Introduction

Alzheimer’s disease (AD) is a neurological condition associated with advanced progressive dementia and is characterized by cognitive impairments and progressive, irreversible memory loss1. As the most common form of dementia, AD accounts for 60–80% of all dementia cases in this century2. Currently, AD affects 50 million patients worldwide, ranking the sixth leading cause of mortality2. The World Health Organization reported that the prevalence of AD increases with age, peaking at 25–50% among the elderly and roughly 5–8% among those aged over 65. By 2050, AD will impact about 115 million individuals globally if current trends continue3. Patients with advanced AD experience severe malnourishment, multiple organ failure due to neuronal necrosis, and even brain death4. Diagnostic methods of AD currently include cerebrospinal fluid protein analysis and positron emission tomography of tracer molecules5. However, since pathological changes begins decades before the manifestation of clinical symptoms6, it is critical to find reliable biomarkers for an early detection of AD in order to improve patients’ pathological changes and cognitive deficits7.

Mitochondria are an organelle that functions as the power plant to cells. It has been discovered that the development of neuroplasticity is significantly influenced by mitochondria. Specifically, mitochondria control axonal development and differentiation during early neural differentiation by buffering cytoplasmic Ca2+ to promote axonal microtubule polymerization8. Hyperphosphorylated tau (p-tau) and amyloid β peptide (Aβ) have been shown to accumulate during AD pathogenesis as a result of mitochondrial malfunction and neuronal bioenergetic abnormalities. Neuronal apoptosis and mitochondrial abnormalities can be facilitated by Aβ and p-tau pathology9,10. Thus, the health of neurons depends on maintaining mitochondrial homeostasis, which can be achieved by various mechanisms of mitochondrial quality control, such as mitophagy. This process involves the phagocytosis and breakdown of damaged mitochondria11,12. Mitophagy has been shown to be essential for neuronal survival and function through preserving a healthy mitochondrial pool and preventing neuronal death13,14. Memory loss and AD pathogenesis can be exacerbated by the buildup of defective mitochondria caused by impaired mitophagy. By preventing Aβ and p-tau buildup, mitophagy restoration improves memory loss in an AD mouse model15,16. Thus, mining mitophagy-related gene (MRG) biomarkers for AD diagnosis could help overcome the treatment bottleneck of the disease.

Based on the transcriptomic data of AD patients collected from the GEO public database, this study conducted an in-depth analysis to develop a MRG-related signature using machine learning algorithms, including weighted gene co-expression network analysis (WGCNA). Additionally, the robustness of the model was confirmed using external data. Furthermore, a TF-mRNA interaction network was established and the expressions of feature genes in AD were determined. Finally, the relationship between immune cell infiltration and biological pathways and the identified genes was analyzed. In conclusion, our study may provide new perspectives and strategies for the detection of AD and its treatment.

Methods

Data acquisition

The scRNA-seq expression profiles of AD in GSE13290317 and GSE12206318 scRNA-seq expression profiles were sourced from GEO (https://www.ncbi.nlm.nih.gov/geo/) database. Specifically, GSE132903 samples were from the middle temporal gyrus and some of the GSE122063 samples were from the frontal lobe. Based on the platform annotation files of microarray data, the probe-level expression profiles were translated into gene-level expression profiles. When numerous probes matched to the same gene, the average of those probes was computed as the gene expression, whereas the probe was eliminated if one probe matched to many genes. The maximum value obtained from several probes matching to one gene was considered as the gene’s expression value. Samples without status or clinical follow-up information were excluded while retaining samples with a survival time of more than 30 days. After preprocessing, the data of GSE132903 (training set: AD = 97; Controls = 98) and GSE122063 (test set: AD = 56; Controls = 44) were acquired. Finally, 34 MRGs from a previous literature were extracted for further analysis19.

WGCNA construction

WGCNA, a systems biology technique, could characterize the patterns of gene correlation among various samples to identify highly synergistically varying gene sets and candidate biomarker genes or therapeutic targets. This study developed a weighted gene co-expression network based on 195 diseased and control samples utilizing the “WGCNA” software20 to find co-expression networks and select genes from various clusters.

The autophagy-related score of the dataset GSE132903 was evaluated using the ssGSEA method, and the autophagy-related gene set was used as the background gene set to calculate the autophagy score of the training set samples. The R software package “WGCNA” software20 was used to identify the autophagy-related gene modules of patients with AD, and the ssGSEA scoring results were used as traits to perform WGCNA analysis on the training set. Using the WGCNA’s “pickSoftThreshold” R function, the ideal soft threshold power was determined to ensure the network’s scalability. Subsequently, the adjacency matrix was transformed into a topology matrix, which was further converted into an adjacency matrix. Average-linkage hierarchical clustering approach was adopted based on the topological overlap measure to cluster the genes21. After identifying the gene modules using the dynamic shear method, we determined the eigenvector values (eigengenes) of each module and combined close modules into a new one by setting height = 0.25, deepSplit = 2, and minModuleSize = 80. Then, based on gene significance (GS) and module affiliation (MM), the “WGCNA” software was employed to analyze the interaction strengths to reflect the relationship between the gene modules and AD patients22. To assess module-trait interactions based on the correlations between the module genes and clinical trait, the “Heatmap” software23 was applied to extract various module-characterized genes according to the first principal component (PC) of module expression. Next, genes with the strongest positive or negative module-trait connections were selected. Furthermore, the genes obtained from the modules were imported into the STRING database (Version: 12.0) for protein–protein interaction (PPI) analysis24 under CONFIDENCE ≥ 0.4.

DEGs analysis

Robust multi-array average (RMA) approach in the “affy” R package25 and Bioconductor package were adopted to perform background correction, quantile normalization, and professional summarization of gene expression profiling data derived from the 195 samples. Using the “limma” program26,27,28, gene expression data were determined to identify those whose expression significantly changed between AD and control samples. Under the thresholds set to |log2FC|≥ 1 and padj < 0.01, significant DEGs were filtered using the “limma” package29. DEGs were visualized using the “ggplot2” program30 into a volcano plot. Finally, the overlapping genes of the DEGs and the module genes obtained by WGCNA were extracted.

Screening MRGs-related feature genes to develop a diagnostic model and validation

Further perform LASSO regression and Recursive Feature Elimination (RFE) on overlapping genes. Specifically, LASSO regression in R language’s “glmnet” package31 was performed to reduce the gene number while retaining genes. Take the characteristic genes at the intersection of LASSO and RAF as the biomarkers of this study and establish a diagnostic model. Then, the R package “timeROC”32 was employed to compute the AUC and show time-dependent receiver operating characteristic (ROC) curves to evaluate the efficacy of the biomarkers in differentiating AD from control samples in the training set GSE132903 and the test set GSE122063.

Establishment of a TF-mRNA regulatory network

By accessing JASPAR ( https://jaspar.elixir.no/), a TF motif database, the TFs of biomarkers were predicted using the NetworkAnalyst (https://www.networkanalyst.ca/) to create a TF-mRNA interaction network. Cytoscape (version 3.7.2)33 was employed to visualize their interactions.

Immune infiltration analysis

We used ssGSEA34 to calculate the enrichment scores of 28 types of immune cells in each sample35. Using Spearman’s correlation coefficient, the relationship between immune cell infiltration and the biomarkers was compared.

Functional enrichment analysis

In functional enrichment analysis, Gene Ontology (GO) uses cellular component (CC), molecular function (MF), and biological process (BP) terms to characterize the roles of genes and proteins. Kyoto encyclopedia of genes and genomes (KEGG) stores the information about systemic, chemical and genetic functions36,37,38,39. Here, the primary activities and signaling pathways of the DEGs were analyzed by performing functional enrichment analyses. Using the R program “clusterProfiler”40, the GO and KEGG functions of DEGs were analyzed. The Top 10 functions enriched in BP, MF, and CC terms in the GO enrichment and the top 20 enriched KEGG pathways were visualized. The bubble plot of the top 20 pathways in the KEGG pathway enrichment results was plotting employing the “GSVA” software package41. And based on the results of functional enrichment analyses, ssGSEA scores were computed to evaluate the correlation between biomarkers and the pathway scores. The pathways of h.all.v2023.2.Hs.symbols.gmt were collected from GSEA website (http: //www.gsea-msigdb.org/gsea/index.jsp).

Statistical analysis

Statistical analyses were performed using R software (version 3.6.0). Differences in continuous variables between two groups were analyzed by the Student’s t-test (parametric test) or the Wilcoxon rank-sum test (non-parametric test). The correlation between biomarker expressions and immune scores was measured by Spearman correlation analysis. *p < 0.05, **p < 0.01, ***p < 0.001, and ****p < 0.0001 denoted a statistical significance, while ns indicated p > 0.05. The flowchart of this study can be seen in Fig. S1.

Results

Screening MRGs-related genes by WGCNA

Using the list of MRGs as the background gene set, it was found that all of the AD samples in the dataset had considerably lower mitophagy-related scores than the control samples (Fig. 1A). Next, we identified gene modules linked to mitophagy in AD patients using WGCNA, with ssGSEA score as the traits against the training set. According to the module correlations, no modules could be merged, and the soft threshold β was determined to be 12 by the pickSoftThreshold function (R2 = 0.9, Fig. 1B). Nine modules were identified when the minimum number of genes in a module was 80, and no notable outlier samples were discovered (Fig. 1C). A heat map of the module-shape correlation was visualized (Fig. 1D). The association between modules and mitophagy data was analyzed to identify clinically significant modules. The nine modules included 2932 genes, with the MEbrown module showing the strongest association with mitophagy (cor = 0.67 and p = 1.17e−26), therefore this significant module was chosen for further study. Using brown color to represent the MEbrown module, a strong positive correlation between MM and GS (cor = 0.66, p < 1.e−200, Fig. 1E) was detected. Next, 39 genes with |GS|> 0.75 and |MM|> 0.75 that were closely related to mitophagy were identified. These genes were imported into the STRING database for analyzing PPIs after removing genes without interactions (CONFIDENCE ≥ 0.4). The gene interaction relationships and confidence scores are shown in Table S1. Finally, 12 genes were included as the modular MRGs for subsequent investigation.

Fig. 1
figure 1

Weighted co-expression network construction and key module identification in AD. (A) Box plot of ssGSEA results. (B) WGCNA soft threshold screening graph. (C) Dendrogram of all genes clustered based on a dissimilarity measure (1-TOM). (D) Module-trait correlation plot. (E) GS-MM plot of MEbrown modules.

Full size image

Identification of DEGs and functional analysis

417 DEGs (257 down-regulated and 160 up-regulated genes) was obtained (Fig. 2A). Next, a gene expression heatmap for the top 100 DEGs from both AD and control samples was generated to investigate the relationships between the DEGs and the clinical features (gender, female/male, age, ≤ 80/ > 80 years). A negative correlation in the expression patterns of up- and down-regulated genes was observed (Fig. 2B). GO and KEGG functional enrichment analysis showed that the DEGs were significantly enriched to neurotransmitter transport, neurotransmitter secretion, synaptic vesicle recycling, signal release, and synaptic signal release in BP term (Fig. 2C). In CC term, the genes were mostly enriched to presynapses, glutamatergic synapses, synapses across neurons, transport vesicle membranes, and transport vesicles (Fig. 2C). In MF term, the DEGs were most significantly enriched to binding, GTPase activity, SNARE binding, calmodulin binding, and protein C-terminal binding (Fig. 2C). Moreover, the DEG were significantly enriched to the pathways of gap junction pathways, insulin secretion, GABAergic synapses, long-term potentiation, and synaptic vesicle cycle (Fig. 2D). It was discovered that there was a notable enrichment of these genes to pathways associated with synaptic function.

Fig. 2
figure 2

Enrichment analysis of DEGs as well as GO and KEGG between AD and control samples. (A) Differential gene volcano plot. (B) Heatmap of the top 100 differential genes. (C) Differently expressed gene GO function analysis bubble chart. (D) Differently expressed gene KEGG functional analysis bubble plot.

Full size image

Screening of MRGs-related signature genes

By taking the intersection of WGCNA-derived genes and DEGs, five overlapping genes were ultimately identified (Fig. 3A). LASSO regression and RFE techniques were employed to compress these 5 genes. Specifically, the model reached its optimal value in LASSO regression at lambda.1se = 0.1198712 (Fig. 3B,C). Then, five candidate genes were selected based on their characteristics by RFE algorithm when the model error was minimized (Fig. 3D). Finally, by taking the intersection of the two machine learning experiments, the YWHAZ and NDE1 genes were identified as biomarkers.

Fig. 3
figure 3

Identification of characterized genes associated with MRGs in AD. (A) Venn diagram of the intersection of mitophagy-related modular genes and differential genes screened by WGCNA. (B) Trajectories of each independent variable with lambda. (C) Confidence intervals under lambda. (D) The plot of SVM-RFE pan error versus the number of features.

Full size image

Diagnostic modeling and validation

A diagnostic prediction model for AD was created based on the YWHAZ and NDE1 in the training set GSE132903. After plotting the ROC curves, the AUC values of YWHAZ and NDE1 were calculated to be 0.777 and 0.793, respectively, which were both higher than 0.7 (Fig. 4A). This indicated that our signature genes can effectively diagnose AD. Subsequently, the results of ROC analysis showed that the area under the curve of the two-gene model was 0.796 (Fig. 4B), suggesting that a strong predictive ability of the model. Moreover, the mRNA expression of YWHAZ was significantly lower in AD samples than in control samples, while NDE1 was high-expressed in AD samples (Fig. 4C). The prediction stability of the model gene in distinguishing AD from control samples was tested in the test set GSE122063 cohort with the same coefficients of the training set. It can be seen that the two genes had a strong predictive performance in the GSE122063 cohort, with an AUC value of 0.792 and 0.795 for YWHAZ and NDE1, respectively (Fig. 4D). The prediction was also accurate in the GSE122063 test (AUC = 0.848, Fig. 4E), further confirming the reliability of the diagnostic model. Similarly, the expression of YWHAZ was significantly lower in AD samples than in control samples, while the expression of NDE1 was significantly higher in AD samples than in control samples (Fig. 4F). This suggested that YWHAZ and NDE1 genes play different regulatory roles in the progression of AD.

Fig. 4
figure 4

Establishment as well as validation of diagnostic model based on MRGs-related signatures. (A) ROC curves for a single feature gene in the GSE132903 training set. (B) ROC curves of the combination of 2 feature genes in the GSE132903 training set. (C) Expression of hub genes in AD and control samples in the GSE132903 training set. (D) ROC curves for a single feature gene in the GSE122063 test set. (E) ROC curves for a combination of 2 feature genes in the GSE122063 test set. (F) Expression of hub genes in AD and control samples in the GSE122063 test set.

Full size image

TF-mRNA regulatory network analysis

The TFs of YWHAZ and NDE1 genes were predicted using the NetworkAnalyst online tool, and a TF-mRNA interaction network was built using the JASPAR database. As shown in Fig. 5, the network consisted of 26 TFs and 2 hub genes. In particular, four TFs (FOXC1, FOXL1, HOXA5, and GATA2) were associated with both YWHAZ and NDE1, suggesting that they may play an important role in the two biomarkers targeting in AD.

Fig. 5
figure 5

Network diagram of TF and mRNA interactions. The purple dots in the figure represent mRNA molecules, while the green squares represent transcription factors.

Full size image

Immune cell infiltration and its correlation with biomarkers

The results of immune cell infiltration analysis showed that half of the immune cells, including natural killer cells, plasmacytoid dendritic cells, effector memory CD8 T cells, and Type 1 T helper cells, were considerably more abundant in control samples than in AD samples (Fig. 6A,B, p < 0.05). Next, the relationship between immunological scores and YWHAZ and NDE1 was analyzed. The expression of NDE1 was negatively correlated with the infiltration of activated CD4 T cells, effector memory CD4 T cells, memory B cells, regulatory T cells, and type 2 T helper cells (Fig. 6C, p < 0.05) but positively correlated with the infiltration of the rest immune cell types (Fig. 6C, p < 0.05). YWHAZ was positively related to the infiltration of activated CD4 T cells, effector memory CD4 T cells, memory B cells, type 2 T helper cells, and eosinophils but negatively related to the rest immune cell types (Fig. 6C, p < 0.05).

Fig. 6
figure 6

Correlation of YWHAZ and NDE1 genes with immune cell infiltration. (A, B) Differences in immune cell infiltration between AD and control. (C) Correlation of YWHAZ and NDE1 genes with immune cells in AD (* represents whether it is significant or not, red–orange represents positive correlation and green represents negative correlation).

Full size image

Correlation between the characteristic genes and pathways

The relationship between the YWHAZ and NDE1 genes and the pathway scores was examined. It was found that NDE1 was negatively correlated with the pathways MYC targets v1, spermatogenesis, pancreas beta-cells, and oxidative phosphorylation but positively correlated with the pathways mitotic spindle, wnt/beta-catenin signaling, p53 pathway and notch signaling (Fig. 7, p < 0.05). On the other hand, YWHAZ was positively correlated with wnt/beta-catenin signaling, p53 pathway, notch signaling and myogenesis but negatively correlated with the pathways MYC targets v1, oxidative phosphorylation, pancreas beta-cells and spermatogenesis (Fig. 7, p < 0.05). Based on the above results, it was speculated that YWHAZ and NDE1 may have different effects on the occurrence and development of AD.

Fig. 7
figure 7

Correlation between characteristic gene expression and pathways (* represents whether it is significant or not, red represents positive correlation and blue represents negative correlation).

Full size image

Discussion

AD is the primary cause of death for patients with dementia and at present there is no cure for it42,43, necessitating the development of preventive and diagnostic methods for the disease. Though we currently still face a lack of molecular indicators for an accurate diagnosis of AD, research showed that AD pathogenesis is significantly influenced by neuronal mitophagy44. Due to their high energy requirements, neurons are the most susceptible cell type to mitochondrial malfunction44,45. Mitophagy may lead to neurological dysfunction because Aβ buildup in the nervous system triggers cellular oxidative stress, a shortage of vital nutrients, and neuronal cell apoptosis46. It has been proven that mitophagy lowers oxidative stress and eliminates damaged mitochondria, showing the potential to be developed as a preventive and therapeutic approach against neurodegenerative disorders. In the present work, YWHAZ and NDE1, which were identified as having potential for diagnosing AD, were identified as two MRGs by integrative analyses based on public databases.

Since the pathophysiology of AD is believed to be significantly influenced by mitophagy in neurons44, building an AD prediction model with MRGs is a reliable method for clarifying the cause of AD and developing new treatment strategies. This study used machine learning experiments to identify YWHAZ and NDE1 for an AD diagnostic model. A major protein implicated in numerous signal transduction pathways is 14–3–3 zeta protein isoform 1 (YWHAZ)47. According to previous report, YWHAZ lowers the likelihood of AD by interacting with the tau gene in NFT, a distinctive neurohistological hallmark of AD48. YWHAZ is directly associated with AD49 because the protein expression of YWHAZ is downregulated in the hippocampus of AD patients and it may be a key regulatory gene in the prefrontal cortex of AD patients50,51. This is consistent with the current finding that the expression of YWHAZ in AD samples was lower than that in control tissues. Neurodevelopmental protein 1 (NDE1) is, a protein-coding gene also known as nude, is thought to be a risk factor for major psychiatric illnesses and is substantially expressed in neurological and psychiatric diseases52,53. Mutations in NDE1 in mitochondria may stimulate the occurrence of multiple biosynthetic pathways that increase mitochondrial quality and quantity, including mitophagy54. It was discovered that NDE1 is crucial for microtubule organization, mitosis, and neural migration55. Microtubule loss observed by Zempel et al. is a risk factor for the deterioration of AD56. According to a study, a higher level of NDE1 in AD might be caused by altered NDE1 copy number in cortical neurons, which leads to poor neurodevelopment caused by secondary genomic mutations57. NDE1 has been shown to recruit LIS1 to dynein58, and knockdown of LIS1 can inhibit the retrograde movement of mitochondria in the axons of rat hippocampal neurons59. Thus, screening crucial genes involved in mitophagy to prevent the development of AD is a promising strategy. The present work confirmed that YWHAZ and NDE1 were closely associated with AD progression and had the potential to serve as biomarkers for AD.

The characterization of putative TF-mRNA regulation networks was also performed. Among the 26 TFs, YWHAZ and NDE1 were linked to FOXC1, FOXL1, HOXA5, and GATA2. Forkhead box C1 (FOXC1), a TF of the forkhead box family, plays a regulatory role in cell viability and oxidative stress tolerance60. Cerebellar structure develops relying on FOXC1, and the pathophysiology of Dandy-Walker deformity is linked to FOXC1 deletion61. The forkhead box L1 (FOXL1) gene induces cell cycle arrest and death to suppressor cells, which facilitates brain development62. A major TF controlling neuronal shape is homology box A5 (HOXA5). In mice, disruption of Hoxa5-mediated neuronal morphogenesis can influence spatial learning and memory63. Furthermore, tau neuropathology is closely involved in brain hypermethylation of the HOXA gene cluster64. Contribution from hematology65,66 proposed GATA-binding protein 2 (GATA2) as a critical TF that controls the expression of monoamine oxidase A in neurological and cardiovascular disorders67. Hence, it is possible that these TFs fulfill unique regulatory functions in the initiation and progression of AD, albeit more research is required for confirmation.

Furthermore, GO and KEGG study revealed that the biological functions of DEGs were primarily enriched in pathways related to signaling and insulin secretion at synapses, long-term potentiation, and gap junction pathways. These findings were consistent with some previous researches. For instance, synapses are found to be the basic building blocks of the brain, and synaptic activity promotes the development of new synapses and the maturation of mushroom-like spines68,69. This process enhances the adaptation of synaptic strength to environmental changes and is crucial for learning and memory. Early AD patients’ brains show a loss of synapses and disruption of synaptic activity70. It has been confirmed that the primary cause of cognitive problems in AD patients is synapse loss and decreased synaptic activity71. According to past research, p-tau accumulation causes mitochondrial abnormalities that trigger neuronal death9, while p-tau deficiency increases the body’s release of insulin. Clinical research also demonstrated that blood glucose levels are lowered in AD patients, and that serum tau protein and blood glucose levels in healthy controls have a favorable correlation72. Furthermore, perturbations in neuronal Ca2+ levels caused by mitochondrial dysfunction in AD may lead to an imbalance in cell membrane, thereby causing neuronal death73. Overall, these pathways may provide theoretical guidance and new insights for subsequent studies in AD.

This study examined the relationship between immune infiltration cells and diagnostic features. It was found that multiple immune cells, including activated CD8 T cell, central memory CD8 T cell, effector memory CD8 T cell, immature B cell, T follicular helper cell, and etc., were positively correlated with NDE1 gene but negatively correlated with YWHAZ. However, the correlations were opposite for activated CD4 T cells and effector memory CD4 T cells. This indicated that these cells might play a part in the onset and progression of AD. Numerous studies indicated that CD4+ T cells, CD8+ T cells, and cytokines are involved in the pathogenesis of AD74,75. CD4+ T cells may orchestrate immunological systems, interact with microglia, and facilitate amyloid clearance, which could all lead to AD pathogenesis and offer therapeutic interventions76. Clonally expanded CD8+ T effector memory CD45RA+ (TEMRA) cells were detected in the cerebrospinal fluid of patients with AD77. CD8+ TEMRA cells are negatively correlated with cognitive functions78. In AD, disrupted balance between CD4+ T cells and CD8+ T cells may cause immune system dysregulation, which could promote or exacerbate the pathological process of AD79. The expansion of B cells in AD promote the disease progression by producing immunoglobulins that are deposited around Aβ plaques, whereas therapeutic elimination of B cells can slow down AD deterioration80,81. In summary, these studies suggested that YWHAZ and NDE1 may influence the occurrence of AD by affecting immune infiltration.

According to the pathway enrichment scores, oxidative phosphorylation was negatively correlated with NDE1 but positively correlated with YWHAZ, while opposite Wnt/β-catenin signaling was positively correlated with NDE1 but negatively correlated with YWHAZ. Activation of the wnt/β-catenin pathway prevents the deposition of Aβ and p-tau, reduces neuroinflammation, and slows down the progression of AD82,83. Wnt signaling can maintain and even restore the blood–brain barrier (BBB) properties of endothelial cells (ECs), thus influencing the occurrence of AD84. Some studies have also found that restoring Wnt/β-catenin signaling is a promising therapeutic strategy for AD85. Abnormal oxidative phosphorylation will cause Aβ deposition and tau hyperphosphorylation, facilitating the onset of AD86. As previously mentioned, NDE1 is closely associated with mitochondrial mutations. Previous research revealed that genes involved in OxPhos, including those participating in complexes II and V, are significantly downregulated in AD subjects87. In fact, mitochondrial dysfunction may lead to energy failure in neurons, thereby inducing synaptic dysfunction in cognitive impairment88. These results suggested that YWHAZ and NDE1 may influence the development of AD through these pathways.

It is also equally important to acknowledge the limitations of our study. Firstly, only a small number of patients from public databases were analyzed as the basis to support our findings. To address this, future studies could expand our sample size by including a more diverse cohort of patients from multiple sources (both domestic and international). Secondly, prospective investigations are required to rule out potential bias caused by retrospective recruitment of patients. In addition, the role of the two MRGs identified by this study for AD mitophagy should be verified by in vivo animal experiment in animals and in vitro cell experiments. Finally, further research is required to examine the relationship between AD hallmark genes and immune cell infiltration.

Conclusion

This study developed a diagnostic model based on two hallmark genes linked to mitophagy in AD using machine learning algorithms. The level of immunological infiltration and biochemical pathways in AD patients were accurately predicted by the two signature genes, which showed a strong predictive potential in both AD and control samples. In summary, investigation into the characteristics of mitophagy may provide theoretical guidance for exploring the mechanisms of AD occurrence and development.