Alzheimer’s disease modification mediated by bone marrow-derived macrophages via a TREM2-independent pathway in mouse model of amyloidosis

Microglia and monocyte-derived macrophages (MDM) are key players in dealing with Alzheimer’s disease. In amyloidosis mouse models, activation of microglia was found to be TREM2 dependent. Here, using Trem2−/−5xFAD mice, we assessed whether MDM act via a TREM2-dependent pathway. We adopted a treatment protocol targeting the programmed cell death ligand-1 (PD-L1) immune checkpoint, previously shown to modify Alzheimer’s disease via MDM involvement. Blockade of PD-L1 in Trem2−/−5xFAD mice resulted in cognitive improvement and reduced levels of water-soluble amyloid beta1–42 with no effect on amyloid plaque burden. Single-cell RNA sequencing revealed that MDM, derived from both Trem2−/− and Trem2+/+5xFAD mouse brains, express a unique set of genes encoding scavenger receptors (for example, Mrc1, Msr1). Blockade of monocyte trafficking using anti-CCR2 antibody completely abrogated the cognitive improvement induced by anti-PD-L1 treatment in Trem2−/−5xFAD mice and similarly, but to a lesser extent, in Trem2+/+5xFAD mice. These results highlight a TREM2-independent, disease-modifying activity of MDM in an amyloidosis mouse model. Using a treatment that activates the peripheral immune system in an animal model of amyloidosis, the authors show that monocyte-derived macrophages can modify Alzheimer’s disease progression via a TREM2-independent mechanism.

Trem2-independent beneficial effect of anti-PD-L1 in 5xFAD mice. The above results encouraged us to use anti-PD-L1 treatment in Trem2 −/− 5xFAD mice, which lack DAM 20 , to gain an insight into the differential role of DAM and MDM in disease modification. To this end, we treated 6-9-month-old Trem2 −/− 5xFAD mice with either anti-PD-L1 or IgG2b control antibody and tested their behavior 27-30 days after treatment; each treated group included both males and females equally distributed between the two treatments. As a positive control for response to anti-PD-L1 antibody, we also treated Trem2 +/+ 5xFAD mice. Mice were tested in the tasks radial arm water maze (RAWM) and novel object recognition (NOR), which measure spatial learning and working memory, respectively ( Fig. 2a and Extended Data Fig. 3a,b). We used wild-type (WT) mice from both Trem2 +/+ and Trem2 −/− genotypes as controls for normal behavior. Because there were no differences in the WT groups (Extended Data Fig. 3c), the results from these two WT control groups were pooled in all behavioral data. Impaired cognitive performance was found in both Trem2 −/− 5xFAD and Trem2 +/+ 5xFAD control mice (treated with IgG2b) with no differences between the two groups (Fig. 2b,c), consistent with previous reports 48 . Following anti-PD-L1 treatment, a significant improvement in cognitive ability, assessed by both RAWM and NOR, was observed in Trem2 −/− 5xFAD mice (Fig. 2b,c) and, as expected 19 , in Trem2 +/+ 5xFAD mice (Fig. 2b,c). While in the NOR task cognitive improvement was similar in Trem2 +/+ 5xFAD and Trem2 −/− 5xFAD mice (Fig. 2b), in RAWM the effect of treatment in the Trem2 +/+ 5xFAD group was slightly but significantly higher than in the Trem2 −/− 5xFAD group (Fig. 2c). These results suggest a Trem2-independent effect of peripheral immune activation on cognitive performance. They further suggest an additional contribution of a Trem2-dependent mechanism, possibly related to the observed elevated levels of DAM shown in Fig. 1c.
It was previously reported that TREM2 is required for the clearance of Aβ plaques 17,48,49 , while no differences in the levels of 'soluble' [Tris-buffered saline (TBS)-soluble] Aβ were found between Trem2 −/− 5xFAD and Trem2 +/+ 5xFAD mice 17,48 . Those reported results, together with our observations, prompted us to test whether improved cognitive ability following anti-PD-L1 treatment could be associated with a reduction in the levels of TBS-soluble Aβ. Thus, after cognitive testing, hippocampi (one hemisphere from each mouse brain) were excised and levels of TBS-soluble Aβ 1-42 measured by ELISA. We found a reduction in the level of TBS-soluble Aβ  following anti-PD-L1 treatment in Trem2 −/− 5xFAD mice (Fig. 3a); in Trem2 +/+ 5xFAD mice, anti-PD-L1 led to a similar reduction in TBS-soluble Aβ   (Fig. 3b). When we analyzed 5xFAD mice of both genotypes (Trem2 −/− and Trem2 +/+ ) treated with either anti-PD-L1 or IgG2b control, as well as WT mice, correlations were found between the levels of hippocampal TBS-soluble Aβ 1-42 and cognitive performance measured in both RAWM (average time to platform of the final three trials, Pearson's r = 0.6919, P < 0.0001; Fig. 3c) and NOR (percentage of novel object exploration time, Pearson's r = −0.3410, P = 0.0181; Fig. 3d). These correlations suggest a potential functional link between reduction in TBS-soluble Aβ 1-42 and cognitive performance, but do not imply that this is the sole mechanism 19,36 . Of note, we observed a clear effect of treatment on TBS-soluble Aβ 1-42 level only in animals showing cognitive impairment, which is known to be manifested earlier in females than in males [50][51][52][53] .
We also prepared non-ionic Triton X-100 extracts of Aβ. According to the manufacturer of the ELISA kit used for these assays (Methods), the Aβ extracted by this detergent and detected by the kit can reach up to 80% of total mouse brain Aβ for the age of mice used here. Using such extracts from the hippocampus of both Trem2 +/+ 5xFAD and Trem2 −/− 5xFAD mice, a reduction in Triton X-100-soluble Aβ 1-42 in Trem2 +/+ 5xFAD mice treated with a-PD-L1 was revealed. Such a reduction was not observed in Trem2 −/− 5xFAD mice treated with anti-PD-L1 (Fig. 3e,f). In addition, no reduction in plaque burden was seen in anti-PD-L1-treated 8-9-month-old ; CD45 + cells were collected for MARS-seq from brains of 6-7-month-old Trem2 +/+ 5xFAD mice 14 days following injection of anti-PD-L1 or IgG2b control. b, 2D projection of 2,931 microglia collected from anti-PD-L1-and IgG2btreated mice, arranged according to transcriptomic similarity and annotated and colored according to microglia/DAM signature of each metacell; n = 3 per group. c, Percentage of DAM (mean ± s.e.m.) of total microglia in mice treated with anti-PD-L1 and IgG2b (n = 6 per group; four females (F), two males (M)). Data derived from two cohorts of mice, pooled after normalization per cohort and further analyzed using a one-tailed unpaired t-test (t (10) = 2.89, P = 0.0081; 95% confidence interval (CI), 0.06497-0.50240). d, Multiple bar graphs showing expression level (average UMI per cell) of selected marker genes across 31 microglia metacells isolated from the brains of both anti-PD-L1-and IgG2b-treated mice. Metacells are arranged from left to right based on activation level and colored accordingly (red, DAM; orange, stage I DAM; yellow, homeostatic microglia); n = 3 per group. e, Schematic presentation of the experimental design used to test the origin of DAM (relates to f,g). Eight-month-old BM-chimeric mice (Trem2 +/+ 5XFAD transplanted with Ub-GFP WT BM) were injected with either anti-PD-L1 or IgG2b and, 14 days later, CD45 + and CD45 + GFP + cells were collected for MARS-seq. f,g, 2D projection of 3,531 cells from both CD45 + and CD45 + GFP + plates, clustered according to transcriptomic similarity; color coding represents cell type annotation in f and the source of each cell, either CD45 + or CD45 + GFP + in g. Data for anti-PD-L1-(n = 3 M) and IgG2b-(n = 2 M) treated AD mice were pooled. **P < 0.01.
Trem2 −/− 5xFAD mice (Fig. 3g), unlike the reduction in Aβ plaque load previously observed in Trem2 +/+ 5xFAD mice 18,19 . These results suggest that anti-PD-L1 treatment is affecting in a Trem2independent manner only the hydrophilic form of Aβ  , which includes the TBS-soluble form, and not the membrane-associated forms of Aβ, nor Aβ plaques. Notably, no change in TBS-or Triton X-100-soluble Aβ 1-40 was observed following treatment (Extended Data Fig. 3d,e). Because the TBS-soluble fraction contains Aβ 1-42 oligomers, which were shown to induce synaptotoxicity 4,8,54 , we further tested, in Trem2 −/− 5xFAD mice, whether the treatment would have an effect on synaptic markers. To this end, we used brain sections from 8-9-month-old Trem2 −/− 5xFAD mice and evaluated levels of the presynaptic marker synaptophysin and postsynaptic marker Homer1 in the hippocampal dentate gyrus (DG) and CA3 following anti-PD-L1 treatment. Higher levels of synaptophysin and Homer1 immunoreactivity were measured in the DG and CA3 of mice treated with anti-PD-L1 compared to the IgG2b-treated group (Fig. 3h,i). Overall, the above results suggest that modification of AD pathology by systemic immune activation could be achieved through a Trem2-independent mechanism, and is correlated with a reduction in TBS-soluble Aβ 1-42 and accompanied by higher expression of pre-and postsynaptic markers.
Transcriptomic signature of MDM. The effects of treatment on cognition and TBS soluble Aβ 1-42 in the absence of Trem2 and, accordingly, in the absence of DAM, led us next to analyze the transcriptomic signature of MDM using MARS-seq. To enrich for MDM, which are rare in the brain compared to resident microglia, we sorted CD45 + CD11b + Tmem119 neg/low cells (Extended Data Fig. 4a) from 7-9-month-old Trem2 +/+ 5xFAD and Trem2 −/− 5xFAD brains, 14 days after treatment with either anti-PD-L1 or IgG2b control antibody. Metacell analysis of 1,181 nonmicroglial myeloid cells from both Trem2 +/+ 5xFAD and Trem2 −/− 5xFAD brains resulted in 12 metacells, which were clustered into three subgroups of monocytes and three MDM subsets ( Fig. 4a and Extended Data Fig. 4b). An equal number of microglia from the same pool of mice was sampled to enable qualitative comparison ( Fig. 4a; Methods). Cell type composition analysis revealed, as expected, the absence of DAM in Trem2 −/− 5xFAD mouse brains, with no differences in the MDM and monocyte subsets between Trem2 +/+ 5xFAD and Trem2 −/− 5xFAD brains (Fig. 4b,c). All three MDM subsets shared a similar transcriptomic profile, with different expression levels of several functional genes, which reflects a gradual change in the state of differentiation or activation of infiltrating monocytes. Among the three subsets, MDM-1 were found to express a unique repertoire of genes, including those  encoding for scavenger receptors such as Stabilin 1 (Stab1), Cd163, Mannose receptor C-type 1 (Mrc1) and Macrophage scavenger receptor 1 (Msr1) (Fig. 4d), which could contribute to both clearance of neurotoxic misfolded proteins from the brain 43 and reduction in local inflammation 55 . MDM-3 were found to express high levels of the C-C chemokine receptor type 2 (Ccr2), suggesting that they are probably the cells most recently migrated, while the MDM-2 profile represented an intermediate stage (Fig. 4d) between MDM-3 and MDM-1.
Because the MDM-1 subset included the most highly activated cells, we compared their signature with that of DAM. Analysis of differential gene expression between the two cell groups revealed their nonredundant profiles (Extended Data Fig. 4c). The majority of key functional genes expressed by MDM-1 were not expressed by DAM (Fig. 4e), except for Apolipoprotein E (Apoe), which was highly expressed by both and is a known AD 'risk factor' gene required for the removal of Aβ 56 . No differences in transcriptomic profile were observed between MDM-1 derived from Trem2 −/− 5xFAD and Trem2 +/+ 5xFAD mice (Extended Data Fig. 4d). Additionally, no profile differences were found between the IgG2b and anti-PD-L1 treatment conditions in this cell population (Extended Data Fig. 4e). Overall, MDM express a unique set of genes distinct from that of DAM, and this gene program is Trem2 independent.

Anti-CCR2 abrogates the beneficial effects of PD-L1 blockade.
Since the signature of MDM in Trem2 −/− 5xFAD mice was similar to that found in Trem2 +/+ 5xFAD mice, we could further use Trem2 −/− 5xFAD mice to evaluate the contribution of MDM to disease modification in the absence of Trem2-dependent mechanisms in general, and DAM, in particular. We hypothesized that anti-PD-L1 treatment, primarily through its effect in the periphery, facilitates MDM homing to the brain which, in turn, leads to disease modification. We first verified target engagement in the periphery after a single injection of anti-PD-L1 antibody relative to IgG2b isotype control, by T-cell receptor occupancy in blood and spleen, as a function of anti-PD-L1 antibody-administered dose (Extended Data Fig. 5a-d), and also demonstrated an increased level of PD-1 + effector memory T (T EM ) cells in blood (Extended Data Fig. 5e,f).
To block monocytes we used an anti-CCR2 antibody that we and others previously calibrated for optimal reduction of monocytes in the circulation, and consequently in the CNS 31,57 . We treated 6-9-month-old Trem2 −/− 5xFAD mice with anti-PD-L1 antibody, along with anti-CCR2 antibody. The anti-CCR2 antibody was injected to Trem2 −/− 5xFAD mice, starting from 3 days before anti-PD-L1 treatment and continuing every 3 days until day 9 post treatment, to ensure monocyte inhibition throughout the entire critical period of their homing to the brain; 31,32 a second group was treated with only anti-PD-L1 and a third received control IgG2b. Of note, both anti-PD-L1 and anti-CCR2 share the same IgG2b isotype. Cognitive performance was assessed by RAWM and NOR 27-30 days after anti-PD-L1 treatment, and subsequently hippocampal tissues from both hemispheres were excised for Aβ 1-42 measurement by ELISA (Fig. 5a). In Trem2 −/− 5xFAD mice, anti-CCR2 antibody led to complete abrogation of the beneficial effect of anti-PD-L1 on cognitive performance (Fig. 5b,c) with no effect on either locomotion or anxiety (Extended Data Fig. 6a,b). The level of TBSsoluble Aβ 1-42 was lower by approximately 40% in mice treated with anti-PD-L1 alone compared to the group treated with anti-PD-L1 and anti-CCR2 (Fig. 5d), suggesting that this effect of anti-PD-L1 was dampened by CCR2 blockade in Trem2 −/− 5xFAD mice. The levels of Triton X-100-soluble Aβ 1-42 were similar in Trem2 −/− 5xFAD mice treated with anti-PD-L1 alone and in those treated with anti-PD-L1 combined with anti-CCR2 (Extended Data Fig. 6c). Of note, a significant positive correlation (Pearson's r = 0.5065, P = 0.03) was found between RAWM performance (average time to platform of the final three trials) and the levels of hippocampal TBS-soluble Aβ 1-42 in Trem2 −/− 5xFAD mice treated with anti-PD-L1 + anti-CCR2 or anti-PD-L1 alone (Fig. 5e). Although these results do not necessarily imply that reduction in TBS-soluble Aβ 1-42 directly underlies the beneficial effect of anti-PD-L1 on cognitive performance, they suggest that it is a critical factor affected by the treatment. To verify that CCR2 blockade resulted in reduced levels of MDM in the brain, we carried out an additional experiment in which we quantified, by flow cytometry, the levels of MDM in 7-8-monthold Trem2 −/− 5xFAD mice treated with either IgG2b, anti-PD-L1 or anti-PD-L1 in combination with anti-CCR2. On day 7 following anti-PD-L1 injection, mouse brains were excised and analyzed for levels of infiltrating MDM expressing CD45 + CD11b + CD44 + with negative selection for neutrophils and border-associated macrophages [(BAM) 58 ], Ly6G − and CD38 − , respectively] (Extended Data Fig. 6d). We observed a significant increase in the number of MDM in brain following anti-PD-L1 treatment, which was lost following CCR2 blockade (Extended Data Fig. 6e). To gain insight to the potential additional contribution of DAM and other Trem2-dependent pathways, beyond that of MDM, we also evaluated cognitive performance following CCR2 blockade in mice expressing normal levels of Trem2. Thus, we treated Trem2 +/+ 5xFAD mice with either anti-PD-L1 combined with anti-CCR2, anti-PD-L1 alone or IgG2b and tested their performance in RAWM and NOR (Fig. 6a). In Trem2 +/+ 5xFAD mice, blockade of CCR2 completely abrogated the beneficial effect of anti-PD-L1 treatment on NOR ( Fig. 6b) but only partially on RAWM (Fig. 6c). The partial abrogation resulting from inhibition of CCR2 detected in RAWM could be attributed to additional mechanisms stimulated by anti-PD-L1 treatment, including the observed increased level of DAM induced by treatment in Trem2 +/+ 5xFAD mice (Fig. 1c), or to other Trem2dependent mechanisms. The fact that in NOR, the beneficial effect T r e m of anti-PD-L1 was completely abrogated by anti-CCR2 antibody could be reflection of the different sensitivity or functions assessed by the two mazes. Overall, we show here that MDM are key players in cognitive improvement following anti-PD-L1 immunotherapy, in both Trem2 +/+ 5xFAD and Trem2 −/− 5xFAD mice.

Discussion
In the present study, we found that MDM play a Trem2-independent role in the cognitive improvement observed in both Trem2 +/+ 5xFAD and Trem2 −/− 5xFAD mice following anti-PD-L1 antibody treatment. The effect on behavior was accompanied by a reduction  The role of MDM in AD modification was demonstrated when Trem2 −/− 5xFAD mice were treated with anti-PD-L1 antibody along with CCR2-blocking antibody. Under these conditions, no beneficial effect of treatment was found on working and spatial memory, nor on the levels of TBS-soluble Aβ 1-42 . In Trem2 +/+ 5xFAD mice, the beneficial effect of anti-PD-L1 treatment was slightly but significantly stronger than in Trem2 −/− 5xFAD mice based on assessment by RAWM, but not by NOR. In addition, the effect of anti-PD-L1 treatment was completely lost following blockade of CCR2 inTrem2 +/+ 5xFAD mice when measured in the NOR task, but only partially in the RAWM task. The differences in observed results between RAWM and NOR in Trem2 +/+ highlight a potential contribution of the Trem2-dependent pathway to the cognitive skills measured by the RAWM task, but apparently not to those measured by the NOR task. Thus, it is possible that the activity of DAM, the level of which was elevated following anti-PD-L1 treatment in Trem2 +/+ 5xFAD mice, contributed more to spatial memory than to working memory via, at least in part, reduction of Aβ plaques 17,20,48 .
Importantly, the MDM-dependent effect in both Trem2 +/+ 5xFAD and Trem2 −/− 5xFAD mice, beyond its impact on TBS-soluble Aβ 1-42 ,  could be attributed to effects on additional disease-escalating factors, including reduction of inflammatory cytokines, as we have previously reported 19,36 . Although MDM-1 share some of their signature with macrophages previously observed in WT mouse brain 59 , they express additional functional genes such as the scavenger receptor Msr1, which was shown to be required for clearance of soluble Aβ 1-42 oligomers 43 . MSR1-expressing macrophages were recently shown to be important in the resolution of sterile inflammation in the brain following stroke, through damage-associated molecular patterns (DAMPs) 55 . Moreover, other studies have demonstrated that MDM reduce inflammation within the brain in mouse models of amyloidosis 37,42 . This is consistent with our MARS-seq data, which revealed that MDM found in the brains of both Trem2 −/− and Trem2 +/+ 5xFAD mice expressed a transcriptomic signature of genes associated with macrophage anti-inflammatory activity (for example, Mrc1 60-62 ). It is possible that newly recruited MDM, by affecting the parenchymal milieu, could also facilitate clearance of Aβ 1-42 , which occurs through a Trem2-dependent pathway.
Following anti-PD-L1 treatment in a tauopathy mouse model, a macrophage signature similar to that of MDM-1, including the expression of Msr1, Mrc1 and Cd163, was found 19 . Those results further support the potential ability of these cells to remove toxic forms of misfolded proteins via recognition of molecular patterns 36 , which could explain why the same treatment, targeting the PD-1/PD-L1 pathway, is effective in different mouse models of dementia 18,19,37 . Additionally, MDM signature in the present study was not affected by treatment and was found to be similar in Trem2 +/+ 5xFAD and Trem2 −/− 5xFAD mice, which suggests that enhancement of MDM levels helps to limit brain pathology regardless of disease etiology. This contention is in line with evidence that, although spontaneous homing of MDM to the brain is limited 30 , they are needed for coping with chronic neurodegeneration in several brain pathologies [57][58][59][60][61][62] . Of note, MDM could also contribute to the clearance of toxic molecules via expression of additional genes, including Apoe. Importantly, in the present study MDM were sorted from the entire brain. Therefore, although MDM were previously observed in brain parenchyma 19 , we cannot rule out the possibility that they can also execute pathology-mitigating activity from the brain's borders 58,59,63 .
Single-cell RNA-seq data revealed an increased level of DAM following anti-PD-L1 treatment that was not accompanied by any molecular changes in their transcriptomic signature. This supports the contention that the DAM elevation is a result of a shift from resting microglia via a Trem2-dependent pathway 20  through cell proliferation. In addition, it is important to note that our experiment using BM-chimeric Trem2 +/+ 5xFAD mice, which enabled us to distinguish between BM-derived cells (unless originating from the skull) and resident microglia, ruled out the possibility that the observed elevation in DAM might be an outcome of infiltration of monocytes that acquired a DAM-like signature. This result is in line with previous studies demonstrating that plaque-associated myeloid cells derive from resident microglia 20,64,65 . Interestingly, the fact that the same treatment was found to be effective in both Trem2 −/− 5xFAD and Trem2 +/+ 5xFAD mice might also be in line with the recent observation that Trem2-expressing microglia could act as a double-edged sword, being protective at early stages of the disease but not at the stage when cognitive impairment is manifested 66,67 .
In the present study, anti-PD-L1 treatment in Trem2 −/− 5xFAD mice affected only the TBS-soluble Aβ 1-42 form but neither membrane-associated Aβ (Triton X-100 extract) nor Aβ plaques, in line with previous findings showing that Trem2, and specifically Trem2expressing microglia, are needed for plaque removal 17,48,49 . Of note, Triton X-100 extraction mainly contains membrane-associated Aβ and thus its use in the measurement of total Aβ accumulation is reduced along disease progression. Accordingly, the Triton X-100soluble fraction could be used for comparison between mice of the same genotype and age, with and without treatment but not across ages 68 .
The beneficial effect of blockade of the PD-1/PD-L1 pathway on cognitive deficit was detected in different mouse models, of both amyloidosis and tauopathy, at different stages of disease progression 19,36,37 . These effects were found to be dependent on MDM expressing Msr1, recognizing DAMPs, such as soluble Aβ 1-42 43 . Since a correlation was found here between cognitive performance and TBS-soluble Aβ 1-42 , the latter could be an informative readout of the effect of blockade of the PD-1/PD-L1 pathway. Accordingly, reduction of plaques as a single readout 69 , without measurement of cognition, could not be informative for assessment of therapeutic capacity of a treatment that does not directly target the plaques; this is even more relevant given the differences in disease progression among various transgenic mouse models, and between males and females [50][51][52][53] . In light of the above, one can explain the apparent failure of anti-PD-1 treatment reported by Latta-Mahieu and colleagues, when using plaque burden as sole readout without assessing cognitive performance 69 . Moreover, the latter work can not be considered as a lack of replication of previous publications, as it presents results using three different mouse models of AD, without including 5xFAD, that was originally used 18,19 . In addition, in two of the three experiments, the antibodies were not identical to the ones previously used, and the doses were lower. In contrast, in the work published by Xing and colleagues, which did measure cognitive performance along with plaque burden, the effect of anti-PD-1 was replicated 37 .
Notably, the TBS-soluble fraction of Aβ measured by ELISA contains a wide range of Aβ assemblies, ranging from monomers essential for synaptic activity and up to dodecamers, which are toxic to synapses 3,4,67 ; the relative reduction of each of these Aβ species cannot be determined by this methodology. Importantly, however, as stated above, disease modification by treatment could involve reduction in additional disease-escalating factors, including inflammation and neuronal loss 19,36 .
Our results are consistent with the most recent consensus regarding the amyloid cascade hypothesis, suggesting that the soluble aggregates of Aβ 1-42 are a major species responsible for neuronal dysfunction, collectively termed oligomeropathy 68 . Our results are also consistent with accumulated evidence from clinical trials highlighting the lack of correlation between amyloid plaque distribution and disease severity, and argue against a direct pathogenic role of Aβ plaques in AD 70 . The stability of Aβ oligomers and their dynam-ics remain unclear, but their overall formation might be closely related to the low-grade inflammation seen during early disease development 71,72 .
Altogether, our findings support the approach of empowering the immune system to facilitate MDM mobilization as a shared mechanism of repair for treatment of AD patients, regardless of whether these individuals carry mutations in the TREM2 gene or in a TREM2-related pathway. They further suggest that targeting the peripheral immune system could potentially overcome TREM2 polymorphism in patients.

Methods
Mice. Two mouse models were used in this study: (1) heterozygous 5xFAD transgenic mice (on a C57/BL6-SJL background), which express familial AD mutant forms of human APP (the Swedish mutation, K670N/M671L; the Florida mutation, I716V; and the London mutation, V717I) and PS1 (M146L/L286V) transgenes under transcriptional control of the neuron-specific mouse Thy-1 promoter 28 (5xFAD line Tg6799; The Jackson Laboratory); and (2) Trem2 −/− 5xFAD mice. Trem2 −/− 5xFAD and Trem2 +/+ 5xFAD mice on a C57/BL6 background were obtained from the laboratory of M. Colonna (Washington University), where they were generated as previously described 73 . For BM transplantation assays, donor cells were isolated from C57/BL6 CD45.2 Ub-GFP mice in which GFP is ubiquitously expressed 47 . All mice were bred and maintained at the animal breeding center of the Weizmann Institute of Science. For the period of cognitive assessment, mice were kept on a reversed light/dark cycle. All experiments described complied with the regulations formulated by the Institutional Animal Care and Use Committee of the Weizmann Institute of Science (application no. 18751119-1). In all experiments, mice were anesthetized and transcardially perfused with PBS before tissue dissection.
Preparation of BM chimeras. Chimeras were prepared by subjecting Trem2 +/+ 5xFAD recipient mice to lethal irradiation (950 rad), directing the beam to irradiate the entire body except for the head, as previously described 31 . Mice received antibiotic supplement in their drinking water (1 ml of 10% ofloxacin for every 200 ml of water) for 1 week, starting immediately after irradiation. The following day, BM cells were isolated from the tibia and femur of Ub-GFP mice and filtered through a 70-µm cell strainer. Each recipient mouse was then reconstituted with 5 × 10 6 BM intravenously injected cells of gender-matched donors. Recipient mice were analyzed 5-8 weeks after BM transplantation to determine the extent of chimerism, and were further monitored to ensure they maintained a chimerism level of at least 30%.
Flow cytometry analysis. After perfusion with PBS, brains were excised excluding the brain stem and manually chopped (0.5-1.0 mm 2 in size), before softwarecontrolled dissociation by gentle magnetic-activated cell separation (MACS) in PBS. For density gradient separation, the pellet was resuspended with 40% Percoll and centrifuged at 750g for 20 min at 4 °C; the supernatant was then discarded. Spleens were mashed with the plunger of a syringe. Both spleen and blood samples were treated with ammonium chloride potassium-lysing buffer to remove erythrocytes. Cells from all samples were suspended in ice-cold sorting buffer (PBS supplemented with 2 mM EDTA and 2% fetal calf serum (FCS)) supplemented with anti-mouse CD16/32 (1:100, no. 101302, BioLegend) to block Fc receptors before labeling with fluorescent antibodies against cell-surface epitopes. Samples were stained using the following antibodies: APC-conjugated CD3e Mice were euthanized on day 7 after injection. Blood was collected by retro-orbital bleeding into Eppendorf tubes containing heparin. Spleens were dissected and mashed using the back of a syringe against a metal mesh in PBS. Blood and spleen samples were divided into two tubes: tube 1 ('saturated' tube) and tube 2 ('tested' group). Tube 1 was used to test for maximal PD-L1 occupancy for each individual sample, achieved by the addition of 10 −7 M anti-PD-L1 antibody (the same antibody that was injected into the animal), thus saturating all PD-L1 expressed by cells with the antibody. Tube 2 was used to test the actual RO of each individual sample. To this end, samples were incubated with IgG2b isotype control antibody (LTF-2; anti-keyhole limpet hemocyanin). Following an incubation period of 30 min, samples were washed three times with flow cytometry buffer and stained with FITC-anti-Rat-IgG2b (1:100, no. 408205, BioLegend) and APC anti-CD3e (1:150, no. 100312, BioLegend). The percentage of PD-L1 RO on the pregated lymphocyte population (CD3 + cells) was assessed for each sample according to the following calculation: %RO = tested tube(actual RO) saturated tube (full RO) ×100.
Single-cell sorting. After perfusion with PBS supplemented with 1% l-glutamine, brains were excised without olfactory bulb and brain stem and manually chopped (0.5-1.0 mm 2 in size), before software-controlled dissociation by gentle MACS in PBS. For density gradient separation, the pellet was resuspended with 40% Percoll and centrifuged at 750g for 20 min at 20 °C; the supernatant was then discarded. Cells were suspended in ice-cold sorting buffer (PBS supplemented with 2 mM EDTA and 2% FCS) supplemented with anti-mouse CD16/32 (1:100, no. 101302, BioLegend) to block Fc receptors before labeling with fluorescent antibodies against cell-surface epitopes. Samples were stained using the following antibodies: BV421-, PE-Cy7-and FITC-conjugated CD45 (1:150, nos. 103134, 103114, 103108, BioLegend), PE-conjugated CD11b (1:200, no. 101208, BioLegend), Rabbit-anti-Tmem119 (1:100, nos. 106-6, ab210405, abcam) and Af647-anti-Rabbit (1:100, no. 406414, BioLegend). For sorting, samples were gated for CD45 + after exclusion of debris and doublets, and further gated according to the experimental design. Cell populations were sorted using either a SORP-aria or ARIA-III instrument (BD Biosciences) and analyzed with BD FACSDiva (BD Biosciences) software. Isolated single cells were sorted into 384-well cell-capture plates containing 2 µl of lysis solution and barcoded poly(T) reverse-transcription (RT) primers for single-cell RNA-seq 44 . Four empty wells were retained in each 384-well plate as a no-cell control. Immediately after sorting, each plate was spun down to ensure cell immersion in the lysis solution, snap-frozen on dry ice and stored at -80 °C until processing.

MARS-seq library preparation.
Single-cell libraries were prepared according to the MARS-seq2.0 protocol 45 . In brief, messenger RNA was isolated from cells sorted into capture plates, barcoded and converted into complementary DNA and pooled using an automated pipeline. The pooled sample was then linearly amplified by T7 in vitro transcription, and the resulting RNA was fragmented and converted into a sequencing-ready library by tagging samples with a pool of barcodes and Illumina sequences during ligation, RT and PCR. Each pool of cells was tested for library quality and concentration, assessed as previously described 45 .
Analysis of MARS-seq data. Single-cell RNA-seq libraries (pooled at equimolar concentrations) were sequenced on an Illumina NextSeq 500 at a median sequencing depth of ~20,000 reads per cell. Sequences were mapped to the mouse genome (mm10), demultiplexed and filtered as previously described 44,45 with the following adaptations: mapping of reads was done using HISAT (v.0.1.6) and reads with multiple mapping positions were excluded. Reads were associated with genes if they were mapped to an exon, using the UCSC Genome Browser for reference. The level of spurious unique molecular identifiers (UMIs) in the data was estimated using statistics on empty MARS-seq wells, and rare cases with estimated noise >5% were excluded. We used the R package MetaCell 46 to generate homogenous and robust groups of cells in each analysis. Cells were filtered for mitochondrial genes using a 20% threshold; we then filtered cells with <400 UMIs. Feature genes were selected based on variance to mean ratio using the function 'mcell_gset_filter_varmean' with parameter Tvm = 0.2 and minimum total UMI count >50. In analysis of all experiments, microglia were identified according to the expression of key marker genes including Hexb, P2ry12 and Tmem119. For the differential gene expression analysis presented in Extended Data Fig. 1b, we used the single-cell RNA-seq dataset published by Keren-Shaul et al. 20 available at Gene Expression Omnibus (GEO) with accession no. GSE176085. In Fig. 4, because microglia were highly abundant compared to MDM and occupied most of the space in a two-dimensional (2D) plot, we randomly sampled an equal number of microglia for the purpose of visualization to obtain a more balanced 2D map focused mainly on the MDM compartment. Aβ 1-40 and Aβ 1-42 ELISA. Following intracardial ice-cold PBS perfusion of mice, hippocampi were collected from one or both brain hemispheres, as indicated in the text, and immediately frozen and stored at −80 °C. Samples were homogenized in TBS solution (Tris pH 7.4 (50 mM), NaCl (150 mM)), EDTA (2 mM) and 1% Protease Inhibitor Cocktail (Sigma-Aldrich) using a Micro Tube homogenizer with plastic pestles. Lysates were then centrifuged for 40 min at 350,000g in 500-μl polycarbonate centrifuge tubes (Beckman Coulter) at 4 °C in an Optima MAX-XP Ultracentrifuge with a TLA 120.1 rotor (Beckman Coulter). Supernatant was collected and stored at −80 °C as the TBS-soluble fraction, ready for further ELISA assay. The pellet was completely resuspended in Triton X-100 solution [Tris pH 7.4 (50 mM), NaCl (150 mM), 1% Triton X-100 (Sigma-Aldrich) and 1% Protease Inhibitor Cocktail (Sigma-Aldrich)]. After 15 min of incubation at 4 °C, samples were centrifuged for 40 min at 350,000g at 4 °C and the supernatant was collected as the 'Triton X-100soluble fraction' and stored at −80 °C. Bicinchoninic acid assay (Pierce BCA Protein Assay Kit) was performed to determine protein concentrations in both TBS-soluble (also referred as 'soluble') and Triton X-100-soluble fractions. Aβ 1-42 levels presented in Fig. 3a,b,e,f were measured using the LEGEND MAX β-Amyloid x-42 kit (no. 842401, BioLegend), while the LEGEND MAX β-Amyloid x-40 ELISA Kit (no. 842301, BioLegend) was used to measure Aβ 1-40 peptide levels in the experiments presented in Extended Data Fig. 3d,e; in both cases, assays were performed following the manufacturer's instructions. In the experiments presented in Fig. 5d and Extended Data Fig. 6c, the human Aβ 42 Ultrasensitive ELISA Kit (no. KHB3544, Invitrogen) was used to measure Aβ 1-42 levels, according to the manufacturer's instructions.

Blockade of CCR2.
For depletion of CCR2-expressing cells, anti-CCR2 monoclonal antibody (MC21), generated by M. Mack et al. 75 , was injected i.p. (400 μg) every 4 days. No effect on behavior was observed in WT animals. Of note, this clone of anti-CCR2 (MC21) maintains full effectiveness after five injections 76 .
Synaptophysin/Homer1 immunoreactivity analyses. Analyses of synaptophysin and Homer1 immunoreactivity levels were performed with Fiji/ImageJ (v.2.0.0, National Institutes of Health) software 78 , using a macro designed for that purpose. Before quantification, images were coded to mask the identity of experimental groups and were quantified by a researcher blinded to the identity of the groups.
Aβ plaque quantitation. From each brain, four sections per mouse were immunostained. Histogram-based segmentation of positively stained pixels was performed using Image-Pro Plus software (v.4.5, Media Cybernetics). The segmentation algorithm was manually applied to each image in the DG hilus and CA3, and the percentage of area occupied by total Aβ immunostaining, selected for minimal plaque size, was determined. Plaque numbers were quantified from the same 6-µm sagittal brain sections and are presented as the average number of plaques per brain region, in the region of interest, identically marked on all slides from all animals examined. Before quantification, images were coded to mask the identity of the experimental groups and were quantified by a researcher blinded to the identity of the groups. RAWM task. RAWM was used to test hippocampal-dependent spatial learning, following the protocol of Alamed and colleagues 79 with some modifications. Briefly, six stainless-steel inserts were placed in a plastic pool, forming six open and connected arms. A hidden platform was placed at the end of a 'goal arm' (arm 6; Extended Data Fig. 3a). Milk powder was used to render the water opaque, maintained at a temperature of 23 ± 1 °C. On day 1, the training phase, mice were subjected to 15 trials. In each trial, mice were given 60 s to find the platform. Mice that failed to find the platform were placed on it by the experimenter. Intertrial interval was, on average, 20 min. Trials alternated between a visible and hidden platform. However, from trial 12 and throughout the second day, the platform was hidden. Spatial learning and memory were measured by an investigator who was blinded to the treatment of the mice, and who recorded the number of arm entry errors (error was defined as entrance to an incorrect arm, or failure to enter any arm within 15 s) as well as the escape latency of the mice on each trial. The 30 trials were grouped into three trial bins-five bins each on days 1 and 2. Data were analyzed by a team member who did not perform the experiment. Animals that showed floating behavior or motor difficulties and were unable to execute the task were excluded from analysis. NOR task. The NOR protocol was modified from Bevins and Besheer 80 , and used a 41.5 × 41.5 cm 2 gray apparatus. The experiment spanned 2 days and included three trials: (1) a habituation trial-a 20-min session in the empty apparatus (day 1); (2) a familiarization trial-a 10-min session presenting two identical objects located 15 cm apart (day 2); and (3) a test trial-following a 1-h training-to-testing interval, each mouse was returned to the apparatus for a 6-min session in which one of the objects was replaced by a novel one (Extended Data Fig. 3b). Mouse behavior was recorded and analyzed by an investigator who was blinded to the treatment group. Novel object preference was defined as 'discrimination ratio' = time (s) spent with novel object/(time spent with familiar object + time spent with novel object). Locomotor activity was measured as distance moved (cm d -1 ) in the empty arena during the day of habituation. The time spent in the center of the arena was recorded to assess anxiety. Experiment design and statistical analysis. All statistical analyses were conducted using either GraphPad Prism (v.9.0.1, GraphPad Software, Inc.), R software (v.3.6.0, R Foundation for Statistical Computing) or MATLAB (R2019b, MathWorks). Animals were randomly distributed between treatment groups, ensuring that mice were equally distributed by sex among groups. All experiments included littermate controls. In all experiments, data collection and analyses were performed by experimentalists that were blind to the identity of the treatment and groups. No statistical methods were used to predetermine sample sizes, but they were based on previous publications 18,19,36 . Due to limited number of animals in a single cohort of mice, we pooled animals from different cohorts as detailed in the relevant legends. Unpaired Student's t-test (one-tailed) was used to assess the difference in percentage of DAM of total microglia cells (Fig. 1c), and to analyze data from immunostaining measurements (Fig. 3g-i) and ELISA with two groups (Fig. 5d and Extended Data Fig. 6c). One-tailed t-tests were performed in these experiments, since we compered only two groups and had a clear hypothesis for the directionality of the expected effect. One-way analysis of variation (ANOVA) was used to analyze the ELISA data ( Fig. 3a,b,e,f and Extended Data Fig. 3d,e) and NOR (Figs. 2b, 5c and 6b), as well as activity (Extended Data Fig. 6a) and anxiety (Extended Data Fig. 6b) measurements collected at the NOR test habituation phase, and to evaluate differences between groups in FlowCytometry analyses (Extended Data Figs. 5c,d,f and 6e). Two-way ANOVA with repeated measures was used to analyze RAWM data (Figs. 2c, 5b and 6c and Extended Data Fig. 3c). ANOVA identifying a significant result was followed by Fisher's least significant difference (LSD) test for multiple comparisons. Pearson correlation was used to determine the correlation between the performance on RAWM (time to platform) and NOR (percentage of novel object exploration time) and levels of soluble Aβ   (Figs. 3c,d and 5e). Differential gene expression analysis was performed following downsampling of the UMI matrix as part of the MetaCell package on molecules per 1,000 UMI by Mann-Whitney U-test with false-discovery rate (FDR) correction ( Fig. 4e and Extended Data Figs. 1b,e and 4c-e). Two mice were excluded from the RAWM behavioral analysis due to floating behavior-these mice did not swim and therefore could not perform the task. In the ELISA assessment, two mice were excluded before analysis due to technical issues during protein extraction. In Extended Data Fig. 5f, one sample from the group of anti-PD-L1 (1.5 mg per mouse) was not included for technical reasons. Only one sample (Fig. 3i (top, DG)) was excluded from analysis based on the results of Grubbs' analysis to identify outliers on GraphPad Prism.
Reporting Summary. Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Data availability
The single-cell RNA-seq data were deposited at the National Center for Biotechnology Information's GEO with accession no. GSE176085. For the differential gene expression analysis presented in Extended Data Fig. 1b, we used the single-cell RNA-seq dataset published by Keren-Shaul et al. 20 available at GEO with accession no. GSE176085. All underlying data used for generation of figures are collated in the associated source files. All other data are available from the corresponding authors upon reasonable request.