Introduction

Alzheimer's Disease (AD) is characterized by the accumulation of amyloid-beta (Aβ) plaques and neurofibrillary tangles1. The current focus of scientific investigation revolves around unraveling the complex function of microglia, the immune cells residing in the central nervous system (CNS), in the development of AD2. Traditionally recognized as sentinels of neural health, microglia undergo activation in response to the presence of Aβ plaques and neurofibrillary tangles, attempting to mitigate or clear these pathological elements3. However, the precise dynamics governing the relationship between microglial activation and the progression of AD remain enigmatic, prompting intensive investigation4. Understanding the interplay between these immune cells and the neurodegenerative processes in AD holds promise for developing targeted therapeutic interventions in the ongoing pursuit of effective treatments for this debilitating condition.

This study employs bioinformatic techniques, including single-cell RNA sequencing (scRNA-seq), Weighted Gene Co-expression Network Analysis (WGCNA), and machine learning, to delve into the cellular-level pathological alterations associated with AD5. By meticulously analyzing publicly available AD and HC (healthy control) scRNA-seq databases, the study identifies distinct subclusters of microglia responding to AD-related changes initially. The subsequent exploration of critical genes, cell–cell communication6, and pseudotime analysis7 enhances our comprehension of the intricate interaction between microglia and neurodegenerative mechanisms in AD. Leveraging the insights from WGCNA, the study extended its investigation to a bulk-RNA seq database (GSE33000), where the least absolute shrinkage and selection operator (LASSO) and COX regression analysis were used to screen 9 hub genes8. Subsequently, AD disease subgroups were studied based on these genes using Non-negative Matrix Factorization (NMF)9. Finally, a Convolutional Neural Network (CNN) model10 was proposed and the identified hub genes were validated using PCR. This research not only enhances our comprehension of the molecular and cellular intricacies underlying AD, but also holds promise for unveiling targeted therapeutic strategies in the ongoing quest for effective treatments against this debilitating condition (Fig. 1A).

Figure 1
figure 1

AD-associated subclusters of microglia from scRNA-seq data. (A) Workflow of the experiment. (B–D) UMAP plot illustrates cell clusters identified from the entire dataset. (E, F), Identification of microglia subclusters in AD and HC. (G–I), Plot of pseudotime analysis of microglia. (J), Plot of cell communication between microglia and other cells. AD, Alzheimer's Disease; scRNA-seq, single-cell RNA sequencing; HC, healthy control; WGCNA, Weighted Gene Co-expression Network Analysis; LASSO, the least absolute shrinkage and selection operator; NMF, Non-negative Matrix Factorization; CNN, Convolutional Neural Network; UMAP, Uniform Manifold Approximation and Projection.

Full size image

Method

scRNA-seq and bulk-seq data acquisition and processing

The scRNA-seq database for AD was obtained from GSE188545 dataset in the GEO database (http://www.ncbi.nlm.nih.gov/). Our analytical focus encompassed 6 AD patients and 6 HCs. The scRNA-seq database was processed using the "Seurat" package, which involved filtering the data for cells and genes, normalization, principal component analysis (PCA), and Uniform Manifold Approximation and Projection (UMAP)11. Quality control measures excluded cells with mitochondrial gene expression exceeding 20%, less than 200 or more than 2500 detected genes. The annotation of cell types was accomplished through the identification of marker genes within different cell clusters. These clusters include oligodendrocytes (identified by ST18 and CTNNA3), astrocytes (identified by SLC1A2 and ADGRV1), oligodendrocyte progenitor cells (OPCs, identified by PCDH15 and LHFPL3), neurons (identified by ROBO2 and SYT1), microglia (identified by DOCK8 and CD74), endothelial cells (identified by ARHGAP29 and CEMIP) and unidentified cells12. All microglia were extracted and subsequently divided into 11 subclusters based on distinct cell marker genes. The bulk-RNA seq database for AD was obtained from two databases, GSE33000 and GSE132903. The GSE33000 comprised 97 AD samples and 98 HC samples, while GSE132903 included 310 AD samples and 157 HC samples. To minimize potential batch effects, the expression profiles were harmonized using the Combat function from the "sva" package13.

WGCNA

"WGCNA" package was performed to analyze entirety of genes identified within microglia from AD and HC samples5. Initially, a clustering step was implemented to evaluate the relevance of all samples in the dataset and eliminate any outliers. A soft thresholding power parameter was determined for constructing a scale-free network. Clustering identified modules, and correlation analysis with microglia subclusters revealed modules with high specificity to AD. Top 30 genes within each module were selected as critical genes in AD-related microglia.

Cell–cell interaction analysis and pseudotime analysis

The "CellChat 1.0.0" algorithm was employed to analyze cell–cell communication at the single-cell level involving AD-related microglia6. This algorithm was specifically designed for examining and quantifying intercellular communication. The study involved the examination, quantification, and visualization of the global intercellular communication network under AD conditions. "Mococle" R package was used for pseudotime analysis, determining the differentiation state of microglia subclusters7. Initially, all microglia subclusters were extracted, and subsequently, the "DDRTree" approach was applied to reduce dimensionality. The "orderCell" function was then employed to determine the cell differentiation state.

Critical genes in AD-related microglia enrichment analysis

The Metascape data resource was utilized for functional enrichment pathways of critical genes using the "clusterProfiler" package (http://metascape.org/gp/index.html)14. This integrative platform combines information from various pathway databases such as the gene oncology (GO) functional database, Kyoto Encylopedia of Genes and Genomes (KEGG), and Reactome, served as the foundation for the functional enrichment analysis.

LASSO and COX regression analysis

LASSO and COX regression analyses were conducted to identify hub genes in AD-associated microglia in the bulk RNA-seq8. The gene set variation analysis (GSVA) method was employed to combine the LASSO result15.

Immune infiltration

The Microenvironment Cell Populations (MCP)-counter algorithm was utilized to assess the level of immune cell infiltration in a specific tissue or disease. This computational method utilizes gene expression data to estimate the relative abundance of various immune cell types by analyzing the correlation between specific gene markers and immune cell populations. By employing this approach, the role of immune cells in specific diseases and potential immunopathological mechanisms can be elucidated16. The assessment of immune infiltration levels for each sample was evaluated utilizing the "MCP-counter" R package.

Identification of molecular subsets of AD with NMF method

Utilizing the NMF method on bulk RNA-seq from the AD group, molecular subgroups of AD were identified, effectively delineating their heterogeneity in complex diseases9. The NMF results were combined using the GSVA method. The analytical procedures were performed by utilizing the "NMF" package within the R software. The infiltration abundance of 10 immune cell species was assessed using MCP-counter.

BV2 microglia culture and treatment

The BV2 microglia cell line was acquired from the Cell Bank of the Institute of Biochemistry and Cell Biology (Shanghai, China). The BV2 cells were cultured in DMEM medium (Procell, China) supplemented with 10% FBS (Procell, China) and 1% penicillin–streptomycin (Solarbio Science & Technology, China). A humidified atmosphere containing 95% air and 5% CO2 was used to incubate the cells at 37 °C. Initially, the BV2 cells were seeded in a 6-well plate and left to expand overnight until they reached a confluency of 70%. Subsequently, cells were serum-starved for 6 h on serum-free medium. Following the starvation period, BV2 cells were stimulated with LPS (5 µug/ml) for 12 h to induce an inflammation model. Changes in RNA levels were verified using qPCR.

Convolutional neural network

A CNN-based model was developed for AD risk prediction. The training set comprised bulk RNA-seq data (GSE33000), and the testing set comprised data from GSE13290310. The CNN model incorporated diverse features, including hub genes expressions, immune infiltration scores, and sample quantity. In order to ensure the best performance during training, we try different training epochs to get the best training model. Training epochs were conducted using "TensorFlow-1.14" in the R programming language.

Quantitative qPCR

Total RNA was isolated from newly collected cell samples using Trizol Reagent (Takara, Japan) and converted into cDNA according to the instructions provided by the manufacturer of Prime Script TM RT Master Mix (Takara, Japan). Primers for USP3, NLRP3, IL-1β, and ACTB (utilized as an internal control) were custom-designed and synthesized by Sangon Biotech (Shanghai, China).

  • The forward primer of USP3: 5′ GCCTGGACATGAAGTGCTACCTAC3′,

  • the reverse primer of USP3: 5′ CCAACGGTCTCCTCATCAGTCAC3′.

  • The forward primer of NLRP3: 5′ GCTGCGATCAACAGGCGAGAC 3′,

  • the reverse primer of NLRP3: 5′ CCATCCACTCTTCTTCAAGGCTGTC 3′.

  • The forward primer of IL-1β: 5′ TCGCAGCAGCACATCAACAAGAG 3′,

  • the reverse primer of IL-1β: 5′ AGGTCCACGGGAAAGACACAGG 3′.

  • The forward primer of ACTB: 5′GTGCTATGTTGCTCTAGACTTCG3′

  • the reverse primer of ACTB: 5′ATGCCACAGGATTCCATACC3′.

The cycling conditions for quantitative PCR were adjusted to a temperature of 95 °C for a duration of 30 s, followed by 45 cycles consisting of 5 s at 95 °C and 34 s at 60 °C. In order to analyze the comparative Ct values for gene levels of USP3, NLRP3, IL-1β and ACTB, the ABI QuantStudio3 system (Applied Biosystems, USA) was employed. For the purpose of evaluating the mRNA's relative expression, the equation  2(−△△Ct) was utilized.

Statistical analysis

The statistical analysis was carried out with R Studio (version 4.1.2). For variables that followed a normal distribution, the independent t-test was utilized, while the non-normally distributed variables were subject to the Wilcoxon test. Significance was considered at a level of P < 0.05. The software employed for the statistical analysis was GraphPad Prism 9.0.

Results

Identified altered subclusters in microglia in AD

We conducted a comprehensive analysis of scRNA-seq data (GSE188545) obtained from 6 AD patients and 6 HCs. Following gene expression filtering and normalization, we identified 7 cell clusters through principal component analysis and graph-based clustering (Fig. 1B–D). With a specific focus on microglial alterations, our analysis further delved into the microglial landscape, identifying 11 subclusters (Fig. 1F)17. Notably, 4 of these subclusters (1, 2, 6, and 10) exhibited significant changes in the context of AD, illustrating the intricate heterogeneity of microglial responses within this neurodegenerative disorder (Fig. 1E). A pseudotime analysis elucidated a unique differentiation pattern in AD-related microglia compared to their healthy counterparts, shedding light on the temporal dynamics of these alterations (Fig. 1G–I).

Cell–cell communication between microglia and other neuronal cells

The utilization of the "CellChat" algorithm allowed us to unravel the complex cell communication network of microglia associated with AD. Our findings underscored the intricate interactions between AD microglia and various cell types in the central nervous system (Fig. 1J). The communication strength and frequency observed between AD microglia and other microglia emphasized the pivotal role of these immune cells in the intricate cross-talk within the neural environment. Additionally, we observed notable communication patterns with astrocytes, neurons, oligodendrocytes, endothelial cells and OPCs, providing insights into the broader impact of microglial responses in AD.

WGCNA identification of critical genes in AD-related microglia

To identify critical genes associated with AD in microglia, we utilized the "WGCNA" package in R software to construct a weighted co-expression network encompassing all detected genes in microglia in both AD and HC groups in scRNA-seq (Fig. 2A). Our analysis led to identify 4 modules (blue, purple, red, and brown) with high specificity to AD (Fig. 2B,C). From each module, we selected the top 30 genes to form a set specific to AD-associated microglia, hereby designated as critical genes in AD-related microglia. Subsequently, we conducted enrichment analysis for pathway and functional annotations using the Metascape online platform (Fig. 2G). Processes such as peptide chain elongation, synapse-related functions, and cell adhesion emerged as significantly affected pathways, highlighting the multifaceted involvement of microglia in AD pathology. Among these, peptide chain elongation held the top rank, followed by protein–protein interactions at synapses, positive regulation of cell junction assembly, cell–cell adhesion, and cell adhesion molecules as the second to fifth ranked processes, respectively. Additionally, a subset of genes was enriched in the construction of the actin cytoskeleton, neuron recognition, and antigen presentation and processing.

Figure 2
figure 2

Identifying hub genes in AD-related microglia. (A), Plot of WGCNA Module Distribution. (B), Feature Plots of different module scores in microglia. (C), The correlations between modules and microglia subclusters. (D,E), Identification of hub genes using LASSO and COX in bulk RNA-seq. (F), GSVA scores of hub genes in the AD group and HC group in bulk RNA-seq. (G), Functional enrichment analysis of critical genes using Metascape. (H), Correlation analysis of immune infiltration in hub genes. AD, Alzheimer's Disease; WGCNA, Weighted Gene Co-expression Network Analysis; LASSO, the least absolute shrinkage and selection operator; HC, healthy control; GSVA, gene set variation analysis.

Full size image

Identification of biomarkers in AD-associated microglia

Employing LASSO and COX regression analysis in the bulk RNA-seq data (GSE33000), we identified 9 hub genes associated with AD microglia. These hub genes comprised HLA-B, RPS15, TPT1, ARHGAP22, NRXN3, RPL41, ANKS1B, USP3, and PDE4D (Fig. 2D,E). The GSVA scores for 9 hub genes are represented in Fig. 2F. We performed immune cell correlation analysis using MCP-counter. The results demonstrated the correlation of hub genes with monocytes, endothelial cells, T cells, CD8+ T cells, cytotoxic lymphocytes, B cells, myeloid dendritic cells, and neutrophils (Fig. 2H). Specifically, there was a notable positive correlation observed between neutrophils and HLA-B, RSP15, TPT1 and USP3 (P < 0.001), while a significant negative correlation was found with ARHGAP22 (P < 0.01). T cells, on the other hand, exhibited a significant positive correlation with ARHGAP22 (P < 0.01) and ANKSVB (P < 0.01), alongside a significant negative correlation with RPL41 (P < 0.001) and NRXN3 (P < 0.01). Correlation analyses revealed intricate associations between these hub genes and various immune cell types, providing a deeper understanding of the immune landscape within AD microglia.

Differential immune microenvironment across subgroups of AD

The application of NMF method allowed us to identify 3 molecular subgroups of AD, each displaying distinct patterns of immune cell infiltration in bulk RNA-seq (Fig. 3A). GSVA was employed to determine the relative expression levels of the hub genes within subgroup 1, 2, and 3 (Fig. 3B). Subgroup 1 exhibited the highest expression levels and subgroup 3 displayed the lowest. Gene heatmaps revealed that subgroup 1 displayed increased expression of PLXNC1, ATG7, and NRXN1, and decreased expression of ROCK1, while subgroup 3 exhibited significantly increased expression of EEF1A1 and decreased expression of ROCK1, PLXNC1, and NRXN1 (Fig. 3C). Subgroup analysis unveiled variations in the levels of infiltrating immune cells, emphasizing the heterogeneity within AD. Conducting immune infiltration analysis using MCP-counter on the three subgroups revealed high infiltration of cytotoxic lymphocytes in subgroup 1, with low infiltration of monocytes and endothelial cells (Fig. 3D). Within the subgroup 3, there was a notable infiltration of neutrophils and endothelial cells, whereas the level of infiltration from cytotoxic lymphocytes and CD8+ T cells remained comparatively low. GSVA scoring of immune infiltration among the three subgroups demonstrated significant differences (P < 2.2e-16) in the types and levels of immune cell infiltration. Notably, subgroup 1 exhibited significantly higher cytotoxic lymphocyte infiltration compared to the other two subgroups (P < 0.001), and monocytes and endothelial cells displayed significantly lower levels than those in subgroup 2 and 3 (P < 0.001) (Fig. 3E). Additionally, monocytes (P < 0.001), neutrophils (P < 0.01), and endothelial cells (P < 0.001) in subgroup 3 were significantly higher than those in the other two subgroups (Fig. 3E). Significant differences in immune cell infiltration levels among the subgroups provided further insights into the nuanced immune microenvironment associated with different molecular subgroups of AD.

Figure 3
figure 3

9 hub genes identified as significant biomarkers for AD. (A), The optimal rank was 3 as the cophenetic coefficient started firstly decreasing. (B), GSVA scores of 9 hub genes in the 3 NMF subgroups. (C), Heat map of genes across 3 subgroups. (D), Heat map of immune infiltration in 3 subgroups. (E), GSVA scores of immune infiltration in 3 subgroups. (F–N), Box plot analysis of 9 hub genes expression in AD and HC groups in bulk RNA-seq. (F), USP3. (G), ANKS1B. (H), RPL41. (I), PDE4D. (J), TPT1. (K), ARHGAP22. (L), RPS15. (M), NRXN3. (N), HLA-B. AD, Alzheimer's Disease; NMF, Non-negative Matrix Factorization; HC, healthy control. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001.

Full size image

LPS enhanced USP3 expression in BV2 cells

Validation of the expression levels of 9 hub genes was conducted in bulk RNA-seq, and all of them exhibited significant differences (Fig. 3F–N). The violin plot highlighted the elevated expression of USP3, a standout hub gene, in microglia in AD group compared to the HC group in scRNA-seq (Fig. 4A). Furthermore, in an inflammation model induced by LPS in BV2 cells, USP3 expression increased, accompanied by elevated RNA levels of NLRP3 and IL-1β, implicating its role in neuroinflammation associated with AD (Fig. 4B–D).

Figure 4
figure 4

Validation of USP3 differential expression and construction of AD model using CNN. (A), Assessment of USP3 expression in microglia of AD and HC groups using violin plots in scRNA-seq. (B–D), PCR analysis of mRNA expression levels in BV2 and LPS groups. (B), USP3. (C), NLRP3. (D), IL-1β. (E), Accuracy Evaluation of CNN Model. (F), Training process of the AD CNN model in 3000 times. (G,H), Performance of the AD CNN model in the training and validation sets in 3000 times. (I), Training process of the AD CNN model in 5000 times. (J–K), Performance of the AD CNN model in the training and validation sets in 5000 times. (L), Training process of the AD CNN model in 10,000 times. (M,N), Performance of the AD CNN model in the training and validation sets in 10,000 times. AD, Alzheimer's Disease; CNN, Convolutional Neural Network; AUC, area under curve.

Full size image

Establishment of an AD convolutional neural network model

To enhance diagnostic accuracy, we developed a CNN model for AD risk prediction. This model incorporated diverse features, including the expression levels of the identified hub genes, immune infiltration scores, and sample size. The training set consisted of bulk RNA-seq data (GSE33000), and the testing set comprised bulk RNA-seq data (GSE132903). The CNN model was trained for 3000, 5000, and 10,000 epochs, respectively (Fig. 4E). Training for 3000 epochs resulted in an area under curve (AUC) of 0.997 in the training set and an AUC of 0.716 in the validation set (Fig. 4F–H). With 5000 epochs, the AUC values were 0.998 in the training set and 0.725 in the validation set (Fig. 4I–K). Extending the training to 10,000 epochs achieved AUC values of 1.000 in the training set and 0.692 in the validation set (Fig. 4L–N). The detailed training epochs analysis provided insights into the model's stability and potential for improving diagnostic precision in identifying AD patients.

Discussion

Microglia play a pivotal role in neuroinflammation and neurodegeneration in AD, serving as the first line of immune defense in the CNS18,19. Joint analysis of GWAS and single-nucleus sequencing has revealed that AD risk loci are closely associated with microglia, as indicated by functional enrichment20. This in-depth investigation into AD provides a nuanced understanding of the involvement of microglia in the progression of the disease. The study employed cutting-edge techniques, scRNA-seq, WGCNA, and machine learning, to dissect the intricate molecular and cellular landscape of AD. This study delves into the intricate dynamics between microglia and AD, shedding light on the potential molecular mechanisms and therapeutic targets associated with this neurodegenerative disorder.

The focus on microglial alterations is paramount, given their pivotal role in responding to the accumulation of Aβ plaques and neurofibrillary tangles, the hallmark features of AD21,22. Traditionally considered guardians of neural health, microglia undergo activation in an attempt to clear or contain these pathological elements. However, the complex interplay between microglial activation and AD progression remains a subject of intense investigation. The utilization of scRNA-seq allowed for an in-depth exploration of cellular-level pathological changes in AD. In this study, the scRNA-seq analysis, particularly focused on microglia, revealed 11 distinct subclusters, 4 of which exhibited significant alterations in AD, underscoring the complexity of microglial responses to the disease. Pseudotime analysis revealed a unique differentiation pattern of AD-related microglia compared to their healthy counterparts. This insight into the dynamic changes in microglia subclusters contributes significantly to our understanding of their role in AD progression.

The examination of cell–cell communication networks using the CellChat algorithm revealed intricate interactions between AD-related microglia and various cell types. The activation of microglia triggered the transformation of astrocytes into the neurotoxic A1 subgroup through secretion of cytokines IL-1α, TNF, and C1q23,24. Oligodendrocytes, responsible for forming myelin sheaths and participating in synaptic transmission among neurons, relied on microglia to maintain the physiological stability of myelin sheaths. However, dysfunctional aging microglia could lead to myelin damage, contributing to cognitive disorders like AD25. Our findings revealed a unidirectional communication pattern from microglia to neurons, with neurons not transmitting signals to microglia.

Integration of WGCNA and scRNA-seq facilitated the identification of critical genes in AD-related microglia. The subsequent enrichment analysis unveiled the involvement of these genes in key processes such as peptide chain elongation, synapse-related functions, and cell adhesion. The enrichment analysis revealed that the peptide chain extension was the most significantly affected process, indicating a substantial alteration in the gene cluster associated with protein synthesis and peptide chain extension in microglia. The enrichment of protein interactions at the synapse confirmed the crucial role of AD-related microglia in processes such as synapse formation, transmission, and plasticity. The increased regulation of cell adhesion suggested that activated microglia perform physiological functions, such as phagocytosis, by migrating to lesion sites26. The enhanced microglia adhesion in the diseased state facilitated the phagocytosis and clearance of pathological substances. This enrichment profile provides valuable insights into the molecular mechanisms underlying microglial responses in AD.

The identification of hub genes, including HLA-B, RPS15, TPT1, ARHGAP22, NRXN3, RPL41, ANKS1B, USP3, and PDE4D, through LASSO and COX regression analyses provided a molecular signature associated with AD. The correlation of these hub genes with various immune cell types underscores the intricate immune landscape in AD. The exploration of AD molecular subgroups using the NMF method revealed distinct patterns of immune cell infiltration. Normally, peripheral immune cells such as neutrophils and monocytes could not penetrate the blood–brain barrier and enter the CNS. However, chronic disease stimulation could weaken the blood–brain barrier, allowing neutrophils to infiltrate the CNS27. Neutrophil infiltration had been observed in AD mouse models28, and depleting neutrophils in the CNS significantly improved cognitive function in AD mice29. The differential immune microenvironment across AD subgroups, particularly the variations in neutrophil infiltration, highlights the heterogeneity within AD and suggests potential immune-mediated mechanisms.

USP3, a standout hub gene and a member of the deubiquitinating enzyme family, emerges as a potential key player in neuroinflammation and a candidate for therapeutic intervention30,31. Notably, USP3 has been found to promote inflammasome activation through ASC ubiquitination32, and negatively regulated type I interferon signaling by activating RIG-I33. USP3, identified as a hub gene, emerges as a promising candidate for further exploration. Its upregulation in AD-associated microglia, validated through both bulk RNA-seq and scRNA-seq, positions it as a potential target for immunotherapy. The observed elevation of USP3 in an LPS-induced inflammation model further supports its role in neuroinflammation, providing a potential avenue for intervention in AD.

For additional hub genes, our findings align with existing studies in AD. Specifically, the PDE4 family has emerged as a pivotal contributor to microglial function in an AD model34. Inhibition of PDE4 has been shown to mitigate microglial responses to inflammation, underscoring its potential as a therapeutic target35. Furthermore, HLA-B gene has been implicated in autoimmune encephalitis within the CNS36,37,38. Recent genomic investigations in individuals with late-onset AD have revealed noteworthy correlations between gene polymorphisms within the HLA family and an elevated risk, suggesting these genes could serve as novel risk loci20,39.

CNN, a deep learning tool, has been widely applicated in bioinformatics, including DNA/RNA sequence analysis, protein structure prediction, gene expression profiling analysis, and disease model construction. The utilization of a CNN model for constructing an AD risk prediction model showcases the potential of deep learning in enhancing diagnostic accuracy in this study. The incorporation of gene expression levels, immune infiltration scores, and individual sample size in the CNN model enhances its robustness in identifying AD patients. The detailed training epochs analysis demonstrates the model's robustness, indicating a promising avenue for significantly enhancing diagnostic accuracy in identifying AD patients. This predictive model for AD holds the potential for refining the precision of AD diagnosis or prediction but also signifies a noteworthy advancement in leveraging deep learning for intricate biomedical applications.

Despite the comprehensive insights provided by this study, certain limitations, such as the absence of in vivo validation experiments and considerations for different AD stages, should be acknowledged. Nevertheless, the identified hub genes and the proposed CNN model hold promise for advancing our understanding of AD pathogenesis and improving diagnostic precision, opening avenues for targeted therapeutic intervention.

In conclusion, this comprehensive study employed bioinformatic techniques, including scRNA-seq, WGCNA, and a CNN model, to unravel the intricate molecular and cellular dynamics of AD. By meticulously analyzing microglial alterations associated with AD, the research identified 11 distinct subclusters, 4 of which exhibited significant changes in the disease context. The investigation into cell–cell communication networks revealed intricate interactions between AD-related microglia and various cell types, shedding light on the complex cross-talk within the CNS. Integration of WGCNA and scRNA-seq identified critical genes in AD-related microglia, providing valuable insights into their functional roles, particularly in processes such as peptide chain elongation, synapse-related functions, and cell adhesion. The identification of 9 hub genes, including USP3, through LASSO and COX analyses, suggests potential targets for therapeutic intervention in AD. Moreover, the development of a CNN-based model showcases the promising application of deep learning in enhancing diagnostic accuracy for AD. Overall, this study significantly contributes to our understanding of the molecular mechanisms underlying microglial involvement in AD and offers potential avenues for targeted therapeutic strategies and improved diagnostic precision.