Introduction
Alzheimer’s disease (AD) manifests as a chronic, progressive neurodegenerative condition marked by cognitive decline.1 As the leading cause of dementia, AD currently affects approximately 2–8% of the global population, creating substantial economic and societal burdens.2 By 2050, the number of individuals diagnosed with AD is projected to exceed 100 million, driven by ongoing population aging.2,3 Epidemiological studies indicate that AD incidence doubles every five years after age 65, with nearly half of all individuals aged 85 or older affected, highlighting aging as a primary risk factor.4
Osteoporosis (OP) is a widespread metabolic bone condition defined by reduced bone mass and structural deterioration, predisposing affected individuals to fractures.5 Like AD, OP represents one of the major geriatric diseases largely driven by accelerated population aging, with its prevalence projected to double between 2010 and 2040.6 Besides aging, both AD and OP share common risk determinants, including female sex, genetic predisposition, and adverse lifestyle patterns,7 all of which contribute substantially to their pathogenesis and progression. Nevertheless, the biological mechanisms by which aging increases susceptibility to AD and OP remain poorly understood.
Emerging research has revealed that these two conditions are mechanistically interconnected, with aging serving as the fundamental driver of their shared pathology. Aging creates a permissive microenvironment marked by chronic low-grade inflammation, oxidative stress, and cellular senescence.8 Age-related alterations in peripheral and tissue-resident immune cells may sustain inflammatory responses, thereby contributing to both neuroinflammation and abnormal bone remodeling. With aging, multiple pathological cascades converge and become mutually reinforcing between AD and OP.
Accumulating evidence suggests bidirectional interactions between brain pathology and skeletal homeostasis.7,9 On the one hand, experimental evidence indicates that Aβ enhances RANKL-induced osteoclast activation by promoting IκB-α degradation, ERK phosphorylation, and intracellular calcium oscillations.10 Consistently, amyloid-β neuropathology has been associated with bone loss characterized by suppressed bone formation and enhanced bone resorption in transgenic mouse models.11 On the other hand, bone-derived factors such as osteocalcin (OCN), whose bioavailability declines with aging, cross the blood-brain barrier to support hippocampal neurogenesis and mitigate Aβ aggregation.7,12 Furthermore, key signaling pathways such as Wnt/β-catenin are progressively impaired during aging, contributing to tau hyperphosphorylation in AD and suppressed osteoblast differentiation in OP.13,14 In parallel, age-related elevation of pro-inflammatory cytokines, notably TNF-α and IL-6, amplifies both neuroinflammation and bone resorption through promoting glial activation and osteoclastogenesis.15 Thus, aging not only serves as a common risk factor but also actively amplifies these interconnected pathological processes, reinforcing the mechanistic overlap between AD and OP.
Although aging is known to amplify the interconnected pathology of AD and OP, the aging-related molecular drivers governing this crosstalk remain poorly defined at a systems level. However, well-validated molecular biomarkers reflecting shared aging-related alterations across AD and OP remain limited, and clinically applicable predictive tools for assessing their shared molecular risk have not yet been established. To address this gap, we employed an integrative bioinformatics approach combining machine learning, immune infiltration profiling, and experimental validation to systematically pinpoint aging-related genes that are dysregulated in both diseases and may serve as shared molecular signatures.
Machine learning was selected for multivariable feature selection in high-dimensional transcriptomic data. Co-differentially expressed genes (Co-DEGs) shared by AD and OP were initially identified through differential expression analysis. We then applied 130 machine learning models to prioritize aging-related DEGs, with feature importance explained via Shapley Additive Explanations (SHAP).
Candidate biomarkers were further evaluated through functional annotation, expression analysis, and immune infiltration analysis. Finally, six core candidate genes were validated via RT-qPCR. The overall workflow is summarized in Figure 1.
Figure 1Flowchart for the identification of Aging-related DEGs in AD and OP.Flowchart sequence: 1. Data Collection, 2. Shared Aging Related DEGs, 3.
Functional Enrichment and Candidate Screening, 4. Machine Learning Screening, 5. Diagnostic Nomogram Model, 6.
Immune Infiltration Analysis, 7. D galactose Induced Cellular Senescence In Vitro Validation. Step 1 text: AD datasets, GSE132903, GSE5281’ OP datasets, GSE35958, GSE7158’ Aging related genes: 2285.
Step 2 text: AD DEGs 5889’ Co DEGs 725’ OP DEGs 2348’ WGCNA analysis’ 228 shared aging related DEGs’ Co DEGs. Step 3 text: GO and KEGG enrichment’ PPI network. Step 4 text: Lasso plus GBM ensemble model’ 6 core aging hub genes: EGR1, NFKBIA, UQCRC2, RB1, CREBBP, NDUFV1.
Step 5 text: Co occurrence risk prediction. Step 6 text: Gene immune cell correlation’ CREBBP, UQCRC2, EGR1, NDUFV1, RB1NFKBIA. Step 7 text: D gal’ HT22 cell’ MC3T3 E1 cell’ CCK 8’ SA beta gal staining’ RT qPCR.Flowchart sequence: Data collection, shared DEGs, enrichment, machine learning, nomogram, immune, validation.
Materials and Methods
Data Source
Gene expression profiles for AD and OP were obtained from the NCBI-GEO (https://www.ncbi.nlm.nih.gov/gds) website. For AD, we included GSE13290316 (platform GPL10558; 97 cases, 98 controls) and GSE528117 (GPL570; 87 cases, 74 controls). For OP, we analyzed GSE3595818 (GPL570; 5 cases, 4 controls) and GSE715819 (GPL570; 12 cases, 14 controls).
In addition, GSE3300020 (GPL4372), comprising prefrontal cortex samples from 310 patients with AD and 157 non-demented controls, was used for immune infiltration analysis. Clinical characteristics of each cohort are summarized in Supplementary Table S1. A set of 2285 aging‑related genes (ARGs) was compiled from the Human Ageing Genomic Resources21 (HAGR) and GeneCards databases.22
Differential Expression and Enrichment Analyses
The microarray datasets GSE132903, GSE5281, GSE35958, and GSE7158 were processed and analyzed separately using the limma package in R under a consistent analytical workflow. Dataset-specific differential expression analysis was performed to avoid directly pooling samples derived from different cohorts, tissue sources, and microarray platforms. Differentially expressed genes (DEGs) were identified using an adjusted P-value < 0.05 and an absolute log2 fold change > 0.585.
Genes showing reproducible differential expression across the independent datasets were subsequently identified through intersection analysis.
Weighted Gene Co-Expression Network Analysis (WGCNA)
WGCNA was performed to construct a scale-free gene co-expression network. After quality control and removal of outliers, an appropriate soft-thresholding power was chosen to achieve a scale-free topology (R2 > 0.85). Gene modules were identified based on the topological overlap matrix (TOM) using dynamic tree cutting (minimum module size of 30, merge cut height of 0.25).
Pearson correlation was used to evaluate modules and traits relationships; modules with |R| ≥ 0.5 and P < 0.05 were deemed significant. Hub genes were identified by strong intramodular connectivity (kME > 0.8).
Identification of Co-DEGs
Co-DEGs shared between AD and OP were derived from the overlap of DEGs across the GSE132903, GSE5281, GSE35958, and GSE7158 datasets. Aging-related DEGs (ARDEGs) were subsequently obtained by overlapping Co-DEGs with curated ARGs, and results were visualized using Venn diagrams.
Functional Enrichment Analysis and Protein-Protein Interaction Network (PPI Network)
To investigate the biological functions and signaling pathways of aging-related DEGs in AD and OP, Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses were carried out with clusterProfiler (v4.16.0), with significance set at P < 0.05. GO analysis encompassed biological processes (BP), cellular components (CC), and molecular functions (MF). The DEGs were then submitted to the STRING database (http://string-db.org) to build a PPI network with a minimum confidence score of 0.7, and the network was visualized in Cytoscape (v3.7.2).
Machine Learning-Based Core Gene Screening
To systematically uncover aging-related diagnostic biomarkers in AD, a multilayered machine learning framework was developed using 11 classical algorithms. Based on training datasets, 130 predictive models (Lasso, SVM, RF, glmBoost, Stepglm, Ridge, Enet, GBM, LDA, XGBoost, and Naive Bayes) were constructed, with hyperparameters optimized via stratified 10-fold cross-validation. Model performance was evaluated using the area under the ROC curve (AUC).
The best-performing models were subsequently combined through a stacking ensemble approach. To reduce overfitting, we employed stratified cross-validation, independent validation, multiple-algorithm comparison, and cross-model feature stability assessment. Models were ranked according to their overall AUC performance across the analyzed datasets.
Candidate genes were subsequently ranked according to their selection frequency across the retained models, and genes showing stable importance across different algorithms were defined as core candidates. Their expression patterns were visualized using the pheatmap package in R.
Diagnostic Value of Hub DEGs
Following the identification of Aging-related biomarkers for AD, a diagnostic nomogram was constructed in the training cohort using the rms package (v6.3–0). Model performance was assessed via ROC analysis using the pROC package (v1.18.0), with the AUC used to measure overall predictive accuracy. Clinical applicability was further evaluated using decision curve analysis (DCA).
Immune Cell Infiltration and Immune Correlation Analysis
CIBERSORT algorithm was applied to the GSE33000 dataset (normal controls and AD samples) to evaluate immune cell infiltration. Based on the LM22 gene signature, relative abundances of 22 immune cell types were inferred from bulk expression profiles (default matrix, 100 permutations), and only samples with deconvolution P < 0.05 derived from Monte Carlo sampling were retained to ensure reliability. In parallel, immune activity was quantified using ssGSEA via the GSVA package (v1.44.5) to generate enrichment scores for 28 immune cell subsets across all samples.
Spearman correlation was applied to assess associations between candidate biomarkers and immune cell types, with results visualized as heatmaps. Additionally, the linkET package was employed to further investigate associations between hub genes and immune cell infiltration.
Cells and Culture Conditions
The HT22 mouse hippocampal neuronal cell line and MC3T3-E1 mouse pre-osteoblast cell line were purchased from Wuhan Procell Life Science and Technology Co., Ltd. (Wuhan, China). HT22 cells were cultured in DMEM, and MC3T3-E1 cells in α-MEM (Procell), both supplemented with 10% fetal bovine serum (FBS; Gibco) and 1% penicillin-streptomycin (Solarbio), under 5% CO2 at 37°C in a humidified atmosphere.
Cell Proliferation Assay
Cell viability was assessed using the Cell Counting Kit-8 (CCK-8, Dojindo, Japan). Cells were seeded into 96-well plates at 8 × 103 cells/well in 100 μL complete medium and allowed to attach for 24h. HT22 and MC3T3-E1 cells were then treated separately with varying concentrations of D-galactose (D-gal).
After treatment, the medium was replaced with 100 μL fresh medium containing 10 μL CCK-8 reagent, followed by incubation at 37°C for 2–4h. Absorbance at 450 nm (OD450) was read using a microplate reader. Cell viability was normalized to the untreated control (NC), and all experiments were performed in triplicate.
Senescence‑Associated β‑Galactosidase (SA‑β‑Gal) Staining
Cellular senescence in HT22 and MC3T3-E1 cells was assessed using a SA-β-gal staining assay, according to the manufacturer’s instructions. After the indicated treatments, cells were washed with prewarmed phosphate-buffered saline (PBS, 37 °C), fixed in 4% paraformaldehyde for 10–15 min at room temperature, and washed three times with PBS. Fixed cells were then incubated overnight at 37 °C with freshly prepared SA-β-gal staining solution (930 μL staining buffer, 10 μL solution A, 10 μL solution B, and 50 μL X-gal solution [20 mg/mL in DMF]).
The assay was performed in three independent biological experiments, with three technical replicate wells per condition in each experiment.
Senescent cells, identified by blue cytoplasmic staining, were visualized using a Nikon XSZ-D2 light microscope (Nikon Corporation, Tokyo, Japan). The proportion of SA-β-gal–positive cells relative to the total cell number was calculated to quantify cellular senescence. All experiments were performed in triplicate.
Quantitative Real-Time Polymerase Chain Reaction
Relative transcriptional levels of six hub genes (EGR1, NFKBIA, UQCRC2, RB1, CREBBP, and NDUFV1) were quantified by RT-qPCR. Total RNA was extracted using TRIzol reagent according to the manufacturer’s protocol from D-galactose-induced senescent HT22 and MC3T3-E1 cells. cDNA was synthesized by reverse transcription at 25 °C for 10 min, 42 °C for 30 min, and 85 °C for 5 min. qPCR was performed with an initial denaturation at 95 °C for 10 min, followed by 40 cycles of 95 °C for 15s and 60 °C for 60s. GAPDH was used as the internal reference gene, and relative gene expression was calculated using the 2^−ΔΔCt method.
Three independent biological experiments were performed, with three technical replicates for each sample. Primer sequences for the six genes are listed in Supplementary Table S2.
Statistical Analysis
All analyses were performed in R software (v 4.4.1). Group differences were assessed using the Wilcoxon test, with significance defined as P < 0.05.
Results
Gene Expression Alterations Associated with AD and OP
Analysis of the AD-related dataset GSE5281 and GSE132903 with the limma R package, identified 5889 AD-related differentially expressed genes (AD-DEGs), which are summarized in Figure 2A–E. The top 10 significantly upregulated and downregulated genes from GSE5281 are displayed in the heatmap (Figure 2B); while a volcano plot shows 124 up-regulated genes and 232 down-regulated genes in GSE132903 based on the log-fold change (logFC) values (Figure 2E). Similarly, differential expression analysis of the OP-related datasets GSE35958 and GSE7158 identified a total of 2348 OP-associated DEGs (OP-DEGs), with their distribution shown in Figures 2F–J.
The heatmap for GSE35958 highlights the top 10 upregulated and downregulated genes (Figure 2G), whereas the volcano plot for GSE7158 reveals 3768 upregulated and 1276 downregulated genes based on logFC values (Figure 2J). Intersection analysis further demonstrated that 725 DEGs were shared between the AD and OP datasets. Figure 2Identification of differentially expressed genes in AD and OP.
(A) Volcano plot illustrating DEGs in GSE5281. (B) Heatmap of GSE5281 DEGs. Red: upregulation; blue: down-regulation.
(C) Heatmap of GSE132903 DEGs. (D) PCA plot of GSE132903. (E) GSE132903 DEGs volcano plot.
(F) Volcano plots of GSE35958 DEGs. (G) Heatmap of GSE35958 DEGs. (H) Heatmap of GSE7158 DEGs.
(I) PCA plot of GSE7158. (J) GSE7158 DEGs volcano plot. Red: significantly upregulated; blue: significantly downregulated.A) Volcano scatter plot.
X axis label, logFC, ticks negative 2, negative 1, 0, 1, 2. Y axis label, minus log10 p value, ticks 0, 5, 10, 15, 20. Legend, Significant: Down, Not, Up.
Points form two wings around logFC 0. B) Clustered heatmap with dendrogram and gene names at right. Scale labeled Type with ticks negative 2, negative 1, 0, 1, 2.
Text, 232 down, 124 up. C) Same heatmap style as B with top annotation labeled group. D) Principal component analysis scatter plot with two groups and ellipses.
X axis, Dim1 (19.8 percent). Y axis, Dim2 (9.9 percent). Dashed cross at 0.
E) Volcano scatter plot. X axis label, limma, ticks negative 2 to 2. Y axis, minus log10 p value, ticks 0 to 20.
Dashed cutoffs near negative 1, 1 and 2. Legends, Groups: ND, AD’ change: DOWN, STABLE, UP. Top text, 232 down, 124 up.
F) Volcano scatter plot. X axis, logFC, ticks negative 1, 0, 1. Y axis, minus log10 p value, ticks 0 to 4.A multi plot figure of volcano plots, heatmaps and principal component analysis plots for gene expression.
WGCNA in AD and OP
The WGCNA was applied to the AD dataset GSE5281 and the OP dataset GSE7158 to investigate associations between gene expression patterns and disease-related traits. Following batch effect correction, hierarchical clustering was conducted to evaluate sample similarity based on transcriptomic profiles. The optimal soft-thresholding power was set at 3 for GSE5281 and 6 for GSE7158 (Figure 3A and B).
Subsequently, gene co-expression networks were constructed, and modules were defined using dynamic tree-cutting. Based on module eigengene similarity, 6 distinct modules were identified in the AD dataset and 10 modules in the OP dataset (Figure 3C and D). The relationships between the identified modules and disease traits were subsequently visualized using module–trait correlation heatmaps (Figure 3E and F).
Module-trait relationship analysis revealed that the MEbrown module was most strongly positively correlated with AD (r = 0.57, P = 5 × 10−15), whereas the MEred module showed the strongest negative correlation with OP (r = −0.50, P = 0.01) (Figures 3E and F). Module membership was positively correlated with gene significance in the MEbrown module (r = 0.76, P = 4.1 × 10−45) and the MEred module (r = 0.25, P = 0.009) (Figures 3G and H). The MEbrown module included 1576 genes and the MEred module included 711 genes, respectively.
By integrating these genes with DEGs, a total of 1660 Co-DEGs were obtained. Intersection of these candidates with 2286 ageing-related genes identified 228 ageing-related common differentially expressed genes (Aging-related Co-DEGs), which are potentially involved in the shared pathological mechanisms underlying AD and OP (Figure 3I). Figure 3WGCNA analysis of AD and OP.
(A and B) Selection of soft-threshold power for GSE5281 and GSE7158 samples. (C and D) Dendrograms showing initial and merged gene modules for GSE5281 and GSE7158. (E and F) Heatmap depicting correlations between the module signature gene with AD and OP, respectively.
(G and H) Association between module signature genes and AD or OP occurrence. (I) Venn plot of Co-DEGs of AD, OP and aging.A) Two line graphs. Left x-axis, Soft Threshold power’ y-axis, Scale Free Topology Model Fit signed R squared.
Right x-axis, Soft Threshold power’ y-axis, Mean Connectivity. B) Two line graphs with the same axis labels as A). C) Dendrogram with y-axis label Height and a strip labeled Dynamic Tree Cut.
D) Dendrogram with y-axis label Height and a strip labeled Dynamic Tree Cut. E) Heatmap with x-axis labels Con and Treat’ y-axis labels MEblue, MEgreen, MEbrown, MEred, MEturquoise, MEyellow’ scale labeled negative 1 to 1. F) Heatmap with x-axis labels Con and Treat’ y-axis labels MEbrown, MEblue, MEpink, MEyellow, MEred, MEgreen, MEmagenta, MEblack, MEturquoise, MEgrey’ scale labeled negative 1 to 1.
G) Scatter plot with x-axis label Module Membership in brown module and y-axis label Gene significance for Treat. H) Scatter plot with x-axis label Module Membership in red module and y-axis label Gene significance for Treat.A multi-plot figure of soft-threshold curves, dendrograms, heatmaps, scatter plots and a Venn diagram.
Functional Enrichment Analysis and Construction of the PPI Network
GO and KEGG enrichment analyses were conducted on the 228 aging-related Co-DEGs. These genes were predominantly involved in multiple aging-associated biological pathways, including the PI3K-Akt, longevity-regulation, cellular senescence, FoxO, p53, adipocytokine, TNF, apoptosis, neurotrophin signaling pathway, and chronic inflammatory responses (Figure 4A and B). Furthermore, the 15 genes most frequently represented among the top 20 significantly enriched pathways were identified and summarized (Figure 4C).
To explore protein-level interactions, a PPI network was generated based on the STRING database, from which 30 hub genes-such as TP53, EGFR, NDUFV1, and UQCRC2-were recognized as central nodes within the network (Figure 4D). Figure 4GO and KEGG enrichment analysis and PPI network construction. (A) GO enrichment analysis of overlapping genes across BP, CC, and MF categories.
The x-axis represents the gene ratio, color intensity reflects adjusted P-values (darker red indicates greater statistical significance), and dot size corresponds to the number of enriched genes. (B) KEGG analysis shows enriched pathways for Co-DEGs. The x-axis indicates gene counts, while the color scale denotes adjusted P-values, with red representing higher significance.
(C) Top 20 enriched KEGG pathways for Co-DEGs and distribution of top 15 frequent genes. (D) PPI network visualizes interactions among top 30 Co-DEGs.A) Bubble plot with three sections labeled, bp, cc and mf. X-axis label, GeneRatio, ticks 0.050, 0.075, 0.100, 0.125.
Y-axis lists terms. bp: intrinsic apoptotic signaling pathway’ gland development’ regulation of apoptotic signaling pathway’ regulation of kinase activity’ regulation of cellular catabolic process’ cellular response to transforming growth factor beta stimulus’ response to transforming growth factor beta’ protein localization to nucleus’ cell cycle G1 slash S phase transition’ gland morphogenesis. cc: focal adhesion’ cell substrate junction’ neuron to neuron synapse’ asymmetric synapse’ RNA polymerase two transcription regulator complex’ postsynaptic density’ oxidoreductase complex’ respiratory chain complex’ perikaryon’ organelle inner membrane lumen.Four-panel enrichment charts and a protein protein interaction network with two separated node groups.
Identification of Core Age Related Genes Shared by AD-OP Genes
To identify genes with the highest diagnostic relevance, a machine learning approach was applied to 30 candidate targets and yielded 130 predictive models. Among these, the Lasso + GBM ensemble model achieved the highest accuracy in both training and validation sets (Figure 5A), ultimately highlighting six pivotal genes: EGR1, NFKBIA, UQCRC2, RB1, CREBBP, and NDUFV1. The Lasso + GBM model demonstrated strong discriminative capacity, with AUC values exceeding 0.9 in both cohorts (Figure 5B).
The identified hub genes demonstrated high diagnostic value for AD with OP, as shown by ROC curve analysis in Figure 5C. The AUC values of the six hub genes were as follows: CREBBP = 0.808, NFKBIA = 0.853, UQCRC2 = 0.832, EGR1 = 0.775, RB1 = 0.646, and NDUFV1 = 0.775 (Figure 5B), indicating favorable predictive performance. Figure 5Identification of core aging genes in OP and AD.
(A) Comparative evaluation of model performance across cohorts, with a heatmap illustrating AUC values (higher values indicate better performance). (B) ROC curves of the six hub genes (EGR1, NFKBIA, UQCRC2, RB1, CREBBP, and NDUFV1), where AUC reflects diagnostic accuracy. (C) ROC curves comparing the predictive performance of the GBM and LASSO models.
(D) Feature importance ranking based on the mean absolute SHAP values. (E) SHAP summary plot showing the contributions of the six hub genes to model predictions. (F) SHAP dependence plots showing the relationships between gene expression and SHAP values; the color gradient indicates the expression level of the interacting gene.
(G) SHAP force plot illustrating the contribution of each gene to the model prediction for an individual sample. Negative SHAP values indicate lower prediction risk, positive values indicate higher risk.A) Heatmap of AUC values with columns, Cohort1, Cohort2, GSE5281 and GSE1297 and a scale labeled, AUC. B) ROC plot with x axis label, 1 - Specificity and y axis label, Sensitivity’ labels list NFKBIA AUC 0.853, CREBBP AUC 0.808, UQCRC2 AUC 0.832, EGR1 AUC 0.775, RB1 AUC 0.646, NDUFV1 AUC 0.775.
C) ROC plot with x axis label, 1 - Specificity and y axis label, Sensitivity’ labels show GBM 0.915 (0.883, 0.967) and LASSO 0.909 (0.853, 0.964). D) Horizontal bar chart with x axis label, mean(|SHAP value|) and y axis gene labels EGR1, NFKBIA, UQCRC2, RB1, CREBBP, NDUFV1. E) SHAP summary plot with x axis label, SHAP value, y axis gene labels and a vertical label, Feature value, with High and Low.
F) Six SHAP dependence plots titled EGR1, NFKBIA, UQCRC2, RB1, CREBBP, NDUFV1’ each has x axis as the titled gene and y axis, SHAP value, with a side scale labeled by another gene name.A composite infographic of AUC heatmap, ROC curves, SHAP importance, SHAP summary, dependence plots and force.
Model interpretability was evaluated using SHAP analysis, which revealed distinct contributions of individual genes to model output. EGR1 (SHAP = 0.104) and NFKBIA (SHAP = 0.095) were identified as the most influential predictors (Figure 5D). Nonlinear dependency patterns were observed, including an inverse association between EGR1 and NFKBIA expression, as well as peak predictive effects of NFKBIA and CREBBP at higher expression levels (>10).
Force plot analysis (Figure 5G) further illustrated that EGR1 (8.47, Δ = +0.0656) served as a primary positive regulator, whereas RB1 (6.73, Δ = −0.193) and NFKBIA (10.4, Δ = −0.0359) exerted negative effects, jointly driving the predicted probability (f(x) = 0.342) below the baseline expectation (E[f(x)] = 0.529) (Figure 5G).
Construction and Evaluation of the Diagnostic Model
Based on six hub genes selected from the Lasso + GBM model, we constructed a nomogram for estimating the probability of AD and OP co-occurrence (Figure 6A). The model exhibited excellent calibration, with a slight error margin between predicted and actual risks (Figure 6B). Its superior clinical applicability was validated by DCA, which showed high net benefits across a reasonable risk threshold range (Figure 6C).
In addition, box plot analyses illustrated the distribution patterns of hub gene expression across groups (Figure 6D). Figure 6Constructing a nomogram. (A) The visible nomogram for diagnosing AD with OP.
(B) The DCA curve, which evaluates the clinical applicability of the nomogram model. (C) Model calibration was examined by comparing predicted and observed outcomes using a calibration curve. (D) Box plots illustrating expression differences of hub genes among groups. ***P < 0.001.A) Nomogram with scales labeled, Points, NFKBIA, CREBBP, NDUFV1, UQCRC2, EGR1, RB1, Total Points and Risk of Disease.
Points scale 0 to 100. NFKBIA ticks 7.5, 8, 8.5, 9, 9.5, 10, 11, 12. CREBBP ticks 5, 6, 7, 8, 9, 10, 11.
NDUFV1 shows 11 and 6. UQCRC2 ticks 11, 10.5, 10, 9.5, 9, 8.5, 8, 7.5. EGR1 ticks 15, 13, 11, 9, 8, 7, 6, 5, 4.
RB1 scale 2 to 10. Total Points 0 to 280. Risk of Disease ticks 0.1, 0.3, 0.5, 0.7, 0.9, 0.99.
B) Line graph. Horizontal axis label, Threshold probability, range 0.0 to 1.0. Vertical axis label, Net Benefit, range 0.0 to 0.5.
Legend, genes, All, None. None is a horizontal line at 0. Genes curve decreases from about 0.5 at 0.0 to near 0 at 1.0.
All curve drops below 0 near 0.55. Additional horizontal axis label, Cost:Benefit Ratio, ticks 1:100, 1:4, 2:3, 3:2, 4:1, 100:1. C) Line graph.
Horizontal axis label, Predicted probability, range 0.0 to 1.0. Vertical axis label, Actual probability, range 0.0 to 1.0. Legend, Apparent, Bias-corrected, Ideal.A mixed figure with a nomogram, a net benefit line graph, a calibration line graph and box plots.
Pathways and Functions Associated with Hub Biomarkers
To further explore the role of hub genes in cellular senescence, gene set enrichment analysis (GSEA) was performed using the SAUL_SEN_MAYO dataset, with particular emphasis on senescence-associated secretory phenotype (SASP)-related genes. The enrichment results indicated that these biomarkers were primarily associated with Cellular Senescence, p53 signaling, NF-κB signaling, PI3K-Akt signaling, FoxO signaling, Oxidative phosphorylation, Mitophagy, and osteoclast differentiation (Figure 7 and Supplementary Table S1). Biomarker-enriched pathways are particularly implicated in the pathogenesis of AD and OP, underpinning key pathological mechanisms such as chronic inflammation, immune dysregulation, and mitochondrial dysfunction.
Delineating this molecular network provides mechanistic insight into disease progression and reveals promising therapeutic targets, facilitating targeted strategy development for disease modification. Figure 7GSEA for biomarkers (A) EGR1, (B) NFKBIA, (C) UQCRC2, (D) RB1, (E) CREBBP, and (F) NDUFV1.A) Line graph of gene set enrichment analysis. Horizontal axis, Rank in Ordered Dataset, unit not shown, ticks at 0, 5000, 10000.
Vertical axis, Running Enrichment Score, unit not shown, ticks at minus 0.50, minus 0.25, 0.00, 0.25. Seven curves with a shared shape: early rise to a peak near rank 1000 to 3000, then a long decline, then an upturn near rank 12000 to 13000. Approximate peaks: 0.30, 0.28, 0.26, 0.22, 0.18, 0.12 and 0.05.
Approximate troughs before the final upturn: minus 0.45, minus 0.40, minus 0.35, minus 0.30, minus 0.25, minus 0.20 and minus 0.10. Legend terms: Circadian entrainment’ Hippo signaling pathway comma multiple species’ NOD like receptor signaling pathway’ Notch signaling pathway’ Oxidative phosphorylation’ PI3K Akt signaling pathway’ TGF beta signaling pathway. Middle rug marks show dense hits across the full rank range.
Bottom plot shows Ranked List Metric, unit not shown, decreasing from about 1.0 at rank 0 to about minus 1.5 at rank 13000.Six line graphs showing gene set enrichment analysis curves for EGR1, NFKBIA, UQCRC2, RB1, CREBBP and NDUFV1.
Immune Cell Infiltration and Correlation with Shared Hub Genes
Given the prominent enrichment of immune-related pathways in aging, we further examined immune cell infiltration patterns in AD using CIBERSORT and ssGSEA approaches on the GSE33000. Visualization with box plots revealed significant differences in specific immune cell populations, including M2 macrophages and neutrophils, between AD patients and healthy controls (Figure 8A). Figure 8Correlation between hub gene expression and immune infiltration in AD.
(A) Proportions of infiltrating immune cells in AD samples. (B) Pearson correlation between core disease genes and immune cell populations. (C) Heatmap showing correlations among immune cell types.
(D) Correlation matrix linking immune cell proportions and six hub genes; red color indicates positive correlation and blue color indicates negative correlation. (E-J) Lollipop plots showing the correlation between hub genes and immune cells. (*p < 0.05, **p < 0.01, ***p < 0.001 and ****P < 0.0001; ns, not significant.
In panels E-J, red values indicate P < 0.05.).A) Box plot comparing Control and AD across immune cell types. X axis, Immune cell type. Y axis, Estimated proportion, unit not shown.
Group labels, Control and AD. Significance labels include ns and asterisk marks. B) Triangular matrix of immune cell populations with links to six genes, CREBBP, UQCRC2, EGR1, RB1, NDUFV1, NFKBIA.
Axes list immune cell types. Side scales labeled Mantel&apos’s r, Pvalue and Pearson&apos’s cor, units not shown. C) Heatmap of correlations among immune cell types.
Both axes list the immune cell types. Color scale labeled Pearson&apos’s cor, unit not shown. D) Heatmap of Pearson&apos’s cor between six genes on the y axis and immune cell types on the x axis.
Color scale labeled Pearson&apos’s cor, unit not shown. E) Lollipop plot for CREBBP. X axis, Correlation Coefficient, unit not shown.
Y axis, immune cell type. F) Same as E for UQCRC2. G) Same as E for EGR1.
H) Same as E for RB1. I) Same as E for NDUFV1. J) Same as E for NFKBIA.A composite infographic of immune cell proportions and gene correlations using box plot, heatmaps, matrix and.
To assess associations between hub genes and immune components, Pearson correlation analysis was applied to the six shared central genes and infiltrating immune cells (Figure 8B). The heatmap displays correlations among different immune cells, with lines connecting genes. The direction of these correlations is represented by the colour, red represents positive associations, while blue indicates negative correlations (Figure 8B). ssGSEA analysis indicated significant positive correlations between CREBBP and Dendritic cells activated; UQCRC2, EGR1, NDUFV1 and Plasma cells, T cells CD8 and activated NK cells.
Conversely, RB1 and NFKBIA showed significant negative correlations with Plasma cells, T cells CD8, NK cells activated and Mast cells resting (Figure 8C and D). Additionally, lollipop plots were used to depict the infiltration levels of various immune cells relative to the six aging-related hub genes, providing a clear visualization of these associations (Figure 8E–J).
D-Galactose-Induced Senescence Validates Key Senescence-Related Genes
CCK-8 assay showed that D-gal induced senescence in HT22 and MC3T3-E1 cells in a concentration- and time-dependent manner. HT22 cells treated with 100 mM/L D-gal for 48h and MC3T3-E1 cells treated with 150mM/L D-gal for 24h were identified as the optimal concentrations for D-gal-induced senescence modeling (Figures 9A and B). SA-β-gal staining confirmed the induction of senescence, with D-gal-treated cells displaying pronounced blue staining, whereas control cells showed minimal staining (Figure 9C).
Quantification revealed that the percentage of senescent cells increased significantly after D-gal treatment (Figure 9D and E). Consistently, RT-qPCR analysis of six core genes demonstrated that EGR1 was upregulated, while NFKBIA, UQCRC2, RB1, CREBBP, and NDUFV1 were downregulated in HT22 cells (p<0.05 or p<0.01) (Figure 9F). In MC3T3-E1 cells, EGR1 was upregulated, NFKBIA, UQCRC2, and NDUFV1 were downregulated, whereas RB1 and CREBBP remained unchanged (ns) (Figure 9G).
Figure 9Validation of six key genes in HT22 and MC3T3-E1 cells after D-galactose (D-gal)-induced senescence. (A and B) CCK-8 assay showing the effect of D-gal concentrations and treatment durations on the viability of HT22 (A) and MC3T3-E1 (B) cells. (C) Representative SA-β-gal staining.
D-gal-treated cells exhibit typical senescence-associated blue staining, whereas control (NC) cells show minimal staining. (D and E) Quantification of SA-β-gal-positive cells. D-gal treatment significantly increased the percentage of senescent cells in HT22 (D) and MC3T3-E1 (E) cells.
(F and G) mRNA expression of six genes in HT22 (F) and MC3T3-E1 (G) cells after D-gal treatment. Data are presented as mean ± SD. * p < 0.05, ** p < 0.01, ***p < 0.001 and ns: not significant.A) Line graph. X axis, Time(d).
Y axis, HT22 Cell viability(percent). Series labels, NC, D-gal 50mM, D-gal 100mM, D-gal 150mM. Curves decrease over time, with lower values at higher D-gal’ asterisk marks shown at the right.
B) Bar chart. X axis, D-gal(mM/L) with 0, 50, 100, 150, 200. Y axis, MC3T3-E1 Cell viability.
Bars decrease as D-gal increases. C) Four micrographs arranged by columns, NC and D-gal and rows, HT22 and MC3T3-E1. Each includes a 100 um scale bar.
D) Bar chart. Y axis, SA-beta-Gal positive cells(percent). X axis categories, NC and D-gal’ label, HT22 cells’ triple asterisk above.
E) Same as D with label, MC3T3-E1 cells. F) Six bar charts with y axis, Relative mRNA expression of EGR1, NFKBIA, UQCRC2, RB1, CREBBP, NDUFV1’ x axis, NC and D-gal. G) Same set as F’ RB1 and CREBBP marked ns.A composite figure of cell viability, SA beta gal staining and mRNA bar charts for HT22 and MC3T3-E1.
Discussion
Aging is a fundamental biological driver of multiple age-related disorders, prominently including AD and OP. As the most prevalent form of dementia worldwide, AD affects approximately one in ten individuals over the age of 65, with incidence rates rising with advancing age.23,24 Similarly, OP shows a comparable age-dependent increase in prevalence.25 In AD, senescent cells accumulate in neuronal and glial populations, driving neuroinflammation and cognitive decline.26 Notably, pharmacological or genetic clearance of senescent cells has been shown to attenuate amyloid-β/tau pathology and improve cognitive decline in AD mouse models.27 Parallel findings in OP reveal that senescent cells within the bone marrow microenvironment secrete SASP factors,28 which promote osteoclast activity while suppressing osteoblast differentiation,29 ultimately leading to bone loss.30 Collectively, these findings position aging-related pathological processes as a central hub linking AD and OP. Unraveling the molecular crosstalk between aging and these two disorders is therefore critical for identifying shared therapeutic targets and early diagnostic biomarkers.
Through integrative analysis, we identified 228 aging-related genes commonly dysregulated in both AD and OP. Functional enrichment revealed their primarily involvement in apoptosis, cell cycle regulation (notably G1/S transition), focal adhesion, neuron-to-neuron synapses, ubiquitin ligase binding, and p53-related binding functions. KEGG pathway analysis further highlighted enrichment in PI3K/Akt, apoptosis, FoxO, TNF, and p53 signaling pathways.
Subsequently, we employed multiple machine learning algorithms to prioritize key aging-related genes and construct diagnostic models, ultimately identifying six significant ARGs (EGR1, NFKBIA, UQCRC2, RB1, CREBBP, and NDUFV1), which demonstrated high diagnostic values as confirmed by nomogram and ROC curve analyses. Experimental validation using D-galactose-induced senescence models revealed cell-type-specific expression patterns. All six genes were significantly altered in HT22 cells, whereas EGR1, NFKBIA, UQCRC2, and NDUFV1 were significantly altered in MC3T3-E1 cells, supporting their functional relevance to the aging process.
Functionally, these genes cluster into three interconnected modules: inflammatory and transcriptional dysregulation, mitochondrial dysfunction, and cell cycle dysregulation, which collectively illustrate a multi-layered aging-related network underpinning the comorbidity of AD and OP.
EGR1, a zinc-finger transcription factor,31 modulates immune homeostasis and stress responses; its aberrant expression has been linked to chronic low-grade inflammation characteristic of aging.32 In AD, EGR1 exacerbates neuroinflammation via microglial and astrocytic activation, accelerating neuronal injury, while in OP, it perturbs bone remodeling by modulating inflammatory cytokine production and disrupting osteoclast-osteoblast balance.33 NFKBIA, a key endogenous inhibitor of NF-κB signaling, serves as a gatekeeper of age-related inflammatory degeneration. Downregulation of NFKBIA in AD leads to NF-κB hyperactivation, promoting Aβ production, tau hyperphosphorylation, and sustained neuroinflammation;34 in OP, NFKBIA deficiency enhances osteoclastogenesis and bone resorption.35 CREBBP, a histone acetyltransferase, regulates chromatin remodeling and synaptic plasticity-related gene expression. Its dysfunction in AD leads to cognitive deficits through the impairment of synaptic plasticity.36 Collectively, these three genes underscore the central role of dysregulated inflammatory and transcriptional programs in driving the aging-related pathophysiology shared by AD and OP.
Ubiquinol-Cytochrome C Reductase Core Protein 2 (UQCRC2) and NADH Dehydrogenase (Ubiquinone) Flavoprotein 1 (NDUFV1) are core components of mitochondrial respiratory chain complexes III and I,37 respectively, and are essential for cellular energy metabolism and redox homeostasis. In AD, decreased expression of UQCRC2 disrupts mitochondrial integrity, leading to bioenergetic deficits and oxidative stress. Similarly, NDUFV1 dysfunction can cause mitochondrial energy deficits and neurological injury.38 Beyond the nervous system, these mitochondrial defects also compromise bone homeostasis: recent studies show that SIRT5-mediated desuccinylation of UQCRC2 in aged bone marrow mesenchymal stem cells (BMSCs) suppresses osteogenic differentiation.39 Thus, UQCRC2 and NDUFV1 represent critical nodes linking mitochondrial dysfunction to the pathophysiological convergence of AD and OP.
Retinoblastoma 1 (RB1) is a canonical tumor suppressor that regulates the G1/S cell cycle checkpoint via E2F binding.40 In the context of aging, loss of RB1 function permits inappropriate cell cycle re-entry in post-mitotic neurons, increasing vulnerability to injury and apoptosis in AD.41 In OP, RB1 deficiency disrupts osteoblast differentiation, as evidenced by defective cranial development in RB1-deficient mice due to uncontrolled osteoblast proliferation and arrested maturation.42 These findings identify RB1 as a key regulator of cell cycle arrest and senescence in both neuronal and bone lineages.
Cellular senescence, a key feature of aging, is marked by loss of proliferative capacity, altered metabolism and apoptosis resistance, and the secretion of pathogenic factors known as the SASP. GSEA results showed that the PI3K-Akt, NF-kappa B, Notch signaling pathway and Oxidative phosphorylation are closely related to multiple biomarkers. The PI3K/Akt pathway is crucial for cell survival and metabolic homeostasis.
Its dysregulated in AD may impair GSK3β phosphorylation,43 leading to heightened tau hyperphosphorylation and promoting the amyloidogenic processing. Moreover, PI3K/Akt intersects with mTOR signaling to influence autophagy flux, a process essential for clearing misfolded proteins such as Aβ and tau.43,44 Oxidative phosphorylation, tightly coupled to mitochondrial integrity and redox balance,43,45 is compromised in AD. The consequent accumulation of dysfunctional mitochondria and oxidative stress can, in turn, induce senescence and a distinct secretory phenotype.46 Moreover, impaired oxidative phosphorylation may trigger lipid peroxidation in neuronal membranes.47
Immune infiltration profiling revealed an expanded presence of 10 immune cell types, collectively indicative of a heightened pro-inflammatory state. Notably, elevated infiltration of CD8⁺ T cells, gamma delta T cells, NK cells, M2 macrophages, monocytes, and neutrophils was observed, suggesting their distinct contributions to sustained neuroinflammation. These observations are consistent with the involvement in immune dysregulation during aging.
Despite the protective role of the blood-brain barrier (BBB),48 peripheral immune cells can still infiltrate the brain and actively influence AD pathology. For example, bone marrow-derived monocytes efficiently infiltrate amyloid plaques, contributing to Aβ clearance.49 Neutrophils can traverse cerebral vessels, and their depletion mitigates neuropathology in 3xTg-AD mice.50 Infiltrating CD8⁺ T cells worsen AD pathology by enhancing microglial activation and neurodegeneration;51 and NK cell functional hyperactivity in AD patients inversely correlates with MMSE score,52 while their ablation in mice reduces neuroinflammation and cognitive deficits.53
There are several study limitations to address. Firstly, this research was based on comprehensive bioinformatic analyses of public datasets. The relatively small sample size of the OP discovery cohort may represent a potential source of bias and limit the generalizability of the findings.
Clinical and experimental investigations are necessary to confirm the differential expression of the six identified aging-related genes in AD patients with concomitant OP. More comprehensive clinical features, including bone mineral density measurements and cognitive assessment scores, must also be obtained to validate the predictive nomogram’s performance in real-world clinical settings. Secondly, although we identified three distinct functional modules (inflammatory/transcriptional, mitochondrial, and cell cycle), larger cohort studies are required to confirm the robustness of this classification and its relevance to disease progression.
The precise mechanistic connections between these hub genes and the observed immune dysregulation warrant further investigation through functional experiments. Finally, while we employed D-galactose-induced senescence models for initial validation, additional experiments using more specific disease models, such as APP/PS1 transgenic mice for AD and ovariectomized or aged mice for OP, are needed to determine the precise causal mechanisms by which these six genes contribute to the shared pathogenesis of Alzheimer’s disease and osteoporosis. Therefore, these findings represent potential diagnostic and therapeutic targets that require further validation in larger clinical cohorts and in vivo models.
Conclusion
This study systematically integrated transcriptomic datasets of AD and OP with curated aging-related genes to identify shared molecular drivers underlying their comorbidity. Six hub aging-related genes (EGR1, NFKBIA, UQCRC2, RB1, CREBBP, and NDUFV1) were identified via WGCNA and machine learning-based biomarker prioritization, showing diagnostic potential and significant associations with immune cell infiltration. RT-qPCR analysis in D-galactose-induced senescence models revealed cell-type-specific expression patterns: all six genes were significantly altered in HT22 cells, whereas EGR1, NFKBIA, UQCRC2, and NDUFV1 were significantly altered in MC3T3-E1 cells, with no significant changes observed in RB1 or CREBBP.
Collectively, these findings provide preliminary insights into shared aging-related molecular alterations in AD and OP. Further validation in larger independent clinical cohorts, particularly cohorts containing patients with clinically confirmed AD-OP comorbidity, and in disease-specific in vivo models is required to establish their diagnostic and therapeutic relevance.