Identification and experimental validation of Alzheimer's disease hub genes via bioinformatics and machine learning

Abstract

Background

Alzheimer's disease (AD) is a complex neurodegenerative disorder.

Objective

To identify diagnostic and predictive biomarkers for AD.

Methods

Based on three GEO datasets of human brain tissue from AD patients and controls, weighted gene co-expression network analysis (WGCNA) and enrichment analysis were used to identify AD-related gene modules. Hub genes were screened via protein-protein interaction (PPI) analysis and three machine learning algorithms. Diagnostic efficacy was evaluated using receiver operating characteristic (ROC) curves. Immune cell infiltration and hub gene expression correlations were analyzed using xCell. In vivo validation was performed using an AD mouse model.

Results

The magenta module was significantly correlated with AD. PPI network analysis identified 15 AD-related genes, mainly enriched in mitochondria and ribosomes. Two hub genes, DLAT and CCDC88b, were identified. DLAT was significantly downregulated in AD, and CCDC88b was upregulated (p < 0.01); both findings were validated via qPCR in AD model mice. ROC analysis showed good diagnostic performance. Immune infiltration analysis revealed macrophages as the dominant cell type, with hub gene expression associated with immune cell presence.

Conclusions

DLAT and CCDC88b are potential novel biomarkers for AD and may serve as targets for therapeutic intervention.

Keywords

Alzheimer's disease clinical diagnostic model experimental validation machine learning WGCNA

Introduction

With the increasing aging of the global population, the prevalence of Alzheimer's disease (AD) is increasing sharply. In 2018, Alzheimer's Disease International estimated that approximately 50 million people worldwide were living with dementia, and the global prevalence is projected to triple by 2050.¹ The clinical manifestations of AD include progressive deterioration of cognitive and memory functions, a gradual decline in the ability to perform activities of daily living, and the presence of various neuropsychiatric symptoms and behavioral disturbances. The most prominent pathological features of AD are amyloid plaques formed by amyloid-β (Aβ) deposition and neurofibrillary tangles (NFTs) composed of hyperphosphorylated tau protein (P-tau).² Moreover, an increasing body of evidence has demonstrated that neuroinflammation³ and mitochondrial dysfunction⁴ play critical roles in the pathogenesis of AD.

Neuroinflammation is one of the key pathological features of AD, and its persistent activation has been shown to accelerate neurodegeneration. The deposition of Aβ, abnormal phosphorylation of tau, activation of glial cells, and initiation of inflammasomes collectively contribute to the formation of a complex inflammatory network.⁵ Moreover, peripheral immune cells can also trigger or exacerbate local inflammatory responses by regulating the central nervous system; moreover, peripheral infections may even promote the innate immune training of microglia, thereby further activating neuroinflammation.^6,7 These phenomena suggest that immune dysfunction not only affects neurons directly but also indirectly accelerates neurodegeneration by regulating inflammatory responses.

Mitochondrial dysfunction is also a major focus in the study of AD pathogenesis. In the brains of AD patients, significant alterations in mitochondrial function, morphology, distribution, and apoptosis-inducing mechanisms have been observed.^4,8 As the central organelle of cellular energy metabolism, mitochondria play a critical role in maintaining neuronal function and antioxidant defense.⁹ Neurons have exceptionally high energy demands, and mitochondrial metabolic dysfunction can lead to synaptic impairment, neuroinflammation, and cell death.^10,11 The tricarboxylic acid (TCA) cycle and the electron transport chain are the primary metabolic pathways of mitochondria, and their dysregulation can lead to insufficient energy supply and exacerbate oxidative damage and apoptosis through excessive generation of reactive oxygen species (ROS).¹² This vicious cycle not only further impairs mitochondrial function but also plays a pivotal role in accelerating neurodegeneration. In drug development, approximately 200 candidate drugs for AD were abandoned in clinical trials between 2003 and 2019, with none receiving FDA approval.¹³ Traditional symptomatic treatments primarily include cholinesterase inhibitors, such as donepezil,¹⁴ galantamine,¹⁵ and rivastigmine,¹⁶ as well as N-methyl-D-aspartate (NMDA) receptor antagonists, such as memantine.¹⁷ However, these drugs have limited efficacy in delaying disease progression.^14,18,19

In recent years, significant progress has been made in the development of anti-amyloid therapies that target Aβ. For example, in 2021, the FDA granted accelerated approval for aducanumab.²⁰ In 2023, lecanemab was also approved, and clinical trials have shown that this class of drugs can significantly reduce the burden of Aβ plaques in the brain.²¹ Additionally, donanemab has also demonstrated the potential to improve cognitive function in early AD patients in phase III clinical trials.²² Despite the promise of various emerging therapies for AD, their long-term efficacy, safety, and sustained impact on improving cognitive function remain insufficiently validated. Therefore, there is an urgent need to develop reliable biomarkers for early diagnosis, monitoring disease progression, and enabling targeted therapies for AD.

With the development of high-throughput sequencing and bioinformatics technologies, systematic analysis of large-scale molecular data has become an important approach to uncover disease mechanisms and advance drug development. The introduction of machine learning has further enhanced the ability to mine and predict biological data.²³ With the expansion of the data scale, the integration of bioinformatics and machine learning has become an important direction in biomedical research. Despite the emergence of various predictive methods for AD, existing bioinformatics approaches still face issues of low accuracy and insufficient efficiency, making it difficult to meet the demands for screening and early diagnosis. For example, traditional differential gene expression analysis often leads to the loss of critical information, limiting the selection and application of biomarkers.²⁴ The present study comprehensively applied weighted gene co-expression network analysis (WGCNA), protein-protein interaction (PPI) analysis, and functional enrichment analysis to screen gene modules and candidate genes closely associated with AD systematically, providing a theoretical basis for subsequent diagnostic and therapeutic target development. Considering that different analytical algorithms may introduce result biases, three machine learning methods—least absolute shrinkage and selection operator (LASSO), random forest, and support vector machine recursive feature elimination (SVM-RFE)—were further employed to cross-select key hub genes. The diagnostic efficacy of the hub genes was subsequently evaluated via receiver operating characteristic (ROC) curves, and their relationship with immune cell infiltration was analyzed via the xCell database. The changes in the expression of key genes were validated experimentally. Ultimately, this study identified DLAT and CCDC88b as potential novel biomarkers for AD, providing new theoretical insights for the early diagnosis and targeted treatment of AD.

Methods

Acquiring data

The transcriptomic data of AD patients and control groups, specifically GSE174367, GSE122063, and GSE159699, were obtained from the GEO database. GSE174367 is based on gene chip data from the GPL24676 platform and includes 44 human prefrontal cortex brain samples from AD patients and 46 control samples. The samples were sourced from the Alzheimer's Disease Research Center at the University of California, Irvine (UCI), and were obtained with approval from the Institutional Review Board of the University of California, Irvine.²⁵ GSE122063²⁶ includes 56 AD samples and 44 control samples, all from the human frontal cortex, sourced from the Alzheimer's Disease Center Brain Bank at the University of Michigan. GSE159699 comprises 12 AD samples and 17 control samples derived from the human lateral temporal lobe, with AD samples selected on the basis of Braak and CERAD staging and clinical diagnosis.²⁷

Weighted gene co-expression network analysis

The gene co-expression network was constructed via the R package “WGCNA”.^28,29 A scale-free co-expression network (scale-free R2 > 0.85) was established with a power of soft thresholding = 8. The minimum number of genes per module was set to 50. Genes with similar expression profiles were first divided into different modules via a dynamic tree-cutting method and then merged with a maximum dissimilarity of 0.25.

Functional enrichment of AD-related module genes

Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis was conducted via KOBAS to explore the biological pathways associated with the selected genes.³⁰ The R package “ClusterProfiler” was used for Gene Ontology (GO) analysis.^31,32 Biological process (BP) and molecular function (MF) GO terms were analyzed in this study. A p value less than 0.05 was considered significant.

AD-related gene identification and the PPI network

Genes in the AD-related module with gene significance (GS) > 0.2 and module membership (MM) > 0.9 were selected as AD-related genes. The online tool GeneMania was used to construct a PPI network and investigate the functions of AD-related genes.³³

Correlation analysis of AD-related genes and immune cell types

The R package “ggpubr” was used to conduct Pearson correlation analysis to illustrate the relationships between the expression levels of AD-related genes and immune abundance.A correlation heatmap was generated with the R package “pheatmap”.

Machine learning

This study utilized the transcriptomic data of the prefrontal cortex from AD patients and normal controls in the GSE122063 dataset and applied three machine learning algorithms to screen for candidate genes associated with AD. LASSO, a feature selection method commonly used in multivariate linear regression analysis, introduces a penalty term to shrink regression coefficients, effectively simplifying the model and preventing collinearity and overfitting.³⁴ RF, an algorithm based on ensemble decision trees, was used to assess the importance of various features and identify the genes that contribute most to classification.³⁵ SVM-RFE is a supervised learning method that iteratively removes low-weight features, gradually optimizing the feature set to improve model performance.³⁶ LASSO analysis was performed via the “glmnet” package (version 4.1.8), with the optimal feature set determined on the basis of the 1-SE criterion to balance model complexity and generalization ability. The RF algorithm was used to score gene importance, selecting genes with an importance score greater than 0.25 as candidates. Finally, a robust set of hub genes was identified by integrating the results from the LASSO, SVM-RFE, and RF feature selection methods.

Nomogram construction and ROC curve evaluation

Nomograms are important tools for clinical prediction and are constructed via the “rms” package (version 6.5-0) along with hub genes. In this nomogram, each gene is assigned a specific score (“point”), with the “total score” representing the cumulative score of all included genes. This scoring system transforms complex genetic data into a user-friendly graphical representation, helping clinicians make decisions. To assess the diagnostic performance of the nomogram, we constructed ROC curves. The area under the curve (AUC) and the corresponding 95% confidence interval (CI) were calculated to quantify diagnostic performance, providing a measure of the model's ability to distinguish between the AD and control groups. The AUC was calculated via the “pROC” package (version 1.18.5) in R, a nonparametric method for estimating the AUC. This method is particularly advantageous because it does not assume a specific distribution of the data, making it suitable for a wide range of datasets. The “pROC” package uses the DeLong method to calculate the AUC, a widely accepted ROC analysis technique in biomedical research. This method compares the observed ROC curve to the null hypothesis (AUC = 0.5) to accurately and unbiasedly estimate the diagnostic ability of the model. The AUC is a measure of the model's ability to distinguish between positive and negative cases, with values closer to 1 indicating higher diagnostic accuracy.³⁷ The predictive utility of the nomogram was externally validated via an independent dataset (GSE159699).²⁷ This validation included constructing ROC curves for the validation set to ensure the robustness and applicability of the model across different sample cohorts.

Immune infiltration analysis

To assess the infiltration characteristics of immune cells in AD brain tissue, this study utilized the xCell method.³⁸ xCell is a computational tool based on gene expression data that can be used to infer the abundance of 64 cell types quantitatively. It is widely used in tumor immunology and brain tissue microenvironment research. We used the “xCellAnalysis” function provided by xCell and performed the analysis via its default parameters, which are consistent with the settings in the original publication and other existing studies, with no additional parameters available for adjustment.^38–40 In this study, on the basis of 44 AD samples from the GSE122063 dataset, xCell was used to calculate the scores for 64 cell types, as well as three composite scores: the immune score, stroma score, and microenvironment score. Since only these three composite scores are provided in the official xCell output, we systematically analyzed and interpreted these scores to comprehensively reflect the immune status, stromal composition, and overall microenvironmental changes of the samples.

qPCR detection of hub gene expression

Experimental animal model grouping

This study utilized 3×Tg-AD transgenic mice (129-Tg (APPSwe, tauP301L) 1LfaPsen1tm1Mpm/Mmjax, 034830-JAX) along with their strain-matched wild-type (WT) control mice (B6129SF2/J, 004807-JAX), both of which were purchased from Jackson Laboratory (Bar Harbor, Maine). All animals were housed in a specific pathogen-free (SPF)-grade animal facility (Certificate No. SYXK 2021-0003) with free access to food and water, maintained on a 12-h light/dark cycle and maintained at a temperature of 22–24°C. The study adhered to the EU Animal Care and Use Guidelines (Directive 2101/63/EU) and received approval from the Animal Use and Care Committee of Zunyi Medical University (Approval No. ZMU21-2107-019). During breeding, a ratio of 1 male to 3 females was maintained per cage to propagate new generations of 3×Tg-AD and WT mice. The genotypes of the offspring were confirmed through polymerase chain reaction (PCR) and agarose gel electrophoresis, following the Jackson Laboratory protocols, to verify that the 3×Tg-AD mice expressed the mutated human APP, PS1 and Tau genes, whereas the WT mice did not carry these mutations. Twelve-month-old 3×Tg-AD male mice were used as the experimental group, while age-matched C57BL/6 WT male mice served as the control group. Each group consisted of six mice, with each mouse weighing approximately 35 grams. The study followed strict ethical standards in line with animal care guidelines.

Behavioral testing

Y maze test: The Y maze consists of three arms: the initial arm, the novel arm, and the other arm. Before the formal test, the mice were trained for 2 d. During training, the entrance to the novel arm was blocked with a partition, and the mice were placed at the starting point of the initial arm, allowing them to freely explore between the initial arm and the other arm for 20 min. After 24 h, the next training session was conducted. After the training sessions, the formal test began. The partition at the entrance to the novel arm was removed, and the mice were placed again at the starting point of the initial arm. The testing period lasted for 5 min. During this time, the activity trajectory and related parameters of the mice, including the spontaneous alternation rate (%) and the number of entries into the novel arm as a percentage of the total number of entries into all three arms (%), were recorded. These data were analyzed to assess the exploratory behavior and spatial memory ability of the mice.

Morris water maze test: The Morris water maze (MWM) test was used to assess the learning and memory abilities of AD model mice. The water maze consisted of a circular pool with a diameter of 1.2 m and a height of 40–50 cm, filled with water to a depth of 30 cm. A circular platform with a diameter of 9 cm and a height of 28 cm was placed in one of the quadrants of the pool. A camera was positioned vertically above the center of the maze, approximately 1.8 m high, to capture the entire test area. The testing system included a camera, computer, and image acquisition card. The water temperature was maintained at 22 ± 1°C, and the pool was divided into four equal quadrants (arranged counterclockwise), with the platform located in the center of the third quadrant. The water level was adjusted so that the platform was submerged by 1 cm. Edible-grade dye was added to the water to make the platform invisible. White curtains are hung around the pool to prevent shadows and lighting variations that might distract the mice. Training sessions are set for 60 s each. Each mouse underwent training in all four quadrants twice a d, with 15–20 min intervals between sessions, for five consecutive d. On the sixth d, the test was conducted by removing the platform and starting a 60-s probe trial. The mice were placed into the pool from the first quadrant, and the number of times they crossed the area where the platform was previously located (target quadrant) was recorded as an indicator of spatial memory. The observed parameters included the escape latency (seconds) from d 1–5 and the number of times the mice crossed the platform area on the sixth day.

Tissue sample extraction

Six mice from each group were randomly selected and anesthetized with 0.3% pentobarbital sodium (30 mg/kg). After decapitation, the brain tissues were extracted, placed in liquid nitrogen for preservation, and subsequently transferred to a −80°C environment.

qPCR detection of hub gene expression

Total RNA was extracted from brain tissue via the TRIzol method, and the RNA concentration was measured. The RNA was then reverse-transcribed into cDNA according to the manufacturer's instructions. qPCR was performed on an Applied Biosystems PCR 7500 instrument (Applied Biosystems, CHN) with ChamQ SYBR qPCR Master Mix (Vazyme, CHN). The mRNA levels of each gene were normalized to those of B-tubulin, and data analysis was conducted via the 2^−ΔΔCt method. The primer sequences are listed in Table 1.

Table 1.

The primer sequence information.

Gene	Primer	Sequence (5′-3′)	PCR Products
Mus GAPDH	Forward	GAGAGTGTTTCCTCGTCCCGTA	290 bp
Mus GAPDH	Reverse	CCTCACCCCATTTGATGTTAGT	290 bp
Mus ccdc88b	Forward	TGCGTCATTTAGTCCAGGGG	140 bp
Mus ccdc88b	Reverse	CTTCATGCTCCGAGAGGGTC	140 bp
Mus Dlat	Forward	ACTCGTCTTGGATGGACACAG	293 bp
Mus Dlat	Reverse	AGCACCGATTGCCAGAATAC	293 bp

Statistical methods

The bioinformatics data were analyzed and visualized via R software version 4.3.1, whereas the animal experimental data were statistically analyzed via SPSS software version 25.0. All the data are presented as the means ± standard deviations (x̄ ± s). Statistical analysis was performed via the t-test, with a P-value of <0.05 indicating statistical significance.

Results

Construction of a co-expression network in AD

WGCNA was used to analyze the expression values of 18,849 genes in 90 samples (46 control samples and 44 AD samples). The soft-thresholding power was 8, which was determined on the basis of the scale-free fit index and average connectivity (Figure 1A). A TOM plot depicting the 400 random genes within the clustering dendrogram is shown in Figure 1B. With the minimum module size set to 50, hierarchical clustering and module combination clustered genes into 8 modules (Figure 1C). The magenta (r = 0.23, p = 0.03) and gray (r = 0.36, p = 6e-04) modules exhibited significant positive correlations with AD (Figure 1D). As the gray module represents unclassified genes, only the magenta module was used for subsequent analysis.

Figure 1.

Identification of AD-related modules. (A) Graphs of scale-free topology, scale independence, mean connectivity, and the appropriate soft power were 8. (B) TOM plot depicting the 400 random genes. (C) Cluster dendrogram of the co-expression network modules (1-TOM). (D) Cluster plot of the relationships between modules and disease status. (TOM, topological overlap matrix)

Functional analysis of AD-related modules

The 1993 genes in the magenta module were functionally enriched (Supplemental Table 1). The molecular functions of catalytic activity, histone binding, and methyltransferase activity were highly enriched in the module (Figure 2A). GO analysis revealed that histone modification, RNA splicing, and methylation were enriched in AD-related modules (Figure 2B). KEGG analysis indicated that genes in the AD-related module were markedly enriched in metabolic pathways and the mTOR signaling pathway.

Figure 2.

Functional analysis of genes related to the AD module. (A, B) GO bar charts: the horizontal axis represents the proportion of magenta module genes under each term, and the color of the items represents the range of corrected P values; MF: Molecular Function; BP: Biological Process. (C) KEGG bar chart: the horizontal axis represents the proportion of magenta module genes under each term, and the color of the items represents the range of corrected p values. KEGG: Kyoto Encyclopedia of Genes and Genomes.

Construction of a protein-protein interaction network for magenta module genes

The relationships between the expression levels of magenta module genes and clinical traits were calculated via Pearson correlation coefficients. A scatter plot of GS versus module MM was drawn (Figure 3A). The scatter plot shows a significant correlation between MM and GS in the magenta module, indicating that these highly trait-related genes play crucial roles within the module. Using the criteria of GS > 0.2 and MM > 0.9, 15 genes (ECHDC1, L3MBTL2, BIRC6, MRPS7, ZBTB1, STARDB, DLAT, CCDC88B, PPID, FOXG1, FAM104B, TMEM186, MIB2, GPIHBP1, and GGNBP2) were identified as AD-related genes. A PPI network was then constructed (Figure 3B). We found that these 15 AD-related genes were highly enriched in mitochondria and ribosomes (Figure 3C).

Figure 3.

Construction of a protein-protein interaction network for magenta module genes. (A) Scatter plot analysis of the magenta module. Hub genes were identified in the upper right region with GS > 0.2 and MM > 0.9. (B) Protein-protein interaction network. (C) Functional analysis of the PPI network.

Comparison of AD-related gene expression profiles

The expression levels of 15 AD-related genes were compared between the AD group and the control group. The results revealed significant differences in the expression levels of DLAT and CCDC88b between the AD group and the control group(*p < 0.05). Specifically, DLAT was downregulated in the AD group, whereas CCDC88b exhibited higher expression levels in the AD group (Figure 4).

Figure 4.

Comparison of the expression levels of 15 genes between AD and normal samples (*p < 0.05; NS: not significant).

Identification of AD biomarkers via machine learning algorithms

Given the differences in feature selection mechanisms among various machine learning algorithms, this study comprehensively applied LASSO, random forest, and SVM-RFE methods to achieve multidimensional validation of the screening results, enhancing the robustness and reliability of the selected core genes. Using the GSE122063 dataset, we applied these three machine-learning methods to screen for hub genes associated with AD from 15 genes. LASSO regression analysis selected 7 predictive genes from statistically significant univariate features (Figure 5A and B); SVM-RFE analysis identified 7 feature genes (Figure 5C and D); random forest, combined with feature selection, was used to determine the relationships among the error rate, number of classification trees (Figure 5E and F), and 15 genes with relative importance. A Venn diagram was used to identify 2 intersecting genes (DLAT, CCDC88b) as hub genes from the intersection of the three methods (Figure 5G).

Figure 5.

Feature gene selection. (A, B) Adjusted feature selection in the least absolute shrinkage and selection operator (LASSO) model. (C, D) Validation of biomarker feature gene expression through the support vector machine recursive feature elimination (SVM-RFE) algorithm. (D) Relationship between the random forest error rate and the number of classification trees. (E) Relative importance ranking of the 15 selected genes. (F) Venn diagram showing the overlap of genes selected by the three algorithms.

Evaluation and validation of predictive model effectiveness

We constructed a nomogram based on two hub genes via the “rms” package (Figure 6A). ROC curves were used to evaluate the diagnostic specificity and sensitivity of each gene as well as the nomogram itself. The AUC for each gene was calculated. The results were as follows: for the GSE159699 validation set, the AUC was 0.612 (Figure 6B), and the AUCs for DLAT and CCDC88b were 0.619 and 0.538, respectively (Figure 6C and D). Therefore, these findings suggest that all genes present substantial diagnostic value for AD, with the constructed nomogram demonstrating the highest diagnostic efficiency. Additionally, we confirmed significant differences in the expression of DLAT and CCDC88b between the AD and control groups in both datasets (GSE122063 and GSE159699) (Figure 6D and F). These findings indicate that DLAT and CCDC88b are strongly associated with the diagnosis of AD and may be related to the progression of AD (p < 0.05).

Figure 6.

Construction and validation of the diagnostic nomogram model. (A) Visualization of the nomogram for AD diagnosis. (B-D) The ROC curve for each gene (DLAT and CCDC88b) and the nomogram in both the training and validation sets show its significant diagnostic value for AD. (E, F) Box plots showing the differences in the expression of the feature genes DLAT and CCDC88b in the GSE122063 and GSE159699 datasets (p < 0.05). ROC: receiver operating characteristic.

Correlations between hub genes and immune cell infiltration in AD

To explore immune infiltration in AD, we used xCell to estimate the relative abundances of 64 types of immune and nonimmune cells (such as hematopoietic progenitor cells, epithelial cells, and extracellular matrix cells). As shown in Figure 7A, macrophages were the most abundant immune cells in AD patients. Additionally, we analyzed the relationship between the proportion of immune cells and the expression levels of the hub genes (Figure 7B). CCDC88b was highly positively correlated with Th1 cells, CD8⁺ Tem cells, CD8⁺ T cells, CD8⁺ naïve T cells, GMP cells, iDCs, and the overall immune score. In contrast, DLAT scores were significantly negatively correlated with monocyte, pre-B cell, pDC, and overall immune scores but significantly positively correlated with CD4⁺ T cell, CD8⁺ Tem, and CLPs.

Figure 7.

Immune infiltration in AD and its correlation with hub genes. (A) Proportion of 64 cell types in AD samples. (B) Correlation analysis between hub gene expression and immune cell abundance.

Experimental validation of hub genes expression via qPCR

Behavioral testing of mouse cognitive ability. The results of the Morris water maze navigation test (Figure 8A and B) revealed that the escape latency of the AD experimental group was significantly greater than that of the wild-type control group. The results of the Morris water maze spatial exploration test indicated (Figure 8C) that the number of times the AD group mice reached the previous platform location and the time they spent on the platform was significantly lower than those of the wild-type control group (**p < 0.01). The Y-maze test results (Figure 8D) revealed that the AD model group mice spontaneously entered the novel arm fewer times than the wild-type group mice did. The cognitive function of the model group mice significantly decreased in both the Morris water maze and Y maze tests (**p < 0.01), indicating the successful establishment of the AD mouse model.

Figure 8.

(A) Representative tracking paths of the control and treatment groups in the Morris water maze test on day 4. (B) Escape latency (seconds) during the Morris water maze test. (C) Number of platform crossings on day 5; the T group exhibited a significant increase (**p < 0.01). (E) The relative expression level of the DLAT gene significantly decreased in the T group (**p < 0.01). (F) The relative expression level of the CCDC88b gene was significantly increased in the T group (**p < 0.01). (x̄ ± SD, n = 6; C = control group, T = experimental group).

qPCR validation results. The mRNA expression of the hub genes (DLAT and CCDC88b) was detected via qPCR (Figure 8E and F). Experimental validation revealed that, compared with that in the control group, the mRNA expression of DLAT was significantly lower (**p < 0.01), whereas the mRNA expression of CCDC88b was significantly greater (**p < 0.01) in the AD group. This finding is consistent with our previous findings.

Discussion

Research into the mechanisms of AD is ongoing, and bioinformatics technologies have proven to be excellent tools for integrating disease research with informatics. By constructing a PPI network combined with machine learning algorithms and clinical diagnostic models, we identified two AD hub genes, DLAT and CCDC88b.

To reveal the complexity of the microenvironment in AD, we assessed the levels of immune cell infiltration across different groups. In the immune infiltration analysis, macrophages were identified as the most abundant immune cells. Macrophages are generally categorized into pro-inflammatory and anti-inflammatory types. A persistent immune response in the brain is considered a third core pathological feature of AD. Long-term activation of microglia and macrophages may lead to neuroinflammation, which further exacerbates neuronal damage.^41,42 The increased infiltration of macrophages is part of the AD immune landscape and is likely related to their pro-inflammatory role. Immune infiltration analysis revealed significant differences in immune cell infiltration associated with DLAT and CCDC88b, indicating that these genes play crucial roles in the development of AD.

Areas of reduced glucose metabolism have been observed in the brain scans of patients with AD and other forms of dementia. Currently, increasing evidence suggests that decreased glucose metabolism may contribute to the progression of AD.³⁵ Researchers used the Alzheimer's Disease Neuroimaging Initiative database to compare brain glucose metabolism between AD patients and healthy controls. The results revealed that the local glucose metabolism rates in multiple brain regions of AD patients were significantly lower than those in healthy individuals.^36,37 DLAT primarily participates in glucose metabolism within the mitochondria, providing a crucial basis for intracellular energy metabolism.⁴³ DLAT is also a key component of the mitochondrial pyruvate dehydrogenase complex (PDHc), which encodes the E2 subunit. PDHc is a mitochondrial enzyme complex and one of the critical enzymes in the mitochondrial citric acid cycle, promoting the conversion of pyruvate to acetyl-CoA, which then enters the TCA cycle.⁴⁴ Dysregulated expression of DLAT, a critical enzyme in mitochondria, can impair PDHc activity and metabolic reprogramming, leading to mitochondrial dysfunction.⁴⁵ This dysfunction not only affects the intracellular energy balance but also may lead to metabolic diseases or neurological disorders.^46,47 Neurons require normal mitochondrial function to maintain homeostasis and proper biological activity. Neuronal cells require proper mitochondrial function to maintain homeostasis and normal biological activities within neurons. Typically, the brain relies on glucose for energy, but in a fasting state, it depends on ketone bodies,⁴⁸ indicating that when the energy supply is compromised, the brain is affected first, as it relies primarily on glucose and ketone bodies as energy sources.⁴⁹ During the progression of AD, mitochondrial dysfunction reduces ATP production, thereby affecting neuronal function and intracellular homeostasis.⁵⁰

Furthermore, mitochondrial dysfunction can lead to excessive production of ROS, reducing the cell's antioxidant defence capacity, triggering intracellular inflammatory responses, and causing oxidative stress.⁵¹ The complex neural network interactions between the olfactory cortex, hippocampus, temporal cortex, and frontal cortex facilitate the formation and storage of memory. Together, they construct memory circuits, allowing us to perceive environmental details in daily life and transform them into episodic memories, playing a crucial role in cognitive function.⁵² Studies have shown that DLAT expression is significantly reduced in the cortical tissues of AD patients and is consistently downregulated in multiple brain regions, including the olfactory cortex, hippocampus, temporal cortex, and frontal cortex.^53,54 In summary, low DLAT expression may disrupt cellular metabolism, leading to mitochondrial dysfunction and excessive ROS production, which in turn causes oxidative stress and results in cell damage or death, thereby contributing to the onset of AD. Research on DLAT has focused mainly on tumor cell metabolic pathways,⁵⁵ with few studies focused on its role in AD. This study revealed that DLAT is significantly downregulated in AD. More studies could help elucidate the exact role of DLAT in AD pathogenesis and provide a new theoretical foundation for future therapeutic strategies, which may include interventions targeting DLAT to regulate mitochondrial function and metabolic balance.⁵⁶

CCDC88b is expressed in brain tissues and various immune cells, particularly CD4⁺ and CD8⁺ T lymphocytes and myeloid cells. The human CCDC88b gene maps to the 11q13 locus, which is associated with susceptibility to inflammatory and autoimmune diseases.⁵⁷ For example, during intestinal injury, the expression of CCDC88b in the colon increases, accompanied by an influx of CCDC88b⁺ lymphocytes and myeloid cells, contributing to the occurrence of ulcerative colitis.⁵⁸ Studies have shown that inactivation of CCDC88b can prevent neuroinflammation.⁵⁹ Given the association of CCDC88b with susceptibility to inflammation, we hypothesize that its expression may play a positive role in neuroinflammation. A team utilizing data from the Alzheimer's Disease Sequencing Project reported a significant association between CCDC88b and AD,⁶⁰ which is consistent with our findings. Overall, CCDC88b may be closely involved in the pathogenesis of AD. CCDC88b is expressed in brain tissues and various immune cells, particularly CD4⁺ and CD8⁺ T lymphocytes and myeloid cells. Genome-wide screening of CCDC88bMut mice revealed significant impacts on T lymphocyte function, including impaired maturation in vivo, markedly reduced activation, decreased cell division, and impaired cytokine production (IFN-γ and TNF), indicating that CCDC88b is a critical regulator of T-cell function.⁵⁷ T cells are essential immune cells in the adaptive immune system, yet they often exhibit functional abnormalities in AD. Abnormal T cells can indirectly promote neuroinflammation by secreting pro-inflammatory mediators, thereby participating in the pathological process of AD.⁶¹ Our immune infiltration results also suggested a strong positive correlation between CCDC88b expression and T-cell infiltration.

In summary, high expression of CCDC88b may exacerbate neuroinflammation by regulating T-cell function, thereby contributing to AD pathology. In addition to regulating T cells, animal experiments have shown that CCDC88b can form a complex with RASAL3 and ARHGEF2 in T cells, regulating DC migration by controlling RHOA activation. This finding demonstrates the importance of CCDC88b in regulating the movement and migration of dendritic cells (DCs), with its inactivation leading to migratory defects in mouse DCs.^59,62 Recent evidence suggests that the peripheral immune system plays an important role in regulating brain inflammation, particularly DCs, whose altered numbers and functions may exacerbate disease progression.⁶³ A decrease in the number of myeloid DCs in the peripheral blood of AD patients is also closely related to AD progression.⁶⁴ The number, function, location, and overactivation of DCs in cerebrospinal fluid can contribute to a pro-inflammatory state in the brain through their intrinsic properties, which can activate pathogenic T cells or release pro-inflammatory cytokines, exacerbating inflammation in the brain. This inflammatory environment can directly or indirectly damage neurons, thereby promoting the progression of various symptoms.^65,66 DCs can regulate brain inflammation by playing a critical role in immune responses. In conclusion, we infer that high expression of CCDC88b may aggravate neuroinflammation by regulating T-cell function and mediating DC migration, contributing to AD development. The relationship between CCDC88b and AD is complex and worthy of further investigation, with significant implications for understanding and treating this disease.

In summary, the two core genes we identified (DLAT and CCDC88b) will broaden our understanding of the molecular mechanisms involved and provide additional potential diagnostic and therapeutic targets for clinical treatment. Although we have already validated the expression of the hub genes in animal experiments, the inherent limitations of bioinformatics techniques still necessitate the use of human samples for further experimental validation of our findings. Additionally, since our data were derived from databases, common covariates such as age, sex, ethnicity, and comorbidities were not considered. However, further clinical studies and higher levels of evidence are still needed.

Conclusion

The identification of the two core genes (DLAT and CCDC88b) expands our understanding of the molecular mechanisms underlying AD and provides potential diagnostic and therapeutic targets for clinical treatment. Although we confirmed the expression of these hub genes in animal experiments, the specific mechanisms by which these genes regulate AD remain unclear. Therefore, our next step is to conduct in vitro cell experiments to elucidate their roles in regulating AD.

Supplemental Material

sj-xlsx-1-alr-10.1177_25424823251356300 - Supplemental material for Identification and experimental validation of Alzheimer's disease hub genes via bioinformatics and machine learning

Supplemental material, sj-xlsx-1-alr-10.1177_25424823251356300 for Identification and experimental validation of Alzheimer's disease hub genes via bioinformatics and machine learning by Ying Hu, Zhaoshuyu Pan, Jianping Li, Bin Tang, Mingbo Luo, Yu Li, Xue Cao, Kaiwen Zheng, Nana Wang and Chuanjie Xu in Journal of Alzheimer's Disease Reports

Footnotes

ORCID iD

Ying Hu

Ethical considerations

The animal study protocol was approved by the Ethics Committee of Zunyi Medical University (ZMU21-2201-167).

Author contributions

Ying Hu: Conceptualization; Data curation; Formal analysis; Methodology; Resources; Software; Validation; Visualization; Writing – original draft; Writing – review & editing.

Zhaoshuyu Pan: Conceptualization; Formal analysis; Supervision; Writing – original draft.

Jianping Li: Conceptualization; Methodology; Resources; Writing – review & editing.

Bin Tang: Conceptualization; Formal analysis; Project administration; Writing – original draft.

Mingbo Luo: Conceptualization; Data curation; Formal analysis; Investigation; Software; Writing – original draft.

Yu Li: Conceptualization; Formal analysis; Funding acquisition; Resources; Software; Writing – original draft.

Xue Cao: Conceptualization; Funding acquisition; Investigation; Methodology; Writing – original draft.

Kaiwen Zheng: Conceptualization; Formal analysis; Investigation; Methodology; Writing – original draft.

Nana Wang: Conceptualization; Formal analysis; Investigation; Methodology; Writing – original draft.

Chuanjie Xu: Conceptualization; Formal analysis; Resources; Supervision; Visualization; Writing – original draft; Writing – review & editing.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability statement

All the data used in the present study are available from the GEO database (). The accession numbers are as follows: GSE174367, GSE122063, and GSE159699.

Supplemental material

Supplemental material for this article is available online.

References

Scheltens

De Strooper

Kivipelto

, et al. Alzheimer's disease. Lancet 2021; 397: 1577–1590.

Blömeke

Rehn

Kraemer-Schulien

, et al. Aβ oligomers peak in early clinical stages of Alzheimer's disease and proceed Tau pathology. Alzheimers Dement 2023; 19: e082907.

Galizzi

Di Carlo

. Mitochondrial DNA and inflammation in Alzheimer’s disease. Curr Issues Mol Biol 2023; 45: 8586–8606.

Ashleigh

Swerdlow

Beal

. The role of mitochondrial dysfunction in Alzheimer's disease pathogenesis. Alzheimers Dement 2023; 19: 333–342.

Heneka

Carson

Khoury

, et al. Neuroinflammation in Alzheimer's disease. Lancet Neurol 2015; 14: 388–405.

Wang

, et al. GSDMD In peripheral myeloid cells regulates microglial immune training and neuroinflammation in Parkinson's disease. Acta Pharm Sin B 2023; 13: 2663–2679.

Jorfi

Park

Hall

, et al. Infiltrating CD8+ T cells exacerbate Alzheimer’s disease pathology in a 3D human neuroimmune axis model. Nat Neurosci 2023; 26: 1489–1504.

Andrews

. Alzheimer's disease is causally associated with mitochondrial dysfunction. Alzheimers Dement 2021; 17: e057480.

Murali Mahadevan

Hashemiaghdam

Ashrafi

, et al. Mitochondria in neuronal health: from energy metabolism to Parkinson's disease. Adv Biol 2021; 5: 2100663.

10.

Bustamante-Barrientos

Luque-Campos

Araya

, et al. Mitochondrial dysfunction in neurodegenerative disorders: potential therapeutic application of mitochondrial transfer to central nervous system-residing cells. J Transl Med 2023; 21: 613.

11.

Trovato

Pennisi

Crupi

, et al. Neuroinflammation and mitochondrial dysfunction in the pathogenesis of Alzheimer's disease: modulation by coriolus versicolor (Yun-Zhi) nutritional mushroom. J Neurol Neuromedicine 2017; 2: 19–28.

12.

Sang

Philbert

Hartland

, et al. Coenzyme A-dependent tricarboxylic acid cycle enzymes are decreased in Alzheimer's disease consistent with cerebral pantothenate deficiency. Front Aging Neurosci 2022; 14: 893159.

13.

The Lancet Healthy Longevity. Hope on the horizon for Alzheimer's disease treatment? Lancet Healthy Longev 2022; 3: e722.

14.

Ren

Lin

, et al. The China Alzheimer Report 2022. Gen Psychiatr 2022; 35: e100751.

15.

Burns

Bernabei

Bullock

, et al. Safety and efficacy of galantamine (reminyl) in severe Alzheimer's disease (the SERAD study): a randomised, placebo-controlled, double-blind trial. Lancet Neurol 2009; 8: 39–47.

16.

Jia

Wei

Chen

, et al. The cost of Alzheimer's disease in China and re-estimation of costs worldwide. Alzheimers Dement 2018; 14: 483–491.

17.

Kwon

Koh

. Neuroinflammation in neurodegenerative disorders: the roles of microglia and astrocytes. Transl Neurodegener 2020; 9: 42.

18.

Hong

Beja-Glasser

Nfonoyim

, et al. Complement and microglia mediate early synapse loss in Alzheimer mouse models. Science 2016; 352: 712–716.

19.

Asby

Boche

Allan

, et al. Systemic infection exacerbates cerebrovascular dysfunction in Alzheimer's disease. Brain 2021; 144: 1869–1883.

20.

Budd Haeberlein

Aisen

Barkhof

, et al. Two randomized phase 3 studies of Aducanumab in early Alzheimer's disease. J Prev Alzheimers Dis 2022; 9: 197–210.

21.

van Dyck

Swanson

Aisen

, et al. Lecanemab in early Alzheimer's disease. N Engl J Med 2023; 388: 9–21.

22.

Mintun

Duggan Evans

, et al. Donanemab in early Alzheimer's disease. N Engl J Med 2021; 384: 1691–1704.

23.

Wang

, et al. Identification of potential feature genes in CRSwNP using bioinformatics analysis and machine learning strategies. J Inflamm Res 2024; 17: 7573–7590.

24.

Zhao

Duan

Cui

, et al. Exploration of biomarkers for systemic lupus erythematosus by machine-learning analysis. BMC Immunol 2023; 24: 44.

25.

Morabito

Miyoshi

Michael

, et al. Single-nucleus chromatin accessibility and transcriptomic characterization of Alzheimer's disease. Nat Genet 2021; 53: 1143–1155.

26.

McKay

Beck

Khoo

, et al. Peri-infarct upregulation of the oxytocin receptor in vascular dementia. J Neuropathol Exp Neurol 2019; 78: 436–452.

27.

Nativio

Lan

Donahue

, et al. An integrated multi-omics approach identifies epigenetic alterations associated with Alzheimer's disease. Nat Genet 2020; 52: 1024–1035.

28.

Langfelder

Horvath

. WGCNA: an R package for weighted correlation network analysis. BMC Bioinformatics 2008; 9: 559.

29.

Zhang

Horvath

. A general framework for weighted gene co-expression network analysis. Stat Appl Genet Mol Biol 2005; 4: Article17.

30.

Luo

Huo

, et al. KOBAS-i: intelligent prioritization and exploratory visualization of biological functions for gene enrichment analysis. Nucleic Acids Res 2021; 49: W317–W325.

31.

, et al. Clusterprofiler 4.0: a universal enrichment tool for interpreting omics data. Innovation (Camb) 2021; 2: 100141.

32.

Wang

Han

, et al. Clusterprofiler: an R package for comparing biological themes among gene clusters. OMICS 2012; 16: 284–287.

33.

Montojo

Zuberi

Rodriguez

, et al. GeneMANIA: fast gene network construction and function prediction for cytoscape [version 1; peer review: 2 approved]. F1000Res 2014; 3: 153.

34.

Bose

Healy

Lokhande

, et al. Early predictors of clinical and MRI outcomes using Least Absolute Shrinkage and Selection Operator (LASSO) in multiple sclerosis. Ann Neurol 2022; 92: 87–96.

35.

Szymczak

. A review on longitudinal data analysis with random forest. Brief Bioinform 2023; 24: bbad002.

36.

Kumari

Akhtar

Shah

, et al. Support matrix machine: a review. Neural Netw 2025; 181: 106767.

37.

Chen

Guo

, et al. Identification of diagnostic markers for moyamoya disease by combining bulk RNA-sequencing analysis and machine learning. Sci Rep 2024; 14: 5931.

38.

Aran

Butte

. Xcell: digitally portraying the tissue cellular heterogeneity landscape. Genome Biol 2017; 18: 220.

39.

Chang

Feng

Hao

, et al. Deciphering the molecular landscape of rheumatoid arthritis offers new insights into the stratified treatment for the condition. Front Immunol 2024; 15: 1391848.

40.

Tang

Wang

Xie

, et al. Classification of distinct tendinopathy subtypes for precision therapeutics. Nat Commun 2024; 15: 9460.

41.

Kinney

Bemiller

Murtishaw

, et al. Inflammation as a central mechanism in Alzheimer's disease. Alzheimers Dement 2018; 4: 575–590.

42.

Munawara

Catanzaro

, et al. Hyperactivation of monocytes and macrophages in MCI patients contributes to the progression of Alzheimer's disease. Immun Ageing 2021; 18: 29.

43.

Goh

Kuznetsov

, et al. DLAT Subunit of the pyruvate dehydrogenase complex is upregulated in gastric cancer-implications in cancer therapy. Am J Transl Res 2015; 7: 1140–1151.

44.

Han

Chua

Baqai

, et al. Pyruvate dehydrogenase inactivation causes glycolytic phenotype in BAP1 mutant uveal melanoma. Oncogene 2022; 41: 1129–1139.

45.

Bhatti

Kaur

Mishra

, et al. Targeting dynamin-related protein-1 as a potential therapeutic approach for mitochondrial dysfunction in Alzheimer's disease. Biochim Biophys Acta Mol Basis Dis 2023; 1869: 166798.

46.

Misrani

Tabassum

Yang

. Mitochondrial dysfunction and oxidative stress in Alzheimer’s disease. Front Aging Neurosci 2021; 13: 617588.

47.

Weidling

Swerdlow

. Mitochondrial dysfunction and stress responses in Alzheimer’s disease. Biology (Basel) 2019; 8: 39.

48.

Midha

Zhou

Queliconi

, et al. Organ-specific fuel rewiring in acute and chronic hypoxia redistributes glucose and fatty acid metabolism. Cell Metab 2023; 35: 504–516.e5.

49.

Mergenthaler

Lindauer

Dienel

, et al. Sugar for the brain: the role of glucose in physiological and pathological brain function. Trends Neurosci 2013; 36: 587–597.

50.

Moreira

Carvalho

Zhu

, et al. Mitochondrial dysfunction is a trigger of Alzheimer's disease pathophysiology. Biochim Biophys Acta 2010; 1802: 2–10.

51.

Pszczołowska

Walczak

Miśków

, et al. Mitochondrial disorders leading to Alzheimer’s disease—perspectives of diagnosis and treatment. Geroscience 2024; 46: 2977–2988.

52.

Aqrabawi

Kim

. Hippocampal projections to the anterior olfactory nucleus differentially convey spatiotemporal information during episodic odour memory. Nat Commun 2018; 9: 2735.

53.

Lai

Lin

, et al. Identification and immunological characterization of cuproptosis-related molecular clusters in Alzheimer's disease. Front Aging Neurosci 2022; 14: 932676.

54.

Jia

Wang

. Systemic alterations of tricarboxylic acid cycle enzymes in Alzheimer’s disease. Front Neurosci 2023; 17: 1206688.

55.

Zhang

Zhao

Yuan

, et al. DLAT Is a promising prognostic marker and therapeutic target for hepatocellular carcinoma: a comprehensive study based on public databases. Sci Rep 2023; 13: 17295.

56.

Eyigoz

Mathur

Santamaria

, et al. Linguistic markers predict onset of Alzheimer's disease. EClinicalMedicine 2020; 28: 100583.

57.

Kennedy

Fodil

Torre

, et al. CCDC88B Is a novel regulator of maturation and effector functions of T cells during pathological inflammation. J Exp Med 2014; 211: 2519–2535.

58.

Fodil

Moradin

Leung

, et al. CCDC88B Is required for pathogenesis of inflammatory bowel disease. Nat Commun 2017; 8: 932.

59.

Olivier

Langlais

Jeyakumar

, et al. CCDC88B Interacts with RASAL3 and ARHGEF2 and regulates dendritic cell function in neuroinflammation and colitis. Commun Biol 2024; 7: 77.

60.

Lee

English

Xia

, et al. Association analysis of common structural variants in the Alzheimer’s disease sequencing project. Alzheimers Dement 2023; 19: e078821.

61.

Dai

Shen

. Insights into T-cell dysfunction in Alzheimer's disease. Aging Cell 2021; 20: e13511.

62.

Jean-Frederic

Nassima

Sara Al

, et al. CCDC88B Is required for mobility and inflammatory functions of dendritic cells. J Leukoc Biol 2020; 108: 1787–1802.

63.

Brezovakova

Valachova

Hanes

, et al. Dendritic cells as an alternate approach for treatment of neurodegenerative disorders. Cell Mol Neurobiol 2018; 38: 1207–1214.

64.

Ciaramella

Salani

Bizzoni

, et al. Myeloid dendritic cells are decreased in peripheral blood of Alzheimer’s disease patients in association with disease progression and severity of depressive symptoms. J Neuroinflammation 2016; 13: 18.

65.

Constant

Maarifi

Blanchet

, et al. Role of dendritic cells in viral brain infections. Front Immunol 2022; 13: 862053.

66.

Han

Zhang

, et al. Understanding immune microenvironment alterations in the brain to improve the diagnosis and treatment of diverse brain diseases. Cell Commun Signal 2024; 22: 132.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.09 MB