Transcriptional profiling of DM was acquired from the Gene Expression Omnibus. The GSE20966 dataset (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE20966) was composed of pancreatic tissues from 10 non-diabetic controls and 10 T2DM patients on the GPL1352 platform (14). The GSE95849 dataset (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE95849) comprised peripheral blood samples from six T2DM patients and six healthy participants on the GPL22448 platform (15). The GSE20966 and GSE95849 datasets were merged as the discovery set, and batch effects were removed using the sva package, which was visualized into the principal component analysis (PCA) (16). The GSE26168 dataset (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE26168) contained peripheral blood specimens from eight healthy subjects and nine T2DM patients on the GPL6883 platform for external verification (17).

Pyroptosis genes were gathered from prior research (18–20). RCircos package was adopted for the visualization of the genomic position of pyroptosis genes (21).

Four machine learning approaches composed of random forest (RF), support vector machine (SVM), extreme gradient boosting (XGB), and generalized linear modeling (GLM) were conducted for selecting the characteristic pyroptosis genes. Receiver operating characteristic curves (ROCs) were plotted for computing the area under the curve (AUC) on each established model or key pyroptosis gene.

A nomogram was generated through the integration of key pyroptosis genes utilizing the rms package. Calibration curves were drawn for the visualization and evaluation of the consistency between the actual observations and the nomogram-predicted results. Decision curve analysis (DCA) was conducted for quantifying the net benefit at distinct threshold probabilities.

Gene set enrichment analysis (GSEA) was conducted for identifying the possible functions of the selected genes (22). The reference gene sets were acquired from the Molecular Signature Database (23), with p <0.05 as the threshold.

Through the CIBERSORT approach (24), the normalized transcriptional profiling of T2DM and normal specimens was transformed to the immune components based upon the LM22 reference signature matrix set at 1,000 permutations.

From the Drug Gene Interaction Database (DGIdb; www.dgidb.org) (25), the interactions of drugs with key pyroptosis genes were acquired. Afterward, a drug–target network was built by using the Cytoscape software (26).

The weighted correlation network analysis (WGCNA) package was adopted for building co-expression modules (27). The optimal soft thresholding value was selected through the pickSoftThreshold function. By using the dynamic tree cut method, highly connected genes were merged into one co-expression module. The structure of the co-expression modules was visualized through a heatmap plot via the TOMplot function. The interactions of modules with key pyroptosis genes were then estimated via Pearson’s test, followed by the evaluation of the module membership versus gene significance.

Module genes were imported onto the STRING website (28), and protein–protein interaction pairs were acquired. The key genes were selected by using molecular complex detection (MCODE) (a plugin in Cytoscape).

By using the clusterProfiler method (29), Gene Ontology (GO) enrichment analysis was carried out, which comprised the biological process (BP), cellular component (CC), and molecular function (MF). Afterward, enrichment of the Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways was implemented. Results with p <0.05 were indicative of significant enrichment.

All the analyses were implemented using the R package (version 3.5.3; https://www.r-project.org/). Two groups were compared by adopting the Wilcoxon test. Pearson’s test was employed for correlation analysis. p <0.05 was indicative of statistical significance.

Figure 1 depicts the workflow of this study. This work combined two T2DM datasets, namely, GSE20966 and GSE95849, for expanding the sample size as much as possible ( Figure 2A ). The removal of batch effects was then implemented ( Figure 2B ). Figure 2C illustrates the genomic position of pyroptosis genes. The detailed information is listed in Table 1 . Next, the expression differences in pyroptosis genes were estimated in T2DM and controls. Most pyroptosis genes including CHMP4A, CHMP6, GSDMD, IL1B, IRF2, TP53, CASP9, NLRC4, NOD1, NOD2, and PYCARD exhibited notable upregulation in T2DM relative to normal specimens ( Figures 2D, E ), indicating the activation of pyroptosis in T2DM. Both in T2DM and control tissues, pyroptosis genes closely interacted ( Figure 2F ).

Four machine learning algorithms—RF, SVM, XGB, and GLM—were implemented for establishing pyroptosis-relevant models for T2DM diagnosis. The selected characteristic pyroptosis genes of each model were as follows: RF (CASP3, TP53, TIRAP, CHMP2B, and NLRP6), SVM (PLCG1, CHMP2B, TIRAP, BAK1, and NLRP6), XGB (CASP1, IRF1, CHMP4A, NLRP7, and CHMP2B), and GLM (CASP4, IRF1, CHMP4A, HMGB1, and TP53). Among the four models, the SVM model exhibited the lowest “residual” ( Figures 3A, B ). We also summarized the feature importance distribution of pyroptosis genes in each machine learning algorithm ( Figure 3C ). The ROCs demonstrated excellent diagnostic efficacy of the SVM model with AUC = 1 ( Figure 3D ). The sensitivity of the RF, SVM, XGB, and GLM models was 1, 1, 1, and 0.75, respectively. The specificity of the four models was 0.75, 1, 0.5, and 0.25, respectively. Thus, the SVM was the best model, and the genes selected by the SVM were considered the key pyroptosis genes.

Five key pyroptosis genes were eventually selected for establishing the nomogram, composed of BAK1, CHMP2B, NLRP6, PLCG1, and TIRAP ( Figure 4A ). Calibration curves proved the good consistency between the nomogram-predictive results and the actual observations ( Figure 4B ). For determining the clinical significance of the nomogram in daily clinical practice, we plotted DCA curves. As depicted in Figure 4C , in comparison to all of the patients or none of them, the application of the nomogram to predict the risk of T2DM might be reasonable and have more clinical net benefit in accordance with the predicted possibilities computed by the nomogram and threshold probabilities.

The excellent diagnostic efficacy of the SVM model was also proven in the GSE26168 dataset (AUC = 1; Figure 5A ). In addition, this work investigated the diagnostic performance of each key pyroptosis gene. It was demonstrated that BAK1, CHMP2B, NLRP6, PLCG1, and TIRAP can individually diagnose T2DM with relatively high AUC values ( Figures 5B–F ). This demonstrated the crucial significance of key pyroptosis genes in T2DM.

GSEA unveiled that BAK1 was negatively connected with basal cell carcinoma, taurine and hypotaurine metabolism, epithelial cell signaling in Helicobacter pylori infection, and maturity-onset diabetes of the young ( Figure 6A ). CHMP2B was negatively related to the mTOR signaling pathway, renal cell carcinoma, progesterone-mediated oocyte maturation, and glycosphingolipid biosynthesis of lacto- and neolacto-series ( Figure 6B ). NLRP6 exhibited positive interactions with hematopoietic cell lineage, basal cell carcinoma, galactose metabolism, hedgehog signaling pathway, epithelial cell signaling in H. pylori infection, and oxidative phosphorylation ( Figure 6C ). PLCG1 was negatively linked with peroxisome, fatty acid metabolism, primary bile acid biosynthesis, propanoate metabolism, O-glycan biosynthesis, and PPAR signaling pathway ( Figure 6D ). TIRAP presented positive connections with Vibrio cholerae infection, long-term potentiation, and hedgehog signaling pathway ( Figure 6E ).

We also evaluated the correlations between the key pyroptosis genes and clinical features (age and BMI). Nevertheless, no significant associations between BAK1, CHMP2B, NLRP6, PLCG1, and TIRAP and age and BMI were observed among T2DM patients ( Figures 6F–O ).

Through the implementation of CIBERSORT, the fraction of diverse immune cell types was estimated across T2DM and control tissues ( Figure 7A ; Supplementary Table 1 ). Their difference was also investigated between the two groups. As illustrated in Figure 7B , natural killer (NK) cells resting exhibited lower abundance in T2DM relative to the normal specimens. In contrast, the higher abundance of neutrophils was investigated in T2DM. Among the key pyroptosis genes, NLRP6 was positively connected with mast cells activated and neutrophils ( Figure 7C ). In addition, CHMP2B had a positive interaction with B cells naive. The assessment of the interactions between key pyroptosis genes was also conducted. In Figure 7D , CHMP2B negatively interacted with BAK1 and PLCG1, while the other genes presented positive interactions.

Drugs that potentially targeted key pyroptosis genes were inferred utilizing the DGIdb. As a result, BAK1 was a druggable target of keracyanin; PLCG1 was a druggable target of 9,10-phenanthrenequinone, diclofenac, phosphomethylphosphonic acid adenosyl ester, acetaminophen, cefixime, and aspirin; and TIRAP was a druggable target of ustekinumab ( Figure 7E ).

To select the key pyroptosis gene-relevant co-expression modules, this work adopted the WGCNA ( Figure 8A ). The appropriate soft thresholding value was set to 13 based on the scale independence as well as mean connectivity ( Figure 8B ). By using the dynamic tree cut method, eight co-expression modules were built ( Figures 8C, D ). Among them, the turquoise module exhibited the strongest interaction with the key pyroptosis gene CHMP2B ( Figure 8E ). The genes in the turquoise module were regarded as CHMP2B-relevant genes ( Figure 8F ; Supplementary Table 2 ).

To determine the key CHMP2B-relevant genes, the MCODE method was adopted ( Supplementary Table 3 ). As a result, 10 key genes were acquired as follows: KNTC1, NCAPG, KIAA0101, DLGAP5, GMNN, CEP55, KIF20B, ZWILCH, MCM6, and MAD2L1 ( Figure 9A ). Most of the genes presented differential expression in T2DM relative to the normal specimens ( Figure 9B ). The biological significance of CHMP2B-relevant genes was further probed. It was noted that they were notably connected with proteasome-mediated ubiquitin-dependent protein catabolic process, proteasomal protein catabolic process, etc. ( Figure 9C ; Table 2 ). In addition, RNA transport, cell cycle, T-cell receptor pathway, etc. were remarkably enriched by CHMP2B-relevant genes ( Figure 9D ; Table 3 ).