Identification and validation of biomarkers related to mitochondria-associated endoplasmic reticulum membranes in type 2 diabetes mellitus using peripheral blood transcriptomics

Abstract

Background

The pathological hallmarks of Type 2 diabetes mellitus (T2DM) are impaired insulin sensitivity and insufficient insulin secretion. As the primary insulin source, pancreatic β-cell decline or dysfunction is key to T2DM progression. Disrupted mitochondria-associated endoplasmic reticulum membranes (MAMs) may compromise β-cell viability and function. This study aimed to identify MAMs-related biomarkers in T2DM pathogenesis.

Methods

Data were obtained from public databases, and the biomarkers related to MAMs in T2DM were identified by differential expression analysis, WGCNA, supervised machine learning, and expression validation. Subsequently, a nomogram for predicting the prevalence of T2DM was developed, and the performance was evaluated. Additionally, we conducted immune infiltration analysis, GSEA, and molecular docking were performed to analyze the underlying mechanisms of the identified biomarkers. Finally, RT-qPCR was used to further validate the expression trends of these biomarkers.

Results

Three key biomarkers—DUSP26, SLC15A1, and TBX1—were discovered, and the nomogram developed using these markers exhibited strong predictive accuracy for T2DM risk. Interestingly, these biomarkers were predominantly associated with the olfactory transduction pathway and neuroactive ligand–receptor interactions. Additionally, five distinct immune cell types were identified (p < 0.05). Among these, Th2 cells showed the highest positive correlation with activated CD4 T cells (r = 0.45), whereas activated dendritic cells displayed the strongest negative correlation with activated CD4 T cells (r = −0.42). Furthermore, all 3 biomarkers displayed favorable binding abilities with all 3 therapeutic agents for T2DM (< −5.0 kcal/mol), suggesting the potential of biomarkers in the treatment of T2DM. Ultimately, the trend of 3 biomarker expression in the clinical samples was consistent with the GSE184050 and GSE15932, with up-regulated expression, revealing the reliability of biomarker identification.

Conclusion

The biomarkers DUSP26, SLC15A1, and TBX1 related to MAMs in T2DM were identified, which supplied a theoretical basis for T2DM-related mechanistic studies and clinical treatment.

Supplementary Information

The online version contains supplementary material available at 10.1186/s40001-025-03681-2.

Background

Diabetes mellitus (DM) is a chronic endocrine disorder characterized by persistently elevated blood glucose levels. Type 2 diabetes mellitus (T2DM), accounting for over 90% of all cases, is essentially a result of the combined effects of insulin resistance and progressive failure of pancreatic β-cell function. The superimposition of genetic susceptibility, environmental exposure, and unhealthy lifestyles further exacerbates the high prevalence of the disease [1, 2]. Global epidemiological data indicating that T2DM has become a major public health burden with approximately 589 million adults (aged 20–79) worldwide suffer from diabetes, with over 90% of them having type 2 diabetes mellitus (T2DM), according to the International Diabetes Federation (IDF)Diabetes Atlas 11th Edition (2025) (https://diabetesatlas.org/). Global medical spending related to diabetes will reach 1.015 trillion US dollars in 2024, an increase of 338% compared to 17 years ago [1, 4]. The current clinical treatment system still faces many unresolved core challenges: existing drugs (such as metformin, thiazolidinediones, SGLT2 inhibitors) can alleviate symptoms by improving blood glucose control, but they cannot reverse the degenerative damage of pancreatic β-cells or prevent the continuous progression of insulin resistance; some patients remain in a state of poor blood glucose control due to drug tolerance or heterogeneous treatment responses, leading to a high incidence of chronic complications such as cardiovascular and cerebrovascular diseases, nephropathy, and neuropathy, significantly increasing the risk of premature death [3, 5]. More importantly, the molecular mechanisms underlying the onset of T2DM have not been fully elucidated, especially at the subcellular level, where the relationship between key pathways regulating insulin signaling and energy metabolism and disease progression remains a research gap, resulting in a lack of treatment methods that can precisely target the pathological root cause. These unmet clinical needs and research gaps in mechanisms highlight the urgency of identifying new regulatory targets for T2DM and understanding its underlying molecular mechanisms, and also provide an important logical basis and value orientation for research focusing on the mitochondria-associated endoplasmic reticulum membranes (MAMs), a cellular functional hub.

The mitochondrial-associated endoplasmic reticulum membranes (MAMs) are the crucial molecular regulatory hubs that connect the above pathological mechanisms with therapeutic needs. MAMs are special membrane microdomains with liquid ordered phase characteristics (a physical state of membrane structure that facilitates the aggregation of signaling molecules), dynamically formed at the contact sites between mitochondria and the endoplasmic reticulum [6]. As specialized organelle connection structures, they mediate bidirectional communication between the two [7]. Their core functions include regulating calcium flux, transporting lipids, maintaining the balance of reactive oxygen species, regulating mitochondrial morphology, and mediating programmed cell death (i.e., apoptosis, a key pathway for orderly cell death), and these functions precisely cover the core links of the T2DM pathological process: Studies have confirmed that calcium signal disorders, imbalance of reactive oxygen homeostasis, endoplasmic reticulum stress, and abnormal autophagy are key driving factors for the formation of peripheral tissue insulin resistance and the decline of pancreatic β-cell function [12–14]. And all these abnormalities directly depend on the structural integrity and functional normality of MAMs. However, there is currently no clear specific association mechanism of the key genes of MAMs in T2DM patients in the current field, and MAMs have not been developed as an intervention target to develop precise treatment plans—this not only leads to a key gap in the research of T2DM pathological mechanisms but also makes the existing treatments unable to break through the limitation of "only controlling symptoms, not blocking progression." Therefore, this study focuses on the association mechanism between MAMs and T2DM, which not only can improve the research paradigm of T2DM pathophysiology but also has the potential to provide scientific basis for developing new precise targets that can block disease progression.

This study employed an integrated bioinformatics approach, leveraging GEO database resources and MAM-related gene sets to systematically identify MAM biomarkers in T2DM. The multi-omics strategy incorporated differential expression analysis, WGCNA, machine learning, nomogram construction, immune infiltration profiling, GSEA, molecular docking, and RT-qPCR validation. These analyses provide novel insights into the role of MAMs in T2DM pathogenesis and establish a theoretical framework for clinical translation, with potential implications for advancing T2DM therapeutic strategies and improving clinical decision-making.

This study aims to identify systemic immune-related transcriptional biomarkers associated with the pathogenesis of type 2 diabetes by analyzing the peripheral blood transcriptome that mainly reflects the state of the immune system. We hypothesize that MAM dysfunction in metabolic organs can trigger systemic responses, which are reflected in the white blood cell transcriptome, thereby providing an accessible observation window for revealing the complex pathophysiological mechanisms of this disease.

Methods

To systematically identify MAMs-related biomarkers for T2DM and clarify their functional mechanisms and clinical application potential, this study integrated multiple bioinformatics and experimental methods. The overall workflow follows a progressive logic of “data foundation”—“candidate gene screening”—“biomarker confirmation”—“multi-dimensional validation”—“mechanism”—“application exploration,” forming a complete evidence chain from gene to clinical application. The core role of each method in this workflow is clarified as follows:

Data foundation: rely on the GSE184050 (training set) and GSE15932 (validation set) datasets (peripheral blood samples of T2DM patients and healthy controls) to ensure the clinical relevance of subsequent analyses;

Candidate gene screening: use differential expression analysis (DESeq2) to screen T2DM-related DEGs, and combine WGCNA to identify MAMs-related key module genes, thereby locking in candidate genes that are “both MAMs-associated and T2DM-differentially expressed”;

Biomarker confirmation: apply machine learning algorithms (LASSO regression, SVM-RFE) to screen core biomarkers from candidate genes, ensuring the specificity and robustness of biomarkers;

Multi-dimensional validation: verify biomarker reliability through three dimensions—external dataset validation (GSE15932) and experimental validation (RT-qPCR);

Mechanism and application exploration: use functional enrichment analysis (GO/KEGG), GSEA, GeneMANIA, and immune infiltration analysis to clarify the biological functions and regulatory mechanisms of biomarkers, through nomogram construction to evaluate their T2DM predictive value, and molecular docking to explore their potential in T2DM drug development.

Data sources

The datasets were all obtained from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/). A total of 131 samples with age and gender information were included in this study, of which 115 (87.8%) were derived from the GSE184050 dataset and used as the model training set, and 16 samples (12.2%) with age and gender information were from the GSE15932 dataset and served as an independent validation set to ensure the objectivity of model validation (Additional File 1). Specifically, the training set (platform: GPL11154) contained 50 peripheral blood samples from patients with T2DM (T2DM cohort) and 65 healthy individuals (control group), respectively, with a case–control ratio of approximately 1:1.3. The validation set (platform: GPL570) included 8 peripheral blood samples from T2DM patients and 8 from healthy individuals, with a case–control ratio of 1:1. Moreover, according to the literature [15], the 28 mitochondria-associated endoplasmic reticulum membranes-related genes (MAM-RGs) were acquired. The flowchart of this study is provided in Additional File 2.

Differential expression analysis

Using DESeq2 (v 1.38.0) [16], differential expression profiling was performed on the GSE184050 dataset, employing stringent selection criteria (adjusted p-value < 0.05 and absolute log2FC > 0.5) to identify differentially expressed genes (DEGs). Subsequent data visualization was achieved using ggplot2 (v 3.3.6) [17] to construct a volcano plot highlighting the top 10 most dysregulated genes (5 up- and 5 down-regulated) based on log2FC magnitude, with the corresponding gene labels. Parallelly, ComplexHeatmap (v 2.14.0) [18] was employed to generate a clustered heatmap that systematically displayed the expression profiles of these key DEGs across samples, facilitating comparative pattern recognition.

Weighted gene co-expression network analysis (WGCNA)

To acquire the most relevant module genes for MAMs, the WGCNA was performed. Firstly, according to the expression of genes in the GSE184050, the ssGSEA algorithm (v 1.42.0) [19] was adopted to rank the genes in order of their expression from largest to smallest. Following this, the plage algorithm (v 1.42.0) was employed to compute ssGSEA scores for MAMs across T2DM and control cohorts. The Wilcoxon test was then utilized to evaluate differences in ssGSEA scores between disease and control groups (p < 0.05). To identify potential outlier samples, hierarchical clustering analysis was performed on T2DM samples using the WGCNA package (v 1.71) [20]. Then, in order to construct scale-free networks, soft threshold (power) and scale-free fit indices (R²) were computed, and the power with mean connectivity converging to 0 was selected (R² = 0.85). Then, the clustering tree was constructed, and the modules were acquired by cutting them according to the minModuleSize of genes 200, utilizing a shear tree algorithm (parameter = 0.4). Moreover, to gain the most relevant module genes for MAMs, the WGCNA package (v 1.71) was applied for Pearson analysis to acquire correlation coefficients (cor) (|cor|> 0.3, p < 0.05). The modules exhibiting the most marked positive and negative relevances were selected. More importantly, hypergeometric/Fisher's exact test (with BH-FDR correction) was used to evaluate the overlap between module genes and core functional gene sets. Moreover, the relevance between module genes and modules, as well as the relevances between module genes and module traits, was analyzed. The selection criteria for identifying gene modules with the strongest associations were established using stringent correlation thresholds. For negative correlations, modules were selected when both |cor.geneModuleMembership| exceeded 0.8 and |cor.geneTraitSignificance| surpassed 0.35. Conversely, positive correlations required |cor.geneModuleMembership|> 0.8 coupled with |cor.geneTraitSignificance|> 0.5. Following this threshold-based screening, key module genes were identified. Subsequently, candidate genes were derived from the intersection between DEGs and these key module genes using the ggvenn package (v 0.1.9) [21].

Functional enrichment and hub gene discovery of diabetes-associated candidate genes

The candidate genes used for functional enrichment analysis were rigorously screened through two sequential steps to ensure their relevance to both T2DM and MAMs. First, DEGs were identified from the GSE184050 dataset via “DESeq2” (v 1.38.0) with stringent criteria (adjusted p-value < 0.05 and absolute log2FC > 0.5) as described in Sect. "Differential expression analysis". Second, key MAMs-related module genes were obtained through WGCNA (v 1.71): after constructing a scale-free co-expression network and identifying gene modules, modules strongly associated with MAMs were selected based on Pearson correlation coefficients (|cor|> 0.3, p < 0.05), followed by filtering genes within these modules using thresholds of |cor.geneModuleMembership|> 0.8 and |cor.geneTraitSignificance|> 0.35 (for negative correlations) or > 0.5 (for positive correlations) as detailed in Sect. "Weighted gene co-expression network analysis (WGCNA)". Finally, the candidate genes for enrichment analysis were determined as the intersection of the aforementioned DEGs and MAMs-related key module genes, using the ggvenn package (v 0.1.9) to visualize and extract the overlapping gene set.

To comprehensively analyze the functional properties of candidate genes, we performed extensive GO and KEGG pathway enrichment analyses using the clusterProfiler toolkit (v 4.6.2) [22], with a significance threshold set at p < 0.05. Analysis results comprised the following: (1) graphical representation of the top five significantly enriched GO terms within three ontological categories (biological processes, cellular components, and molecular functions), and (2) comprehensive listing of statistically significant KEGG pathways. To construct protein–protein interaction networks, gene identifiers were queried against the STRING database (https://string-db.org/) applying a medium-stringency interaction score cutoff (> 0.15), with subsequent elimination of isolated nodes and network visualization using Cytoscape (v 3.91) [23]. Topological analysis was subsequently performed using CytoHubba [24] with four distinct centrality algorithms (Degree, MNC, DMNC, MCC). Consensus hub genes were identified through intersection of the top 20 candidates from each algorithm, with resultant overlaps graphically represented via ggvenn (v 0.1.9).

Identification of biomarkers

Least Absolute Shrinkage and Selection Operator (LASSO) regression, leveraging its L1 regularization property, can shrink the coefficients of redundant genes in high-dimensional gene expression data to zero, enabling embedded variable selection. It excels at identifying parsimonious and reproducible features from high-dimensional, multi-correlated features, which aligns with the characteristics of the gene data in this study [25]. As a complementary wrapper method, SVM-RFE iteratively eliminates low-information features based on a supervised SVM classifier. It is particularly suitable for transcriptomic data with “small sample size and high feature dimensionality,” and can efficiently identify compact gene sets with discriminative power [26]. The combined use of these two methods reduces method-specific bias, ensuring that the screened core biomarkers of MAMs-related T2DM are more robust and accurately tailored to the research objectives. To further screen candidate genes, the LASSO logistic regression was first applied to the candidate genes (family = “binomial,” type.measure = “deviance” (default)). The glmnet package (v 4.1-4.1) [27] with 10-fold cross-validation was used to determine the optimal penalty parameter λ, which corresponds to the minimum mean cross-validation error. At this λ value, genes with non-zero regression coefficients were retained as potential LASSO-derived features. Additionally, the SVM-RFE algorithm implemented in the caret package (v 6.0–93.0) [28] was employed to identify feature genes through 5-fold cross-validation, with optimal features determined at the point of lowest error rate. The intersection of genes selected by both LASSO and SVM-RFE was then obtained using the ggvenn package (v 0.1.9) to define candidate biomarkers. Expression patterns of these candidates were compared between disease and control groups across the GSE184050 and GSE15932 datasets using ggplot2 (v 3.3.6), and biomarkers demonstrating significant differential expression (Wilcoxon test, p < 0.05) with consistent trends in both datasets were identified.

Construction and evaluation of a nomogram

To evaluate the predictive capacity of the identified biomarkers for T2DM risk, a nomogram model was constructed using the rms package (v 6.5.0) [29] with the GSE184050 dataset. Model calibration was assessed through calibration curves to verify prediction accuracy. Additionally, decision curve analysis (DCA) was performed via the RMDA package (v 1.6) [30] to examine the clinical utility of the nomogram.

GSEA and GeneMANIA analysis

To elucidate the functional significance of the identified biomarkers in T2DM pathogenesis, we initially performed gene correlation analysis on the GSE184050 dataset using the psych R package (v 2.2.9) [31], wherein genes were systematically ranked based on their correlation coefficients with each biomarker. Subsequently, pathway enrichment analysis was conducted through GSEA implemented in clusterProfiler (v 4.6.2), employing the KEGG gene set ("c2.cp.kegg.v7.4.symbols.gmt") from MSigDB (https://www.gsea-msigdb.org/gsea/msigdb) as reference. Significant pathways were filtered using stringent criteria (|NES|> 1 and adjusted p-value < 0.05), with the top five enriched pathways graphically represented via enrichplot (v 1.18.0) [32]. For deeper functional annotation, we leveraged GeneMANIA (http://www.genemania.org/) to construct a PPI network encompassing both the biomarkers and their predicted functional partners, followed by topological analysis and visualization in Cytoscape (v 3.91).

Immune infiltration analysis

To investigate immune microenvironment changes in diseased samples, utilizing 28 immune cell types [33], the ssGSEA algorithm (v 1.42.0) was employed to assess cellular abundance and infiltration enrichment scores across disease and control cohorts in GSE184050. Subsequently, Wilcoxon testing was implemented to identify significantly altered immune populations between groups (p < 0.05). Additionally, Spearman correlation analysis was performed using the psych package (v 2.2.9) to examine associations among altered immune populations and their relationships with biomarkers (|cor|> 0.3, p < 0.05).

Construction of the molecular regulatory network

To investigate the transcription factors (TFs) that had interactions with biomarkers, the NetworkAnalyst database (http://www.jaspar.genereg.net) was applied to anticipate the TFs targeting biomarkers. Additionally, to investigate the functions and regulatory mechanisms of biomarkers at a deeper level, the MicroCosm Targets database (MicroCosm, https://mycocosm.jgi.doe.gov/mycocosm/home) and MicroRNA Target Prediction database (miRDB, https://mirdb.org/) were applied to anticipate the microRNAs (miRNAs) that target biomarkers, respectively. Subsequently, the predicted miRNAs were intersected to get the target miRNAs. Eventually, the Cytoscape (v 3.9.1) was applied to visualize the results.

Molecular docking

To investigate the potential binding capacity between the proteins encoded by the biomarkers and commonly used therapeutic drugs (dapagliflozin, pioglitazone, and metformin) for T2DM, computational simulations of molecular docking were performed based on their three-dimensional protein structures. However, this does not imply that these biomarkers necessarily act as direct therapeutic targets of the drugs in vivo. The results served only to provide hypothetical directions for subsequent experimental validation. Protein structures were retrieved from the RCSB PDB database (https://www.rcsb.org/), whereas drug molecules were acquired from PubChem (https://pubchem.ncbi.nlm.nih.gov/). Among these, the retrieved 3D protein structures were sorted by resolution, and the one with the highest resolution was selected for subsequent analysis. Molecular docking was conducted via the CB-Dock2 online platform (https://cadd.labshare.cn/cb-dock2/), and binding affinity calculations were employed to assess interaction strength.

The CB-Dock2 server implements automated blind protein–ligand docking calculations through the following four steps: (1) Input and quality inspection of protein and ligand structural data; (2) Structural data preprocessing, including repairing missing atoms and hydrogen atoms, and removing water molecules and other heteroatoms; (3) Protein pocket detection, spatial parameter estimation, molecular docking based on Vina scoring, and molecular docking based on template matching and conformation transfer; (4) Comprehensive processing and visualization of calculation results.

Reverse transcription-quantitative PCR (RT-qPCR)

For experimental validation of computational results, RT-qPCR analysis was conducted to evaluate biomarker expression levels. Ethical approval was obtained from Shenzhen Qianhai Shekou Free Trade Zone Hospital (approval no. 2025-KY-015-01), and informed consent was secured from all participants. Blood specimens were acquired from matched participant groups: 5 treatment-naïve T2DM patients and 5 age-/sex-matched healthy controls (n = 10 total). RNA isolation was accomplished using Trizol reagent (Ambion, 15596018CN) [34], followed by cDNA generation via the SweScript First Strand cDNA Synthesis System (Servicebio, G3333-50) as per manufacturer's instructions. Expression quantification utilized the 2−ΔΔCt method [35] with GAPDH serving as the internal reference. Statistical analysis employed Student's t-test via GraphPad Prism 8.0 [36] to establish significance (p < 0.05), with primer details provided in Additional file 3.

Statistical analysis

The R software (v 4.2.2) was adopted to proceed with bioinformatics analysis. The Wilcoxon test and t-test were utilized to evaluate paired sample comparisons, with statistical significance defined as p-value < 0.05.

Results

Identification of 47 candidate genes related to MAMs

Differential expression analysis identified 3678 DEGs between disease and control cohorts, with the T2DM cohort exhibiting 103 up-regulated and 3575 down-regulated DEGs (Fig. 1a). The top 10 most significantly up-regulated and down-regulated genes demonstrated distinct expression profiles (Fig. 1b). Subsequently, MAM-RGs scores were significantly reduced in the T2DM cohort compared to controls (p < 0.05) (Fig. 1c), indicating an association between MAMs and T2DM pathogenesis. After confirming there were no samples needed to be excluded, all T2DM samples were employed to WGCNA (Fig. 1d). The optimal soft threshold was 18 (R² = 0.85) (Fig. 1e), and 4 modules were identified (including the gray module) (Fig. 1f). Next, the turquoise module (cor = 0.31, p = 8 × 10⁻⁴) and the blue module (cor = −0.3, p = 1 × 10⁻³) were identified as key modules related to the MAMs-RGs scores (Fig. 1g). Furthermore, in this study, the turquoise module was found to be significantly enriched in the Calcium signaling pathway (k = 71, OR = 2.16, FDR = 0.0288); this result is consistent with the role of MAMs as a hub for Ca²⁺ signaling. In contrast, no overlap was observed between the “GO: Lipid transport” term and the modules, primarily due to insufficient coverage of this GO term in the gene background of the current platform/after processing. The genes within these key modules were further screened using module membership and gene significance. Following this, 208 key module genes were identified, among which 74 and 134 genes were obtained from the blue and turquoise modules, respectively (Fig. 1h, i). Finally, a sum of 47 candidate genes were retained from the intersection of DEGs and key module genes (Fig. 1j).

Fig. 1.

Identification and WGCNA was performed in the GSE184050. a Volcano plots of DEGs in the GSE184050. b Heatmap plots of DEGs in the GSE184050.c The inter-group differences in ssGSEA scores of MAMs. d Clustering dendrogram of 115 samples after removing outlier in the GSE184050. e Determination of soft-threshold power in the WGCNA for the GSE184050. f Clustering dendrogram of soft-threshold power in the WGCNA for the GSE184050.g Heatmap showing correlation between modules and disease characteristics for the GSE184050.h Scatter plot of the correlation between module members and gene significance in key modules(negative correlation).i Scatter plot of the correlation between module members and gene significance in key modules(positive correlation).j Gene intersection Venn diagram

Related functional pathways of candidate genes and identification of 16 key genes

Functional annotation of the candidate genes revealed significant enrichment across 336 GO terms, with distinct distribution patterns observed among the three major categories: 299 terms in biological processes (BP), 9 in cellular components (CC), and 28 in molecular functions (MF) (Additional file 4). Detailed analysis identified prominent enrichment for specific functional annotations, including pattern specification processes under BP, membrane-spanning components in CC, and DNA-binding transcription factor binding activities within MF (Fig. 2a). Pathway analysis further demonstrated significant involvement in two key metabolic/signaling pathways: lysine degradation and tight junction regulation (Fig. 2b). PPI network analysis identified 41 candidate gene products forming an interconnected network, with SOX9, PITX2, PAX2, and ISL1 emerging as highly connected nodal proteins (Fig. 2c). Integration of results from four distinct topological algorithms through intersection analysis ultimately identified 16 consensus hub genes (Fig. 2d).

Fig. 2.

Functional enrichment analysis of DEGs in the GSE184050. a GO analysis of DEGs in the GSE184050. b KEGG analysis of DEGs in the GSE184050. c PPI protein interaction network in the GSE184050. d Venn diagram of key genes in the GSE184050

Identification of biomarkers DUSP26, SLC15A1, and TBX1

The 9 LASSO-feature genes were acquired after analysis (lambda.min = 0.0116) (Fig. 3a), including PRRX1, IRX2, ISL1, CDX1, SOX7, DUSP26, SLC15A1, PAX1 and TBX1. Meanwhile, 11 SVM-RFE-feature genes were obtained when the model reached the lowest error rate, including TBX1, DUSP26, PAX1, IRX2, CBLN4, PRRX1, CDX1, RAX, LHX6, PITX2, and SLC15A1 (Fig. 3b). Then, 7 candidate biomarkers were acquired by intersection, namely, TBX1, DUSP26, PAX1, IRX2, PRRX1, CDX1, and SLC15A1 (Fig. 3c). Notably, 3 biomarkers whose expression trends were the same in GSE184050 and GSE15932 were then acquired, and in the T2DM cohort, the DUSP26, SLC15A1, and TBX1 were markedly up-regulated (p < 0.05) (Fig. 3d, e).

Fig. 3.

Identification of core diagnostic biomarker based on intersection genes among different genes lists. a Lasso Cox regression analysis was conducted to screen candidate biomarkers based on 16 key genes in the GSE184050. b SVM-RFE analysis was conducted to screen candidate biomarkers in the GSE184050. c Candidate biomarker Venn diagram. d The expression levels of candidate biomarkers vary in the GSE184050. e The expression differences of candidate biomarkers in the GSE15932

Construction of a nomogram with outstanding predictive ability

Based on DUSP26, SLC15A1, and TBX1, predicting incidence of T2DM was displayed in the nomogram, with the higher the total point, the higher the probability of T2DM (Fig. 4a). The calibration analysis revealed no statistically significant deviation between predicted and observed outcomes (p > 0.05), indicating excellent agreement and validating the nomogram's predictive accuracy (Fig. 4b). Furthermore, DCA demonstrated superior clinical utility of the nomogram, with net benefit values consistently exceeding: (1) the "treat all" and "treat none" strategies, and (2) individual biomarker predictions across the entire probability threshold range (Fig. 4c). In summary, the biomarker-integrated nomogram demonstrated robust predictive performance for T2DM incidence, providing a clinically actionable tool for diabetes risk stratification and early intervention.

Fig. 4.

Evaluation of the specificity and accuracy of the biomarkers by constructing a nomogram and drawing calibration curves and decision curves. a Construction of biomarker-based nomogram. b Calibration Curve. c Decision Curves

Crucial functional pathways and associated interactions of biomarkers

DUSP26, SLC15A1, and TBX1 were enriched in 27, 24, and 25 pathways, respectively. Notably, the DUSP26, SLC15A1, and TBX1 were all markedly enriched in olfactory transduction, neuroactive ligand–receptor interaction, calcium signaling pathway, ECM receptor interaction, and retinol metabolism pathways, suggesting a possible function of biomarkers in T2DM (Fig. 5a–c) (Additional file 5). Moreover, the biomarker interacted with the associated 20 genes mainly through physical interactions and predicted modes, and the 20 genes were mainly involved in oligopeptide transport, peptide transmembrane transporter activity, and positive regulation of muscle cell differentiation. (Fig. 5d). To summarize, understanding how these interactions and the pathways that may be involved can deepen the understanding of the role that biomarkers may play in the occurrence and development of T2DM, and supply theoretical guidance for achieving targeted therapies for T2DM through the combination of other relevant genes.

Fig. 5.

Understanding the potential functions of biomarkers by KEGG analysis and GeneMANIA analysis. a KEGG analysis of DUSP26. b KEGG analysis of SLC15A1. c KEGG analysis of TBX1. d GGI Network of DUSP26 and SLC15A1and TBX1

Altered immune microenvironment of T2DM and the associations with biomarkers

Comparative immune profiling between disease and control groups revealed significant differences in immune cell infiltration (Fig. 6a). Five immune cell types exhibited marked variation (p < 0.05): activated CD4 T cells, activated dendritic cells, central memory CD4 T cells, Th1 cells, and Th2 cells. Among these, activated dendritic cells and Th1 cells demonstrated significantly higher ssGSEA scores in the T2DM cohort (Fig. 6b). Correlation analysis identified robust associations between differential immune cell populations. Th2 cells showed the strongest positive correlation with activated CD4 T cells (r = 0.45, p < 0.001), while activated dendritic cells exhibited the strongest negative correlation with activated CD4 T cells (r = −0.42, p < 0.001). Notably, biomarker-immune cell correlations followed consistent trends. SLC15A1 displayed the strongest positive correlation with activated dendritic cells (r = 0.30, p < 0.001), whereas TBX1 showed the strongest negative correlation with central memory CD4 T cells (r = −0.35, p < 0.05) (Fig. 6c). These findings suggest that the identified biomarkers may contribute to T2DM pathogenesis by modulating the immune microenvironment.

Fig. 6.

Infiltration differences in 28 types of immune cells between diseased and normal samples in T2DM. a The relative proportion of immune cell content in the T2DM and non- T2DM samples. b Infiltration differences in 28 types of immune cells between T2DM and non- T2DM samples. c Correlation heatmap of different immune cells and different immune cells with biomarkers. d TF-mRNA-miRNA network diagram of biomarkers

Complex molecular regulatory network of biomarkers

A total of 16 TFs targeting biomarkers were acquired, among which 4, 9, and 8 TFs were predicted by DUSP26, SLC15A1, and TBX1, respectively. At the same time, 30 miRNAs targeting biomarkers were predicted by the MicroCosm and miRDB databases, of which 10 were predicted by DUSP26, 14 by SLC15A1, and 7 by TBX1. Then, the 49 nodes and 52 interaction pairs were included in the TF–mRNA–miRNA network, and compared with the TBX1 and DUSP26, the SLC15A1 revealed a more complex regulatory network (Fig. 6d). Besides, the SLC15A1 and TBX1 co-predicted 4 TFs, MAX, USF1, FOXC1, and USF2, and also co-predicted YY1 with DUSP26. Furthermore, among the 3 biomarkers, only TBX1 and DUSP26 co-predicted the miRNA (hsa-miR-139-5p). Overall, this supplied the basis for the targeted regulation of biomarkers.

Strong binding energies between dapagliflozin, pioglitazone, and metformin with biomarkers

The molecular docking results showed that the proteins encoded by the three biomarkers had good binding activity (|total score|> 5.0 kcal/mol) with dapagliflozin, pioglitazone, and metformin (Table 1). Among them, the best binding activity was displayed between DUSP26 (pdb_00004hrf) and dapagliflozin (−8.6 kcal/mol), and SLC15A1 (pdb_00007pmx) displayed consistent binding ability with pioglitazone and dapagliflozin (−7.8 kcal/mol), as well as the high binding ability was also displayed between TBX1 (pdb_00004a04) and pioglitazone (−8.3 kcal/mol) (Fig. 7a–i). Molecular docking to evaluate the binding potential of DUSP26, SLC15A1, and TBX1 protein products with drugs commonly used in the treatment of T2DM further reveals the potential of biomarkers in the clinical management of T2DM.

Table 1.

Molecular docking of dapagliflozin, pioglitazone and metformin and DUSP26, SLC15A1 and TBX1

Biomarkers—Drugs	Binding energry(kcal/mol)	Center
DUSP26-pioglitazone	−8	−7, −9, 22
DUSP26-dapagliflozin	−8.6	−7, −9, 22
DUSP26-metformin	−5.2	−4,−29,10
SLC15A1-pioglitazone	−7.8	88, 112, 120
SLC15A1-dapagliflozin	−7.8	88, 112, 120
SLC15A1-metformin	−5	88, 112, 120
TBX1-pioglitazone	−8.3	−35, 39, −26
TBX1-dapagliflozin	−7.8	−35, 39, −26
TBX1-metformin	−5.4	−34, 41, −23

Fig. 7.

Verification of the binding capabilities of three common T2DM drugs to target biomarkers through molecular docking. a Overall and detailed diagrams molecular docking of the DUSP26–Pioglitazone. b Overall and detailed diagrams molecular docking of the DUSP26–Dapagliflozin. c Overall and detailed diagrams molecular docking of the DUSP26–Metformin. d Overall and detailed diagrams molecular docking of the SLC15A1–Pioglitazone. e Overall and detailed diagrams molecular docking of the SLC15A1–Dapagliflozin. f Overall and detailed diagrams molecular docking of the SLC15A1–Metformin. g Overall and detailed diagrams molecular docking of the TBX1–Pioglitazone. h Overall and detailed diagrams molecular docking of the TBX1–Dapagliflozin. i Overall and detailed diagrams molecular docking of the TBX1–Metformin.

Validation of clinical samples

The RT-qPCR results of clinical samples showed that DUSP26, SLC15A1, and TBX1 were all up-regulated in T2DM (Fig. 8a–c) and the expression of SLC15A1 and TBX1 reached a significance level (p < 0.05). Overall, the expression of these biomarkers was consistent with the expression trend of GSE184050 and GSE15932, which indicated that the biomarkers obtained through the analysis were reliable and could also indicate that these biomarkers have a diagnostic ability for T2DM and control samples.

Fig. 8.

Verification of the expression differences of the target biomarker between diabetic and non-diabetic patients through PCR. a The expression differences of DUSP-26 in diabetic and non-diabetic patients. b The expression differences of SLC15A1 in diabetic and non-diabetic patients. c The expression differences of TBX1 in diabetic and non-diabetic patients

Discussion

DM is a persistent endocrine-metabolic condition marked by chronic hyperglycemia resulting from defects in insulin secretion, insulin action, or both [1]. MAMs, serving as functional contact sites between the ER and mitochondria, regulate multiple cellular functions and affect the normal physiological functions of the pancreas [7]. At present, the specific regulatory mechanism of MAMs in the occurrence of T2DM remains unclear, which limits our in-depth understanding of their potential therapeutic value in T2DM. This study systematically identified the key biomarkers of MAMs in T2DM through integrating transcriptomics, immune infiltration analysis, and molecular docking technology. These biomarkers, namely, DUSP26, SLC15A1, and TBX1, were used to construct a network model, which demonstrated excellent clinical predictive performance. Molecular docking analysis further confirmed that these biomarkers have high-affinity binding characteristics with various anti-diabetic drugs. GSEA revealed that these biomarkers were significantly enriched in the olfactory transduction and ligand–receptor interaction pathways involved in neural activity. Meanwhile, immune microenvironment analysis found that five immune cell subpopulations, including CD4 T cells and activated dendritic cells, showed significant changes in T2DM patients. These findings not only elucidated the important role of MAM dysfunction in the pathogenesis of T2DM from a multi-omics perspective but also identified three potential intervention targets for developing novel therapeutic strategies targeting MAM, with significant translational medical value.

This study found that the expression levels of DUSP26, SLC15A1, and TBX1 were significantly up-regulated in the T2DM group (p < 0.05), suggesting that these three genes may serve as excellent discriminatory biomarkers for the disease. DUSP26, a 24 kDa atypical dual-specificity phosphatase (DUSP), is a group of heterogeneous protein phosphatases. It can dephosphorylate the phosphorylated tyrosine and phosphorylated serine/threonine residues within a substrate, and is involved in various biological and pathological processes such as cell growth, differentiation, inflammation, and apoptosis [37]. Consistent with prior findings, our study observed elevated DUSP26 expression in T2DM patients. Nicolas et al. demonstrated that DUSP26 can regulate the survival of pancreatic β cells and glucose homeostasis. Pharmacological inhibition of DUSP26 can improve hyperglycemia in diabetic mice and protect human pancreatic cells from cell death [38]. Similarly, Huang et al. demonstrated that the up-regulated DUSP26 expression accelerated kidney injury and dysfunction, leading to the accumulation of reactive oxygen species (ROS) in the renal cortex or glomeruli of mice, excessive production of hydrogen peroxide, and a decrease in superoxide dismutase (SOD) levels. The generation of ROS activated the mitogen-activated protein kinase (MAPKs) signaling pathway in the kidneys, exacerbating the onset of diabetic nephropathy [39]. Notably, in db/db diabetic mice, cardiac-specific overexpression of DUSP26 demonstrated significant cardioprotective effects, including improved left ventricular systolic function (evidenced by elevated ejection fraction and fractional shortening) along with attenuated myocardial remodeling (reduced fibrosis and hypertrophy). At the mitochondrial level, DUSP26 overexpression promoted bioenergetic efficiency through increased ATP generation, while enhancing mitochondrial dynamics via augmented fusion processes and structural stabilization [40]. These observations underscore the tissue-specific regulatory functions of DUSP26 in diabetic complications, positioning it as a promising molecular target for organ-protective interventions in T2DM.

SLC15A1 (PepT1), a proton-coupled oligopeptide transporter, plays a key role in glucose homeostasis through intestinal protein sensing [41]. Helen et al. demonstrated that pharmacological inhibition of PepT1 disrupts the improvement in glucose tolerance induced by a high-protein diet, emphasizing its role in metabolic regulation. Emerging evidence indicates that PepT1-dependent nutrient sensing pathways in the proximal small intestine demonstrate glucoregulatory properties in preclinical models of developing metabolic dysfunction. Specifically, these preabsorptive protein detection mechanisms have been shown to enhance systemic glucose control during the early phases of both insulin resistance and diet-induced obesity. This suggests that upper intestinal luminal nutrient sensing may represent a previously underappreciated modulator of whole-body glycemic regulation [42]. In pancreatic β cells, a high glucose environment can induce an increase in the expression of ornithine decarboxylase 1 (ODC1), and ODC1 has a synergistic regulatory relationship with the gene expression of SLC family members (including SLC15A1) [43]. Huang et al. demonstrated that up-regulating the expression of Lnc-SLC15A1-1 could induce an increase in the CXCL10/CXCL8 cascade reaction signal, leading to vascular endothelial damage and aggravating endothelial cell inflammation––a potential mechanism linking diabetes and chronic kidney disease (DM/CKD) to elevated cardiovascular disease (CVD) risk [44]. These findings collectively highlight the dual role of SLC15A1 in metabolic regulation and vascular pathology, positioning it as a promising yet context-dependent therapeutic target for diabetes-associated complications.

TBX1, a T-box transcription factor critical for mesoderm development, belongs to the T-box family, serves as a key regulator in the growth and differentiation of multicellular organisms [45]. Although there is no direct evidence proving the role of TBX1 in T2DM, previous studies have shown that the expression of TBX1 in mature fats is necessary for the full activity of UCP1 and the maintenance of body temperature during cold periods, but the overexpression of TBX1 is insufficient to drive obesity and prevent the development of diet-induced obesity. In addition, fat TBX1 is necessary for appropriate subcutaneous fat insulin signaling, β-adrenergic sensitivity, and glucose homeostasis in the body [46]. In addition, Liu [47] et al. demonstrated that exercise can regulated the differentiation of Th17 cells through the STAT3/RORγt signaling pathway, thereby improving the aortic endothelial function of diabetic mice. Moreover, the expression of TBX1 was enhanced in the exercise samples [48], which regulates the expression of the Vegfr2 gene and affects the differentiation of cardiomyocytes, smooth muscle cells, and endothelial cells [49]. This suggests that enhancing the expression of TBX1 may improve the endothelial function in diabetic patients. Ozcan L et al. demonstrated that TBX1 inhibits the expression of MK2, thereby improving the glucose homeostasis of obese mice and enhancing insulin sensitivity. At the same time, it reduces plaque necrosis and increases the thickness of the fibrous cap at the lesion site of the mouse aortic root, indicating that TBX-1 not only improves the metabolic status but also enhances the stability of atherosclerotic plaques [50]. These data indicate that TBX1 is a novel regulator of adipose tissue function and systemic metabolism, with potential for further clinical development. Based on the above literature, this study infers the potential mechanisms of DUSP26, SLC15A1, and TBX1 in T2DM (Additional File 6).

GSEA analysis revealed that the core MAM-related biomarkers (DUSP26, SLC15A1, TBX1) in this study were significantly enriched in the "olfactory transduction" and "neuroactive ligand–receptor interaction" pathways. These results are not isolated functional annotations but are directly related to the role of biomarker-mediated MAM dysfunction in the pathological process of T2DM. For instance, in the "olfactory transduction" pathway, the abnormal expression of DUSP26 may be involved in the early regulation of T2DM-related cognitive impairment through a dual mechanism: first, DUSP26 can activate the JNK signaling pathway to regulate vesicle transport and protein processing in neurons [51], and olfactory neuron signal transmission is highly dependent on vesicle-mediated neurotransmitter release, with functional abnormalities directly associated with olfactory dysfunction [52, 53]; second, as a core hub for calcium signal transmission, MAM dysfunction leading to calcium homeostasis imbalance can exacerbate the abnormal excitability of olfactory neurons, and it can be speculated that DUSP26 indirectly affects this process by regulating signal transduction in the MAM region [54–56]. Previous studies have confirmed that olfactory dysfunction and cognitive decline are associated in T2DM patients, and MAM dysfunction is a key link between metabolic abnormalities and neurodegenerative changes [57, 58]. All these suggest that DUSP26-mediated MAM dysfunction may become an early molecular event in T2DM cognitive impairment by regulating the olfactory transduction pathway.

Concurrently, immune microenvironment shifts, particularly altered activated dendritic cells and central memory CD4⁺ T cells, correlate with chronic inflammation and β-cell autoimmunity. These cells secrete pro-inflammatory cytokines (e.g., TNF-α, IL-6), which exacerbate insulin resistance and β-cell dysfunction. The observed biomarker-immune cell correlations suggest their potential role in modulating immune infiltration, thereby offering therapeutic targets to restore immune homeostasis. These findings collectively highlight the complex interplay between metabolic dysregulation, immune dysfunction, and tissue-specific pathology in T2DM, establishing a multi-dimensional framework for understanding disease progression and developing targeted interventions.

In this study, we identified DUSP26, SLC15A1, and TBX1 as biomarkers of T2DM based on peripheral blood mRNA expression data, and further explored the potential of their encoded proteins to bind commonly used anti-diabetic drugs by molecular docking analysis. The results showed that all the three proteins showed good binding energy, suggesting that these proteins may be involved in the drug pathway, but whether they are the direct targets of the drug still needs to be verified by further experiments. As a sodium-glucose cotransporter-2 (SGLT2) inhibitor, dapagliflozin functions by enhancing urinary glucose excretion, while pioglitazone acts as a PPARγ agonist that improves insulin sensitivity and lipid profiles [59]. Clinical evidence demonstrates pioglitazone's efficacy as adjunct therapy in poorly controlled T2DM patients on metformin and dapagliflozin. Versus placebo, pioglitazone showed greater HbA1c reduction at 24 weeks (p < 0.05) without increasing hypoglycemia risk, while also improving triglycerides, HDL cholesterol, and insulin resistance [59]. These biomarker–drug interactions confirm their biological relevance and therapeutic potential. The strong binding affinities suggest they could directly target or modulate current anti-diabetic drugs, enabling personalized T2DM treatments. Furthermore, the binding energy only reflects static binding capacity and has no absolute positive correlation with in vivo efficacy. For instance, high binding-energy complexes may dissociate rapidly due to excessively high k_off values. In subsequent studies, surface plasmon resonance (SPR) experiments can be conducted to verify the dissociation rate of complexes, thereby enhancing the biological significance of the docking results.

The detection based on fasting blood glucose and HbA1c is currently the cornerstone of T2DM diagnosis and management, with the absolute advantages of simplicity and low cost. The Nomogram prediction model developed in this study based on DUSP26, SLC15A1, and TBX1 is not intended to replace these conventional diagnostic methods, but to provide a supplementary tool that can reflect the potential pathophysiological state of T2DM. This group of MAMs-related biomarkers reveals the possible roles of intracellular organelle communication disorders and immune–metabolic interactions in the disease. Therefore, the value of this nomogram model lies in deepening the diagnosis from merely “judging blood glucose levels” to “identifying molecular subtypes of the disease.” Based on this model, we anticipate its potential clinical applications including early risk warning, disease subtype classification, and complication prediction. For individuals in the prediabetic stage, this model may identify those whose blood glucose indicators are still at the critical value but have significant molecular-level disorders, thus enabling earlier intervention. Moreover, T2DM is highly heterogeneous. This nomogram may help distinguish subtypes characterized by MAMs dysfunction and immune disorders, providing a basis for future targeted therapy. Additionally, whether patients with high model scores are more prone to developing complications such as diabetic nephropathy and cardiovascular diseases is worthy of further research verification. This model can serve as a potential risk stratification tool.

This study was mainly based on peripheral blood samples. The reasons for using peripheral blood samples are as follows: Firstly, peripheral blood, as an alternative tissue in metabolic disease research, is of great value in clinical and basic research due to its accessibility and the representativeness of systemic information [60, 61]. Secondly, type 2 diabetes is a systemic disease characterized by chronic low-grade inflammation. Immune cells in circulation are not only affected by this pathological state but also play an active role in it [62]. Dysfunctional MAMs in metabolic organs can release damage-associated molecular patterns (DAMPs) and alter the circulating cytokine environment, which in turn can reprogram gene expression in blood cells [63–65]. Thirdly, and most importantly, previous studies have successfully identified and validated transcriptomic biomarkers of other organ-specific pathologies (such as neurodegenerative diseases) in blood samples, indicating that the peripheral blood transcriptome can capture meaningful biological signals related to distant organ dysfunction [66, 67]. Therefore, the MAM-related gene expression signatures we discovered in the blood may represent a repeatable and systematic reflection of the underlying pathophysiological processes occurring in metabolic tissues during the progression of T2DM. This approach not only provides a feasible strategy for biomarker discovery but also highlights the interconnection between the metabolic and immune systems in diabetes. However, we are also aware of the limitation that this study mainly relies on transcriptomic data from peripheral blood. Therefore, in the future, we will validate the roles of these MAM-related biomarkers in primary metabolic tissues (such as islets and liver) from animal models or human donors.

However, this study also has some limitations. First, the biomarkers are derived from the peripheral blood transcriptome and may not fully reflect their expression and function in key target tissues of T2DM, such as the pancreas and liver. Future studies need to verify this with tissue samples. Second, the clinical sample size for RT-qPCR validation is small, and protein expression also needs to be verified. Third, although a combination of machine learning algorithms and independent datasets was used for validation, there is still a risk of overfitting in feature selection with a limited sample size, and the generalizability of the model needs to be further evaluated in prospective studies. Finally, the predictive results of molecular docking and the specific mechanism of action of biomarkers in T2DM still need further experimental verification. To address these limitations, we plan to actively collect key tissue samples such as the pancreas and liver in the future and compare them with peripheral blood data to confirm the consistency of biomarker expression in tissues. At the same time, we will expand the clinical sample size and conduct protein-level validation to enhance the reliability of the results. In terms of model optimization, we plan to incorporate multi-center prospective cohort data to improve the generalization ability of the prediction model and reduce the risk of overfitting. In addition, the regulatory mechanisms suggested by molecular docking will be further analyzed through in vitro and in vivo functional experiments to clarify the specific action pathways of biomarkers in T2DM and provide a basis for subsequent intervention studies.

Conclusion

This study employed transcriptomics, immune infiltration analysis, and molecular docking technology to identify MAM-associated biomarkers in T2DM-DUSP26, SLC15A1, and TBX1––which were significantly linked to disease pathogenesis. Using bioinformatics methods, the MAM-related genes from GEO database were analyzed and verified; GSEA showed that they were significantly enriched in pathways such as olfactory transduction; immune infiltration analysis revealed that there were five types of differentially infiltrating immune cells in T2DM; these biomarkers showed strong binding affinity with T2DM therapeutic drugs. In summary, our findings establish MAM dysfunction as a novel mechanism in T2DM and provide a theoretical foundation for targeted therapies. However, database limitations and incomplete mechanistic exploration warrant further clinical and functional studies to elucidate these biomarkers' roles in MAM regulation and T2DM progression.

Abbreviations

DM

Diabetes Mellitus

T2DM

Type 2 diabetes mellitus

T1DM

Type 1 Diabetes Mellitus

MAMs

Mitochondria-associated membranes

GEO

Gene Expression Omnibus

MAM-RGs

Membranes-related genes

DEGs

Differentially expressed genes

WGCNA

Weighted Gene Co-expression Network Analysis

DCA

Decision curve analysis

TFs

Transcription factors

miRNAs

MicroRNAs

RT-qPCR

Reverse transcription-quantitative PCR

BP

Biological processes

CC

Cellular components

MF

Molecular functions

DUSP

Dual-specificity phosphatase

ROS

Reactive oxygen species

SOD

Superoxide dismutase

MAPK

Mitogen-activated protein kinase

ODC1

Ornithine decarboxylase 1

CKD

Chronic kidney disease

CVD

Cardiovascular disease

SGLT2

Sodium-glucose cotransporter-2

GO

Gene Ontology

GSEA

Gene Set Enrichment Analysis

KEGG

Kyoto Encyclopedia of Genes Genomes

Author contributions

S.F.L and Y.Q.Y contributed to the study design, data processing, and analysis. Q.J.L,R.F.T and J.H.Y contributed equally to this work. S.F.L contributed to manuscript writing. Y.Q.Y contributed to the design and oversight of this study, manuscript writing, and revision. All authors read and approved the final manuscript.

Funding

This research received no external funding.

Data availability

The datasets generated during and analysed during the current study are available in the [Gene Expression Omnibus (GEO)] database (GSE184050, GSE15932), [https://www.ncbi.nlm.nih.gov/geo/].

Declarations

Ethics approval and consent to participate

This study was conducted in accordance with the Declaration of Helsinki, and approved by the Institutional Review Board of Shenzhen Qianhai Shekou Free Trade Zone Hospital (approval no. 2025-KY-015-01).

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.