The single-cell RNA sequencing (scRNA-seq) data used in this study were obtained from the GSE214695 dataset, which includes samples from 6 patients with UC, 6 patients with CD, and 6 healthy controls (HC) (10). Based on previous studies (3, 4), we selected 10 pyroptosis-related genes associated with the canonical inflammasome pathway (NLRP1, NLRP3, NLRC4, AIM2, Pyrin, PYCARD, CASP1, IL1B, IL18, and GSDMD). After quality control, cells with fewer than 200 or more than 2,500 detected genes or with a mitochondrial gene proportion exceeding 5% were excluded, resulting in high-quality cells for downstream analysis. The top 2,000 highly variable genes were identified using the “vst” method to provide a stable basis for subsequent dimensionality reduction and clustering (11). To mitigate potential technical variation across samples, batch effects were corrected using the Harmony algorithm implemented in Seurat, using principal components derived from PCA with default parameters (theta = 2, lambda = 1) (12). Cells were clustered using Seurat’s graph-based clustering approach with a resolution parameter set to 0.5. Based on the batch-corrected data, unsupervised clustering was performed using the Seurat package with a resolution of 0.5, and UMAP was applied for dimensionality reduction and visualization, revealing diverse intestinal mucosal cell subpopulations (13). Cell annotation was conducted by integrating the original nanostring_reference information with known cell marker genes, ultimately identifying 54 distinct cell subpopulations based on the nanostring_reference for downstream analysis, encompassing various cell types (10). For bulk RNA-seq data, external validation was performed using peripheral blood bulk transcriptome data (GSE3365) (14).

pyroptosis-related gene expression signatures was quantified using AUCell (15), AddModuleScore (16), and GSVA (17) based on a curated pyroptosis gene set. AUCell calculated the AUC of ranked gene expression per cell; AddModuleScore generated normalized module scores; GSVA estimated pathway-level enrichment for each cell or cluster. Cells were divided into high- and low-pyroptosis groups according to score distribution. Differential expression was analyzed using Seurat’s FindMarkers (Wilcoxon test). Gene–pyroptosis associations were assessed using Spearman correlation between gene expression and pyroptosis scores across cells or clusters. Subsequently, Gene Ontology (GO) enrichment analysis was performed using the clusterProfiler R package to explore the potential biological mechanisms associated with these differentially expressed genes.

To evaluate the diagnostic potential of candidate genes identified from single-cell analysis, we applied an integrative machine learning–based feature selection framework to an independent peripheral blood transcriptomic dataset. Gene expression normalization and feature selection were conducted on the complete peripheral blood dataset in the context of gene prioritization.

To identify the most robust diagnostic markers, we integrated five machine learning algorithms, including LASSO (18), GBM (19), SVM-RFE (20), Boruta (21), and random forest (22). To ensure reproducibility, all machine learning analyses were performed using a fixed random seed (set.seed(123)). LASSO regression was implemented using the glmnet package, in which an L1 penalty was applied to shrink coefficients and eliminate redundant variables, with 10-fold external validation used to determine the optimal regularization parameter (λ). Gradient boosting machine (GBM) models were constructed using default parameters (n.trees = 100, interaction.depth = 1, shrinkage = 0.1), and random forest models were built with default settings (including ntree = 500), to rank feature importance based on ensemble learning and thereby improve the stability of feature selection. SVM-RFE was used to iteratively remove the least informative genes and retain the most discriminative features, and was implemented using a linear kernel with default settings. In parallel, the Boruta algorithm was employed to identify all relevant variables by comparing their importance with that of randomly permuted shadow features, and was applied using default parameters (maxRuns = 100, pValue = 0.01), thereby ensuring comprehensive feature extraction. Finally, the intersecting genes identified by all five algorithms were defined as the optimal diagnostic gene set for further analysis.

Rather than constructing a standalone predictive classifier, these machine learning methods were applied primarily for robust feature selection and ranking. Genes consistently identified by all five algorithms were defined as the final diagnostic gene set. ROC analyses were performed using the same peripheral blood transcriptomic dataset employed for feature selection, without a separate held-out testing set, with the primary objective of gene prioritization.

All statistical analyses were conducted in R (version 4.4.0). Group differences were assessed using Student’s t-test or Wilcoxon rank-sum test as appropriate, and multiple-group comparisons used one-way ANOVA or Kruskal–Wallis tests. Correlations between pyroptosis scores and gene expression were evaluated using Spearman’s correlation. Differentially expressed genes were identified with Benjamini–Hochberg FDR–adjusted P values < 0.05. A two-sided P < 0.05 was considered statistically significant.

Prior to downstream analyses, we performed stringent quality control and selected highly variable genes, resulting in a high-quality dataset (Figures 1A, B). Batch effects across samples were then corrected, leading to a more stable distribution (Figure 1C). Following the standard Seurat workflow, we conducted clustering and identified 18 distinct cell subpopulations (Figure 1D). Given that the GSE214695 dataset provides high-quality cell-type annotations that are highly consistent with our clustering results, all subsequent analyses were based on the original annotations (Figure 1E). Comparing cell compositions across different pathological states revealed substantial differences among healthy controls, ulcerative colitis, and Crohn’s disease (Figure 1F). Notably, several immune cell subsets were markedly increased in inflammatory conditions, suggesting that these cells may play key roles in the pathogenesis of IBD.

We evaluated the expression of pyroptosis-related gene sets at the single-cell level using the AUCell and AddModuleScore algorithms. pyroptosis-related gene expression signatures exhibited substantial heterogeneity across different cell types; specifically, these genes were more highly enriched in macrophages, intestinal epithelial cells, and neutrophils (Figure 2A). Both the Crohn’s disease and ulcerative colitis groups showed higher overall pyroptosis scores than healthy controls (HC) across multiple immune and epithelial cell populations (Figures 3B, C). Using AUCell, AddModuleScore, and GSVA, we assessed pyroptosis-related gene expression signatures across the three groups within each cell type (Supplementary Material 1). Across all three scoring methods, six cell populations consistently exhibited higher pyroptosis-related gene expression in both the UC and CD groups: colonocytes, inflammatory monocytes, N1, N2, N3, and M1. By stratifying cells into high- and low-pyroptosis groups according to average expression levels, we further demonstrated that these genes were predominantly expressed within the neutrophil clusters (Figure 2E).

To identify genes associated with pyroptosis-related gene expression signatures, we categorized cells into high-pyroptosis and low-pyroptosis groups based on the AUCell and AddModuleScore pyroptosis scores, respectively. Differential expression analysis between the two groups was performed using Seurat’s FindMarkers function (Wilcoxon rank-sum test). Using the AUCell-derived scores, we identified 118 upregulated genes, whereas the AddModuleScore-based scoring yielded 14 upregulated and 140 downregulated genes (FDR < 0.05; FindMarkers criteria: |log2FC| > 1, min.pct > 0.1). To further determine genes most closely associated with pyroptosis, we conducted correlation analysis and identified 96 genes significantly correlated with pyroptosis-related gene expression signatures (r > 0.3, FDR < 0.05) (Figure 2H). To avoid over-filtering and potentially missing important candidates, all DEGs identified by either scoring method (FDR < 0.05, |log2FC| > 0) were retained for constructing the intersections shown in Figure 2I. The intersection between the correlated genes and the differentially expressed genes resulted in 19 pyroptosis-associated genes (Figure 2I).

Given that IBD is characterized by both localized intestinal mucosal inflammation and systemic immune activation, we performed external validation by evaluating candidate genes identified from intestinal mucosal scRNA-seq analysis in an independent peripheral blood bulk transcriptomic dataset (GSE3365). Although the two datasets originate from different biological sources, the consistency of key gene expression patterns at both local (mucosa) and systemic (peripheral blood) levels would further support their stable and broad biological relevance in IBD. To verify the reliability of the 19 selected genes, we further evaluated their expression profiles in the bulk transcriptome dataset. Cross-comparison between scRNA-seq–identified genes and bulk data revealed 16 overlapping genes. All 16 genes were significantly upregulated in the IBD group (Figure 3A). GO and KEGG enrichment analyses indicated that these genes were primarily involved in the regulation of intrinsic and overall apoptotic signaling pathways and were closely associated with reactive oxygen species–mediated chemical carcinogenesis (Figure 3B). A heatmap further supported these findings, showing a consistent high-expression pattern of the 16 genes in the IBD samples (Figure 3C). Among these upregulated genes, SOD2, PPIF, HIF1A, NAMPT, ICAM1, and PLEK exhibited higher DC scores and formed a tightly connected interaction network with the other genes (Figure 3D).

We applied five machine learning algorithms to screen for signature genes, including the LASSO algorithm (Figure 4A), GBM algorithm (Figure 4B), SVM-RFE algorithm (Figure 4C), Boruta algorithm (Figure 4D), and random forest model (Figure 4E). Intersection analysis of the marker genes identified by all five algorithms yielded six characteristic genes: BASP1, LITAF, NAMPT, PHACTR1, PLAUR, and PPIF (Figure 4F). We further evaluated the diagnostic performance of these six genes in peripheral blood. ROC curve analysis (Figure 4G) showed that all six genes exhibited good diagnostic value, with BASP1 displaying the strongest diagnostic ability (AUC = 0.935), followed by PLAUR (AUC = 0.840), NAMPT (AUC = 0.825), PPIF (AUC = 0.742), LITAF (AUC = 0.733), and PHACTR1 (AUC = 0.666).