Deconvolution reveals cell-type-specific transcriptomic changes in the aging mouse brain

Mounting evidence highlights the crucial role of aging in the pathogenesis of Alzheimer's disease (AD). We have previously explored human apoE-targeted replacement mice across different ages and identified distinct molecular pathways driven by aging. However, the specific contribution of different brain cell types to the gene modules underlying these pathways remained elusive. To bridge this knowledge gap, we employed a computational deconvolution approach to examine cell-type-specific gene expression profiles in major brain cell types, including astrocytes (AS), microglia (MG), oligodendroglia (OG), neurons (NEU), and vascular cells (VC). Our findings revealed that immune module genes were predominantly expressed in MG, OG, and VC. The lipid metabolism module genes were primarily expressed in AS, MG, and OG. The mitochondria module genes showed prominent expression in VC, and the synapse module genes were primarily expressed in NEU and VC. Furthermore, we identified intra- and inter-cell-type interactions among these module genes and validated their aging-associated expression changes using published single cell studies. Our study dissected bulk brain transcriptomics data at the cellular level, providing a closer examination of the cell-type contributions to the molecular pathways driven by aging.


Cell type-specific analysis of gene modules related with immune responses
The dysregulation of the immune responses is a defining feature of aging 19 and plays a substantial role in the development of AD 20 .In our prior study, we identified an immune gene module characterized by upregulation with aging and enriched for genes related to immune responses 11 .To gain insights into the specific cell types and their associated genes driving this collective immune response, we assigned the immune module genes to different cell types based on their estimated cell-type-specific expressions (Fig. 3).Our analysis revealed that the immune module genes were predominantly expressed in microglia (MG), oligodendroglia (OG), and vascular cells (VC) (Fig. 3A).Pathway analysis showed that MG genes were enriched for innate immune response (-logP = 29.4),positive regulation of response to stimulus (-logP = 26.7),and cell activation (-logP = 17.6) (Fig. 3B).Among these pathways, well-known MG markers such as Trem2, Tyrobp, Cd74, Cd14, and Csf1r contributions to these pathways 21 .OG genes also displayed enrichment in immune system processes, sharing genes with MG such as Stat1, Ifit1, Ifit3, and Appl2, consistent with the recent findings suggesting that OG is active immunomodulator in response to disease 22,23 .Additionally, OG genes were implicated in specialized pathways including glial cell development (− logP = 7.6), myelination (− logP = 6.2), and axon ensheathment (− logP = 6.1) (Fig. 3C).Key genes in these pathways included Smo, Tspan2, Itgb4, Ugt8a, and Sh3tc2.Furthermore, numerous VC genes were co-expressed in MG and exhibited functions related to immune responses, with examples including Gbp2, Gbp3, P2rx7, Trim30a, and Fcer1g.VC-specific genes, on the other hand, displayed enrichment in processes such as hemostasis (-logP = 6), blood vessel morphogenesis (− logP = 4.9), and angiogenesis (− logP = 4) (Fig. 3D).Notable genes in these pathways encompassed Tgfa, Cmtm6, and Fmnl3.The complete pathway analysis results can be found in Supplementary Table S2.
To gain a deeper understanding of the interactions among these genes, we constructed a gene correlation network encompassing the three cell types.Notably, we observed strong connections between MG genes (Trem2, Stat1, Tyrobp, Trim30a), OG genes (Ugt8a, Itgb4, Car2), and a VC gene (Psmb9) (Fig. 3E).
To validate the age-related expression changes of these key genes and others, we calculated fold changes between old and young animals in MG, OG, and VC, respectively, using our deconvoluted cell-type-specific expressions.We then compared these changes with data from three prominent mouse aging single-cell RNA sequencing studies [12][13][14] (Fig. 3F).Specifically, for the study by Ximerakis et al. 12 , we extracted the fold changes    12 , Allen et al. 13 and Lee et al. 14 .MG genes are denoted in pink, OG genes in gold, and VC genes in green.www.nature.com/scientificreports/ between old and young animals in each cell type as listed in their publication.In the studies of Allen et al. 13 , and Lee et al. 14 , we performed differential expression analyses comparing old versus young animals within each cell type using their respective Seurat data objects.Out of the 50 immune genes examined, 31 (62%) displayed the same change in expression direction with age as reported by Ximerakis et al. 12 , 33 (66%) exhibited consistent changes with aging as observed in Allen et al. 13 , and 26 (52%) showed similar age-related changes as reported by Lee et al. 14 .In total, 45 out of these 50 genes (90%) were validated by at least one of the single-cell studies.Additionally, 15 genes (30%) were validated by all three studies.Five genes displayed opposite changes in expression direction with aging compared to all three single-cell studies or were expressed in less than 10% of cells in these studies.These genes include Ifit3, Isg15, Csf1r, Smo, and Plxdc2 (Fig. 3F).It's worth noting that the discrepancies may be attributed to variations arising from factors such as animal sexes, ages, brain regions, tissue preparations, and cell dissociation procedures employed in these single-cell studies.Collectively, our deconvoluted data indicate that microglia, together with oligodendroglia and vascular cells, play a pivotal role in orchestrating the immune response network within the brain during the aging process.

Cell type-specific analysis of gene modules related with lipid metabolism
Lipids play pivotal roles in cell signaling and various physiological processes within the brain and central nervous system.Disruptions in lipid homeostasis and impaired lipid metabolism have been implicated in both aging and AD pathogenesis 3,24 .In our previous study 11 , we identified an upregulated lipid metabolism module (part of the extracellular vesicle module) associated with aging.Given the potential relevance of these genes to aging and AD, we conducted further investigations focusing on this module.Our findings revealed that the lipid metabolism module predominantly comprised genes expressed in astrocytes (AS), microglia (MG), and oligodendroglia (OG) (Fig. 4A).Only a limited number of genes were attributed to neuronal population (NEU) and vascular cells (VC).Specifically, in AS, genes were significantly enriched for processes such as phospholipid metabolism (− logP = 48.4),fatty acid metabolism (− logP = 43.9),and phosphatidylinositol metabolism (− logP = 24.3)(Fig. 4B).Key genes contributing to these pathways included Mboat2, Mecp2, Pik3c2a, Gpld1, Inpp5a, Chpt1, Ocrl, and Lpin1.In MG, genes were notably enriched for fatty acid metabolism (− logP = 25.1), the PPAR signaling pathway (− logP = 8.4), and the regulation of cholesterol metabolic process (− logP = 5.8) (Fig. 4C).Genes associated with these pathways included C3, Spp1, Pnpla7, Apobec1, Hcar2, Ggta1, and Gpx4.OG genes exhibited enrichment in glycerolipid metabolic processes (− logP = 41.5),phospholipid metabolic processes (− logP = 38.7),and sphingolipid metabolic processes (− logP = 22.1) (Fig. 4D).Genes contributing to these pathways encompassed Cers2, Asah1, St8sia1, Smpdl3b, and Alg2.The complete pathway analysis results can be found in Supplementary Table S2.
To validate the aging-associated expression changes of these key genes, we again compared fold changes between old and young animals in our deconvoluted data with the three single-cell RNA sequencing studies [12][13][14] .We found that, out of the 38 lipid metabolism genes, 25 (65.7%)exhibited consistent directional changes with age as reported by Ximerakis et al. 12 , 17 (44.7%)with Allen et al. 13 , and 14 (36.8%) with Lee et al. 14 .In total, 34 out of these 38 genes (89.5%) were validated by at least one single-cell study.Of note, two genes-Lpin1 and Nus1-were validated by all three studies, while four genes displayed opposite aging-related changes compared to these three single-cell studies or were expressed in less than 10% of cells in these single-cell studies, including Pigv, Fitm2, Smpdl3b, and Bdh2 (Fig. 4F).
Altogether, our deconvoluted data indicate both similarities and differences in lipid dysregulation across various cell types.Further investigation is required to elucidate how these lipid changes influence cellular functions.

Cell type-specific analysis of gene modules related with mitochondrial function
Mitochondrial dysfunction and diminished energy metabolism are prominent neuropathological characteristics of AD 25 and represent key hallmarks of the aging process 26 .In our prior research, we identified a module of genes associated with mitochondria that displayed a downregulation pattern with aging and enriched for respiratory chain pathways.Notably, our deconvolution analyses revealed that these genes were predominantly expressed in vascular cells (VC) (Fig. 5A).This module comprised pivotal hub genes (Fig. 5B), including members of the Cytochrome C Oxidase gene family such as Cox6a1, Cox7a2, Cox8a, and Cox7b, alongside genes from the mitochondrial NADH: ubiquinone oxidoreductase gene family, including Ndufa1, Ndufa8, Ndufb3, and Ndufb6.Additionally, prominent genes from the Ubiquinol-Cytochrome C Reductase gene family, such as Uqcrb and Uqcr10, were also featured within this module (Fig. 5B).
The further validation with the three single-cell RNA sequencing studies showed that, among the 28 mitochondria function genes, 23 (82.1%) exhibited concordant changes with age as reported by Ximerakis et al. 12 , 6 genes (21.4%) aligned with Allen et al. 13 , and 16 genes (57.1%) were consistent with Lee et al. 14 (Fig. 5C).Importantly, 26 out of these 28 genes (92.9%) were validated by at least one single-cell study.Two genes, Cox7a1 and Cox7b2, displayed opposing changes with aging when compared to these three single-cell studies or were expressed in less than 10% of cells in these single-cell studies.It is worth noting that the module of mitochondria-related genes exhibited the most variation when compared to Allen et al. 13 .This divergence is likely due to inadequate sampling of certain vascular cell types, including pericytes and endothelial cells, in the Allen et al. 13 study, potentially leading to inaccuracies in gene expression profiling.
Overall, the consistent downregulation of these mitochondrial function genes with aging, as reported in our study and corroborated by Ximerakis et al. 12 and Lee et al. 14  www.nature.com/scientificreports/mitochondrial signaling in the endothelium constitutes an early event in the aging process and contributes causally to the development of various age-related diseases 27 .

Cell type-specific analysis of gene modules related with synapses
Accumulated research over the years has firmly established that synaptic transmission undergoes changes as individuals age.Indeed, studies have shown that genes regulated by age are associated with synaptic functions 28 .

Lpin1
Regulation of cholesterol metabolic process Gpld1   12 , Allen et al. 13 , and Lee et al. 14 .AS genes are denoted in dark blue, MG genes in pink, and OG genes in gold.
In our investigation, we deconvoluted the synapse module, which exhibited downregulation with aging.Our analysis revealed that the majority of genes within this module were expressed in neurons (NEU), with some also expressed in vascular cells (VC) (Fig. 6A).A subset of genes was associated with astrocytes (AS), demonstrating  12 , Allen et al. 13 , and Lee et al. 14 .VC genes are denoted in green.
In the validation with single-cell RNA sequencing data sets, we found that out of the 30 neuronal genes, 21 (70%) exhibited consistent directional changes with Ximerakis et al. 12 , and 15 (53.3%) with Allen et al. 13 .Of note, there were not many genes captured by Lee et al. 14 due to the lack of neuronal population in this dataset.In total, 27 out of these 30 neuronal genes (90%) were validated by at least one single-cell study (Fig. 6C).Three genes displayed opposing changes with aging compared to these three single-cell studies or were expressed in less than 10% of cells in these single-cell studies, including Lzts3, Oprd1, and Pak1.Regarding the genes associated with vascular cells (VC), our validation using all three single-cell studies indicated that 3 out of the total 9 (33.3%)VC genes exhibited consistent aging-related changes with Ximerakis et al. 12 , 8 (88.9%) with Allen et al. 13 , and 5 (55.6%) showed similar aging-related changes with Lee et al. 14 .In total, 8 out of 9 VC genes (88.9%) were validated by at least one single-cell study, with gene Adrbk1 being the only exception, showing opposite changes with aging compared to these three single-cell studies or being expressed in less than 10% of cells in these single-cell studies.
Together, these data provide support for the concept of synaptic alterations occurring with aging, as well as for the critical network changes of neurovascular coupling and neuronal-glial-endothelial interactions observed during aging 29,30 .

Discussion
In this study, we applied cell type deconvolution to bulk RNA sequencing data from apoE-TR mice to identify key cell types involved in gene modules related to aging.We also explored the molecular pathways and interactions between module genes within and across cell types.Additionally, we identified hub genes with significant connections to others, providing a further understanding of the contributions of different cell types to important aging-regulated pathways.Most of our deconvoluted aging-associated gene expression changes were validated by one or more previously published single cell studies.
During the aging process, the brain undergoes various immune responses, such as neuroinflammation, microglial activation, blood-brain barrier dysfunction, myelin disintegration, immune cell senescence, and altered interactions between immune cells and other brain cells 2 .These immune changes contribute to cognitive decline and increase the risk of neurodegenerative diseases.Our deconvolution data consistently highlights the involvement of microglia, oligodendrocytes, and vascular cells in the immune response during aging.It was reported that microglia and oligodendrocytes have mutual interactions and influence each other's functions 31 .Microglia play a role in modulating immune responses and impacting the survival and differentiation of oligodendrocytes.Additionally, brain vascular cells, including endothelial cells and pericytes, have a dual role in maintaining the integrity of the blood-brain barrier and participating in immune responses 32 .They interact with immune cells and influence inflammation in the brain.Further understanding the intricate crosstalk among these brain cells and their interactions with immune responses is crucial for unraveling disease mechanisms and developing targeted therapies.
Furthermore, our deconvolution analysis reveals dysregulation of lipids in astrocytes, microglia, and oligodendrocytes, with variations in the specific lipid processes involved in each cell type.Both astrocytes and microglia show affected fatty acid metabolism, which has been linked to neuroinflammation [33][34][35] .It was reported that the pro-inflammatory microglia suppress fatty acid oxidation and synthesis, while anti-inflammatory microglia enhance these processes 35 .Additionally, microglia show dysregulation in the PPAR signaling pathway, known to modulate microglial innate immunity and fatty acid metabolism 36,37 .These findings suggest a potential link between lipid dysregulation and neuroinflammation in astrocytes and microglia.In oligodendrocytes, we observe dysregulation in glycerolipid, phospholipid, and sphingolipid processes.While the understanding of glycerolipid metabolism in oligodendrocytes is limited, it is crucial to acknowledge the significant roles of phospholipids and sphingolipids in myelin composition 38,39 .Importantly, lipid metabolism in oligodendrocytes has been implicated in the development of demyelination, a key factor contributing to the risk of AD 40 .Taken together, these findings highlight age-related lipid dysregulation that may contribute to neuroinflammation and demyelination during the aging process and in AD.
Mitochondrial dysfunction is a key feature of aging, characterized by structural and functional changes in the mitochondria 2 .This leads to reduced energy production, increased production of reactive oxygen species, and impaired mitochondrial quality control.Interestingly, our deconvolution data suggests strong mitochondrial respiration dysregulation in brain vascular cells.It is known that endothelial mitochondrial dysfunction is an important factor causing abnormal function of the endothelium, which plays a central role during atherosclerosis development 41 .These changes can also disrupt the integrity and function of the blood-brain barrier, resulting in increased permeability and inflammation in the brain.
Our deconvoluted data imply that the downregulated synapse module with aging includes genes primarily originating from neurons, vascular cells, and astrocytes.This suggests that during the aging process, it's not just neurons experiencing synaptic loss; the interactions or networks between neurons, glial cells, and vascular cells are also disrupted.This observation aligns well with our current understanding that the brain is a complex tissue, where vascular cells and glial cells closely collaborate with neurons to maintain brain functions 29,30 .These network gene changes can be challenging to detect in single-cell RNA sequencing data if specific cell types are not adequately captured during the cell isolation process.
In our prior study, we reported several gene module changes associated with the APOE genotype, including lysosome functions and RNA splicing 11 .However, our deconvolution results indicate that the majority of these genes are attributed to neurons (data not shown).Unfortunately, because the neuronal population was not adequately represented in our single-cell validation dataset 14 , we were unable to validate and present this data    12 , Allen et al. 13 and Lee et al. 14 .NEU genes are denoted in purple, and VC genes in green.www.nature.com/scientificreports/It has long been recognized that cell heterogeneity in tissue samples leads to the caveat in bulk RNA sequencing experiments that it is often hard to dissociate true transcriptomics landscape changes from the contribution of cell types.In the study of diseases where certain cell types play a major role in disease pathogenesis, cell-type specific information is particularly pertinent.Computational deconvolution methods have the unique advantage of being readily available as open-source packages, and can be applied to old or previously published data where samples are no longer available or difficult to re-obtain.The success of our application of the deconvolution algorithm in this study serves as proof of principal for the potential of applying computational deconvolution methods to the vast available bulk public dataset to gain novel insights of disease mechanisms.

Limitations of the study
Our study has several limitations, some of which are inherent to the use of computational deconvolution.First, it is important to acknowledge that computational methods are susceptible to generating false positives.In our case, the transition from bulk gene expression to cell-type-specific gene expression involved multiple layers of machine learning processes, which collectively could lead to inaccurate estimations of gene expression in certain cell types.To mitigate this, it is essential to cross-reference our results with publicly available databases and existing literature, ensuring that we identify and rectify any such estimations and avoid overgeneralization of our findings.For genes with limited public information, experimental validations, such as co-staining, will be crucial to confirm the accuracy of the sample-level deconvolution results in the future.Second, it's important to recognize that different computational deconvolution algorithms may be better suited for specific types of bulk transcriptomics data.Many deconvolution methods have been developed for or evaluated on blood, immune, or tumor samples, and their performance may not be equivalent for tissues with high complexity, such as the brain 7,10,42 .Depending on the specific characteristics of the bulk tissue being studied, it becomes imperative to carefully evaluate the choice of deconvolution method and exercise caution in its application.Acknowledging these limitations, future study should aim to minimize potential errors through rigorous validation strategies.It is also important to critically interpret our findings within the context of existing knowledge, recognizing that computational deconvolution, while a powerful tool, must be employed judiciously and complemented with experimental evidence when necessary.

Dataset description
We used our previously published cerebral cortex transcriptomics data of 141 apoE-TR animals, including male and female APOE2, APOE3 and APOE4 mice at 3, 12, and 24 months of age 11 .The animal husbandry, tissue processing, RNA extraction, RNA sequencing, quality control and gene count normalization were previously described 11 .The CQN normalized RPKM gene expression values of 19,120 genes were used for deconvolution analyses.

Cell type deconvolution and gene correlation network analysis
We used the CIBERSORTx algorithm high resolution mode to deconvoluted the bulk transcriptomics data based on the author's recommendations 8 .Specifically, we used the mouse brain single cell RNA sequencing data from GSE129788 to construct the single cell signature gene matrix due to the data's close match in both age and brain region to our bulk data 12 .From the cell-type-specific gene expression results, we removed genes that were either "NA" or "− 1", or had the same deconvoluted expression values across all samples, which indicated insufficient evidence of estimation.Module genes were scaled and visualized via hierarchical clustering in heatmaps using the ComplexHeatmap R package 43 .Pathway analyses of cell type specific module genes were performed using Metascape (https:// metas cape.org) 44 .Gene correlation network analysis of selected modules were performed using WGCNA 45,46 and visualized using Cytoscape version v3.10.0 47 .

Validation of cell-type-specific module genes using public data
To validate the deconvoluted cell-type-specific modules genes, we compared their cell type assignments against the following sources: (1) top 50 mouse brain cell type marker genes from each brain cell type from the BRETI-GEA R package, which compared and contrasted five human and mouse cell type-specific transcriptomics datasets to identify consensus brain cell-type marker genes 9 , (2) brain single cell RNA sequencing expression profiles from the Human Brain Atlas v22.0 (proteinatlas.org),(3) published mouse brain MG gene expression profiles across different activation states 15 , (4) published AS gene expression profiles in mouse brain cortical regions 16 , and (5) published OG gene expression profiles across various mouse brain regions 17 .
To validate aging-associated changes in our cell-type-specific module genes, we compared fold changes of old versus young animals between our deconvoluted data and three prominent mouse aging single cell studies [12][13][14] .For our deconvoluted genes in each cell type, fold changes were calculated between 24 and 3 months using an ANOVA model while adjusting for APOE genotype and sex.The fold changes between old (21-23 months) and young (2-3 month) mice from Ximerakis et al. 12 were obtained from their supplementary Table 6, which was calculated using the MAST model 48 for each annotated single cell type.To calculate fold changes in Allen et al. 13 , we downloaded the integrated and annotated R data object GSE207848_Cell.rdsfrom CELL x GENE repository (https:// cellx gene.czisc ience.com/ colle ctions/ 31937 775-0602-4e52-a799-b6acd d2bac 2e), and used the MAST model to compare 90 weeks (22.5 months) and 4 weeks (1 month) for each single cell type using Seurat version 5 49 .To calculate fold changes in Lee et al. 14 , we obtained the integrated and annotated R data file from the authors and used the MAST model to compare 24 months and 3 months for each single cell type while adjusting for APOE genotype using Seurat version 5.

Figure 1 .
Figure 1.Computational deconvolution of bulk transcriptomics data and identification of major cell types.(A) Analytical workflow.(B) Heatmap showing the expression of known cell type marker genes deconvoluted from our previously published bulk RNA-seq data.

Figure 2 .
Figure 2. The proportion of the five major brain cell types in each age, APOE and sex group.(A) The estimated cell type proportions in 3 months (3 m) and 24 months (24 m) mice.(B) The estimated cell type proportions in APOE2 (E2), APOE3 (E3), and APOE4 (E4) mice.(C) The estimated cell type proportions in female and male mice.T test was used to detect differences between groups.Bonferroni correction was applied to correct for multiple testing.****p < 0.0001; n.s.not significant.

Figure 3 .
Figure 3.Immune gene module deconvolution.(A) Heatmap showing the expression of immune module genes in the five cell types.(B) Top GOs enriched in MG genes.All 3 GO terms were based on Biological Processes (BP) (C) Top GOs enriched in OG genes based on BP. (D) Top GOs enriched in VC genes.Hemostatis was based on Reactome Gene Sets, while Blood vessel morphogenesis and Angiogenesis were based on BP. (E) Correlation network of selected module genes from MG, OG and VC.(F) Dot plot showing gene fold changes of old versus young animals from our deconvoluted data, Ximerakis et al.12 , Allen et al.13 and Lee et al.14 .MG genes are denoted in pink, OG genes in gold, and VC genes in green. https://doi.org/10.1038/s41598-023-44183-7

Figure 4 .
Figure 4. Lipid metabolism gene module deconvolution.(A) Heatmap showing the lipid metabolism module genes in the five cell types.(B) Top GOs enriched in AS genes.All 3 GO terms were based on Biological Processes (BP).(C) Top GOs enriched in MG genes.Fatty acid metabolic process and Regulation of cholesterol metabolic process were based on BP, while PPAR signaling pathway was based on WikiPathways.(D) Top GOs enriched in OG genes based on BP. (E) Correlation network of selected module genes from AS, MG and OG.(F) Dot plot showing gene fold changes of old versus young animals from our deconvoluted data, Ximerakis et al.12 , Allen et al.13 , and Lee et al.14 .AS genes are denoted in dark blue, MG genes in pink, and OG genes in gold.

Figure 5 .
Figure 5. Mitochondria gene module deconvolution.(A) Heatmap showing the respiration chain module genes in the five cell types.(B) Correlation network of VC genes.(C) Dot plot showing gene fold changes of old versus young animals from our deconvoluted data, Ximerakis et al.12 , Allen et al.13 , and Lee et al.14 .VC genes are denoted in green.
https://doi.org/10.1038/s41598-023-44183-7www.nature.com/scientificreports/ in the current study.Therefore, the analysis and validation of APOE genotype-related gene module changes will be a subject of future investigation.