Abstract

Objectives

   Neuroinflammation signaling has been identified as an important
   hallmark of Alzheimer’s disease (AD) in addition to amyloid β plaques
   (Aβ) and neurofibrillary tangles (NFTs). However, the molecular
   mechanisms and biological processes of neuroinflammation remain unclear
   and have not well delineated using transcriptomics data available. Our
   objectives are to uncover the core neuroinflammation signaling pathways
   in AD using integrative network analysis on the transcriptomics data.

Materials and methods

   From a novel perspective, i.e., investigating weakly activated
   molecular signals (rather than the strongly activated molecular
   signals), we developed integrative and systems biology network analysis
   to uncover potential core neuroinflammation signaling targets and
   pathways in AD using the two large-scale transcriptomics datasets,
   i.e., Mayo Clinic (77 controls and 81 AD samples) and ROSMAP (97
   controls and 260 AD samples).

Results

   Our analysis identified interesting core neuroinflammation signaling
   pathways, which are not systematically reported in the previous studies
   of AD. Specifically, we identified 7 categories of signaling pathways
   implicated on AD and related to virus infection: immune response,
   x-core signaling, apoptosis, lipid dysfunctional, biosynthesis and
   metabolism, and mineral absorption signaling pathways. More
   interestingly, most of the genes in the virus infection, immune
   response, and x-core signaling pathways are associated with
   inflammation molecular functions. The x-core signaling pathways were
   defined as a group of 9 signaling proteins: MAPK, Rap1, NF-kappa B,
   HIF-1, PI3K-Akt, Wnt, TGF-beta, Hippo, and TNF, which indicated the
   core neuroinflammation signaling pathways responding to the low-level
   and weakly activated inflammation and hypoxia and leading to the
   chronic neurodegeneration. It is interesting to investigate the
   detailed signaling cascades of these weakly activated neuroinflammation
   signaling pathways causing neurodegeneration in a chronic process, and
   consequently uncover novel therapeutic targets for effective AD
   treatment and prevention.

Conclusions

   The potential core neuroinflammation and associated signaling targets
   and pathways were identified using integrative network analysis on two
   large-scale transcriptomics datasets of AD.

   Keywords: neuroinflammation, signaling network, Alzheimer’s disease,
   molecular mechanism, signaling targets

Introduction

   A major challenge limiting effective treatments for Alzheimer’s disease
   (AD) is the complexity of AD. More than 42 genes/loci have been
   associated with AD ([45]Sims et al., 2017; [46]Verheijen and Sleegers,
   2018). Unfortunately, only few of these genes, such as CD33 ([47]Zhao,
   2019), TREM2 ([48]Gratuze et al., 2018), and MS4A ([49]Deming et al.,
   2019), are being evaluated as therapeutic targets for AD management
   ([50]Verheijen and Sleegers, 2018). Over 240 drugs have been tested in
   AD clinical trials, but only one drug, i.e., Aduhelm (aducanumab) was
   recently approved for the treatment of AD since 2003 ([51]Alzheimer’s
   Association, 2018; [52]Cummings et al., 2018; [53]Dunn et al., 2021).
   One major challenge is that the complicated pathogenesis and core
   signaling pathways of AD remain unclear. Therefore, it is significant
   to uncover the core signaling pathways implicated on AD pathogenesis
   and novel therapeutic targets of AD for identifying effective drugs and
   synergistic drug combinations (targeting multiple essential targets on
   the cores signaling network) for AD prevention or treatment. Our
   knowledge of the molecular mechanisms and signaling pathways that
   ultimately lead to the chronic neurodegeneration in AD is limited. For
   example, there are only a few strong genetic biomarkers for AD that
   have been identified, including the APOE, APP, and PSEN1/2 genes.
   However, the signaling consequence of these biomarkers, as they relate
   to the accumulation of dysfunctional A-beta and p-Tau proteins, as well
   as neuron degeneration and immune response, remains unclear.

   Over the last 10 years, neuroinflammation signaling has been identified
   as the third core feature or a central pathogenesis mechanism of AD
   ([54]Akiyama et al., 2000; [55]Combs et al., 2000; [56]Ekdahl et al.,
   2003; [57]Kinney et al., 2018; [58]Knezevic and Mizrahi, 2018;
   [59]Newcombe et al., 2018), in addition to amyloid β plaques (Aβ) and
   neurofibrillary tangles (NFTs) pathologies. The detailed occurrence and
   roles of neuroinflammation are complex and remain unclear, though set
   of inflammation and immune genes, such as TNF ([60]Decourt et al.,
   2017), IL-1beta ([61]Shaftel et al., 2008; [62]Ng et al., 2018), IL-6
   ([63]Ng et al., 2018), and NFkB ([64]Chen et al., 2012), have been
   reported ([65]Guzman-Martinez et al., 2019). For example, the
   accumulation of altered proteins, such as the Aβ and tau protein, are
   toxic to neuron cells and are believed to trigger the neurodegeneration
   diseases ([66]Guzman-Martinez et al., 2019). Not restricted to the
   neurons, the neuroinflammation and immune response are reported to be
   regulated and affected by the neuron-immune cell interactions
   ([67]Heneka et al., 2015; [68]Leng and Edison, 2021), such as macroglia
   and astrocytes([69]Gratuze et al., 2018; [70]Shi et al., 2019;
   [71]Mahan et al., 2022). Moreover, peripheral infections are also
   identified to be associated with degeneration diseases ([72]McManus and
   Heneka, 2017).

   On the other hand, large-scale transcriptomics datasets of AD and
   control samples have been generated, such as ROSMAP ([73]Bennett et
   al., 2018; [74]de Jager et al., 2018) and Mayo Clinic ([75]Allen et
   al., 2016), to investigate the essential targets and signaling pathways
   in AD. Network analysis models were proposed to identify the potential
   dysfunctional signaling pathways and biomarkers using these related
   RNA-seq datasets. For example, the molecular signatures and networks
   under different brain regions were reported using integrative
   co-expression network analysis, and the myelin signaling dysregulation
   was identified in AD ([76]Wang et al., 2016; [77]Wan et al., 2020). In
   addition, the co-splicing network using the WGCNA (co-expression
   network analysis model) was conducted to identify the altered splicing
   in AD, which indicated that the altered splicing is the mechanism for
   the effects of the AD-related CLU, PTK2b, and PICALM alleles ([78]Raj
   et al., 2018). Moreover, the molecular subtypes and potential driver
   genes, such as CABRB2, LRP10, and ATP6V1A of AD, were identified by
   combing key driver analysis (KDA) and multiscale embedded gene
   expression network analysis (MEGENA) ([79]McKenzie et al., 2017;
   [80]Neff and Vatansever, 2020; [81]Neff et al., 2021). Moreover, the
   meta-analysis of multi-datasets of human and mouse AD samples has been
   studied to understand the essential biomarkers and signaling modules of
   AD ([82]Wan et al., 2020).

   However, the systematic delineation of core neuroinflammation signaling
   pathways has not been reported yet in these previous computational data
   analysis models using the transcriptomics data ([83]Guzman-Martinez et
   al., 2019). Therefore, it is important to continue to pursue systematic
   investigations, including the use of network analysis techniques, in
   order to understand the details of core neuroinflammation and immune
   signaling pathways that are associated with the neurodegeneration of
   AD. In response to the preceding gap in knowledge, herein, we
   systematically sought to identify the potential core neuroinflammation
   and related signaling pathways that potentially cause neuron
   degeneration in AD by analyzing the large-scale transcriptomics, i.e.,
   RNA-seq data of human AD samples ([84]Bennett et al., 2018; [85]de
   Jager et al., 2018). Inspired by the observation of that strongly
   activated (hyper-) inflammation and immune ([86]Catanzaro et al., 2020;
   [87]Patel et al., 2021) signaling pathways are often associated with
   relatively faster/acute disease progression, such as what is observed
   in COVID-19 ([88]Gordon et al., 2020; [89]Li et al., 2020) here we
   hypothesize that the neuroinflammation in AD (a chronic disease) is
   weakly activated (i.e., with small fold change values compared with
   control samples). In another word, the transcriptomics changes in
   neuroinflammation are at a low level, which might be missed by only
   detecting the strongly activated signaling targets and pathways. This
   hypothesis was also supported by the that there are only a few strongly
   activated signals in AD samples vs. control, which is defined as large
   gene expression fold-change, e.g., fold-change ≥2.0 (see the results)
   ([90]Andrews, 2010; [91]Dobin et al., 2013; [92]Patro et al., 2017;
   [93]Gordon et al., 2020; [94]Li et al., 2020). Specifically, we
   employed the RNA-seq data of neuropathology-free controls and AD
   samples from two datasets: ROSMAP ([95]Bennett et al., 2018; [96]de
   Jager et al., 2018) and Mayo Clinic ([97]Allen et al., 2016).
   Leveraging these data, we then identified all of the weakly activated
   and inhibited genes with very low fold-change thresholds. Subsequently,
   we conducted network analyses to identify relevant core
   neuroinflammation signaling pathways. More importantly, the potential
   core neuroinflammation and associated signaling pathways were
   identified using large-scale transcriptomics datasets.

Materials and methods

Gene expression data and analysis of Alzheimer’s disease samples

   In this study, 77 control tissue samples and 81 AD pathological aging
   samples in Mayo dataset^[98]1; and 97 control samples and 260 AD
   dorsolateral prefrontal cortex samples in ROSMAP dataset^[99]2 that
   passed the quality control using FastQC ([100]Andrews, 2010) were used.
   The AD and control samples in Mayo dataset were identified using the
   diagnosis result in the clinical data. The definition of AD and control
   cases was provided in the study description, which is available
   at.^[101]3 As introduced, two experts, i.e., Dr. Dennis Dickson or Dr.
   Thomas Beach, conducted the neuropathologic evaluation of Mayo samples,
   respectively. The AD cases were diagnosed according to the NINCDS–ADRDA
   criteria and with Braak NFT stage of IV or greater. The control cases
   had Braak NFT stage of III or less; and the CERAD neuritic and cortical
   plaque densities are less than 1 (1 represents “sparse” and 0 means
   “none”) and had no any of the dementia-related diseases. The clinical
   and neuropathological data available from the ROSMAP cohort were used
   to classify controls and AD cases in our analyses. Specifically,
   subjects that exhibit no decline in cognition and scored Low or No AD
   on the Reagan scale, and with a BraakTau score of ≤3, were identified
   as controls. The AD cases were identified as patients who exhibit clear
   cognition decline as well as scored high or intermediate on the Reagan
   scale ([102]Bennett et al., 2018). For RNA-seq data analysis, these
   RNA-seq data were first aligned to reference genome GRCh38 using STAR
   (v.2.7.1a) ([103]Dobin et al., 2013). We excluded ALT, HLA, and Decoy
   contigs from the reference genome due to the lack of RNA-Seq tools that
   allow to handle these regions properly. Then, reads alignment was
   further evaluated by applying Picard CollectRnaSeqMetrics^[104]4 based
   on the GENCODE^[105]5 annotations to examine reads distribution on the
   genome. Samples that showed a Percent Duplication of >50% were flagged
   for removal. Then, the transcripts per million (TPM) values of 16,132
   common protein-coding genes were then obtained in the two datasets by
   applying the Salmon quantification tool ([106]Patro et al., 2017) in
   alignment-based mode using the aligned RNA-seq data using the coding
   transcripts of the reference genome.

Differentially expressed genes

   To identify the upregulated and downregulated genes in AD samples vs.
   control samples, the limma-voom ([107]Law et al., 2014; [108]Ritchie et
   al., 2015) approach, taking the normalized [using trimmed mean of M
   values (TMM) ([109]Robinson and Oshlack, 2010)] log-transformed TPM
   values as the input, in edgeR ([110]Robinson et al., 2010) R package,
   was employed.

Inflammation genes

   A set of 1,043 inflammation genes were obtained by extracting genes
   from the inflammatory response category as defined in the Gene Ontology
   (GO:0006954) ([111]Gene Ontology Consortium et al., 2000).
   Subsequently, 485 inflammation genes were obtained from the 5,191 kyoto
   encyclopedia of genes and genomes (KEGG) signaling genes, i.e., the
   interaction between the KEGG signaling pathways and GO inflammation
   genes.

Alzheimer’s disease GWAS data

   The GWAS data of AD was obtained from NIAGADS database ([112]Kunkle et
   al., 2019).^[113]6 The Stage 1 p-Value Data (updated by 26 February
   2019) and Stage 2 p-Value Data (updated by 27 February 2019) were
   downloaded. The 553 candidate GWAS genes and also available in the KEGG
   signaling pathways were obtained by a filter with p-value ≤ 1.0 ×
   10^–5.

Kyoto encyclopedia of genes and genomes signaling pathway enrichment analysis

   The KEGG signaling pathways were extracted using the graphite R package
   ([114]Sales et al., 2012), which consists of 311 signaling pathways
   ([115]Ogata et al., 1999; [116]Kanehisa and Goto, 2000). There are
   59,242 signaling interactions among 5,191 genes in these pathways,
   which were used for network enrichment analysis and network inference
   analysis in this study. For the network enrichment analysis, a Fisher’s
   exact test ([117]Fisher, 1932; [118]Kim, 2017) was used based upon the
   upregulated genes.

KEGG signaling network inference analysis

   To infer the signaling cascades among a set of genes of interest, we
   developed a network inference approach. First, we divided the genes
   into two groups: signaling sources (like the inflammation signaling
   genes) and signaling targets (like the apoptosis signaling genes).
   Second, a signaling network was constructed by linking the signaling
   source genes to the signaling target genes iteratively. Specifically,
   the signaling source genes were used as the initial signaling source
   nodes set: V0. The signaling target genes were used as the target nodes
   set: V1. In the iterative process, the shortest signaling
   cascades/paths between the nodes in V0 and V1 were calculated and
   identified: P[ij] ≤ g[i], g[k1], g[k2],…, g[j] >, where g[i] belongs to
   V0, and g[j] belongs to V1. Third, all of the genes on the signaling
   path P[ij] and belong to V1 were selected and added to V0, and removed
   from V1. This process was repeated until all the genes were added to
   V0. The data analysis codes are available upon request.

Results

Normal and Alzheimer’s disease tissue samples are barely separable in the
gene expression data space

   There were 77 control subjects and 81 AD cases in Mayo dataset; and 97
   control samples and 260 AD cases in ROSMAP dataset. [119]Table 1 shows
   the epidemiology information of Mayo and ROSMAP datasets. The
   transcripts per million (TPM) values of 16,132 protein-coding genes
   were obtained by applying the Salmon quantification tool ([120]Patro et
   al., 2017) in alignment-based mode using the STAR-aligned RNA-seq data.
   A multidimensional scaling (MDS) model was used to generate the 2D
   clustering plots of control and AD samples in the Mayo and ROSMAP
   datasets, respectively (see [121]Figure 1). As seen in these
   visualizations, the control and AD samples are barely separable,
   especially in the ROSMAP dataset, which of note, has more control
   samples than Mayo dataset.

TABLE 1.

   Epidemiology information of Mayo and ROSMAP datasets.
   Mayo               Control         AD           ROSMAP        Control    AD
   Total                 77           81           Total            97      260
   Male                  40           33            Male            44      82
   Female                37           48           Female           53      178
   Age, mean (SD)   82.65 (8.70) 82.57 (7.62)  Age, mean (SD)  84.24 (6.82)
   90.34 (5.75)
   Braak, mean (SD) 2.18 (0.86)  5.53 (0.52)  Braak, mean (SD) 1.88 (0.98)
   4.25 (0.85)
   APOE_22               0            0           APOE_22           2        0
   APOE_23               12           4           APOE_23           13      22
   APOE_33               56           34          APOE_33           72      141
   APOE_24               1            0           APOE_24           1       10
   APOE_34               8            36          APOE_34           8       83
   APOE_44               0            7           APOE_44           1        3
   [122]Open in a new tab

FIGURE 1.

   [123]FIGURE 1
   [124]Open in a new tab

   Alzheimer’s disease (AD) and control tissue samples are not well
   separable in the first (x-axis) and second (y-axis) principal
   components of MDS analysis on the RNA-seq protein-coding genes in Mayo
   (top-panel) and ROSMAP (bottom-panel) datasets.

   We further conducted a widely used differential expression analysis
   method to identify differentially expressed genes (DEGs) between the AD
   and control samples. To identify the common set of DEGs between the two
   datasets, we applied a number of fold-change and p-value thresholds. As
   seen in [125]Supplementary Table 1 and expected from [126]Figure 1,
   only a few common upregulated and downregulated DEGs were identified
   with fold change thresholds ≥1.5 and p-value ≤ 0.05. With the
   fold-change threshold ≥1.25, only about 230 upregulated and about 60
   downregulated genes were identified (out of the 16,132 protein-coding
   genes, ∼1.85%), in both studies. When relaxing both thresholds to fold
   change ≥1.1 and p-value ≤ 0.1, 1,120 upregulated genes and 689
   downregulated genes were identified (∼11.2% of the 161,32
   protein-coding genes). Based on these observations, we hypothesized
   that the AD-associated signaling pathways are weakly activated or
   inhibited. The “weakly” term is defined as that the differentially
   expressed genes have small fold-change values.

Weak inflammation and hypoxia are the potential major factors in the
Alzheimer’s disease brain microenvironment causing neuron cell death

   As was noted previously, we believe it is important to identify
   AD-associated weakly activated signaling pathways and understand their
   roles in AD disease progression, as well as their potential roles as
   targets for AD therapeutics. Among the 1,120 common upregulated genes
   (identified from Mayo and ROSMAP datasets), 417 genes were included in
   the 311 KEGG signaling pathways, and the rest genes were not available
   in the KEGG signaling pathways. To this end, we first conducted an
   enrichment analysis of KEGG signaling pathways using Fisher’s exact
   test applied to the 417 upregulated genes. [127]Table 2 showed the
   enriched signaling pathways with p-value ≤ 0.15. We then clustered
   these activated signaling pathways empirically into 7 categories. Using
   these 417 upregulated genes, the first principal component values in
   the MDS analysis of the AD and control samples were used to compare the
   difference between AD and control samples. The OR, absolute beta
   values, and p-values of logistic regression analysis (see [128]Table 3)
   indicated that these selected genes [p-value = 1.22 × 10^–13 (Mayo) and
   p-value = 4.2 × 10^–6 (ROSMAP)] can separate the AD and control samples
   much better than using all protein genes [p-value = 0.036 (Mayo) and
   p-value = 0.027 (ROSMAP)] in the two datasets, respectively. The box
   plots are also provided in [129]Figure 2, which indicated that the
   control and AD samples are more separable using the selected genes.

TABLE 2.

   Upregulated genes in TNF and apoptosis signaling pathways.
   Apoptosis BCL2, RELA, BIRC3, FADD, GADD45G, TNFRSF1A, NFKBIA,
   TNFRSF10B, CAPN2, TUBA1C, IL3RA, CTSH, FOS, CASP6, CASP7, TNFRSF10A,
   PARP4
   TNF signaling pathway RELA, BIRC3, FADD, MAP2K3, TNFRSF1A, NFKBIA,
   CREB3L2, FOS, CASP7, MLKL, IRF1, CEBPB
   [130]Open in a new tab

TABLE 3.

   Odds ratio (OR), beta, and p-values of logistic regression using all
   genes and 417 upregulated genes.
   All genes 417 genes
     __________________________________________________________________
     __________________________________________________________________

   OR abs(beta) P-value OR abs(beta) P-value
   Mayo 1.42 0.35 0.037 5.9 1.78 9.5 × 10^–9
   ROSMAP 1.31 0.27 0.027 1.9 0.63 9.7 × 10^–5
   [131]Open in a new tab

FIGURE 2.

   [132]FIGURE 2
   [133]Open in a new tab

   Box plots of the first principal component of MDS analysis in control
   and AD cases. Left and right columns are Mayo and ROSMAP samples,
   respectively. Upper and lower panels represent the MDS analysis using
   all genes and 417 upregulated genes, respectively. Y-axis is the first
   principal component of MDS analysis.

   As seen in our results (see [134]Figure 3 and [135]Table 4), a set of
   signaling pathways were activated, such as those involved in virus
   infection signaling [including Epstein-Barr virus, Human T-cell
   leukemia virus 1 infection, Legionellosis, Pathogenic Escherichia coli
   infection, Staphylococcus aureus infection, Yersinia infection, Human
   cytomegalovirus infection, Human papillomavirus infection, Malaria,
   Human immunodeficiency virus 1 infection, Rheumatoid arthritis, and
   Inflammatory bowel disease (IBD)]. Out of the 417 upregulated genes,
   111 genes were in common across these pathways highlighting a set of
   core genes implicated on these processes. These results indicated that
   weakly activated inflammation-related signaling pathways, such as
   inflammation, cytokine, and immune response, may be represent activated
   signaling pathways in the AD brain microenvironment.

FIGURE 3.

   [136]FIGURE 3
   [137]Open in a new tab

   Seven categories of weakly activated signaling pathways in AD.

TABLE 4.

   The seven categories of enriched kyoto encyclopedia of genes and
   genomes (KEGG) signaling pathways.
   Name P-value Name P-value
   Virus-related signaling pathways x-core signaling pathways
   Viral protein interaction with cytokine and cytokine receptor 0.0019
   PI3K-Akt signaling pathway 0.0011
   Epstein-Barr virus infection 0.0056 MAPK signaling pathway 0.0059
   Human T-cell leukemia virus 1 infection 0.0188 NF-kappa B signaling
   pathway 0.0085
   Staphylococcus aureus infection 0.0249 Hippo signaling pathway 0.0132
   Human papillomavirus infection 0.0299 TGF-beta signaling pathway 0.0137
   Pertussis 0.0375 TNF signaling pathway 0.0434
   Yersinia infection 0.0397 Rap1 signaling pathway 0.0571
   Pathogenic Escherichia coli infection 0.0430 HIF-1 signaling pathway
   0.1009
   Human cytomegalovirus infection 0.0603 Wnt signaling pathway 0.1043
   Malaria 0.0758 Apoptosis 0.0658
   Legionellosis 0.0906
   Human immunodeficiency virus 1 infection 0.1075
   Rheumatoid arthritis 0.1192 Mineral absorption 2.56E-05
   Inflammatory bowel disease (IBD) 0.1321
   Immune signaling pathways Diabetic/Lipid signaling pathways
   IL-17 signaling pathway 0.0104 AGE-RAGE signaling pathway in diabetic
   complications 0.0021
   Complement and coagulation cascades 0.0214 Adipocytokine signaling
   pathway 0.0060
   NOD-like receptor signaling pathway 0.0401 Insulin resistance 0.0283
   Th17 cell differentiation 0.1275 Glucagon signaling pathway 0.1179
   Th1 and Th2 cell differentiation 0.1368 Cushing syndrome 0.1356
   Natural killer cell-mediated cytotoxicity 0.1410
   Biosynthesis/Metabolism signaling pathways Adhesion signaling pathways
   Sulfur metabolism 0.0758 Focal adhesion 1.39E-05
   Galactose metabolism 0.0812 ECM-receptor interaction 0.0002
   Glycosaminoglycan degradation 0.0905 Adherens junction 0.0669
   Steroid hormone biosynthesis 0.1084
   Starch and sucrose metabolism 0.1084
   Primary bile acid biosynthesis 0.1437
   [138]Open in a new tab

   In addition, a group of activated signaling pathways or factors that
   are not clustering to a specific biological function or disease
   (referred to as the x-signaling pathway: the Hippo, PI3K-Akt, AGE-RAGE,
   MAPK, Adipocytokine, NF-kappa B, IL-17, TGF-beta, NOD-like receptor,
   TNF, Apoptosis, HIF-1, and Wnt signaling pathways, as well as apoptosis
   signaling) were identified. [139]Figure 4 shows the associations
   between these upregulated genes and activated signaling pathways. As
   seen in [140]Figure 4, a set of genes in the center areas of the
   network are associated with a set of signaling pathways, which could
   represent therapeutic signaling targets that could be used to inhibit
   or otherwise perturb these activated signaling pathways. In addition,
   there are a number of metabolisms signaling pathways, such as sulfur
   metabolism, galactose metabolism, starch and sucrose metabolism,
   steroid hormone biosynthesis, and glycosaminoglycan degradation,
   implicated in this model. Moreover, Th1/2/17 (T helper, CD4 + cells)
   cell differentiation and natural killer cell-mediated cytotoxicity
   signaling pathways were activated. [141]Supplementary Table 2 lists
   these associated upregulated genes and the involved signaling pathways.
   All the observations suggest a potential novel hypothesis that the
   inflammation, immune signaling, and hypoxia signaling in AD
   microenvironment activated the MAPK, PI3K-Akt, and mTOR signaling
   pathways, and then activated the HIF-1 signaling pathway. However, the
   activation of HIF-1 may fail to bring enough oxygen to protect against
   hypoxic injury to the involved neurons. The dysfunction of blood vessel
   functions, leading to hypoxia, might be partially indicated by the
   recent study showing that blood and cerebrospinal fluid flow cleaning
   the brain during sleeping ([142]Fultz et al., 2019).

FIGURE 4.

   [143]FIGURE 4
   [144]Open in a new tab

   The upregulated gene (green circles)-pathway (purpose squares)
   interaction network, including 1,021 interactions between 291
   upregulated genes and 61 enriched pathways.

Weak inflammation and hypoxia are the major factors in the Alzheimer’s
disease brain microenvironment causing neuron cell death

   As was introduced above, many genes that are activated as a function of
   virus infection, immune response, and the x-core signaling pathways are
   inflammation-related genes. It is well known that virus infection and
   immune response signaling pathways respond to inflammation. Our
   analyses identified 1,043 inflammation response genes in the gene
   ontology (GO) database (GO:0006954), which includes 492 genes in the
   KEGG signaling pathways. Interestingly, among the 417 upregulated
   genes, 66 genes were inflammation related. The p-value of observing the
   66 upregulated inflammation signaling targets from 417 upregulated
   genes identified in the AD vs. control samples was 8.34E-05 (calculated
   using Fishers’ exact test), and the odds ratio (OR) = 1.77, which
   indicates that the activation of inflammation signaling is concomitant
   with AD progression. Furthermore, there are 66 overlapping upregulated
   genes spanning the virus infection (from 111 upregulated genes) and
   x-core signaling pathways (from 136 upregulated genes), which indicate
   that the x-core signaling pathways are the likely pathways being
   activated in response to this inflammation. In addition, the activation
   of HIF-1 signaling pathway indicates the presence of hypoxia in the AD
   brain environment.

   To further investigate the network signaling cascades involving
   inflammation and apoptosis genes, we conducted the network analysis
   incorporating the activated signaling pathways and apoptosis signaling
   genes. [145]Figure 5 shows the inferred signaling network that links
   the upregulated inflammation-related genes (cyan) in virus infection
   and x-core signaling pathways, respectively, to the apoptosis signaling
   genes (red). Among the 338 signaling network genes in [146]Figure 5,
   there were 18 reported GWAS genes (with p-value ≤ 1.0 × 10^–5): PIK3CB,
   AKT3, RAF1, MAPK10, PPP2R2B, ERBB4, MECOM, IL1R1, MYD88, CAMK2D, GNB4,
   VAV3, PRKD3, PRKCE, THRB, FN1, LTBP1, and WWTR1, which were reported in
   the GWAS analysis ([147]Kunkle et al., 2019). The signaling network
   might indicate the roles of GWAS genes in neuroinflammation and
   neurodegeneration. For example, the PIK3CB, AKT3, MAPK10, and ERBB4
   genes are in the x-core signaling pathways; IL1R1 and MYD88 are
   cytokine, inflammation, and immune response genes; CAMK2D belongs to
   serine/threonine protein kinase and Ca2 + protein kinase subfamily.
   MECOM is involved in apoptosis and is a transcription factor that
   interacts with SMAD3 and MAPK genes. The GNB4, PRKD3, and PRKCE are
   related to the cAMP response element-binding protein signaling, which
   is associated with AD ([148]Saura and Valero, 2011).

FIGURE 5.

   [149]FIGURE 5
   [150]Open in a new tab

   Signaling cascades that links the upregulated genes in the virus
   infection pathways (cyan) (top) and x-core signaling pathways (bottom),
   respectively, to the upregulated apoptosis signaling genes (red) via
   the linking genes (pink).

   Further, we also compared the distance distribution among the
   inflammation-related upregulated genes and apoptosis genes as shown in
   [151]Figure 6. As can be seen, the inflammation signaling genes are
   much closer, based on the shortest path metric calculated using
   Dijkstra’s algorithm, on the signaling network (see green, blue, and
   red nodes) to the apoptosis genes compared with other signaling genes
   (see gray lines). These results indicate a potential signaling
   interaction between the inflammation signaling genes and apoptosis
   signaling. In other words, the results suggest a potential association
   that the weak inflammation and hypoxia signaling in the AD brain
   environment led to chronic neurodegeneration process via the activation
   of the x-core signaling pathways. Therefore, drugs and drug
   combinations that can perturb the x-core signaling pathways have the
   potential to be effective for AD prevention and treatment.

FIGURE 6.

   [152]FIGURE 6
   [153]Open in a new tab

   The distribution of average shortest distance on the kyoto encyclopedia
   of genes and genomes (KEGG) signaling network between the upregulated
   genes in the inflammation related signaling pathways, including virus
   infection, immune response, and x-core signaling pathways to the
   apoptosis signaling genes. As seen, the inflammation signaling genes
   are much closer (see green, blue, and red) to the apoptosis genes
   compared with all other signaling genes (see gray lines). X-axis
   represents the shortest path distance; and y-axis is the density
   function.

Activated tumor necrosis factor signaling might lead the programmed apoptosis
of neurons

   Of note, our results show that among the x-core signaling pathways, the
   TNF signaling pathways are also activated. Particularly, the TNF (Tumor
   Necrosis Factor) receptors (TNFRSF1A TNFRSF10A, and TNFRSF10B) were
   upregulated. We reconstructed these signaling pathways linking the TNF
   receptors to the upregulated genes in TNF and apoptosis signaling
   pathways (see in [154]Figure 7) using the proposed KEGG network
   inference model. As seen, the activation of these TNF signaling
   pathways might be one possible molecular mechanism causing the
   activation of apoptosis signaling via the CASP6 and CASP7 cascades.

FIGURE 7.

   FIGURE 7
   [155]Open in a new tab

   Signaling cascades, causing neuron degeneration, from the 3 TNF
   receptors (cyan) to the upregulated genes (red) in TNF and apoptosis
   signaling pathways via the linking genes (pink).

Discussion and conclusion

   In addition to Aβ and NFTs, neuroinflammation has been identified as
   another important AD pathogenesis ([156]Heneka et al., 2015; [157]Shi
   et al., 2017; [158]Kinney et al., 2018). A set of inflammation and
   immune genes, such as TREM2 ([159]Guerreiro et al., 2012; [160]Jonsson
   et al., 2012), TNF ([161]Decourt et al., 2017), IL-1beta ([162]Shaftel
   et al., 2008; [163]Ng et al., 2018), IL-6 ([164]Ng et al., 2018), and
   FkB ([165]Chen et al., 2012) have been reported ([166]Guzman-Martinez
   et al., 2019). The neural–immune interactions were also reported to
   exacerbate the AD disease progress ([167]Heneka et al., 2015; [168]Shi
   et al., 2017; [169]Kinney et al., 2018). Also, the analysis in the
   human brain endothelial cells and in Alzheimer’s brain, the JNK-AP1
   signaling pathway was identified that mediates the activation and
   expression of inflammatory genes induced by beta-amyloid peptides
   ([170]Vukic et al., 2009). The network analysis models were proposed to
   identify the potential dysfunctional signaling pathways and biomarkers
   using the same RNA-seq datasets ([171]Wang et al., 2016; [172]McKenzie
   et al., 2017; [173]Raj et al., 2018; [174]Neff and Vatansever, 2020;
   [175]Wan et al., 2020; [176]Neff et al., 2021). However,
   neuroinflammation signaling pathways have not been systematically
   uncovered and analyzed in these reported computational models using the
   same datasets. Compared with existing studies, the unique contributions
   are the discovery of systematic core neuroinflammation signaling
   pathways in AD by investigating the weakly activated molecular signals.
   We generated supportive evidence of linking these targets with
   neuroinflammation using enrichment analysis, GO terms, and network
   analysis systematically. The neuroinflammation signaling pathways,
   including the virus infection, immune response, x-core signaling
   pathways, and apoptosis signaling pathways indicated potentially novel
   targets and mechanisms of neuroinflammation in neurodegeneration.

   There are some limitations in the analysis. Through integrative and
   network analysis, we uncovered a set of core signaling pathways of
   neuroinflammation. However, first, the detailed signaling cascades
   among these neuroinflammation signaling genes in AD brain niche remain
   unclear. The general known signaling pathways in KEGG are not
   specifically designed for AD. For example, though the activation of T
   and NK cell signaling pathways can indicate the activation of some
   immune signaling–related proteins, the roles of these proteins in
   neuroinflammation remain unclear because the T and NK cells are not
   involved in the brain. In addition to the identified inflammation
   signaling pathways, it is interesting to further investigate the
   specific and detailed inflammation signaling cascades among the
   identified signaling targets that are potentially caused by the
   abnormal β-amyloid and tau protein aggregations. The other limitation
   is that the identified inflammation-related targets were not
   experimentally validated at the RNA or/and protein levels, such as qPCR
   or/and Western blotting/IHC. We will look for wet-lab collaborators for
   the potential experimental validations. Also, different brain regions
   might have different dysfunctional signaling pathways or different
   levels of neuroinflammation. Moreover, the current RNA-seq datasets are
   bulk tissue based, including multiple cell types, which confounds the
   gene expression change. Therefore, cell type composition correctness
   might affect the differential expression analysis. In addition, the
   neuroinflammation signaling pathways are reported to be affected by
   multiple cell types, such as microglia and astrocytes, in addition to
   neuron cells. Thus, it is important to analyze brain-region specific,
   single nuclear or cell RNA-seq data ([177]Mathys et al., 2019) to
   uncover the finer-scaled neuroinflammation signaling pathways in
   different cell types and their signaling interactions in AD brain
   microenvironment and niche. To evaluate the relatively weak and strong
   inflammation levels, it is interesting to investigate the inflammation
   activation levels in control cases with different pathology levels
   measured by the clinical data. Some clinical information, such as
   beta-amyloid plaques, cerebral amyloid angiopathy, or/and tau
   pathologies in the brains of AD patients, of the two publicly available
   datasets were not available, thus some signaling targets and pathways,
   like HIF-1 signaling pathway, might be associated with subtypes of AD
   patients with vascular dementia. Therefore, it is interesting and
   critical to differentiate patients with vascular dementia from AD
   patients.

Data availability statement

   Publicly available datasets were analyzed in this study. This data can
   be found here:
   [178]https://adknowledgeportal.synapse.org/Explore/Studies?Study=syn321
   9045, [179]https://www.synapse.org/#!Synapse:syn23277389, and
   [180]https://www.synapse.org/#!Synapse:syn3191087.

Author contributions

   Project was conceived by FL and CC. Transcriptomics datasets of AD
   samples were collected and pre-analyzed by AE, CS, OH, and CC.
   Methodology was developed and data were analyzed by FL. Results were
   discussed by FL, CC, AE, OH, EM, CK, and PP. The manuscript was
   prepared by FL and revised by CC, AE, WB, GW, CX, PP, EM, CK, and OH.
   All authors contributed to the article and approved the submitted
   version.

Footnotes

   ^1

   [181]https://www.synapse.org/#!Synapse:syn23277389
   ^2

   [182]https://www.synapse.org/#!Synapse:syn3191087
   ^3

   [183]https://www.synapse.org/#!Synapse:syn5550404
   ^4

   [184]http://broadinstitute.github.io/picard/
   ^5

   [185]https://www.gencodegenes.org/human/releases.html
   ^6

   [186]https://www.niagads.org/igap-rv-summary-stats-kunkle-p-value-data

Funding

   This work was partially supported by the National Institute of Ageing
   (NIA) R56AG065352 to Dr. Li. We thank all the participants and their
   families, as well as the many involved institutions and their staff,
   grants from the National Institutes of Health [R01AG044546 (CC),
   P01AG003991 (CC), RF1AG053303 (CC), RF1AG058501 (CC), and U01AG058922
   (CC)] and Chuck Zuckerberg Initiative (CZI), access to equipment made
   possible by the Hope Center for Neurological Disorders, and the
   Departments of Neurology and Psychiatry at Washington University School
   of Medicine. CC receives research support from: Biogen, EISAI, Alector,
   and Parabon. CC was a member of the advisory board of Vivid Genomics,
   Halia Therapeutics, and ADx Healthcare.

Conflict of interest

   The authors declare that the research was conducted in the absence of
   any commercial or financial relationships that could be construed as a
   potential conflict of interest.

Publisher’s note

   All claims expressed in this article are solely those of the authors
   and do not necessarily represent those of their affiliated
   organizations, or those of the publisher, the editors and the
   reviewers. Any product that may be evaluated in this article, or claim
   that may be made by its manufacturer, is not guaranteed or endorsed by
   the publisher.

Supplementary material

   The Supplementary Material for this article can be found online at:
   [187]https://www.frontiersin.org/articles/10.3389/fnagi.2022.935279/ful
   l#supplementary-material
   [188]Click here for additional data file.^ (321KB, docx)

References