The Landscape of m6A Regulators in Multiple Brain Regions of Alzheimer’s Disease

Alzheimer’s disease research has been conducted for many years, yet no effective cure methods have been found. N6-methyladenosine (m6A) RNA methylation, an essential post-transcriptional regulation mechanism, has been discovered to affect essential neurobiological processes, such as brain cell development and aging, which are closely related to neurodegenerative diseases such as Alzheimer’s disease. The relationship between Alzheimer’s disease and the m6A mechanism still needs further investigation. Our work evaluated the alteration profile of m6A regulators and their influences on Alzheimer’s disease in 4 brain regions: the postcentral gyrus, superior frontal gyrus, hippocampus, and entorhinal cortex. We found that the expression levels of the m6A regulators FTO, ELAVL1, and YTHDF2 were altered in Alzheimer’s disease and were related to pathological development and cognitive levels. We also assessed AD-related biological processes influenced by m6A regulators via GSEA and GSVA method. Biological Processes Gene Ontology terms including memory, cognition, and synapse-signaling were found to potentially be affected by m6A regulators in AD. We also found different m6A modification patterns in AD samples among different brain regions, mainly due to differences in m6A readers. Finally, we further evaluated the importance of AD-related regulators based on the WGCNA method, assessed their potential targets based on correlation relationships, and constructed diagnostic models in 3 of all 4 regions using hub regulators, including FTO, YTHDC1, YTHDC2, etc., and their potential targets. This work aims to provide a reference for the follow-up study of m6A and Alzheimer’s disease.


Introduction
Alzheimer's disease was first reported by Dr. Alzheimer more than 100 years ago [1], yet researchers are still trying to find a cure and it remains a huge burden on society. Population aging is making Alzheimer's disease (AD) one of the greatest challenges globally [2]. The diagnosis of Alzheimer's disease is complex. From the earliest NINCDS-ADRDA criteria [3] to subsequent revisions [4], the unique diagnosis pipeline reflects the specificity of the disease. Patients should first be diagnosed with dementia and then require a comprehensive evaluation, including assessment of the levels of biomarkers and the presence of pathological damage to be diagnosed with AD. There are various hypotheses about the pathological cause of Alzheimer's disease. The NINCDS-ADRDA guideline includes Aβ, neurofibrillary tangles, tau protein phosphorylation, and neuritic plaques as pathological diagnostic markers [5], which are also the most recognized hypotheses for Zi Jie Liu and Qing Xia contributed equally to this work. the causes of neuropathology and targets of current drug development research [6]. Last year, the Aβ monoclonal antibody drug aducanumab became the first FDA-approved drug targeting the neuropathological cause of Alzheimer's disease [7]. Although the results of aducanumab to improve the cognitive status of patients are still controversial, it does represent an advance in the concept of AD treatment study from symptomatic treatment to targeting neuropathology causes. However, the failure of many other neuropathology-targeting drugs [8,9] also suggests the need to research the pathological mechanism and drug targets of Alzheimer's disease from different perspectives.
N6-methyladenine (m6A) modification is the most abundant epigenetic modification of eukaryotic mRNA. The m6A mechanism has been reported to participate in the regulation of various mRNA behaviors, including translation [10], degradation [11], splicing [12], and nuclear export [13]. Regulators of m6A that have been discovered to date are divided into 3 types: writers, erasers, and readers. The m6A modification is reversible and dynamic and has been reported to play an indispensable role in the brain, which must execute complex processes that require a flexible regulation method. There have been several recent studies related to m6A and brain function: whole transcriptome sequencing of methylated RNA performed in the mouse medial prefrontal cortex (mPFC) revealed that m6A is involved in the consolidation of fearrelated memory [14]. Me-RIP sequencing of mouse forebrain synapses also suggested the essential role of m6A modification in synaptic functioning [15]. FTO, a m6A eraser, has also been reported to be essential for memory formation in the hippocampus [16]. There are also studies reporting a relationship between m6A modification and cell senescence through several genes, including p21 [17], p53, p27 [18], and AGO2 [19]. Considering the relationship between m6A modification, brain development, and aging, there may be important links between m6A and Alzheimer's disease.
There are already several studies on the relationship between m6A regulators and Alzheimer's disease, including METTL3 [20,21] and FTO [21][22][23], without considering their function as m6A regulators. The reader hnRNPG has been reported to be associated with frontotemporal dementia, another neurodegenerative disease that represents tauopathy [24]. There are also recent studies about m6A modification participating in processes of brain development and aging [25], which are closely related to Alzheimer's disease. A study also suggested that m6A modification profiles are altered in the brains of an Alzheimer's disease mouse model [20,25]. Although studies have suggested a relationship between Alzheimer's disease and the m6A mechanism, a detailed mechanism between m6A regulators and AD in the human brain is still absent.
In this study, we systematically evaluated the relationship between m6A regulators and AD and found potential target genes of hub m6A regulators that might affect disease progression and pathology among 4 different brain regions. We also assessed the validity of the relations between regulators and targets by constructing a diagnostic model (Fig. 1). We hope this research can provide a preliminary understanding of AD etiology research and drug development from an epitranscriptome perspective.

Data Collection
AD transcriptome datasets were obtained from the Gene Expression Omnibus (GEO) database (Supplementary Table 1). GSE48350 was used for the primary analysis, containing 253 samples of 4 brain regions (postcentral gyrus, superior frontal gyrus, hippocampus, and entorhinal cortex). Samples from patients under the age of 60 were not used in the study to balance the age difference between the AD group and the control group. Part of the GSE28146 (30 samples from the hippocampus divided into four groups due to Braak stages) and GSE118553 (98 samples from the entorhinal cortex and 115 samples from the frontal cortex divided into four groups due to Braak stages) datasets were used as validation datasets. The frontal cortex (63 samples) and hippocampus (55 samples) data from GSE84422 were also included as validation datasets. The R package GEOquery [26] was used to obtain the array matrix and detailed sample information.

Screening AD-Related Regulators
In this study, we considered regulators that were differentially expressed in AD groups and genes related to Braak stages and Mini-Mental State Examination (MMSE) scores as AD-related regulators. The R package limma [31] was used to identify the DEGs between the AD and control groups. Limma is an R package commonly used in transcriptome analyses to screen for differentially expressed genes (DEGs) based on linear models. In this study, genes with p-value ≤ 0.05 were considered DEGs. The R package Hmisc was used to analyze the Spearman correlation between regulators and Braak stages [32] and MMSE scores [33], and correlations with p ≤ 0.05 were considered significant. Braak stages are established based on neurofibrillary tangles, which contain stages I to VI, and a higher stage indicates more severe disease. MMSE scores indicate the cognitive level, and higher scores indicate better cognition levels.

Correlation Analysis Among Regulators
The R package Hmisc was used to analyze the Pearson correlation between m6A regulators in different brain regions. Correlations with p ≤ 0.005 were considered significant. The R package corrplot was used for visualization.

Evaluation of the Alteration of Biological Processes
The alteration of genes in biological processes was analyzed through GSEA method using the GSEA function from the clusterProfiler R package [34,35]. GSEA analysis was conducted using the fold change ranking obtained in previous DEG analysis. And the correlation between m6A regulators and Biology Process GO (GO:BP) terms were analyzed by analyzing the Pearson correlation between the transcript level of regulators and GSVA scores of GO:BP terms (p ≤ 0.05 and r ≥ 0.25). GSVA was conducted using the expression data in different brain regions. GSVA [36] is a GSE method that can evaluate the activity of gene sets in an unsupervised manner and can be used to estimate geneset activity. The GO:BP term gene sets were downloaded from the MSigDB database using msigdbr package.

Consensus Clustering
To evaluate whether there were different m6A modification patterns in AD samples, the r package ConsensusCluster-Plus [37] was used to perform consensus clustering in AD samples from the 4 brain regions. Consensus clustering is a method for evaluating the adequate number and robustness of clustering using iterative subsampling and clustering.

WGCNA
A weighted coexpression network was constructed using the WGCNA [38] package. Significant regulators in the network (GS ≥ 0.2 and MM ≥ 0.65 with a module) and significant modules (p ≤ 0.05 for phenotype correlation) were analyzed. WGCNA is a method to weigh the coexpression network so that the whole can be aggregated into different modules, and then the genes and phenotypes can be studied through the modules [39]. The threshold parameter (beta) to construct scale-free weighted networks was determined using the pickSoftThreshold function from WGCNA package. And the minimum threshold allowing R^2 ≥ 0.8 (Supplementary Figure 1) was used to obtain high-confidence scale-free networks in PG region (beta = 4), SFG region (beta = 4), and EC region (beta = 4). In the HP region, as the threshold when R^2 reaching 0.8 was too high (above 18), we chose beta = 3 considering both R^2 and mean connectivity of the WGCNA network (Supplementary Figure 1). The hub genes (GS ≥ 0.2, MM ≥ 0.65) in the corresponding modules of key regulators were selected and considered important regulatory relationship pairs for the disease, and follow-up analysis was carried out.

Potential m6A-Regulated Genes and Confidence Evaluation
For the potential regulator-gene regulation relationship discovery, the Pearson correlation between regulators and these genes was measured, and correlations with p ≤ 0.05 and r ≥ 0.25 were considered potential targets. For these relations, we assessed their confidence in the following aspects: whether the gene transcripts were m6A or m6Am modified; if there was evidence that the regulators could bind the target transcripts or if perturbation of the regulators could affect the target genes in transcript abundance, m6A modification level, and translation level. The m6A2Targets database [40] was used to determine the regulators' perturbation effects and binding potential was evaluated from high-throughput sequencing data. The human m6A modification profile was obtained from Liu et al.'s work [41]. The supplementary tables that contain information about m6A-or m6Am-modified genes were downloaded for the research.

Lasso Regression
The expression data of key regulators and their potential targets were used to fit a LASSO penalized logistic model to diagnose AD. LASSO [42] penalized logistic regression was fitted by the glmnet package in R, and the penalty coefficient of log(λ) minimizing the binomial deviance was chosen. To retain more variables, log(λ)−1 was used as the penalty coefficient to construct the final model. For the constructed model for each brain region, the same-brain region data of GSE28146 and GSE118553 were used as the validation dataset.

AD-Related m6A Regulators Vary Among Different Brain Regions
To evaluate whether m6A regulators are related to AD, we performed gene differential expression analysis between AD and healthy cases in different brain regions ( Fig. 2A , p values shown in Supplementary Table 2). The differential Fig. 2 Identification of AD-related regulators. A Heatmap of m6A regulator expression and changes among 4 brain regions. B Heatmap of correlations between m6A regulators, Braak stages, and MMSE scores (*p<0.05; **p<0.01; ***p<0.001) expression of regulators varied among different brain regions. In the postcentral gyrus region, we found that m6A writer RBM15 was downregulated in the AD group. In the superior frontal gyrus, m6A reader protein HNRNPNPA2B1 was significantly upregulated in AD. Four regulators of different types were found to be differentially expressed in the hippocampi of AD patients: one writer (RBM15, downregulated in AD group), two readers (ELAVL1, upregulated in AD group; YTHDF2, downregulated in AD group), and one eraser (FTO, downregulated in AD group) in the hippocampus. In the entorhinal cortex region, two readers (YTHDC1 and YTHDC2) were upregulated. Only RBM15 was differentially expressed in multiple brain regions (PG and HP), and were downregulated in AD group in both regions. In addition to the hippocampus, the other three brain regions had an obvious tendency towards certain regulator types. In the PG region, m6A writers tended to be downregulated in the AD group. Reader protein tended to be upregulated in the AD group in the SFG region. And m6A readers showed upregulated tendency in the EC region.
In addition to DEG exploration, we also assessed which regulators might influence the pathological progression of AD. We investigated the relationship between the expression levels of m6A regulators and cognition level (MMSE score) and neuropathology progression stage (Braak stage). Multiple regulators were considered to affect the cognitive level or pathological progression among different brain regions (Fig. 2B , p values shown in Supplementary Table 3). For erasers, the level of the ALKBH5 transcript in the PG region was found to be associated with both cognitive and neuropathological progression and negatively correlated with disease progression. For readers, the expression levels of YTHDF2 in both HP and EC regions were correlated with pathological progression and tended to have opposite effects. Among writers, the levels of HAKAI were negatively correlated with Braak score in HP region and MMSE score in EC regions, respectively. Surprisingly, there were no genes with both alteration in expression level in AD patients and also correlated with pathology and cognition other than YTHDF2 in the HP region, which may indicate the different role of m6A regulators between the occurrence and development of AD.
Considering that the interactions between regulators are very common, we also measured the correlation between regulators in different brain regions (p ≤ 0.005, Fig. 3A-D, Supplementary Table 4). Correlations between regulators were common in all brain regions and were mostly positive. However, the correlation status between different types of regulators varied in different brain regions. In the PG region, the correlations between writers and readers were extensive, while erasers were only related to readers. In the SFG region, writers and readers were still commonly correlated. There was an increase in the number of correlations among writers compared to the PG region, while the correlation between readers decreased. Erasers were not only related to readers; there was a correlation between FTO and a writer, HAKAI. The interaction profile in the HP region was similar to that in the SFG region, presenting a strong correlation between writers and writers. Although the correlation between readers in the HP region was weaker, ALKBH5 (eraser) was strongly related to m6A writers. In the EC region, the correlations among regulators were the weakest, but the erasers represented a strong correlation with writers and readers, which might indicate a special role of the m6A eraser. In conclusion, the correlation of regulators varies in different brain regions. Considering the different functions of each brain region, this may be due to different pathogenic mechanisms of different types of m6A regulators.

m6A Regulators Influence Multiple Essential Biological Processes
To explore which biological processes may be affected by AD-related m6A regulators, we used the GSEA method to measure the alteration of genes in GO:BP terms in AD. Then we computed the GSVA score of these terms and assessed the correlation between these terms and m6A regulators (Fig. 4, Supplementary Fig. 2, Supplementary Table 5-6).
GSEA analysis showed changed GO:BP terms in AD group in different brain regions (p ≤ 0.05, 535 in PG, 610 in SFG, 606 in HP, and 734 in EC, Supplementary Table 5). Among them, 55 GO:BP terms were altered in all 4 brain regions, including terms related to memory formation, synaptic interaction, and protein catabolic processes (Supplementary Fig. 2A-B). We also evaluated which terms were potentially regulated by m6A regulators related to AD pathology (DEGs and genes associated with Braak stage and MMSE score) in the corresponding brain regions. The Pearson correlations between regulators and the GSVA scores of GO:BP terms were calculated. Among the 55 terms, 16 were found to be associated with at least one of the diseaserelated regulators in all brain regions (p ≤ 0.05, Fig. 4A, Supplementary Fig. 2C-F, Supplementary Table 6). The terms included several GO:BP terms closely related to AD, including terms GOBP_COGNITION and GOBP_MEM-ORY, and terms related to synaptic signaling. And the term GOBP_REGULATION_OF_GENE_EXPRESSION_EPI-GENETIC indicated the important roles of m6A regulators in epigenetic regulation of above terms (Fig. 4B).
Regulators closely related to most of these terms in 4 regions were most readers, including YTHDF2, YTHDF3, IGFBP2, ELAVL1, and HNRNPA2B1. While in the PG region, eraser ALKBH5 was found to be correlated with 12 of all 16 terms. And another eraser FTO was also found to play an important role in the HP region. The m6A reader METTL3 was related to 14 of all 16 terms. Considering that ELAVL1 played important roles in 2 of the 4 brain regions, there might be an important relationship between ELAVL1 and AD.

Identification of Different m6A Modification Patterns in AD Patients
To investigate whether there were different patterns in AD patients' m6A profiles, we performed consensus clustering using the m6A regulator expression profile of AD patients in different brain regions to discover an adequate number of groups. The clustering results among the 4 brain regions were divided into 2 patterns. The PG and SFG regions were divided into 3 clusters, while the HP and EC regions were divided into 2 clusters (Fig. 5, Supplementary Fig. 3, Supplementary Table 7). To explore the reason for the above groupings, we assessed the differentially expressed regulators in different clusters of each brain region and evaluated similarities and the differences between regions (Supplementary Table 8). For the 3 clusters in each PG region and SFG region, 8 regulators (HNRNPA2B1, HNRNPC, IGFBP2, IGFBP3, YTHDC1, YTHDC2, YTHDF1, and ZC3H13) had similar expression patterns (Fig. 5B, D). For the 2 clusters in each of the HP and EC regions, HNRNPA2B1 and IGFBP3 had similar expression patterns (Fig. 5F, H). The above regulators were all readers except ZC3H13. To explore the biological differences in the different modification patterns, we conducted GO term enrichment analysis to DEGs between different m6A modification patterns in each brain region (p ≤ 0.05, Fig. 6A-D). In the PG region, three groups shared 40 DEGs (Fig. 6A), and were enriched in GO terms related to extracellular matrix organization, which plays an important role in AD etiology. In the SFG region, 463 DEGs were shared among three m6A modification patterns (Fig. 6B), and showed enrichment in multiple terms related to amino acid, lipid, and glycolipid process. In the HP region, DEGs were enriched in terms related to neuron activity, like synapse development and axon development (Fig. 6C). In the EC region, DEGs also showed enrichment in terms related to extracellular matrix organization (Fig. 6D).

Evaluation of the Regulatory Importance of m6A Regulators Based on WGCNA
To assess the importance of AD-related m6A regulators in the regulatory network, we used the WGCNA method to construct a weighted coexpression network for each brain region. The importance of the regulators was evaluated based on the gene significance (GS) of the AD-control group and module membership (MM) (Fig. 7), representing the regulators' relation with disease and gene modules, respectively. Regulators with GS ≥ 0.2 and MM ≥ 0.65 for a module were selected, and their GS values with all phenotypes and MM values with all modules are shown in Supplementary Fig. 4. Several important regulators related to the AD/control phenotype were found (PG: FTO; SFG: RBM15B and FTO; Hippocampus: RBM15, FTO, YTHDF2, YTHDC1, and IGFBP1; EC: YTHDC1 and YTHDC2). To assess the disease-related biological processes that might be modulated by the above regulators, we evaluated whether these hub regulators were related to modules related to the AD/ control phenotype (p ≤ 0.05, |r| ≥ 0.35). After the assessment, the PG region did not obtain disease-related modules related to essential regulators. In the SFG region, RBM15B was found to be related to the grey60 module, and FTO was related to the blue module. In the HP region, RBM15, YTHDC1, and IGFBP1 were related to the turquoise module, while FTO and YTHDF2 were related to the blue module. In the EC region, YTHDC1 and YTHDC2 were found to be related to the light cyan and dark red modules, respectively. The hub genes (MM ≥ 0.65 and GS ≥ 0.2) in each corresponding module were selected for the following analysis. We selected the hub genes correlated with the corresponding regulators above (p ≤ 0.05) and obtained potential relationships between regulators and hub genes. We also evaluated the reliability of the possible regulatory relation based on the Me-RIP seq analysis result of the human brain from Liu et al.'s work [41] and the m6A2targets database. After evaluation, we obtained reliable regulation relations between regulators, disease-related gene modules, and target genes in the corresponding modules that correlated with AD-related genes in each brain region (Supplementary Fig. 5-7). Of these genes, several genes, including TMEM59L, RTN4RL2, SNCB, etc., have been reported to affect AD pathology in past studies [43][44][45][46] and are potentially regulated by regulators, including FTO. In conclusion, we obtained key AD-related regulators and regulatory relationships with their target genes that may have a decisive impact on AD.

Construction of the Diagnosis Model Using Hub Regulators and Related Genes
For the hub regulators and their potential targets obtained in the previous step, we tried to use the above genes to fit diagnostic models to evaluate the importance and relationship with AD in each brain region. Specifically, the hub regulators and genes related to corresponding regulators (p ≤ 0.005) were used to fit diagnostic models, which were disease-related and highly correlated with disease-related modules. We performed Lasso penalized logistic regression to construct the diagnostic model in each brain region. Then, the diagnostic model was validated using the validation dataset.
In conclusion, we obtained different diagnostic models in the regions above (Fig. 8, Supplementary  Table 9). No diagnostic model was constructed in the PG region because no key regulators or target modules were found in this region. For the validation dataset, we also constructed a diagnostic model using the genes from the penalized model obtained from the training dataset of the corresponding brain region. For the SFG region, RBM15B, FTO, and a total of 55 related genes were used to fit the model. With log(λ)=−3.9, the model consisted of 13 genes, and the AUC was 0.958 ( Fig. 8A-B). We used the frontal lobe data of GSE118553 and the frontal cortex data GSE84422 as validation datasets and obtained AUC of 0.98 and 0.735, respectively (Fig. 8A-B). For the HP region, we used RBM15-, FTO-, YTHDF2-, YTHDC1-, IGFBP1-, and 4524-related genes for penalized regression. With log(λ)=−6.1, the final model consists of 21 genes, and the AUC was 1 (Fig. 8C-D). The AUCs of the hippocampal validation datasets GSE28146 and GSE84422 were 1 and 0.853, respectively (Fig. 8C-D). For the EC region training dataset, we used YTHDC1 and YTHDC2, and a total of 33 related genes for model fitting. With log(λ)= −4.9, the model consisted of 8 genes, and the AUC was 1 (Fig. 8E-F). The entorhinal cortex data of GSE118553 were used for validation and had an AUC of 0.731 (Fig. 3F). In conclusion, highly reliable diagnostic models for the SFG, HP, and EC regions were obtained, indicating the consistency of the relationship between AD, regulators, and their potential targets.

Discussion
m6A RNA methylation is a very important epigenetic regulatory mechanism, and its role in neurodegenerative diseases is still under investigation. Although there have been some reports about the potential roles of m6A modification and regulators in AD, no studies have discussed the different mechanisms of regulators from a multi-region perspective. Here, we revealed key m6A regulators and their influence on AD in 4 brain regions through transcriptome mining and further identified important regulators and their potential targets based on weighted coexpression network analysis. We hope these differences and similarities in m6A modification between the brain regions in AD can provide guidance for future studies.
Among the 4 brain regions investigated in this study, the superior frontal gyrus, hippocampus, and entorhinal cortex are regions reported to be pathologically invaded in AD progression, while the postcentral gyrus is relatively less pathologically affected [32]. However, a study also showed changes in aging-related genes in the postcentral gyrus [47]. We first explored regulators that might affect AD through a variety of means. Differential expression analysis is one of the most common methods used in transcriptome research. We found that regulators obtained by DEG analysis varied among different brain regions. Although the changes varied among different kinds of regulators in the postcentral gyrus, superior frontal gyrus, and entorhinal cortex, their changes each focused on certain types of m6A regulators. Only the hippocampus showed different trends in regulator types, so there may be a more complex pathological mechanism of the m6A regulators in the hippocampus, which needs further investigation. Braak staging and MMSE scores are commonly used as pathological progression and cognition indicators in AD, and were also considered here to evaluate the relationship between m6A regulators and AD neuropathology. As a result, we found multiple regulators were found related to AD pathology in the postcentral gyrus, while none of which other than YTHDF2 were found differentially expressed between AD and control group. It may be due to the different roles of m6A regulators between occurrence and development of AD. It should be further studied to determine whether it is related to the early stage of AD or disease state transition.
We then assessed which genes were changed in GO:BP terms in each brain region using the GSEA, and evaluated their relationship with m6A regulators using the GSEA and GSVA method. The GSEA method focus on a set of gene, and can gain insights into alteration of biological mechanisms. The GSVA method has been commonly used in cancer research to study biological processes changes in diseases, which use a pathway-based dimensionality Fig. 5 Identification of different modification patterns in AD samples among four brain regions. A Heatmaps of consensus clustering results based on m6A regulator profiles in the PG region, B the SFG region, C the HP region, D and the EC region. E The m6A regulators that have significant changes among different modification patterns (p ≤ 0.1) in the PG region, F the SFG region, G the HP region, and H the EC region (*p<0.05; **p<0.01; ***p<0.001) ◂ reduction method to improve the biological interpretability of transcriptome data in bioinformatic research.
We obtained biological processes that were commonly changed across brain regions and were potentially affected by AD-related m6A regulators. We found that AD-related processes were regulated by m6A regulators including terms related to memory, cognition, and synaptic signaling. The influence of m6A modification on memory formation has already been reported before, and is indeed an important part in AD [14,16]. But the impact of m6A regulators on memory in AD has not been reported yet, and we assume it is worth exploring basing on our analysis. The role of m6A regulators in synaptic function has already been found using MeRIP-seq sequencing in mice [15]. Here, we found genes related to synaptic function also altered in AD patients and related to m6A regulators.
During medical researches, high-throughput data have become very helpful for understanding genes that may affect diseases. Tools and data have come out to help researchers collect data about their corresponding research fields [40,41,48,49]. In this study, we used m6A2Targets [40] database and Liu et al.'s research [41] to help narrow the target genes of m6A regulators. For regulators found to be correlated with these biological processes, ELAVL1, HNRNPA2B1, IGFBP2, YTHDC2, YTHDF2, and RBM15 were found to play important roles in different brain regions. Some regulators have been reported to have an impact on AD or related pathology. HNRNPA2B1 was recently reported to affect the endocytosis of tau in microglial cells [50]. FTO has also been reported to be associated with AD [50]. Links between AD pathology and other regulators have not been reported and might need further study. Several regulators have been reported to be associated with neurological functions. FTO was reported to be associated with memory [51] and neurogenesis [52]. YTHDF2's relationship with neurogenesis has also been reported [53], so their alteration might represent an underlying cause for neurodegeneration.
We also aimed to determine whether there were specific m6A modification patterns in different brain regions of AD patients. The consensus clustering method was used to provide a way to find a suitable clustering coefficient k by iterative clustering, which gives the adequate number of m6A clusters in each brain region. Similar work has also been conducted in several diseases, such as bladder cancer [54] and pancreatic cancer [55]. In this study, the number of modification patterns obtained in different brain regions differed, and the related biological processes also varied, but several biological processes were found to be commonly affected by m6A regulators. Samples of both the postcentral gyrus and superior frontal gyrus were divided into three clusters, while samples of the hippocampus and entorhinal cortex were divided into two clusters. The expression patterns of several m6A readers were similar in regions that were divided into the same number of clusters, mainly readers.
In the PG region, DEGs among different modification patterns were enriched in GO terms related to extracellular matrix organization, which related with APP protein that is the core proteins of current AD pathological hypotheses [56]. In the SFG region, DEGs among different modification showed enrichment in multiple terms related to lipid processing. It has been reported that an allele of apolipoprotein E (APOE) is a risk factor for familial AD [57]. Our analysis suggests that the alteration of m6A regulators may have an impact on the lipid metabolism process, which is related to AD. In the HP region, DEGs between different modification were enriched in terms related to neuron activity, like synapse development and axon development, closely related to AD. In the EC region, DEGs between different modifications also showed enrichment in terms related to extracellular matrix organization. Our analysis showed that there may be different modification patterns in AD patients' different brain region, which may have effects on these biological processes related to AD.
We also attempted to identify the hub regulators related to AD and assess their association with biological processes and target genes using the WGCNA method. WGCNA has been widely used in studies of a variety of diseases [54,55]. The hub regulators related to AD were selected based on their correlations with the disease phenotypes and modules significantly related to AD. No regulators were screened out in the PG region, which may be associated with the milder influence of AD pathology in the PG region. Two regulators each in the SFG region (RBM15B and FTO) and EC regions (YTHDC1 and YTHDC2) were considered important, while multiple regulators were screened out in the HP region, including In each column, the figures above show the correlation between module and phenotype, and the figures below show the p value of the correlation RBM15, FTO, YTHDF2, YTHDC1, and IGFBP1. Among these regulators, FTO, YTHDC1, and YTHDC2 are all thought to play important roles in multiple brain regions. We used the existing evidence to identify and assess essential target genes related to regulators, including evidence from the m6A2Targets database and the recently published m6A sequencing data from the m6A and m6Am modification landscape of human organs [41]. The m6A2Targets database [40] provides evidence for the regulation of regulators to target genes based on a large amount of highthroughput sequencing data, covering the influence of the perturbation of regulators on changes in the expression, methylation, and translation levels of downstream genes and the interaction between regulators and target genes. Finally, to evaluate the importance of these genes, we attempted to fit diagnostic models with regulators and Fig. 8 Construction of diagnostic models in each brain region using key m6A regulators and their target genes. A LASSO punished logistic regression, showing the relation between log(λ) and AUC in the SFG region. B The AUC of the final model in the test dataset and validation dataset. The same results are shown in C-D the HP region and E-F the EC region related hub genes and obtained stable models in the SFG, HP, and EC regions, indicating the confidence of the relationship between key genes and AD.
The study did have several shortcomings. First, only 4 brain regions were analyzed in the study. Considering the complexity of the human brain, more research is needed to determine the changes in m6A regulators in more brain regions. It is also important to compare the impact of m6A modification alterations in AD patients between different brain regions. Second, the m6A modification pattern clusters identified in the study may not reflect the actual state due to the lack of a sample size in each brain region, but we hope this study can provide insight into how to divide AD patients into different groups based on their m6A profile. Furthermore, the assessment of alteration of biological processes using transcriptome data may not reflect the real status and requires further validation using other omics data and certain experiments. Finally, although there are studies between m6A modification and Alzheimer's disease based on mouse models, there are still no large-scale m6A sequencing data of AD patients' brains, so the evidence collected to assess the effectiveness of m6A regulation was only based on other platforms; therefore, they cannot reflect the real regulatory relationship, and further validation is needed.

Conclusion
In conclusion, this study provides insight into the potential regulatory mechanism of m6A regulators in AD, as well as their similarities and differences among brain regions. We hope this study can be used a reference for future research on the relationship between m6A modification and Alzheimer's disease.