Abstract

Background

   Alzheimer’s disease (AD) is the most common neurodegenerative disease,
   imposing huge mental and economic burdens on patients and society. The
   specific molecular pathway(s) and biomarker(s) that distinguish AD from
   other neurodegenerative diseases and reflect the disease progression
   are still not well studied.

Methods

   Four frontal cortical datasets of AD were integrated to conduct
   differentially expressed genes (DEGs) and functional gene enrichment
   analyses. The transcriptional changes after the integrated frontal
   cortical datasets subtracting the cerebellar dataset of AD were further
   compared with frontal cortical datasets of frontotemporal dementia and
   Huntingdon’s disease to identify AD-frontal-associated gene expression.
   Integrated bioinformatic analysis and machine-learning strategies were
   applied for screening and determining diagnostic biomarkers, which were
   further validated in another two frontal cortical datasets of AD by
   receiver operating characteristic (ROC) curves.

Results

   Six hundred and twenty-six DEGs were identified as AD frontal
   associated, including 580 downregulated genes and 46 upregulated genes.
   The functional enrichment analysis revealed that immune response and
   oxidative stress were enriched in AD patients. Decorin (DCN) and
   regulator of G protein signaling 1 (RGS1) were screened as diagnostic
   biomarkers in distinguishing AD from frontotemporal dementia and
   Huntingdon’s disease of AD. The diagnostic effects of DCN and RGS1 for
   AD were further validated in another two datasets of AD: the areas
   under the curve (AUCs) reached 0.8148 and 0.8262 in [29]GSE33000, and
   0.8595 and 0.8675 in [30]GSE44770. There was a better value for AD
   diagnosis when combining performances of DCN and RGS1 with the AUCs of
   0.863 and 0.869. Further, DCN mRNA level was correlated to CDR
   (Clinical Dementia Rating scale) score (r = 0.5066, p = 0.0058) and
   Braak staging (r = 0.3348, p = 0.0549).

Conclusion

   DCN and RGS1 associated with the immune response may be useful
   biomarkers for diagnosing AD and distinguishing the disease from
   frontotemporal dementia and Huntingdon’s disease. DCN mRNA level
   reflects the development of the disease.

   Keywords: Alzheimer’s disease, neurodegenerative diseases,
   bioinformatic analysis, machine learning strategies, biomarkers

Introduction

   Alzheimer’s disease (AD) is the most common neurodegenerative disease
   that is pathologically characterized by brain β-amyloid (Aβ) deposition
   forming extracellular plaques, Tau hyperphosphorylation aggregating to
   intracellular neurofibrillary tangles, and the progressive reduction of
   synapses and neurons causing brain atrophy ([31]Alzheimer’s
   Association, 2021). More and more studies also suggest that glial
   activation underlying neuroinflammation, brain glucose hypometabolism,
   mitochondrial dysfunction, and oxidative stress also participate in the
   occurrence and progression of the disease ([32]Liu et al., 2019;
   [33]Long and Holtzman, 2019). Yet, the pathogenesis of AD remains
   elusive and the ideal diagnostic biomarkers reflecting the outcome of
   the disease are still rare.

   Genome-wide studies have disclosed the complexity of networks in
   altered expressions of genes contributing to the pathogenesis of AD.
   The abnormal modules in immune and microglia-specific activities
   ([34]Zhang et al., 2013), neural communication, amyloid-β clearance
   ([35]Magistri et al., 2015), cerebral vasculature ([36]Magistri et al.,
   2015), and synaptic transmission ([37]Williams et al., 2021) have been
   indicated the involvement in the disease pathogenesis according to
   transcriptomics studies on postmortem brain samples of AD patients. The
   selective regional vulnerability of these involvements and the
   correlation with clinical outcomes further enhance the complexity
   ([38]Wang et al., 2016). It leads to significant difficulties in
   exploring the real pathogenic factor(s) and ideal diagnostic biomarkers
   for the disease. The recently emerging integrated bioinformatic
   analysis and machine learning strategy may help to address this
   challenge. AD, frontotemporal lobar degeneration (FTD), and
   Huntington’s disease (HD) are all classified into neurodegenerative
   diseases and they share common neurodegeneration and some clinical
   symptoms such as cognitive decline ([39]Walker, 2007; [40]Boeve et al.,
   2022), metabolic changes ([41]Vercruysse et al., 2018; [42]Moll et al.,
   2020), behavioral and psychological disorders ([43]Santacruz Escudero
   et al., 2019; [44]Menculini et al., 2021) and so forth, which causes
   clinically misdiagnosis and mistreatment ([45]Boeve et al., 2022).
   Accurate diagnostic differentiation also requires expensive PET or
   aggressive cerebrospinal fluid tests. Meanwhile, these diseases are
   heterogeneous in pathological mechanisms such as susceptibility of
   different types of synapses and neuronal cell death ([46]Kamat et al.,
   2016; [47]Smith-Dijak et al., 2019; [48]John and Reddy, 2021; [49]Wang
   et al., 2022), which provides the theoretical basis for identifying the
   differential biomarkers and AD specific pathogenesis. Previous studies
   have indicated that plasma phosphorylated tau 217 and phosphorylated
   tau 181 could differentiate AD from FTD ([50]Thijssen et al., 2020,
   [51]2021), while a few studies suggested potential biomarkers only on
   diagnosis of HD without reporting the differential effect from AD
   ([52]Wild et al., 2008; [53]Battaglia et al., 2011; [54]Caron et al.,
   2022). Therefore, analyzing the differences in gene expressions of AD,
   FTD, and HD may provide help for further understanding the pathogenesis
   and exploring new diagnostic biomarkers of AD differentiating from
   other neurodegenerative disease.

   In this study, we integrated four frontal cortical datasets from the
   GEO database into an integrated dataset which was used to be compared
   to one cerebellar dataset of AD, one frontal cortical dataset of FTD,
   and one frontal cortical dataset of HD, and the results were further
   verified in another two frontal cortical datasets of AD. We aim at
   discovering novel pathways and key genes that may serve as valuable
   diagnostic biomarkers for distinguishing AD from normal patients, FTD,
   and HD by bioinformatic analysis combined with machine learning
   strategies. Furthermore, the correlations between the biomarkers with
   the CDR (Clinical Dementia Rating scale) scores and Braak staging of AD
   patients were analyzed to explore whether they could also reflect the
   development of the disease.

Materials and methods

Data collection and data processing

   The microarray datasets referred to profiles of gene expressions with
   frontal cortical tissue of brain samples of AD patients were retrieved
   in National Center for Biotechnology Information (NCBI) Gene Expression
   Omnibus (GEO) database^[55]1 with the following keywords including
   “Alzheimer’s disease,” “Expression profiling by array” and the species
   was selected as “Homo sapiens.” Four datasets including [56]GSE5281 (10
   control and 23 AD samples), [57]GSE131617 (52 control and 19 AD
   samples), [58]GSE48350 (48 control and 21 AD samples), and [59]GSE84422
   (15 control and 24 AD samples) originated from frontal cortical samples
   with the diagnosis of AD, which are all from the same platform of
   Affymetrix. Based on the annotation from corresponding datasets, probes
   were transformed into gene symbols. And if there were more than one
   probe referring to the same gene symbol, the gene expression values
   would be assigned as an average value of them. The four AD frontal
   cortical datasets were normalized with the “SVA” package in R to
   compensate for the batch effect and merged into an integrated dataset
   ([60]Leek et al., 2020). The “SVA” package can help remove batch effect
   in two ways, including estimating surrogate variables for unknown
   sources of variation and directly removing known batch effects using
   the “Combat” function ([61]Leek and Storey, 2007, [62]2008). The
   integrated AD frontal cortical dataset subtracted the expressed genes
   in dataset [63]GSE44768 of AD cerebellum to screen frontal-associated
   genes. Then, the frontal-associated genes of AD were further compared
   with that in the FTD frontal cortical dataset [64]GSE13162 and HD
   frontal cortical dataset [65]GSE3790 to identify the candidates of
   diagnostic biomarkers for AD. More detailed information about the
   included datasets is summarized in [66]Supplementary Table 1.

Differential gene expression analysis

   Differentially expressed genes (DEGs) were analyzed with the “limma”
   package in R and identified with the thresholds that | log2FC| (fold
   change) was larger than 2 and the adjusted p-value was less than 0.05.
   After subtracting the DEGs of the AD cerebellar dataset and comparing
   them with that in FTD and HD frontal cortical datasets, DEGs only
   expressed in AD frontal cortex were identified. Besides, heatmaps and
   volcano plots were conducted with “pheatmap” and “EnhancedVolcano”
   packages in R. Venn diagrams were used to visualize the overlapping or
   unique genes among AD cortex, AD cerebellum, HD, and FTD ([67]Bardou et
   al., 2014).

Functional enrichment analysis

   Gene ontology (GO) enrichment analysis was performed under three
   hierarchical categories of biological process, molecular function, and
   cellular component with the “clusterProfiler” package in R ([68]Yu et
   al., 2012). Kyoto Encyclopedia of Genes and Genomes (KEGG) database was
   applied to do the pathway enrichment analysis with the
   “clusterProfiler” package in R ([69]Yu et al., 2012; [70]Yu and He,
   2016; [71]Minoru et al., 2017; [72]Jassal et al., 2020).

Machine learning for potential diagnostic biomarkers

   We employed three machine learning strategies to identify potential
   diagnostic biomarkers differentiating Alzheimer’s disease from other
   neurodegenerative diseases. Least absolute shrinkage and selection
   operator (LASSO) logistic regression was applied to identify the
   diagnostic genes associated with discrimination with the “glmnet”
   package in R ([73]Tibshirani, 1996; [74]Bühlmann, 2011). Support vector
   machine-recursive feature elimination (SVM-RFE) was conducted in R
   using the “e1071” package with 5-fold cross-validation ([75]Suykens and
   Vandewalle, 1999; [76]Weston and Guyon, 2012). Random forest (RF) was
   performed in R with the “randomForest” package ([77]Breiman, 2001;
   [78]Qi, 2012). Then, the overlapping genes of the above three
   machine-learning strategies were identified and would be further
   verified as candidate biomarkers.

Validation and evaluation of candidate biomarkers

   Two additional AD frontal cortical datasets ([79]GSE33000 and
   [80]GSE44772) were applied as the validation datasets. The overlapping
   genes with differential expression identified in the above three
   mentioned machine-learning strategies were validated. The receiver
   operating characteristic curves (ROCs) analysis was conducted and the
   area under the curves (AUCs) was calculated to evaluate the diagnostic
   efficacy of the selected biomarkers ([81]Seshan et al., 2013). Besides,
   CDR scale is an informant-based global clinical instrument which has
   been widely applied for clinical grading of dementia severity
   ([82]Schafer et al., 2004; [83]Huang H. et al., 2021). Based on the
   neurofibrillary tangle topographic distribution in the brain, the Braak
   staging is strongly associated with the degree of cognitive impairment
   ([84]Braak and Braak, 1991; [85]Nelson et al., 2012; [86]Lowe et al.,
   2018). Thus, the correlation analysis between the diagnostic biomarkers
   and the CDR scores and Braak staging was also performed in dataset
   [87]GSE84422 to explore whether they could reflect the development of
   the disease.

Results

Identification of differentially expressed genes in the frontal cortex of AD

   A diagram of the workflow of the bioinformatics analyses combined with
   machine learning strategies is shown in [88]Figure 1.

FIGURE 1.

   [89]FIGURE 1
   [90]Open in a new tab

   The workflow of the analysis process.

   After removing batch effects, the data of expression profiles of
   frontal cortical samples of AD from four GEO datasets ([91]GSE5281,
   [92]GSE131617, [93]GSE48350, and [94]GSE84422) were merged into an
   integrated dataset, including frontal cortical samples of 87 AD
   patients and 125 control subjects. Firstly, the DEGs in the frontal
   cortical samples of AD from the integrated dataset, cerebellar samples
   of AD ([95]GSE44768), frontal samples of FTD ([96]GSE13162), and HD
   ([97]GSE3790) patients were all analyzed. Secondly, the
   frontal-associated genes of AD were screened, and selected the DEGs in
   the frontal cortex of AD patients as compared with control subjects,
   while not in the cerebellum of AD patients as compared with control
   subjects. Then, the frontal-associated genes were further compared with
   those in FTD and HD datasets to finally identify the candidates of
   diagnostic biomarkers for AD. The results showed that 626 DEGs were
   AD-associated, including 580 downregulated genes and 46 upregulated
   genes. The Venn diagram showed the intersection of the DEGs from the
   above-mentioned datasets ([98]Figure 2A). The volcano plot described AD
   frontal associated gene distribution ([99]Figure 2B), and the heatmap
   illustrated the remarkable differences ([100]Supplementary Figure 1).

FIGURE 2.

   [101]FIGURE 2
   [102]Open in a new tab

   Identification of specific differentially expressed genes in the
   frontal cortex of Alzheimer’s disease. (A) Venn diagram showed the
   intersection of differentially expressed genes in frontal cortical
   samples of Alzheimer’s disease (AD), cerebellar samples of Alzheimer’s
   disease (CR), frontal cortical samples of frontotemporal lobar
   degeneration (FTD), and frontal cortical samples of Huntington’s
   disease (HD). (B) Volcano plot of frontal specific differentially
   expressed genes between Alzheimer’s disease and controls.

Functional enrichment analyses

   Based on the DEGs in AD frontal cortex, GO analysis was performed to
   explore the main cell functions, most of which were associated with the
   immune response and oxidative stress-related such as response to
   lipopolysaccharide, neutrophil activation, response to oxidative
   stress, reactive oxygen species metabolic process, etc. ([103]Figure 3A
   and [104]Supplementary Table 1). KEGG pathway analysis was also
   conducted and suggests significant enrichment in pathways of the
   cGMP-PKG signaling pathway, Peroxisome, Ras signaling pathway, etc.
   ([105]Figure 3B).

FIGURE 3.

   [106]FIGURE 3
   [107]Open in a new tab

   Functional enrichment analysis to investigate the potential function of
   differentially expressed genes (DEGs). (A) GO analyses of DEGs. (B)
   Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis of
   DEGs.

Screening and verification of the diagnostic markers

   Least absolute shrinkage and selection operator logistic regression
   algorithm identified 17 potential diagnostic markers from AD
   frontal-associated DEGs ([108]Figure 4A). SVM-RFE and RF analysis
   showed there were 25 and 30 potential diagnostic markers for AD,
   respectively ([109]Figures 4B, C). Among these potential biomarkers,
   there were two overlapping genes including decorin (DCN), and regulator
   of G protein signaling 1 (RGS1) ([110]Figure 4D).

FIGURE 4.

   [111]FIGURE 4
   [112]Open in a new tab

   Screening candidate biomarkers of Alzheimer’s disease via
   machine-learning strategies. (A) Least absolute shrinkage and selection
   operator (LASSO) logistic regression. (B) Support vector
   machine-recursive feature elimination (SVM-RFE) analysis. (C) Random
   forest (RF) analysis. (D) Venn diagram showed the intersection of
   diagnostic markers obtained by the three machine learning strategies.

   The differential expressions of DCN and RGS1 were further validated in
   two additional datasets [113]GSE33000 and [114]GSE44770, and the
   results suggested that both of them were significantly upregulated
   ([115]Figure 5 and [116]Supplementary Table 1). The ROC analysis of DCN
   and RGS1 was conducted to assess the diagnostic efficacy. The AUCs of
   DCN and RGS1 were 0.8148 and 0.8262 in dataset [117]GSE33000, and
   0.8595 and 0.8675 in [118]GSE44770 ([119]Figure 5E). The AUCs of the
   combination of the two biomarkers as a diagnostic tool for AD were
   0.863 and 0.869 ([120]Figure 5F). DCN mRNA level was correlated to CDR
   (Clinical Dementia Rating scale) score (r = 0.5066, p = 0.0058) and
   Braak staging (r = 0.3348, p = 0.0549) ([121]Figures 6A, B). RGS1 mRNA
   level correlates with Braak staging (r = 0.3141 p = 0.0675) and CDR
   score (r = 0.1184, p = 0.2908) ([122]Figures 6C, D).

FIGURE 5.

   [123]FIGURE 5
   [124]Open in a new tab

   Validating the differential expressions of diagnostic biomarkers in
   verification datasets. (A) Validation of the expression levels and
   diagnostic efficacy of DCN in dataset [125]GSE33000. (B) Validation of
   the expression levels of RGS1 in dataset [126]GSE33000. (C) Validation
   of the expression levels of DCN in dataset [127]GSE44770. (D)
   Validation of the expression levels of RGS1 in dataset [128]GSE44770.
   (E) ROC analysis of DCN, RGS1, and the combination of the two
   biomarkers as a diagnostic tool in the validation dataset
   [129]GSE33000. (F) Receiver operating characteristic (ROC) analysis of
   DCN, RGS1, and the combination of the two biomarkers as a diagnostic
   tool in the validation dataset [130]GSE44770.

FIGURE 6.

   [131]FIGURE 6
   [132]Open in a new tab

   Correlation analysis of the diagnostic biomarkers and the development
   of disease. (A) Correlation analysis of DCN and CDR score. (B)
   Correlation analysis of DCN and Braak staging. (C) Correlation analysis
   of RGS1 and CDR score. (D) Correlation analysis of RGS1 and Braak
   staging.

Discussion

   Cerebellum has always been viewed as a less affected or even neglected
   area of AD because autopsy studies found that the two main
   neuropathological hallmarks of AD including amyloid plaques and
   neurofibrillary tangles are both present first in the cerebral cortex
   while the cerebellum could escape from neurofibrillary tangles
   ([133]Larner, 1997). Although the emerging role of the cerebellum in AD
   has gained some attention, recent studies revealed that the genomic
   changes mainly occurred in the cortex instead of the cerebellum,
   especially excitatory neurons are preferentially vulnerable to
   pathological changes in AD ([134]Fu et al., 2019; [135]Leng et al.,
   2021; [136]Miller et al., 2022). Multiple cerebral regions have been
   reported to be related to AD ([137]Mu and Gage, 2011; [138]Aggleton et
   al., 2016; [139]Cogswell et al., 2021), among which the frontal cortex
   is more predictive of cognitive impairments and early disturbance of
   daily activities, which gives rise to more sensitivity to subclinical
   changes ([140]Stoeckel et al., 2013; [141]Marshall et al., 2019).
   Therefore, to identify biomarkers that are more related to the
   pathogenesis and clinical manifestations of AD and avoid interference
   from non-AD-associated genes, we selected DEGs in the frontal cortex of
   AD patients as compared with control subjects, while not in cerebellum
   AD patients as compared with control subjects. Those DEGs were defined
   as AD-associated genes.

   In the current studies, in order to solve the problem of the small
   sample size of existing individual databases originating from frontal
   samples of AD patients, we first integrated four homogeneous databases
   into an integrated database with a larger sample size. Then, the
   frontal-associated DEGs of AD were screened by comparing the
   differences in the DEGs between the integrated and cerebellar datasets
   ([142]Figure 2A). AD, HD, and FTD are neurodegenerative diseases with
   cognitive impairment, which clinically need differential diagnosis. To
   further identify the potential biomarkers for AD diagnosis, the
   frontal-associated DEGs and the DEGs from the datasets of FTD and HD
   frontal samples were analyzed. The 626 AD-associated DEGs were
   identified, including 580 upregulated genes and 46 downregulated genes
   ([143]Figure 2B). The functional enrichment analysis revealed that the
   immune response and oxidative stress-related processes were highly
   enriched such as neutrophil activation, immunological synapse
   formation, response to oxidative stress, reactive oxygen species, and
   so forth ([144]Figure 3). As mounting genetic and functional evidence
   indicates, neuroinflammation is a prominent manifestation of AD, and
   the immune response does matter a lot in the pathogenesis of the
   disease ([145]Gjoneska et al., 2015; [146]Shi and Holtzman, 2018). The
   immune response in AD is mobilized by the immune components in the
   central nervous system (CNS) mainly including complement and microglia,
   which function as a “double-edged sword.” Complement proteins not only
   act as the first line of defense to help synaptic pruning, mediate Aβ
   clearance and maintain neural circuits ([147]Schafer et al., 2012;
   [148]Hong et al., 2016) but associate with Aβ plaques causing the
   surrounding neuronal atrophy and synapse loss ([149]Hong et al., 2016;
   [150]Jevtic et al., 2017; [151]Shi et al., 2017). As one of the
   prominent immune cells of the CNS, microglia rapidly respond to stimuli
   and pathogens performing immune surveillance, phagocytosis, and
   neuroprotection ([152]Streit et al., 1999; [153]Davalos et al., 2005;
   [154]Nimmerjahn et al., 2005). However, the inflammatory cytokines and
   oxidative stress produced by microglia also lead to neural impairment
   ([155]Heneka et al., 2013; [156]McIntosh et al., 2019;
   [157]Marschallinger et al., 2020; [158]Hashioka et al., 2021). Recent
   studies also suggest that oxidative stress causes neuronal damage in
   various pathways, and plays a critical or even center role pathogenesis
   and mechanism of AD ([159]Jiang et al., 2016; [160]Bai et al., 2022).
   Oxidative stress promotes the expression of amyloid precursor protein
   (APP) and upregulates the activity of β-secretase contributing to the
   aggregation of Aβ ([161]Tamagno et al., 2012; [162]Butterfield and
   Boyd-Kimball, 2018). Oxidative stress can also enhance the Tau
   phosphorylation by producing reactive oxygen species (ROS) and directly
   interacting with glycogen synthase kinase-3 (GSK3) ([163]Su et al.,
   2010; [164]Du et al., 2022).

   Most importantly, DCN and RGS1 were identified to be useful diagnostic
   biomarkers by integrated bioinformatic analysis and machine-learning
   strategies. The diagnostic efficacies of DCN and RGS1 were verified by
   ROC analysis with the AUCs of 0.8148 and 0.8262 in dataset
   [165]GSE33000, and 0.8595 and 0.8675 in [166]GSE44770. And the
   combination of the two biomarkers as a diagnostic tool for AD owned
   even higher AUCs of 0.863 and 0.869 in the validated datasets.
   Furthermore, the correlation analysis indicated that DCN was
   significantly correlated to the CDR score and may be correlated to the
   Braak stage, and the RGS1 showed a significant correlation to the Braak
   stage. The CDR is one of the most widely used global clinical rating
   scales to evaluate the development of disease ([167]Morris, 1993).
   Braak staging is based on the neurofibrillary tangle topographic
   distribution in the brain and is strongly associated with cognitive
   impairment ([168]Braak and Braak, 1991). Therefore, DCN and RGS1 may
   reflect the development of the disease.

   Decorin is an extracellular matrix proteoglycan associated with
   collagen fibril formation ([169]Reed and Iozzo, 2002). Recent
   researches have shown that it also plays a critical role in autoimmune
   and inflammatory diseases ([170]Dong et al., 2022), antifibrotic
   ([171]Järvinen and Ruoslahti, 2019), antioxidant, and antiangiogenic
   properties ([172]Sofeu Feugaing et al., 2013; [173]Järveläinen et al.,
   2015). DCN was found significantly increased in both AD mouse models
   and CSF of AD patients, predicting well innate immune activation and
   potential choroid plexus dysfunction ([174]Jiang et al., 2022). Other
   studies also demonstrated that endothelial-activated DCN-positive
   astrocytes contributed to vascular amyloid deposits but not parenchymal
   amyloid plaques in AD mouse models and AD/CAA patients ([175]Taylor et
   al., 2022). Our study here confirmed that DCN was significantly
   increased in AD brain parenchyma at the mRNA level and could perform as
   a useful diagnostic marker. Future study is still needed to clarify the
   source and function of AD pathogenesis.

   RGS1 is one of the members of the RGS family and is identified as an
   immediate early gene responding to B cell activation signals
   ([176]Liang et al., 2018). RGS1 serves as a negative regulator to
   inhibit the chemotaxis of cytotoxic T lymphocytes and TH1 cells toward
   tumor-associated chemokines ([177]Huang D. et al., 2021). It was also
   found to regulate the homeostasis and trafficking of B and T cells to
   inflammatory diseases and attract macrophages to atherosclerotic
   plaques ([178]Patel et al., 2015). At present, research on the
   association between RGS1 and AD is limited. Only one study has shown
   that RGS1 was upregulated in the PBMCs of AD patients ([179]Leandro et
   al., 2018). Therefore, RGS1 is expected to be a blood-based diagnostic
   marker, and the role of RGS1 in AD is worthy of further study.

   There are some limitations to our study. First, although we carefully
   select the same frontal lobe of brain samples of AD from the same
   platforms of Affymetrix and have validated them in another two
   datasets, the results still need to be thoroughly confirmed
   experimentally. Second, limited by the sample size, the availability of
   the biomarkers in differentiating from other neurodegenerative diseases
   and reflecting the development of the disease needs further clinical
   exploration. Third, lacking appropriate datasets of the frontal cortex
   with a diagnosis of HD and FTD from the same platform as the selected
   datasets of AD, we failed to further evaluate the differential effect
   of biomarkers distinguishing AD from HD and FTD. We would like to
   update this evaluation, once there were appropriate datasets.

Conclusion

   We applied an approach of integrated bioinformatic analysis combined
   with machine learning strategies to identify AD frontal-associated
   biomarkers differentiating from normal subjects, FTD and HD. The
   results revealed that DCN and RGS1 associated with immune response are
   useful biomarkers for diagnosing AD and distinguishing from FTD and HD,
   and their mRNA levels may also reflect the development of the disease.

Data availability statement

   The original contributions presented in this study are included in the
   article/[180]Supplementary material, further inquiries can be directed
   to the corresponding authors.

Author contributions

   SS and CZ designed the study and accessed the funding. BJ downloaded
   the data and perform bioinformatics analysis. SS, GF, and BJ wrote the
   manuscript. All authors contributed to the article and approved the
   submitted manuscript.

Funding Statement

   This study was supported by grants from the National Natural Science
   Foundation of China (81901081, 81870822, 91332201, and 82171408),
   Shanghai Municipal Science and Technology Major Project, and the
   Natural Science Foundation of Fujian Province (2020CXB049).

Footnotes

   ^1

   [181]https://www.ncbi.nlm.nih.gov/geo/

Conflict of interest

   CZ holds shares of Shanghai Raising Pharmaceutical Co., Ltd., which is
   dedicated to developing drugs for the prevention and treatment of AD.
   SS was employed by the company Shanghai Raising Pharmaceutical
   Technology Co., Ltd. The remaining 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:
   [182]https://www.frontiersin.org/articles/10.3389/fnmol.2023.1152279/fu
   ll#supplementary-material
   [183]Click here for additional data file.^ (493.2KB, PDF)
   [184]Click here for additional data file.^ (175.9KB, XLSX)
   [185]Click here for additional data file.^ (280.4KB, XLSX)

References