Diverse regulated cell death patterns and immune traits in kidney allograft with fibrosis: a prediction of renal allograft failure based on machine learning, single-nucleus RNA sequencing and molecular docking

Abstract

Objectives: Post-transplant allograft fibrosis remains a challenge in prolonging allograft survival. Regulated cell death has been widely implicated in various kidney diseases, including renal fibrosis. However, the role of different regulated cell death (RCD) pathways in post-transplant allograft fibrosis remains unclear.

Methods and Results: Microarray transcriptome profiling and single-nuclei sequencing data of post-transplant fibrotic and normal grafts were obtained and used to identify RCD-related differentially expressed genes. The enrichment activity of nine RCD modalities in tissue and cells was examined using single-sample gene set enrichment analysis, and their relations with immune infiltration in renal allograft samples were also assessed. Parenchymal and non-parenchymal cells displayed heterogeneity in RCD activation. Additionally, cell–cell communication analysis was also conducted in fibrotic samples. Subsequently, weighted gene co-expression network analysis and seven machine learning algorithms were employed to identify RCD-related hub genes for renal fibrosis. A 9-gene signature, termed RCD risk score (RCDI), was constructed using the least absolute shrinkage and selection operator and multivariate Cox regression algorithms. This signature showed robust accuracy in predicting 1-, 2-, and 3-year allograft survival status (area under the curve for 1-, 2-, and 3-year were 0.900, 0.877, 0.858, respectively). Immune infiltration analysis showed a strong correlation with RCDI and the nine model genes. Finally, molecular docking simulation suggested rapamycin, tacrolimus and mycophenolate mofetil exhibit strong interactions with core RCD-related receptors.

Conclusions: In summary, this study explored the activation of nine RCD pathways and their relationships with immune traits, identified potential RCD-related hub genes associated with renal fibrosis, and highlighted potential therapeutic targets for renal allograft fibrosis.

Introduction

Kidney transplantation is often the last resort for patients with end-stage kidney disease. However, post-transplantation allograft fibrosis remains a challenge. This is clinically termed interstitial fibrosis and tubular atrophy (IFTA) under the Banff classification [1]. Immunological responses of various etiologies greatly affect allograft fibrosis. The crosstalk between specific sub-clusters of injured parenchyma and non-parenchymal cells in the fibrotic niche has garnered considerable recent attention, though the specific mechanisms remain unclear [2–4].

Regulated cell death (RCD) maintains homeostasis, but disproportionate responses can lead to unnecessary cell loss, resulting in glomerulosclerosis and tubular atrophy, and inflammation and fibrosis stem from excess leukocyte and myofibroblast infiltration[5]. RCD modalities – such as ferroptosis, necroptosis, and pyroptosis – have been well-established in the development of kidney diseases, including ischemia–reperfusion injury (IRI), acute kidney injury to chronic kidney disease (AKI-CKD) transition, and IRI-induced fibrosis. This occurs through the promotion of necroinflammation, epithelial–mesenchymal transition, and release of profibrotic factors [6–10]. Senescence, characterized by permanent cell growth arrest and the secretion of senescence-associated secretory phenotype (SASP), induces maladaptive repair and renal fibrosis [11]. Autophagy-dependent cell death, the process by which cytoplasmic materials are transported and degraded in the lysosome, leads to cell death. The role of autophagy in renal fibrosis seems controversial, depending on the cell type involved [12, 13]. NETotic cell death and PANoptosis, primarily induced by pathogen-associated molecular patterns, have increasingly shown involvement in kidney diseases via autoimmunity and autoinflammation[14–16]. Cuproptosis and disulfidptosis are relatively newly termed cell-death modalities. Cuproptosis is characterized by intracellular copper accumulation, aggregation of mitochondrial lipoylated proteins, and destabilization of Fe-S cluster proteins [17], whereas disulfidptosis is caused by aberrant disulfide SLC7A11^high cells, leading to the collapse of cytoskeleton proteins and F actin [18]. Recent exploratory bioinformatic research has also revealed the participation of cuproptosis and disulfidptosis in organ fibrosis [19, 20]. Additionally, the interplay of multiple RCD pathways contributes to maladaptive repair and tissue regeneration [8]. For instance, Martin-Sanchez et al. demonstrated that ferroptosis initiates the first wave of regulated necrosis, followed by necroptosis as the subsequent wave during AKI [21].

Given RCD’s regulatory roles in immune infiltration and inflammatory responses, and the potential interrelation of diverse RCD forms, investigating how they synchronously affect different cell types is crucial. Although bioinformatic analyses of the relationship between the immune microenvironment or individual RCD and renal allograft fibrosis have been conducted [22–24], exploration is lacking into non-presumptive RCD activation and immune infiltration in post-transplant fibrotic allografts using techniques such as machine learning, single-nucleus RNA sequencing (snRNA-seq), and molecular docking.

In this study, we used transcriptome and snRNA data from patients with allograft fibrosis to identify RCD-related markers for kidney fibrosis. We examined and compared the activity of nine RCD modalities in samples with and without IFTA to identify the dominant RCD and further determine their activity level in different cells. Using multiple machine-learning algorithms, we also identified nine genes that could distinguish IFTA. From this, we constructed an RCD risk score based on these genes, which could predict graft survival. Additionally, molecular docking simulation was used to repurpose existing drugs that could target core identified genes. In summary, RCD showed varied activity in different groups and cell types. The identified RCD model genes and risk score could serve as therapeutic targets for treating allograft fibrosis, and conventional immunosuppressants may have therapeutic effects by targeting core RCD-related molecules.

Methods

Data collection

Four microarray datasets were obtained from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/): GSE22459 [25], GSE53605 [26], GSE76882 [27], and GSE21374 [28]. All datasets were from human kidney allograft biopsies, classified into normal (non-IFTA) and IFTA groups (Table 1). The GSE21374 dataset additionally included clinical information such as kidney transplant rejection events and the time interval from biopsy to rejection or the next follow-up.

Table 1.

Clinical information of GEO datasets.

GEO identifier	Platform	Tissue	Time post-transplantation	Normal versus IFTA	Group
GSE22459	GPL570	Kidney allograft biopsies	1 year	25 vs 40	Diagnostic
GSE53605	GPL571	Kidney allograft biopsies	2 years	18 vs 10	Diagnostic
GSE76882	GPL13158	Kidney allograft biopsies	1 year	99 vs 81	Diagnostic
GSE21374	GPL570	Kidney allograft biopsies	1 to 31 years	NA	Prognostic
GSE195718	GPL24676	Kidney allograft biopsies	≥15-months	3 vs 6	snRNA-seq

GEO: gene expression omnibus; IFTA: interstitial fibrosis and tubular atrophy; snRNA-seq: single-nucleus RNA sequencing.

The RCD gene list was summarized using various sources, including the Kyoto Encyclopedia of Genes and Genomes (KEGG) Database, the Molecular Signature Database, the Reactome Database, the GeneCards Database, and the previously summarized gene list from Su et al. [29]. The final gene list included nine cell-death pathways: autophagy, cuproptosis, disulfidptosis, ferroptosis, necroptosis, NETosis, PANoptosis, pyroptosis, and senescence (Supplementary Table 1).

Single-nucleus RNA-sequencing (snRNA-seq) dataset GSE195718 from the GEO database included six biopsies from chronic allograft dysfunction patients with IFTA and three from patients with stable graft function with normal or nonspecific histopathology [30].

Data preprocessing

The ‘affy’ R package was used to standardize the four datasets. Using matching platform files, each gene probe was annotated. Based on the presence or absence of IFTA, the three bulk RNA-seq datasets (GSE22459, GSE53605, GSE76882) were combined to build a predictive model. The batch effect was removed using the ‘removeBatchEffect’ from the ‘limma’ R package using the default parameter (Supplementary Figure 1). ‘removeBatchEffect’ is used to correct batch effects by adjusting systematic biases in the data, removing non-biological variations caused by differences in experimental batches. The core principle of this function is the use of linear models to fit batch information and identify and remove its influence on gene expression data, allowing subsequent analysis to focus more accurately on biological differences.

For snRNA-seq data, each of the nine samples was generated as a Seurat object by the function ‘Read10 x’ using the Seurat package (version 4.3.0), then merged into one Seurat object. Based on the reference paper inclusion criteria [30], cells with more than 400 and less than 5,000 expressed genes, and samples with <2.5% mitochondrial genes, were included. Next, normalization and batch effect correction were performed by implementing Harmony (version 1.1) within the Seurat workflow using the R packages ‘Seurat’ and ‘Harmony.’ Since we had samples representing different conditions in our snRNA-seq dataset, FindConservedMarkers from Seurat was used to identify conserved cell cluster markers conserved between the two groups. A log fold change (FC) threshold of greater than 0.25 and adjusted p < 0.05 were set as the criteria for cell cluster markers. The top 10 markers with the highest average FC per cluster were extracted for cell type annotation (Supplementary Table 5). Cell clusters were then manually annotated based on the expression of these top 10 cell cluster markers, using a published human kidney atlas [31]. Lastly, the uniform manifold approximation and projection (UMAP) method was employed to generate a 2D map of the identified clusters.

Identification of RN-DEGs and functional enrichment analysis

Differential expression analysis was performed using the ‘limma’ R package to identify genes playing a crucial role in post-transplant fibrosis. The filtering criteria for differentially expressed genes (DEGs) include an adjusted p-value of less than 0.05 and |log2FC| of greater than 1. Gene set enrichment analysis (GSEA) was used to identify the functional enrichment pathways of these DEGs using the ‘GSEABase’ R package, with the gene annotations obtained via the R package ‘org.Hs.eg.db’ and the reference gene sets (c2.cp.all.v2022.1.Hs.symbols.gmt) obtained from the Molecular Signatures Database (MSigDB; https://www.gsea-msigdb.org/gsea/msigdb/collections.jsp). The false detection rate was <0.25 and the adjusted p-value was <0.05. RCD-related DEGs (RCD-DEGs) were obtained by intersecting the aforementioned DEGs and RCD genes from the list.

Analysis of immune cell infiltration

Various immune cell infiltration levels were acquired and calculated using the CIBERSORT [32] algorithm to investigate the infiltration of immune cells across conditions within the combined bulk RNA-seq cohort. We used three additional methods to verify the reliability of our results: single-sample gene set enrichment analysis (ssGSEA) via the ‘gsva’ R package, MCPcounter via the ‘IOBR’ R package, and the ImmuneCellAI algorithm from the website [33–35].

Calculation of RCD pathway activity

To quantify and explore changes in RCD pathway activity across conditions, we implemented ssGSEA for bulk microarray data and the ‘escape’ R package for snRNA-seq data [36, 37]. ssGSEA is a non-parametric method that calculates a gene set enrichment score per sample. Additionally, we used the ‘AUCell’ R package to analyze the state of RCD gene sets within snRNA-seq data. AUCell uses the area under the curve (AUC) to calculate a rank-based gene set enrichment score [38].

Identification of key gene modules related to fibrosis and RCD pathways with WGCNA

To explore key gene modules highly correlated with fibrosis and the nine RCD pathways, we employed weighted gene co-expression network analysis (WGCNA). This can cluster highly correlated genes into the same module across many samples [39]. The ‘hclust’ function was used to cluster samples and identify and remove outliers. The soft-thresholding power was then calculated using the ‘pickSoftThreshold’ function. The dynamic tree-cut method was used for module identification, and the minimum module size was set to 50. Lastly, correlation with traits was assessed. The ‘clusterProfiler’ R package was used to identify potential Gene Ontology (GO) and KEGG pathways based on the hub genes obtained from the identified key gene module.

Cell–cell communication analysis with CellChat

Cell–cell communication among the cell types was evaluated using the ‘CellChat’ package, following the vignette provided by the author [40]. Briefly, CellChatDB, a literature-supported repository of ligand–receptor complexes, was used to calculate the interactions among the 13 cell types.

Construction of fibrosis-related RCD risk score based on multiple machine-learning approaches

Machine learning has been widely used for identifying biological processes and building predictive models for risk prediction [41]. We employed seven machine-learning models to evaluate the predictive potential of RCD-DEGs and select core genes. These models included extreme gradient boosting (XGBoost), decision trees, logistic regression, naïve Bayes, K-nearest neighbors, random forests, and support vector machines (SVMs).

SVMs with the radial-basis function kernel transform non-separable problems into separable ones, making them easier to solve. This maximizes the separation between different classes by finding the widest possible margin. K-nearest neighbors is a classification approach where a data point is classified based on the known classes of the k most similar points in the training set. It then predicts unknown samples based on the category of these k samples.

Decision trees create a tree-based structure, with each node representing a decision based on a feature, and the leaves representing the predicted output. Random forests and XGBoost are another class of robust non-linear models with the advantages of providing feature importance metrics and requiring minimal hyperparameter tuning. These models are an excellent choice when understanding which features contribute most to prediction crucial for biological insight.

Naïve Bayes fits a model using Bayes’ theorem to compute the probability of each class, given the predictor values. Logistic regression via ‘glmnet’ predicts a generalized linear model for binary outcomes. It uses a linear combination of predictive variables to model the log odds of an event. The ‘tidymodels’ R package was used for modeling.

The workflow for generating signatures proceeded as follows:

1. The combined bulk microarray dataset, containing 131 IFTA samples and 142 normal samples, was randomly split into a training (75%) and a testing (25%) group to avoid overfitting.

2. Previously identified hub RCD-DEGs were used to construct predictive models. Additionally, each model was assessed based on the 10-fold cross-validation approach to reduce bias arising from imbalanced sample sizes.

3. Model discrimination ability was assessed by AUC, and calibration capability was evaluated by accuracy (proportion of data predicted correctly). Overall performance was measured using the Brier score. The model with higher accuracy and AUC and a lower Brier score was deemed optimal. Variable importance scores for the predictors were calculated and plotted using the ‘vip’ function.

4. The optimal model was validated in the testing group.

To construct a fibrosis-related RCD risk score, we subjected the RCD-related model genes from our previous analysis to univariate Cox regression. This second screening identified survival-related genes. We then used the least absolute shrinkage and selection operator (LASSO) to eliminate redundant predictors, building a prognostic model using the GSE21374 dataset. Allograft samples were randomly divided into training and validation sets in a 70:30 ratio [42–44]. Significant RCD-related model genes (HR ≠ 1 and p-value < 0.05) were filtered and put into the LASSO algorithm with 10-fold cross-validation using the ‘glmnet’ R package. Fibrosis-related RCD risk score was calculated as follows: ${\displaystyle \sum }_{i=1}^n\beta i\ast \text{exp}(i)$, with the ‘exp’ representing each gene expression and βi representing the LASSO coefficient. Using the median risk score, samples were classified into high- and low-risk groups. The Kaplan–Meier (K–M) survival curve compared the graft survival rates of recipients between the high- and low-risk groups. The time-dependent receiver operating characteristic (ROC) curve was used to assess the performance of the prognostic model.

For validation of the RCD risk score, the score was calculated for each sample from the combined microarray dataset and divided into high- and low-risk score groups based on the median RCD risk score.

Molecular docking simulation

Molecular docking is a computational method used to ­predict the binding affinity between ligands and target receptors. In this study, we repurposed traditional post-transplantation drugs, including mycophenolate mofetil, tacrolimus, and rapamycin, to identify their potential interactions with key RCD-related model genes. The crystal protein structures of three targeted RCD-related model genes – CCL5, VWF, and SOX9 – were obtained from the RCSB Protein Data Bank (https://www.rcsb.org/) [45] and AlphaFold Protein Structure Database (https://alphafold.ebi.ac.uk/) [46]. The protein conformations were modified by PyMOL 2.5.8 and AutoDockTools 1.5.7. The three key RCD-related model genes were deemed receptors, and the drugs were treated as ligands. AutoDock Vina was used to predict the binding conformation between the ligands and receptors. The binding energy was used to assess the ligand–receptor binding affinity, with affinity values of less than −7 kcal/mol indicating strong affinity, −4 to −7 kcal/mol indicating moderate affinity, and values of greater than −4 kcal/mol indicating weak interactions. The molecular interactions between the receptors and ligands were visualized using PyMOL 2.5.8.

Statistical analysis and visualization

Statistical analyses were performed using R software (version 4.2.1). The Wilcoxon rank sum test, Welch t-test, and Student t-test were used for comparison between the two groups. Spearman correlation analysis was used to determine correlations between variables. Statistical significance was defined as p < 0.05.

Results

Study workflow

We explored the roles of multiple cell-death patterns on allograft fibrosis. Additionally, we trained and validated our prognostic model on previously published microarray and snRNA-seq cohorts. A detailed workflow of the study is illustrated in Figure 1.

Figure 1.

Schematic workflow of the study. RCD: regulated cell death; snRNA-seq: single-nuclei RNA sequencing; WGCNA: weighted correlation network analysis; ssGSEA: single-sample gene set enrichment analysis; RCD-DEGs: regulated cell death-related differentially expressed genes; XGBoost: eXtreme gradient boosting; Lasso: least absolute shrinkage and selection operator.

Landscape of RCD pathways and immune characteristics in fibrotic allograft

We identified a total of 325 fibrosis-related DEGs, among which 205 were upregulated and 120 were downregulated. Allografts with IFTA exhibited a distinct scaled transcriptomic landscape compared with the normal group (Figure 2(A–B)). The complete list of fibrosis-related DEGs can be found in Supplementary Table 2. GSEA enrichment analysis revealed that these fibrosis-related DEGs are positively involved in various immune-associated pathways, such as innate immune system, neutrophil degranulation, interleukin signaling, and cytokine signaling. However, they are negatively involved in metabolism-related pathways, such as lipid metabolism, amino acid and derivative metabolism, and proximal tubule transport (Figure 2(C)).

Figure 2.

Functional analysis of fibrosis-related DEGs, RCD pathway enrichment and immune infiltration in control and IFTA groups. (A) Volcano plot displaying 120 downregulated genes and 205 upregulated genes. (B) Heatmap of the top 25 changed gene expression between control and IFTA samples. (C) GSEA enrichment analysis of fibrosis-related DEGs. (D) Bean plot of ssGSEA scores of 9 RCD pathways. (E) Immune infiltration scores between control and IFTA groups by CIBERSORT. (F) Spearman correlation analysis between RCD ssGSEA scores and various immune cells demonstrated by bubble plot. Not sig: not significant; IFTA: interstitial fibrosis and tubular atrophy; NES: normalized enrichment score; Cor: correlation coefficient. *p < 0.05; **p < 0.01; ***p < 0.001.

To understand the dynamics of multiple RCD pathways, ssGSEA revealed changes in all pathways except for ferroptosis. Autophagy, necroptosis, NETosis, PANoptosis, pyroptosis, and senescence all increased in the IFTA group. Although cuproptosis and disulfidptosis were most enriched in both groups, the IFTA group showed declined enrichment of these two pathways (Figure 2(D)). These findings suggest that RCD might play a critical role in the development of IFTA in allografts.

Since previous findings suggest that the immune response plays a key role in allograft fibrosis, we also explored differences in immune cell infiltration between the two groups. In general, an increased pro-inflammatory and cytotoxic immune response trend existed with decreased regulatory and repair cells in the IFTA group. Infiltration of activated CD4 memory T cells, γδT cells, naïve B cells, and eosinophils increased, whereas activated NK cells, Tregs, M2 macrophages, and memory B cells decreased (Figure 2(E)). Differences in the abundance of cytotoxic T cells, memory T cells, and Treg cells were verified by MCPcounter, ssGSEA, and ImmuneCell AI analysis (Supplementary Figure 2). Furthermore, the ssGSEA scores of necroptosis, NETosis, pyroptosis, PANoptosis, and senescence are positively correlated with increased infiltration of activated CD4 memory T cells, γδ T cells, and naïve B cells. They are negatively associated the decreased infiltration with activated memory B cells, NK cells, M2 macrophages, Tregs, and mast cells (Figure 2(F)). Our analysis revealed increased enrichment of necroptosis, NETosis, pyroptosis, PANoptosis, and senescence in the IFTA group. These findings were positively associated with a pro-inflammatory and dysregulated immune repertoire. Consequently, we focused our attention on these core RCD pathways.

Identification of hub RCD-related genes that drive fibrosis

We performed a WGCNA analysis to narrow down the genes correlated with both group differences and RCD. We removed outliers by setting the cut height to 120 through dynamic tree cutting (Figure 3(A)). We selected the soft-thresholding power (β) to ensure scale independence, setting it to 7 (Figure 3(B)). We then built the adjacency matrix and calculated topological overlap matrices, visualizing them through hierarchical clustering of genes and color labeling (Figure 3(C)). Lastly, we conducted a correlation analysis between module eigengenes and sample traits, revealing strong relationships between the green module and necroptosis, NETosis, PANoptosis, pyroptosis, senescence, and group differences (Figure 3(D); Supplementary Figure 2). Enrichment analysis by GO and KEGG analysis showed that module hub genes are enriched in leukocyte cell–cell adhesion, leukocyte-mediated immunity, virus infection, and allograft rejection (Figure 3(E)), emphasizing potential interactions among RCD pathways, immune pathways, pathways related to virus infection, and post-transplant renal fibrosis. The green module contained 939 genes. After overlapping the green module with RCD-DEGs, we obtained 244 genes (hub RCD-DEGs). The full list can be accessed in Supplementary Table 3.

Figure 3.

Identification Of hub RCD-related genes closely associated with allograft fibrosis by WGCNA. (A) Dendrogram and trait heatmap by WGCNA. (B) Scale-free exponent and average connectivity for each soft threshold. (C) Dendrogram of gene clusters colored by modules. (D) Heatmap displaying correlations between gene modules and group, along with ssGSEA scores of 9 RCD pathways. (E) GO, KEGG enrichment analysis of the genes extracted from the green module. (F) Venn plot of the intersection of green module genes and RCD-DEGs. RCD: regulated cell death; ssGSEA: single-sample gene set enrichment analysis; BP: biological process; CC: cellular component; MF: molecular function; KEGG: Kyoto Encyclopedia of Genes and Genomes; RCD-DEGs: regulated cell death-related differentially expressed genes.

snRNA reveals distinct RCD pathway enrichment in different cell clusters

After the initial correction of batch effects, samples from the two groups mixed harmoniously, indicating no remaining batch effects. The FindVariableFeatures function from Seurat was then used to identify highly variable genes while mitochondrial percent are well below 2.5%, indicating sufficient quality control (Supplementary Figure 3A, C, D). Following Harmony and UMAP downscaling analysis, we obtained 20 cell clusters. These were later merged and manually categorized into 13 cell types based on the top 10 gene markers for each cell cluster. These cell types were proximal tubule cell (PT), the thick ascending limb of Henle’s loop (TAL), endothelial cell (EC), collecting duct principal cell (CDP), distal convoluted cell (DCT), connecting tubule, kidney collecting duct intercalated cell, leukocyte, podocyte (POD), interstitial fibroblast, interstitial pericyte, parietal epithelial cell (PEC), and an unknown cell cluster (Figure 4(A–C); Supplementary Figure 3B). The cell proportion was similar in each sample, except for one sample with podocytes as the majority of cells (Figure 4(D)).

Figure 4.

Dissection of RCD pathway activities based on snRNA-seq. (A, B) The heatmap and violin plot demonstrate distinct gene expression patterns and cell type markers in 13 cell types, respectively. (C) The UMAP depicts 13 cell types in both normal and fibrotic samples. (D) The bar plot indicates cell proportions in each sample. (E) The heatmap shows ssGSEA scores of 9 RCD pathways in 13 cell types. (F) Global interactions of cell-cell interactions are shown in fibrotic samples. (G) The bubble plot illustrates ligand-receptor pairs of leukocytes. (H) The signals contributing to outgoing or incoming signaling across cell groups are represented. The colored bar on top indicates the total signaling strength of a cell group, while the right grey bar shows the total signaling strength of a signaling pathway. RCD: regulated cell death; UMAP: uniform manifold approximation and projection; ssGSEA: single-sample gene set enrichment analysis.

RCD pathway enrichment, as determined by ssGSEA analysis, showed varying enrichment patterns across different cell types. NETosis and pyroptosis were strongly enriched in leukocytes. CDP and PEC shared similar cell-death patterns, with senescence, necroptosis, and autophagy pathways being the most enriched. Interestingly, dominant nephron cells including PT, TAL, and DCT displayed low cell-death activity. PT only showed moderate ferroptosis activity. EC exhibited similar pyroptosis and NETosis activation to leukocytes (Figure 4(E). AUCell analysis further validated the RCD activity levels of different renal cell types (Supplementary Figure 5). Furthermore, the analysis revealed statistical differences in pathway activity between the IFTA and normal groups for each RCD pathway, except for cuproptosis (Supplementary Table 4).

We conducted cell–cell communication analysis to further illuminate the integral role of cell crosstalk in a fibrotic environment. Global interactions among these cells are visualized in Figure 4(F). The signaling patterns among cells are mostly involved in tissue repair and inflammatory response (Figure 4(H)). Specifically, leukocytes showed strong outgoing communication patterns in the SEMA4D-PLXNB1 pathway, with DCT being the most probable targeted cell. This indicates potential crosstalk between immune cells and nephron cells (Figure 4(G)).

Construction and validation of RCD-related prognostic gene signature via multiple machine-learning-based selection for long-term allograft survival prediction

We used a machine-learning-based procedure to develop a prognostic RCD index (RCDI) based on the expression profiles of the 244 hub RCD-DEGs obtained. Seven machine-learning algorithms were fitted to the combined microarray dataset, and their AUCs were calculated. The top three algorithms were XGBoost, decision tree, and LASSO regression. XGBoost demonstrated robust diagnostic capability, with a mean AUC of 0.896, Brier score of 0.14, and accuracy of 82% (Figure 5(A–B); Supplementary Table 6). XGBoost also demonstrated robust performance in the testing set, with a mean AUC of 0.891 (Figure 5(C)), Brier score of 0.13, and accuracy of 85% (Supplementary Table 6). Therefore, we chose XGBoost for gene selection and identified the top 20 genes with the highest importance (Figure 5(D)).

Figure 5.

Construction And validation of RCDI. (A, B) Comparison of the performance of 7 machine learning algorithms. (C) Performance of XGBoost in the test set measured by AUC. (D) Bar plot of the contributing variables selected by XGBoost, in descending order. (E) Univariate analysis of the model genes in the GSE21374 dataset. (F) Multivariate analysis of the genes in the GSE21374 cohort. (G, H) Regularization path and coefficient path in the lasso model. (I–J) ROC curve of RCDI in predicting 1-, 2-, and 3-year graft survival in the train and test set, respectively. (K–L) K–M curve of high or low RCDI groups in the train and test set, respectively. (M) Risk map of graft survival status based on risk group. RCDI: regulated-cell death index; ROC: receiver operating characteristic curve; TPR: true positive rate; FPR: false positive rate; boost_tree: XGBoost; decision_tree: decision tree; logistic_reg: logistic regression; naïve_Bayes: naïve bayes; nearest_neighbor: K-nearest neigbors; rand_Forest: random Forest; svm_rbf: radial basis function support vector machines.

To build a prognostic model, kidney transplant recipients from GSE21374 were randomly divided into training and testing sets in a 7:3 ratio. A univariate Cox regression analysis revealed that 18 out of 20 genes had a hazard ratio greater than 1 and were associated with graft survival (Figure 5(E)). We used LASSO regression analysis for further screening, and then constructed a prognostic RCDI model using multivariate Cox regression analysis (Figure 5(F–H)). RCDI was calculated for each patient using the formula: RCDI= 1.44 × exp(TYROBP) + 0.75 × exp(CP) + 1.31 × exp(VWF) + (−1.40) × exp(NCF4) + 1.46 × exp(CCL5) + (−1.30) × exp(RAC2) + (−2.92) × exp(DOCK2) + 0.86 × exp(SOX9) +0.61 × exp(IGHM). In the training dataset, time-related ROC curves confirmed the robust predictive ability of the RCDI model for predicting graft survival rates at 1 year (AUC= 0.902), 2 years (AUC= 0.904), and 3 years (AUC= 0.862; Figure 5(I)). In the test dataset, time-related ROC curves confirmed the robust predictive ability of the RCDI model for predicting graft survival rates at 1 year (AUC= 0.913), 2 years (AUC= 0.849), and 3 years (AUC= 0.896; Figure 5(J)). Based on the median RCDI, each sample was classified into the high- or low-risk group. The high-risk group was associated with poorer graft survival outcomes, and the expression of the nine model genes was significantly higher in this group (Figure 5(K–M)). The Sankey diagram summarizes the relationship between the high- and low-risk groups, graft loss, and graft rejection (Supplementary Figure 6).

Role of RCDI gene signature in IFTA

We used a combined microarray dataset and the snRNA-seq dataset as validation cohorts. Samples were also divided into low- and high-risk groups based on the median RCDI risk score. The IFTA group had a higher RCDI risk score, which showed statistical significance compared to the normal group (Figure 6(A)). The expression of the nine model genes was increased in the IFTA group (Figure 6(B)). The model genes and RCDI risk score were highly correlated with naïve B cells, activated CD4 memory T cells, and γδT cells; they were negatively associated with Tregs and activated NK cells (Figure 6(C)). The snRNA-seq showed that most of the model genes were highly expressed in leukocytes and endothelial cells (Figure 6(D)).

Figure 6.

Validation of RCDI in combined microarray dataset and snRNA-seq dataset. (A, B) Comparison of RCD risk score and expressions of the 9 model genes in control and IFTA samples. (C) Bubble plot of the relationship between immune cell infiltration, risk score, and model genes. (D) Feature plot showing the distribution of model genes. RCDI: regulated-cell death index; Cor: correlation coefficient. *p < 0.05; **p < 0.01; ***p < 0.001.

Immunosuppressants interacts with key fibrosis-related RCD genes by molecular docking analysis

Mycophenolate mofetil (MMF), tacrolimus, and rapamycin are commonly used immunosuppressants for post-transplantation patients. Molecular docking was performed to explore potential interactions between these drugs and key fibrosis-related RCD genes.

For CCL5, MMF had a binding affinity of −5.8 kcal/mol, primarily interacting with ASN-63 and ARG-59. Tacrolimus exhibited a stronger affinity of −7.9 kcal/mol, with interactions involving SER-5 and TYR-29. Rapamycin displayed the highest affinity at −8.4 kcal/mol, interacting with GLN-48, SER-5, and THR-7 polar contacts (Figure 7(A–C)).

Figure 7.

Molecular docking simulations. (A–C) MMF, tacrolimus, and rapamycin with CCL5, respectively. (D–F) MMF, tacrolimus, and rapamycin with SOX9, respectively. (G–I) MMF, tacrolimus, and rapamycin with VWF, respectively. MMF: mycophenolate mofetil; CCL5: C–C motif chemokine ligand 5; SOX9: SRY-box transcription factor 9; VWF: Von Willebrand factor.

For SOX9, MMF exhibited a binding affinity of −5.9 kcal/mol, interacting with ALA-74, ARG-70, and ARG-125 polar contacts. Tacrolimus had a binding affinity of −7.3 kcal/mol, interacting with LEU-105. Rapamycin had an affinity of −8.2 kcal/mol, with the key interacting amino acid being ARG-115 polar contacts (Figure 7(D–F)).

In the case of VWF, MMF showed a binding affinity of −6.8 kcal/mol, with interactions at LYS-554, GLN-793, and THR-791 polar contacts. Tacrolimus had an affinity of −9.3 kcal/mol, interacting with ALA-297 and TYR-127. Rapamycin again demonstrated the highest affinity at −10.4 kcal/mol, with interactions at ASP-1040 and SER-1042 polar contacts (Figure 7(G–I)).

These results indicate that rapamycin generally has the strongest binding affinity across all three target proteins, followed by tacrolimus, with MMF having moderate binding affinity. Overall, these findings suggest that rapamycin might be the most effective in terms of binding strength to fibrosis-related RCD proteins. This could lead to more pronounced therapeutic effects in inhibiting allograft fibrosis.

Discussion

Our study provides novel insights into the role of RCD pathways and immune characteristics in the progression of post-transplant fibrosis. Based on the hub RCD-DEGs, we developed a prognostic RCDI employing machine-learning algorithms.

Diverse RCD types have been proposed at the crossroads between injurious stimulation and immunologic response in kidney transplantation. For instance, Martin Sanchez et al. demonstrated in a folic acid-induced AKI model that the primary cause of tubule cell death – ferroptosis – triggers a secondary wave of cell death through immunogenicity, a process known as necroinflammation [47]. On the other hand, these RCD pathways are interconnected. Inhibiting one pathway may activate another, complicating targeted cell-death therapies. Therefore, mapping RCD activation synchronously is essential. Previous studies based on bioinformatic analysis have focused solely on one specific RCD type (necroptosis, pyroptosis) and fibrosis [22, 24]. We found that all forms of RCD studied, except for ferroptosis, cuproptosis, and disulfidptosis, showed statistically increased trends in enrichment scores between the two groups (Supplementary Table 4). Immune infiltration analysis revealed an overall increased activation of cytotoxic T cells and B lineage cells, which corresponds to a previous computational analysis using machine learning to identify core immune infiltrating cells in the fibrotic niche [23]. Considering that necroptosis, NETosis, pyroptosis, PANoptosis, and senescence had significant correlations with all the differentially expressed immune infiltrating cells, we focused on these RCD forms. Necroptosis was predominantly observed in CDP, whereas NETosis and pyroptosis demonstrated enriched scores in leukocytes. Surprisingly, our results showed low cell-death activity in proximal tubules. These results align with a previous study on diabetic kidney disease, which reported low cell-death activation in proximal tubules, high activation of pyroptosis and necroptosis in immune cells, and elevated autophagy pathway activity in connecting and collecting tubule cells [19]. Prior research has emphasized RCD in renal tubular epithelial cells (RTECs) as a contributing factor to the AKI-CKD transition or delayed graft function [48, 49]. Our findings further reveal that (1) RCD, particularly necroptosis, ferroptosis, and autophagy, occurring in collecting duct tubules is dominant in post-transplant allografts, and (2) RCD activation is also observed in leukocytes in post-transplant fibrosis. A previous study noted that P2Y14 on the apical membrane of collecting duct tubules contributes to ischemic AKI because it binds to its ligand UDP-glucose, causing chemokine release and neutrophil and monocyte infiltration in the kidney [50]. Overall, we highlight consistent cell-death dynamics across different kidney cells, even following the initial wave of cell death in RTEC in IRI.

We suspected that the implications of leukocyte cell death could either involve the downstream release of cytokines or be immunogenic. The latter could trigger autoimmunity or alloimmunity in kidney grafts, which may induce different forms of cell death in their target cells [6]. We revealed that the main signaling pathway of leukocytes is the SEMA4D-PLXNB1 pathway. Sema4D, also known as CD100, is a protein that belongs to class IV semaphorin. Its primary physiological functions include T-cell priming. Much attention has been given to its pro-carcinogenic role in tumor progression [51], with limited research exploring its role in fibrosis. A recent study showed that knockout of Sema4D alleviates CCl4-induced liver fibrosis, and this process was partially mediated by effector T cells and Tregs [52]. Therefore, the SEMA4D/PLXNB1 signaling pathway in leukocytes requires future investigation.

Various RCDs are associated with ischemia–reperfusion injury and its long-term complication, chronic allograft dysfunction, characterized by interstitial inflammation and fibrosis [53, 54]. In this study, we first identified core RCD genes using XGBoost, the optimal diagnostic model among seven machine-learning algorithms. We then developed a gene signature comprising nine RCD-related hub genes (TYROBP, CP, VWF, NCF4, CCL5, RAC2, DOCK2, SOX9, IGHM) using LASSO. The gene signature was further calculated as an RCDI to predict graft survival in an external cohort, which showed great accuracy in predicting graft loss at 1, 2, and 3 years. Each model gene and RCDI showed a positive correlation with activated memory CD4 T cell, γδ T cell, and naïve B-cell infiltration scores.

TYROBP, a transmembrane adaptor protein, activates NK cells and may serve as a biomarker for acute kidney transplant rejection and diabetic nephropathy identified by bioinformatic approach [55, 56]. Ceruloplasmin (CP), a metalloprotein that primarily binds copper in plasma, plays a role in the peroxidation of Fe(II) transferrin to Fe(III) transferrin. Studies have found an inverse association between the expression of CP and liver fibrosis [57]. Furthermore, CP has been linked to the protection of ischemia–reperfusion injury (IRI) in kidney transplantation by regulating iron and copper metabolism [58]. Von Willebrand factor (VWF) is crucial for hemostasis and inflammation through platelet adhesion mediation. A recent study showed that regulating VWF function can mitigate IRI-induced AKI [59]. Teixeira et al. identified VWF as a strong predictor of kidney graft failure, potentially serving as an independent risk factor due to its impact on immune endothelial activation [60]. Neutrophil cytosolic factor 4 (NCF4) and RAC2 both generate reactive oxygen species and are crucial for immunological regulation. The direct role of NCF4 in fibrosis remains unclear, with only indirect evidence suggesting that neutrophil cytosolic factor 1 reduced high-fat diet–induced steatohepatitis, a condition characterized by fibrosis [61]. RAC2 has been linked to pulmonary fibrosis, with fibrosis-related signals observed to be downregulated in Rac2 knockout mice [62, 63].

C–C Motif Chemokine Ligand 5 (CCL5) induces inflammatory responses and prompts migration and differentiation of T-cell subsets as part of the immune response. Its elevated expression in IFTA samples suggests its potential as an early biomarker for renal interstitial fibrosis [64, 65]. B cells within the allograft also produce significant amounts of CCL5 and other cytokines. Reduction of these B cells results in decreased interstitial fibrosis and fewer infiltrations of macrophages and fibroblasts [66]. Dedicator of cytokinesis 2 (DOCK2), predominantly expressed in hematopoietic cells, contributes to pulmonary fibrosis by promoting fibroblast–myofibroblast transition [67]. Similarly, SRY-Box transcription factor 9 (SOX9), a protein that recognizes the HMG-box class DNA-binding proteins, plays a crucial role in kidney epithelial repair, with its silencing leading to regeneration without fibrosis [68]. Significantly, another bioinformatic analysis echoed our findings, identifying SOX9 and DOCK2 as potential biomarkers for renal fibrosis [69]. Immunoglobulin heavy constant mu (IGHM) has shown promise in predicting graft loss at 3 years, with its differential expression observed in patients exhibiting operational tolerance versus those with chronic antibody-mediated rejection [70].

Computational prediction by molecular docking simulation has shown the possibility of repurposing existing drugs to target three core RCD-related model genes. The results revealed that rapamycin exhibited a generally stronger binding affinity with all three molecules, followed by tacrolimus and MMF. Tacrolimus and MMF inhibit cell proliferation. Specifically, MMF hampers the de novo pathway of purine formation, hinders DNA synthesis, and subsequently reduces lymphocyte proliferation. Tacrolimus inhibits calcineurin by binding to an immunophilin known as FK506 binding protein (FKBP), thereby inhibiting NF-kB mediated transcription of pro-inflammatory genes involved in T-cell activation. On the other hand, rapamycin inhibits the mTOR pathway to prevent rejection, with additional benefits in reducing risks of developing malignancies and viral infections.

MMF, tacrolimus, and rapamycin are emerging as potential therapeutic drugs for treating fibrosis in autoimmune diseases, such as interstitial lung disease in systemic rheumatic diseases, skin fibrosis in systemic sclerosis, and renal fibrosis in lupus nephritis [71, 72]. However, the evidence of their involvement in post-transplant fibrosis and the specific target or mechanism is limited. A retrospective study suggested that a higher average dose of MMF over 1 year was associated with improved renal function and slowed IF/TA progression, partly independent of its immunosuppressive effects [73]. Clinical and in vivo studies have suggested that MMF can inhibit fibroblast activation and proliferation in rheumatic diseases and lung transplantation [71]. Additionally, both rapamycin and MMF can alleviate fibrotic responses by mesangial cells by downregulating mTOR and extracellular signal-regulated kinase phosphorylation [72]. However, the interactions between these immunosuppressants and RCD-related molecules, and their mechanisms, require further validation through in vitro and in vivo experiments.

Our study has limitations. The findings lack laboratory verification due to our inability to obtain a sufficient number of surveillance biopsies within the given timeframe. However, further experiments investigating the potential mechanisms linking RCD and immune traits are currently underway. Moreover, relying solely on the RCDI developed in this study may oversimplify the issue; a more comprehensive nomogram including clinical features would be preferable.

In conclusion, our study provides exploratory but novel insights into the potential role of RCD pathways in the development and progression of post-transplant fibrosis. We also constructed the RCDI using multiple machine-learning approaches, which holds promise in distinguishing normal and IFTA allograft samples, as well as predicting allograft outcomes.

Funding Statement

The study was supported by grants from the Beijing Natural Science Foundation (No. 7234372) and the National Natural Science Foundation of China (Grant No. 82070764).

Author contributions

All authors read and approved the final manuscript.

Disclosure of interest

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

Data availability statement

The study did not generate any original data. Codes used in the study are available upon reasonable request.