Integrated Bioinformatics and Ensemble Learning Reveal Diagnostic Modeling and Drug Discovery in Alzheimer’s Disease

Integrated Bioinformatics and Ensemble Learning Reveal Diagnostic Modeling and Drug Discovery in Alzheimer’s Disease | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Integrated Bioinformatics and Ensemble Learning Reveal Diagnostic Modeling and Drug Discovery in Alzheimer’s Disease WANG ZIFU, Hou Jinqi, Zhu Yuxuan, Chenyun Guan This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8200314/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background: Alzheimer’s disease (AD) is driven by complex molecular and immune dysregulation, yet reliable diagnostic biomarkers and druggable targets remain limited. This study aimed to identify key AD-associated regulatory genes, characterize their immune and spatial expression features, and prioritize small-molecule compounds with therapeutic potential. Methods: Multiple AD-related transcriptomic datasets—including bulk RNA-seq, microarray, and spatial transcriptomic profiles—were retrieved from GEO and systematically partitioned into discovery (GSE5281, GSE66333), validation (GSE110226, GSE28146, GSE29378), independent testing (GSE29378), and spatial validation cohorts (GSE147047). Differential expression analysis and weighted gene co-expression network analysis (WGCNA) were used to construct co-expression networks and define AD-associated gene modules. Protein–protein interaction (PPI) analysis and multiple network centrality measures were then applied to prioritize candidate key genes. Twelve machine-learning algorithms were combined into 127 classification models, and SHAP-based interpretability analysis was used to quantify feature contributions and identify diagnostic genes. Single-cell and spatial transcriptomic data were further used to validate the cell type specificity and spatial localization of the hub genes. Drug–gene enrichment analysis (DSigDB), compound retrieval (PubChem), ADMET and drug-likeness profiling, and molecular blind docking were integrated to screen and evaluate potential lead compounds. Results: We identified 2,534 differentially expressed genes (DEGs) between AD and control samples, and their intersection with WGCNA-derived modules yielded 848 candidate genes. PPI-based network analysis prioritized 15 key genes, on which 127 machine-learning models were constructed; the random forest model achieved the best overall performance with an average AUC of 0.957. SHAP analysis identified 11 key diagnostic genes, among which IGF1R and SPP1 emerged as stable hub genes with AUCs greater than 0.70 across multiple external cohorts. Immune infiltration, single-cell, and spatial transcriptomic analyses demonstrated distinct immune associations and cell type– and region-specific expression patterns of these hub genes. Drug–gene enrichment identified 176 drug signatures and 445 related compounds, of which 37 grade-A molecules remained after ADMET and drug-likeness filtering. Molecular docking revealed four top-ranked compounds with binding energies better than −9.0 kcal/mol, including one ligand with a minimum binding energy of −10.5 kcal/mol and extensive non-covalent interactions with the target protein. Conclusion: A systematic methodological framework from gene discovery and diagnostic modeling to lead drug screening was developed in this study. IGF1R and SPP1 were identified as stable and biologically interpretable AD hub genes, which can be used as potential diagnostic markers, and various high-affinity small molecule compounds based on the hub genes provide new drug candidates for targeted AD therap. Bioinformatics Artificial Intelligence and Machine Learning Drug Discovery, Design, & Development Cellular & Molecular Neuroscience Alzheimer’s disease machine learning SHAP spatial transcriptomics molecular docking Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Introduction Alzheimer's disease (AD), a neurodegenerative disease, is characterized by progressive cognitive decline, memory impairment and irreversible neuronal loss (1). It has become the most common type of dementia in the elderly population and constitutes a major public health challenge worldwide. Epidemiologic data show that the number of people with dementia has reached approximately 55 million worldwide and is projected to increase to 139 million by 2050, continuing to increase the burden on healthcare and social security systems worldwide (2). In terms of the economic burden of disease, annual healthcare expenditures related to dementia have exceeded US$1.3 trillion and are expected to exceed US$2.8 trillion by 2030 (3). In addition, AD-related mortality rates in the United States have risen by more than 140% between 2000 and 2021, underscoring the urgency of upgrading the diagnostic and therapeutic capacity of the disease (4). Despite significant progress in the study of AD pathogenesis, early diagnosis of the disease remains a challenge, with the main obstacle being the lack of specificity of clinical symptoms in the early stages of the disease, which often leads to delayed recognition and lagging interventions (5). Accurate and timely diagnosis is essential for the development of clinical strategies and patient prognosis. In this process, reliable biomarkers play a central role in early screening, precise staging, disease progression monitoring, and identification of therapeutic targets.. However, the molecular regulatory network of AD has not been fully elucidated, and there is still a lack of consensus biomarkers with sufficient diagnostic efficacy that can be generalized to clinical applications (6). Machine learning (ML) methods have shown unique advantages in parsing high-dimensional complex biological data (7). Algorithms such as random forests, support vector machines, neural networks and deep learning are widely used in genomics and bioinformatics to mine key biological features from multi-omics data (8). By integrating large-scale molecular phenotyping and clinical information, ML is able to effectively identify nonlinear association patterns, and thus promote the systematic discovery of disease-related biomarkers and predictive models (9, 10). The deep integration of bioinformatics and ML provides a powerful framework for the systematic identification of key regulatory genes in diseases, which has demonstrated significant value in disease molecular typing, prognosis prediction, and development of targeted therapeutic strategies (11, 12). In the field of AD research, Base-ML analysis methods integrating multi-omics data have made substantial progress in diagnostic marker identification and potential therapeutic target mining. However, existing studies still have obvious limitations in feature selection strategies—especially the lack of systematic evaluation of multiple algorithm combinations to ensure the robustness and reproducibility of diagnostic features (13, 14). To address this challenge, this study proposes a new research paradigm: to construct a more reliable screening system for AD-related genetic features by integrating a large number of supervised machine learning models (15, 16). We hypothesize that this multi-model integration strategy can identify key gene features with high stability and clinical applicability. On this basis, we further explore the multidimensional features of these diagnostic genes, including their regulatory roles in the immune microenvironment, potential functional pathways, and spatial expression patterns (17). These analyses are expected to provide new insights into the elucidation of AD pathogenesis and provide a theoretical basis for targeted drug development. In this study, we identified differentially expressed genes (DEGs) and WGCNA-derived module genes, and the intersecting genes were screened using 12 machine learning algorithms and 127 combinations selection techniques to determine reliable diagnostic biomarkers for AD (18, 19). The diagnostic performance of these biomarkers was validated through nomogram construction, calibration analysis, and clinical decision curve analysis. Immune cell infiltration and regulatory features were analyzed to elucidate immune-related mechanisms (20). Spatial transcriptomics was applied to confirm the spatial distribution of key genes in AD brain tissue. Finally, therapeutic compounds were identified through drug enrichment analysis, ML–based ADMET profiling, and Smina automated blind docking to prioritize high-affinity compound–target complexes for potential AD treatment (21, 22). Methods Study design Based on the GEO database, this study identified AD-related DEGs (23). Hub genes screening was performed using protein-protein interaction analysis and an ensemble MLapproach, with key candidates validated through transcriptional regulatory networks and spatial transcriptomics. Promising therapeutic agents were then identified via drug-gene interaction analysis, ML-based ADMET profiling, and molecular docking for high-affinity complexes screening (Fig. 1) (24). ============================ Fig. 1 ============================ Fig. 1 Flowchart of the study Data sources used for analysis Five distinct gene expression profiles (GSE5281, GSE66333, GSE110226, GSE28146, GSE29378, and GSE29378), as well as the spatial transcriptomic dataset GSE147047, were retrieved from the GEO database. GSE5281 and GSE66333 were designated as the training group for this study, comprising a sample of 91AD patients and 78 non-AD controls, while GSE110226, GSE28146, and GSE29378 served as the validation and independent testing groups. Following logarithmic transformation, each dataset was normalized using the “normalizeBetweenArrays” function of the R package “limma” (version 3.66.0) (25). To eliminate batch effects between datasets, the expression matrices were corrected and merged using the “sva” package (version 3.58.0). Principal component analysis (PCA) is mainly used for data dimensionality reduction and feature extraction, and reduce the complexity of data (15). The merged matrix was then used for differential expression analysis with “limma,” and the filtering criteria for DEGs were set as |log₂Fold-Change| > 0.585 and false discovery rate (FDR) < 0.05 (26). Weighted gene co-expression network analysis (WGCNA) of DEGs To build a WGCNA for DEGs, we utilized the R package "WGCNA" (version 1.73.0) (27). The objective was to identify gene modules significantly linked to AD. Quality control and sample clustering were performed on the gene expression data to exclude abnormal samples. The soft threshold function was used to determine the optimal power value, construct the neighbor-joining matrix between genes, and further calculate the topological overlap matrix (TOM). Hierarchical clustering was performed based on the dissimilarity of the TOM matrix, and gene co-expression modules were identified by a dynamic shear tree algorithm. Modules with highly correlated feature genes were merged to obtain final independent modules (28). Key modules with significant phenotypic associations were screened by calculating the correlation between module feature genes and clinical traits, and the genes that intersect between the significant coexpressed genes identified through WGCNA and DEG were analyzed for protein-protein interactions (PPI) analysis (27). Construction of PPI networks In this study, we utilized the STRINGdb package (version 2.22.0) to obtain protein interactions from the "org.Hs.eg.db" package (versioin 3.22.0) (29). This network specifically focused on the intersection of DEGs and those genes identified through WGCNA. combined_score≥400 was set as the threshold for reliable interactions. Gene ID conversion was performed using the bitr function of the "clusterProfiler" package (version 4.18.1) to convert gene symbols to ENTREZID for database mapping (30). The network was constructed using the "igraph" package (versioin 2.2.1), which computes five network centrality metrics: degree centrality, betweenness, closeness, eigen_centrality, and eigen_centrality. centrality) and page_rank algorithm (31). A standardized weighted scoring model (weight coefficients of 0.25, 0.25, 0.20, 0.15, 0.15, respectively) was used to calculate the composite hub scores and screen the top 15 key genes. Visual analysis was performed using the ggraph package (version 2.2.2) for network layout, and the node positions were optimized using the stress layout. In the node attribute mapping, size was positively correlated with degree centrality (rescale function normalized to the range of 3-15), and color was divided into blue, orange, and red gradients according to the median centrality tertiles. Edge attribute settings for width and transparency were positively correlated with interactions scores (rescale to the 0.3-2 and 0.3-0.8 ranges). Chord diagrams were drawn using the circlize package (version 0.4.16), which demonstrates the top 30% of high-intensity interactions via the chordDiagram function, and the represented genes were considered as key genes for subsequent Gene Ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG), and Disease Ontology (DO) analyses (32). Functional Enrichment Analysis of Hub Genes: GO, KEGG and DO Integration Gene function enrichment analysis is an important method for elucidating the biological significance of gene sets (32). Annotation related to human diseases, biological functions, and signaling pathways helps reveal the connections and roles among overlapping genes associated with Alzheimer's disease (AD). The target gene set was systematically functionally annotated using the clusterProfiler, gene IDs were converted using the org.Hs.eg.db database, and three levels of enrichment analysis were performed: GO analysis, including biological processes (BP), molecular functions (MF), and cellular components (CC), to reveal the functional characteristics of genes; KEGG pathway analysis to identify key signaling pathways and metabolic pathways involved by genes; DO analysis to find the correlation between genes and human diseases. Enrichment results were presented using p < 0.05 and FDR < 0.05 as significance thresholds, and visualized using various methods such as bubble charts, bar charts, network diagrams, and pie charts. Model Construction Method Based on Integrated ML and SHAP Key genes for AD interactions identified by PPI network analysis algorithms were subsequently integrated into different machine learning frameworks. In this study, we used to build a two-stage machine learning framework. In the first stage, 12 algorithms were applied for initial screening of features, including regularization methods (Lasso, Ridge, ElasticNet), integrated learning models (Random Forest, GBM, XGBoost), generalized linear models (Stepglm, glmBoost, plsRglm), and pattern recognition models (SVM, LDA, Naive Bayes). Hyperparameter optimization was performed by grid search combined with 10-fold cross-validation, where Random Forest tuned the mtry parameters (2-10), and XGBoost optimized the learning rate (0.01-0.3), maximum tree depth (3-7), and regularization parameters (33). In the second stage, a stacked generalization strategy is used to integrate models from the initial screening results, construct 127 algorithm combinations, and evaluate the performance under a rigorous cross-validation framework, with AUC as the main evaluation metric, and validate the generalization ability of the models in multiple independent cohorts. SHAP interpretability analysis is implemented for the best performing models, and the permSHAP method is used to calculate the feature contribution and generate 95% confidence intervals by Bootstrap resampling (100 times) to ensure the stability of the results (34). The evaluation of feature importance is based on a multi-dimensional index system, including the average absolute SHAP value, the cumulative importance curve (with a double threshold of 80% and 90%), and the statistical significance test. The swarm plot shows the distribution of feature importance, the waterfall plot analyzes the prediction mechanism of individual samples, and the dependency graph reveals the nonlinear relationship between features and predicted output. The core AD diagnostic genes with statistical significance and biological interpretability were selected. Validation analysis of hub genes To assess the diagnostic performance of key diagnostic genes screened by 127 integrated algorithms in AD, the R software package "pROC" (version 1.19.0.1) was used to perform subject work characterization curves (ROCs) analysis and to calculate the area under the curve (AUC) as a measure of the ability of each gene to discriminate AD from control samples (35). Ability. Genes with AUC values greater than 0.60 in both the training and validation cohorts were considered to be AD-related biomarkers with stable diagnostic potential. Subsequently, multigene logistic regression models were constructed using the R software package "glmnet" (version 4.1.10) to evaluate the combined diagnostic efficacy of key genes in the training cohort, validation cohort, and spatial transcriptome data. Meanwhile, "PerformanceAnalytics" (version 2.0.8) was used to perform correlation analysis and visualization of the hub genes in the validation cohort to reveal the interrelationships among genes and potential co-regulatory features. Analysis of transcription factor-hub genes regulatory networks To investigate the upstream regulatory mechanisms of hub genes, this study was based on the TRUST v2 database (www.grnpedia.org/trust, accessed on November 1, 2025) to screen transcription factors with potential interactions with hub genes (36). Based on the predictive confidence scores provided by the database, the top 10 transcription factors with the highest scores were selected as candidate key regulators for subsequent analysis. Transcription factors that were up-regulated in AD tissues were regarded as likely to exert a positive regulatory effect on the hub genes. Subsequently, the expression levels of the candidate transcription factors and the hub genes were correlated to assess their potential regulatory relationships, and a schematic diagram of the transcription factor-target gene regulatory network was constructed accordingly. Correlation between Hub Genes and Immune Cells To assess the immune microenvironment characteristics of Alzheimer's disease (AD) tissues, this study used CIBERSORT-based deconvolution (https://cibersortx.stanford.edu/) to perform immune cell deconvolution analysis on Discovery data (37). CIBERSORTx, based on a linear support vector regression (SVR) model, estimates the proportion of cell infiltration, including 22 immune cell types. Before analysis, the expression matrix was quantile-normalized and its gene intersection was matched with an LM22 reference matrix. Samples with p < 0.05 were selected as reliable results using a permutation test (permutation = 1000), and Spearman correlation analysis was used to assess the association between pivotal genes and key immune cell types to identify AD-related changes in immune infiltration. The analysis results were visualized using box plots, correlation plots, and other methods. Single-cell and spatial transcriptomics validation analysis In order to further validate the expression pattern of hub genes systematically, this study integrated single-cell RNA sequencing and spatial transcriptomics data for multilevel analysis (38). The scRNA-seq data were acquired from GEO and processed for quality control, normalization and clustering using Seurat (version 5.3.1). The quality control thresholds were set at nFeature_RNA 15%; LogNormalize (scale factor = 10,000) was used for normalization, and 1,500 highly variable genes were screened for downstream analysis. This was followed by initial PCA downscaling and correction for batch effects using harmony (version 1.2.4); shared nearest neighbor (SNN) based mapping and clustering using Louvain's algorithm (resolution 0.6). Cell type annotation was done by SingleR, using HumanPrimaryCellAtlas as the reference database. For the screened hub genes, t-SNE, violin plot and dot plot were used to evaluate their expression distribution and cell specificity in different cell types. In order to further validate the spatial localization and expression patterns of hub genes, 10x Genomics Visium spatial transcriptome data were integrated, and the raw data were pre-processed by Space Ranger to map the single-cell annotation results to the tissue spatial coordinates, and the spatial expression characteristics of the hub genes were demonstrated by spatial feature maps and analyzed in conjunction with the images of tissue slices. The spatial expression characteristics of Hub genes are shown in the spatial feature map, and their localization and enrichment in specific tissue regions are analyzed in conjunction with the tissue slices, thus realizing the spatial validation of the single-cell analysis results. Drug enrichment analysis and candidate compound screening After obtaining the pivotal genes, this study conducted drug enrichment analysis based on the DSigDB database to evaluate potential drug-gene interactions and screen candidate therapeutic compounds (39). Drug target enrichment analysis was performed using the enricher function of the “clusterProfiler” toolkit with a significance threshold set at p < 0.05 and corrected p < 0.05 to reduce the false positive rate. Significantly enriched drug entries were visualized by bubble plots to present the strength of association between drugs and gene sets. Based on the candidate compounds obtained from the enrichment analysis, the corresponding SMILES structural formulas were further obtained in this study via the PubChem database (https://pubchem.ncbi.nlm.nih.gov/) to provide key inputs for subsequent virtual screening and drug repositioning studies (40). Computational Prediction of Drug-likeness In the present study, a systematic pharmacophore evaluation of compounds obtained from the PubChem database was carried out using computational chemistry. Based on the RDKit (version 2024.3.5) toolkit in the Python (version 3.8.15) environment, we calculated key molecular descriptors, including molecular weight, lipid-water partition coefficient, and topopolar surface area, and constructed a seven-dimensional weighted scoring system (41). The system integrates the following evaluation dimensions: the Lipinski rule (15% weight) to assess oral absorption properties, the Veber rule (10%) to determine bioavailability, the Ghose rule (10%) to validate drug-like criteria, the quantitative drug similarity index QED (20%) to comprehensively evaluate drug similarity, the synthetic accessibility score SA Score (15%) to measure synthetic feasibility, ADMET Characterization (20%) to predict pharmacokinetic profiles, and Structure Alert Screening (10%) to identify potentially toxic moieties (42). After obtaining the comprehensive scores through weighted calculation, the compounds were classified into five grades of A (Excellent), B (Good), C (Medium), D (Poor), and F (Unqualified), which provided a reliable basis for prioritizing the compounds for subsequent molecular docking studies. Intelligent molecular docking and multidimensional validation To validate the binding potential of the lead compounds obtained from the enrichment analysis, the molecular docking method based on Python's Smina (version 2020.12.10) library was used in this study for systematic validation (43). After obtaining the crystal structure of the target protein (PDB ID: [3D94, 5FXS]) from the RCSB PDB database, the protein structure was extracted and processed using the Python's MDAnalysis (version 2.2.0) library, followed by protein protonation and PDBQT formatting via Open Babel at various physiological pH conditions (6.4, 7.4, 8.4) (44). The ligand molecules are based on SMILE. For ligand molecules, the PDBQT files for docking were generated based on SMILES expressions, through 3D structure generation, force field optimization and format conversion. A dual strategy based on the automatic detection of eutectic ligands and the designation of key active residues (e.g., HIS:114, ASP:165, GLU:166) was used to define the binding pockets during the docking process, and different exhaustion levels (8, 16, 32, 64) were systematically evaluated to balance the efficiency and adequacy of the conformational search. In order to improve the computational efficiency, an intelligent task distribution system is constructed, which can automatically detect the computational resources and optimize the parallel processing strategy, and each task covers the complete pre-processing-docking-post-processing flow, and is equipped with task-level timeout protection and automatic retry mechanism, which guarantees the stable operation of large-scale computational tasks. The docking results are verified by RMSD redocking, which evaluates the reliability of docking by comparing the deviation between the initial and redocked conformations, and lays the foundation for future cell and animal experiments. Results A total of 2534 DEGs associated with AD were identified Two datasets (GSE5281 and GSE66333) were selected as training models. We observed baseline batch differences between the included data sets (Fig. 2A). In order Immune microenvironment of AD brain tissues, PCA was applied, all batch differences are eliminated (Fig. 2B). A total of 2534 DEGs were identified with log2FC > 0.585 and FDR < 0.05 between the AD group (n=91) and the Control group (n=78), including 1325 up-regulated genes and 1209 down-regulated genes (Fig. 2C-D) (S1). ============================ Fig. 2 ============================ Fig. 2 Data preprocessing and DEGs screening. using principal component analysis for batch calibration of GSE5281and GSE66333. (A) Before batch calibration. (B) after batch calibration. (C) Volcano map. (D) Heat map WGCNA identified key modules on DEGs that are closely related to AD The goal of WGCNA is to identify co-expressed gene modules closely associated with AD. WGCNA focuses on network mining of genes based on expression pattern consistency. This method first calculates the Pearson correlation coefficient between any two genes, then constructs a scale-free weighted adjacency matrix using a soft threshold, generating a TOM and performing hierarchical clustering to identify co-expressed modules. The weighted network constructed based on the training set expression matrix shows that, with the determination of the soft threshold, both the adjacency matrix and the topological overlap matrix exhibit good scale-free network characteristics (Fig. 3A–C). Dynamic tree pruning and average clustering methods ultimately identified a significant module (Fig. 3D), where the MEbrown module (R = 0.62, p < 0.05) showed the strongest correlation with the AD phenotype, containing 1854 genes (Figure 3E) (S2). Taking the intersection of these modules with DEGs yielded 848 candidate genes, providing a basis for subsequent PPI screening (Fig. 3F). ============================ Fig. 3 ============================ Fig. 3 DEGs and WGCNA analysis Results for display of AD. (A) Analysis of the scale-free index for various soft-threshold powers(β). (B) Analysis of the mean connectivity for various soft-threshold powers. C: threshold for WGCNA analysis. (D) Heatmap of the correlation between module characterized genes and the level of DEGs. Each cell contains the corresponding correlation and p-value. (E) Module trait relationship and module features of AD. (F) Intersection gene between module gene and DEGs in WGCNA analysis Identification of biomarkers for AD using PPI networks and multifunctional enrichment analysis A high-confidence PPI network was constructed by integrating DEGs and WGCNA genes, from which the top 15 key genes were screened based on their overall centrality scores (Fig. 4A) (S3). To elucidate the potential roles of these genes in AD, these genes were subsequently analyzed for functional enrichment of GO, KEGG and DO. The results showed that glial cell differentiation, gliogenesis in the pathogenesis of AD (Fig. 4B). the GO analyzed gene set was significantly enriched for a variety of biological processes, including endoplasmic reticulum lumen, vesicle lumen, chromatin DNA binding, integrin binding, glial cell differentiation, gliogenesis (Fig. 4C). Further KEGG pathway analysis showed that these genes were significantly enriched in Adherens junction, FoxO signaling pathway, Human papillomavirus infection, JAK-STAT signaling pathway, MicroRNAs in cancer, Notch signaling pathway signaling pathway. DO analysis showed that these genes were significantly associated with brain ischemia, cognitive disorder, retinal disease, and endocrine gland cancer (Fig. 4D). ============================ Fig. 4 ============================ Fig 4 Key Gene Screening and Functional Enrichment in AD. (A) Top 15 key genes from the PPI network. (B) GO enrichment. (C) KEGG pathway enrichment. (D) DO enrichment. Constructing a diagnostic model of AD by machine learning Based on the 15 candidate key genes screened by the preliminary PPI network analysis, this study further integrated 12 machine learning algorithms and constructed a total of 127 predictive models with different combinations to screen the most robust diagnostic classifiers. These models were all built in the training cohort and systematically evaluated on 4 external validation datasets, and the whole process was completed based on a cross-validation framework. Among the 127 models constructed, the RF integration algorithm demonstrated the highest average AUC in both the training and test sets (training set AUCs with significant overfitting have been excluded), with an average AUC of 0.957 (Fig. 7A), and was therefore identified as the optimal classification model. Subsequently, we performed feature importance parsing on the RF model based on the SHAP (SHapley Additive exPlanations) method, which was used to quantify the contribution of each gene to the classification results. This analysis finally filtered out 11 feature genes: APP, CREBBP, SUZ12, CXCR4, SNAP23, BCL6, NFKBIA, IGF1R, CEBPB, SPP1, and EZR (Fig. 7B) . To further validate the generalization ability and diagnostic efficacy of different feature selection strategies and their constructed models, we apply each model to an external validation cohort and compare their classification performance in the validation set. As shown in Fig. 7C, the AUC values of the IGF1 and SPP1 correlation models were consistently greater than 0.70 in multiple validation datasets (GSE110226, GSE28146, GSE29378, GSE29378), suggesting that these two genes are hub genes with stable diagnostic value (S4). ============================ Fig. 5 ============================ Fig 5 127 machine learning algorithm combinations evaluated via 10-fold cross-validation and SHAP screening key genes and verification evaluation of hub genes. (A) 127 machine learning algorithm combinations evaluated via 10-fold cross-validation. (B) SHAP Analysis of the Best Model. (C) The ROC curve and expression level of the characteristic gene screening hub genes. Analysis of upstream transcriptional regulation of IGF1R and SPP1 and their association with the immune microenvironment In order to deeply investigate the upstream transcriptional regulatory mechanisms of hub genes IGF1R and SPP1 and their roles in the tumor immune microenvironment, we systematically constructed a transcription factor-target gene regulatory network (Figure 7A). The results showed that IGF1R was at the center of this regulatory network and was co-regulated by several key transcription factors, including TP53, SP1, POU2F1, NR3C1, KLF6, and WT1, suggesting that IGF1R is subjected to a complex and multi-level regulation at the transcriptional level and may play a central role in the disease process. Further, we evaluated the association between the expression levels of IGF1R and SPP1 and the degree of infiltration of various types of immune cells in the tumor microenvironment by immune infiltration analysis (Fig. 6B-D). IGF1R expression was moderately negatively correlated with the infiltration levels of some naive and memory B cells and moderately negatively correlated with the infiltration levels of most innate immune cells (e.g., monocytes and T cells gamma delta) and correlated weakly (Fig. 6B). On the contrary, SPP1 expression showed a significant positive correlation with the infiltration degree of various myeloid-derived immune cells, such as M0/M1/M2 macrophages and neutrophils, and a negative correlation with B cell memory (Fig. 6C). As shown in Figure 6D, the differential association pattern between IGF1R and SPP1 with different immune cell subpopulations was further visualized. ============================ Fig. 6 ============================ Fig 6 Transcriptional regulatory network of hub genes IGF1R and SPP1 and their association with immune cell infiltration. (A) TF–gene regulatory network of hub genes IGF1R and SPP1. (B) Correlations between IGF1R expression and infiltrating immune cell subsets. (C) Correlations between SPP1 expression and infiltrating immune cell subsets. (D) Network visualization of hub genes and significantly correlated immune cell subsets. Single-cell transcriptomic analysis reveals cellular heterogeneity and expression patterns of IGF1R and SPP1 To further validate the expression characteristics of the aforementioned hub genes and their cellular origin at the single-cell transcriptomic level, we implemented a systematic quality control analysis of single-cell RNA sequencing data. By assessing the number of genes detected (nFeature_RNA), UMI count (nCount_RNA), and mitochondrial gene percentage ( percent.mt ) in each cell, it was confirmed that the cellular quality of the control group and the treatment group was good overall and met the criteria for subsequent analysis (Fig. 7A). Subsequently, principal component analysis was performed based on the highly variable genes obtained from the screening, and the highly loaded genes corresponding to principal components 1 to 4 (PC1-PC4) were extracted for resolving the major transcriptional heterogeneity among cellular subpopulations (Fig. 7B). The results of t-SNE downscaling visualization showed that the cells were classified into major subtypes such as neuroepithelial cells versus neurons. Although there was a partial overlap between the control and treated cells in the two-dimensional space, there was a significant difference in the proportion of their cellular composition (Fig. 7C). Further, we applied the Slingshot algorithm to construct cell differentiation trajectories, and the proposed time-series analysis revealed a potential pathway of gradual differentiation from neuroepithelial cell to neuron state (Fig. 7D). On this basis, hub genes (IGF1R vs. SPP1) were mapped to single-cell expression profiles, and their expression dynamics were compared between the two groups. The results showed that IGF1R showed a clear pattern of high expression in specific cell subpopulations, and the overall expression level was significantly upregulated in the treated groups; in contrast, SPP1 was expressed at a lower level in most cells, with no significant trend in the difference between groups (Fig. 7E). ============================ Fig. 7 ============================ Fig 7 Single-cell transcriptomic analysis reveals cellular heterogeneity and expression patterns of IGF1R and SPP1）.(A) A Distribution of single-cell RNA-seq quality control metrics (nFeature_RNA, nCount_RNA and percent.mt) in control and treated groups. (B) Top loading genes of principal components PC1–PC4. (C) t-SNE visualization of control and treated cells and annotation of major cell subpopulations. (D) Slingshot-inferred pseudotime trajectory of cell differentiation. (E) Spatial distribution and differential expression of IGF1R and SPP1 in single-cell profiles between control and treated groups. Screening of potential therapeutic small molecules based on gene–drug enrichment and ADMET profiling To uncover potential therapeutic drugs based on hub genes at the molecular level, we systematically carried out drug enrichment and drug-like assessment analysis. The enrichment results based on the public drug-gene interaction database showed that hub genes were significantly associated with 176 drug components (P < 0.05), suggesting that these drugs may be involved in the disease process by modulating key genes (Fig. 8A) (S5). Based on this, we collected and organized 445 relevant small molecule compounds from the PubChem database and conducted a multidimensional systematic evaluation of their ADMET properties and drug-like properties, covering key parameters such as oral bioavailability, lipid solubility, polarity, molecular weight, metabolic stability, and potential toxicity. According to the comprehensive scoring system, all the compounds were categorized into six grades from A to F, in which grade A represents the candidate molecules with optimal pharmacokinetic and drug-like properties. Through this evaluation system, a total of 37 A-grade compounds were screened, and they were identified as the key candidate ligands for subsequent molecular docking studies (Fig. 8B) (S6). ============================ Fig. 8 ============================ Fig 8 Screening of candidate small-molecule compounds for molecular docking based on gene–drug enrichment and ADMET/drug-likeness analysis. (A) Significantly enriched drug ingredients associated with candidate genes (n = 176, P < 0.05). (B) ADMET and drug-likeness profiling of 445 candidate compounds. Molecular docking-based analysis of binding modes and interactions between candidate small molecules and target proteins To further evaluate the binding ability and mode of action between the candidate class A small molecule compounds and the target proteins, we carried out Cartesian product molecular blind docking analysis based on the target proteins with PDB IDs (3D94 and 5FXS) (the docking parameters were set as pH = 7.4, Exhaustiveness = 16, and Box mode = ligand) (S7). In the docking results, the compound C1CN(CCC12C(=O)NCN2C3=CC=CC=C3)CCCC(=O)C4=CC=C(C=C4)F, which ranked at the top of the affinity list, showed the best free energy of binding to the target protein 3D94 of all ligands at -10.5 kcal/mol (Figure 9A). Conformational analysis showed that the molecule is deeply embedded in the hydrophobic pocket of the protein, forming extensive hydrophobic interactions with residues LEU975, VAL983, ALA1021, and PHE1124, and establishing a network of multiple hydrogen bonds with residues CYS1111, VAL1113, ASP1123, and PHE1124, with a bond spacing ranging from about 2.2 to 3.8 Å. The compound CN1CCN(CC1)C2=NC3=C(C=CC(=C3)Cl)NC4=CC=CC=C42 has a binding energy of -9.2 kcal/mol with 5FXS (Figure 9B). The ligand mainly formed hydrophobic stacking with hydrophobic residues such as LEU1005 and ALA1031 through the aromatic ring and formed multiple hydrogen bonds with ASP1086, with an interaction distance of about 2.0-4.0 Å. In addition, its chlorine atom formed a stable halogen bond with the carboxylate oxygen of ASP1153, which further enhanced the ligand binding to the target protein. The binding affinity and specificity between the ligand and the target protein are further enhanced. Compound CC1(CCC(C2=C1C=CC(=C2)NC(=O)C3=CC=C(C=C3)C(=O)O)(C)C)C binds to 5FXS with a binding energy of -9.1 kcal/mol (Fig. 9C), and its hydrophobic skeleton is tightly encapsulated by residues such as LEU1005, LYS1033, and MET1079. Its hydrophobic backbone was tightly wrapped by residues such as LEU1005, LYS1033, and MET1079 and further stabilized the binding mode by forming a hydrogen bond with MET1082. Another compound, COC1=C2C3=C(C(=O)CC3)C(=O)OC2=C4C5C=COC5OC4=C1, also has a binding energy of -9.1 kcal/mol with 5FXS (Figure 9D), and its polycyclic aromatic structure forms multi-point hydrophobic contacts with residues such as LEU1005, ALA1031, and so on. Its polycyclic aromatic structure formed multi-point hydrophobic contacts with residues such as LEU1005 and ALA1031, and despite the relatively small number of hydrogen bonds, its overall affinity was still at a favorable level. The above four representative small molecules exhibited docking affinities better than -9.0 kcal/mol and were stably bound to the protein binding pocket through a variety of non-covalent interactions, including hydrophobic interactions, hydrogen bonding, and halogen bonding (S8). ============================ Fig. 9 ============================ Fig 9 Binding poses and key interactions of top-ranked candidate small molecules with target proteins 3D94 and 5FXS. (A) Compound C1CN(CCC12C(=O)NCN2C3=CC=CC=C3)CCCC(=O)C4=CC=C(C=C4)F Docking conformation and its hydrophobic/hydrogen bonding interactions with the target protein 3D94. (B) Compound CN1CCN(CC1)C2=NC3=C(C=CC(=C3)Cl)NC4=CC=CC=C42 Docking conformation and its hydrophobic, hydrogen- and halogen-bonding interactions with target protein 5FXS. (C) Docking conformation and predominantly hydrophobic binding mode of the compound CC1(CCC(C2=C1C=CC(=C2)NC(=O)C3=CC=C(C=C3)C(=O)O)(C)C)C with 5FXS. (D) Compound COC1=C2C3=C(C(=O)CC3)C(=O)OC2=C4C5C=COC5OC4=C1 Docking conformation with 5FXS and its hydrophobic-dominant binding interactions. Discussion In this study, we systematically identified key biomarkers of AD and screened small molecule compounds with therapeutic potential by integrating multi-omics data and machine learning methods. The findings provide new insights and candidate targets for the diagnosis and treatment of AD. The integrated analysis of multiple cohorts in this study firstly suggests that there are not only a large number of differentially expressed genes in AD brain tissues, but also that these genes are highly clustered in glial cell-related processes and key signaling pathways, and the GO/KEGG/DO enrichment results point to glial cell differentiation, gliomagenesis, JAK-STAT, Notch and FoxO pathways, as well as disease phenotypes such as cerebral ischemia and cognitive deficits, which is highly compatible with the current understanding of the pathological features of AD (45, 46). That neuronal damage often occurs in the context of a long-term glial cell response and chronic inflammation. In other words, our network and enrichment analyses at the transcriptomic level reinforce the view that AD is not a single "neuronal disease" but a systemic pathology centered on the neuronal-glial-immune axis, and that key drivers are most likely located at the intersection of this network (47). The key drivers are likely to be located at the intersection of this network. On this basis, through topology centrality and integrated machine learning, we converge layer by layer on two hub genes, IGF1R and SPP1, which have several noteworthy features. First, the RF models constructed by these 11 characterized genes still maintain an AUC close to 0.96 in multiple cohorts, cross-validation, and external validation, indicating that the expression patterns of these genes are not only discriminative in a single cohort but also stable and migratory across different platforms and sample sources (48). This is not common in highly heterogeneous diseases such as AD and suggests that these genes may be located relatively "upstream" or at "key nodes" in the disease process. Secondly, the AUCs of both IGF1R and SPP1 were > 0.70 in the multi-cohort ROC, suggesting that there is some diagnostic value at the single-gene level—providing a realistic basis for the development of simplified translational tools (e.g., qPCR panels or digital PCR diagnostic markers) in the future. Biologically, the IGF1R-associated signaling axis is closely related to insulin/IGF signaling, metabolic imbalance, cell survival, and stress response, which have long been reported in "brain insulin resistance," impaired synaptic function, and aberrant tau phosphorylation in AD, while SPP1 (osteopontin) is associated with microglia activation. SPP1 (osteopontin), on the other hand, is highly associated with microglia activation, chemotaxis, inflammatory amplification, and tissue remodeling, which is more inclined to an "inflammatory-immune-driven" pathology (49). The TF-target network in this study further showed that IGF1R is located at the intersection of multiple high-confidence transcription factors (TP53, SP1, POU2F1, NR3C1, etc.) (50), suggesting that it is subject to transcriptional regulation at multiple pathways and levels and may act as a "signaling integrator" between stress, metabolism, and inflammation signals. integrator" between stress, metabolic, and inflammatory signals. To some extent, this explains why IGF1R has always been the top contributor in machine learning models: it is not only a "read" but also a "convergence point" of multiple pathological pathways. Immune infiltration and single-cell/spatial transcriptome results provided further clues to the functions of IGF1R and SPP1 from a "cell-microenvironment" perspective (51). CIBERSORT analysis showed that high IGF1R expression was moderately negatively correlated with the infiltration of some B and T cell subsets, while SPP1 was significantly positively correlated with myeloid cells such as M0/M1/M2 macrophages and neutrophils, which formed a differential pattern of "relative suppression of adaptive immunity and significant activation of myeloid/inflammation." This pattern is very consistent with the characteristic of AD in which innate immunity is over-activated, microglia are chronically activated, and adaptive immune involvement is relatively limited. Single-cell analysis further revealed that IGF1R expression was concentrated in neuroepithelial cells and some neuronal subpopulations and changed dynamically with the temporal differentiation from precursors to mature neurons, suggesting that IGF1R was more involved in "neuronal fate determination/survival and plasticity," whereas SPP1 was more focally and sparsely expressed and appeared mainly in a limited number of inflammation-associated cells. SPP1 expression was more focal and sparse, occurring mainly in a limited population of inflammation-associated cells. Combined with spatial transcriptomic evidence, IGF1R and SPP1 showed specific spatial enrichment in AD-susceptible brain regions, implying that the same molecular marker does not play the same role in different cell types and spatial regions, which reminds us that the subsequent intervention strategy needs to take into account the cellular and spatial specificity and to avoid "one-size-fits-all" targeting in the whole brain. This also reminds us that subsequent intervention strategies need to take into account cellular and spatial specificity to avoid "one-size-fits-all" targeted inhibition. The molecular docking results of the four candidate compounds with the target proteins revealed their binding modes and key structure-activity relationships. Among them, compound C1CN(CCC12C(=O)NCN2C3=CC=CC=CC=C3)CCCC(=O)C4=CC=C(C=C4)F exhibited optimal binding affinity (ΔG = -10.5 kcal/mol) with 3D94 protein. Structural analysis indicated that its cyclic backbone constitutes a rigid support; the hydrophobic groups penetrate deep into the hydrophobic cavity of the protein to form stable van der Waals interactions, whereas the amide bonds and fluorine atoms form a multicenter hydrogen bonding network with residues such as CYS1111, ASP1123, and others (bond distances in the range of 2.2-3.8 Å). The introduction of fluorine atoms further enhanced the metabolic stability of the compounds. In the binding mode of compound CN1CCN(CC1)C2=NC3=C(C=CC(=C3)Cl)NC4=CC=CC=CC=C42 with 5FXS, the chlorine atom forms a directional halogen bond with the carboxyl oxygen of ASP1153, a non-covalent interaction that contributes to the enhancement of the binding specificity and free energy contribution. The remaining two polycyclic aromatic compounds have slightly lower binding energies (ΔG = -9.1 kcal/mol), but their thick ring systems are stabilized by extensive hydrophobic stacking with residues such as LEU1005 and ALA1031, and the aromatic planes also have the potential to form π-π or cation-π interactions, which are often helpful in improving the binding specificity and free energy contribution of the molecules. Structural features usually contribute to the improved blood-brain barrier permeability of the molecule. It should be noted that molecular docking is mainly based on the static structure to assess the thermodynamic stability, which is not yet able to adequately reflect the conformational dynamics, dissociation rate, and full-atom scale dynamic behavior during the binding process. Therefore, the actual drug potential of these compounds needs to be systematically evaluated in conjunction with molecular dynamics simulations, cellular-level functional validation, and animal experiments. From a translational perspective, an important value of this study is that instead of merely proposing that "certain genes are up-/down-regulated in AD," multiple approaches point to a core logic chain: Glial and immune-related transcriptional abnormalities in AD to hub genes such as IGF1R/SPP1 are at key regulatory nodes to stable diagnostic value and correlation with specific cellular subpopulations and spatial structures to mapping to specific small molecule sites of action and derivation of drug candidates (21). This linkage makes the process from "discovering markers" to "proposing targets for intervention" to "giving lead compounds" more coherent and interpretable. From biomarker discovery to clinical application faces many problems and challenges, and we need to develop more efficient and cost-effective research programs. There is a need to develop assays that are noninvasive, reproducible, and highly correlated with intracerebral states. Blood or cerebrospinal fluid assays are ideal directions, but their consistency with central expression needs to be systematically verified. Second, standardized testing procedures and uniform diagnostic thresholds need to be established through multicenter studies, and the interference of age, gender, and comorbidities needs to be assessed (52). In terms of therapeutic development, although the compounds screened in this study provide lead molecules for drug innovation, the discovery-to-clinical translation of new drugs usually requires long lead times, high cost investment, and the need to address issues such as blood-brain barrier permeability, long-term safety, and adaptive therapeutic windows. AD pathology often exists for many years prior to symptom onset, which means that the most effective therapies often need to be implemented in the preclinical phase, and the clinical trials face both ethical and practical difficulties in the design and implementation of clinical trials at this stage. This study provides potentially translational targets and drug candidates, but their clinical application requires rigorous systematic validation. Future research should focus on mechanistic insight and technological innovation. We need to further elucidate the dynamic roles of IGF1R signaling dysregulation and SPP1-mediated immunoregulation in AD (53); develop brain-targeted drug delivery systems and multimodal diagnostic platforms; and rely on artificial intelligence to integrate multi-omics and clinical data to build an accurate AD typing and treatment decision-making system, which will ultimately promote the establishment of individualized prevention and treatment strategies (54). Conclusion In this study, we systematically established a screening and validation pipeline for key AD-related genes by integrating multi-omics data with machine-learning approaches. By combining differential expression analysis with WGCNA-derived modules and applying 127 algorithm combinations together with SHAP-based interpretability analysis, we ultimately identified two hub genes, IGF1R and SPP1, whose diagnostic value was validated at multiple levels. Subsequent drug screening based on these hub genes yielded several high-affinity small-molecule compounds, providing both candidate biomarkers and lead compounds for precision diagnosis and targeted therapy in AD. Declarations Funding Not applicable. Compliance with Ethical Standards Not applicable. Consent for publication All authors have declared their consent for this publication. Competing interests The authors declare that they have no conflict of interest. CRediT authorship contribution statement WZF : Investigation, Conceptualization, Methodology development, Coding, Data Analysis, Writing - Original Draft, Writing - Review & Editing, Software implementation, Validation of results. HJQ : Conceptualization, Project Administration, Writing - Review & Editing, Supervision, Drawing, Preparation of figures and tables, Resource allocation. ZY : Algorithm programming, data set testing. GC : Supervisors, project management. Data availability The data will be provided as required. Acknowledgements Not applicable. References Botella Lucena P, Heneka MT. Inflammatory aspects of Alzheimer’s disease. Acta neuropathologica. 2024;148(1):31. Chen S, Cao Z, Nandi A, et al. The global macroeconomic burden of Alzheimer's disease and other dementias: estimates and projections for 152 countries or territories. The Lancet Global Health. 2024;12(9):e1534-e43. Zhang J, Na X, Li Z, et al. Sarcopenic obesity is part of obesity paradox in dementia development: evidence from a population-based cohort study. BMC medicine. 2024;22(1):133. Benjamin SR, Lima Fd, Nunes PIG, Dutra RF, Andrade GMd, Oriá RB. Advanced Biosensing Technologies: Leading Innovations in Alzheimer’s Disease Diagnosis. Chemosensors. 2025;13(6):220. Liu Y, Xia X, Zheng M, Shi B. Bio–Nano Toolbox for Precision Alzheimer's Disease Gene Therapy. Advanced Materials. 2024;36(29):2314354. Frisoni GB, Aho E, Brayne C, et al. Alzheimer's disease outlook: controversies and future directions. The Lancet. 2025;406(10510):1424-42. Sarma M, Chatterjee S. Machine Learning-Based Alzheimer’s Disease Stage Diagnosis Utilizing Blood Gene Expression and Clinical Data: A Comparative Investigation. Diagnostics. 2025;15(2):211. Ballard JL, Wang Z, Li W, Shen L, Long Q. Deep learning-based approaches for multi-omics data integration and analysis. BioData Mining. 2024;17(1):38. Bao J, Lee BN, Wen J, et al. Employing Informatics Strategies in Alzheimer's Disease Research: A Review from Genetics, Multiomics, and Biomarkers to Clinical Outcomes. Annual review of biomedical data science. 2024;7. Ahuja SK, Shrimankar DD, Durge AR. A study and analysis of disease identification using genomic sequence processing models: An empirical review. Current Genomics. 2023;24(4):207-35. Cardillo M, Katam K, Survajahala P. Advancements in Multi-Omics Research to Addressing Challenges in Alzheimer's Disease: A Systems Biology Approach Utilizing Molecular Biomarkers and Innovative Strategies. Frontiers in Aging Neuroscience. 2025;17:1591796. Abdelwahab MM, Al-Karawi KA, Semary HE. Deep learning-based prediction of Alzheimer’s disease using microarray gene expression data. Biomedicines. 2023;11(12):3304. Vemulavada S, Karthikvatsan S, Babu A, et al. Role of Artificial Intelligence in the Diagnosis and Therapy of Gynecological Disorders: Opportunities and Challenges. Translational Research in Biomedical Sciences: Recent Progress and Future Prospects. 2024:121-44. Aghdam MA, Bozdag S, Saeed F. Machine-learning models for Alzheimer’s disease diagnosis using neuroimaging data: survey, reproducibility, and generalizability evaluation. Brain Informatics. 2025;12(1):1-27. Sudharsan M, Thailambal G. Alzheimer's disease prediction using machine learning techniques and principal component analysis (PCA). Materials Today: Proceedings. 2023;81:182-90. Zhang X, Li Z, Zhang Q, Yin Z, Lu Z, Li Y. A new weakly supervised deep neural network for recognizing Alzheimer’s disease. Computers in Biology and Medicine. 2023;163:107079. He H, Yang Y, Wang L, et al. Combined analysis of single-cell and bulk RNA sequencing reveals the expression patterns of circadian rhythm disruption in the immune microenvironment of Alzheimer’s disease. Frontiers in Immunology. 2023;14:1182307. Lv H, Tan L, Miao Y, Zhou Y, Wang L, Liang Z. Machine learning identifies lactate metabolism biomarkers and deciphers immune infiltration landscapes in Parkinson’s disease. Scientific Reports. 2025;15(1):37665. Sagulkoo P, Chuntakaruk H, Suratanee A, Plaimas K, Teerakulkittipong N. Integrative bioinformatics frameworks for abdominal aortic aneurysm using GWAS meta-analysis, biological network construction, and structural modeling. Scientific Reports. 2025;15(1):22331. Zhu Y, Kong L, Han T, Yan Q, Liu J. Machine learning identification and immune infiltration of disulfidptosis‐related Alzheimer's disease molecular subtypes. Immunity, Inflammation and Disease. 2023;11(10):e1037. Alshabrmi FM, Alkhayl FFA, Rehman A. Novel drug discovery: Advancing Alzheimer's therapy through machine learning and network pharmacology. European Journal of Pharmacology. 2024;976:176661. Chhabra N, Matore BW, Lakra N, et al. Multilayered screening for multi-targeted anti-Alzheimer’s and anti-Parkinson’s agents through structure-based pharmacophore modelling, MCDM, docking, molecular dynamics and DFT: a case study of HDAC4 inhibitors. In Silico Pharmacology. 2025;13(1):16. Gan Y, Sun J, Yang D, Fang C, Zhou Z, Yin J. Exploration and validation of biomarkers for Alzheimer's disease based on GEO database. IBRO Neuroscience Reports. 2025. Alhassan HH. Integrative machine learning and molecular simulation approaches identify GSK3β inhibitors for neurodegenerative disease therapy. Scientific Reports. 2025;15(1):21632. Elghaish RA, Attallah NE, Khaled H, Mekawy AS, Elserafy M, Badr E. A computational framework for identifying cytoskeletal genes associated with age-related diseases. Scientific Reports. 2025;15(1):14590. Ferrari RR, Fantini V, Garofalo M, et al. A map of transcriptomic signatures of different brain areas in Alzheimer’s disease. International Journal of Molecular Sciences. 2024;25(20):11117. Feng G, Zhong M, Huang H, et al. Identification of UBE2N as a biomarker of Alzheimer’s disease by combining WGCNA with machine learning algorithms. Scientific Reports. 2025;15(1):6479. Xie Z, Li L, Hou W, et al. Critical role of Oas1g and STAT1 pathways in neuroinflammation: insights for Alzheimer’s disease therapeutics. Journal of Translational Medicine. 2025;23(1):182. Liu M-j, Zhu J-y, Han S-y, Zhu S-j, Yang D-d. Single-Cell Sequencing Screening of Biomarkers Associated with Lipophagy in the MPTP Model Mice of Parkinson’s Disease. Molecular Neurobiology. 2025:1-18. Wu T, Hu E, Xu S, et al. clusterProfiler 4.0: A universal enrichment tool for interpreting omics data. The innovation. 2021;2(3). Ünal B. Causality analysis for COVID-19 among countries using effective transfer entropy. Entropy. 2022;24(8):1115. Muley VY. Functional Insights Through Gene Ontology, Disease Ontology, and KEGG Pathway Enrichment. Computational Virology: Springer; 2025. p. 75-98. Ullah I, Liu K, Yamamoto T, Zahid M, Jamal A. Modeling of machine learning with SHAP approach for electric vehicle charging station choice behavior prediction. Travel Behaviour and Society. 2023;31:78-92. Mitra U, Sarkar P, Mondal J, Kundu J, editors. Enhancing Interpretability in Diabetics Prediction: A Comparative Study of SHAP, LIME and Permutation Feature Importance. 2025 AI-Driven Smart Healthcare for Society 5.0; 2025: IEEE. Nilchi AN, Dehghanian F, Vallian S, Bahreini A. Dissecting lncRNA-mRNA regulatory network in type 2 diabetes as the risk factor of pancreatic cancer. Scientific Reports. 2025;15(1):37288. Paprotny D, Terefenko P, Śledziowski J. HANZE v2. 1: an improved database of flood impacts in Europe from 1870 to 2020. Earth System Science Data. 2024;16(11):5145-70. Zamore Mn, Junior SM, Andree SL, Altunbulakli C, Lindstedt M, Levander F. Considerations and Software for Successful Immune Cell Deconvolution Using Proteomics Data. Journal of Proteome Research. 2025;24(8):3751-61. Jain S, Rick JW, Joshi RS, et al. Single-cell RNA sequencing and spatial transcriptomics reveal cancer-associated fibroblasts in glioblastoma with protumoral effects. The Journal of clinical investigation. 2023;133(5). Li Q, Qu L, Miao Y, et al. A gene network database for the identification of key genes for diagnosis, prognosis, and treatment in sepsis. Scientific Reports. 2023;13(1):21815. Kim S, Bolton EE. Pubchem: a large‐scale public chemical database for drug discovery. Open Access Databases and Datasets for Drug Discovery. 2024:39-66. Lai CH-L, Kwok APK, Wong K-C. Cheminformatic Identification of Tyrosyl-DNA Phosphodiesterase 1 (Tdp1) Inhibitors: A Comparative Study of SMILES-Based Supervised Machine Learning Models. Journal of Personalized Medicine. 2024;14(9):981. de Oliveira LHD, Cruz JN, Dos Santos CBR, de Melo EB. Multivariate QSAR, similarity search and ADMET studies based in a set of methylamine derivatives described as dopamine transporter inhibitors. Molecular Diversity. 2023:1-16. Andersson C, Bejsjö E, Magnusson SB, Carlmark A, Fuciu F, Nordwall E. Protein Engineering in Python Proteinmodifiering i Python. 2025. Burley SK, Bhikadiya C, Bi C, et al. RCSB Protein Data Bank (RCSB. org): delivery of experimentally-determined PDB structures alongside one million computed structure models of proteins from artificial intelligence/machine learning. Nucleic acids research. 2023;51(D1):D488-D508. Song T, Zhang Y, Zhu L, Zhang Y, Song J. The role of JAK/STAT signaling pathway in cerebral ischemia-reperfusion injury and the therapeutic effect of traditional Chinese medicine: A narrative review. Medicine. 2023;102(46):e35890. Wu J-W, Wang B-X, Shen L-P, et al. Investigating the Potential Therapeutic Targeting of the JAK-STAT Pathway in Cerebrovascular Diseases: Opportunities and Challenges. Molecular Neurobiology. 2025:1-27. Maiese K. The metabolic basis for nervous system dysfunction in alzheimer’s disease, parkinson’s disease, and huntington’s disease. Current neurovascular research. 2023;20(3):314-33. He R, Cheng J, Qiu Y, et al. IGF1R and FLT1 in female endothelial cells and CHD2 in male microglia play important roles in Alzheimer's disease based on gender difference analysis. Experimental Gerontology. 2024;194:112512. Engel MG. Deciphering the Role of Insulin and IGF-1 in Brain Aging, Alzheimer’s Disease, and Vascular Dementia: Albert Einstein College of Medicine; 2023. Sengupta P, Mukhopadhyay D. IGF1R/ARRB1 Mediated Regulation of ERK and cAMP Pathways in Response to Aβ Unfolds Novel Therapeutic Avenue in Alzheimer’s Disease. Molecular Neurobiology. 2025;62(6):8065-83. Johnston KG, Berackey BT, Tran KM, et al. Single-cell spatial transcriptomics reveals distinct patterns of dysregulation in non-neuronal and neuronal cells induced by the Trem2 R47H Alzheimer’s risk gene mutation. Molecular Psychiatry. 2025;30(2):461-77. Liu Q, Ling J, Li Z, Bi L. Advances in lymphoma biomarkers research based on proteomics technology. Oncology Reports. 2025;54(3):1-17. Argandona Lopez C, Brown AM. Microglial-neuronal crosstalk in chronic viral infection through mTOR, SPP1/OPN and inflammasome pathway signaling. Frontiers in immunology. 2024;15:1368465. Günaydın T, Varlı S. Computer-Aided Decision Support Systems of Alzheimer's Disease Diagnosis-A Systematic Review. Current Medical Imaging. 2025;21(1):E15734056359358. Additional Declarations The authors declare no competing interests. Supplementary Files Supplementarydocuments.zip Supplementary documents Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8200314","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":550490705,"identity":"d330c97f-8926-4cb7-ad57-8d3f4a110f2b","order_by":0,"name":"WANG ZIFU","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABBElEQVRIiWNgGAWjYBAC+/sHGx98+PG/vh/IOfAAKiqBTwubRPJhw5k9zIwzG4BaEojTkpYmzcPGzLjhAJBHpJYcY2MeoB7ja4cfAm2xyzM4wHzwNg/DtsQGXFrkzxg+nGPBw2Z2O80AqCW52OAAW7I1D8Nt3FokeIwN3vBI8JjdTgBpYU7ccIDHTJqAFjMJHjYDCePZ6R+AWuqBWvi/EdDClibJw5ZgYCCdA7LlMMgWNgJamEGBfCBB4nZOwYEEg+PFkofZjC3nGNw2xq2FERSVBxL4Z6dv/vChojqP73jzwxtvKm7L4tKCBgyAUcMMYTgSqQURmwz2xOoYBaNgFIyCYQ8AwpJa/cKFaGgAAAAASUVORK5CYII=","orcid":"https://orcid.org/0009-0007-4166-569X","institution":"Asia Pacific University of Technology \u0026 Innovation","correspondingAuthor":true,"prefix":"","firstName":"WANG","middleName":"","lastName":"ZIFU","suffix":""},{"id":550491324,"identity":"291987e7-d2c2-4c1e-a6fa-a4c31edde966","order_by":1,"name":"Hou Jinqi","email":"","orcid":"https://orcid.org/0009-0005-0256-4373","institution":"Carnegie Mellon University","correspondingAuthor":false,"prefix":"","firstName":"Hou","middleName":"","lastName":"Jinqi","suffix":""},{"id":550491325,"identity":"c650b46b-f1c1-4847-81b4-9bc5fb478633","order_by":2,"name":"Zhu Yuxuan","email":"","orcid":"","institution":"Carnegie Mellon University","correspondingAuthor":false,"prefix":"","firstName":"Zhu","middleName":"","lastName":"Yuxuan","suffix":""},{"id":550491326,"identity":"a41d8483-68a0-428e-9980-48605ea7efcb","order_by":3,"name":"Chenyun Guan","email":"","orcid":"https://orcid.org/0009-0007-7893-7556","institution":"University of Pittsburgh","correspondingAuthor":false,"prefix":"","firstName":"Chenyun","middleName":"","lastName":"Guan","suffix":""}],"badges":[],"createdAt":"2025-11-25 07:42:00","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":true,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":true},"doi":"10.21203/rs.3.rs-8200314/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8200314/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":96794498,"identity":"249a6aa8-7031-4e0d-aa86-5079d562bc49","added_by":"auto","created_at":"2025-11-26 07:16:15","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":78525,"visible":true,"origin":"","legend":"","description":"","filename":"Manuscript.docx","url":"https://assets-eu.researchsquare.com/files/rs-8200314/v1/f506e5a88b3d0aaf607d50a4.docx"},{"id":96794497,"identity":"2c42726a-c7f2-4108-a9bc-e7dad7950d0c","added_by":"auto","created_at":"2025-11-26 07:16:15","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":342,"visible":true,"origin":"","legend":"","description":"","filename":"rs8200314.json","url":"https://assets-eu.researchsquare.com/files/rs-8200314/v1/d61219b5390538345dcb70ab.json"},{"id":96794504,"identity":"646f6cbd-15b6-4051-b758-73a19cbb894f","added_by":"auto","created_at":"2025-11-26 07:16:16","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":132861,"visible":true,"origin":"","legend":"","description":"","filename":"rs82003141enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8200314/v1/4297a65db86e3e53a9e5c5a4.xml"},{"id":96794500,"identity":"ab299cdd-a48f-407d-8211-eabff07e865c","added_by":"auto","created_at":"2025-11-26 07:16:16","extension":"xml","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":130511,"visible":true,"origin":"","legend":"","description":"","filename":"rs82003141structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8200314/v1/ac3f221ed705fb5ae3aa8edb.xml"},{"id":96794502,"identity":"a573d264-6038-446c-87ab-8b25dd81694a","added_by":"auto","created_at":"2025-11-26 07:16:16","extension":"html","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":141563,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8200314/v1/c807827903113626a54f1fcb.html"},{"id":96794499,"identity":"6950f231-6612-457e-b75d-3a6c56b10148","added_by":"auto","created_at":"2025-11-26 07:16:15","extension":"jpeg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":4049061,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart of the study\u003c/p\u003e","description":"","filename":"Fig1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8200314/v1/fea918163d6dbb9f2a7720e6.jpeg"},{"id":96794501,"identity":"c23d11da-a6e8-4188-8e76-6efc3858c232","added_by":"auto","created_at":"2025-11-26 07:16:16","extension":"jpeg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":9512614,"visible":true,"origin":"","legend":"\u003cp\u003eData preprocessing and DEGs screening. using principal component analysis for batch calibration of GSE5281and GSE66333. (A) Before batch calibration. (B) after batch calibration. (C) Volcano map. (D) Heat map\u003c/p\u003e","description":"","filename":"Fig2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8200314/v1/c0056b346f2381495660f2d7.jpeg"},{"id":96794503,"identity":"39bfbbb1-2306-45da-bc4f-d3752eddb545","added_by":"auto","created_at":"2025-11-26 07:16:16","extension":"jpeg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":10986186,"visible":true,"origin":"","legend":"\u003cp\u003eDEGs and WGCNA analysis Results for display of AD. (A) Analysis of the scale-free index for various soft-threshold powers(β). (B) Analysis of the mean connectivity for various soft-threshold powers. C: threshold for WGCNA analysis. (D) Heatmap of the correlation between module characterized genes and the level of DEGs. Each cell contains the corresponding correlation and p-value. (E) Module trait relationship and module features of AD. (F) Intersection gene between module gene and DEGs in WGCNA analysis\u003c/p\u003e","description":"","filename":"Fig3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8200314/v1/8c38072a32269ec2500fe75d.jpeg"},{"id":96794510,"identity":"6c23b745-38c8-4de1-b9ae-a27b7ed608ca","added_by":"auto","created_at":"2025-11-26 07:16:16","extension":"jpeg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":35482942,"visible":true,"origin":"","legend":"\u003cp\u003eKey Gene Screening and Functional Enrichment in AD. (A) Top 15 key genes from the PPI network. (B) GO enrichment. (C) KEGG pathway enrichment. (D) DO enrichment.\u003c/p\u003e","description":"","filename":"Fig4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8200314/v1/5deed73e4ae3f9e0c9e9415b.jpeg"},{"id":96794511,"identity":"49d0b5eb-64ba-4bc7-8322-f990665de9c7","added_by":"auto","created_at":"2025-11-26 07:16:17","extension":"jpeg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":39095578,"visible":true,"origin":"","legend":"\u003cp\u003e127 machine learning algorithm combinations evaluated via 10-fold cross-validation and SHAP screening key genes and verification evaluation of hub genes. (A) 127 machine learning algorithm combinations evaluated via 10-fold cross-validation. (B) SHAP Analysis of the Best Model. (C) The ROC curve and expression level of the characteristic gene screening hub genes.\u003c/p\u003e","description":"","filename":"Fig5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8200314/v1/4feea3ec6c3e501f4dda3e22.jpeg"},{"id":96915603,"identity":"7b95d8ae-24e6-40b4-af2c-b55e00067932","added_by":"auto","created_at":"2025-11-27 14:07:26","extension":"jpeg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":11759775,"visible":true,"origin":"","legend":"\u003cp\u003eTranscriptional regulatory network of hub genes IGF1R and SPP1 and their association with immune cell infiltration. (A) TF–gene regulatory network of hub genes IGF1R and SPP1. (B) Correlations between IGF1R expression and infiltrating immune cell subsets. (C) Correlations between SPP1 expression and infiltrating immune cell subsets. (D) Network visualization of hub genes and significantly correlated immune cell subsets.\u003c/p\u003e","description":"","filename":"Fig6.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8200314/v1/8d843ba0a0f7259c783d2d7a.jpeg"},{"id":96794508,"identity":"62b7b406-a15a-4d10-b110-8ce786459cc0","added_by":"auto","created_at":"2025-11-26 07:16:16","extension":"jpeg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":18826388,"visible":true,"origin":"","legend":"\u003cp\u003eSingle-cell transcriptomic analysis reveals cellular heterogeneity and expression patterns of IGF1R and SPP1）.(A) A Distribution of single-cell RNA-seq quality control metrics (nFeature_RNA, nCount_RNA and percent.mt) in control and treated groups. (B) Top loading genes of principal components PC1–PC4. (C) t-SNE visualization of control and treated cells and annotation of major cell subpopulations. (D) Slingshot-inferred pseudotime trajectory of cell differentiation. (E) Spatial distribution and differential expression of IGF1R and SPP1 in single-cell profiles between control and treated groups.\u003c/p\u003e","description":"","filename":"Fig7.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8200314/v1/8071dbd0695dfc0081916d38.jpeg"},{"id":96794506,"identity":"8c970e21-e595-47cc-b53b-22d500f21a74","added_by":"auto","created_at":"2025-11-26 07:16:16","extension":"jpeg","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":9601249,"visible":true,"origin":"","legend":"\u003cp\u003eScreening of candidate small-molecule compounds for molecular docking based on gene–drug enrichment and ADMET/drug-likeness analysis. (A) Significantly enriched drug ingredients associated with candidate genes (n = 176, P \u0026lt; 0.05). (B) ADMET and drug-likeness profiling of 445 candidate compounds.\u003c/p\u003e","description":"","filename":"Fig8.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8200314/v1/aa3839b170ea627116d5a59e.jpeg"},{"id":96794509,"identity":"fab24eb0-f944-4f40-b087-235fd3a31261","added_by":"auto","created_at":"2025-11-26 07:16:16","extension":"jpeg","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":25668863,"visible":true,"origin":"","legend":"\u003cp\u003eBinding poses and key interactions of top-ranked candidate small molecules with target proteins 3D94 and 5FXS. (A) Compound C1CN(CCC12C(=O)NCN2C3=CC=CC=C3)CCCC(=O)C4=CC=C(C=C4)F Docking conformation and its hydrophobic/hydrogen bonding interactions with the target protein 3D94. (B) Compound CN1CCN(CC1)C2=NC3=C(C=CC(=C3)Cl)NC4=CC=CC=C42 Docking conformation and its hydrophobic, hydrogen- and halogen-bonding interactions with target protein 5FXS. (C) Docking conformation and predominantly hydrophobic binding mode of the compound CC1(CCC(C2=C1C=CC(=C2)NC(=O)C3=CC=C(C=C3)C(=O)O)(C)C)C with 5FXS. (D) Compound COC1=C2C3=C(C(=O)CC3)C(=O)OC2=C4C5C=COC5OC4=C1 Docking conformation with 5FXS and its hydrophobic-dominant binding interactions.\u003c/p\u003e","description":"","filename":"Fig9.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8200314/v1/f908dccebafe3b5e9cbf93b2.jpeg"},{"id":96794512,"identity":"c2f2117c-cb03-4bea-a733-d30f25a9c62f","added_by":"auto","created_at":"2025-11-26 07:16:17","extension":"zip","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":89280279,"visible":true,"origin":"","legend":"\u003cp\u003eSupplementary documents\u003c/p\u003e","description":"","filename":"Supplementarydocuments.zip","url":"https://assets-eu.researchsquare.com/files/rs-8200314/v1/135456d9b756429c88c02479.zip"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eIntegrated Bioinformatics and Ensemble Learning Reveal Diagnostic Modeling and Drug Discovery in Alzheimer’s Disease\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"Introduction","content":"\u003cp\u003eAlzheimer\u0026apos;s disease (AD), a neurodegenerative disease, is characterized by progressive cognitive decline, memory impairment and irreversible neuronal loss (1). It has become the most common type of dementia in the elderly population and constitutes a major public health challenge worldwide. Epidemiologic data show that the number of people with dementia has reached approximately 55 million worldwide and is projected to increase to 139 million by 2050, continuing to increase the burden on healthcare and social security systems worldwide (2). In terms of the economic burden of disease, annual healthcare expenditures related to dementia have exceeded US$1.3 trillion and are expected to exceed US$2.8 trillion by 2030 (3). In addition, AD-related mortality rates in the United States have risen by more than 140% between 2000 and 2021, underscoring the urgency of upgrading the diagnostic and therapeutic capacity of the disease (4).\u003c/p\u003e\n\u003cp\u003eDespite significant progress in the study of AD pathogenesis, early diagnosis of the disease remains a challenge, with the main obstacle being the lack of specificity of clinical symptoms in the early stages of the disease, which often leads to delayed recognition and lagging interventions (5). Accurate and timely diagnosis is essential for the development of clinical strategies and patient prognosis. In this process, reliable biomarkers play a central role in early screening, precise staging, disease progression monitoring, and identification of therapeutic targets.. However, the molecular regulatory network of AD has not been fully elucidated, and there is still a lack of consensus biomarkers with sufficient diagnostic efficacy that can be generalized to clinical applications (6).\u003c/p\u003e\n\u003cp\u003eMachine learning (ML) methods have shown unique advantages in parsing high-dimensional complex biological data (7). Algorithms such as random forests, support vector machines, neural networks and deep learning are widely used in genomics and bioinformatics to mine key biological features from multi-omics data (8). By integrating large-scale molecular phenotyping and clinical information, ML is able to effectively identify nonlinear association patterns, and thus promote the systematic discovery of disease-related biomarkers and predictive models (9, 10). The deep integration of bioinformatics and ML provides a powerful framework for the systematic identification of key regulatory genes in diseases, which has demonstrated significant value in disease molecular typing, prognosis prediction, and development of targeted therapeutic strategies (11, 12).\u003c/p\u003e\n\u003cp\u003eIn the field of AD research, Base-ML analysis methods integrating multi-omics data have made substantial progress in diagnostic marker identification and potential therapeutic target mining. However, existing studies still have obvious limitations in feature selection strategies\u0026mdash;especially the lack of systematic evaluation of multiple algorithm combinations to ensure the robustness and reproducibility of diagnostic features (13, 14). To address this challenge, this study proposes a new research paradigm: to construct a more reliable screening system for AD-related genetic features by integrating a large number of supervised machine learning models (15, 16). We hypothesize that this multi-model integration strategy can identify key gene features with high stability and clinical applicability. On this basis, we further explore the multidimensional features of these diagnostic genes, including their regulatory roles in the immune microenvironment, potential functional pathways, and spatial expression patterns (17). These analyses are expected to provide new insights into the elucidation of AD pathogenesis and provide a theoretical basis for targeted drug development.\u003c/p\u003e\n\u003cp\u003eIn this study, we identified differentially expressed genes (DEGs) and WGCNA-derived module genes, and the intersecting genes were screened using 12 machine learning algorithms and 127 combinations selection techniques to determine reliable diagnostic biomarkers for AD (18, 19). The diagnostic performance of these biomarkers was validated through nomogram construction, calibration analysis, and clinical decision curve analysis. Immune cell infiltration and regulatory features were analyzed to elucidate immune-related mechanisms (20). Spatial transcriptomics was applied to confirm the spatial distribution of key genes in AD brain tissue. Finally, therapeutic compounds were identified through drug enrichment analysis, ML\u0026ndash;based ADMET profiling, and Smina automated blind docking to prioritize high-affinity compound\u0026ndash;target complexes for potential AD treatment (21, 22).\u003c/p\u003e"},{"header":"Methods","content":"\u003ch3\u003e\u003cstrong\u003eStudy design\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003eBased on the GEO database, this study identified AD-related DEGs (23). Hub genes screening was performed using protein-protein interaction analysis and an ensemble MLapproach, with key candidates validated through transcriptional regulatory networks and spatial transcriptomics. Promising therapeutic agents were then identified via drug-gene interaction analysis, ML-based ADMET profiling, and molecular docking for high-affinity complexes screening (Fig. 1) (24).\u003c/p\u003e\n\u003cp\u003e============================\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFig. 1\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e============================\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFig. 1\u003c/strong\u003e Flowchart of the study\u003c/p\u003e\n\u003ch3\u003e\u003cstrong\u003eData sources used for analysis\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003eFive distinct gene expression profiles (GSE5281, GSE66333, GSE110226, GSE28146, GSE29378, and GSE29378), as well as the spatial transcriptomic dataset GSE147047, were retrieved from the GEO database. GSE5281 and GSE66333 were designated as the training group for this study, comprising a sample of 91AD patients and 78 non-AD controls, while GSE110226, GSE28146, and GSE29378 served as the validation and independent testing groups. Following logarithmic transformation, each dataset was normalized using the \u0026ldquo;normalizeBetweenArrays\u0026rdquo; function of the R package \u0026ldquo;limma\u0026rdquo; (version 3.66.0) (25).\u003c/p\u003e\n\u003cp\u003eTo eliminate batch effects between datasets, the expression matrices were corrected and merged using the \u0026ldquo;sva\u0026rdquo; package (version 3.58.0). Principal component analysis (PCA) \u0026nbsp;is mainly used for data dimensionality reduction and feature extraction, and reduce the complexity of data (15). The merged matrix was then used for differential expression analysis with \u0026ldquo;limma,\u0026rdquo; and the filtering criteria for DEGs were set as |log₂Fold-Change| \u0026gt; 0.585 and false discovery rate (FDR) \u0026lt; 0.05 (26).\u003c/p\u003e\n\u003ch3\u003e\u003cstrong\u003eWeighted gene co-expression network analysis (WGCNA) of DEGs\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003eTo build a WGCNA for DEGs, we utilized the R package \u0026quot;WGCNA\u0026quot; (version 1.73.0) (27). The objective was to identify gene modules significantly linked to AD. Quality control and sample clustering were performed on the gene expression data to exclude abnormal samples. The soft threshold function was used to determine the optimal power value, construct the neighbor-joining matrix between genes, and further calculate the topological overlap matrix (TOM). Hierarchical clustering was performed based on the dissimilarity of the TOM matrix, and gene co-expression modules were identified by a dynamic shear tree algorithm. Modules with highly correlated feature genes were merged to obtain final independent modules (28). Key modules with significant phenotypic associations were screened by calculating the correlation between module feature genes and clinical traits, and the genes that intersect between the significant coexpressed genes identified through WGCNA and DEG were analyzed for protein-protein interactions (PPI) analysis (27).\u003c/p\u003e\n\u003ch3\u003e\u003cstrong\u003eConstruction of PPI networks\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003eIn this study, we utilized the STRINGdb package (version 2.22.0) to obtain protein interactions from the \u0026quot;org.Hs.eg.db\u0026quot; package (versioin 3.22.0) (29). This network specifically focused on the intersection of DEGs and those genes identified through WGCNA. combined_score\u0026ge;400 was set as the threshold for reliable interactions. Gene ID conversion was performed using the bitr function of the \u0026quot;clusterProfiler\u0026quot; package (version 4.18.1) to convert gene symbols to ENTREZID for database mapping\u0026nbsp;(30).\u003c/p\u003e\n\u003cp\u003eThe network was constructed using the \u0026quot;igraph\u0026quot; package (versioin 2.2.1), which computes five network centrality metrics: degree centrality, betweenness, closeness, eigen_centrality, and eigen_centrality. centrality) and page_rank algorithm (31). A standardized weighted scoring model (weight coefficients of 0.25, 0.25, 0.20, 0.15, 0.15, respectively) was used to calculate the composite hub scores and screen the top 15 key genes.\u003c/p\u003e\n\u003cp\u003eVisual analysis was performed using the ggraph package (version 2.2.2) for network layout, and the node positions were optimized using the stress layout. In the node attribute mapping, size was positively correlated with degree centrality (rescale function normalized to the range of 3-15), and color was divided into blue, orange, and red gradients according to the median centrality tertiles. Edge attribute settings for width and transparency were positively correlated with interactions scores (rescale to the 0.3-2 and 0.3-0.8 ranges). Chord diagrams were drawn using the circlize package (version 0.4.16), which demonstrates the top 30% of high-intensity interactions via the chordDiagram function, and the represented genes were considered as key genes for subsequent Gene Ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG), and Disease Ontology (DO) analyses (32).\u003c/p\u003e\n\u003ch3\u003e\u003cstrong\u003eFunctional Enrichment Analysis of Hub Genes: GO, KEGG and DO Integration\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003eGene function enrichment analysis is an important method for elucidating the biological significance of gene sets (32). Annotation related to human diseases, biological functions, and signaling pathways helps reveal the connections and roles among overlapping genes associated with Alzheimer\u0026apos;s disease (AD). The target gene set was systematically functionally annotated using the clusterProfiler, gene IDs were converted using the org.Hs.eg.db database, and three levels of enrichment analysis were performed:\u003c/p\u003e\n\u003col class=\"decimal_type\"\u003e\n \u003cli\u003eGO analysis, including biological processes (BP), molecular functions (MF), and cellular components (CC), to reveal the functional characteristics of genes;\u003c/li\u003e\n \u003cli\u003eKEGG pathway analysis to identify key signaling pathways and metabolic pathways involved by genes;\u003c/li\u003e\n \u003cli\u003eDO analysis to find the correlation between genes and human diseases.\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eEnrichment results were presented using p \u0026lt; 0.05 and FDR \u0026lt; 0.05 as significance thresholds, and visualized using various methods such as bubble charts, bar charts, network diagrams, and pie charts.\u003c/p\u003e\n\u003ch3\u003e\u003cstrong\u003eModel Construction Method Based on Integrated ML and SHAP\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003eKey genes for AD interactions identified by PPI network analysis algorithms were subsequently integrated into different machine learning frameworks. In this study, we used to build a two-stage machine learning framework. In the first stage, 12 algorithms were applied for initial screening of features, including regularization methods (Lasso, Ridge, ElasticNet), integrated learning models (Random Forest, GBM, XGBoost), generalized linear models (Stepglm, glmBoost, plsRglm), and pattern recognition models (SVM, LDA, Naive Bayes). Hyperparameter optimization was performed by grid search combined with 10-fold cross-validation, where Random Forest tuned the mtry parameters (2-10), and XGBoost optimized the learning rate (0.01-0.3), maximum tree depth (3-7), and regularization parameters (33).\u003c/p\u003e\n\u003cp\u003eIn the second stage, a stacked generalization strategy is used to integrate models from the initial screening results, construct 127 algorithm combinations, and evaluate the performance under a rigorous cross-validation framework, with AUC as the main evaluation metric, and validate the generalization ability of the models in multiple independent cohorts. SHAP interpretability analysis is implemented for the best performing models, and the permSHAP method is used to calculate the feature contribution and generate 95% confidence intervals by Bootstrap resampling (100 times) to ensure the stability of the results (34).\u003c/p\u003e\n\u003cp\u003eThe evaluation of feature importance is based on a multi-dimensional index system, including the average absolute SHAP value, the cumulative importance curve (with a double threshold of 80% and 90%), and the statistical significance test. The swarm plot shows the distribution of feature importance, the waterfall plot analyzes the prediction mechanism of individual samples, and the dependency graph reveals the nonlinear relationship between features and predicted output. The core AD diagnostic genes with statistical significance and biological interpretability were selected.\u003c/p\u003e\n\u003ch3\u003e\u003cstrong\u003eValidation analysis of hub genes\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003eTo assess the diagnostic performance of key diagnostic genes screened by 127 integrated algorithms in AD, the R software package \u0026quot;pROC\u0026quot; (version 1.19.0.1) was used to perform subject work characterization curves (ROCs) analysis and to calculate the area under the curve (AUC) as a measure of the ability of each gene to discriminate AD from control samples (35). Ability. Genes with AUC values greater than 0.60 in both the training and validation cohorts were considered to be AD-related biomarkers with stable diagnostic potential.\u003c/p\u003e\n\u003cp\u003eSubsequently, multigene logistic regression models were constructed using the R software package \u0026quot;glmnet\u0026quot; (version 4.1.10) to evaluate the combined diagnostic efficacy of key genes in the training cohort, validation cohort, and spatial transcriptome data. Meanwhile, \u0026quot;PerformanceAnalytics\u0026quot; (version 2.0.8) was used to perform correlation analysis and visualization of the hub genes in the validation cohort to reveal the interrelationships among genes and potential co-regulatory features.\u003c/p\u003e\n\u003ch3\u003e\u003cstrong\u003eAnalysis of transcription factor-hub genes regulatory networks\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003eTo investigate the upstream regulatory mechanisms of hub genes, this study was based on the TRUST v2 database (www.grnpedia.org/trust, accessed on November 1, 2025) to screen transcription factors with potential interactions with hub genes (36). Based on the predictive confidence scores provided by the database, the top 10 transcription factors with the highest scores were selected as candidate key regulators for subsequent analysis. Transcription factors that were up-regulated in AD tissues were regarded as likely to exert a positive regulatory effect on the hub genes. Subsequently, the expression levels of the candidate transcription factors and the hub genes were correlated to assess their potential regulatory relationships, and a schematic diagram of the transcription factor-target gene regulatory network was constructed accordingly.\u003c/p\u003e\n\u003ch3\u003e\u003cstrong\u003eCorrelation between Hub Genes and Immune Cells\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003eTo assess the immune microenvironment characteristics of Alzheimer\u0026apos;s disease (AD) tissues, this study used CIBERSORT-based deconvolution (https://cibersortx.stanford.edu/) to perform immune cell deconvolution analysis on Discovery data (37). CIBERSORTx, based on a linear support vector regression (SVR) model, estimates the proportion of cell infiltration, including 22 immune cell types. Before analysis, the expression matrix was quantile-normalized and its gene intersection was matched with an LM22 reference matrix. Samples with p \u0026lt; 0.05 were selected as reliable results using a permutation test (permutation = 1000), and Spearman correlation analysis was used to assess the association between pivotal genes and key immune cell types to identify AD-related changes in immune infiltration. The analysis results were visualized using box plots, correlation plots, and other methods.\u003c/p\u003e\n\u003ch3\u003e\u003cstrong\u003eSingle-cell and spatial transcriptomics validation analysis\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003eIn order to further validate the expression pattern of hub genes systematically, this study integrated single-cell RNA sequencing and spatial transcriptomics data for multilevel analysis (38). The scRNA-seq data were acquired from GEO and processed for quality control, normalization and clustering using Seurat (version 5.3.1). The quality control thresholds were set at nFeature_RNA \u0026lt; 100 and percent.mt \u0026gt; 15%; LogNormalize (scale factor = 10,000) was used for normalization, and 1,500 highly variable genes were screened for downstream analysis. This was followed by initial PCA downscaling and correction for batch effects using harmony (version 1.2.4); shared nearest neighbor (SNN) based mapping and clustering using Louvain\u0026apos;s algorithm (resolution 0.6). Cell type annotation was done by SingleR, using HumanPrimaryCellAtlas as the reference database.\u003c/p\u003e\n\u003cp\u003eFor the screened hub genes, t-SNE, violin plot and dot plot were used to evaluate their expression distribution and cell specificity in different cell types. In order to further validate the spatial localization and expression patterns of hub genes, 10x Genomics Visium spatial transcriptome data were integrated, and the raw data were pre-processed by Space Ranger to map the single-cell annotation results to the tissue spatial coordinates, and the spatial expression characteristics of the hub genes were demonstrated by spatial feature maps and analyzed in conjunction with the images of tissue slices. The spatial expression characteristics of Hub genes are shown in the spatial feature map, and their localization and enrichment in specific tissue regions are analyzed in conjunction with the tissue slices, thus realizing the spatial validation of the single-cell analysis results.\u003c/p\u003e\n\u003ch3\u003e\u003cstrong\u003eDrug enrichment analysis and candidate compound screening\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003eAfter obtaining the pivotal genes, this study conducted drug enrichment analysis based on the DSigDB database to evaluate potential drug-gene interactions and screen candidate therapeutic compounds (39). Drug target enrichment analysis was performed using the enricher function of the \u0026ldquo;clusterProfiler\u0026rdquo; toolkit with a significance threshold set at p \u0026lt; 0.05 and corrected p \u0026lt; 0.05 to reduce the false positive rate. Significantly enriched drug entries were visualized by bubble plots to present the strength of association between drugs and gene sets.\u003c/p\u003e\n\u003cp\u003eBased on the candidate compounds obtained from the enrichment analysis, the corresponding SMILES structural formulas were further obtained in this study via the PubChem database (https://pubchem.ncbi.nlm.nih.gov/) to provide key inputs for subsequent virtual screening and drug repositioning studies (40).\u003c/p\u003e\n\u003ch3\u003e\u003cstrong\u003eComputational Prediction of Drug-likeness\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003eIn the present study, a systematic pharmacophore evaluation of compounds obtained from the PubChem database was carried out using computational chemistry. Based on the RDKit (version 2024.3.5) toolkit in the Python (version 3.8.15) environment, we calculated key molecular descriptors, including molecular weight, lipid-water partition coefficient, and topopolar surface area, and constructed a seven-dimensional weighted scoring system (41). The system integrates the following evaluation dimensions: the Lipinski rule (15% weight) to assess oral absorption properties, the Veber rule (10%) to determine bioavailability, the Ghose rule (10%) to validate drug-like criteria, the quantitative drug similarity index QED (20%) to comprehensively evaluate drug similarity, the synthetic accessibility score SA Score (15%) to measure synthetic feasibility, ADMET Characterization (20%) to predict pharmacokinetic profiles, and Structure Alert Screening (10%) to identify potentially toxic moieties (42). After obtaining the comprehensive scores through weighted calculation, the compounds were classified into five grades of A (Excellent), B (Good), C (Medium), D (Poor), and F (Unqualified), which provided a reliable basis for prioritizing the compounds for subsequent molecular docking studies.\u003c/p\u003e\n\u003ch3\u003e\u003cstrong\u003eIntelligent molecular docking and multidimensional validation\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003eTo validate the binding potential of the lead compounds obtained from the enrichment analysis, the molecular docking method based on Python\u0026apos;s Smina (version 2020.12.10) library was used in this study for systematic validation (43). After obtaining the crystal structure of the target protein (PDB ID: [3D94, 5FXS]) from the RCSB PDB database, the protein structure was extracted and processed using the Python\u0026apos;s MDAnalysis (version 2.2.0) library, followed by protein protonation and PDBQT formatting via Open Babel at various physiological pH conditions (6.4, 7.4, 8.4) \u0026nbsp;(44). The ligand molecules are based on SMILE. For ligand molecules, the PDBQT files for docking were generated based on SMILES expressions, through 3D structure generation, force field optimization and format conversion. A dual strategy based on the automatic detection of eutectic ligands and the designation of key active residues (e.g., HIS:114, ASP:165, GLU:166) was used to define the binding pockets during the docking process, and different exhaustion levels (8, 16, 32, 64) were systematically evaluated to balance the efficiency and adequacy of the conformational search. In order to improve the computational efficiency, an intelligent task distribution system is constructed, which can automatically detect the computational resources and optimize the parallel processing strategy, and each task covers the complete pre-processing-docking-post-processing flow, and is equipped with task-level timeout protection and automatic retry mechanism, which guarantees the stable operation of large-scale computational tasks. The docking results are verified by RMSD redocking, which evaluates the reliability of docking by comparing the deviation between the initial and redocked conformations, and lays the foundation for future cell and animal experiments.\u003c/p\u003e"},{"header":"Results","content":"\u003ch3\u003e\u003cstrong\u003eA total of 2534 DEGs associated with AD were identified\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003eTwo datasets (GSE5281 and GSE66333) were selected as training models. We observed baseline batch differences between the included data sets (Fig. 2A). In order Immune microenvironment of AD brain tissues, PCA was applied, all batch differences are eliminated (Fig. 2B). \u0026nbsp;A total of 2534 DEGs were identified with log2FC \u0026gt; 0.585 and FDR \u0026lt; 0.05 between the AD group (n=91) and the Control group (n=78), including 1325 up-regulated genes and 1209 down-regulated genes (Fig. 2C-D) (S1).\u003c/p\u003e\n\u003cp\u003e============================\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFig. 2\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e============================\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFig. 2\u003c/strong\u003e Data preprocessing and DEGs screening. using principal component analysis for batch calibration of GSE5281and GSE66333. (A) Before batch calibration. (B) after batch calibration. (C) Volcano map. (D) Heat map\u003c/p\u003e\n\u003ch3\u003e\u003cstrong\u003eWGCNA identified key modules on DEGs that are closely related to AD\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003eThe goal of WGCNA is to identify co-expressed gene modules closely associated with AD. WGCNA focuses on network mining of genes based on expression pattern consistency. This method first calculates the Pearson correlation coefficient between any two genes, then constructs a scale-free weighted adjacency matrix using a soft threshold, generating a TOM and performing hierarchical clustering to identify co-expressed modules.\u003c/p\u003e\n\u003cp\u003eThe weighted network constructed based on the training set expression matrix shows that, with the determination of the soft threshold, both the adjacency matrix and the topological overlap matrix exhibit good scale-free network characteristics (Fig. 3A\u0026ndash;C). Dynamic tree pruning and average clustering methods ultimately identified a significant module (Fig. 3D), where the MEbrown module (R = 0.62, p \u0026lt; 0.05) showed the strongest correlation with the AD phenotype, containing 1854 genes (Figure 3E) (S2). Taking the intersection of these modules with DEGs yielded 848 candidate genes, providing a basis for subsequent PPI screening (Fig. 3F).\u003c/p\u003e\n\u003cp\u003e============================\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFig. 3\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e============================\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFig. 3\u003c/strong\u003e DEGs and WGCNA analysis Results for display of AD. (A) Analysis of the scale-free index for various soft-threshold powers(\u0026beta;). (B) Analysis of the mean connectivity for various soft-threshold powers. C: threshold for WGCNA analysis. (D) Heatmap of the correlation between module characterized genes and the level of DEGs. Each cell contains the corresponding correlation and p-value. (E) Module trait relationship and module features of AD. (F) Intersection gene between module gene and DEGs in WGCNA analysis\u003c/p\u003e\n\u003ch3\u003e\u003cstrong\u003eIdentification of biomarkers for AD using PPI networks and multifunctional enrichment analysis\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003eA high-confidence PPI network was constructed by integrating DEGs and WGCNA genes, from which the top 15 key genes were screened based on their overall centrality scores (Fig. 4A) (S3). To elucidate the potential roles of these genes in AD, these genes were subsequently analyzed for functional enrichment of GO, KEGG and DO. The results showed that glial cell differentiation, gliogenesis in the pathogenesis of AD (Fig. 4B). the GO analyzed gene set was significantly enriched for a variety of biological processes, including endoplasmic reticulum lumen, vesicle lumen, chromatin DNA binding, integrin binding, glial cell differentiation, gliogenesis (Fig. 4C). Further KEGG pathway analysis showed that these genes were significantly enriched in Adherens junction, FoxO signaling pathway, Human papillomavirus infection, JAK-STAT signaling pathway, MicroRNAs in cancer, Notch signaling pathway signaling pathway. DO analysis showed that these genes were significantly associated with brain ischemia, cognitive disorder, retinal disease, and endocrine gland cancer (Fig. 4D).\u003c/p\u003e\n\u003cp\u003e============================\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFig. 4\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e============================\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFig 4\u003c/strong\u003e Key Gene Screening and Functional Enrichment in AD. (A) Top 15 key genes from the PPI network. (B) GO enrichment. (C) KEGG pathway enrichment. (D) DO enrichment.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConstructing a diagnostic model of AD by machine learning\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eBased on the 15 candidate key genes screened by the preliminary PPI network analysis, this study further integrated 12 machine learning algorithms and constructed a total of 127 predictive models with different combinations to screen the most robust diagnostic classifiers. These models were all built in the training cohort and systematically evaluated on 4 external validation datasets, and the whole process was completed based on a cross-validation framework.\u003c/p\u003e\n\u003cp\u003eAmong the 127 models constructed, the RF integration algorithm demonstrated the highest average AUC in both the training and test sets (training set AUCs with significant overfitting have been excluded), with an average AUC of 0.957 (Fig. 7A), and was therefore identified as the optimal classification model. Subsequently, we performed feature importance parsing on the RF model based on the SHAP (SHapley Additive exPlanations) method, which was used to quantify the contribution of each gene to the classification results. This analysis finally filtered out 11 feature genes: APP, CREBBP, SUZ12, CXCR4, SNAP23, BCL6, NFKBIA, IGF1R, CEBPB, SPP1, and EZR (Fig. 7B) .\u003c/p\u003e\n\u003cp\u003eTo further validate the generalization ability and diagnostic efficacy of different feature selection strategies and their constructed models, we apply each model to an external validation cohort and compare their classification performance in the validation set. As shown in Fig. 7C, the AUC values of the IGF1 and SPP1 correlation models were consistently greater than 0.70 in multiple validation datasets (GSE110226, GSE28146, GSE29378, GSE29378), suggesting that these two genes are hub genes with stable diagnostic value (S4).\u003c/p\u003e\n\u003cp\u003e============================\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFig. 5\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e============================\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFig 5\u003c/strong\u003e 127 machine learning algorithm combinations evaluated via 10-fold cross-validation and SHAP screening key genes and verification evaluation of hub genes. (A) 127 machine learning algorithm combinations evaluated via 10-fold cross-validation. (B) SHAP Analysis of the Best Model. (C) The ROC curve and expression level of the characteristic gene screening hub genes.\u0026nbsp;\u003c/p\u003e\n\u003ch3\u003e\u003cstrong\u003eAnalysis of upstream transcriptional regulation of IGF1R and SPP1 and their association with the immune microenvironment\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003eIn order to deeply investigate the upstream transcriptional regulatory mechanisms of hub genes IGF1R and SPP1 and their roles in the tumor immune microenvironment, we systematically constructed a transcription factor-target gene regulatory network (Figure 7A). The results showed that IGF1R was at the center of this regulatory network and was co-regulated by several key transcription factors, including TP53, SP1, POU2F1, NR3C1, KLF6, and WT1, suggesting that IGF1R is subjected to a complex and multi-level regulation at the transcriptional level and may play a central role in the disease process.\u003c/p\u003e\n\u003cp\u003eFurther, we evaluated the association between the expression levels of IGF1R and SPP1 and the degree of infiltration of various types of immune cells in the tumor microenvironment by immune infiltration analysis (Fig. 6B-D). IGF1R expression was moderately negatively correlated with the infiltration levels of some naive and memory B cells and moderately negatively correlated with the infiltration levels of most innate immune cells (e.g., monocytes and T cells gamma delta) and correlated weakly (Fig. 6B). On the contrary, SPP1 expression showed a significant positive correlation with the infiltration degree of various myeloid-derived immune cells, such as M0/M1/M2 macrophages and neutrophils, and a negative correlation with B cell memory (Fig. 6C). As shown in Figure 6D, the differential association pattern between IGF1R and SPP1 with different immune cell subpopulations was further visualized.\u003c/p\u003e\n\u003cp\u003e============================\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFig. 6\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e============================\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFig 6\u0026nbsp;\u003c/strong\u003eTranscriptional regulatory network of hub genes IGF1R and SPP1 and their association with immune cell infiltration. (A) TF\u0026ndash;gene regulatory network of hub genes IGF1R and SPP1. (B) Correlations between IGF1R expression and infiltrating immune cell subsets. (C) Correlations between SPP1 expression and infiltrating immune cell subsets. (D) Network visualization of hub genes and significantly correlated immune cell subsets.\u003c/p\u003e\n\u003ch3\u003e\u003cstrong\u003eSingle-cell transcriptomic analysis reveals cellular heterogeneity and expression patterns of IGF1R and SPP1\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003eTo further validate the expression characteristics of the aforementioned hub genes and their cellular origin at the single-cell transcriptomic level, we implemented a systematic quality control analysis of single-cell RNA sequencing data. By assessing the number of genes detected (nFeature_RNA), UMI count (nCount_RNA), and mitochondrial gene percentage (\u003ca href=\"http://percent.mt\" target=\"_blank\"\u003epercent.mt\u003c/a\u003e) in each cell, it was confirmed that the cellular quality of the control group and the treatment group was good overall and met the criteria for subsequent analysis (Fig. 7A). Subsequently, principal component analysis was performed based on the highly variable genes obtained from the screening, and the highly loaded genes corresponding to principal components 1 to 4 (PC1-PC4) were extracted for resolving the major transcriptional heterogeneity among cellular subpopulations (Fig. 7B).\u003c/p\u003e\n\u003cp\u003eThe results of t-SNE downscaling visualization showed that the cells were classified into major subtypes such as neuroepithelial cells versus neurons. Although there was a partial overlap between the control and treated cells in the two-dimensional space, there was a significant difference in the proportion of their cellular composition (Fig. 7C). Further, we applied the Slingshot algorithm to construct cell differentiation trajectories, and the proposed time-series analysis revealed a potential pathway of gradual differentiation from neuroepithelial cell to neuron state (Fig. 7D).\u003c/p\u003e\n\u003cp\u003eOn this basis, hub genes (IGF1R vs. SPP1) were mapped to single-cell expression profiles, and their expression dynamics were compared between the two groups. The results showed that IGF1R showed a clear pattern of high expression in specific cell subpopulations, and the overall expression level was significantly upregulated in the treated groups; in contrast, SPP1 was expressed at a lower level in most cells, with no significant trend in the difference between groups (Fig. 7E).\u003c/p\u003e\n\u003cp\u003e============================\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFig. 7\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e============================\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFig 7\u003c/strong\u003e Single-cell transcriptomic analysis reveals cellular heterogeneity and expression patterns of IGF1R and SPP1）.(A) A Distribution of single-cell RNA-seq quality control metrics (nFeature_RNA, nCount_RNA and percent.mt) in control and treated groups. (B) Top loading genes of principal components PC1\u0026ndash;PC4. (C) t-SNE visualization of control and treated cells and annotation of major cell subpopulations. (D) Slingshot-inferred pseudotime trajectory of cell differentiation. (E) Spatial distribution and differential expression of IGF1R and SPP1 in single-cell profiles between control and treated groups.\u003c/p\u003e\n\u003ch3\u003e\u003cstrong\u003eScreening of potential therapeutic small molecules based on gene\u0026ndash;drug enrichment and ADMET profiling\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003eTo uncover potential therapeutic drugs based on hub genes at the molecular level, we systematically carried out drug enrichment and drug-like assessment analysis. The enrichment results based on the public drug-gene interaction database showed that hub genes were significantly associated with 176 drug components (P \u0026lt; 0.05), suggesting that these drugs may be involved in the disease process by modulating key genes (Fig. 8A) (S5).\u003c/p\u003e\n\u003cp\u003eBased on this, we collected and organized 445 relevant small molecule compounds from the PubChem database and conducted a multidimensional systematic evaluation of their ADMET properties and drug-like properties, covering key parameters such as oral bioavailability, lipid solubility, polarity, molecular weight, metabolic stability, and potential toxicity. According to the comprehensive scoring system, all the compounds were categorized into six grades from A to F, in which grade A represents the candidate molecules with optimal pharmacokinetic and drug-like properties. Through this evaluation system, a total of 37 A-grade compounds were screened, and they were identified as the key candidate ligands for subsequent molecular docking studies (Fig. 8B) (S6).\u003c/p\u003e\n\u003cp\u003e============================\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFig. 8\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e============================\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFig 8\u003c/strong\u003e Screening of candidate small-molecule compounds for molecular docking based on gene\u0026ndash;drug enrichment and ADMET/drug-likeness analysis. (A) Significantly enriched drug ingredients associated with candidate genes (n = 176, P \u0026lt; 0.05). (B) ADMET and drug-likeness profiling of 445 candidate compounds.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003ch3\u003e\u003cstrong\u003eMolecular docking-based analysis of binding modes and interactions between candidate small molecules and target proteins\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003eTo further evaluate the binding ability and mode of action between the candidate class A small molecule compounds and the target proteins, we carried out Cartesian product molecular blind docking analysis based on the target proteins with PDB IDs (3D94 and 5FXS) (the docking parameters were set as pH = 7.4, Exhaustiveness = 16, and Box mode = ligand) (S7).\u003c/p\u003e\n\u003cp\u003eIn the docking results, the compound C1CN(CCC12C(=O)NCN2C3=CC=CC=C3)CCCC(=O)C4=CC=C(C=C4)F, which ranked at the top of the affinity list, showed the best free energy of binding to the target protein 3D94 of all ligands at -10.5 kcal/mol (Figure 9A). Conformational analysis showed that the molecule is deeply embedded in the hydrophobic pocket of the protein, forming extensive hydrophobic interactions with residues LEU975, VAL983, ALA1021, and PHE1124, and establishing a network of multiple hydrogen bonds with residues CYS1111, VAL1113, ASP1123, and PHE1124, with a bond spacing ranging from about 2.2 to 3.8 \u0026Aring;.\u003c/p\u003e\n\u003cp\u003eThe compound CN1CCN(CC1)C2=NC3=C(C=CC(=C3)Cl)NC4=CC=CC=C42 has a binding energy of -9.2 kcal/mol with 5FXS (Figure 9B). The ligand mainly formed hydrophobic stacking with hydrophobic residues such as LEU1005 and ALA1031 through the aromatic ring and formed multiple hydrogen bonds with ASP1086, with an interaction distance of about 2.0-4.0 \u0026Aring;. In addition, its chlorine atom formed a stable halogen bond with the carboxylate oxygen of ASP1153, which further enhanced the ligand binding to the target protein. The binding affinity and specificity between the ligand and the target protein are further enhanced. Compound CC1(CCC(C2=C1C=CC(=C2)NC(=O)C3=CC=C(C=C3)C(=O)O)(C)C)C binds to 5FXS with a binding energy of -9.1 kcal/mol (Fig. 9C), and its hydrophobic skeleton is tightly encapsulated by residues such as LEU1005, LYS1033, and MET1079. Its hydrophobic backbone was tightly wrapped by residues such as LEU1005, LYS1033, and MET1079 and further stabilized the binding mode by forming a hydrogen bond with MET1082. Another compound, COC1=C2C3=C(C(=O)CC3)C(=O)OC2=C4C5C=COC5OC4=C1, also has a binding energy of -9.1 kcal/mol with 5FXS (Figure 9D), and its polycyclic aromatic structure forms multi-point hydrophobic contacts with residues such as LEU1005, ALA1031, and so on. Its polycyclic aromatic structure formed multi-point hydrophobic contacts with residues such as LEU1005 and ALA1031, and despite the relatively small number of hydrogen bonds, its overall affinity was still at a favorable level.\u003c/p\u003e\n\u003cp\u003eThe above four representative small molecules exhibited docking affinities better than -9.0 kcal/mol and were stably bound to the protein binding pocket through a variety of non-covalent interactions, including hydrophobic interactions, hydrogen bonding, and halogen bonding (S8).\u003c/p\u003e\n\u003cp\u003e============================\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFig. 9\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e============================\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFig 9\u003c/strong\u003e Binding poses and key interactions of top-ranked candidate small molecules with target proteins 3D94 and 5FXS. (A) Compound C1CN(CCC12C(=O)NCN2C3=CC=CC=C3)CCCC(=O)C4=CC=C(C=C4)F Docking conformation and its hydrophobic/hydrogen bonding interactions with the target protein 3D94. (B) Compound CN1CCN(CC1)C2=NC3=C(C=CC(=C3)Cl)NC4=CC=CC=C42 Docking conformation and its hydrophobic, hydrogen- and halogen-bonding interactions with target protein 5FXS. (C) Docking conformation and predominantly hydrophobic binding mode of the compound CC1(CCC(C2=C1C=CC(=C2)NC(=O)C3=CC=C(C=C3)C(=O)O)(C)C)C with 5FXS. (D) Compound COC1=C2C3=C(C(=O)CC3)C(=O)OC2=C4C5C=COC5OC4=C1 Docking conformation with 5FXS and its hydrophobic-dominant binding interactions.\u0026nbsp;\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we systematically identified key biomarkers of AD and screened small molecule compounds with therapeutic potential by integrating multi-omics data and machine learning methods. The findings provide new insights and candidate targets for the diagnosis and treatment of AD.\u003c/p\u003e\n\u003cp\u003eThe integrated analysis of multiple cohorts in this study firstly suggests that there are not only a large number of differentially expressed genes in AD brain tissues, but also that these genes are highly clustered in glial cell-related processes and key signaling pathways, and the GO/KEGG/DO enrichment results point to glial cell differentiation, gliomagenesis, JAK-STAT, Notch and FoxO pathways, as well as disease phenotypes such as cerebral ischemia and cognitive deficits, which is highly compatible with the current understanding of the pathological features of AD (45, 46). That neuronal damage often occurs in the context of a long-term glial cell response and chronic inflammation. In other words, our network and enrichment analyses at the transcriptomic level reinforce the view that AD is not a single \u0026quot;neuronal disease\u0026quot; but a systemic pathology centered on the neuronal-glial-immune axis, and that key drivers are most likely located at the intersection of this network\u0026nbsp;(47). The key drivers are likely to be located at the intersection of this network.\u003cbr\u003eOn this basis, through topology centrality and integrated machine learning, we converge layer by layer on two hub genes, IGF1R and SPP1, which have several noteworthy features. First, the RF models constructed by these 11 characterized genes still maintain an AUC close to 0.96 in multiple cohorts, cross-validation, and external validation, indicating that the expression patterns of these genes are not only discriminative in a single cohort but also stable and migratory across different platforms and sample sources (48). This is not common in highly heterogeneous diseases such as AD and suggests that these genes may be located relatively \u0026quot;upstream\u0026quot; or at \u0026quot;key nodes\u0026quot; in the disease process. Secondly, the AUCs of both IGF1R and SPP1 were \u0026gt; 0.70 in the multi-cohort ROC, suggesting that there is some diagnostic value at the single-gene level\u0026mdash;providing a realistic basis for the development of simplified translational tools (e.g., qPCR panels or digital PCR diagnostic markers) in the future.\u003c/p\u003e\n\u003cp\u003eBiologically, the IGF1R-associated signaling axis is closely related to insulin/IGF signaling, metabolic imbalance, cell survival, and stress response, which have long been reported in \u0026quot;brain insulin resistance,\u0026quot; impaired synaptic function, and aberrant tau phosphorylation in AD, while SPP1 (osteopontin) is associated with microglia activation. SPP1 (osteopontin), on the other hand, is highly associated with microglia activation, chemotaxis, inflammatory amplification, and tissue remodeling, which is more inclined to an \u0026quot;inflammatory-immune-driven\u0026quot; pathology (49). The TF-target network in this study further showed that IGF1R is located at the intersection of multiple high-confidence transcription factors (TP53, SP1, POU2F1, NR3C1, etc.) (50), suggesting that it is subject to transcriptional regulation at multiple pathways and levels and may act as a \u0026quot;signaling integrator\u0026quot; between stress, metabolism, and inflammation signals. integrator\u0026quot; between stress, metabolic, and inflammatory signals. To some extent, this explains why IGF1R has always been the top contributor in machine learning models: it is not only a \u0026quot;read\u0026quot; but also a \u0026quot;convergence point\u0026quot; of multiple pathological pathways.\u003c/p\u003e\n\u003cp\u003eImmune infiltration and single-cell/spatial transcriptome results provided further clues to the functions of IGF1R and SPP1 from a \u0026quot;cell-microenvironment\u0026quot; perspective (51). CIBERSORT analysis showed that high IGF1R expression was moderately negatively correlated with the infiltration of some B and T cell subsets, while SPP1 was significantly positively correlated with myeloid cells such as M0/M1/M2 macrophages and neutrophils, which formed a differential pattern of \u0026quot;relative suppression of adaptive immunity and significant activation of myeloid/inflammation.\u0026quot; This pattern is very consistent with the characteristic of AD in which innate immunity is over-activated, microglia are chronically activated, and adaptive immune involvement is relatively limited. Single-cell analysis further revealed that IGF1R expression was concentrated in neuroepithelial cells and some neuronal subpopulations and changed dynamically with the temporal differentiation from precursors to mature neurons, suggesting that IGF1R was more involved in \u0026quot;neuronal fate determination/survival and plasticity,\u0026quot; whereas SPP1 was more focally and sparsely expressed and appeared mainly in a limited number of inflammation-associated cells. SPP1 expression was more focal and sparse, occurring mainly in a limited population of inflammation-associated cells. Combined with spatial transcriptomic evidence, IGF1R and SPP1 showed specific spatial enrichment in AD-susceptible brain regions, implying that the same molecular marker does not play the same role in different cell types and spatial regions, which reminds us that the subsequent intervention strategy needs to take into account the cellular and spatial specificity and to avoid \u0026quot;one-size-fits-all\u0026quot; targeting in the whole brain. This also reminds us that subsequent intervention strategies need to take into account cellular and spatial specificity to avoid \u0026quot;one-size-fits-all\u0026quot; targeted inhibition.\u003c/p\u003e\n\u003cp\u003eThe molecular docking results of the four candidate compounds with the target proteins revealed their binding modes and key structure-activity relationships. Among them, compound C1CN(CCC12C(=O)NCN2C3=CC=CC=CC=C3)CCCC(=O)C4=CC=C(C=C4)F exhibited optimal binding affinity (\u0026Delta;G = -10.5 kcal/mol) with 3D94 protein. Structural analysis indicated that its cyclic backbone constitutes a rigid support; the hydrophobic groups penetrate deep into the hydrophobic cavity of the protein to form stable van der Waals interactions, whereas the amide bonds and fluorine atoms form a multicenter hydrogen bonding network with residues such as CYS1111, ASP1123, and others (bond distances in the range of 2.2-3.8 \u0026Aring;). The introduction of fluorine atoms further enhanced the metabolic stability of the compounds.\u003c/p\u003e\n\u003cp\u003eIn the binding mode of compound CN1CCN(CC1)C2=NC3=C(C=CC(=C3)Cl)NC4=CC=CC=CC=C42 with 5FXS, the chlorine atom forms a directional halogen bond with the carboxyl oxygen of ASP1153, a non-covalent interaction that contributes to the enhancement of the binding specificity and free energy contribution.\u003c/p\u003e\n\u003cp\u003eThe remaining two polycyclic aromatic compounds have slightly lower binding energies (\u0026Delta;G = -9.1 kcal/mol), but their thick ring systems are stabilized by extensive hydrophobic stacking with residues such as LEU1005 and ALA1031, and the aromatic planes also have the potential to form \u0026pi;-\u0026pi; or cation-\u0026pi; interactions, which are often helpful in improving the binding specificity and free energy contribution of the molecules. Structural features usually contribute to the improved blood-brain barrier permeability of the molecule.\u003c/p\u003e\n\u003cp\u003eIt should be noted that molecular docking is mainly based on the static structure to assess the thermodynamic stability, which is not yet able to adequately reflect the conformational dynamics, dissociation rate, and full-atom scale dynamic behavior during the binding process. Therefore, the actual drug potential of these compounds needs to be systematically evaluated in conjunction with molecular dynamics simulations, cellular-level functional validation, and animal experiments.\u003c/p\u003e\n\u003cp\u003eFrom a translational perspective, an important value of this study is that instead of merely proposing that \u0026quot;certain genes are up-/down-regulated in AD,\u0026quot; multiple approaches point to a core logic chain:\u003c/p\u003e\n\u003cp\u003eGlial and immune-related transcriptional abnormalities in AD to hub genes such as IGF1R/SPP1 are at key regulatory nodes to stable diagnostic value and correlation with specific cellular subpopulations and spatial structures to mapping to specific small molecule sites of action and derivation of drug candidates (21).\u003c/p\u003e\n\u003cp\u003eThis linkage makes the process from \u0026quot;discovering markers\u0026quot; to \u0026quot;proposing targets for intervention\u0026quot; to \u0026quot;giving lead compounds\u0026quot; more coherent and interpretable.\u003c/p\u003e\n\u003cp\u003eFrom biomarker discovery to clinical application faces many problems and challenges, and we need to develop more efficient and cost-effective research programs. There is a need to develop assays that are noninvasive, reproducible, and highly correlated with intracerebral states. Blood or cerebrospinal fluid assays are ideal directions, but their consistency with central expression needs to be systematically verified. Second, standardized testing procedures and uniform diagnostic thresholds need to be established through multicenter studies, and the interference of age, gender, and comorbidities needs to be assessed (52).\u003c/p\u003e\n\u003cp\u003eIn terms of therapeutic development, although the compounds screened in this study provide lead molecules for drug innovation, the discovery-to-clinical translation of new drugs usually requires long lead times, high cost investment, and the need to address issues such as blood-brain barrier permeability, long-term safety, and adaptive therapeutic windows. AD pathology often exists for many years prior to symptom onset, which means that the most effective therapies often need to be implemented in the preclinical phase, and the clinical trials face both ethical and practical difficulties in the design and implementation of clinical trials at this stage. This study provides potentially translational targets and drug candidates, but their clinical application requires rigorous systematic validation. Future research should focus on mechanistic insight and technological innovation. We need to further elucidate the dynamic roles of IGF1R signaling dysregulation and SPP1-mediated immunoregulation in AD (53); develop brain-targeted drug delivery systems and multimodal diagnostic platforms; and rely on artificial intelligence to integrate multi-omics and clinical data to build an accurate AD typing and treatment decision-making system, which will ultimately promote the establishment of individualized prevention and treatment strategies (54).\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn this study, we systematically established a screening and validation pipeline for key AD-related genes by integrating multi-omics data with machine-learning approaches. By combining differential expression analysis with WGCNA-derived modules and applying 127 algorithm combinations together with SHAP-based interpretability analysis, we ultimately identified two hub genes, IGF1R and SPP1, whose diagnostic value was validated at multiple levels. Subsequent drug screening based on these hub genes yielded several high-affinity small-molecule compounds, providing both candidate biomarkers and lead compounds for precision diagnosis and targeted therapy in AD.\u003c/p\u003e\n"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompliance with Ethical Standards\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll authors have declared their consent for this publication.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no conflict of interest.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCRediT authorship contribution statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eWZF\u003c/strong\u003e: Investigation, Conceptualization, Methodology development, Coding, Data Analysis, Writing - Original Draft, Writing - Review \u0026amp; Editing, Software implementation, Validation of results. \u003cstrong\u003eHJQ\u003c/strong\u003e: Conceptualization, Project Administration, Writing - Review \u0026amp; Editing, Supervision, Drawing, Preparation of figures and tables, Resource allocation. \u003cstrong\u003eZY\u003c/strong\u003e: Algorithm programming, data set testing. \u003cstrong\u003eGC\u003c/strong\u003e: Supervisors, project management.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe data will be provided as required.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eBotella Lucena P, Heneka MT. Inflammatory aspects of Alzheimer\u0026rsquo;s disease. Acta neuropathologica. 2024;148(1):31. \u003c/li\u003e\n\u003cli\u003eChen S, Cao Z, Nandi A, et al. The global macroeconomic burden of Alzheimer\u0026apos;s disease and other dementias: estimates and projections for 152 countries or territories. The Lancet Global Health. 2024;12(9):e1534-e43. \u003c/li\u003e\n\u003cli\u003eZhang J, Na X, Li Z, et al. Sarcopenic obesity is part of obesity paradox in dementia development: evidence from a population-based cohort study. BMC medicine. 2024;22(1):133. \u003c/li\u003e\n\u003cli\u003eBenjamin SR, Lima Fd, Nunes PIG, Dutra RF, Andrade GMd, Ori\u0026aacute; RB. Advanced Biosensing Technologies: Leading Innovations in Alzheimer\u0026rsquo;s Disease Diagnosis. Chemosensors. 2025;13(6):220. \u003c/li\u003e\n\u003cli\u003eLiu Y, Xia X, Zheng M, Shi B. Bio\u0026ndash;Nano Toolbox for Precision Alzheimer\u0026apos;s Disease Gene Therapy. Advanced Materials. 2024;36(29):2314354. \u003c/li\u003e\n\u003cli\u003eFrisoni GB, Aho E, Brayne C, et al. Alzheimer\u0026apos;s disease outlook: controversies and future directions. The Lancet. 2025;406(10510):1424-42. \u003c/li\u003e\n\u003cli\u003eSarma M, Chatterjee S. Machine Learning-Based Alzheimer\u0026rsquo;s Disease Stage Diagnosis Utilizing Blood Gene Expression and Clinical Data: A Comparative Investigation. Diagnostics. 2025;15(2):211. \u003c/li\u003e\n\u003cli\u003eBallard JL, Wang Z, Li W, Shen L, Long Q. Deep learning-based approaches for multi-omics data integration and analysis. BioData Mining. 2024;17(1):38. \u003c/li\u003e\n\u003cli\u003eBao J, Lee BN, Wen J, et al. Employing Informatics Strategies in Alzheimer\u0026apos;s Disease Research: A Review from Genetics, Multiomics, and Biomarkers to Clinical Outcomes. Annual review of biomedical data science. 2024;7. \u003c/li\u003e\n\u003cli\u003eAhuja SK, Shrimankar DD, Durge AR. A study and analysis of disease identification using genomic sequence processing models: An empirical review. Current Genomics. 2023;24(4):207-35. \u003c/li\u003e\n\u003cli\u003eCardillo M, Katam K, Survajahala P. Advancements in Multi-Omics Research to Addressing Challenges in Alzheimer\u0026apos;s Disease: A Systems Biology Approach Utilizing Molecular Biomarkers and Innovative Strategies. Frontiers in Aging Neuroscience. 2025;17:1591796. \u003c/li\u003e\n\u003cli\u003eAbdelwahab MM, Al-Karawi KA, Semary HE. Deep learning-based prediction of Alzheimer\u0026rsquo;s disease using microarray gene expression data. Biomedicines. 2023;11(12):3304. \u003c/li\u003e\n\u003cli\u003eVemulavada S, Karthikvatsan S, Babu A, et al. Role of Artificial Intelligence in the Diagnosis and Therapy of Gynecological Disorders: Opportunities and Challenges. Translational Research in Biomedical Sciences: Recent Progress and Future Prospects. 2024:121-44. \u003c/li\u003e\n\u003cli\u003eAghdam MA, Bozdag S, Saeed F. Machine-learning models for Alzheimer\u0026rsquo;s disease diagnosis using neuroimaging data: survey, reproducibility, and generalizability evaluation. Brain Informatics. 2025;12(1):1-27. \u003c/li\u003e\n\u003cli\u003eSudharsan M, Thailambal G. Alzheimer\u0026apos;s disease prediction using machine learning techniques and principal component analysis (PCA). Materials Today: Proceedings. 2023;81:182-90. \u003c/li\u003e\n\u003cli\u003eZhang X, Li Z, Zhang Q, Yin Z, Lu Z, Li Y. A new weakly supervised deep neural network for recognizing Alzheimer\u0026rsquo;s disease. Computers in Biology and Medicine. 2023;163:107079. \u003c/li\u003e\n\u003cli\u003eHe H, Yang Y, Wang L, et al. Combined analysis of single-cell and bulk RNA sequencing reveals the expression patterns of circadian rhythm disruption in the immune microenvironment of Alzheimer\u0026rsquo;s disease. Frontiers in Immunology. 2023;14:1182307. \u003c/li\u003e\n\u003cli\u003eLv H, Tan L, Miao Y, Zhou Y, Wang L, Liang Z. Machine learning identifies lactate metabolism biomarkers and deciphers immune infiltration landscapes in Parkinson\u0026rsquo;s disease. Scientific Reports. 2025;15(1):37665. \u003c/li\u003e\n\u003cli\u003eSagulkoo P, Chuntakaruk H, Suratanee A, Plaimas K, Teerakulkittipong N. Integrative bioinformatics frameworks for abdominal aortic aneurysm using GWAS meta-analysis, biological network construction, and structural modeling. Scientific Reports. 2025;15(1):22331. \u003c/li\u003e\n\u003cli\u003eZhu Y, Kong L, Han T, Yan Q, Liu J. Machine learning identification and immune infiltration of disulfidptosis‐related Alzheimer\u0026apos;s disease molecular subtypes. Immunity, Inflammation and Disease. 2023;11(10):e1037. \u003c/li\u003e\n\u003cli\u003eAlshabrmi FM, Alkhayl FFA, Rehman A. Novel drug discovery: Advancing Alzheimer\u0026apos;s therapy through machine learning and network pharmacology. European Journal of Pharmacology. 2024;976:176661. \u003c/li\u003e\n\u003cli\u003eChhabra N, Matore BW, Lakra N, et al. Multilayered screening for multi-targeted anti-Alzheimer\u0026rsquo;s and anti-Parkinson\u0026rsquo;s agents through structure-based pharmacophore modelling, MCDM, docking, molecular dynamics and DFT: a case study of HDAC4 inhibitors. In Silico Pharmacology. 2025;13(1):16. \u003c/li\u003e\n\u003cli\u003eGan Y, Sun J, Yang D, Fang C, Zhou Z, Yin J. Exploration and validation of biomarkers for Alzheimer\u0026apos;s disease based on GEO database. IBRO Neuroscience Reports. 2025. \u003c/li\u003e\n\u003cli\u003eAlhassan HH. Integrative machine learning and molecular simulation approaches identify GSK3\u0026beta; inhibitors for neurodegenerative disease therapy. Scientific Reports. 2025;15(1):21632. \u003c/li\u003e\n\u003cli\u003eElghaish RA, Attallah NE, Khaled H, Mekawy AS, Elserafy M, Badr E. A computational framework for identifying cytoskeletal genes associated with age-related diseases. Scientific Reports. 2025;15(1):14590. \u003c/li\u003e\n\u003cli\u003eFerrari RR, Fantini V, Garofalo M, et al. A map of transcriptomic signatures of different brain areas in Alzheimer\u0026rsquo;s disease. International Journal of Molecular Sciences. 2024;25(20):11117. \u003c/li\u003e\n\u003cli\u003eFeng G, Zhong M, Huang H, et al. Identification of UBE2N as a biomarker of Alzheimer\u0026rsquo;s disease by combining WGCNA with machine learning algorithms. Scientific Reports. 2025;15(1):6479. \u003c/li\u003e\n\u003cli\u003eXie Z, Li L, Hou W, et al. Critical role of Oas1g and STAT1 pathways in neuroinflammation: insights for Alzheimer\u0026rsquo;s disease therapeutics. Journal of Translational Medicine. 2025;23(1):182. \u003c/li\u003e\n\u003cli\u003eLiu M-j, Zhu J-y, Han S-y, Zhu S-j, Yang D-d. Single-Cell Sequencing Screening of Biomarkers Associated with Lipophagy in the MPTP Model Mice of Parkinson\u0026rsquo;s Disease. Molecular Neurobiology. 2025:1-18. \u003c/li\u003e\n\u003cli\u003eWu T, Hu E, Xu S, et al. clusterProfiler 4.0: A universal enrichment tool for interpreting omics data. The innovation. 2021;2(3). \u003c/li\u003e\n\u003cli\u003e\u0026Uuml;nal B. Causality analysis for COVID-19 among countries using effective transfer entropy. Entropy. 2022;24(8):1115. \u003c/li\u003e\n\u003cli\u003eMuley VY. Functional Insights Through Gene Ontology, Disease Ontology, and KEGG Pathway Enrichment. Computational Virology: Springer; 2025. p. 75-98.\u003c/li\u003e\n\u003cli\u003eUllah I, Liu K, Yamamoto T, Zahid M, Jamal A. Modeling of machine learning with SHAP approach for electric vehicle charging station choice behavior prediction. Travel Behaviour and Society. 2023;31:78-92. \u003c/li\u003e\n\u003cli\u003eMitra U, Sarkar P, Mondal J, Kundu J, editors. Enhancing Interpretability in Diabetics Prediction: A Comparative Study of SHAP, LIME and Permutation Feature Importance. 2025 AI-Driven Smart Healthcare for Society 5.0; 2025: IEEE.\u003c/li\u003e\n\u003cli\u003eNilchi AN, Dehghanian F, Vallian S, Bahreini A. Dissecting lncRNA-mRNA regulatory network in type 2 diabetes as the risk factor of pancreatic cancer. Scientific Reports. 2025;15(1):37288. \u003c/li\u003e\n\u003cli\u003ePaprotny D, Terefenko P, Śledziowski J. HANZE v2. 1: an improved database of flood impacts in Europe from 1870 to 2020. Earth System Science Data. 2024;16(11):5145-70. \u003c/li\u003e\n\u003cli\u003eZamore Mn, Junior SM, Andree SL, Altunbulakli C, Lindstedt M, Levander F. Considerations and Software for Successful Immune Cell Deconvolution Using Proteomics Data. Journal of Proteome Research. 2025;24(8):3751-61. \u003c/li\u003e\n\u003cli\u003eJain S, Rick JW, Joshi RS, et al. Single-cell RNA sequencing and spatial transcriptomics reveal cancer-associated fibroblasts in glioblastoma with protumoral effects. The Journal of clinical investigation. 2023;133(5). \u003c/li\u003e\n\u003cli\u003eLi Q, Qu L, Miao Y, et al. A gene network database for the identification of key genes for diagnosis, prognosis, and treatment in sepsis. Scientific Reports. 2023;13(1):21815. \u003c/li\u003e\n\u003cli\u003eKim S, Bolton EE. Pubchem: a large‐scale public chemical database for drug discovery. Open Access Databases and Datasets for Drug Discovery. 2024:39-66. \u003c/li\u003e\n\u003cli\u003eLai CH-L, Kwok APK, Wong K-C. Cheminformatic Identification of Tyrosyl-DNA Phosphodiesterase 1 (Tdp1) Inhibitors: A Comparative Study of SMILES-Based Supervised Machine Learning Models. Journal of Personalized Medicine. 2024;14(9):981. \u003c/li\u003e\n\u003cli\u003ede Oliveira LHD, Cruz JN, Dos Santos CBR, de Melo EB. Multivariate QSAR, similarity search and ADMET studies based in a set of methylamine derivatives described as dopamine transporter inhibitors. Molecular Diversity. 2023:1-16. \u003c/li\u003e\n\u003cli\u003eAndersson C, Bejsj\u0026ouml; E, Magnusson SB, Carlmark A, Fuciu F, Nordwall E. Protein Engineering in Python Proteinmodifiering i Python. 2025. \u003c/li\u003e\n\u003cli\u003eBurley SK, Bhikadiya C, Bi C, et al. RCSB Protein Data Bank (RCSB. org): delivery of experimentally-determined PDB structures alongside one million computed structure models of proteins from artificial intelligence/machine learning. Nucleic acids research. 2023;51(D1):D488-D508. \u003c/li\u003e\n\u003cli\u003eSong T, Zhang Y, Zhu L, Zhang Y, Song J. The role of JAK/STAT signaling pathway in cerebral ischemia-reperfusion injury and the therapeutic effect of traditional Chinese medicine: A narrative review. Medicine. 2023;102(46):e35890. \u003c/li\u003e\n\u003cli\u003eWu J-W, Wang B-X, Shen L-P, et al. Investigating the Potential Therapeutic Targeting of the JAK-STAT Pathway in Cerebrovascular Diseases: Opportunities and Challenges. Molecular Neurobiology. 2025:1-27. \u003c/li\u003e\n\u003cli\u003eMaiese K. The metabolic basis for nervous system dysfunction in alzheimer\u0026rsquo;s disease, parkinson\u0026rsquo;s disease, and huntington\u0026rsquo;s disease. Current neurovascular research. 2023;20(3):314-33. \u003c/li\u003e\n\u003cli\u003eHe R, Cheng J, Qiu Y, et al. IGF1R and FLT1 in female endothelial cells and CHD2 in male microglia play important roles in Alzheimer\u0026apos;s disease based on gender difference analysis. Experimental Gerontology. 2024;194:112512. \u003c/li\u003e\n\u003cli\u003eEngel MG. Deciphering the Role of Insulin and IGF-1 in Brain Aging, Alzheimer\u0026rsquo;s Disease, and Vascular Dementia: Albert Einstein College of Medicine; 2023.\u003c/li\u003e\n\u003cli\u003eSengupta P, Mukhopadhyay D. IGF1R/ARRB1 Mediated Regulation of ERK and cAMP Pathways in Response to A\u0026beta; Unfolds Novel Therapeutic Avenue in Alzheimer\u0026rsquo;s Disease. Molecular Neurobiology. 2025;62(6):8065-83. \u003c/li\u003e\n\u003cli\u003eJohnston KG, Berackey BT, Tran KM, et al. Single-cell spatial transcriptomics reveals distinct patterns of dysregulation in non-neuronal and neuronal cells induced by the Trem2 R47H Alzheimer\u0026rsquo;s risk gene mutation. Molecular Psychiatry. 2025;30(2):461-77. \u003c/li\u003e\n\u003cli\u003eLiu Q, Ling J, Li Z, Bi L. Advances in lymphoma biomarkers research based on proteomics technology. Oncology Reports. 2025;54(3):1-17. \u003c/li\u003e\n\u003cli\u003eArgandona Lopez C, Brown AM. Microglial-neuronal crosstalk in chronic viral infection through mTOR, SPP1/OPN and inflammasome pathway signaling. Frontiers in immunology. 2024;15:1368465. \u003c/li\u003e\n\u003cli\u003eG\u0026uuml;naydın T, Varlı S. Computer-Aided Decision Support Systems of Alzheimer\u0026apos;s Disease Diagnosis-A Systematic Review. Current Medical Imaging. 2025;21(1):E15734056359358. \u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":false,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Alzheimer’s disease, machine learning, SHAP, spatial transcriptomics, molecular docking","lastPublishedDoi":"10.21203/rs.3.rs-8200314/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8200314/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground:\u003c/strong\u003e Alzheimer’s disease (AD) is driven by complex molecular and immune dysregulation, yet reliable diagnostic biomarkers and druggable targets remain limited. This study aimed to identify key AD-associated regulatory genes, characterize their immune and spatial expression features, and prioritize small-molecule compounds with therapeutic potential.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods:\u003c/strong\u003e Multiple AD-related transcriptomic datasets—including bulk RNA-seq, microarray, and spatial transcriptomic profiles—were retrieved from GEO and systematically partitioned into discovery (GSE5281, GSE66333), validation (GSE110226, GSE28146, GSE29378), independent testing (GSE29378), and spatial validation cohorts (GSE147047). Differential expression analysis and weighted gene co-expression network analysis (WGCNA) were used to construct co-expression networks and define AD-associated gene modules. Protein–protein interaction (PPI) analysis and multiple network centrality measures were then applied to prioritize candidate key genes. Twelve machine-learning algorithms were combined into 127 classification models, and SHAP-based interpretability analysis was used to quantify feature contributions and identify diagnostic genes. Single-cell and spatial transcriptomic data were further used to validate the cell type specificity and spatial localization of the hub genes. Drug–gene enrichment analysis (DSigDB), compound retrieval (PubChem), ADMET and drug-likeness profiling, and molecular blind docking were integrated to screen and evaluate potential lead compounds.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults:\u003c/strong\u003e We identified 2,534 differentially expressed genes (DEGs) between AD and control samples, and their intersection with WGCNA-derived modules yielded 848 candidate genes. PPI-based network analysis prioritized 15 key genes, on which 127 machine-learning models were constructed; the random forest model achieved the best overall performance with an average AUC of 0.957. SHAP analysis identified 11 key diagnostic genes, among which IGF1R and SPP1 emerged as stable hub genes with AUCs greater than 0.70 across multiple external cohorts. Immune infiltration, single-cell, and spatial transcriptomic analyses demonstrated distinct immune associations and cell type– and region-specific expression patterns of these hub genes. Drug–gene enrichment identified 176 drug signatures and 445 related compounds, of which 37 grade-A molecules remained after ADMET and drug-likeness filtering. Molecular docking revealed four top-ranked compounds with binding energies better than −9.0 kcal/mol, including one ligand with a minimum binding energy of −10.5 kcal/mol and extensive non-covalent interactions with the target protein.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusion:\u003c/strong\u003e A systematic methodological framework from gene discovery and diagnostic modeling to lead drug screening was developed in this study. IGF1R and SPP1 were identified as stable and biologically interpretable AD hub genes, which can be used as potential diagnostic markers, and various high-affinity small molecule compounds based on the hub genes provide new drug candidates for targeted AD therap.\u003c/p\u003e","manuscriptTitle":"Integrated Bioinformatics and Ensemble Learning Reveal Diagnostic Modeling and Drug Discovery in Alzheimer’s Disease","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-11-26 07:16:11","doi":"10.21203/rs.3.rs-8200314/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"768ee73e-bb12-4096-ac9f-21d52091c730","owner":[],"postedDate":"November 26th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":58574119,"name":"Bioinformatics"},{"id":58574120,"name":"Artificial Intelligence and Machine Learning"},{"id":58574121,"name":"Drug Discovery, Design, \u0026 Development"},{"id":58574122,"name":"Cellular \u0026 Molecular Neuroscience"}],"tags":[],"updatedAt":"2025-11-26T07:16:11+00:00","versionOfRecord":[],"versionCreatedAt":"2025-11-26 07:16:11","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8200314","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8200314","identity":"rs-8200314","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}