Immunometabolic transcriptomic signatures of diabetes-related genes predict kidney transplant rejection

Abstract

Objectives

Kidney transplantation remains the gold standard for end-stage renal disease, yet long-term graft survival is limited by acute and chronic rejection. Emerging evidence highlights the critical role of immunometabolism, the bidirectional relationship between alloimmunity and metabolic dysregulation, specifically in the context of post-transplant diabetes mellitus (PTDM). This study investigates whether the expression of genes traditionally associated with diabetes susceptibility can predict kidney transplant rejection.

Methods

We performed a transcriptomic reanalysis of a publicly available dataset (GSE49198) comprising 596 peripheral blood samples from kidney transplant recipients. A panel of 13 candidate genes (INSR, CAT, TNF, MMP2, TGFB1, VEGFA, IGF1, PPARG, PPARGC1A, HLA-DQB1, CTLA4, ABCA1, DPP4) was selected based on prior domain knowledge. Patients were stratified into a High-Risk group (Acute and Chronic Rejection, n = 95) and a Low-Risk group (Stable, Tolerant, Minimal Immunosuppression, n = 439). Statistical analysis included Mann–Whitney U tests with Bonferroni correction, Spearman’s rank correlation, hierarchical clustering, and 3D Principal Component Analysis (PCA).

Results

Three genes (PPARG, INSR, and DPP4) showed statistically significant differential expression between High-Risk and Low-Risk cohorts (p < 0.05, corrected). PPARG and INSR were upregulated in rejection, while DPP4 was downregulated. Inter-gene correlation analysis revealed low redundancy. 3D PCA revealed a topological distinction: Low-Risk patients formed a dense homeostatic cluster, whereas High-Risk patients exhibited vector space dispersion (‘metabolic scattering’).

Conclusions

We identified a distinct metabolic gene signature capable of predicting kidney transplant rejection. The downregulation of DPP4 with the upregulation of PPARG and INSR suggests that metabolic reprogramming and loss of homeostasis are hallmarks of immune-mediated graft failure.

Introduction

Kidney transplantation offers the best survival and quality of life outcomes for patients with end-stage renal disease (ESRD) [1]. However, despite advances in immunosuppressive therapies that have drastically reduced early acute rejection rates, long-term graft survival remains challenged by Chronic Antibody-Mediated Rejection (CAMR), Interstitial Fibrosis/Tubular Atrophy (IF/TA), and late Acute Rejection (AR) episodes [2]. Current monitoring methods, such as serum creatinine and proteinuria, often detect graft injury only after significant structural damage has occurred, necessitating the discovery of sensitive, noninvasive molecular biomarkers [3].

Concurrently, metabolic complications, particularly Post-Transplant Diabetes Mellitus (PTDM), are major risk factors for graft loss and patient mortality [4]. The relationship between alloimmunity and metabolism, a field known as immunometabolism, is increasingly recognized as bidirectional: systemic inflammation can induce insulin resistance [5], while hyperglycemia triggers oxidative stress, which promotes inflammation [6]. The clinical finding that PTDM is associated with a higher incidence of acute rejection episodes [7] suggests that the genes regulating these metabolic pathways might be differentially expressed, also in peripheral blood mononuclear cells, during an alloimmune response.

Previous transcriptomic studies, such as the meta-analysis by Baron et al. [8], have primarily focused on immune cell subsets, identifying B-cell signatures associated with operational tolerance. In this study, we hypothesize that a specific panel of diabetes-related genes can distinguish recipients with acute or chronic rejection from stable or tolerant patients. By employing a knowledge-driven feature selection approach rather than automated screening [9], we aim to ensure clinical interpretability and investigate potential molecular patterns linking transplant outcomes with diabetes pathways.

Materials and methods

Dataset and study population

Transcriptomic data were obtained from the Gene Expression Omnibus (GEO) dataset GSE49198, originally published by Baron et al. [8]. This dataset comprises 596 peripheral blood samples profiled using microarray technology. To facilitate a binary risk assessment model, the original multiclass labels were mapped into two clinical categories:

1. High Risk (Rejection): Includes patients with Chronic Rejection (CR) and Acute Rejection (AR).

2. Low Risk (Stable/Tolerant): Includes Tolerant patients (TOL), Stable patients (STA, standard immunosuppression), and Minimal Immunosuppression (MIS) patients.

Healthy Volunteers (HV) were excluded to focus the analysis specifically on the transplant context. After filtering for missing metadata and applying the grouping logic, the final cohort consisted of 534 samples (High Risk: n = 95; Low Risk: n = 439). Given sample size considerations, Acute and Chronic Rejection were grouped to represent a shared clinical phenotype of rejection-associated graft injury.

Feature selection

The number and type of features included in the model were defined exclusively based on domain knowledge. We selected 13 candidate genes implicated in diabetes susceptibility, insulin signaling, and inflammation: INSR (Insulin Receptor) [10], CAT (Catalase) [11], TNF (Tumour Necrosis Factor) [12], MMP2 (Matrix Metallopeptidase 2) [13], TGFB1 (Transforming Growth Factor Beta 1) [14], VEGFA (Vascular Endothelial Growth Factor A) [15], IGF1 (Insulin-like Growth Factor 1) [16], PPARG (Peroxisome Proliferator Activated Receptor Gamma) [17], PPARGC1A (Peroxisome Proliferator-Activated Receptor Gamma Coactivator 1 Alpha) [18], HLA-DQB1 (Major Histocompatibility Complex, Class II, DQ Beta 1) [19], CTLA4 (Cytotoxic T-Lymphocyte Associated Protein 4) [20], ABCA1 (ATP Binding Cassette Subfamily A Member 1) [21], and DPP4 (Dipeptidyl Peptidase 4) [22].

Statistical analysis

Data preprocessing and analysis were performed using Python 3.12 (pandas [23], scipy [24], seaborn [25], scikit-learn [26]).

• Descriptive statistics: Mean, Standard Deviation (SD), Median, and Interquartile Ranges (IQR) were calculated for all gene expression values, which were pre-processed (including log-transformation and normalization) according to the methodology detailed in the meta-analysis of kidney allograft tolerance [8].

• Hypothesis testing: The Shapiro–Wilk test indicated a non-normal distribution of gene expression data. Consequently, the non-parametric Mann–Whitney U test was employed to compare expression levels between High-Risk and Low-Risk groups. P-values were adjusted for multiple comparisons using the Bonferroni correction method. Significance was defined as an adjusted p-value < 0.05.

• Correlation analysis: Spearman’s rank correlation coefficient was used to evaluate pairwise relationships between genes to assess redundancy.

Visualization

• Hierarchical clustering: A clustered heatmap (Euclidean distance, Ward’s method) was generated using Z-score normalized expression values to visualize patterns across patient groups.

• Dimensionality reduction: 3D Principal Component Analysis (PCA) was performed to project samples into a lower-dimensional vector space. We compared a global model (all genes) against a restricted model using only the statistically significant genes to assess stratification capability.

Results

Cohort characteristics and gene expression distribution

The analysis included 534 transplant recipients. The distribution of the target variable was imbalanced, reflecting typical clinical prevalence, with 439 (82.2%) patients in the Low-Risk group and 95 (17.8%) in the High-Risk group. The descriptive statistics for the expression of the 13 candidate genes are summarized in Table 1.

Table 1.

Gene expression summary (log-transformed normalized units).

Gene	Mean ± SD	Median [IQR]
INSR	−0.053 ± 1.258	−0.104 [1.375]
CAT	−0.083 ± 1.160	−0.002 [1.300]
TNF	0.039 ± 1.062	−0.031 [1.199]
MMP2	0.061 ± 1.217	−0.035 [1.357]
TGFB1	−0.078 ± 1.120	−0.155 [1.288]
VEGFA	0.004 ± 1.182	−0.102 [1.495]
IGF1	0.188 ± 1.280	−0.050 [1.261]
PPARG	0.381 ± 1.924	0.017 [1.621]
PPARGC1A	0.035 ± 1.141	−0.096 [1.303]
HLA-DQB1	0.041 ± 1.141	0.014 [1.385]
CTLA4	0.161 ± 1.093	0.092 [1.513]
ABCA1	0.015 ± 1.316	−0.062 [1.553]
DPP4	−0.224 ± 1.330	−0.114 [1.466]

Differential expression analysis

The Mann–Whitney U test, corrected for multiple testing (Figure 1), identified three genes with statistically significant differences between the High-Risk and Low-Risk groups:

Figure 1.

Differential gene expression analysis between high-risk (rejection) and low-risk (stable/tolerant) kidney transplant recipients.

1. PPARG (p = 0.043): Observed to be upregulated in the High-Risk (Rejection) group compared to the Low-Risk group.

2. DPP4 (p < 0.001): Showed marked downregulation in the High-Risk group.

3. INSR (p = 0.016): Exhibited upregulation in the High-Risk group.

The remaining ten genes (CAT, TNF, MMP2, TGFB1, VEGFA, IGF1, PPARGC1A, HLA-DQB1, CTLA4, ABCA1) did not reach statistical significance after Bonferroni correction.

Correlation and clustering

Pairwise intercorrelation was computed among the diabetes-related candidate genes using Spearman’s rank correlation (Figure 2). The resulting heatmap demonstrated a lack of strong correlation between the selected features (r < 0.5 for all significant pairs), indicating that PPARG, DPP4, and INSR provide complementary, non-redundant information regarding the molecular state of the graft.

Figure 2.

Spearman’s rank inter-gene correlation matrix of 13 candidate genes.

Unsupervised hierarchical clustering was performed using the Z-score normalized expression of the gene panel (Figure 3). The resulting Clustermap highlights distinct expression patterns for the significant markers. PPARG and INSR exhibit a trend of upregulation (red hues) in the High-Risk cluster, a profile consistent with the metabolic reprogramming described in activated lymphocytes, which require enhanced nutrient uptake and anabolic pathways to sustain the immune response [27].

Figure 3.

Unsupervised hierarchical clustermap of 13 diabetes-related gene expression profiles.

3D principal component analysis

To overcome the limitations of 2D projection and better visualize patient stratification, we performed 3D PCA.

1. Global analysis (Figure 4): When analyzing the full panel of diabetes-related genes, the High Risk and Low Risk populations showed substantial overlap, indicating that the global expression profile contains significant heterogeneity unrelated to rejection status.

2. Selected biomarker analysis (Figure 5): When restricted to the three statistically significant markers (PPARG, DPP4, INSR), the 3D projection preserved the total variance of the subset (PC1: 66.7%, PC2: 22.6%, PC3: 10.7%). As expected, when PCA is applied to three variables, the three resulting components jointly explain 100% of the variance. Therefore, this total variance should be interpreted strictly as a mathematical consequence of the dimensionality, rather than an indicator of biological completeness. Despite this, the visualization revealed a topological distinction: the Low Risk (Stable) patients formed a denser core cluster, whereas the High Risk (Rejection) patients exhibited a ‘scattering’ effect, occupying a much larger volume of the PCA vector space.

Figure 4.

3D principal component analysis of the global gene profile.

Figure 5.

3D principal component analysis restricted to the selected genes (PPARG, DPP4, INSR).

Discussion

This study demonstrates that a knowledge-driven panel of diabetes-related genes can identify molecular signatures associated with kidney transplant rejection. By revisiting the transcriptomic data, we identified PPARG, DPP4, and INSR as significant differentiators of graft status. A key finding of this work is the topological distribution observed in the 3D PCA. The Low-Risk group (comprising stable and tolerant patients) clustered tightly, suggesting a conserved homeostatic regulation of these metabolic pathways. In contrast, the High-Risk group displayed high variability and dispersion. This scattering effect suggests that rejection is not defined by a single static state, but rather by a loss of metabolic coordination and control. This mirrors concepts in systems biology where disease states are often characterized by increased entropy and loss of systemic homeostatic control [28]. The downregulation of DPP4 (CD26) in the High-Risk cohort suggests a distinct immunomodulatory profile in the circulating alloimmune response. While the kidney proximal tubules are the primary source of soluble DPP4 protein, our analysis quantifies mRNA expression within peripheral blood mononuclear cells. Consequently, this downregulation likely reflects the shedding or transcriptional suppression of CD26 on activated T-lymphocytes, a phenomenon previously associated with chronic immune stimulation and T-cell exhaustion [29]. Rather than a direct measure of tubular atrophy, this signals a systemic lymphocyte adaptation to the ongoing graft injury. The upregulation of PPARG and INSR indicates a state of systemic metabolic reprogramming within the recipient’s immune compartment rather than the graft tissue itself. Activated lymphocytes undergo significant metabolic shifts (immunometabolism) to sustain the bioenergetic demands of rejection. The upregulation of INSR suggests that circulating immune cells in rejecting patients are priming their machinery for enhanced glucose uptake, mirroring the metabolic avidity required for effector T-cell function [30]. Similarly, PPARG upregulation, traditionally viewed as anti-inflammatory [31], may represent a compensatory feedback mechanism by the host immune system attempting to dampen the systemic inflammatory surge triggered by the allograft rejection. According to our findings, classic immune transcripts such as TNF and CTLA4 did not discriminate High- vs Low-risk groups. This can be related to different factors. First of all, TNF activity is usually up-regulated in acute and strongly active processes, conditions not fully present in the cohort studied [32].

Moreover, CTLA4 usually is associated with regulatory and tolerance processes and not in effector rejection processes [33,34]. The absence of significance of these immune markers might also support the concept of the role of immunometabolism specificity in modulating immune cell function. It is important also to remember that our study focused on blood sample analyses. A difference between blood and graft microenvironment could be hypothesize. All these concepts could be deeply explored in prospective studies, possibly involving a higher number of patients.

Unlike our metabolic findings, Roedder et al. characterize kSORT as a transcriptional profile driven by monocytes that reflects the activation of the allogeneic immune response [35]. Similarly, the Common Rejection Module (Khatri et al.) constitutes a network involved in immune cell trafficking and cellular movement linked to the regulation of the immune response [36]. Finally, Rabant et al. describe the CXCL9 urinary level as a diagnostic marker of acute T cell-mediated rejection, while identifying CXCL10 as a surrogate for subhistological inflammatory burden [37]. While these established biomarkers primarily detect the presence, trafficking, and activity of immune effectors, our panel captures the metabolic reprogramming and bioenergetic support required to sustain this immune function.

Our findings complement the original analysis by Baron et al. [8], which identified a B-cell signature in tolerant patients. While their study focused on tolerance, our analysis highlights that the rejection phenotype may carry a metabolic footprint. From a translational perspective, integrating this metabolic signature into noninvasive monitoring protocols could refine risk stratification, potentially identifying graft injury before functional decline occurs.

This study has several limitations: firstly, it is retrospective and relies on transcriptomic data from peripheral blood, which may not perfectly mirror the intra-graft environment. Additionally, we lacked specific clinical data regarding the glucose status (e.g. HbA1c) of these patients at the time of sampling. Therefore, we cannot definitively separate gene expression changes caused by rejection from those caused by prodromal diabetes. However, the exclusion of Healthy Volunteers ensures that the comparison is strictly between transplant phenotypes.

Conclusion

In summary, we identified a metabolic gene signature composed of PPARG, INSR, and DPP4 that correlates with kidney transplant rejection risk. The analysis suggests that rejecting patients are characterized by metabolic instability and scattering rather than a uniform profile. These markers provide a bridge between alloimmunity and metabolic regulation, supporting the concept that graft rejection involves systemic metabolic reprogramming. Further validation in prospective cohorts is required to determine if these markers can serve as early predictors of rejection or potential therapeutic targets.

Funding Statement

This work was supported by the PNRR2023 (Piano Nazionale di Ripresa e Resilienza) ‘Post-transplant diabetes outcomes prediction through machine learning and deep phenotyping (PerCeive)’ – PNRR-MCNT2-2023-12378319.

Ethics approval statement

The study involves the re-analysis of publicly available transcriptomic data (GSE49198). As this study utilizes anonymized, open-access data, specific ethical approval or patient consent was not required for this secondary analysis.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The data that support the findings of this study are openly available in the Gene Expression Omnibus (GEO) repository (reference number GSE49198) and are associated with the publication available at https://doi.org/10.1038/ki.2014.395.