Identification of m5C-associated gene signatures in Alzheimer's disease and their relationship with immune infiltration.

Save

Background

Alzheimer's disease (AD) is a progressive neurodegenerative disorder with complex genetic, epigenetic, and immune components. This study investigated m5C-associated gene-expression signatures in AD and explored their biological and immune-cell associations

Methods

Two GEO datasets (GSE132903 and GSE118553) were used for discovery, with GSE33000 and GSE122063 used for external expression assessment. Differential-expression, GO/KEGG enrichment, GSEA, protein-protein interaction, regulatory-network, CIBERSORTx, and ROC analyses were performed. Selected candidates were further examined by RT-qPCR in cortex and hippocampus from 4-month-old male 5 ×FAD mice and age-matched WT controls (n = 5 per group)

Results

Five m5C-associated differentially expressed candidate genes (ZBTB4, MBD4, ZBTB33, NEIL1, and TDG) were identified. ZBTB4 showed reproducible expression differences across several human datasets, but its ROC AUC values (0.657, 0.605, and 0.569) indicated limited standalone discriminatory performance. Gene-immune-cell correlations were statistically significant but weak.

In the 5 ×FAD model, Zbtb4 and Mbd4 mRNA levels were significantly reduced in both cortex and hippocampus compared with WT controls.

Conclusion

The integrated analyses identify candidate methylation/m5C-associated expression changes in AD and provide in vivo transcript-level corroboration for Zbtb4 and Mbd4. The ROC and immune-infiltration findings should be considered exploratory and do not establish clinical diagnostic utility or causal immune mechanisms.

Introduction

Alzheimer's disease (AD) is the prevailing neurodegenerative disorder affecting the elderly population worldwide, often resulting in dementia. The symptoms of AD are neuronal loss, memory impairment, and declining cognitive abilities. The disease is progressive, irreversible, and incurable at this time (Jack, 2022).

Globally, there are more than 55 million AD patients, and this figure continues to grow, presenting a public health challenge (ANON, 2022). From 1990–2019, the incidence and prevalence of AD increased by 147.95%, with the burden increasing worldwide (Li et al., 2022). Besides, AD patients are more vulnerable to SARS-CoV-2 infection and have more severe clinical outcomes and higher mortality rates than that of healthy elderly (Mok et al., 2020).

Thus, early diagnosis of AD could help to develop better patient care and control methods. Effective intervention at the early stages of AD might delay or reduce the prevalence of disease onset due to accurate early diagnosis (Ma et al., 2019). Although the recent technological advancements in genetic testing (Gene Ontology, 2021) and neuroimaging bring the dawn of early AD diagnosis, nearly half of AD patients remain undiagnosed (Li et al., 2022).

Hence, it is crucial to acquire a more profound comprehension of AD mechanisms in order to detect novel biomarkers for timely detection.

The mechanisms underlying AD remain incompletely understood and involve interacting genetic and epigenetic processes (De Jager et al., 2014). RNA modifications add an additional layer of post-transcriptional regulation. Among them, 5-methylcytosine (m5C) occurs in several RNA species and can influence RNA stability, processing, and translation (Roundtree et al., 2017).

Altered RNA-modification pathways have been implicated in neurological disorders and cancer. NSUN2, a well-characterized RNA m5C methyltransferase, has been linked to tau-related neurodegeneration, supporting investigation of m5C-associated molecular changes in AD (Blanco et al., 2011). At the same time, the broader methylation-associated gene panel used in the present study includes genes involved in DNA methylation recognition and base-excision repair as well as genes more directly connected with RNA-m5C biology (Cortellino et al., 2011); these categories are therefore distinguished in the revised interpretation.

Recent studies have linked m5C-related regulation to immune and inflammatory processes (Cui et al., 2022). For example, NSUN2-mediated RNA methylation has been reported to increase interleukin-17A expression under hyperhomocysteinemic conditions (Wang et al., 2017), and altered m5C patterns have been associated with immune-, cytokine-, and interferon-related pathways (Guo et al., 2020). However, comparable evidence in AD remains limited.

We therefore examined whether expression of the selected methylation/m5C-associated genes was related to computationally inferred immune-cell abundance (Newman et al., 2015), while treating these associations as exploratory rather than evidence of direct immune regulation.

We used public brain transcriptomic datasets to identify AD-associated expression changes overlapping a literature-derived panel of m5C/methylation-associated genes. We then examined functional enrichment, molecular interaction networks, immune-cell deconvolution, and the discriminatory performance of selected candidates across independent datasets (Elsheikh et al., 2020). To add experimental corroboration, Zbtb4 and Mbd4 expression was subsequently measured by RT-qPCR in cortex and hippocampus from 4-month-old 5 ×FAD mice and WT controls.

The study is intended as an exploratory, hypothesis-generating analysis rather than a demonstration of a blood biomarker, clinical diagnostic test, or causal mechanism.

Materials and methods

Datasets and preprocessing

Four AD-associated datasets were obtained from the Gene Expression Omnibus (GEO): GSE132903 (Piras et al., 2019), GSE118553 (Patel et al., 2019), GSE33000 (Narayanan et al., 2014), and GSE122063 (McKay et al., 2019). GSE132903 contained 195 samples (93 controls and 102 AD), GSE118553 contained 266 samples (100 controls and 166 AD), GSE33000 contained 467 samples (157 controls and 310 AD), and GSE122063 contained 100 samples (44 controls and 56 AD). Each dataset was processed separately.

Expression values were normalized within each dataset using the normalizeBetweenArrays function in the limma package; no ComBat-based cross-dataset batch correction was performed. The pre- and post-normalization distributions for GSE132903 and GSE118553 are shown in Fig. 2.

Identification of methylation/m5C-associated differentially expressed candidates

Differential expression between AD and control samples was analyzed separately in GSE132903 and GSE118553 using limma. The original exploratory screening criterion was a nominal P value < 0.05 together with a non-zero log2 fold change (|log2FC= > 0); genes with log2FC > 0 were classified as upregulated and those with log2FC < 0 as downregulated. We acknowledge that this is a permissive effect-size threshold and therefore interpret downstream overlaps as candidate-generating rather than as a set of large-effect DEGs.

The intersection of DEGs from the two datasets was obtained using a Venn analysis (Song et al., 2022). A literature-derived panel of 20 genes (DNMT1, DNMT3A, DNMT3B, MBD1, MBD2, MBD3, MBD4, MECP2, NEIL1, NTHL1, SMUG1, TDG, UHRF1, UHRF2, UNG, ZBTB33, ZBTB4, TET1, TET2, and TET3) was then intersected with the shared DEGs. This panel contains canonical DNA methylation/base-excision-repair genes in addition to genes linked more directly to RNA-m5C biology; accordingly, the term 'm5C-associated' is used operationally and does not imply that every gene in the panel is a direct RNA m5C writer, reader, or eraser.

Enrichment Analysis of m5CRDEGs

Gene Ontology (GO) (Mi et al., 2019) and Kyoto Encyclopedia of Genes and Genomes (KEGG) (Mandrekar, 2010) enrichment analyses were performed with clusterProfiler, with multiple-testing-adjusted P values reported where available. Gene set enrichment analysis (GSEA) (Subramanian et al., 2005) was performed using curated pathway gene sets from MSigDB (Liberzon et al., 2011). In the revised analysis and interpretation, pathways with adjusted P < 0.05 are treated as statistically supported after multiple testing, whereas pathways meeting only nominal P < 0.05 are presented as exploratory and are not interpreted as AD-specific mechanisms.

Protein-protein interaction network analysis

Protein-protein interaction (PPI) analysis was performed to characterize potential functional relationships among the candidate genes. Interactions were queried in STRING (Szklarczyk et al., 2019) using the medium-confidence threshold (minimum interaction score > 0.400), and the resulting network was visualized in Cytoscape (version 3.8.0) (Shannon et al., 2003). The network analysis was used for candidate prioritization rather than as evidence of functional causality.

Construction of mRNA-miRNA, mRNA-TF, and mRNA-compound interaction networks

Potential mRNA-miRNA interactions were identified using ENCORI/starBase (version 3.0) (Li et al., 2014) and miRDB (version 5.0) (Chen and Wang, 2020). miRDB predictions with a Target Score > 80 were intersected with mRNA-miRNA relationships retrieved from ENCORI, and the retained relationships were used to construct the mRNA-miRNA interaction network.

Potential transcription-factor (TF)-target relationships were obtained from ChIPBase (version 2.0) (Zhou et al., 2017) and hTFtarget (Zhang et al., 2020a) and intersected with the five candidate genes to construct the mRNA-TF network. Candidate gene-compound associations were retrieved from the Comparative Toxicogenomics Database (CTD) (Davis et al., 2021). All regulatory and compound-association networks were visualized in Cytoscape (version 3.7.1).

Analysis of immune-cell infiltration by CIBERSORTx

CIBERSORTx was used with the LM22 signature matrix to estimate the relative proportions of 22 immune-cell populations in GSE132903 and GSE118553 (Newman et al., 2015). All available samples were retained for the descriptive analyses. The exported GSE132903 matrix contained immune-cell fractions but did not include per-sample CIBERSORTx P values, whereas the GSE118553 output included P values ranging from 0.26 to 0.98 (0/266 samples with P < 0.05).

Therefore, a CIBERSORTx P-value threshold was not used as a uniform sample-exclusion criterion. Spearman correlations between candidate-gene expression and estimated immune-cell fractions were calculated, and associations with P < 0.01 were visualized. Because the correlations were small in magnitude and deconvolution significance could not be uniformly applied, these analyses are considered exploratory.

Receiver operating characteristic analysis

Candidate-gene expression was compared between AD and control samples across GSE132903, GSE118553, GSE33000, and GSE122063. For candidates showing reproducible group differences, receiver operating characteristic (ROC) curves were generated to quantify discrimination between AD and control samples. AUC values are reported descriptively and are not interpreted as evidence of clinical diagnostic performance in the absence of prospective clinical validation.

Animal model and RT-qPCR validation

Four-month-old male 5 ×FAD mice and age-matched male WT controls were used for experimental validation (n = 5 biological replicates per group) (Oakley et al., 2006). Mice were obtained from Capital Medical University and maintained under standard laboratory housing conditions in accordance with institutional animal-care requirements. Euthanasia, brain removal, regional dissection, and tissue preservation were performed using routine procedures approved under the institutional animal-welfare protocol.

Cortex and hippocampus were collected for RNA analysis.

Total RNA was extracted using the FastPure Cell/Tissue Total RNA Isolation Kit V2 (Vazyme Biotech Co., Ltd., Nanjing, China; Cat. No. RC112–01). cDNA was synthesized using the HiScript III 1st Strand cDNA Synthesis Kit (+gDNA wiper) (Vazyme; Cat.

No. R312). Quantitative PCR was performed with 2 ×RealStar Fast SYBR qPCR Mix (GenStar; Cat.

No. A301) on an Applied Biosystems QuantStudio 6 Flex Real-Time PCR System with a 96-well block. Each biological sample was analyzed in three technical replicates.

Hprt1 and Gapdh were used as reference genes, and the arithmetic mean of their Ct values was used as the reference Ct. Relative expression was calculated by the 2^-ΔΔCt method (Livak and Schmittgen, 2001), with ΔCt defined as Ct(target) minus the mean Ct of Hprt1 and Gapdh and ΔΔCt calculated relative to the mean WT ΔCt within the same brain region. Primer sequences were: Zbtb4-F, 5′-GTTGGCTATGCTCTCAGAGGT-3′; Zbtb4-R, 5′-TGCCAGCCTAAACATAGGGAA-3′; Mbd4-F, 5′-TTGGGATCGCTGCAAGGAAG-3′; Mbd4-R, 5′-CAGTGGGCTCTTGGAGAAGG-3′; Hprt1-F, 5′-TTGGGCTTACCTCACTGCTT-3′; Hprt1-R, 5′-CTAATCACGACGCTGGGACT-3′; Gapdh-F, 5′-ACCCAGAAGACTGTGGATGG-3′; and Gapdh-R, 5′-CACATTGGGGGTAGGAACAC-3′.

Statistical analysis

R (version 4.2.0) was used for the bioinformatic statistical analyses. Between-group transcriptomic comparisons were performed as described above, and Spearman coefficients were used for gene-immune-cell correlations. For RT-qPCR, WT and 5 ×FAD groups were compared using two-sided Welch's t-tests on ΔCt values.

The individual mouse was the biological replicate; technical replicates were averaged and were not treated as independent observations. RT-qPCR data are displayed as mean ± SEM in Fig. 10.

A two-sided P value < 0.05 was considered statistically significant unless an adjusted P-value criterion is specified.

Results

Overview of the study design (Fig. 1)

Within-dataset normalization

GSE132903 and GSE118553 were normalized separately using limma. Fig. 2 shows the expression-value distributions before and after normalization for each dataset.

These plots demonstrate within-dataset normalization; they should not be interpreted as evidence of cross-dataset batch-effect removal because no ComBat or comparable cross-study batch-correction procedure was applied.Fig. 1Flow Chart for the Comprehensive Analysis of Co-DEGs. Co-DEGs, Common differentially expressed genes, m5CRDEGs, m5C related differentially expressed genes; GO, Gene Ontology; BP, Biological Process; CC, Cellular Component; MF, Molecular Function; KEGG, Kyoto Encyclopedia of Genes and Genomes; GSEA, Gene Set Enrichment Analysis; ROC, Receiver Operating Characteristic Curve.Fig.

2Normalization of GSE132903 and GSE118553. (A) Pre-normalization expression-value distributions for GSE132903. (B) Post-normalization distributions for GSE132903.

(C) Pre-normalization distributions for GSE118553. (D) Post-normalization distributions for GSE118553. AD and control samples are shown using the colors indicated in the figure.

These panels demonstrate within-dataset normalization and do not represent cross-dataset batch correction.

Identification of methylation/m5C-associated candidate genes

Differential-expression analyses were performed separately in GSE132903 and GSE118553, with volcano plots and heatmaps shown in Fig. 3A-D. The intersection of nominally significant DEGs from the two datasets yielded the shared DEG set (Fig.

3E). Intersecting these genes with the 20-gene literature-derived methylation/m5C-associated panel identified five candidates: ZBTB33, TDG, ZBTB4, NEIL1, and MBD4 (Fig. 3F) (Song et al., 2022, Song et al., 2021).

Using the original exploratory criterion of P < 0.05 and non-zero log2FC, 15,388 genes met the threshold in GSE132903 (8240 upregulated and 7148 downregulated) and 11,005 genes met the threshold in GSE118553 (5396 upregulated and 5609 downregulated). Because the fold-change threshold was deliberately permissive, these counts and downstream overlaps are interpreted as exploratory candidate screening.Fig. 3Analyzing gene expression differentially.

(A) A volcanic plot of differential gene expression analysis between AD and normal groups in the dataset GSE132903. (B) A volcanic plot of differential gene expression analysis between AD and normal groups in the dataset GSE118553. (C) Analyzing the dataset GSE132903 heat maps to identify differentially expressed genes in AD and normal groups.

(D) The dataset GSE118553 heat map analysis of differentially expressed genes among AD and normal groups. (E) Venn diagram of DEGs identified for the above two data sets. (F) Venn diagram of intersection and m5C-related genes.

DEGs: differentially expressed genes.

Analysis of the GO and KEGG pathways enriched in m5CRDEGs

GO and KEGG analyses of the five candidate genes are summarized in Table 1 and Fig. 4. Enriched terms included pyrimidine nucleotide catabolic processes, pyrimidine deoxyribonucleotide metabolism, base-excision repair/AP-site formation, and DNA N-glycosylase activity.

These annotations are consistent with the known DNA-repair functions of several genes in the panel and should not be interpreted as direct evidence of altered RNA m5C modification. Dataset-specific gene-concept displays based on logFC values are shown in Fig. 4C-F.Table 1Results of GO and KEGG Enrichment.ONTOLOGYIDDescriptionGeneRatioBgRatiopvaluep.adjustBPGO:0006244pyrimidine nucleotide catabolic process3/514/186703.3536E-094.1361E-08BPGO:0009219pyrimidine deoxyribonucleotide metabolic processe3/520/186701.0498E-087.552E-09BPGO:0006285base-excision repair, AP site formation3/510/186701.1059E-094.0919E-08CCGO:0016605PML body1/599/197170.024856890.04971378CCGO:0016607Nuclear speck1/5397/197170.096710630.09671063MFGO:0016798hydrolase activity, acting on glycosyl bonds3/5117/176972.7984E-062.2315E-05MFGO:0016799hydrolase activity, hydrolyzing N-glycosyl compounds3/521/176971.4379E-081.7254E-07MFGO:0019104DNA N-glycosylase activity3/514/176973.9375E-099.45E-08KEGGhsa03410Base excision repair3/333/80766.2172E-086.2172E-08Fig.

4GO and KEGG enrichment analysis of the five candidate genes. (A) Bar plot of selected GO biological process (BP), cellular component (CC), molecular function (MF), and KEGG terms. (B) Network connecting selected enriched terms and candidate genes.

(C-D) Gene-concept displays using GSE132903 logFC values. (E-F) Corresponding displays using GSE118553 logFC values. Adjusted P values are shown where applicable.

GSEA

GSEA results are shown in Fig. 5 and Table 2. After adjustment for multiple testing, WP_METABOLIC_REPROGRAMMING_IN_COLON_CANCER remained significant in GSE132903 (adjusted P = 0.0286), whereas PID_BETA_CATENIN_NUC_PATHWAY and BIOCARTA_P38MAPK_PATHWAY remained significant in GSE118553 (adjusted P = 0.0386 for each).

The displayed apoptosis, fatty-acid metabolism, circadian-rhythm, hedgehog-ligand-biogenesis, and autophagy gene sets met nominal P < 0.05 but did not meet adjusted P < 0.05 and are therefore presented as exploratory patterns. Pathway names derived from cancer-oriented databases reflect the source annotation and do not imply that a cancer-specific mechanism is operating in AD.Fig. 5GSEA results for GSE132903 and GSE118553.

(A-D) Selected gene sets in GSE132903: (A). WP_METABOLIC_REPROGRAMMING_IN_COLON_CANCER. (B).

WP_APOPTOSIS. (C). REACTOME_FATTY_ACID_METABOLISM.

(D). KEGG_CIRCADIAN_RHYTHM_MAMMAL. (E-H) Selected gene sets in the GSE118553 dataset: (E).

PID_BETA_CATENIN_NUC_PATHWAY. (F). BIOCARTA_P38MAPK_PATHWAY.

(G). REACTOME_HEDGEHOG_LIGAND_BIOGENESIS. (H).

KEGG_REGULATION_OF_AUTOPHAGY. AD: Alzheimer's disease; GSEA: Gene Set Enrichment Analysis.Table 2Results of Combined Datasets GSEA.Set SizeEnrichment ScoreNESpvaluep.adjustWP_METABOLIC_REPROGRAMMING_IN_COLON_CANCER410.6964112.3909680.0027320.028643WP_APOPTOSIS820.3680811.4283310.0326970.150503REACTOME_FATTY_ACID_METABOLISM141-0.369892-1.4384670.0076920.050822KEGG_CIRCADIAN_RHYTHM_MAMMAL12-0.672961-1.6216290.0203380.105192PID_BETA_CATENIN_NUC_PATHWAY750.4390381.6712910.00320.038587BIOCARTA_P38MAPK_PATHWAY360.5256691.7340110.0032840.038587REACTOME_HEDGEHOG_LIGAND_BIOGENESIS63-0.351430-1.3955620.0387590.163271KEGG_REGULATION_OF_AUTOPHAGY34-0.430018-1.4820230.0493820.184964

PPI network and hub-gene prioritization

To examine interactions among the five candidate genes, we constructed a PPI network using STRING and visualized it in Cytoscape (Fig. 6A). Network ranking with five algorithms—MCC (Fig.

6B), EPC (Fig. 6C), DMNC (Fig. 6D), BottleNeck (Fig.

6E), and Clustering Coefficient (Fig. 6F)—prioritized MBD4 as one of the more highly connected candidates. These network results are interpreted as prioritization signals rather than evidence of functional causality.Fig.

6PPI Network and m5CRDEGs Analysis. (A) Entire PPI network. (B) PPI network of m5CRDEGs computed by the MCC algorithm.

(C) PPI network of m5CRDEGs computed by the EPC algorithm. (D) PPI network of m5CRDEGs computed by the DMNC algorithm. (E) PPI network of m5CRDEGs computed by the BottleNeck algorithm.

(F) PPI network of m5CRDEGs computed by the Clustering Coefficient algorithm. The colors from red to yellow represent the higher to lower assessment scores. PPI: protein-protein interaction.

Construction of mRNA-TF, mRNA-compound, and mRNA-miRNA interaction networks

ChIPBase and hTFtarget were used to identify predicted TF-target relationships involving the five candidate genes. After integrating the database outputs, 153 TF-mRNA interactions were retained and visualized in Cytoscape (Fig. 7A).

Candidate mRNAs are shown as orange nodes and predicted TFs as green nodes.Fig. 7Predicted mRNA-TF, mRNA-compound, and mRNA-miRNA interaction networks. (A) Candidate mRNAs are shown as orange nodes and predicted transcription factors as green nodes.

(B) Candidate mRNAs are shown as light-blue rectangles and associated compounds as yellow ovals. (C) Candidate mRNAs are shown as purple rectangles and predicted miRNAs as pink ovals. Network edges represent database-derived predicted or curated associations and do not establish functional regulation.

CTD was used to identify candidate gene-compound associations. Ten compounds were retained and visualized in the mRNA-compound network (Fig. 7B), with candidate mRNAs shown as light-blue rectangles and compounds as yellow ovals.

ENCORI and miRDB were used to identify predicted mRNA-miRNA relationships. Using the screening criteria described in Methods, 87 mRNA-miRNA pairs were retained and visualized in Cytoscape (Fig. 7C), with candidate mRNAs shown as purple rectangles and predicted miRNAs as pink ovals.

Exploratory immune-infiltration analysis (CIBERSORTx)

CIBERSORTx was used to estimate 22 immune-cell fractions in GSE132903 and GSE118553. Correlation heatmaps summarize relationships between candidate-gene expression and inferred immune-cell proportions (Fig. 8A-B).

Because a uniform CIBERSORTx P-value filter could not be applied to both datasets, all samples were retained and the correlation analysis was treated as exploratory. Associations meeting P < 0.01 were selected for scatter-plot visualization (Fig. 8C-H).Fig.

8Exploratory immune-infiltration analysis using CIBERSORTx. (A-B) Heatmaps of associations between candidate-gene expression and estimated immune-cell fractions in GSE132903 and GSE118553. (C-H) Scatter plots for selected gene-immune-cell pairs with P < 0.01.

Correlation coefficients are small in magnitude and are interpreted as weak associations rather than functional relationships.

The statistically significant gene-immune-cell associations were weak in magnitude. In GSE132903, ZBTB33 correlated positively with monocyte abundance (r = 0.300, P < 0.001), ZBTB4 correlated positively with CD8 + T-cell abundance (r = 0.186, P = 0.009), and NEIL1 correlated negatively with monocyte abundance (r = -0.207, P = 0.004). In GSE118553, ZBTB33 showed a weak positive correlation with monocytes (r = 0.187, P = 0.002), ZBTB4 showed a weak positive correlation with CD8 + T cells (r = 0.176, P = 0.004), and NEIL1 showed a weak negative correlation with monocytes (r = -0.258, P < 0.001).

These associations do not demonstrate a functional or causal immune mechanism.

External expression assessment and ROC analysis in GSE33000 and GSE122063

Fig. 9A-D compares candidate-gene expression between AD and control groups across the four datasets. ZBTB4 showed consistent differential expression in GSE132903, GSE118553, and GSE33000, while ZBTB4 and MBD4 also differed between groups in GSE122063.

ROC analysis of ZBTB4 yielded AUC values of 0.657 in GSE132903, 0.605 in GSE118553, and 0.569 in GSE33000 (Fig. 9E-G). These values indicate limited standalone discrimination and do not support use of ZBTB4 as a clinical diagnostic biomarker on the basis of the present data.Fig.

9Candidate-gene expression and ROC analysis. (A-D) Expression comparisons between AD and control groups across GSE132903, GSE118553, GSE33000, and GSE122063. (E-G) ROC curves for ZBTB4 in GSE132903, GSE118553, and GSE33000, with AUC values of 0.657, 0.605, and 0.569, respectively.

These AUC values indicate limited standalone discrimination and are not presented as evidence of clinical diagnostic utility.

Experimental RT-qPCR validation in 5 ×FAD mice

To provide independent experimental corroboration of the transcriptomic findings, Zbtb4 and Mbd4 mRNA expression was measured in cortex and hippocampus from 4-month-old male 5 ×FAD mice and age-matched WT controls (n = 5 per group; Fig. 10). Zbtb4 expression was significantly reduced in the 5 ×FAD cortex (P = 0.033) and hippocampus (P < 0.001).

Mbd4 expression was also reduced in the 5 ×FAD cortex (P = 0.024) and hippocampus (P = 0.0013). Statistical comparisons were performed using two-sided Welch's t-tests on ΔCt values.Fig. 10Experimental RT-qPCR validation of Zbtb4 and Mbd4 expression in 4-month-old 5 ×FAD mice.

(A) Experimental workflow. Cortex and hippocampus were collected from male WT and 5 ×FAD mice (n = 5 biological replicates per group), followed by RNA extraction, reverse transcription, and RT-qPCR. (B) Zbtb4 expression in cortex (P = 0.033).

(C) Mbd4 expression in cortex (P = 0.024). (D) Zbtb4 expression in hippocampus (P < 0.001). (E) Mbd4 expression in hippocampus (P = 0.0013).

Hprt1 and Gapdh were used as reference genes, with the arithmetic mean Ct used for ΔCt calculation. Data are presented as mean ± SEM; each dot represents one mouse. P values were calculated using two-sided Welch's t-tests on ΔCt values.

These experiments provide transcript-level, cross-species corroboration for altered expression of the selected candidate genes. They do not establish protein-level changes, clinical diagnostic performance, or causal roles in AD pathogenesis.

Discussion

AD is a progressive neurodegenerative disorder that causes substantial morbidity and mortality (Jack, 2022). Therapeutic trials targeting core AD pathology have historically shown limited benefit when initiated late in the disease course (Aisen et al., 2022), supporting continued efforts to identify earlier molecular changes and treatment-relevant pathways. The present study therefore focuses on exploratory transcriptomic signals and candidate biomarkers, with conclusions constrained by the retrospective datasets and limited experimental validation.

Changes in peripheral and central immune compartments have been implicated in AD, and RNA modifications are increasingly recognized as potential regulators of neuronal and immune biology (Heneka et al., 2015, Roundtree et al., 2017). The present study therefore explored relationships between a literature-derived methylation/m5C-associated gene panel, AD transcriptomic changes, and inferred immune-cell abundance. Because several members of the panel are canonical DNA methylation readers/repair factors rather than direct RNA-m5C regulators, the findings should be interpreted as methylation-associated transcriptomic signals rather than evidence of a unified RNA-m5C mechanism.

Five candidate genes (ZBTB33, TDG, ZBTB4, NEIL1, and MBD4) were identified at the intersection of the shared AD-associated DEGs and the literature-derived methylation/m5C-associated panel. ZBTB4 showed relatively consistent expression differences across multiple human datasets; however, the observed AUC values of 0.657, 0.605, and 0.569 demonstrate limited standalone discrimination. Accordingly, ZBTB4 should be considered a candidate expression marker for further study rather than a validated diagnostic biomarker.

Prior studies linking ZBTB4 to genome stability (Weber et al., 2008), p53-related pathways, and age-associated cognitive phenotypes provide biological context, but they do not establish that altered ZBTB4 expression causes AD.

GO/KEGG enrichment emphasized pyrimidine metabolism and base-excision-repair-related terms, which is consistent with the DNA-repair functions of TDG, MBD4, and NEIL1. GSEA identified a limited set of pathways that remained significant after multiple-testing correction. In particular, p38MAPK and beta-catenin-related gene sets in GSE118553 and the pathway labeled 'metabolic reprogramming in colon cancer' in GSE132903 met adjusted P < 0.05 (Inestrosa and Arenas, 2009, Pei et al., 2001).

These gene-set names originate from curated databases and should be interpreted as shared transcriptional programs rather than disease-specific mechanistic assignments. Other displayed pathways were nominally significant only and are treated as hypothesis-generating observations.

The immune-cell analysis identified statistically significant but weak correlations between selected genes and inferred monocyte or CD8 + T-cell abundance. Although prior work supports involvement of CD8 + T cells and monocyte-lineage cells in AD biology (Gate et al., 2020), the present correlations are too small to support claims of a strong relationship or functional coupling. In addition, the CIBERSORTx quality metrics could not be applied uniformly across the two datasets.

These observations should therefore be viewed as exploratory associations that warrant validation with orthogonal immune-profiling approaches.

Network analyses placed MBD4 among the more highly connected candidates in the predicted regulatory networks. MBD4 is a DNA glycosylase involved in repair of deamination events at methylated cytosines and has established roles in genome maintenance (Hashimoto et al., 2012, Hendrich and Bird, 1998). Its role in AD remains uncertain.

The network results therefore provide candidate regulatory relationships for follow-up rather than direct evidence that MBD4 drives AD pathology.

The added RT-qPCR experiments strengthen the study by showing that Zbtb4 and Mbd4 mRNA levels are reduced in both cortex and hippocampus of 4-month-old male 5 ×FAD mice (Oakley et al., 2006). This provides in vivo transcript-level corroboration across species and brain regions. Nevertheless, the mouse experiments do not validate human diagnostic performance and do not establish protein-level regulation or causal function.

Differences in species, disease stage, cell composition, and brain region may also influence expression patterns.

Several limitations should be emphasized. First, the human analyses are retrospective and rely on public transcriptomic datasets with heterogeneous clinical and technical characteristics. Second, the original DEG screen used a permissive non-zero fold-change criterion, which increases sensitivity at the expense of effect-size specificity.

Third, the 20-gene panel combines direct RNA-m5C-related biology with DNA methylation and base-excision-repair genes and therefore should not be interpreted as a canonical RNA-m5C regulator set. Fourth, the CIBERSORTx correlations are weak and exploratory, and deconvolution quality metrics were not uniformly available across datasets. Fifth, the RT-qPCR validation involved a small cohort of male 5 ×FAD mice (n = 5 per group) and measured mRNA only.

Larger independent human cohorts, sex-inclusive animal studies, protein-level validation, and functional perturbation experiments will be required to determine biological and clinical relevance.

Conclusions

In summary, integrated transcriptomic analyses identified five methylation/m5C-associated candidate genes (ZBTB33, TDG, ZBTB4, NEIL1, and MBD4) that differ between AD and control brain samples. ZBTB4 showed reproducible expression differences but only limited standalone ROC discrimination. RT-qPCR in the 5 ×FAD mouse model corroborated lower Zbtb4 and Mbd4 mRNA expression in cortex and hippocampus.

Associations with inferred monocyte and CD8 + T-cell abundance were weak and should be considered exploratory. These findings provide candidate molecular signals and hypotheses for further mechanistic and clinical validation rather than establishing diagnostic biomarkers or causal immune pathways.

Abbreviations

ADAlzheimer's diseasem5C5-methylcytosineDEGsdifferentially expressed genesGEOGene Expression OmnibusGOGene OntologyKEGGKyoto Encyclopedia of Genes and GenomesPPIsProtein–protein interactionsSTRINGSearch Tool for the Retrieval of Interacting Genes/ProteinsROCReceiver operating characteristic curvesIMEImmune microenvironmentBPBiological processMFMolecular functionCCCellular componentGSEAGene set enrichment analysisTFTranscription factorsCTDComparative Toxicogenomics DatabaseSDStandard deviationAMLAcute myeloid leukemiaRT-qPCRreverse transcription quantitative polymerase chain reaction

CRediT authorship contribution statement

Pan Ai: Conceptualization. Xuyang Wang: Writing – original draft. Hui Shi: Investigation.

Anshi Wu: Writing – review & editing. Jing Wang: Writing – review & editing.

Ethics declaration

This study was conducted in accordance with the ARRIVE (Animal Research: Reporting of In Vivo Experiments) guidelines. This study was approved by the Beijing Chaoyang Hospital. (Approval No.

2026–628)

Ethics statement

Public human transcriptomic data were analyzed in de-identified form. The animal validation study was approved by the Animal Welfare and Ethics Committee of Beijing Chao-Yang Hospital, Capital Medical University (Approval No. 2026–628).

Ethics approval and consent to participate

The human component of this study consisted exclusively of secondary analysis of de-identified, publicly available GEO datasets; no new human participants were recruited. Animal procedures for the RT-qPCR validation were approved by the Animal Welfare and Ethics Committee of Beijing Chao-Yang Hospital, Capital Medical University (Approval No. 2026–628) and were performed in accordance with the approved institutional animal-welfare protocol.

Consent for publication

Not applicable.

Funding information

This research was funded by 10.13039/501100001809National Natural Science Foundation of China, grant number 82271210 and 10.13039/501100004826Beijing Natural Science Foundation, grant number 7222074.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Related articles

PubMed

Retraction Note: Modulation of aberrant CDK5 signaling rescues impaired neurogenesis in models of Alzheimer's disease.

Retraction Note to: Cell Death & Diseasehttps://doi.org/10.1038/cddis.2011.2, published online 10 February 2011 The Editors have retracted this article. After publication, concerns were raised regarding some of the data presented in the figures and supplementary materials, specifically:In Fig. 4j–l, the merged image (l) doesn’t fully match the DCX (j, green) and CDK5 (k, red) channels, and there appears to be a duplicated pattern in the top left of the merged image;Fig. 8 panels f and g appear...

PubMed

Glial Cell Mechanisms in Alzheimer's Disease Pathology and Progress in Biomarker Research.

Abstract Alzheimer’s disease (AD) is the most common neurodegenerative disorder in older adults and is increasingly recognized as a systems-level disease involving dynamic dysfunction and crosstalk among multiple glial cell populations. Microglia, astrocytes, and oligodendrocytes play distinct but interconnected roles across the AD continuum. Microglia respond to amyloid-β (Aβ) and tau pathology by adopting disease-associated phenotypes that contribute to neuroinflammation, complement-dependent...