Our integrated analysis of snRNA-seq datasets comprising brain biopsies from various neurological pathologies and healthy subjects, has resulted in the creation of the first comprehensive human atlas of microglia called HuMicA, which includes 60,557 microglial nuclei. The clustering analysis of the integrated object confirmed that microglia encompass a range of distinct cell subpopulations rather than a one-directional transition from homeostasis to activation. In mice, several studies have already developed atlases covering the spectrum of microglia throughout embryonic development, aging, and in pathological conditions27,31,32. The translation of these findings into human biology has become feasible due to the availability of snRNA-seq studies, which have significantly contributed to the more precise characterization of the human cerebral environment, by identifying specific cellular subpopulations. However, microglia often take on secondary attention in human studies primarily due to their low abundance within bulk brain samples. In our analysis, the average proportion of the immune cell population relative to the entire dataset was approximately 5%, which aligns with the reported amounts in the original studies. This limitation has previously impacted the scalability of the studies, requiring larger number of samples, and hinders the identification of smaller microglia subpopulations, which remain underrepresented in smaller datasets. In this study, we have successfully overcome this significant limitation by employing a bioinformatic strategy that maximizes the utilization of already published data.
The HuMicA comprises eleven microglia subpopulations, with ten of them containing a substantial number of cells, allowing for detailed characterization. For instance, we successfully described the classical homeostatic microglia phenotype, which accounted for the most abundant population, Homeos1. In addition, we identified three additional populations with a homeostatic gene profile. Two of these subpopulations were smaller in size and likely represent two distinct stages of pre-activation, one leaning towards DAM activation (Homeos3) and the other tending towards general pro-inflammation (Homeos4). Special attention must also be given to Homeos2, which expresses high levels of P2RY12 and CX3CR1, and we annotated as “homeostatic”. Indeed, Homeos2 presents a distinct profile specifically driven by high expression of GRID2 and ADGRB3, which segregates it clearly from the other subpopulations Interestingly, a similar microglia population, denominated AD2 was identified in an external snRNA-seq study. This AD2 subpopulation was found to be expanded in AD compared to controls and was characterized by upregulation of P2RY12, CX3CR1, GRID2 and ADGRB3. Additionally, this gene expression pattern was also associated with synapse assembly and specifically related to a response to tau-bearing dying neurons33. The true nature of the Homeos2 is subpopulation still requires further studies to elucidate whether it behaves as a homeostatic microglia subpopulation or represents a distinct pathology-related state.
Previous studies had already demonstrated an association between microglia activation and neuropathology. In these, the typical strategy often involves assessing differential gene expression in microglia in from patients compared to healthy controls7–12, 14,16,19,20, with a general description of a pathology-related pro-inflammatory and activated state. In addition, some studies have shown a differential distribution of microglia subpopulations between patients and controls, with the transcriptomic profile of these subpopulations enriched in genes associated with activation. This includes microglia subpopulation resembling the DAM phenotype observed in mice, as well as those expressing genes and in genes associated with genetic risk factors for neurological diseases7,8,11,15. The DAM phenotype has indeed been a valuable tool in evaluating the role of microglia in neurodegeneration. It has been shown to be prevalent not only in AD mice, but also in a model of amyotrophic lateral sclerosis5. Furthermore, the aforementioned DAM enrichment in pathology-related microglia extend beyond AD7,8,11, encompassing other conditions such as autism15 and severe COVID-1919, emphasizing the inter-disease and unspecific nature of DAM. Indeed, the role of the DNA phenotype in neuropathology, particularly in human AD, has been a topic of debate, and its replicability has been questioned12. Moreover, from the methodological standpoint, it has been shown that snRNA-seq microglia express lower levels of main DAM genes in comparison to scRNA-seq microglia26. However, in the case of HuMicA, we exclusively used snRNA-seq data and were still able to widely demonstrate detectable and varying levels of DAM genes across the microglia populations of our datasets.
It is reasonable to conclude that the DAM phenotype alone may not be sufficient to fully understand microglia’s behaviour in pathology. The original study proposed a sequential two-stage process for DAM activation, with the first stage being Trem2-dependent stage 1 and the second stage being Trem2-independent5. However, a recent groundbreaking study by Silvin and colleagues (2022) provided further insights “deconstructing” the DAM signature. These authors developed an encompassing myeloid cell map of mice brain samples by integrating scRNA-seq data from different neuropathological settings. Within this “M-verse”, they observed two DAM populations with different ontogenies and concluded that one of these populations is composed by infiltrating monocyte-derive “disease-inflammatory macrophages” (DIMs). The authors demonstrated that bona fide DAMs were Trem2-dependent, while DIMs were Trem2-independent27. However, the authors did not elaborate on whether the DIMs could represent what was initially described as the stage 1 DAM based on the dual profile of Trem2-dependency. Here, we have shed light on that matter by describing the simultaneous prevalence of all these phenotypes, including the bona fide DAM, with the stage 1 DAM (Intermediate.DAM) and stage 2 phenotype (Final.DAM), and the DIM population. It is worth noting that the DIM population clustered closely with the macrophage population in the HuMicA. Regarding the Trem2-dependency, our experimental design did not allow for its direct evaluation, and we observed the expression of across all TREM2 populations.
An important aspect used to interpret distinct microglia subpopulations is the differential expansion or depletion in different pathological conditions. We observed a decrease in homeostatic clusters in a generalized way across the studied diseases. As for the DAM and DIM subpopulations, we noticed altered distribution of smaller and more precise subpopulations within the Final.DAM, Intermediate.DAM and DIM clusters, with differential patterns across pathologies. For example, we described two DIM subpopulations, associated with differentiation, cytokine binding and response to TGF-B, which were expanded exclusively in epilepsy. In addition, a Final.DAM subpopulation with migration-related phenotype is increased in AD and MS, while an Intermediate.DAM subpopulation was more prevalent in patients with ASD and MS. DAM populations have consistently been shown to be expanded in pathology, particularly in AD. In our study, we were able to identify niche substates within this spectrum with distinct profiles differentially distributed in pathology, rather than observing a general expansion of the DAM phenotype. We must consider the potential confounding effect caused by over-normalization of the data, which is inherent to the integration of multiple datasets. Therefore, subsequent validations are required to evaluate whether these associations between pathologies and populations are genuine and not random outputs of statistical bias.
The HuMicA cluster 6 showed a clear and challenging-to-refute expansion in MS and LBD. This subpopulation exhibited detectable levels of oligodendrocyte markers, such as ST18, PLP1, and MBP. Interestingly, this pattern has been previously observed and validated in vitro in a study whose data we included in the HuMicA. In that study, the authors incubated human primary microglia with myelin particles and observed the localization of oligodendrocyte-derived mRNA within the phagocyting microglia and close to the nuclei14. Similar patterns of phagocytic microglia engulfing oligodendrocyte-derived mRNA were described in mice scRNA-seq data32. This suggests that cluster 6 in the HuMicA could potentially be related to microglia ingulfing oligodendrocyte-derived material. By the same rationale, one could speculate that clusters 3 and 7 of the HuMicA may be associated with excessive pruning of synaptic terminals. However, we decided not to highlight these subpopulations in our subsequent analyses due to the potential implications of non-microglial RNA presence within the microglia transcriptomes impact the interpretation of the data. Nevertheless, the demyelinating nature of MS raises questions on the biological significance of these results.
The findings described in any integrated object must be considered within the limitations intrinsically related to in silico integration of multi-source data, which may potentially mask or exaggerate biologically relevant traits. We acknowledge the necessity for validation of our findings through additional approaches. In situ hybridization and immunohistochemistry of brain tissue slides would be valuable in corroborating the prevalence and coexistence of these microglia subpopulations, and further elucidating their potential pathology-related functions. Nonetheless, we successfully validated HuMicA in an independent scRNA-seq dataset, which was published as an atlas of human microglia subpopulations on its own30. We observed that the transcriptomic profiles characteristic of the main subpopulations in the HuMicA were indeed associated with specific microglia clusters in the external dataset. This approach not only validated the presence of the HuMicA patterns but also revealed a “clean” translation without major overlapping of its signatures on the clusters of the external data.
Using the in-silico approach in our study, we successfully compensated for the common limitation of low sample size in individual human microglia snRNA-seq datasets, which have significantly prevented the ability to fully characterize microglial subpopulations. We have described and validated the HuMicA, in which we describe multiple homeostatic microglia states as well as pathology-related phenotypes, only described in animal models up to now. We intend that the HuMicA is used as a public resource to study microglia under multiple experimental settings thus contributing for the evolution of the overall knowledge on microglial biology.