oligomeric Aβ exposure, obtained by an unbiased pseudotime ordering with Monocle 2 and coloured by clusters as in d. H9-microglia followed a trajectory from H9.HM and H9.PM, to H9.CRM. The heatmap shows the differential expression of representative genes from each cluster, ordered by pseudotime. (c, d) Analysis of the response of endogenous (Rag2^-/- Il2rγ^-/-) mouse microglia upon oligomeric Aβ challenge. (d) t-SNE plot visualizing the 9942 endogenous mouse microglia passing quality control, and after removal of peripheral cells, CNS-Associated Macrophages (CAM) cycling cells and doublets. Cells are coloured according to clusters identified with Seurat’s kNN (upper panel; ms.HM: (mouse) Homeostatic Microglia, ms.CRM: Cytokine Response Microglia, ms.ARM: Activated Response Microglia) and treatment (lower panel; Scr: scrambled peptide, oAb: oligomeric Ab). (c) Plot of the phenotypic trajectory followed by endogenous mouse microglia upon oligomeric Aβ exposure, obtained by an unbiased pseudotime ordering with Monocle 2 and coloured by clusters as in a. Mouse microglia followed a trajectory from ms.HM to ms.CRM to ms.ARM. The heatmap shows the differential expression of representative genes from each cluster, ordered by pseudotime. (e) Correlation analysis of the log-fold change (logFC) in H9 (y-axis) and host (Rag2^-/- Il2rγ^-/-) mouse (x-axis) microglia upon oligomeric Aβ challenge relative to scrambled peptide (Pearson correlation, R=0.4. Differentially expressed genes are highlighted in green when significant in both species, blue only in H9-microglia or orange only in mouse microglia. Numbers between brackets in the legend represent the amount of up and downregulated genes in each group, respectively. (f) Expression changes induced by Aβ challenge in the selected candidate AD-risk genes (Figure1b). (g) Extension of the table shown in Figure 1a highlighting the important number of putative AD-risk genes in humans that lack good orthologues in mice or show an opposite behaviour upon Aβ challenge (highlighted by red dots). Expression profile of 44 putative AD genes in our datasets (H9-microglia; primary human microglia from 7 patients; and mouse host Rag2^-/- Il2rγ^-/-microglia, mouse RM), and wild type mouse microglia from 2 independent datasets of 12-week-old immunocompetent C57Bl/6 mice (Sala Frigerio et al.,¹¹, SF; and Keren-Shaul et al., KS¹⁵). We identified 15 genes with observed expression in human but not mouse microglia and, that were also observed in H9-microglia.

At the same time, we isolated and sequenced 9942 host mouse microglia (after exclusion of CNS-associated macrophages and other immune or cycling cells) from the same animals to compare their reaction to that of H9-microglia (Figure 3c and d; Extended Data 6). Whereas we acknowledge that the genetic background of the host might cause (unknown) developmental abnormalities, analysis of different wild-type mouse microglial datasets did not reveal expression of Rag2 or Il2ry¹⁵ and, although the effect of Il2ry deficiency on microglia is not documented, Rag2 deficiency does not affect microglial number, morphology or gene expression profiles¹⁶. Clustering analysis yielded a homeostatic (ms.HM), a cytokine (ms.CRM), and an activated (ms.ARM) response cluster. The HM cluster was significantly enriched with control cells (70%), whereas the CRM and ARM clusters mostly consisted of cells from the mice treated with oAβ, (69% and 77%, Chi² test, p-value<10^-250) (Figure 3d; Extended Data 8). The ARM cluster showed a similar profile to that of microglia responding to amyloid plaques¹¹ (Figure 3c and d; Extended Data 8). Trajectory analysis showed that mouse microglia transition from homeostatic to cytokine-response to activated-response cells (Figure 3c; Extended Data 8), suggesting that they form a single successive response of mouse microglia to oAβ. We also assessed whether the CRM transcriptomic signature identified here is uniquely elicited by oAβ, as it has not yet been described in the response to Aβ plaques^11,15 (Supplementary Table 3). Reanalysis of previous data on microglial cells from 3 to 21 months old APP^NL-G-F mice revealed a small number of cells, previously embedded in the ARM cluster, that displayed a CRM profile (Extended Data 9a-c). In addition, these cells were positioned in the early ARM phase of the trajectory analysis, suggesting that they are part of a common early response to both oAβ and Aβ plaques (Extended Data 9d). We acknowledge that the current work only provides proof of concept, while further more systematic work is ongoing to fully dissect the acute and chronic responses of mouse and human microglia to oAβ and Aβ plaques.

We finally evaluated whether this chimeric model covers the human expressome better than the classical mouse models. We extracted 10,914 one-to-one, bidirectional orthologs between mouse and human (Supplementary Table 2)^1,17 and performed a correlation analysis comparing log-fold changes in gene expression in the CRM vs HM comparison done in each species (FRD < 0.05). We observed a significant, but rather limited correlation in the response to oAβ (R=0.4, Pearson correlation, p-value ≈ 0) with a number of genes changed in mouse or human alone (logFC > 0.2; Figure 3e; Extended Data 10), 207 of them showing opposite behaviour (Figure 3f; Extended Data 10c), such as TYMP, NFKB, PPARG, LIMK2 and TGFBR1, a homeostatic microglia marker in mouse¹⁸ (Extended Data 10c), and the AD-risk genes ABI3, BIN1 and PICALM (Figure 3f). We also explored how the 8266 human genes with no clear mouse ortholog reacted to oAβ and found 79 and 127 uniquely up- and down-regulated human genes, mainly involved in cytokine/chemokines responses (Extended Data 10d and e). The human response was particularly strong for IL1B and CCL2 (Figure 3e, red arrows), which have been experimentally implicated in the pathology of AD^19,20 (Figure 3e, Extended Data 10). Remarkably, 12 of the 15 AD-risk genes identified as lacking 1:1 mouse orthologs (in Figure 1a), were expressed in primary microglia from surgical samples (Figure 3g), confirming the association of genetic risk of AD with microglia. Reassuringly, all these genes were also detected in the transplanted human H9-microglia (including APOC, CD33, CR1, MS4A and TREM2). The similarities in gene expression between Rag^-/- Il2rγ^-/- and wild type mouse microglia (Figure 3g) further supports the proof of concept study presented here.

Discussion

Although in vitro studies may provide some mechanistic insights into the function of human microglia, it is also clear that signals from the CNS microenvironment are required to sustain microglial specification, and that a loss of those cues dramatically disrupts the microglia phenotype driving them towards an activated state⁵. In addition, some AD-linked genes (e.g TREM2-membrane phospholipids/APOE, CD33-sialic acid, etc.) play a role in the cross-talk between microglia and other brain cells. The main challenge is to understand this cellular phase of AD²¹ and therefore introducing those complex aspects into a model of disease is extremely important. We present here a novel model using ESC-derived human microglia transplantation into the mouse brain providing the human cells with the crucial environment that defines microglial identity. Given the limited similarity between mouse and human microglia in terms of candidate AD-risk genes, this model provides a very useful alternative to study the response of human microglia in vivo in the context of AD and other diseases affecting the CNS, opening important new routes to understand the role of the many genes identified in the GWAS and other genetic studies which are not well modelled in mouse cells.

ESC-derived human microglia transplanted into mouse brain represents clearly a step forward to model part of the GWAS defined risk of AD. Despite certain limitations that should be considered (e.g. lack of adaptive immune system, variability in the grafting efficiency of different pluripotent stem cells, iPSC), we anticipate that our approach will be widely applicable to study other neurological diseases. The use of human H9 cells in combination with CRISPR/Cas9 technology opens unanticipated possibilities to model human specific genetic aspects of brain disease.

Methods

Sample size was estimated based on previous experiments performed in the lab¹¹. No samples were excluded from the analysis and all attempts at replication were successful. The experimental groups were ramdomised to avoid gender, litter and cage effects. Investigators were blinded when performing all experiments

Human vs. mouse gene orthology and selection of putative AD-risk genes

We determined the orthology between human and mouse genomes using Ensembl Biomart¹. We defined “good orthology” as every gene with one-to-one, bidirectional orthology between the two species and >60% protein sequence similarity. This resulted in a total of 10,914 genes. The full list is shown in Supplementary Table 3.

We based our selection of putative AD-risk genes on several recent publications^2–4. We focused on 44 genome-wide significant loci (p<5x10-8) described in these publications and selected as being the nearest gene to the lead SNP. We used the union of these gene sets for our analysis. In summary, we extracted 23 genes from Lambert et al. (Table 2 of the original report)³, 33 genes from Jansen et al. (Table 1 of the original report)² and 21 genes from Kunkle et al. (Table 1 of the original report)⁴. Figure 1a shows the distribution of these genes across the different reports and illustrates how they overlap.

In vitro generation of microglia from ESCs

In vitro microglia differentiation from embryonic stem cells was based on previously described protocols⁶. On days 17, 21, 25, 28, and 32, non-adherent cells were harvested and selected using CD14-labelled magnetic beads (Miltenyi) following manufacturer specifications. Briefly, cells were collected and centrifuged for 5 min at 300g. Then, cells were incubated for 15 min at 4°C in 80 µl MACS buffer (AUTOMACS + 5% MACS serum, Miltenyi) with 20 µl of CD14-beads (Miltenyi), and passed through a LS column (QuadroMACS, Miltenyi). The CD14+ fraction was collected and centrifuged for 5min at 300g. Monocytes were then differentiated into microglia-like cells using microglia differentiation medium (TIC) (DMEM/F12, Glutamine (2mM), N-Acetyl Cysteine (5µg/mL), Insulin (1:2000), Apo-Transferrin (100 µg/mL), Sodium Selenite (100 ng/mL), Cholesterol (1.5 µg/mL), Heparan Sulphate (1 µg/mL)) supplemented with 50 ng/ml IL34, 50 ng/mL M-CSF, 10 ng/ml CX3CL1 and 25 ng/mL TGF-β, based on Abud et al. (2017)²². The medium was changed every other day.

Mice

Rag2^-/- IL2rγ^-/- hCSF1^KI mice were purchased from Jacksons Labs (strain 017708), and bred and maintained in local facilities. All the experiments were performed in these mice as human microglia require hCSF1 for their growth and survival⁷. Mice were housed in groups of 2-5, under a 14 h light/10 h dark cycle at 21°C, with food and water ad libitum. All experiments were conducted in 8-12 weeks old male and female according to protocols approved by the local Ethical Committee of Laboratory Animals of the KU Leuven (government licence LA1210591, ECD project number P177/2017) following local and EU guidelines.

Endogenous mouse microglia depletion

The CSF1R inhibitor BLZ945 was dissolved in 20% (2-hydroxypropyl)-β-cyclodextrin (Sigma-Aldrich). Newborns were injected (i.p.) 24 and 48h prior to human cell transplantation at a dose of 200 mg/kg bodyweight.

Transplantation of human microglia into the mouse brain

Grafting of human PSC-derived microglia was performed as previously described²³. Briefly, human microglia were dissociated and suspended at a concentration of 100,000 cells/μl in PBS. At P4, mice were anaesthetized by hypothermia and bilaterally injected with 1μl of cell suspension at coordinates from Bregma: anteroposterior, -1mm; lateral, ±1mm. After the injections, mice were allowed to recover on a heating pad at 37°C, and then transferred back to their cage.

Isolation of human primary microglia

Human primary microglia were isolated from brain tissue samples resected from the temporal cortex during neurosurgery. All samples represented lateral temporal neocortex and were obtained from patients who underwent amygdalo-hippocampectomy for medial temporal lobe seizures. The mesial temporal specimens were sent to pathology and thus not available for study purposes.

Samples were collected at the time of surgery and immediately transferred to the lab for tissue processing, with post sampling intervals of 5-10 min. All procedures were conducted to protocols approved by the local Ethical Committee (protocol number S61186).

Preparation and intracerebral injection of oligomeric amyloid

Oligomeric Aβ 1-42 (oAβ, 5μl 10μM) or scrambled peptide (Scr, 5μl 10μM) were prepared as previously described by Kuperstein et al.²⁴ Briefly, recombinant amyloid beta 1-42 peptide (rPeptide; #A-1163-1) or scrambled amyloid beta 1-42 (rPeptide; #A-1004-1) were thawed at room temperature 30 minutes before preparation. Peptides were solubilized in 99% hexafluoroisopropanol (HFIP) (Sigma-Aldrich; #105228) at 1 mg/ml concentration. The HFIP was evaporated using a stream of nitrogen gas, the resulting peptide pellet was resolved in dimethylsulfoxide (DMSO; Sigma-Aldrich; #D4540), at final concentration of 1 mg/ml. DMSO was exchanged with Tris-EDTA (50 mM Tris and 1 mM EDTA, pH 7.5) using 5-ml HiTrapTM desalting columns (GE Healthcare; #17-408-01). The eluted peptide concentration was determined using Bradford reagent (Bio-Rad; #5000006) according to the manufacturer’s instructions. The eluted peptide was left to oligomerize at room temperature for two hours in Tris-EDTA buffer. oAβ or scrambled peptide was further diluted to 10 µM in Tris-EDTA buffer and stored at -80°C until use. At 8-10 weeks of age, grafted mice were anesthetized with a ketamine/xylazine mixture (85 and 13 mg/kg), and 5 μl of either oAβ (10 μm) or scrambled peptide (10 μm) were stereotactically injected in the left ventricle at the following coordinates from bregma: anteroposterior, -0.1 mm; mediolateral, +1 mm; dorsoventral, -3 mm. Mice were allowed to recover in a thermo-regulated chamber and then transferred back to their original cage. Isolation of microglia was performed 6h after the intracerebral injection of the peptides.

Isolation of human and mouse microglia from the mouse brain

Mice were terminally anesthetized with an overdose of sodium pentobarbital and transcardially perfused with heparinized PBS. Brains were harvested in PBS 2%, FCS, 2mM EDTA (FACS buffer), mechanically triturated and enzymatically dissociated using the Neural Tissue Dissociation Kit (P) (Miltenyi) following manufacturer specifications. Then, samples were passed through a cell strainer of 70μm mesh (BD2 Falcon) with FACS buffer, and centrifuged twice at 500g for 10 min at 4°C. Next, cells were resuspended in 35% Percoll (GE Healthcare) and centrifuged at 500g for 15 min at 4°C. The supernatant and myelin layers were discarded, and the cell pellet enriched in microglia was resuspended in FcR blocking solution (Miltenyi) in cold FACS buffer, following manufacturer specifications. After a wash, primary antibody labelling was performed for 30 min at 4 °C, using the anti-CD11b (Miltenyi) and anti-hCD45 (BD Bioscience), adding e780 (eBiocience) as a cell viability marker. Moreover, unstained cells and isotype-matched control samples were used to control for autofluorescence and/or non-specific binding of antibodies. Samples were run on a BD FACS Aria II Flow Cytometer and data were analysed using FlowJo and FCS express software. Human cells were sorted according to the expression of CD11b, hCD45, and GFP, whereas mouse cells only expressed CD11b but were negative for hCD45 and GFP (Extended Data 2). For each experimental condition, we pooled the same number of cells from three mice.

Histological analysis

Mice were terminally anesthetized with an overdose of sodium pentobarbital and transcardially perfused with heparinized PBS and 4% PFA in PBS. Brains were harvested, post fixed in 4% PFA overnight, and cut in transverse serial sections (35 μm thick) with a vibrating microtome (Leica). For each sample, 6 series of sections were sequentially collected in free-floating conditions and kept in cryoprotectant solution at −20°C. Sections were blocked with 5% normal serum in PBS-0.2% Tween 20 for nonspecific binding. After rinses with PBS-0.1% Tween 20 (PBST), sections were incubated overnight at 4°C with anti-GFP (Abcam, ab13970), anti-Iba1 (Wako, 019-19741), anti-P2RY12 (Sigma Aldrich, HPA014518) and anti-TMEM119 (Abcam, ab185333). After washes with PBST, sections were incubated with the appropriated biotinylated (Vector Labs) or Alexa 488- and 594-conjugated secondary antibodies (Invitrogen) for 1h at RT. When necessary, sections where incubated with Alexa 488-conjugated Streptavidin (Invitrogen) for 1h at RT. Finally, sections were counterstained with DAPI and mounted with Mowiol/DABCO (Sigma-Aldrich) mixture. Sections were visualized on a Nikon A1R Eclipse confocal system. Nearest neighbour distance (NND) analysis was performed in 20X microphotographs by using a script for Fiji (ImageJ) as previously described by Davis et al. (2017).⁹

Single cell mRNA libraries preparation and sequencing

After microglial isolation, we performed single cell RNA sequencing by using 10X Genomics single cell gene expression profiling kit. cDNA libraries were produced following manufacturer instructions. cDNA libraries were then sequenced in an Illumina HiSeq platform 4000 with the sequencing specification recommended by 10X Genomics workflow. For each experimental condition, we pooled the same number of cells from three mice.

Human-mouse orthologs

Human to mouse and mouse to human orthologs tables were downloaded from Ensembl/Biomart (release 94)¹. From these tables, only those genes were extracted that have a clean one-to-one bidirectional ortholog. After filtering out genes that do not express in our human and mouse microglia datasets, the table resulted in 10914 genes (Supplementary Table 4).

Statistics

Analysis of histological data

Nearest neighbor distance (NND) and microglial density data (from Figure 1) were analysed with a two-tailed t-test. Data distribution was assumed to be normal but this was not formally tested. P values < 0.05 were consider statistically significant at a confidence interval of 95%. Data were represented as mean±SEM.

Analysis of single cell RNA sequencing datasets

Alignment. The raw BCL files were demultiplexed and aligned by Cellranger (version 2.1.1) against a human genome database (build hg38 build 84) and mouse database (mm10 build 84). Raw count matrices were imported in R (version 3.4.4) for data analysis.

Quality control of cells - step 1. For each dataset, to exclude poorly sequenced cells, damaged cells and dying cells, we filtered out cells with less than 1000 reads or less than 100 genes detected; moreover, we excluded cells with more than 10% of reads aligning to mitochondrial genes. Cells with a number of reads or genes above 3 standard deviations from the sample mean were considered as doublets and removed. Genes detected in less than 3 cells were excluded from the count matrices. Data were analysed by principal component analysis (PCA) to identify any obvious batch effects. For the joint analysis of H9-derived microglia and primary microglia from surgical resections (Figure 2), the mean depth of sequencing was 102,000 reads/cell, while the mean number of genes detected per cell was 2072. For the analysis of mouse microglia (Figure 3), the mean depth of sequencing was 68,000 reads/cell, while the mean number of genes detected per cell was 1777. For the analysis of H9-derived microglia (Figure 3), the mean depth of sequencing was 96,000 reads/cell, while the mean number of genes detected per cell was 1964.

Quality control of cells - step 2. We analysed each dataset using the R package Seurat (version 2.3.4)²⁵ for the mouse and H9-derived microglia datasets, and version 3.0²⁶ for the joint analysis). We performed principal component analysis (PCA) on both the mouse and H9-derived microglia datasets, after data normalization and scaling and selection of the most variable genes, respectively 2000 and 1390. We selected the first principal components (PCs), 20 for mouse and 20 for H9-derived cells, based on a scree plot (i.e. a plot of the PC eigenvalues in decreasing order) as input for the downstream calculations. Clusters are identified using Seurat's FindClusters function. Further non-linear dimensionality reduction for visualization is done using t-SNE. The standard workflow was followed also for the joint analysis, see Data integration and Joint clustering section.

In the joint dataset integrating in vitro H9 MNC and MG, in vivo H9 microglia and primary microglia from human cases, unbiased clustering by Seurat identified 13 major cellular populations (integrated clustering resolution = 0.8) after removal of B cells (28 cells, marked by CD52, CD48 expression), NK/T cells (288 cells, marked by NKG7, CD247, CD7 expression), oligodendrocytes (1159 cells, marked by MBP, PLP1 expression), cycling cells (60 cells, marked by TOP2A expression), doublets (172 cells, co-expressing microglial and neuronal/astrocyte markers) and a microglial cluster with very low number of reads and genes (168 cells, with mean genes = 545.9/cell, mean reads = 967.5/cell), probably reflecting damaged or low-quality cells. Post-QC a total of 32973 microglia, CNS-associated macrophages, monocytes and neutrophils cells were retained for further analysis. Seurat clusters were merged in 6 main cell types/states (Figure 2a) according to transcriptomic profile similarities as indicated by differential expression analyses and signature scoring of cells based on published single-cell microglia datasets (Extended Data 3; Supplementary Table 3). Stability of the clustering was assessed by multiple runs of analysis exploring different combinations of parameters and clusters-correlation analyses, in order to avoid over- or under-clustering of the data.

For the mouse microglia dataset, we identified 12 major cellular populations, most of them showing a tight distribution on the t-SNE plot (Extended Data 6a), with two main clusters (6 and 7) clearly separating, as well as four other very small clusters (9,10,11,12). Clusters 0 to 5 expressed high levels of homeostatic microglia markers, which were not expressed in the other, separated, clusters (Extended Data 6b). Cluster 8 expressed activated microglia and cytokines markers (Extended Data 6b). Based on a panel of marker genes (Extended Data 6c), we could identify enrichment for markers of different cell types other than microglia in the six separated clusters. Clusters 6 and 7 showed high expression levels of gene markers of neutrophils (Ccrl2) and monocytes (Ccr2), respectively. Clusters 9, 10, and 12, all composed by very small number of cells, expressed gene signatures of other brain cells (astrocytes (Clu), neurons (Npy), oligodendrocytes (Mbp). Cluster 11 was enriched in markers of cycling cells (Top2a). Overall, 89% of cells (13342/15036) in our post-QC dataset were microglia, and only these cells were retained for further analysis. The final analysis was performed on oAβ and scrambled peptide-treated cells (Figure 3), consisting of a final dataset of 9942 cells.

For the H9-derived microglia dataset, we identified 8 major cellular populations, distributed in two main groups of cells on the t-SNE plot (Extended Data 6d), both showing a treatment-associated distribution of cells (Extended Data 6e). Clusters 0, 2, 3, 5 expressed homeostatic microglia markers (Extended Data 6f), while clusters 1 and 4 expressed gene markers of CNS-associated macrophages (MRC1, CD163). Cluster 7 expressed low level of macrophage markers and some activation markers (CD74), while cluster 6 was enriched in markers of cycling cells (MKI67). Cluster 8 counted few cells, was very different from all the others and had no clear markers, probably reflecting a small population of doublets. Overall, 72% of cells (6444/8998) in our post-QC dataset were microglia, and only these cells were retained for further analysis. The final analysis was performed on oAβ and scrambled peptide-treated cells (Figure 3), consisting of a final dataset of 4880 cells, after excluding CNS-associated macrophages.

Independent clustering of mouse and H9-derived microglia. Cells passing QC were analysed using functions provided with the Seurat package, version 2.3.4. Data was log normalized and we regressed out the variable of read count. Next, we identified the genes with highest variability and performed PCA on such gene set. We identified the most informative principal components based on a scree plot and we used these to perform cell clustering. Identification of differential expressed genes was performed using the Wilcox test implemented by Seurat’s FindMarker. t-SNE plots were prepared using Seurat’s t-SNE implementation. For the mouse microglia dataset, we considered 1020 highly variable genes for PCA and the first 15 PCs for clustering. The H9-derived human microglia dataset was analysed similarly as described above, by performing PCA on the 1886 most variable genes and by using the first 15 PCs to perform cluster analysis.

Data integration and joint clustering. Cells passing QC were analysed using the Seurat package, version 3.0. The combined object (H9-derived naive microglia and primary microglia from patients) was split into a list, with each dataset as an element. Standard preprocessing (log-normalization) was performed individually for each of the two datasets, and variable features (nfeatures = 2000) that were identified based on a variance stabilizing transformation (selection.method = "vst"). Next, we identified anchors using the FindIntegrationAnchors function, giving the list of Seurat objects as input. We used all default parameters, including the dimensionality of the dataset (dims = 1:30). We passed these anchors to the IntegrateData function, in order to get an integrated (or ‘batch-corrected’) expression matrix for all cells, enabling them to be jointly analysed. We used the new integrated matrix for downstream analysis and visualization using the standard workflow.

Pseudotime analysis. To infer the pseudotime of microglia progression towards phenotypic change in response to oAβ challenge, we used the Monocle 2 package (version 2.6.4)^14,27. We performed an unsupervised identification of cell trajectories and states, based on the top 200 marker genes identified with a differential expression analysis between oAβ treated cells and scramble-treated cells.

Differential Expression. Differential expression was performed using functions provided with the Seurat package; p values were calculated using the Wilcoxon rank-sum test. In Seurat’s function FindAllMarkers, no threshold for the min.pct parameter was applied, in order not to miss marker genes of rare cell populations. All the other parameters were set to default. Genes with adjusted p values (using a Bonferroni correction) < 0.05 were considered significantly differentially expressed. Differential expression was used to find cluster markers in all datasets. For Figure 3, differential expression was performed with the FindMarkers function of Seurat comparing CRM and HM clusters, both in mouse and H9-derived human datasets, with no logFC or min.pct thresholds.

Scores of cell types/states signatures. For Extended Data 3 and 7-9, signatures were calculated using Seurat’s AddModuleScore function using a list of marker genes identified for each cell type or cell state, based on previous experimental data from our lab¹⁹ and recent description of microglial transcriptional modules¹². See Supplementary Table 3 for a complete list of all genes defining the different signatures.

Distribution of samples across clusters. We compared the distribution of samples between different clusters by two different tests. We used two-dimensional contingency table (Pearson’s Chi-squared test) to test the overall distribution of treatments across clusters (null hypothesis assuming that the joint distribution of the cell counts in a 2-dimensional contingency table is the product of the row and column marginals). In addition, we used goodness-of-fit test (Chi-squared test for given probabilities) to test distribution within each cluster (null hypothesis assuming that the observed population probabilities in each cluster equal the expected ones; human microglia: Aβ 45.6%, Scr 54.4%; mouse microglia: Aβ 43.8%, Scr 56.2%).

Pathway enrichment analysis. Pathway analysis was performed with GOrilla (Gene Ontology enRIchment anaLysis and visuaLizAtion tool)²⁸, with single ranked list of genes as running mode. For both mouse and H9-derived microglia, genes were ranked by p-value adjusted taken from the Differential Expression analysis performed between CRM and HM clusters. The enriched ontology terms were then grouped by major functional categories, and the most significant terms (after multiple correction by FDR) in the H9-microglia dataset were compared to the same terms in the mouse host microglia dataset (Extended Data 10a). Each gene that was found significant in Differential Expression was then annotated with the functional categories it belongs to (Extended Data 10b and c), considering only the terms that were found significantly enriched in the GOrilla analysis.

Extended Data

Extended Data 1. Gating strategy for the isolation of H9-microglia from the mouse brain and graft efficiency.

(a) Human cells were sorted according to the expression of CD11b, hCD45, and GFP, whereas mouse cells only expressed CD11b but were negative for hCD45 and GFP. (b) H9-microglia graft efficiency. Percentage of CD11b cells in the total sample, and proportion of human cells amongst them. Graph shows mean±SEM, n=6 mice per group.

Extended Data 2. H9-microglia showed a widespread distribution across multiple areas of the brain.

(a) Representative overview of the extent of H9-microglia graft in the mouse brain. Human microglia are stained for P2RY12 across consecutive sections separated by 500μm to capture multiple anatomical areas. Scale bar, 1mm. (b) Higher magnification images of multiple anatomical areas including meninges, cortex, striatum, white matter, choroid plexus and hippocampus. Labelling shows DAPI (in blue), GFP (in green) and P2RY12 (in cyan). Images are representative of a staining performed in n=4 mice. Scale bar, 100 μm.

Extended Data 3. Extended clustering and distribution of in vitro, in vivo (engrafted) H9 and primary microglia.

(a) PCA shows clear separation between in vitro (MNC and MG) and in vivo (engrafted H9 and primary) microglia. The colours correspond to the clustering shown in Figure 2a. (b) t-SNE plots as in Figure 2a, coloured by the combined level of expression of groups of genes that characterise distinct microglial states. The original clusters from Figure 2a are outlined. (c) Selected genes defining the different transcriptomic scores shown in b. The full list of genes is shown in Supplementary Table 3. (d) Distribution of the different samples across the tSNE plot, and (e) percentage of each sample across the different clusters. All the data shown represents 2,246 transplanted H9-microglia (in vivo) (n=3/1, 3 mice in 1 combined sequencing pool), 4496 H9-derived monocytes (n=2/1, 2 differentiations in 1 combined sequencing pool) and 3385 microglia in vitro (n=2/1), and 22,846 human primary microglia obtained from cortical surgical resections (n=7/7 online Methods).

Extended Data 4. Direct comparison of in vitro, in vivo (engrafted) H9 and primary microglia.

(a) Volcano plots showing paired comparisons between average gene expression in vitro MNC, in vitro MG, in vivo (engrafted) MG and primary cells. (b) Individual comparisons of in vivo (engrafted) H9-microglia and each human subject (human cases 1-7). The dashed line corresponds to an arbitrary threshold logFC of 0.2. Blue labels correspond to homeostatic genes whereas red labels correspond to microglial activation genes (Supplementary Table 3). All the data shown represents 2,246 transplanted H9-microglia (in vivo) (n=3/1, 3 mice in 1 combined sequencing pool), 4496 H9-derived monocytes (n=2/1, 2 differentiations in 1 combined sequencing pool) and 3385 microglia in vitro (n=2/1), and 22,846 human primary microglia obtained from cortical surgical resections (n=7/7). In all cases, Wilcoxon Rank Sum test, p-values adjusted with Bonferroni correction based on the total number of genes in the dataset.

Extended Data 5. Characterization of oAβ preparation.

Freshly eluted recombinant Aβ1-42 monomers follow a rapid aggregation course in Tris-EDTA buffer. (a) After 2 hours of incubation, Aβ1-42 monomers oligomerize and run as dimers and trimers (indicated with *) on SDS-PAGE/Coomassie staining, and they are proteinase-K sensitive. (b) Early Aβ1-42 oligomers form A11 and OC-positive aggregates. Two µl of either scrambled or amyloid beta 1-42 from different time points (0 hours, 2 hours and 2 weeks) of incubation was spotted on blots. These dot blots were probed with A11 antibody (Invitrogen; #AHB0052), which recognizes amino acid sequence-independent oligomers of proteins or peptides. A11 epitope is transient and is present only in the early oligomers (2 h), in contrast to the mature fibers (2 w) formed after 2 weeks of incubation. No fibrillary material is detected after 2 hours of incubation. (c) OC antibody (Millipore; #AB2286) recognize epitopes common to monomers, amyloid oligomers, and fibrils. (d) 4G8 antibody (Eurogentec; #SIG-39220) detects N-terminal of the amyloid aggregates (epitope between amino acids 17-24).

Extended Data 6. Preparation of the datasets for the analysis of (a-c) host mouse and (d-f) H9-microglia response to oAβ (see Figure 3).

(a) t-SNE plot of the 13342 cells passing quality control, coloured by clusters. (b) t-SNE plots as in a, coloured by the level of ln normalized expression of selected genes for microglia (Cx3cr1, Tmem119), monocytes (Ccr2) and neutrophils (Ccrl2). (c) Violin plots of selected marker genes for homeostatic microglia (Cx3cr1, Tmem119), CRM (Il1b), ARM (Cd74, H2-Eb1, Ifit3), neutrophils (Ccrl2), monocytes (Ccr2, Ly6c1), astrocytes (Clu), oligodendrocytes (Mbp), neurons (Npy), and cycling cells (Top2a). Analysis shown in Figure 2 was performed after removal of clusters 4 (neutrophils), 7 (monocytes), 10 (astrocytes), 12 (oligodendrocytes) and 9 (neurons). (d) t-SNE plot of the 6444 H9-microglia cells passing quality control, coloured by clusters. (e) t-SNE plot as in a, coloured by treatment (naïve; scrambled peptide, Scr; and oligomeric Aβ, oAβ). (f) t-SNE plot as in a, coloured by the level of ln normalized expression of selected genes for microglia (CX3CR1), cycling cells (MKI67) and brain resident macrophages (MRC1, CD163). Analysis shown in Figure 2 was performed after removal of clusters 1 and 4 (brain resident macrophages), 6 (cycling cells) and 8 (doublets).

Extended Data 7. Expanded analysis, clustering and trajectory inference of the analysis of the response of H9-microglia upon oAβ.

(a) PCA of 4880 H9-microglia isolated from the mouse brain (n=3 mice in 1 combined sequencing pool) shows clear separation of the different clusters identified in our analysis in PC1 and PC2. (b) t-SNE plots as in Figure 3a, coloured by the combined level of expression of groups of genes that characterise distinct microglial states. (c) Selected genes defining the different transcriptomic scores shown in Figure 3b: homeostatic score (1), cytokine score (2) activated score (3). The full list of genes is shown in Supplementary Table 3. (d) Volcano plots showing paired comparisons between H9.HM, H9.CRM and H9.PM clusters ((Wilcoxon Rank Sum test, p-values adjusted with Bonferroni correction based on the total number of genes in the dataset). (e) Proportion of the different experimental groups across clusters in Figure 3a (Chi² test, *** p < 10^-250). (f, g) Phenotypic trajectory inferred by Monocle 2 as shown in Figure 3a, coloured by (f) treatment and (g) clusters from Figure 3a.

Extended Data 8. Expanded analysis, clustering and trajectory inference of the analysis of the response of mouse host microglia upon oAβ.

(a) PCA of 9942 endogenous mouse cells (n=3 mice in 1 combined sequencing pool) shows clear separation of the different clusters identified in our analysis in PC1 and PC2. (b) t-SNE plots as in Figure 3d, coloured by the combined level of expression of groups of genes that characterise distinct microglial states. (c) Selected genes defining the different transcriptomic scores shown in b: homeostatic score (1), cytokine score (2) activated score (3). The full list of genes is shown in Supplementary Table 3. (d) Volcano plots showing paired comparisons between ms.HM, ms.CRM and ms.ARM clusters (Wilcoxon Rank Sum test, p-values adjusted with Bonferroni correction based on the total number of genes in the dataset). (e) Proportion of the different experimental groups across clusters in Figure 3d (Chi² test, *** p < 10^-250). (f, g) Phenotypic trajectory inferred by Monocle 2 as shown in Figure 3c, coloured by (f) treatment and (g) clusters from Figure 3d.

Extended Data 9. Cytokine response microglia (CRM) are also present in APP^NL-G-F mice.

(a) Original clustering analysis from Sala Frigerio et al. (2019)¹¹ consisting of 10,801 microglial cells from 3 to 21 months old APP^NL-G-F mice and aged matched wild type controls. (b) Clusters shown in a, coloured with CRM, HM, ARM and IRM transcriptomic signatures. Note the small population of cells displaying CRM features embeded into the ARM response in APP^NL-G-F microglia. (c) Significant enrichment of either homeostatic (HM) or activated (ARM) microglia gene sets from Sala Frigerio et al. (2019)¹¹ in our ms.HM and ms.ARM clusters, respectively (ANOVA with Turkey HSD multiple comparisons correction, *** p≈0; box plots represent median, with 25^th and 75^th percentiles and 1.5 times the inter-quartile range as minima and maxima). (d) Subselection of CRM cells from the main clusters shown in a. (e) Microglia cells enriched with a CRM transcriptomic profile are largely located at early stages of the response to amyloid in APP^NL-G-F mice¹¹. The left panel shows the trajectory analysis coloured by clusters as represented in panel a, whereas the right panel highlights the cells displaying a CRM profile.

Extended Data 10. Differential responses of human and host mouse microglia to oligomeric Aβ.

(a) Pathway enrichment analysis (GOrilla) shows that the differentially expressed genes in CRM vs. HM clusters are involved in immune and inflammatory processes. (b) Top differentially expressed genes in H9-microglia upon Aβ challenge relative to scrambled peptide, and expression of their mouse orthologs by endogenous mouse cells. Coloured marks indicate the functional category as shown in b. (c) Differentially expressed genes that show opposite behaviour in H9-and mouse host (Rag2^-/- Il2rγ^-/-) microglia. Coloured marks indicate the functional category as shown in b. (d) Volcano plots showing paired comparisons between H9.HM, H9.CRM, but including all genes (even those with no clear orthology to mouse, Wilcoxon Rank Sum test, p-values adjusted with Bonferroni correction based on the total number of genes in the dataset). (e) Further pathway enrichment analysis (GOrilla) performed on the human-specific (with no clear orthology) differentially expressed genes in H9.CRM vs. H9.HM clusters are involved in cytokine/chemokine responses.

Data availability

The data generated in this study is available from GEO (identifier to be provided). The data and code are also available at scope.bds.org. Data from Karen-Shaul et al.¹⁵ is available from GEO (identifier GSE98969). Data from Sala Frigerio et al. ¹¹ is available from GEO (identifier GSE127893).