Abstract

Objective

   This study aimed to identify DNA methylation patterns associated with
   Very Late-Onset Schizophrenia-like Psychosis (VLOSLP) and to develop
   methylation-based biomarkers that differentiate VLOSLP from
   Schizophrenia (SCZ) and Alzheimer’s Disease (AD).

Methods

   We analyzed methylation microarray datasets (n = 1218) from SCZ and AD
   patients obtained from the GEO database. We then collected blood
   samples from VLOSLP patients and age-matched healthy controls (n = 80)
   at the Wuxi Mental Health Center for methylation microarray profiling
   and bisulfite sequencing validation. Differential methylation analysis
   and Gene Ontology (GO) enrichment analysis identified candidate loci.
   We prioritized key methylation sites through integrated analysis of
   methylation quantitative trait loci (meQTL), linkage disequilibrium
   (LD) patterns, and blood-brain methylation correlations. Machine
   learning algorithms generated diagnostic models, with classification
   performance evaluated using Area Under the Curve (AUC) metrics.

Results

   Analysis revealed distinct DNA methylation signatures in VLOSLP
   patients compared to controls. The GNB5 gene exhibited shared
   epigenetic modifications across SCZ, AD, and VLOSLP, suggesting a
   common pathogenic mechanism. The diagnostic model discriminating AD
   from VLOSLP demonstrated high accuracy, achieving an AUC of 1.0 in the
   training set and 0.958 in the test set (95% CI: 0.875–1.000). The AD
   versus SCZ classification model showed similar robustness, with AUCs of
   0.995 and 0.955 in training and test sets, respectively (95% CI:
   0.926–0.983). The SCZ versus VLOSLP model achieved perfect
   discrimination (AUC = 1.0) in both training and test sets, with
   substantial clinical utility. Additional analyses suggested distinct
   molecular subtypes within VLOSLP.

Conclusion

   Specific DNA methylation alterations in VLOSLP are identified as
   potential diagnostic biomarkers. These findings may contribute to the
   development of molecular diagnostic tools, though further validation in
   larger, independent cohorts is warranted.

   Keywords: very late-onset schizophrenia-like psychosis, epigenetics
   schizophrenia, Alzheimer’s disease, biomarkers

Introduction

   Very-late-onset schizophrenia-like psychosis (VLOSLP) represents a
   distinct psychiatric condition that manifests initially after age
   60.[40]^1 This disorder constitutes approximately 15% of geriatric
   psychiatric presentations and associates with elevated mortality
   rates[41]^2 and increased dementia incidence.[42]^3^,[43]^4 VLOSLP
   exhibits a phenomenological profile that differentiates it from
   early-onset schizophrenia, characterized predominantly by delusional
   positive symptoms while maintaining relative cognitive
   preservation.[44]^5 Despite these distinctive clinical features, the
   underlying biological mechanisms remain incompletely understood,
   creating substantial challenges for accurate diagnosis and therapeutic
   intervention.

   DNA methylation and other epigenetic modifications appear to mediate
   critical pathogenic processes in psychiatric disorders.[45]^6 Research
   has demonstrated characteristic epigenetic alterations in both
   schizophrenia (SCZ)[46]^7–9 and Alzheimer’s disease[47]^10–13 (AD).
   However, investigators have not yet systematically examined the
   epigenetic profile of VLOSLP. The distinctive clinical presentation and
   age-specific onset of VLOSLP suggest that understanding its epigenetic
   architecture may illuminate disease mechanisms and identify potential
   diagnostic biomarkers.

   This investigation presents the first comprehensive analysis of
   genome-wide DNA methylation patterns in VLOSLP patients, incorporating
   comparative analyses with SCZ and AD cohorts. The research aims to
   elucidate VLOSLP-specific epigenetic signatures and their potential
   diagnostic and therapeutic implications. [48]Figure 1 presents the
   methodological framework for this investigation.

Figure 1.

   [49]Figure 1
   [50]Open in a new tab

   Research flow diagram. The flow diagrams illustrate the research
   design.

Materials and Methods

Subjects

   We retrieved datasets from the National Center for Biotechnology
   Information (NCBI) GEO database ([51]https://www.ncbi.nlm.nih.gov/geo;
   accessed December 2023) using the keywords Schizophrenia and
   Alzheimer’s Disease. The inclusion criteria for dataset selection were
   as follows: (1) Study design: Methylation profiling by array, with a
   minimum coverage of 450,000 probes; (2) Specimen type: Post-mortem
   brain tissue or peripheral blood samples from Homo sapiens; and (3)
   Sample size: Each dataset must include at least 10 samples from either
   schizophrenia (SCZ) or Alzheimer’s disease (AD) patients, along with a
   comparable number of control samples.[52]^14

   Based on these criteria, six datasets were included in the final
   analysis: three derived from peripheral blood samples
   ([53]GSE41169:[54]^15 n [SCZ] = 62, n [HC] = 33; [55]GSE153712:[56]^16
   n [AD] = 255, n [HC] = 471; [57]GSE157252:[58]^17 n [SCZ] = 41, n [HC]
   = 50) and three from post-mortem brain tissue samples
   ([59]GSE61107:[60]^18 n [SCZ] = 23, n [HC] = 25; [61]GSE66351:[62]^19 n
   [AD] = 106, n [HC] = 84; [63]GSE112179:[64]^20 n [SCZ] = 35, n [HC] =
   33). In total, the study encompassed 161 SCZ patients, 361 AD patients,
   and 696 healthy controls. Detailed inclusion and exclusion criteria for
   each dataset are described in the respective publications.

   A total of 80 (44 VLOSLP patients and 36 matched health controls)
   participants from the Wuxi Mental Health Center, affiliated with
   Jiangnan University, were included in the final analysis. Six
   individuals were excluded from the study: four declined to participate,
   and two did not meet the inclusion criteria, which are detailed below.
   All participants provided complete medical histories, underwent
   physical examinations, and completed laboratory testing. Informed
   consent was obtained from each participant following a thorough
   explanation of the study’s purpose and procedures. The research was
   conducted in accordance with the ethical principles outlined in the
   Declaration of Helsinki and received approval from the Ethics Committee
   of the Wuxi Mental Health Center (Ethics No. WXMHCIRB2023LLky001). The
   specific inclusion and exclusion criteria are: The inclusion criteria
   were: 1) Age 60–85 years, Han ethnicity, right-handed; 2) VLOSLP
   diagnosis according to international expert consensus criteria,[65]^21
   with initial psychiatric symptoms (delusions and hallucinations) onset
   after 60 year-old, confirmed by two senior psychiatrists. 3) No serious
   physical illness as well as normal laboratory results for past three
   months. 4) Cognitive screening criteria: Initial phase (first two
   participants): Mini-Mental State Examination (MMSE) score ≥24 points;
   Formal study phase (all participants): Montreal Cognitive Assessment
   (MoCA) score ≥ 26 points; 5) Capable of providing informed consent. The
   exclusion criteria: 1) DSM-5 Axis I disorders: organic mental
   disorders, primary emotional disorders, dementia, consciousness
   disturbances, delirium, memory impairments, or substance-induced mental
   disorders; 2) Unstable medical conditions in acute phase (diabetes,
   thyroid disorders, hypertension, cardiovascular diseases); 3) were
   treated with electroconvulsive therapy during the last month; 4) had
   abnormal liver or renal function; 5) had QTc prolongation, a history of
   congenital QTc prolongation, or recent (ie, within the past 6 months)
   myocardial infarction; 6) Has a history of previous treatment with
   antipsychotic medications, antidepressant agents, and mood stabilizing
   drugs; 7) The patient or their legally designated guardian is unable to
   provide signed consent on the informed consent form.

Genome-Wide DNA Methylation Profiling

   Our study examined whole-genome DNA methylation patterns of peripheral
   blood using the methylation EPIC v2.0 BeadChip platform in a cohort
   comprising 5 VLOSLP patients and 3 matched healthy controls. Blood
   collection occurred at 0600 hours on enrollment day following informed
   consent protocols. We obtained 5 mL fasting blood samples from the
   antecubital vein and stored them in EDTA-coated tubes at −80°C.

   DNA extraction from peripheral blood samples utilized the DNeasy Blood
   and Tissue Kit (Qiagen Cat. No. 69504). Nanodrop 2000 spectrophotometry
   (Thermo, Beijing, China) enabled DNA purity and concentration
   assessments. We processed 500 ng of DNA from each sample through
   bisulfite conversion using EZ DNA Methylation Kits (Zymo Research,
   Tustin, CA, USA). The processed samples underwent analysis via the
   Infinium Methylation EPIC 2.0 BeadChip platform according to
   manufacturer specifications (Illumina, Hayward, CA, USA). This platform
   enables quantitative analysis of over 935,000 CpG sites across the
   genome at single-nucleotide resolution.

   Shanghai Bohao Biotechnology Co., Ltd. conducted all laboratory
   procedures, including DNA extraction, quality control assessment,
   bisulfite conversion, and BeadChip-specific protocols (amplification,
   labeling, hybridization, washing, and scanning) under standardized
   conditions.

Data Preprocessing and Analysis

   To ensure the reliability and precision of our data analysis, each
   dataset was independently preprocessed and analyzed. After downloading
   the raw data, we utilized the ChAMP package[66]^22 in R software for
   preliminary processing.

   Specifically, we employed the champ.load () function to import and
   screen the data using the following parameters: methValue was set to
   “B” for beta values, autoimpute was enabled to automatically impute
   missing values, and filterDetP was enabled with a detection p-value
   threshold (detPcut) of 0.01 to exclude probes with detection p-values
   exceeding this limit. Probes with bead counts below 3% in at least 10%
   of the samples were removed, along with non-CpG probes, probes with
   multiplex mapping, and probes located on sex chromosomes (X and Y) or
   associated with single nucleotide polymorphisms (SNPs).

   After initial filtering, stringent quality control was performed using
   the champ.QC () function. To normalize the beta value matrix and
   correct for biases between type I and type II probes, we applied the
   BMIQ normalization method through the champ.norm () function. To
   mitigate batch effects, we implemented the champ.runCombat () function
   for batch correction. Additional preprocessing parameters included the
   exclusion of probes based on bead count cutoffs (beadCutoff set to
   0.05) and the adjustment for population-specific genetic variations
   (population parameter set to “EAS”).

Differential Methylation Analysis

   Following the preprocessing of genomic DNA methylation microarray data,
   we utilized the DMP () function from the ChAMP package in R to identify
   differentially methylated positions (DMPs). To determine the
   statistical significance of the detected DMPs, we applied the following
   criteria: a p-value < 0.05, an absolute beta value difference (|β
   difference|) > 0.1,[67]^23 and false discovery rate (FDR) correction
   using the Benjamini-Hochberg procedure, with the adjusted p-value
   threshold set at < 0.05.

Gene Functional Enrichment Analysis

   To gain a comprehensive understanding of the biological functions
   associated with the differentially methylated sites in our dataset, we
   employed the DAVID database ([68]https://david.ncifcrf.gov). Gene
   Ontology (GO) analysis and Kyoto Encyclopedia of Genes and Genomes
   (KEGG) pathway enrichment analysis were subsequently performed on the
   genes containing these differentially methylated sites. Entries with a
   p-value < 0.05 were considered significantly enriched. This approach
   aimed to elucidate the functional roles and biological pathways
   influenced by these genes.

Bisulfite Pyrosequencing

   To validate the reliability of array-based methylation measurements, we
   performed technical replication using bisulfite pyrosequencing on
   candidate CpG sites. Genomic DNA from an independent cohort of 44
   VLOSLP patients and 36 matched healthy controls was bisulfite-converted
   using the EZ DNA Methylation-Gold Kit (Zymo Research) following
   manufacturer protocols.

Correlation Analysis of DNA Methylation in Multiple Tissues

   Recognizing tissue specificity as a defining feature of DNA
   methylation, we utilized the Blood Brain DNA Methylation Comparison
   Tool ([69]https://epigenetics.essex.ac.uk/bloodbrain/) to assess the
   concordance of methylation levels at differentially methylated loci
   between peripheral blood and selected brain regions. Comparative
   analyses were performed to evaluate methylation levels in peripheral
   blood relative to four brain regions: the prefrontal cortex (PFC),
   entorhinal cortex (EC), superior temporal gyrus (STG), and cerebellum
   (CER). A p-value < 0.05 was considered indicative of a statistically
   significant correlation.

The Methylation Quantitative Trait Loci Analysis

   To explore the interplay between alleles and methylation sites, we
   extracted data from two well-established methylation quantitative trait
   loci (meQTL) databases. Specifically, we utilized the meQTL Database
   ([70]https://epicmeqtl.kcl.ac.uk) to perform a detailed analysis of
   meQTL associations in the acquired dataset. Our investigation focused
   exclusively on cis-meQTLs, defining statistically significant single
   nucleotide polymorphism (SNP)-CpG site pairs as those with a p-value ≤
   1 × 10^-8.

Analysis of Linkage Disequilibrium

   In our study, we utilized the Linkage Disequilibrium Score Regression
   (LDSC) tool ([71]https://github.com/bulik/ldsc) to calculate linkage
   disequilibrium (LD) scores for single nucleotide polymorphisms (SNPs)
   within our dataset. The East Asian cohort from the Thousand Genomes
   Project served as the reference population for these computations. SNPs
   were classified as being in linkage disequilibrium if their R² > 0.5 or
   D’ > 0.5. To further elucidate the observed LD patterns, we employed
   the LDmatrix tool to generate graphical representations of the linkage
   disequilibrium structure.

Dimensionality Reduction Analysis

   We utilized three advanced dimensionality reduction techniques to
   extract latent features indicative of mental illness risk: Principal
   Component Analysis (PCA), t-distributed Stochastic Neighbor Embedding
   (t-SNE), and Uniform Manifold Approximation and Projection (UMAP). PCA
   was performed using the prcomp () function within the R statistical
   environment. The t-SNE algorithm was implemented via the R t-SNE ()
   function provided by the corresponding R package
   ([72]https://github.com/jkrijthe/Rtsne), while UMAP was executed using
   the UMAP () function from the dedicated R package
   ([73]https://github.com/tkonopka/umap). To ensure methodological
   consistency and result reliability, all datasets were preprocessed with
   centering and scaling using the preprocess () function from the caret
   package in R. This preprocessing step enabled a rigorous assessment of
   methylation profiles and clinical characteristics using these
   dimensionality reduction techniques.

Machine‑Learning Model Development

   Using the R software and the caret package
   ([74]https://github.com/topepo/caret), we incorporated the observed
   differentially methylated positions (DMPs) and chip type as independent
   variables, with disease classification as the dependent variable. The
   dataset of methylated samples was divided into training and testing
   subsets in a 1:1 ratio. Data in the training set were centered, scaled,
   and transformed to follow a Gaussian distribution using the preprocess
   () function from the caret package prior to model development. To
   address potential underfitting or overfitting, we employed either
   ten-fold cross-validation or leave-one-out validation during the
   training phase.

   Model construction was performed using the XGBoost
   algorithm.[75]^24^,[76]^25 Key hyperparameters were optimized via a
   random search approach, including the number of trees (n estimators),
   maximum tree depth (max depth), minimum samples required for a split
   (min samples split), minimum samples required at a leaf node (min
   samples leaf), learning rate (eta), regularization parameter (C),
   feature sampling rate per tree (colsample bytree), kernel coefficient
   (gamma), sample sampling rate per tree (subsample), and the number of
   boosting iterations (n rounds).

   The classification performance of the predictive models was evaluated
   using the area under the curve (AUC). Clinical utility was further
   assessed via decision curve analysis, implemented using the rmda
   package in R ([77]https://github.com/mdbrown/rmda).

Statistical Analysis

   Statistical analyses were performed using SPSS version 26.0. Continuous
   variables were analyzed using t-tests, ANOVA, or Kruskal–Wallis tests,
   depending on their distribution. Categorical variables were assessed
   with chi-square tests, while Pearson correlation was employed for
   evaluating relationships between continuous variables. A statistical
   power analysis was conducted using G*Power 3.1.9.2 software. Based on
   preliminary experimental data, we set α = 0.05, β = 0.1 (Type II error
   rate of 10%), and estimated an effect size of 0.79. The analysis
   determined that a sample size of 70 participants provided sufficient
   power (1−β = 0.90) for all statistical tests. To account for multiple
   comparisons, we applied false discovery rate (FDR) correction using the
   Benjamini-Hochberg method, with a threshold of q = 0.05.

Results

Overview of Whole Genome DNA Methylation

   Our genome-wide DNA methylation analysis identified significant
   differential methylation sites across 5 VLOSLP patients compared to
   3 healthy control samples. The analysis revealed 2213 hypermethylated
   and 3274 hypomethylated sites ([78]Figure 2A), suggesting distinct
   epigenetic modification patterns between experimental groups.
   Functional domain analysis of hypermethylated sites demonstrated their
   distribution across genomic regions: 296 sites (63%) within promoter
   regions, 32 sites (7%) in 5’ untranslated regions (5’UTR), 13 sites
   (3%) in 3’ untranslated regions (3’UTR), and 125 sites (27%) within
   exonic regions. Similarly, hypomethylated sites exhibited a comparable
   distribution pattern: 488 sites (62%) in promoter regions, 36 sites
   (5%) in 5’UTR, 45 sites (6%) in 3’UTR, and 211 sites (27%) in exonic
   regions ([79]Figure 2B).

Figure 2.

   [80]Figure 2
   [81]Open in a new tab

   Overview of Whole Genome DNA Methylation. (A) Volcano plot illustrating
   the analysis of differential methylation sites in VLOSLP. (B), left:
   Functional domain analysis of hypermethylated sites (B), right:
   Functional domain analysis of hypomethylated sites. (C) Analysis of
   differential methylation sites in the promoter region. (D) Hierarchical
   clustering analysis of VLOSLP and healthy group.

   Further analysis of the 1246 differentially methylated sites in
   promoter regions revealed specific patterns of CpG-related features.
   The distribution comprised 11% CpG islands, while flanking regions
   showed distinct patterns: North (N) and South (S) shores each
   represented 36%, and N and S shelves constituted 7% and 10%,
   respectively ([82]Figure 2C). Hierarchical clustering analysis revealed
   substantial similarities among samples within their respective groups.
   Notably, the disease group samples exhibited higher methylation pattern
   homogeneity and clustered distinctly from control group samples
   ([83]Figure 2D). Additional details are provided in [84]Supplementary
   Figure S1.

Specific Differentially Methylated Sites of VLOSLP

   This investigation examined genome-wide DNA methylation patterns across
   four distinct groups: patients with VLOSLP (5 cases), healthy controls
   (696 cases), individuals with SCZ (161 cases), and patients with AD
   (361 cases). Subsequent gene mapping analysis of these differentially
   methylated sites, visualized through Venn diagram analysis ([85]Figure
   3A), identified 983 uniquely differentially methylated genes,
   suggesting distinct epigenetic signatures associated with VLOSLP.

Figure 3.

   [86]Figure 3
   [87]Open in a new tab

   Specific differential methylation site analysis. (A) Venn diagram
   illustrating the overlap of differentially expressed genes across
   datasets, revealing 983 genes commonly altered with the intersecting
   regions highlighting the shared differentially expressed genes among
   the datasets. (B) Analysis of Specific Differential Methylation Sites
   in VLOSLP.

   To noted, we identified six VLOSLP-specific sites ([88]Figure 3B).
   Statistical analysis revealed varying effect sizes and confidence
   intervals for these sites: cg06324048 (effect size: 2.05 [95% CI:
   0.174–3.824]; cg27661264 (effect size: 2.15 [95% CI: 0.24–3.96];
   cg19428235 (effect size: 2.19 [95% CI: 0.26–4.01]; cg18380528 (effect
   size: −1.60 [95% CI: −3.24–0.13]; cg00638797 (effect size: 0.406 [95%
   CI: −1.06–1.84]; cg00256593 (effect size: −0.482 [95% CI: −1.92–0.99]).

Functional Enrichment Analysis

   Gene Ontology (GO) functional enrichment analysis revealed distinct
   biological focuses in the datasets. The SCZ dataset ([89]Figure 4A) was
   primarily enriched for processes related to glutamate synapse function
   and neuron cell body structure. In contrast, the AD dataset ([90]Figure
   4B) demonstrated enrichment in synapse activity, glutamate synapse,
   neuron cell body, nervous system development, axonogenesis, and
   neuronal protrusion growth.

Figure 4.

   [91]Figure 4
   [92]Open in a new tab

   GO and KEGG enrichment analysis. (A) GO functional enrichment analysis
   of SCZ. (B) GO functional enrichment analysis of AD. The red column
   denotes Biological Processes (BP), the green signifies Cellular
   Components (CC), while the blue column is indicative of Molecular
   Functions (MF). (C) GO functional enrichment analysis of VLOSLP. (D)
   KEGG pathway enrichment analysis of VLOSLP. (E) Network analysis of
   GNB5 related diseases.

   In the independent VLOSLP dataset, functional enrichment analysis of
   the identified differentially methylated positions (DMPs) highlighted
   significant involvement in key biological processes essential to
   nervous system function. These included G protein-coupled receptor
   activity, protein tyrosine kinase activator activity, and the
   regulation of neurotransmitter secretion and phagocytosis mediated by
   calcium ions ([93]Figure 4C).

   KEGG pathway enrichment analysis indicated that the identified DMPs
   were significantly enriched in pathways such as adrenergic signaling in
   cardiomyocytes, cAMP signaling, cGMP-PKG signaling, linoleic acid
   metabolism, arachidonic acid metabolism, and cocaine addiction
   signaling pathways ([94]Figure 4D).

   Furthermore, disease-associated gene network analysis revealed notable
   interactions involving GNB5, GNAS, GNB1, GNAI1, and ADCYAP1R1, forming
   a robust interaction network ([95]Figure 4E).

Bisulfite Pyrosequencing Validation

   To validate our genome-wide methylation findings, we conducted
   bisulfite pyrosequencing on seven neurologically relevant CpG sites
   selected by effect size magnitude (effect size>0.4): (GNB1
   (cg16645714), GNAI1 (cg19428235), RYR2 (cg18380528), GNAS (cg06324048,
   cg27661264), GNB5 (cg00638797), ADCYAP1R1 (cg00256593). The analysis
   revealed locus-specific methylation alterations in VLOSLP patients
   compared to healthy controls. A statistically significant
   hypermethylation was observed at the GNB5 cg00638797 site (Δβ = +14.7%,
   p < 0.001, Mann–Whitney U) ([96]Figure 5). In contrast, the remaining
   sites showed no significant intergroup differences: GNB1 (cg16645714, p
   = 0.753); GNAI1 (cg19428235, p = 0.881); RYR2 (cg18380528, p = 0.097);
   GNAS (cg06324048, p = 0.498); GNAS (cg27661264, p = 0.958); ADCYAP1R1
   (cg00256593, p = 0.923), as shown in [97]Supplementary Table S1.

Figure 5.

   [98]Figure 5
   [99]Open in a new tab

   Pyrophosphate verification of GNB5. Asterisks denote statistical
   significance: ***p < 0.001.

Correlation Analysis of Methylation Levels Between Peripheral Blood and Brain
Tissue

   In VLOSLP peripheral blood, the methylation status of cg13934625
   demonstrated a significant association with the prefrontal cortex (p =
   0.000331) and the entorhinal cortex (p = 0.0025), as shown in
   [100]Figure 6A. Similarly, the methylation level of cg11871050 was
   significantly correlated with the frontal lobe (p = 0.0394), entorhinal
   lobe (p = 0.016), superior temporal lobe (p = 0.0126), and cerebellum
   (p = 0.0408), as illustrated in [101]Figure 6B. For cg00638797, a
   strong correlation was observed with the frontal lobe (p = 0.00091),
   entorhinal lobe (p = 0.0249), superior temporal lobe (p = 0.000696),
   and cerebellum (p = 4.2×10^−16), as depicted in [102]Figure 6C.
   Additional correlations are provided in [103]Supplementary Figure S2.

Figure 6.

   [104]Figure 6
   [105]Open in a new tab

   Correlation of GNB5-linked methylation patterns in blood and brain
   tissue. (A–C) show methylation for cg13934625, cg11871050, and.
   cg00638797.

Genetic-Epigenetic Interaction Analysis

   Analysis of genetic-epigenetic interactions within the GNB5 gene in
   VLOSLP patients, specifically examining the relationship between
   differential methylation sites and methylome quantitative trait loci
   (mQTL), revealed 286 significant single nucleotide polymorphism
   (SNP)-CpG pairs functioning as mQTLs across 349 differentially
   methylated regions (see [106]Supplementary Table S2).

   Linkage disequilibrium analysis, conducted with reference to East Asian
   population allele frequencies, indicates potential associations between
   the GNB5-meQTL and multiple neurobiological phenotypes. These
   associations encompass structural brain characteristics (including
   volumetric variations in specific brain regions), calcium homeostasis,
   and white matter integrity, with particular emphasis on the posterior
   cingulate cortex. Moreover, the GNB5-meQTL appears to correlate with
   pharmacological response to antipsychotic agents and may contribute to
   the pathophysiology of severe depressive disorder (comprehensive data
   presented in [107]Supplementary Table S3).

Development, Validation, and Evaluation of the Predictive Model

   We developed a diagnostic classification model predicated on DNA
   methylation patterns at five specific CpG sites (cg13934625,
   cg14120436, cg19366178, cg19747632, cg20557935) within the GNB5 gene to
   differentiate among SCZ, AD, and VLOSLP. The model’s classification
   performance underwent evaluation using both training and test datasets
   ([108]Figure 7A and [109]B). In the binary classification of AD versus
   VLOSLP (n train = 184(n [AD] = 181; n [VLOSLP] = 3), n test = 182(n
   [AD]= 180; n [VLOSLP] = 2)), the model demonstrated an area under the
   curve (AUC) of 1.0 in the training set and 0.958 (95% CI: 0.875–1.000)
   in the test set. The AD versus SCZ binary classification (n train =
   264(n [AD] = 181; n [SCZ] = 83), n test = 263(n [AD] = 180; n [SCZ] =
   83)) yielded an AUC of 0.995 in the training set and 0.955 (95% CI:
   0.926–0.983) in the test set. For SCZ versus VLOSLP differentiation (n
   train = 84(n [SCZ] = 81; n [VLOSLP] = 3), n test = 82(n [SCZ] = 80; n
   [VLOSLP] = 2)), the model achieved an AUC of 1.0 in both training and
   test sets.

Figure 7.

   [110]Figure 7
   [111]Open in a new tab

   Performance evaluation of predictive models. (A) Performance of the
   training set model. (B) Test set model performance. AUC. Area under the
   curve. (C) Decision Curve Evaluation for Identifying Distinctive
   Features of Three Models in the Test Data. The green curve shows the AD
   vs VLOSLP binary classification model, the red curve shows the AD vs
   SCZ binary classification model, and the blue curve shows the SCZ vs
   VLOSLP binary classification model. (D) Dimensionality reduction
   analysis of GNB5 methylation profile, Montreal Cognitive Assessment
   (MoCA) language function, and Activities of Daily Living (ADL). Error
   bars show data variability (standard error or confidence intervals).

   Note: Asterisks denote statistical significance: *p < 0.05, **p < 0.01.

   Decision curve analysis ([112]Figure 7C) indicates substantial net
   benefits in the test set for all three binary classification models (AD
   vs SCZ, AD vs VLOSLP, and SCZ vs VLOSLP). These findings suggest that
   GNB5 methylation patterns may serve as potential biomarkers for
   distinguishing among these neuropsychiatric conditions, though further
   validation studies are warranted.

   The outcomes of the sensitivity assessment, subgroup examination, and
   Bootstrap testing indicate a high degree of robustness for the model.
   Details are provided in [113]Supplementary Figure S3.

   Dimensionality reduction analysis revealed associations between GNB5
   methylation profiles and specific clinical parameters. Notably,
   variations in language performance scores on the Montreal Cognitive
   Assessment (MoCA) and Activities of Daily Living (ADL) scores,
   stratified by GNB5 methylation levels, appear to contribute to the
   identification of VLOSLP subtypes ([114]Figure 7D).

Discussion

   This study identified significant DNA methylation changes in the blood
   of patients with VLOSLP, primarily affecting gene promoter regions. Our
   enrichment analysis suggests that pathways related to neurotransmitter
   regulation, inflammation, and cognitive function may play central roles
   in VLOSLP pathogenesis. Notably, we observed a strong
   genetic-epigenetic interaction involving the GNB5 gene,[115]^26 which
   is critical for brain structure and has been implicated in various
   psychiatric disorders. By leveraging the methylation status of GNB5, we
   developed and validated a predictive model that demonstrated excellent
   classification performance and clinical utility. Moreover, we
   identified distinct VLOSLP subtypes with unique clinical features,
   emphasizing the disorder’s heterogeneity.

   The observed DNA methylation alterations primarily affected promoter
   regions of genes involved in neurotransmitter pathways. As DNA
   methylation modulates gene expression, increased methylation levels may
   downregulate genes associated with the synthesis, release, and
   metabolism of neurotransmitters,[116]^27 thereby affecting the balance
   of the neurotransmitter system. For instance, the methylation status of
   GNB5 is closely related to VLOSLP occurrence. Studies have shown that
   GNB5 is significantly enriched in the brain, particularly in the
   hippocampus and striatum[117]^28 and is integral to neurotransmitter
   signal transduction.[118]^29 Altered GNB5 methylation may thus
   contribute to neurotransmitter dysfunction and the pathogenesis of
   VLOSLP.

   Our study revealed distinct neurotransmitter pathway-specific
   methylation alterations that significantly differ from findings in SCZ
   epigenetic research. Previous genome-wide methylation studies of SCZ
   primarily identified methylation abnormalities in glutamatergic
   receptors (GRM3,[119]^30 GRIN2A[120]^31) and dopamine receptor
   DRD2,[121]^32 whereas our findings demonstrate predominant methylation
   changes in GNB5, a key gene in the G protein signaling transduction
   pathway. This discrepancy suggests that VLOSLP may involve different
   regulatory targets compared to typical SCZ. Specifically, VLOSLP
   appears to predominantly involve G protein-mediated transmembrane
   signaling transduction, while typical SCZ more frequently associates
   with receptor functional impairments.

   Comparative analysis with AD revealed that although both conditions
   exhibit epigenetic modifications in cognition-related pathways, AD
   characteristic methylation loci predominantly cluster in amyloid
   precursor protein metabolism[122]^33^,[123]^34 (ANK1, APP, PSEN1) and
   tau phosphorylation regulation[124]^35 (MAPT). In contrast, the GNB5
   methylation signature identified in our study may reflect synaptic
   plasticity impairments - a potential VLOSLP-specific pathological
   mechanism.

   Besides, the methylation characteristics of VLOSLP demonstrate a
   distinct pattern from late-onset major depressive disorder (MDD) and
   anxiety disorders. In MDD, where elevated methylation of
   neuroplasticity-related genes (BDNF) and stress-responsive loci (NR3C1)
   is linked to depression risk.[125]^36 What’s more, generalized anxiety
   disorder[126]^37^,[127]^38 (GAD), shows methylation changes at
   neuroinflammation-associated CpG sites (RIOK3 cg21515243, PSMB4
   cg01334186). However, VLOSLP predominantly involves G protein-coupled
   receptor (GNB5) and calcium-mediated synaptic plasticity pathways.

   Notably, the epigenetic dysregulation observed in VLOSLP exhibits
   pathway-level convergence with neurodegenerative disorders beyond AD.
   This is particularly evident in GPCR signaling pathways — as evidenced
   by the involvement of inhibin-mediated GPCR modulation in
   frontotemporal dementia pathogenesis.[128]^39 Furthermore, calcium
   signaling alterations in VLOSLP show functional parallels to
   neurodegenerative mechanisms, such as the calcium dyshomeostasis
   observed in striatal neurons in Huntington’s disease,[129]^40 though
   these disorders diverge in their specific molecular targets within
   shared calcium-regulated neurotransmitter pathways.

   Aging-related epigenetic changes appear to play a key role in VLOSLP.
   Epigenetic modifications accumulate with age,[130]^41 influencing gene
   expression and potentially leading to cognitive decline[131]^42 and
   neuropsychiatric symptoms.[132]^43 In this study, DNA methylation
   alterations associated with VLOSLP suggest that age-related epigenetic
   remodeling may underlie the disorder’s late onset. Shifts in
   methylation patterns of neurotransmitter-related genes might compromise
   cognitive function and promote neuropsychiatric manifestations as
   individuals grow older.

   Epigenetic regulation, including DNA methylation, significantly affects
   neurocognitive processes. It can impair neuronal metabolism, synaptic
   regulation, neurogenesis, and neuroplasticity,[133]^44^,[134]^45 all of
   which are vital for maintaining cognitive function. Consistent with
   previous findings,[135]^5 our MoCA assessments revealed cognitive
   deficits in VLOSLP patients, aligning with the observed methylation
   alterations at specific loci. These changes may further disturb
   neurotransmitter equilibrium, exacerbating cognitive dysfunction.

   From a clinical perspective, our findings underscore the potential of
   DNA methylation markers as biological tools for diagnosis, treatment
   planning, and prognosis.[136]^46 In this study, DNA methylation markers
   exhibited high performance (AUC: 0.958 to 1) for diagnosing VLOSLP and
   can be integrated with existing clinical scales for more accurate early
   screening. Additionally, they show promise for predicting drug
   responses and monitoring treatment efficacy. Identification of
   prognosis-related methylation sites could guide early interventions,
   inform personalized treatment strategies, and support dynamic
   monitoring of disease progression and recurrence risk.

   Although numerous epigenetic studies in SCZ focus on late adolescence
   or early adulthood,[137]^47 most emphasize isolated loci, non-coding
   RNAs, epigenetic aging, or telomere shortening.[138]^8 There remains a
   paucity of empirical research specifically examining first-episode SCZ
   in individuals aged 40 or above, or 60 and older, with existing studies
   often limited to clinical follow-up or cross-sectional
   comparisons.[139]^5 A PubMed search (2000–August 2024) for “late-onset
   schizophrenia-like psychosis” yielded 95 publications, of which only 3
   were clinical trials focused solely on antipsychotic
   treatments.[140]^48–50 By conducting a whole-genome DNA methylation
   study on VLOSLP, we address this gap, demonstrating unique epigenetic
   signatures and highlighting the importance of neurotransmitter-,
   inflammatory-, and cognition-related pathways in disease pathology.

   VLOSLP presents particular diagnostic challenges, especially when
   distinguishing it from other age-related neurodegenerative conditions
   such as AD, frontotemporal dementia, and Lewy body de
   entia.[141]^51^,[142]^52 Cognitive impairments in SCZ and AD overlap
   across domains including attention, memory, and language.[143]^53 Our
   predictive model, based on five CpG sites within GNB5, effectively
   differentiated VLOSLP from AD, providing a novel biomarker that could
   refine differential diagnosis and deepen our understanding of cognitive
   dysfunction across these conditions.

   Importantly, the epigenetic patterns we identified in VLOSLP differ
   markedly from known aging-related epigenetic markers.[144]^54 These
   findings suggest that aging-related epigenetic mechanisms might be
   central to VLOSLP pathogenesis. Recent epigenetic meta-analyses[145]^55
   further support the relevance of factors such as age, gender, disease
   duration, and polygenic risk in SCZ, reinforcing the complexity of
   epigenetic influences in late-onset presentations.

   This study offers several innovations. First, we systematically
   characterized VLOSLP’s DNA methylation landscape for the first time,
   identifying distinct epigenetic markers that differentiate VLOSLP from
   SCZ and AD. Second, we developed a DNA methylation analysis framework
   for elderly psychiatric disorders, integrating multi-omics strategies
   (genomics and epigenetics) with clinically useful predictive models.
   Theoretically, we proposed specific hypotheses regarding age-related
   epigenetic mechanisms in VLOSLP, establishing patterns of association
   between epigenetic changes and clinical phenotypes. Practically, we
   proposed diagnostic, therapeutic, and prognostic models based on
   epigenetic markers.

   Despite these strengths, some limitations must be acknowledged. The
   relatively small sample size reduces statistical power, potentially
   explaining why certain VLOSLP-specific sites did not reach significance
   after correction. Larger cohorts are needed for validation. In
   addition, SCZ vs VLOSLP, AD vs VLOSLP classifiers need to be
   interpreted with caution and require external dataset validation. In
   addition, the prediction model includes blood and brain tissue samples,
   which may cause some confusion. Methodological constraints persist in
   DNA methylation testing, bioinformatics approaches, and
   phenotype-genotype association studies.[146]^56 Confounding variables,
   including age, gender, medication use, comorbidities, and lifestyle,
   may influence our results. Additionally, the follow-up lacked and
   incomplete phenotypic data collection limit our capacity for deeper
   mechanistic understanding. Finally, our research is still in the
   conceptual validation stage, and the directionality of the observed
   associations is still unclear without functional experiments.

   Future research should thus expand sample sizes, extend follow-up
   durations, incorporate functional validations, enrich clinical data,
   integrate multi-omics peripheral blood data, explore multi-class
   models, and apply multivariate analytical methods to mitigate
   confounders. Further explore the heterogeneity of VLOSLP subtypes.
   Through these steps, we can refine our understanding of VLOSLP’s
   epigenetic underpinnings, ultimately improving diagnostic accuracy,
   treatment efficacy, and prognostic precision.

Conclusions

   This study identifies novel DNA methylation alterations in VLOSLP,
   revealing distinct epigenetic biomarkers and a highly accurate
   diagnostic model (85%). These findings differentiate VLOSLP from other
   psychiatric disorders, implicate key pathways, and offer new targets
   for personalized treatment, advancing both diagnostic and therapeutic
   strategies.

Acknowledgments