To investigate the intersection of N-glycosylation and AD through bibliometric analysis, data were retrieved from the Web of Science Core Collection, a gold-standard database for such research. English-language research articles and reviews published between 2001 and 2025 were included using the optimized search query:

“TS = ((“N-glycosylat*” OR “N-linked glycosylat*” OR “N-glycan*” OR “N-glycoprotein*” OR “protein N-glycosylat*” OR “n-glycoform*”) AND (Alzheimer* OR “Alzheimer’s disease” OR “Alzheimer’s” OR “Alzheimer disease” OR “Alzheimer’s-like” OR “Alzheimer pathology” OR “Alzheimer pathogenesis” OR “tau protein” OR “beta-amyloid”)).”

The initial search yielded 239 documents, which were filtered to retain only research articles and reviews. The complete list of these documents and detailed search outcomes are provided in Supplementary Table 1.

Three analytical tools were employed: VOSviewer (v1.6.20), CiteSpace (v6.4 R1), and R (v4.3.3). VOSviewer was used to generate network visualizations, including thematic evolution mappings (to track shifts in research focus over time) and thematic quadrant mappings (to classify themes by centrality and density) (Zhou et al., 2025). CiteSpace constructed circular thematic landscapes (to identify clusters using silhouette coefficients) and performed targeted cluster analyses (for in-depth exploration of specific clusters) (Sebastian et al., 2020). Additionally, keyword burst detection was applied to characterize dynamic changes in research themes. R was utilized for quantitative analysis of publication data, including thematic trend quantification and keyword correlation calculations, to ensure rigorous examination of the literature.

The GSE5281 dataset was analyzed to investigate N-glycosylation-related genes in AD. This dataset comprises gene expression profiles from 150 brain samples collected across three Alzheimer’s Disease Centers (ADCs): Arizona ADC, Duke University ADC, and Washington University ADC (Zhang et al., 2022). The dataset covers six brain regions relevant to AD and aging, including the entorhinal cortex, hippocampus, medial temporal gyrus, posterior cingulate, superior frontal gyrus, and primary visual cortex. Samples were stratified by diagnostic group, age group, and APOE genotype to ensure comprehensive coverage of molecular mechanisms in AD and normal aging.

Differentially expressed genes (DEGs) were identified using the LIMMA package in R, a widely used tool for microarray data analysis. The workflow included data preprocessing to normalize expression values and reduce technical variability, followed by fitting linear models to assess differential expression between AD-affected and control samples. DEGs were defined using stringent criteria: an absolute log2 fold change (logFc) > 0.4 and an adjusted p-value (Benjamini-Hochberg correction) < 0.05.

To investigate the overlap between DEGs and N-glycosylation-related genes, a list of N-glycosylation-associated genes was retrieved from the GeneCards database (Supplementary Table 1). The intersection between DEGs and this list was computed to identify genes implicated in both AD pathogenesis and N-glycosylation processes. All analyses, including data preprocessing, statistical modeling, and visualization, were performed using R.

Machine learning techniques were employed to identify genes associated with AD from an initial set of 39 genes derived from the intersection of DEGs and N-glycosylation-related genes in the GSE5281 and GSE122063 datasets. Three regularization methods were applied: Lasso, Elastic-Net, and Adaptive Lasso. Lasso regression was configured with an alpha parameter (α) set to 1, Elastic-Net with α = 0.5, and Adaptive Lasso without a predefined α, allowing adaptive adjustment of regularization strength. These methods were implemented using the glmnet package in R, with cross-validation to determine the optimal regularization parameter (λ) that minimized deviance. Genes with non-zero coefficients at the optimal λ were considered significant. Results were visualized using coefficient path plots and feature impact bar charts, and a Venn diagram was constructed to illustrate the overlap of selected genes across the three methods.

The relationship between five specific genes and AD was investigated using the GSE48350 and GSE5281 datasets. The workflow began with data preprocessing, including data normalization and removal of features with zero variance. Feature selection was performed using the Random Forest algorithm, configured to grow 100 trees to assess feature importance. Three machine learning models were constructed: logistic regression, Random Forest, and XGBoost. The logistic regression model was implemented using the “lrm” function from the “rms” package, the Random Forest model with 100 trees and importance parameter set to TRUE, and the XGBoost model trained for 50 rounds using the “xgb.train” function with the objective parameter set to “binary:logistic.” Model performance was evaluated using the Area Under the Curve (AUC) metric.

The GSE5281 and GSE122063 datasets were utilized to investigate molecular interactions associated with TMEM59, focusing on 11 genes implicated in N-glycosylation and AD. These genes—STT3A, MLEC, CALR, CANX, CTSD, MGAT1, MGAT2, MGAT3, MGAT5, APP, and STT3B—were selected based on their roles in glycosylation processes, calcium signaling, and protein quality control. For example, STT3A and STT3B are involved in glycosylation initiation (Lu et al., 2018), MGAT family members in glycan chain modification (Cha et al., 2017), and MLEC in glycan recognition (Dos Santos Silva et al., 2019). CALR and CANX act as chaperones in protein folding (Knee et al., 2003), while CTSD is linked to protein degradation (Marques et al., 2020). Eight feature selection algorithms were employed: LASSO, Elastic Net, Random Forest, XGBoost, Boruta, Stepwise Regression, Genetic Algorithm, and Decision Tree. Data preprocessing included normalization and removal of samples with missing values. The analysis was implemented in R with parallel computing for efficiency. Core genes selected by at least four algorithms were aggregated as the final output.

To identify pivotal TFs associated with AD, the ChIPBase database was leveraged to predict 44 TFs regulating the expression of four genes of interest: GFAP, FOS, MLEC, and TMEM59. These genes were selected based on their relevance to glial cell differentiation, a process intricately linked to AD pathogenesis (Giesen et al., 2003). The VennDiagram package in R was used to perform intersection analysis, identifying common TFs among GFAP, FOS, MLEC, and TMEM59.

Subsequently, the GSE5281 and GSE48350 datasets were analyzed, providing comprehensive gene expression profiles from brain regions relevant to AD and aging. Seven machine learning algorithms—Lasso, Elastic Net, Adaptive Lasso, XGBoost, Boruta, Genetic Algorithm, and Linear Discriminant Analysis (LDA)—were applied to assess the significance of the predicted TFs. Each algorithm was optimized for feature selection, balancing model complexity and predictive accuracy. The performance of these algorithms was quantified by the selection rate, defined as the ratio of selected genes to the total number of genes considered. Core biomarkers consistently recognized by at least four algorithms were identified as key contributors to AD pathogenesis. The R programming language was used for these analyses, with specific packages such as “glmnet,” “xgboost,” “Boruta,” “GA,” and “MASS” playing critical roles.

The diagnostic performance of machine learning algorithms in AD was evaluated using a panel of genes associated with N-glycosylation and AD pathogenesis. The analysis focused on 13 molecular features: APP, CALR, CANX, CTSD, MAX, MGAT1, MGAT2, MGAT3, MGAT5, MLEC, STT3A, STT3B, and TMEM59. Machine learning algorithms, including logistic regression, support vector machines (linear and radial kernels), random forest, XGBoost, k-nearest neighbors (KNN), LDA, naive Bayes, and decision trees, were implemented to assess diagnostic accuracy. Feature selection was performed using random forest-based importance scoring, and models were trained and validated using five-fold repeated cross-validation. Key performance metrics, including AUC, accuracy, sensitivity, specificity, and F1 score, were calculated for each algorithm and gene combination. The analysis was conducted in R, leveraging packages such as “caret,” “pROC,” “xgboost,” and “randomForest.”

The GSE5281 and GSE48350 datasets were used to analyze 13 molecular features selected based on their roles in N-glycosylation pathways and potential involvement in AD pathogenesis. These features included APP, CALR, CANX, CTSD, MAX, MGAT1, MGAT2, MGAT3, MGAT5, MLEC, STT3A, STT3B, and TMEM59. Data preprocessing involved normalization, median imputation for missing values, and removal of zero-variance features. Feature selection was conducted using a hybrid approach combining random forest importance scoring (ntree = 100) and logistic regression coefficients. Predictive models were built using random forest (ntree = 100), XGBoost (nrounds = 50), and logistic regression with interaction terms. Model performance was evaluated using AUC-ROC, with random forest and XGBoost achieving perfect AUC values of 1.0. SHAP values were calculated to interpret feature contributions, and molecular interactions were analyzed using heatmaps and interaction plots. All analyses were implemented in R, with visualization supported by packages such as “ggplot2,” “pheatmap,” and SHAP tools.

The workflow implemented in our study is depicted in Figure 1, detailed accession numbers and information of the GEO datasets we used can be found in Supplementary Figure 1.

Thematic evolution analysis of the N-glycosylation–AD field from 2001 to 2025 revealed sustained focus on AD’s interactions with core biological and biochemical processes, with N-glycosylation serving as a central connecting theme (Figure 2A). Between 2021 and 2024, research narrowed its focus to granular molecular and biochemical topics, including “cleavage,” “protein,” “mechanisms,” “mass-spectrometry,” and “peptide,” reflecting a shift toward dissecting AD pathogenesis at the molecular level to advance diagnostics and therapeutics. Thematic quadrant analysis (2021–2024) further clarified research priorities (Figure 2B): the upper-right quadrant (“Motor Themes”) included high-centrality and high-density topics such as “Alzheimer’s disease expression” and “glycosylation,” which bridge research directions; the upper-left quadrant (“Niche Themes”) featured specialized subfields like “nervous-system” and “receptor subunits”; the lower-right quadrant (“Basic Themes”) highlighted foundational concepts such as “biomarker” and “dementia”; and the lower-left quadrant (“Emerging/Declining Themes”) included evolving interests like “tau-protein” and “peptide.”

CiteSpace-generated circular thematic landscapes identified high-confidence clusters (Figure 2C): Cluster #0 (“cerebrospinal fluid,” “N-glycan profile,” silhouette coefficient: 0.726) emphasized glycomic analysis of AD clinical samples; Cluster #1 (“neurodegenerative diseases,” silhouette coefficient: 0.712) contextualized N-glycosylation within broader neurodegeneration; and Cluster #3 (“GNT-III activity,” silhouette coefficient: 0.909) focused on enzymatic regulation of glycan maturation. Nodes such as “cerebrospinal fluid” and “N-glycosylation” exhibited high centrality, underscoring their role in connecting clusters. Deep analysis of Cluster #3 revealed high-degree keywords—“glycosylation,” “glycan,” and “Alzheimer’s beta secretase”—highlighting their relevance to GNT-III–mediated AD pathways (Figure 2D). Temporal tracking (2015–2025) showed sustained interest in keywords such as “oxidative stress” and “cell adhesion,” indicating growing focus on GNT-III’s role in AD-related pathology.

The three-field plot visualized the interconnections between key literature, institutions, and keywords in the N-glycosylation–AD field (Figure 3A). Influential studies, including “Vassar R (1999) Science,” “Kizuka Y (2016) Biochem J,” “Ohtsubo K (2006) Cell,” “Gizaw ST (2016) BBA-Gen Subjects,” and “Kizuka Y (2015) EMBO Mol Med,” formed the backbone of current understanding of N-glycosylation’s role in AD pathogenesis. These works were closely linked to leading institutions such as Karolinska Institutet, the University of California System, and Fukushima Medical University, underscoring their dominant contributions to the field. Core keywords included “glycosylation,” “Alzheimer’s disease,” and “amyloid-beta,” while “bisecting GlcNAc” emerged as a prominent focus, signaling growing interest in this specific glycan modification’s role in AD.

The keyword burst detection chart (Figure 3B) tracked high-impact temporal trends (2001–2025): early bursts included “N-linked glycosylation” (strength = 3.65) and “protein glycosylation” (strength = 3.85), reflecting foundational interest in glycan biology. More recently, “pathology” (strength = 3.63) and “tau” (strength = 3.72) have sustained bursts since 2020 (projected through 2025), indicating current focus on disease mechanisms. “Bisecting GlcNAc” also showed a notable burst (strength = 2.74) from 2016 to 2018, underscoring its historical relevance as a specialized research topic. Network visualizations further delineated thematic connections in N-glycosylation–AD research (Figures 3C,D). In Figure 3C, “Alzheimer’s disease” was the most central node (57 occurrences, total link strength = 134), reflecting its status as the field’s core focus, while “N-glycosylation” (24 occurrences, total link strength = 58) confirmed its pivotal role in AD pathological processes. Figure 3D highlighted recent thematic priorities via warm red node coloring: “N-glycosylation” (24 occurrences, link strength = 58) remained central, “Bisecting GlcNAc” (8 occurrences, link strength = 27) showed sustained relevance in recent studies, and “GNT-III (MGAT3)” (2 occurrences, link strength = 7) emerged as a novel focus, consistent with its role in regulating bisecting GlcNAc formation.

Differential expression analysis of the GSE5281 dataset identified 6,845 differentially expressed genes (DEGs) (Supplementary Table 1), visualized in the volcano plot (Figure 4A). The plot highlights significant upregulated genes (red) and downregulated genes (blue), with gray dots representing non-significant genes. Notably, TMEM59 was identified as a significantly downregulated gene, while MLEC and MAX were significantly upregulated in AD-affected samples compared to controls. The intersection of DEGs with N-glycosylation-related genes from the GeneCards database yielded 39 overlapping genes.

In Figure 4B, the Lasso regression coefficient path reveals the influence of various genes on the model, with genes like UNG and MLEC showing the most significant impact as indicated by their steep paths. Figure 4C presents the Elastic-Net regression coefficient path, where genes such as UNG, MLEC, and MAN1C1 are prominently featured, demonstrating their importance in the model with considerable standardized coefficients. Figure 4D showcases the Lasso feature impact, highlighting the top genes selected by absolute coefficient. UNG, MLEC, and MAN1C1 are at the forefront, with UNG having the highest positive coefficient, suggesting its strong association with the disease. Figure 4E illustrates the Adaptive Lasso feature impact, where UNG again leads with the highest positive coefficient, followed by MAN1C1 and MLEC. This figure emphasizes the genes’ significance in the context of AD, with a clear distinction between positive and negative contributions. Lastly, Figure 4F provides a Venn diagram of the key gene intersections across the three methods, showing a core set of genes that are consistently selected: ADPRH, ALG12, ALG8, ARSG, CXCR4, DNPH1, MAGT1, MAN1C1, MLEC, NOX4, NTRK1, ORAI1, SLC3A2, SRD5A3, TMEM59, and UNG. This overlap underscores the robustness of these genes as potential biomarkers for AD, as they are identified across different regularization techniques.

The analysis of the GSE122063 dataset using machine learning algorithms identified a set of genes with potential significance in AD. Figure 4G illustrates the Lasso regression results, where genes such as DAD1 and MLEC exhibit the highest absolute standardized coefficients, indicating their strong influence on the model. Figure 4H presents the Elastic-Net regression outcomes, showing a similar pattern with DAD1 and MLEC being prominent, alongside other genes like ALG8 and MOGS. Figure 4I highlights the top features selected by Lasso regression, with DAD1, MLEC, and ALG8 having the most substantial impact. The Adaptive Lasso results in Figure 4J show DAD1 as the most influential gene, followed by TMEM59 and MLEC, pointing to their crucial role in the disease pathology. The Venn diagram in Figure 4K reveals the intersection of genes selected by all three methods, with ALG1, ALG3, ALG8, DAD1, MLEC, MOGS, ORAI1, PRKCSH, RPN1, SLC3A2, TMEM59, and TMEM59L being consistently identified across Lasso, Elastic-Net, and Adaptive Lasso, underscoring their potential as biomarkers. Finally, Figure 4L compares the gene selection between the GSE5281 and GSE122063 datasets, showing an overlap in ALG8, MLEC, ORAI1, SLC3A2, and TMEM59, which further validates their importance in AD research.

The analysis of the GSE48350 dataset focusing on five genes yielded several significant findings. The logistic regression model achieved a test AUC of 0.594, indicating moderate predictive accuracy. Figure 5A presents the model performance comparison across various machine learning algorithms, showcasing the logistic regression model’s moderate predictive power. Figure 5B illustrates the ranking of the five genes by their importance across different models, with TMEM59 showing the highest mean importance. Figure 5C depicts the molecular signature expression patterns, highlighting a clear distinction between control and disease groups, particularly for TMEM59, which exhibited the most pronounced differential expression. Figure 5D presents the clinical prediction nomogram, which achieved a model performance AUC of 0.699, indicating a reasonable ability to discriminate between control and disease states. This nomogram assigns points to each gene based on its standardized coefficient, with TMEM59 receiving the highest point allocation, significantly contributing to the total points and thus the risk probability of AD. These results underscore the potential utility of these genes as diagnostic biomarkers for AD, with TMEM59 emerging as a particularly promising candidate due to its significant differential expression and high feature importance across models.

The in-depth analysis of the GSE5281 dataset shed light on the roles of five key genes—ALG8, MLEC, ORAI1, SLC3A2, and TMEM59—in AD. The performance of various machine learning models was evaluated in Figure 5E, with logistic regression and Random Forest standing out, achieving test AUC scores of 0.8566 and 0.8505, respectively. XGBoost also performed well, with a test AUC of 0.8304. The clinical prediction nomogram in Figure 5F, which incorporates these genes, shows a high model performance AUC of 0.899, indicating its strong predictive power. This nomogram assigns points to each gene based on its expression level, with TMEM59 receiving the most points, suggesting it plays a crucial role in determining the risk probability of AD. Figure 5G highlights the average feature importance across models, with TMEM59 being identified as the most important gene, followed by ALG8 and MLEC. This underscores the significant impact of these genes on the model’s predictive capabilities. Figure 5H, which includes MLEC and TMEM59, shows molecular signature expression patterns.

The heatmap (Figure 6A) resulting from the analysis of the GSE5281 dataset illustrates the pairwise correlations among the nine selected genes. A significant negative correlation of –0.46 was observed between TMEM59 and MLEC, as indicated by the deep blue color in the heatmap. This finding suggests an inverse relationship between the expression levels of these two genes. The hierarchical clustering dendrogram positioned TMEM59 and MLEC in separate clusters, reinforcing the distinctness of their expression profiles among the analyzed genes. In the analysis of the GSE122063 dataset (Figure 6B), a notable correlation was identified between TMEM59 and MLEC, with a correlation coefficient of 0.52, indicating a moderate positive relationship. This finding suggests that the expression levels of TMEM59 and MLEC may be associated in the samples analyzed from individuals with AD. Analyzing the GSE122063 dataset, Figure 6C shows that MLEC expression levels are increased in the disease group compared to controls.

In the GSE122063 dataset, the feature selection algorithms identified STT3A, CALR, and MLEC as the top three genes associated with TMEM59. Figure 6D shows that the Genetic Algorithm selected all genes, achieving a 100% selection rate, while XGBoost selected 9 genes, indicating an 81.8% selection rate. The heatmap in Figure 6E demonstrates the normalized importance of each gene across algorithms, with STT3A, CALR, and MLEC showing the highest importance. Figure 6F summarizes the core genes, highlighting STT3A, CALR, and MLEC as frequently selected by multiple algorithms. In the GSE5281 dataset, the top three genes associated with TMEM59 were APP, MLEC, and STT3B. Figure 6G details the number of genes selected by each algorithm, with the Genetic Algorithm achieving a 100% selection rate and XGBoost selecting 9 genes, indicating an 81.8% selection rate. The heatmap in Figure 6H illustrates the normalized importance of each gene across algorithms, with APP, MLEC, and STT3B consistently showing high importance. Figure 6I confirms these findings, emphasizing APP, MLEC, and STT3B as core genes frequently selected by multiple algorithms. Across both datasets, MLEC is consistently identified as one of the top three genes associated with TMEM59.

The Venn diagram analysis identified 44 shared TFs among the genes GFAP, FOS, MLEC, and TMEM59 (Figure 7A), as detailed in Supplementary Table 1. This set of TFs is indicative of a regulatory network that may be involved in the biological processes associated with these genes. The rigorous application of machine learning algorithms to the GSE5281 dataset has yielded a refined list of 17 key TFs from an initial pool of 44 predicted TFs. This selection was based on the consistent recognition by multiple algorithms, indicating their potential significance in AD. The 17 TFs identified as critical include MAX, SMC3, ZBTB7A, TCF12, SMARCC1, EP300, SMARCA4, BRD1, EGR1, BRD9, RXRA, MBD4, ARID2, FLI1, TRIM24, SRF, and CTCF. As depicted in Figure 7B, the Elastic Net and Lasso algorithms demonstrated the highest gene selection rates, suggesting their effectiveness in this context. Figure 7C provides a comparative view of the number of genes selected by each method, highlighting the variability in gene selection across different algorithms. Figure 7D presents a heatmap that ranks gene importance across various methods, offering a visual representation of the relative significance of each TF. Figure 7E specifically emphasizes the core biomarkers (MAX ranked top), selected by at least four methods, underscoring their potential role as key regulators in AD.

The analysis leveraging the GSE48350 dataset, as depicted in Figure 7F, demonstrates that the Elastic Net algorithm selected the highest number of genes (37), followed by Lasso (32), and Genetic Algorithm and Adaptive Lasso (20 each). In contrast, the Random Forest and XGBoost algorithms selected fewer genes, indicating a lower frequency of gene importance attribution by these methods. Figure 7G provides a comparative overview of the number of genes selected by each method, with the bar chart visually representing the variability in gene selection across different algorithms. Figure 7H presents a heatmap that ranks gene importance across various methods, offering a visual representation of the relative significance of each TF. Notably, genes such as STAG1, MAZ, BRD9, and FLI1 were consistently identified as important across multiple methods, as highlighted in Figure 7I, which also shows the frequency of gene selection by the algorithms. The Venn diagram in Figure 7J illustrates the intersection of key TFs identified from two distinct datasets, GSE5281 and GSE48350. The intersection, indicating the common TFs significant to both datasets, consists of 8 factors: MAX, ZBTB7A, BRD9, RXRA, MBD4, ARID2, FLI1, and TRIM24.

The analysis of the GSE5281 dataset (Figures 8A–D) has been conducted with a focus on eight TFs, utilizing machine learning models to predict Alzheimer’s disease risk. The results are summarized as follows: Figure 8A presents the ROC curves for three predictive models. The logistic regression model achieved a test AUC of 0.8444, indicating good model performance. The random forest model demonstrated a slightly higher discriminative ability with a test AUC of 0.8715, while the XGBoost model showed a test AUC of 0.8601. These AUC values suggest that all models are effective in distinguishing between disease and control samples, with the random forest model performing the best among the three. In Figure 8B, the boxplot analysis reveals significant differences in the expression levels of TFs between control and disease groups. MAX, in particular, shows a notable higher median expression in control samples compared to disease samples, suggesting its potential role in the disease’s pathogenesis.

Figure 8C features a clinical prediction nomogram that integrates the influence of various TFs to estimate disease risk. The nomogram is based on a logistic regression model with an AUC of 0.985, indicating excellent predictive power. For instance, a patient with a MAX expression level of 6.5 would receive 30 points, which, when combined with points from other factors, contributes to a total score that predicts disease risk. Figure 8D illustrates the average feature importance across the three models, with MAX being the most influential, followed by ZBTB7A, BRD9, and RXRA. The normalized mean importance scores highlight MAX’s predominant role in the model’s predictive accuracy, with a score of 1.000, significantly higher than the other factors.

The current study (Figures 8E–H) analyzed the expression patterns and diagnostic potential of eight TFs (MAX, ZBTB7A, BRD9, RXRA, MBD4, ARID2, FLI1, and TRIM24) in AD using the GSE48350 dataset, which comprises 253 samples from postmortem brain tissues. Figure 8E illustrates the receiver operating characteristic (ROC) curves for the three machine learning models. The logistic regression model demonstrated the highest area under the curve (AUC) of 0.772, followed by RF (AUC = 0.734) and XGBoost (AUC = 0.707). The performance gap between training and testing datasets was minimal for all models, indicating robust generalizability. Figure 8F presents the differential expression analysis of five TFs (ARID2, BRD9, FLI1, MAX, and RXRA) between control and disease groups. Significant differences in expression levels were observed for all analyzed TFs. Notably, MAX exhibited elevated expression in the disease group compared to the control group. Figure 8G displays a clinical prediction nomogram based on the logistic regression model, with an AUC of 0.74. The nomogram assigns points to each TF based on its expression level, and the total points are converted into a risk probability. MAX contributes significantly to the risk score, alongside BRD9 and ARID2. Specifically, higher expression levels of MAX are associated with an increased risk probability, underscoring its role as an important predictor in the model. Figure 8H highlights the average feature importance of the eight TFs across the three machine learning models. BRD9 ranked highest with a normalized importance of 1.000, followed by ARID2 (0.860) and MAX (0.633). The integrated importance scores reflect the relative contribution of each TF to the diagnostic model. MAX emerges as one of the most critical features, with its importance supported by both its differential expression and its substantial contribution to the model’s predictive power.

The diagnostic performance of machine learning algorithms was evaluated using the GSE5281 dataset, with a focus on achieving optimal diagnostic efficacy with fewer gene combinations. Figure 9A illustrates a radar plot comparing key performance metrics across algorithms. Logistic regression demonstrated superior performance in accuracy, specificity, and F1 score, while XGBoost and SVM linear showed strong AUC values. Figure 9B highlights algorithm stability through boxplots of AUC distributions. Logistic regression and SVM linear exhibited the highest median AUC values with minimal variability, indicating robustness. Figure 9C displays a performance matrix heatmap for different gene combinations and algorithms. Notably, logistic regression achieved an AUC of 0.947 using only three genes (MAX, APP, MLEC), and an AUC of 0.948 with five genes (MAX, APP, MLEC, TMEM59, MGAT3). Even with a single gene (MAX), logistic regression demonstrated an AUC of 0.898, underscoring its diagnostic utility. The findings highlight the importance of MAX as a critical feature in gene panels, with logistic regression emerging as the most stable and accurate algorithm for minimizing the number of genes while maximizing diagnostic efficacy.

The analysis of the GSE48350 dataset, comprising 253 samples, evaluated the diagnostic performance of machine learning algorithms across different gene combinations related to N-glycosylation and AD. Figure 9D illustrates the multi-metric performance of algorithms, with Naive Bayes achieving the highest mean AUC of 0.723 using the five-gene combination of MAX, MGAT3, APP, STT3A, and CTSD, supported by an accuracy of 0.702, sensitivity of 0.873, specificity of 0.331, and an F1 score of 0.8. Figure 9E shows the stability of algorithms through boxplots of AUC distributions, with Naive Bayes demonstrating robust stability with a median AUC above 0.70, while algorithms such as Decision Tree and SVM Linear exhibited greater variability. Figure 9F provides a heat map matrix comparing algorithm performance across gene combinations, indicating that Naive Bayes achieved the highest AUC of 0.723 with five genes (MAX, MGAT3, APP, STT3A, CTSD). Even with fewer genes, Naive Bayes maintained strong diagnostic efficacy, achieving an AUC of 0.704 with three genes (MAX, MGAT3, APP) and an AUC of 0.644 with a single gene (MAX). While the five-gene combination yielded the highest diagnostic performance, the three-gene combination offers a balance between diagnostic accuracy and clinical practicality. The MAX gene was identified as the most critical single gene in this analysis.

The analysis of Figures 10A–K provides a comprehensive overview of the key molecular features and interactions driving AD classification in the GSE5281 dataset. The random forest and XGBoost models demonstrated exceptional performance with perfect AUC values of 1.0, outperforming other algorithms such as SVM Radial (AUC = 0.976) and Elastic Net (AUC = 0.953) (Figure 10A). Feature importance analysis identified MAX, APP, MLEC, and TMEM59 as the most impactful predictors, with MAX showing the highest mean absolute SHAP value of 0.195 (Figures 10B,C). The SHAP value distribution (Figure 10B) and feature value vs. SHAP impact analysis (Figure 10E) revealed nonlinear relationships between feature expression levels and their contributions to predictions. MAX and APP exhibited strong positive impacts, with their contributions increasing sharply at higher expression levels. MLEC and TMEM59 also showed significant contributions, though their dependencies on expression levels were weaker. The force plot (Figure 10D) and individual prediction decomposition (Figure 10F) further highlighted the dominant roles of MAX, MLEC, APP, and TMEM59 in driving prediction scores, with MAX contributing the most significantly. Interaction analysis uncovered several key relationships. The interaction between MLEC and TMEM59 was highly significant (p = 0.00019, β = 2.469), with higher expression of both genes correlating with increased predicted probabilities (Figure 10I). Conversely, the interaction between MAX and MGAT3 (p = 0.0288, β = –4.491) showed a strong negative effect, where higher MAX expression reduced predicted probabilities when MGAT3 expression was low (Figure 10H). The interaction between APP and MLEC (p = 0.03068, β = –1.473) also demonstrated a significant negative effect (Figure 10G). While some interactions, such as MAX × MLEC (p = 0.3604, β = –1.439) and APP × MGAT3 (p = 0.3216, β = –0.827), did not reach statistical significance, they suggested potential regulatory trends worth further investigation (Figures 10J,K).

The analysis of N-glycosylation-related molecular features in AD using the GSE48350 dataset provided a comprehensive understanding of their roles and interactions through 11 key visualizations. The Random Forest model demonstrated exceptional classification performance with an AUC of 1.0 (Figure 11A), outperforming other algorithms such as XGBoost (AUC = 0.976) and SVM Radial (AUC = 0.852). This result was supported by SHAP value distributions (Figure 11B), which identified APP and MAX as the most impactful features due to their high SHAP values. The feature importance analysis (Figure 11C) further confirmed APP and MAX as the top predictors, followed by MGAT3 and STT3A. Force plots (Figure 11D) and waterfall plots (Figure 11F) detailed the directional contributions of these features to prediction scores, with MAX showing a strong positive impact and CTSD a significant negative impact. The non-linear relationships between feature expression values and SHAP impacts (Figure 11E) revealed that APP and MAX had positive correlations with AD classification, while STT3A demonstrated a negative correlation. MGAT3 exhibited a non-linear trend, highlighting its complex role in AD pathology. Interaction analyses uncovered significant synergies between molecular pairs. The interaction between MAX and MGAT1 (Figure 11G) showed a strong positive effect (P = 0.0016, β = 7.7), while the interaction between MGAT3 and STT3A (Figure 11I) demonstrated a highly significant negative effect (P < 0.001, β = –6.864). The interaction between MGAT3 and MGAT1 (Figure 11H) also showed a near-significant effect (P = 0.0775, β = 2.767). Other interactions, such as APP × MGAT3 (Figure 11K) and MAX × STT3A (Figure 11J), showed limited or non-significant effects (P = 0.1743, β = –2.298 and P = 0.4841, β = –1.745, respectively).