Abstract

Background

   MicroRNAs (miRNAs), small and highly conserved non-coding RNA
   molecules, have emerged as promising molecular biomarkers due to their
   regulatory roles in gene expression and stability in blood.

Methods

   We used measurements of 64 plasma miRNAs from 145 participants in the
   Alzheimer’s disease Neuroimaging Initiative cohort, including 74
   probable AD patients and 71 cognitively normal (CN) older adults. We
   performed principal component analysis (PCA) with factor rotation for
   dimension reduction to identify AD-associated principal components
   (PCs) and their key miRNAs with factor loadings higher than 0.8. We
   investigated their association with amyloid/tau/neurodegeneration
   (A/T/N) biomarkers and cognition. After identifying the candidate
   target genes of key miRNAs, we performed pathway enrichment analysis.
   We conducted mediation analyses to assess the effect of the
   associations between miRNAs and A/T/N biomarkers on AD diagnosis and
   cognition. Finally, we used a machine learning approach to evaluate the
   performance of key miRNAs for AD classification.

Results

   PCA identified one PC as significantly associated with AD. The PC was
   also significantly associated with CSF p-tau levels, hippocampal
   volume, and cognition. Two key miRNAs (miR-423-5p and miR-92a-3p) in
   the PC were associated with AD. Lower levels of miR-423-5p and
   miR-92a-3p were associated with reduced hippocampal volume and worse
   cognition, and lower levels of miR-423-5p were associated with higher
   brain amyloid deposition. Pathway enrichment analysis identified
   several significant biological processes, including memory, protein
   phosphorylation, and the phosphatidylinositol-3-phosphate biosynthetic
   process. Mediation analysis revealed that miR-423-5p, but not
   miR-92a-3p, had indirect effects on AD diagnosis and memory performance
   through brain amyloid deposition and brain atrophy. Machine learning
   analysis demonstrated that incorporating two key miRNAs improved the
   performance of demographic information for AD classification.

Conclusion

   Plasma miR-423-5p and miR-92a-3p are implicated in AD pathology and
   cognitive decline, providing insights into their roles in disease
   mechanisms. This study suggests the potential of these miRNAs as
   blood-based molecular biomarkers for AD.

   Keywords: Alzheimer’s disease, cognitive decline, machine learning,
   MicroRNAs, miR-423-5p, miR-92a-3p

1. Introduction

   Alzheimer’s disease (AD) is a progressive neurodegenerative disorder
   that stands as a common cause of dementia among older adults ([35]Long
   and Holtzman, 2019). AD manifests through a constellation of
   neuropathological features, including the extracellular accumulation of
   amyloid-β (Aβ) plaques along with intracellular neurofibrillary tangles
   (NFTs) composed of hyperphosphorylated tau protein aggregates ([36]Long
   and Holtzman, 2019). Despite extensive research efforts, the
   pathogenesis of AD remains elusive, largely due to its complex
   neurochemical and genetic landscape ([37]Hampel et al., 2010).

   Significant advancements have been made in antemortem diagnosis of AD
   through the use of biomarkers obtained from cerebrospinal fluid (CSF)
   or visualized via positron emission tomography (PET) imaging ([38]Long
   and Holtzman, 2019; [39]Schindler et al., 2024). These biomarkers have
   proven to be highly effective in identifying the pathological hallmarks
   of the disease ([40]Hampel et al., 2018; [41]Long and Holtzman, 2019;
   [42]Schindler et al., 2024). However, the invasive nature of CSF
   collection and the high costs of PET imaging pose significant
   limitations to their widespread application in routine clinical
   practice ([43]Hampel et al., 2018; [44]Schindler et al., 2024).
   Consequently, there is a critical need for the identification of more
   accessible and less invasive biomarkers, such as blood-based ones, for
   AD diagnosis.

   In recent years, microRNAs (miRNAs), which are non-coding RNA molecules
   approximately 20–23 nucleotides long, have emerged as potential
   biomarkers for AD ([45]Wang et al., 2019; [46]Kou et al., 2020). These
   miRNAs are involved in AD pathogenesis, playing important roles in
   post-transcriptional gene expression regulation by binding the
   3′-untranslated regions of messenger RNAs (mRNAs) ([47]Bartel, 2004;
   [48]Winter et al., 2009; [49]Wang et al., 2019; [50]Kou et al., 2020).
   Notably, alterations in the levels of specific miRNAs in peripheral
   blood have been reported in AD patients, suggesting their potential
   utility as non-invasive biomarkers for AD ([51]Kou et al., 2020;
   [52]Guévremont et al., 2022). Given the stability of miRNAs in
   circulation and their detectability in various biological fluids
   including plasma, they represent promising candidates for the
   development of blood-based diagnostic biomarkers for AD ([53]Weber et
   al., 2010; [54]Turchinovich et al., 2011).

   A critical step in analyzing high-dimensional biological data, such as
   miRNA expression data, is dimension reduction. This step is essential
   for simplifying complex datasets to enhance data interpretability while
   retaining the biological significance of the data ([55]Meng et al.,
   2016). Principal component analysis (PCA) is a widely utilized method
   for dimension reduction, enabling the identification of underlying
   patterns in data that may not be apparent in high-dimensional space
   ([56]Ringnér, 2008; [57]Meng et al., 2016). Through the application of
   PCA, essential features are extracted from an extensive set of miRNAs,
   which helps elucidate the specific roles of miRNAs in AD and deepens
   understanding of their dysregulated expression in the context of the
   disease ([58]Ringnér, 2008; [59]Jolliffe and Cadima, 2016; [60]Meng et
   al., 2016).

   In this study, we investigated the association of circulating miRNAs
   with central AD biomarkers and their potential as candidate blood-based
   molecular biomarkers for AD using expression data from 64 individually
   measured plasma miRNAs from 145 participants in the Alzheimer’s Disease
   Neuroimaging Initiative (ADNI) dataset. We performed PCA to identify
   AD-associated principal components (PCs) and their key miRNAs.
   Subsequently, we also investigated their association with
   amyloid/tau/neurodegeneration (A/T/N) biomarkers and cognitive
   function, aiming to contribute to the identification of potential
   blood-based molecular biomarkers for AD. After identifying the
   candidate target genes of these miRNAs, we investigated their
   involvement in biological pathways to provide insights into the
   molecular mechanisms underpinning AD. Furthermore, we performed
   mediation analysis to assess the effect of the associations between
   miRNAs and A/T/N biomarkers on AD diagnosis and cognitive function.
   Finally, we conducted a machine learning approach to evaluate the
   performance of miRNAs for the classification of AD.

2. Materials and methods

2.1. Study samples

   The ADNI was launched in 2003 through a collaborative effort including
   the National Institute on Aging, National Institute of Biomedical
   Imaging and Bioengineering, Food and Drug Administration, private
   pharmaceutical companies, and nonprofit organizations. The initial
   phase, ADNI-1 ([61]Petersen et al., 2010), aimed to assess the
   feasibility of using serial magnetic resonance imaging (MRI), PET
   scans, various biological markers, and clinical and neuropsychological
   assessments as reliable indicators of AD pathogenesis. This was
   followed by subsequent phases—ADNI-GO ([62]Aisen et al., 2010), ADNI-2
   ([63]Aisen et al., 2015), and ADNI-3 ([64]Weiner et al., 2017)—which
   expanded on the initial phase by allowing for the continuous follow-up
   of existing participants and incorporating new enrollments. Detailed
   information on ADNI, including latest updates, eligibility criteria,
   protocols for clinical and neuroimaging evaluations, and an outline of
   diagnostic criteria is available at [65]https://www.adni-info.org. The
   ADNI Laboratory of Neuro Imaging (LONI) website[66]^1 provided access
   to demographic and clinical information, raw neuroimaging data, CSF
   biomarker data, apolipoprotein E (APOE) ε4 genotyping, and cognitive
   scores. All participants provided written informed consent at the time
   of enrollment, which included consent for data analysis and sharing.
   The study received approval from the Institutional Review Board at each
   participating site. Participants diagnosed with probable AD and
   cognitively normal (CN) were followed up prospectively with clinical
   data, neuroimaging studies, and biological samples for molecular
   biomarker measurements, as previously described ([67]Aisen et al.,
   2010; [68]Petersen et al., 2010; [69]Aisen et al., 2015; [70]Weiner et
   al., 2017). Diagnostic classifications were based on assessments
   including the Logical Memory from the Wechsler Memory Scale—Revised
   ([71]Wechsler, 1987), Mini-Mental State Examination (MMSE)
   ([72]Folstein et al., 1975), and Clinical Dementia Rating scale
   ([73]Morris, 1993).

2.2. miRNA expression data

   To assess concordance of miRNA expression between CSF and plasma, the
   same RT-qPCR workflow was applied to both biofluids using a custom
   TaqMan Advanced Human miRNA Low-Density Array (TLDA) card containing 64
   preselected miRNAs. These miRNAs were selected based on prior
   literature demonstrating their detectability in CSF ([74]Lusardi et
   al., 2016; [75]Saugstad et al., 2017; [76]Wiedrick et al., 2019;
   [77]Sandau et al., 2020b; [78]Sandau et al., 2022) and plasma
   ([79]Sandau et al., 2020a) and their established relevance to AD. RNA
   isolation was conducted in 24-sample batches, including matched CSF and
   plasma samples to mitigate batch effects. Total RNA was extracted from
   250 μL each of CSF and plasma using the miRNA Purification kit,
   following the manufacturer’s protocol ([80]Abdalla et al., 2011). A 3
   pM concentration of the exogenous spike-in cel-miR-39-3p was added
   following the lysis step. RNA elution was performed with 30 μL of
   nuclease-free water for CSF samples and 50 μL for plasma samples.
   Quantification of miRNA was performed using the Qubit miRNA Assay kit
   and Qubit 4.0 fluorometer ([81]Scientific, 2017). For reverse
   transcription quantitative polymerase chain reaction, miRNA was reverse
   transcribed to complementary DNA (cDNA) using the TaqMan Advanced miRNA
   cDNA Synthesis kit according to the manufacturer’s instructions
   ([82]TaqMan, 2011). This process included poly-adenylation and adaptor
   ligation, followed by reverse transcription, and a universal 14-cycle
   miRNA amplification. The diluted miRNA amplification reaction mixture
   was combined with TaqMan Fast Advanced master mix and then loaded onto
   a custom TLDA card for assay on a QuantStudio 7 Flex. Data analysis
   included a detailed examination of quantification cycle (Cq) values for
   artifacts and quality control (QC) using array card well-level
   statistics. For both CSF and plasma samples, Cq values were normalized
   using the exogenous spike-in control cel-miR-39-3p and the endogenous
   reference hsa-miR-16-5p ([83]Moldovan et al., 2014). Handling of
   censored values (Cq > 34) was performed through censoring-aware
   methods, as previously described in detail ([84]Wiedrick et al., 2019;
   [85]Sandau et al., 2024).

   All 64 miRNAs passed predefined QC thresholds and were retained for
   downstream analysis. A total of 64 miRNAs were analyzed from 80
   probable AD patients and 80 CN individuals. Participants who lacked one
   or more miRNA expression data were excluded for subsequent PCA,
   resulting in 74 probable AD patients and 71 CN individuals included for
   further analysis.

2.3. Neuroimaging data

   Brain MRI scans were obtained from participants using 3 T scanners,
   following the ADNI standardized protocol for 3D Magnetization Prepared
   Rapid Gradient Echo sequences. The scans were processed through the
   longitudinal pipeline of FreeSurfer version 5.3, as previously
   described ([86]Jack et al., 2008; [87]Jack et al., 2010), enabling the
   extraction of critical regions of interest, such as the bilateral
   hippocampal volumes and total intracranial volume.

   Additionally, preprocessed [18F] florbetapir PET scans were sourced
   from the ADNI database via the LONI platform (see text footnote 1,
   respectively) and were acquired in line with the protocols established
   by previous studies ([88]Jagust et al., 2010; [89]Jagust et al., 2015).
   The quantification of amyloid burden was conducted by analyzing the
   standard uptake value ratio (SUVR) values from [18F] florbetapir PET
   scans, which is essential for evaluating amyloid deposits in the brain.
   These SUVR values were normalized to the intensity measurements from
   the cerebellum, which served as a reliable reference region ([90]Joshi
   et al., 2012). To correct for skewness in the distribution of
   florbetapir PET SUVR data, a logarithmic transformation was applied.

2.4. CSF biomarkers

   CSF samples were collected through lumbar puncture in the morning,
   following an overnight fast. These samples were frozen within 1 h of
   collection and transported to the ADNI Biomarker Core Laboratory at the
   University of Pennsylvania Medical Center, according to the ADNI
   protocol ([91]Kang et al., 2015). Concentrations of Aβ42 and
   phosphorylated tau181 (p-tau) in CSF were measured using the Roche
   Elecsys Aβ42 and p-tau CSF immunoassays, respectively, as previously
   outlined ([92]Bittner et al., 2016; [93]Hansson et al., 2018). The CSF
   biomarker data were accessed and downloaded from the ADNI LONI website
   (see text footnote 1, respectively). To correct for data skewness, a
   logarithmic transformation of the CSF biomarker data was performed.

2.5. Comprehensive neuropsychological assessment

   Domain-specific cognitive composite scores were calculated for memory
   and executive function. The composite score for memory integrated
   results from the Alzheimer’s Disease Assessment Scale-Cognitive
   Subscale ([94]Mohs et al., 1997), the Rey Auditory Verbal Learning Test
   ([95]Rey, 1983), the memory components of the MMSE ([96]Folstein et
   al., 1975), and the Logical Memory task ([97]Wechsler, 1987). For
   executive function, the composite score included assessments from the
   Wechsler Adult Intelligence Scale–Revised Digit Symbol Substitution
   task ([98]Wechsler and De Lemos, 1981) and the Digit Span backward task
   ([99]Wechsler, 1987), the Trail Making Test Parts A and B ([100]Reitan
   and Wolfson, 1985), category fluency (animals and vegetables)
   ([101]Morris et al., 1989), and five clock drawing tasks. To ensure
   comparability and consistency across individuals within the ADNI
   cohorts, these composite scores were standardized to a mean value of 0
   and a standard deviation of 1.

2.6. A/T/N biomarkers

   We used CSF Aβ42 levels and global cortical amyloid deposition measured
   from amyloid PET scans as biomarkers of Aβ (A), CSF p-tau levels as a
   biomarker of tau (T), and structural hippocampal atrophy on brain MRI
   as a biomarker of neurodegeneration (N), as described in the National
   Institute on Aging–Alzheimer’s Association Research Framework
   ([102]Jack et al., 2024).

2.7. Dimension reduction in miRNA expression analysis using PCA

   The procedure for PCA commenced with the extraction of PCs followed by
   an orthogonal rotation to maximize the variance explained by each PC,
   thus streamlining the data structure while preserving critical
   biological information ([103]Jolliffe, 2002; [104]Ringnér, 2008). The
   selection process for PCs was based on rigorous criteria, only
   retaining components with eigenvalues over 1. This threshold is widely
   accepted for ensuring the inclusion of PCs representing significant
   variance and relevance ([105]Jolliffe, 2002; [106]Peres-Neto et al.,
   2005). A factor loading threshold exceeding 0.8 was applied to identify
   the key miRNAs for each rotated PC, thereby focusing exclusively on
   miRNAs that have a substantial impact on the PCs. This selective
   approach aligns with methodologies that aim to improve data
   interpretability and significance ([107]Fabrigar et al., 1999),
   excluding miRNAs with lower loadings from further analysis.

2.8. Identification of target genes for key miRNAs in AD-associated PCs

   To understand the biological implications of the key miRNAs identified
   in the PCs associated with AD, their target genes were predicted using
   two well-known databases: TargetScan ([108]Lewis et al., 2005) and
   miRDB ([109]Wong and Wang, 2015). These databases are pivotal in the
   field of miRNA research, offering a comprehensive collection of both
   predicted and experimentally validated interactions that outline the
   relationships between miRNAs and their corresponding target genes. To
   ensure the robustness of our analysis, we applied a rigorous selection
   criterion for further analysis, focusing only on target genes
   consistently identified in both databases.

2.9. Pathway-based enrichment analysis of target genes

   To explore intricate molecular pathways and biological processes
   implicated by the target genes of AD-associated key miRNAs, a gene-set
   enrichment analysis was performed using the Database for Annotation,
   Visualization, and Integrated Discovery (DAVID) online resource
   ([110]Huang et al., 2009). DAVID is widely recognized for its
   comprehensive analytical framework, enabling the systematic
   identification of functionally related gene groups and their associated
   biological pathways. We specifically focused on the Gene Ontology (GO)
   - particularly the Biological Processes (BP) category ([111]Ashburner
   et al., 2000). To adjust for multiple testing, Bonferroni correction
   ([112]Dunn, 1961) was applied by dividing the significance threshold by
   the number of GO-BP terms tested within this specific category.

2.10. Statistical analysis

   For the PCs of miRNAs, logistic regression models were used to evaluate
   the association of the PCs and key miRNAs with AD diagnosis to identify
   the AD-associated PCs and key miRNAs. Linear regression models were
   used to assess the association of the AD-associated PCs and key miRNAs
   with global cortical amyloid deposition, hippocampal volume, and CSF
   biomarkers for AD, including Aβ42 and p-tau. Moreover, linear
   regression models were used to investigate the association of the
   AD-associated PCs and key miRNAs with baseline composite scores for
   memory and executive function, while linear mixed effects models were
   used to evaluate their association with longitudinal changes in the
   composite scores of memory and executive function. In all association
   analyses, age and sex were consistently included as covariates.
   Additional covariates included years of education for cognitive
   performance, APOE ε4 carrier status for global cortical amyloid
   deposition, CSF Aβ42, and CSF p-tau, and intracranial volume for
   hippocampal volume. For the association analysis of PCs with AD
   diagnosis, the false discovery rate (FDR) correction with the
   Benjamini-Hochberg procedure was used to adjust for multiple testing
   ([113]Benjamini and Hochberg, 1995).

   To assess the potential modifying effects of sex on the associations
   between miRNAs and A/T/N biomarkers, we conducted stratified linear
   regression analyses separately in males and females. The stratification
   variable sex was not included as a covariate in the sex-stratified
   analyses. Covariate-adjusted regression coefficients were transformed
   into sample size-weighted standardized effect sizes (Cohen’s d) to
   formally compare the strength of associations between subgroups
   ([114]Paternoster et al., 1998). A two-sided p-value was calculated
   from the t-statistic using a standard normal distribution to assess
   sex-specific differences in A/T/N biomarkers based on standardized
   effect sizes. Additionally, we quantified the magnitude of effect
   heterogeneity between subgroups using the I^2 statistic ([115]Higgins
   and Thompson, 2002). An I^2 value above 50% was considered indicative
   of substantial heterogeneity. To complement the subgroup difference
   analyses, we also included an interaction term between sex and miRNA in
   the pooled linear regression model. The p value for the interaction
   term was calculated to evaluate whether the association between each
   miRNA and A/T/N biomarkers significantly differed between male and
   female subgroups. In accordance with previously established methodology
   ([116]Arnold et al., 2020), we classified miRNA-A/T/N biomarker
   associations into homogeneous effects if miRNAs showed very similar
   effects in their association to the biomarker for both sexes (i.e.,
   estimated heterogeneity p > 0.05), and heterogeneous effects if miRNAs
   showed different effects in both sexes leading to significant
   heterogeneity (i.e., estimated heterogeneity p < 0.05) and/or the
   interaction between sex and miRNA. Effects that were significant in
   only one sex with either significant effect heterogeneity between males
   and females or significant interaction between sex and miRNA were
   considered sex-specific.

   For the key miRNAs, we performed Pearson correlation analyses of their
   expression levels between plasma and CSF in participants with paired
   biofluid samples.

   We performed mediation analysis on the key miRNAs demonstrating a
   significant correlation with AD diagnosis and a composite score of
   memory, following established methodologies ([117]Hayes, 2009). Our
   investigation aimed to determine whether the identified relationships
   could be mediated by biomarkers for AD. For these analyses, we used the
   Mediation R package, which allowed for the evaluation of indirect
   effects through a bootstrapping approach ([118]Tingley et al., 2014).

   To investigate the classification performance of AD-associated key
   miRNAs in distinguishing AD patients from CN individuals, we applied
   the STREAMLINE tool ([119]Urbanowicz et al., 2023), using an extreme
   gradient boosting machine learning approach. This machine learning
   approach is selected for integrating multi-view biological data and
   heterogeneous feature sets for the analysis of complex biological
   systems ([120]Li et al., 2018). The dataset was randomly partitioned
   into 70% for training the model and 30% for testing, with a 5-fold
   cross-validation process to enhance the robustness and reliability of
   model evaluation. The classification performance of two different
   models in differentiating AD from CN was assessed through the receiver
   operating characteristic (ROC) curve and the area under the receiver
   operating characteristic curve (AUC). In terms of feature selection,
   Model 1 included age, sex, and APOE ε4 carrier status, while Model 2
   incorporated all features from Model 1 and additionally included
   AD-associated key miRNAs, aiming to investigate their additional
   classification performance values in distinguishing AD from CN.

   All statistical analyses were performed using R software, version
   4.2.2. We defined statistical significance at a p value threshold of
   less than 0.05, incorporating adjustments for multiple comparisons. The
   workflow of all analysis steps used in this study is shown in
   [121]Figure 1.

Figure 1.

   [122]Flowchart illustrating a study of plasma miRNA expression profiles
   in an ADNI cohort of 145 individuals. Steps include miRNA isolation,
   RT-qPCR, and expression profiling. Principal component analysis
   identifies key miRNAs linked to PC and biomarkers. Data analysis
   involves gene prediction, pathway enrichment, mediation analysis, and
   machine learning classification. Various icons and graphs represent
   each stage, showing the process from sample preparation to data
   analysis and interpretation.
   [123]Open in a new tab

   Schematic overview of the workflow of our analysis. A/T/N,
   amyloid/tau/neurodegeneration; AD, Alzheimer’s disease; ADNI,
   Alzheimer’s Disease Neuroimaging Initiative; CN, cognitively normal;
   miRNA, microRNA; mRNA, messenger RNA; PC, principal component; RT-qPCR,
   reverse transcription quantitative polymerase chain reaction.

3. Results

3.1. Demographic and clinical characteristics

   We used miRNA expression data from 145 ADNI participants, including 74
   probable AD patients and 71 CN individuals, as detailed in [124]Table
   1. The median age was 74.5 years, with 56.3% being females. No
   significant differences were observed in sex distribution and age
   across the diagnostic groups. However, a substantial difference in APOE
   ε4 carrier status was noted, with a higher prevalence in AD patients
   compared to CN individuals. For the A/T/N biomarkers, CSF Aβ42 and
   hippocampal volume were significantly lower, and CSF p-tau levels and
   global cortical amyloid deposition were significantly higher in AD
   patients compared to CN individuals. Furthermore, composite scores for
   memory and executive function were significantly lower in AD patients
   compared to CN individuals.

Table 1.

   Demographic information of participants from the ADNI cohort.
   CN (n = 71) AD (n = 74) Total (n = 145) p value^b
   Male (%) 36 (50.7) 44 (59.5) 80 (55.2) 0.289
   Age, years^a 73.7 (69.7–78.8) 75.7 (69.1–80.7) 74.5 (69.6–79.7) 0.335
   Education, years^a 16 (14–18) 16 (13–16) 16 (14–18) 0.011
   APOE ε4 carrier (%) 21 (29.6) 50 (67.6) 71 (49.0) <0.001
   CSF Aβ42, pg/mL^a 1255.5 (826.5–1800.5) 650.2 (482.7–814.0) 809.7
   (577.6–1411.0) <0.001
   CSF p-tau, pg/mL^a 19.2 (16.7–24.2) 33.6 (27.4–47.5) 26.5 (18.7–36.7)
   <0.001
   Amyloid PET global SUVR^a 1.04 (0.98–1.19) 1.46 (1.32–1.58) 1.24
   (1.03–1.47) <0.001
   Hippocampal volume, mm^3a 3,753 (3,520 – 4,000) 3,020 (2,794 – 3,266)
   3,385 (2,974 – 3,777) <0.001
   Intracranial volume, mm^3a 1,481,766 (1,366,350 – 1,633,833) 1,458,122
   (1,371,248 – 1,592,867) 1,469,878 (1,369,858 – 1,629,653) 0.899
   Baseline memory, Z scores^a 0.93 (0.56–1.23) −0.75 (−1.00 – −0.58)
   −0.07 (−0.75–0.88) <0.001
   Baseline executive function, Z scores^a 0.68 (0.42–1.07) −0.39
   (−0.97–0.07) 0.24 (−0.46–0.73) <0.001
   [125]Open in a new tab

   Values are n (%), unless indicated otherwise. ^aData are presented as
   median followed by (interquartile range). ^bThe Mann–Whitney U test or
   chi-square test was used to determine the p value for comparisons
   between AD and CN groups, as appropriate. Aβ, amyloid-β; AD,
   Alzheimer’s disease; ADNI, Alzheimer’s Disease Neuroimaging Initiative;
   APOE, apolipoprotein E; CN, cognitively normal; CSF, cerebrospinal
   fluid; p-tau, phosphorylated tau; PET, positron emission tomography;
   SUVR, standard uptake value ratio.

3.2. Principal component analysis for dimension reduction of miRNA expression
data

   PCA was performed on 64 miRNAs in 74 probable AD patients and 71 CN
   individuals from the ADNI cohort. Dimension reduction with PCA resulted
   in 13 PCs with eigenvalues > 1 ([126]Supplementary Table S1). After
   selecting key miRNAs with a factor loading ≥ 0.8 in each principal
   component, 7 of 13 PCs remained for further analysis
   ([127]Supplementary Table S2).

   Association analysis of the 7 PCs with AD diagnosis identified PC2 as
   significantly associated with AD after multiple comparison correction
   (odds ratio (OR) = 0.59, FDR-corrected p = 0.038; [128]Figure 2;
   [129]Supplementary Table S3).

Figure 2.

   Box plot with violin plots comparing PC2 component scores between CN
   (control group) and AD (affected group). CN shows a broader score
   distribution centered around zero, while AD has a higher concentration
   of scores above zero. An asterisk denotes a significant difference
   between groups.
   [130]Open in a new tab

   Association analysis results of PC2 with AD diagnosis Violin and box
   plots represent the PC2 component score for AD diagnosis. AD,
   Alzheimer’s disease; CN, cognitively normal; PC, principal component.
   Note: significance stars indicating the p values of the correlations *
   p-value < 0.05.

3.3. Association of AD-associated PC2 with A/T/N biomarkers

   The PC2 was investigated to assess its associations with central A/T/N
   biomarkers for AD. Lower PC2 scores were significantly associated with
   higher levels of CSF p-tau (β = −0.035 ± 0.016, p = 0.033) and reduced
   hippocampal volume (β = 90.105 ± 39.647, p = 0.025; [131]Figures
   3A,[132]B; [133]Supplementary Table S4).

Figure 3.

   [134]A set of four scatter plots displays relationships between PC2
   component scores and various cognitive and biological measures. (A)
   Shows a negative trend between PC2 scores and CSF P-tau levels. (B)
   Displays a positive relationship with hippocampal volume. (C) Indicates
   a positive trend with composite scores for executive function. (D)
   Shows a positive trend with composite scores for memory. Data points
   are scattered with trend lines and shaded confidence intervals in each
   plot.
   [135]Open in a new tab

   Association analysis results of PC2 with CSF p-tau, hippocampal volume,
   and cognitive function The scatter plot represents a correlation of the
   PC2 component score with CSF p-tau (A), hippocampal volume (B),
   baseline composite score for executive function (C), and baseline
   composite score for memory (D). The gray zone around the linear
   regression line represents the 95% confidence interval. Hippocampal
   volume was measured in cubic millimeters (mm^3). CSF biomarker
   concentrations were log-transformed and reported in picograms per
   milliliter (pg/mL). Cognitive scores were standardized as Z-scores. AD,
   Alzheimer’s disease; CN, cognitively normal; CSF, cerebrospinal fluid;
   p-tau, phosphorylated tau; PC, principal component.

3.4. Association of AD-associated PC2 with cognition at baseline and
longitudinal change of cognition

   The PC2 was investigated to assess its associations with composite
   scores for memory and executive function at baseline and longitudinal
   changes of composite scores for memory and executive function. Higher
   PC2 scores were associated with better baseline composite scores for
   memory (β = 0.222 ± 0.074, p = 0.003) and executive function
   (β = 0.135 ± 0.064, p = 0.038; [136]Figures 3C,[137]D;
   [138]Supplementary Table S5), and slower longitudinal decline of
   composite scores for memory (β = 0.234 ± 0.086, p = 0.008; [139]Figure
   4; [140]Supplementary Table S5).

Figure 4.

   [141]Graphs A and B show scatter plots with trend lines for composite
   scores over five years. Graph A depicts executive function scores, and
   Graph B shows memory scores. Each graph includes two groups: 1st
   quartile (blue) and 4th quartile (red) of PC2 component scores. Trend
   lines indicate a slight decline over time, particularly in the 1st
   quartile.
   [142]Open in a new tab

   Association analysis results of PC2 with longitudinal changes of
   cognition (A, B) The x-axis represents years after the initial clinical
   evaluation, while the y-axis indicates composite scores for executive
   function (A) and memory (B). The two lines represent the different
   slopes of changes in composite scores for subjects with 1st quartile
   (green color) versus 4th quartile (red color) stratified by PC2
   component score. Cognitive scores were standardized as Z-scores. PC,
   principal component.

3.5. Association of key miRNAs in the AD-associated PC2 with AD diagnosis and
A/T/N biomarkers

   In the AD-associated PC2, two miRNAs, miR-423-5p and miR-92a-3p, with a
   factor loading higher than 0.8 were identified as key miRNAs
   ([143]Supplementary Table S2). Further analyses were performed to
   investigate the association of these two key miRNAs with AD diagnosis
   and A/T/N biomarkers for AD. Expression levels of miR-423-5p and
   miR-92a-3p were significantly lower in AD patients compared to CN
   individuals (miR-423-5p, β = −44.438 ± 15.286, p = 0.004; miR-92a-3p,
   β = −41.606 ± 13.018, p = 0.001; [144]Figures 5A,[145]B). Lower
   expression levels of miR-423-5p and miR-92a-3p were significantly
   associated with reduced hippocampal volume (miR-423-5p, β = 8.162 ×
   10^3 ± 3.279 × 10^3, p = 0.014; miR-92a-3p, β = 6.366 × 10^3 ± 2.632 ×
   10^3, p = 0.017), and lower expression levels of miR-423-5p were
   associated with higher global cortical amyloid deposition
   (β = −1.118 ± 0.561, p = 0.048; [146]Figures 5C, [147]6A).

Figure 5.

   [148]Graphs comparing miR expression levels and their correlations. (A)
   miR-423-5p expression in CN vs AD, indicating significant difference.
   (B) miR-92a-3p expression in CN vs AD with similar results. (C)
   Correlation matrix showing relationships between miR levels and
   clinical measures such as amyloid PET and memory, with significance
   levels indicated by stars.
   [149]Open in a new tab

   Association analysis results of miR-423-5p and miR-92a-3p with amyloid,
   tau, and neurodegeneration (A/T/N) biomarkers as well as cognition (A,
   B) Violin and box plots represent the miR-423-5p (A) and miR-92a-3p
   expression levels (B) for AD diagnosis (C). Heatmap of association of
   miR-423-5p and miR-92a-3p expression levels with the “A/T/N” biomarkers
   for AD and cognition. p values were estimated from linear regression
   analyses. Circle color indicates the direction of the regression
   coefficient (β), while circle size and color intensity reflect the
   significance level of the association, quantified as –log (p value).
   The color bar represents –log (p value) × Sign(β). Red colors represent
   positive associations, and blue colors denote negative associations.
   The darker the color and larger the circle, the stronger the
   association. Hippocampal volume was measured in cubic millimeters
   (mm^3). CSF biomarker concentrations were log-transformed and reported
   in picograms per milliliter (pg/mL). Cognitive scores were standardized
   as Z-scores. Aβ, amyloid-β; AD, Alzheimer’s disease; CN, cognitively
   normal; CSF, cerebrospinal fluid; miR, microRNA; p-tau, phosphorylated
   tau; PET, positron emission tomography. Note: significance stars
   indicating the p values of the correlations adjustment for multiple
   comparisons * p-value < 0.05. ** p-value < 0.01.

Figure 6.

   [150](A) A table displaying results from pooled analysis for biomarkers
   related to miR-423-5p and miR-92a-3p. Findings include β, SE, and P
   values for both males and females, with sex differences and
   interactions noted. Various biomarkers, including Amyloid PET and CSF
   levels, are classified by effect type.(B) A scatter plot showing the
   effect Z-scores for males and females for different biomarkers, such as
   Amyloid PET and CSF Aβ42. Differentiates larger group effects by shape
   and color.(C) Similar scatter plot to (B) for other markers,
   highlighting sex-specific effects and significance levels.
   [151]Open in a new tab

   Association analysis results and sex-based effect heterogeneity of
   miR-423-5p and miR-92a-3p with amyloid, tau, and neurodegeneration
   (A/T/N) biomarkers (A) Linear regression results examining the
   associations between miRNAs and A/T/N biomarkers are presented for the
   full sample and stratified by sex, along with estimates of effect
   heterogeneity and p value for sex-by-miRNA interaction terms. Scatter
   plots display Z-scores of effect estimates for the associations of the
   miR-423-5p (B) and miR-92a-3p (C) with A/T/N biomarkers in males
   (x-axis) and females (y-axis). Homogeneous effects, defined as
   associations with comparable effect size in both sexes, appear along
   the diagonal. Heterogeneous effects appear along the anti-diagonal.
   Sex-specific effects are located near the x-axis (male-specific) or
   y-axis (female-specific effects). Homogeneous and overall significant
   results are indicated by diamonds, while effects significant in only
   one sex are represented as triangles. Sex-specific effects are further
   visualized using a color scale (purple for males; green for females).
   Hippocampal volume was measured in cubic millimeters (mm^3). CSF
   biomarker concentrations were log-transformed and reported in picograms
   per milliliter (pg/mL). Aβ, amyloid-β; CSF, cerebrospinal fluid; miR,
   microRNA; p-tau, phosphorylated tau; PET, positron emission tomography;
   SE, Standard Error; SUVR, standardized uptake value ratio.

   We conducted sex-stratified association analyses and subsequently
   assessed heterogeneity in effect estimates between males and females.
   [152]Figure 6 details the results of miRNA associations with A/T/N
   biomarkers, including sex-by-miRNA interaction analyses. We identified
   a significant homogeneous association between miR-423-5p and global
   cortical amyloid deposition. In contrast, both miR-423-5p and
   miR-92a-3p exhibited male-specific associations with reduced
   hippocampal volume.

3.6. Association of key miRNAs in the AD-associated PC2 with cognition at
baseline and longitudinal change of cognition

   The two key miRNAs, miR-423-5p and miR-92a-3p, were investigated to
   assess their associations with composite scores for memory and
   executive function at baseline and longitudinal changes of composite
   scores for memory and executive function. Higher expression levels of
   miR-423-5p and miR-92a-3p were significantly associated with better
   baseline composite scores for memory (miR-423-5p, β = 20.164 ± 6.098,
   p = 0.001; miR-92a-3p, β = 13.652 ± 4.941, p = 0.006) and executive
   function (miR-423-5p, β = 13.348 ± 5.317, p = 0.013; miR-92a-3p,
   β = 9.624 ± 4.280, p = 0.026; [153]Figure 5C; [154]Supplementary Table
   S6). Higher expression levels of miR-423-5p and miR-92a-3p were also
   significantly associated with slower longitudinal decline of composite
   scores for memory (miR-423-5p, β = 21.573 ± 7.133, p = 0.003;
   miR-92a-3p, β = 15.684 ± 5.775, p = 0.007), and lower expression levels
   of miR-423-5p were associated with slower longitudinal decline of
   composite scores for executive function (β = 12.483 ± 5.927, p = 0.037;
   [155]Figures 5C, [156]7; [157]Supplementary Table S6).

Figure 7.

   [158]Graphs show composite scores over time for executive function and
   memory based on miR-423-5p and miR-92a-3p expression levels. Panels A
   and C depict executive function. Panels B and D depict memory. Red
   lines represent the 4th quartile, and blue lines represent the 1st
   quartile of expression levels. Scores generally decrease over five
   years.
   [159]Open in a new tab

   Association analysis results of miR-423-5p and miR-92a-3p with
   longitudinal changes of cognition The x-axis represents years after the
   initial clinical evaluation, while the y-axis indicates composite
   scores for executive function (A,C) and memory (B,D). The two lines
   represent the different slopes of changes in composite scores for
   subjects with 1st quartile (green color) versus 4th quartile (red
   color) stratified by miR-423-5p (A,B) and miR-92a-3p expression levels
   (C,D). Cognitive scores were standardized as Z-scores. miR, microRNA.

3.7. Correlation between miRNA expression in plasma and CSF

   The two key miRNAs, miR-423-5p and miR-92a-3p, were further
   investigated to evaluate their correlation between plasma and CSF
   samples. The correlation analysis between plasma and CSF expression
   levels of miR-423-5p showed a significant correlation (r = 0.236,
   p = 0.004), while the significant correlation was not observed with
   miR-92a-3p ([160]Figure 8).

Figure 8.

   Scatter plots showing the correlation between CSF and plasma expression
   levels. Panel A depicts a positive correlation between CSF miR-423-5p
   and plasma miR-423-5p with r = 0.236 and P = 0.004. Panel B shows a
   weaker correlation between CSF miR-92a-3p and plasma miR-92a-3p with r
   = 0.084 and P = 0.314. Blue trend lines indicate the correlation with
   shaded areas representing confidence intervals.
   [161]Open in a new tab

   Correlation plot between plasma and CSF expression levels of miR-423-5p
   and miR-92a-3p (A,B) The scatter plot represents a correlation between
   plasma and CSF expression levels of miR-423-5p (A) and miR-92a-3p (B).
   The gray zone around the linear regression line represents the 95%
   confidence interval. CSF, cerebrospinal fluid; miR, microRNA.

3.8. Predicted target genes for two key miRNAs and their pathway enrichment
analysis

   Target genes for the two miRNAs, miR-423-5p and miR-92a-3p, were
   obtained from the miRDB and TargetScan databases, resulting in 700 and
   985 genes from the miRDB database, and 232 and 1,041 genes from the
   TargetScan database, respectively. 106 and 701 overlapping target genes
   for miR-423-5p and miR-92a-3p were then identified between the two
   databases ([162]Supplementary Additional file). The GO analysis of the
   target genes for miR-423-5p and miR-92a-3p revealed significant
   enrichment in BP pathway related to positive regulation of
   transcription from RNA polymerase II promoter, negative regulation of
   transcription from RNA polymerase II promoter, brain development,
   protein phosphorylation, intracellular signal transduction, negative
   regulation of transcription, peptidyl-serine phosphorylation, nervous
   system development, neuron migration, memory, palate development,
   negative regulation of cytoplasmic translation, ventricular septum
   morphogenesis, cell migration, peptidyl-threonine phosphorylation,
   positive regulation of transcription, regulation of transcription from
   RNA polymerase II promoter, phosphatidylinositol-3-phosphate
   biosynthetic process, and cellular response to amino acid stimulus
   ([163]Table 2).

Table 2.

   Gene ontology analysis of target genes of miR-423-5p and miR-92a-3p.
   Biological process Number of genes from study data/Number of genes in
   the pathway (Population Hits) Fold enrichment p value^a
   Positive regulation of transcription from RNA polymerase II promoter
   98/1232 2.088 <0.001
   Negative regulation of transcription from RNA polymerase II promoter
   78/1016 2.015 <0.001
   Brain development 31/264 3.083 <0.001
   Protein phosphorylation 40/444 2.365 <0.001
   Intracellular signal transduction 41/462 2.330 <0.001
   Negative regulation of transcription, DNA-templated 48/591 2.132 <0.001
   Peptidyl-serine phosphorylation 23/184 3.282 <0.001
   Nervous system development 37/425 2.285 0.002
   Neuron migration 17/124 3.599 0.006
   Memory 14/89 4.130 0.010
   Palate development 12/66 4.773 0.012
   Negative regulation of cytoplasmic translation 6/12 13.126 0.017
   Ventricular septum morphogenesis 9/37 6.386 0.020
   Cell migration 26/278 2.455 0.022
   Peptidyl-threonine phosphorylation 12/72 4.375 0.027
   Positive regulation of transcription, DNA-templated 50/728 1.803 0.027
   Regulation of transcription from RNA polymerase II promoter 97/1736
   1.467 0.039
   Phosphatidylinositol-3-phosphate biosynthetic process 7/22 8.353 0.044
   Cellular response to amino acid stimulus 11/64 4.512 0.047
   [164]Open in a new tab

   ^aAdjusted P value for Bonferroni correction. AD, Alzheimer’s dementia;
   miRNAs, microRNAs; mRNA, messenger RNAs.

3.9. Mediation analysis of two key miRNAs on AD or composite scores for
memory

   Mediation analysis revealed that miR-423-5p expression levels had both
   direct and indirect effects on AD through global cortical amyloid
   deposition (direct effect: β = −2.61 × 10^−10; p = 0.04; indirect
   effect: β = −2.61 × 10^−10; p = 0.02), while miR-423-5p expression
   levels had only indirect effects on AD through hippocampal volume
   (direct effect: β = −9.86 × 10^−16; p = 0.16; indirect effect:
   β = −9.86 × 10^−16; p = 0.02; [165]Figure 9; [166]Supplementary Table
   S7). Additionally, miR-423-5p expression levels were found to affect
   composite scores for memory through both direct and indirect effects
   mediated by global cortical amyloid deposition (direct effect:
   β = 10.29; p = 0.02; indirect effect: β = 4.44; p = 0.04) and
   hippocampal volume (direct effect: β = 8.52; p = 0.02; indirect effect:
   β = 7.80; p = 0.02) ([167]Figure 9; [168]Supplementary Table S7).
   However, miR-92a-3p expression levels had no indirect effects on AD or
   composite scores for memory through hippocampal volume
   ([169]Supplementary Table S7).

Figure 9.

   [170]Diagram showing two pathways involving miR-423-5p. Panel A
   illustrates effects on amyloid PET and Alzheimer's disease (AD), with
   direct, indirect, and total effects noted. Panel B shows effects on
   hippocampal volume and AD, with similar metrics. Both panels include
   pathways to memory, highlighting statistical values and significance
   for each effect.
   [171]Open in a new tab

   Mediation analysis results of miR-423-5p on AD diagnosis and composite
   scores for memory through global cortical amyloid deposition and
   hippocampal volume (A,B) Mediation analysis shows total, direct, and
   indirect effects of miR-423-5p expression levels on AD and composite
   scores for memory through global cortical amyloid deposition (A), and
   hippocampal volume (B). Hippocampal volume was measured in cubic
   millimeters (mm^3). Cognitive scores were standardized as Z-scores. AD,
   Alzheimer’s disease; miR, microRNA; PET, positron emission tomography;
   SUVR, standard uptake value ratio.

3.10. Machine learning analysis for AD classification

   A machine learning approach, extreme gradient boosting, for the
   classification of AD patients from CN individuals was used to evaluate
   two different classification models. Results of the 5-fold
   cross-validation are presented in [172]Figure 10. Model 1, including
   age, sex, and APOE ε4 carrier status, achieved a mean AUC value of
   0.708. The mean AUC value of Model 2, obtained by adding two
   AD-associated key miRNAs, miR-423-5p and miR-92a-3p, to Model 1,
   increased to 0.757.

Figure 10.

   [173]Two ROC curves compare the performance of models. Chart (A) shows
   a mean ROC with an area under the curve (AUC) of 0.708 plus/minus
   0.110. Chart (B) shows a mean ROC with an AUC of 0.757 plus/minus
   0.129. Both charts plot true positive rate against false positive rate,
   with a diagonal reference line.
   [174]Open in a new tab

   The ROC curves and mean AUC of machine learning approach Sensitivity is
   on the y-axis and 1-specificity is on the x-axis. 5-fold cross
   validation was used to investigate and compare the classification
   performance of two different classification models for differentiating
   AD from CN. The gray zone around the mean ROC curve represents ± 1
   standard deviation. Two different classification models were evaluated
   using the following training features: (A) age, sex, and APOE ε4
   carrier status; (B) age, sex, APOE ε4 carrier status, miR-423-5p, and
   miR-92a-3p. AD, Alzheimer’s disease; APOE, apolipoprotein E; AUC, area
   under the curve; CN, cognitively normal; miRNAs, microRNAs, ROC,
   receiver operating characteristic.

4. Discussion

   This study highlights the association of circulating miRNAs with
   central biomarkers for AD and their potential utility as blood-based
   molecular biomarkers for AD. Using PCA, we identified one PC that
   showed significant associations with AD, CSF p-tau levels, hippocampal
   volume, and both baseline levels and longitudinal changes in cognition.
   Among the miRNAs that contributed significantly to the PC, miR-423-5p
   and miR-92a-3p were identified as key miRNAs. Both miRNAs were found to
   be significantly downregulated in AD patients compared to CN
   individuals, and their lower expression levels were associated with
   central biomarkers of AD, including reduced hippocampal volume and
   higher global cortical amyloid deposition. Furthermore, lower levels of
   these miRNAs were correlated with worse cognitive performance and a
   faster rate of cognitive decline, indicating their potential role not
   only as diagnostic biomarkers but also as indicators of disease
   progression.

   Accumulating evidence indicates that miRNAs play important roles in AD
   pathogenesis through the regulation of genes involved in amyloid
   processing, tau phosphorylation, and neurodegeneration ([175]Wang et
   al., 2019; [176]Kou et al., 2020; [177]Guévremont et al., 2022). Our
   findings align with this evidence by demonstrating associations of
   lower plasma levels of miR-423-5p and miR-92a-3p with higher global
   cortical amyloid deposition, reduced hippocampal volume, and worse
   cognitive performance. Moreover, our study showed a significant
   correlation between plasma and CSF expression levels of miR-423-5p, but
   not for miR-92a-3p. This suggests that miR-423-5p may have potential as
   a blood-based molecular biomarker reflective of central nervous system
   changes.

   Lower expression levels of miR-423-5p were associated with AD in human
   cerebral white matter ([178]Wang et al., 2011; [179]Bhattacharyya and
   Bandyopadhyay, 2013), aligning with findings of our study. Importantly,
   our study uniquely demonstrates an association between plasma
   miR-423-5p levels and AD and cognitive function, which has not been
   previously reported. Furthermore, overexpression of miR-423-5p has been
   found to enhance cell proliferation, migration, and angiogenesis, and
   to reduce apoptosis in vivo ([180]Li S. et al., 2017). Lower expression
   levels of plasma miR-92a-3p have been associated with AD
   ([181]Peña-Bautista et al., 2022; [182]Piscopo et al., 2023),
   supporting the findings of our study. Higher expression levels of
   plasma miR-92a-3p were also found to be significantly associated with
   better cognitive function ([183]Yaqub et al., 2023; [184]Melas et al.,
   2024), reinforcing the results of our study. Notably, miR-92a-3p is
   involved in maintaining blood–brain barrier integrity by regulating
   tight junctions and endothelial cell function, often compromised in AD
   ([185]Lu et al., 2021). Additionally, miR-92a plays a critical role in
   regulating Aβ clearance, accumulation of tau protein, synaptic and
   gamma-aminobutyric acid-ergic dysfunction in AD ([186]Li X. et al.,
   2017; [187]Siedlecki-Wullich et al., 2019; [188]Peña-Bautista et al.,
   2022). Overexpression of miR-92a-3p, one of the miRNAs identified to
   interact with microtubule-associated protein tau mRNAs, significantly
   reduced tau protein levels in neuroblastoma cell lines ([189]Piscopo et
   al., 2023). We observed significant effect heterogeneity between males
   and females for the key miRNAs, miR-423-5p and miR-92a-3p, indicating
   that sex may influence miRNA-related hippocampal atrophy. Our results
   demonstrate the importance of stratified analyses in uncovering
   AD-related neurodegeneration that appear to be male-specific.

   Pathway enrichment analyses conducted in this study provided insights
   into the significant biological processes that may be influenced by the
   target genes of the two AD-associated key miRNAs. Notably,
   dysregulation in cell migration, proliferation, and differentiation
   plays a significant role in both AD and Parkinson’s disease,
   underscoring the importance of these pathways in neurodegeneration
   ([190]Awuson-David et al., 2023). Furthermore, pathways such as protein
   phosphorylation, including peptidyl-serine and peptidyl-threonine
   phosphorylation, are critical, as hyperphosphorylation of tau at these
   residues contributes to its aggregation into NFTs, a hallmark of AD
   pathology ([191]Wegmann et al., 2021). Lastly, a deficiency in
   phosphatidylinositol-3-phosphate was identified as a contributing
   factor to AD pathogenesis by disrupting amyloid precursor protein
   trafficking and processing ([192]Morel et al., 2013).

   The mediation analysis of our study suggests potential roles of
   miR-423-5p and miR-92a-3p in AD and cognitive function. Specifically,
   miR-423-5p had both direct and indirect effects on AD through global
   cortical amyloid deposition and hippocampal volume, while miR-92a-3p
   did not demonstrate any significant indirect effects. These findings
   underscore the potential role of miR-423-5p in modulating brain Aβ
   deposition and neurodegeneration, suggesting its involvement in the
   molecular mechanisms underlying AD. The lack of significant mediation
   effects for miR-92a-3p suggests that its contribution to AD might
   involve alternative pathways or mechanisms not addressed in our
   analysis, which requires further exploration to fully understand its
   role.

   The machine learning analysis for differentiating AD patients from CN
   individuals demonstrated that incorporating miR-423-5p and miR-92a-3p
   improved the classification performance of demographic information and
   APOE ε4 carrier status for AD classification from the mean AUC value of
   0.708 to that of 0.757. This enhancement underscores the potential of
   miR-423-5p and miR-92a-3p as blood-based molecular biomarkers for AD
   diagnosis. Integrating these miRNA expression data with demographic
   information could pave the way for personalized medicine by supporting
   the development of strategies that account for individual differences
   in miRNA expression. The clinical relevance of miRNAs extends beyond
   their role as diagnostic biomarkers, as they are involved in the
   regulation of molecular pathways implicated in AD pathogenesis. This
   mechanistic involvement has led to growing interest in miRNA-based
   therapeutic strategies, including the use of synthetic miRNA mimics and
   anti-miRNA oligonucleotides aimed at modulating multiple gene
   expressions and disrupting pathological pathways ([193]Miya Shaik et
   al., 2018; [194]Bhatnagar et al., 2023).

   While this study provides a comprehensive analysis of circulating miRNA
   expression data with central A/T/N biomarkers and cognitive function
   within a well-characterized cohort, several limitations need to be
   addressed in future studies. The modest sample size and lack of
   external datasets may restrict generalizability, underscoring the need
   for replication in future large-scale studies. This limitation also
   affects the clinical applicability of our miRNA-based diagnostic model,
   which demonstrated the modest classification performance. Moreover,
   information on comorbid conditions and medication use was not available
   in the ADNI dataset, limiting our ability to adjust for these potential
   confounders. Future longitudinal studies incorporating repeated miRNA
   measurements will be essential to elucidate the temporal dynamics
   underlying the observed associations and to establish causality through
   mediation. Finally, functional studies are crucial to unravel the
   specific roles of miR-423-5p and miR-92a-3p in AD pathogenesis and to
   validate their potential as therapeutic targets.

5. Conclusion

   In conclusion, this study shows associations of circulating miRNAs
   (miR-423-5p and miR-92a-3p) with central AD biomarkers and disease
   progression as well as the improvement in the accuracy of AD
   classification, shedding light on their potential as blood-based
   molecular biomarkers for AD. Further investigations are warranted to
   validate our findings and clarify the underlying mechanisms.

Acknowledgments