2.1. Transcriptomic Data Obtained and Acquisition of the SUMOylation-Related Genes

RNA expression profile data (fragments per kilobase million [FPKM]) and corresponding clinical data for CRC patients and the corresponding clinical information of CRC patients (The Cancer Genome Atlas [TCGA]–COAD cohort) were downloaded from TCGA (https://portal.gdc.cancer.gov/) database. Additionally, the RNA expression profile data and the corresponding clinical information of CRC patients were also downloaded from Gene Expression Omnibus (GEO, https://www.ncbi.nlm.nih.gov/geo/) using the getGEO function in the GEOquery package. The GSE17538 and GSE29623 datasets were used for further validation.

A total of 187 SUMOylation-related genes were identified from the MSigDB (https://www.gsea-msigdb.org/gsea/msigdb/index.jsp). The gene lists are shown in Table S1.

2.2. Screening Differentially Expressed Genes (DEGs) Between CRC Tumor Tissues and Normal Tissues

Limma package was conducted to identify the DEGs between CRC tumor tissues and normal tissues, and the differences were identified with parameters as |log2 fold change (FC)| > 1 and false discovery rate (FDR) < 0.01. The ggplot2 and heatmap packages were used to draw the volcano plots and heatmap.

2.3. Gene Set Enrichment Analysis (GSEA)

GSEA was performed with the SUMOylation-related gene sets from the “MsigDB/msigdb.v2023.1.Hs.symbols.gmt” using the GSVA package. Pathways with an adjusted p < 0.05 were considered significantly enriched.

2.4. Nonnegative Matrix Factorization (NMF) Clustering

Based on the expression of 187 SUMOylation-related genes, the NMF package was used to identify the SUMOylation clusters for CRC patients, the number of the clusters was determined by the k value, and the optimal k value was selected when the cophenetic correlation coefficient started to decrease. Then, principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE) were performed to evaluate the clustering results. Meanwhile, the differences in Kaplan–Meier OS and clinical characteristics for separated clusters were investigated.

2.5. Identification of the DEGs and Biological Functions Between SUMOylation Clusters

The DEGs among the three clusters were identified by comparing each cluster with the rest clusters according to the thresholds of |log2 FC| > 0.585 and FDR > 0.05. The Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) were performed to evaluate the biological functions of each cluster using the clusterProfiler package.

2.6. Investigation of the Immune Characteristics Between SUMOylation Clusters

For immune characteristic analyses, the ESTIMATE package was conducted to evaluate the stromal score, immune score, ESTIMATE score, and tumor purity. Meanwhile, the single-sample GSEA (ssGSEA) method was used to estimate the infiltration of the immune cells. The differences among separated clusters were detected by the Kruskal–Wallis test. p < 0.05 was considered statistically significant. GSEA was performed with the HALLMARK gene sets from the “MsigDB/h.all.v2022.1.Hs.symbols.gmt” using the GSVA package. Pathways with an adjusted p < 0.05 were considered significantly enriched.

2.7. Prediction of the Response to Chemotherapy and Immunotherapy Between SUMOylation Clusters

The differentially expressed HLA Class I and II genes, IFN-γ signaling signatures, and the immune checkpoints among separated clusters were explored by the Kruskal–Wallis test. p < 0.05 was considered statistically significant. The responses to cytotoxic T lymphocyte antigen-4 (CTLA-4) and programmed cell death protein 1 (PD-1) blockers were predicted based on GenePattern (https://www.genepattern.org/). Half-maximal inhibitory concentration (IC50) values of antitumor drugs were calculated using oncoPredict package based on https://www.cancerrxgene.org (GDSC2, https://www.cancerrxgene.org). The differences in drug responses were tested by using the Kruskal–Wallis test. p < 0.05 was considered statistically significant.

2.8. Construction and Validation of the Prognostic SUMOylation Score

The hub SUMOylation-related genes were selected by intersecting the DEGs among separated clusters and 187 SUMOylation-related genes collected from the MSigDB database. The intersected SUMOylation-related genes were incorporated into a univariate Cox regression model to identify the prognostic SUMOylation-related genes with the p < 0.1. These prognostic SUMOylation-related genes were shrunk by the least absolute shrinkage and selection operator (LASSO) Cox regression model using the glmnet package. The SUMOylation risk score was calculated according to the expression and LASSO coefficient of each gene. CRC patients were then divided into high- and low-SUMOylation groups according to the median SUMOylation score, and the differences in Kaplan–Meier OS between high- and low-risk groups were investigated. The receiver operating characteristic (ROC) curves for 1, 3, and 5 years were drawn, and the area under the curve (AUC) was used to test the prediction efficiency of the SUMOylation risk score. Two external datasets (GSE17538 and GSE29623) also were used to validate the prediction efficiency of the SUMOylation risk score.

2.9. Construction of a Nomogram

Univariate and multivariate Cox regression analyses were performed to evaluate whether the SUMOylation risk model was independent of several clinical characteristics by incorporating clinical characteristics. The regplot package was used to develop a nomogram to predict the survival of CRC patients. ROC curves, decision curves, and calibration curves were drawn using survivalROC, d curves, and rms packages to determine the prediction efficiency, reliability, and sensitivity, respectively.

2.10. MR Analysis

Two-sample MR was used to assess the potential association between the SUMOylation gene signature and the risk of colon cancer. The expression of the SUMOylation gene signature was considered as the exposure, CRC was considered as the outcome, and the cis-genome-wide expression quantitative trait loci (cis-eQTLs) were used as the instrument variables (IVs). The cis-eQTLs summary data of SUMOylation gene signature were collected from the eQTLGen Consortium (31,684 cases) [34]. The GWAS summary statistics for CRC were obtained from the IEU OpenGWAS project (https://gwas.mrcieu.ac.uk/, 32,072 cases and 19,948 controls) [35]. All data were from individuals of European ancestry. The IVs selected obeyed the rules: (1) The IVs are strongly connected to the exposure; (2) the IVs are not connected to confounders; and (3) the IVs are connected to the outcome only through the exposure. The cis-eQTLs were selected as IVs with the parameters as follows: the cis-eQTLs associated with the exposures at genome-wide significance (p < 5 × 10⁻⁸), and the cutoff for linkage disequilibrium (LD) was R² < 0.3 at a 10,000-kb window to ensure independence between IVs, minor allele frequency > 0.01, and F-statistic > 10. The causality between SUMOylation and CRC was investigated using the Wald ratio or inverse-variance weighted (IVW) [36], weighted median (WM) estimator [37], MR-Egger regression [38], MR robust adjusted profile score (MR.RAPS) [39], and Mendelian Randomization Pleiotropy RESidual Sum and Outlier (MR-PRESSO) methods [40]. MR analyses were performed by the “TwoSampleMR” and “mr.raps” packages in R (4.2.3), and p value < 0.05 was considered a significant association. Additionally, the MR-Egger intercept test and Cochran’s Q statistic were performed to test the pleiotropy and the heterogeneity, respectively [41].

The Steiger filtering method was conducted to examine the causal direction of IVs, and the leave-one-out method was used to assess the sensitivity of the MR results by sequentially removing an SNP to investigate whether the MR estimate was driven by a single SNP. Due to multiple analyses, Bonferroni correction p value < 0.05/7 (the number of genes) = 0.00714 was considered a significant difference. A p value ranging from 1 0.05 to 0.05/7 was considered a suggestive association.

3.1. Identification of the DEGs and the SUMOylation-Related Pathways in CRC

A total of 3444 DEGs (1885 upregulated and 1559 downregulated) were identified between CRC tumor tissues and normal tissues with |log2 FC| > 1 and FDR < 0.01 (Figures 1(a), 1(b), Table S2). Moreover, we conducted GSEA to explore the SUMOylation pathway enrichment in CRC compared to the normal tissues by ranking the 3444 DEGs. The results indicated that the SUMOylation-related pathways were significantly enriched in CRC, such as SUMOylation of DNA damage and repair proteins, DNA replication proteins, RNA binding proteins, ubiquitination proteins, SUMOylation proteins, chromatin organization proteins, SUMOylation, DNA methylation proteins, intracellular receptors, and transcription cofactors and factors (Figure 1(c), Table S3).

3.2. The Characteristics of the Three SUMOylation-Related Clusters in CRC

The above results revealed that SUMOylation played a vital role in CRC progression. Therefore, the NMF algorithm was performed to explore the relationship between SUMOylation-related gene expression and tumor classification. Based on the 187 SUMOylation-related gene expression, the patients were clustered into three clusters (Cluster 1, n = 57 cases; Cluster 2, n = 121 cases, and Cluster 3, n = 258 cases) according to the k value, which was determined by the cophenetic correlation coefficient (Figures 2(a), 2(b), Table S4). Meanwhile, the PCA and t-SNE analyses indicated that the patients were distributed into three groups, which is consistent with the NMF results (Figures 2(c), 2(d)). Kaplan–Meier survival curve revealed the worst survival status for patients in Cluster 2 (Figure 2(e)). Moreover, we also found a correlation between SUMOylation clustering and microsatellite instability (MSI) in CRC (Figure 2(f)).

3.3. Biological Functions of Each SUMOylation-Related Cluster in CRC

A total of 5511 DEGs (1094 upregulated and 4417 downregulated), 2711 DEGs (2331 upregulated and 380 downregulated), and 1526 DEGs (291 upregulated and 1235 downregulated) were screened out in Cluster 1, Cluster 2, and Cluster 3, respectively (Tables S5–S7). Biological function analyses indicated that three clusters were involved in differential biological processes, such as DEGs of Cluster 1, and were involved in nucleosome assembly, protein localization to chromosome, and negative regulation of myeloid cell differentiation (Figures 3(a), 3(b), 3(c)). KEGG analysis indicated that Cluster 1 is associated with viral carcinogenesis, alcoholism, and neutrophil extracellular trap formation (Figure 3(d)). The DEGs of Cluster 2 were associated with several biological processes, such as the intracellular steroid hormone receptor signaling pathway and intracellular receptor signaling pathway (Figures 3(e), 3(f), 3(g)) and involved in chemical carcinogenesis-receptor activation (Figure 3(h)). The DEGs of Cluster 3 were mainly enriched in nuclear transport, nucleocytoplasmic transport, protein SUMOylation, and nucleobase-containing compound transport (Figures 3(i), 3(j), 3(k)). KEGG analysis revealed that Cluster 3 was associated with platinum drug resistance and nucleocytoplasmic transport (Figure 3(l)).

3.4. Immune Characteristics in Tumor Microenvironment (TME) in SUMOylation-Related Clusters

We further investigated the immune characteristics in TME for each cluster. The ESTIMATE algorithm was used to assess the ESTIMATE score, stromal score, immune score, and tumor purity, which indicated the number of tumor cells, infiltrated immune cells, and stromal cells in TME, respectively. Here, we found the highest ESTIMATE score, immune score, and stromal score but the lowest tumor purity in Cluster 2 (Figures 4(a), 4(b), 4(c), 4(d)). Inversely, there were the lowest ESTIMATE score, immune score, and stromal score but the highest tumor purity in Cluster 3 (Figures 4(a), 4(b), 4(c), 4(d)). Then, we explored the different fractions of the infiltrated immune cells in TME using ssGSEA, and most of infiltrated immune cells were observed in Cluster 2, such as activated dendritic cells (DCs), CD56bright natural killer (NK) cells, CD56dim NK cells, eosinophil, immature DCs, myeloid-derived suppressor cells (MDSCs), macrophage, mast cells, NKT cells, neutrophil, plasmacytoid DCs, activated B cells, activated CD4 T cells, activated CD8 T cells, central memory CD4 T cells, central memory CD8 T cells, effector memory CD4 T cells, effector memory CD8 T cells, gamma delta T cells, immature B cells, memory B cells, regulatory T cells, T follicular helper cells, Type 1 T helper cells, and Type 2 T helper cells. Notably, the fraction of Type 17 T helper cells was the lowest (Figures 4(e), 4(f)). However, the opposite result was observed in Cluster 3. Moreover, different HALLMAKER pathways were enriched in different clusters; for example, DNA repair, MYC targets V1 and V2, oxidative phosphorylation, mTORc1 signaling, E2F targets, and G2M checkpoint were mainly enriched in Cluster 3 (Figure 4(g)). But most other pathways, including several immune-related pathways, were enriched in Cluster 2, such as the P53 pathway, TNF-α signaling vis NF-κB, inflammatory response, IL-6/JAK/STAT3 signaling, IL2/STAT5 signaling, interferon-gamma (INF-γ) response, interferon-alpha (INF-α) response, TGF-β signaling, and Notch signaling (Figure 4(g)).

3.5. Differential Response to Chemotherapy and Immunotherapy for SUMOylation Clusters

We further explored the differential response to therapy in different SUMOylation clusters. The boxplots indicated that the differential expression of the HLA I and II genes, HLA-B, HLA-C, HLA-DMA, HLA-DOA, HLA-DOB, HLA-DPA1, HLA-DPB1, HLA-DQA1, HLA-DQB1, HLA-DRA, HLA-DRB5, HLA-E, HLA-F, and HLA-F, was increased in Cluster 2 than other clusters, but the opposite results were observed in Cluster 1 (Figure 5(a)). We also explored the expression of IFN-γ signaling signatures [42], and the expression of most signatures was increased in Cluster 2 but decreased in Cluster 3, including IFGN, IFNGR1, JAK1, JAK2, and STAT1 (Figure 5(b)). We also analyzed the immune activation signatures and observed the highest expression of APAR, NLR, TCR, and TLR in Cluster 2 (Figure 5(d)). Then, we investigated the expression of the immune checkpoint genes in different SUMOylation clusters, and we found the increased expression of ADORA2A, BTLA, BTN2A1, CD160, CD209, CD226, CD27, CD274, CD276, CD28, CD40, CD40LG, CD47, CD80, CD86, CD96, CTLA 4, HAVCR2, HAL-DRB1, ICOS, IDO1, LAG3, PDCD1, PDCD1LG2, TIGIT, TNFRSF9, TNFSF14, TNFSF18, and TNFSF4 in Cluster 2 compared with other clusters (Figure 5(c)). Furthermore, we predicted the response to anti-CTLA-4 and anti-PD-1 therapies; submap results indicated that Cluster 2 may respond to anti-CTLA-4 and anti-PD-1 immunotherapies (Figures 5(d), 5(e)). Meanwhile, we observed that increased the sensitivity of types of antitumor drugs (AZD1332_1463, AZD8186_1918, BMS-754807_2171, JQ1_2172, NU7441_1038, and WZ4003_1614) in Cluster 2 compared with other clusters (Figure 5(f)).

3.6. Construction and Validation of a SUMOylation Score in CRC

To identify the prognostic SUMOylation-related signatures for CRC, the 77 differentially expressed SUMOylation-related genes were selected by overlapping the 7548 DEGs (Table S8) among SUMOylation-related clusters and 187 SUMOylation-related genes (Figure 6(a)). Univariate Cox analysis was used to identify 11 survival-related SUMOylation-related genes (Figure 6(b)), and then, 9 SUMOylation-related signatures (NDC1, PPARGC1A, CDKN2A, UHRF2, NUP54, PIAS3, H4C4, CHD3, and SUMO2) were selected using LASSO regression analysis (Figures 6(c), 6(d), 6(e)). The SUMOylation risk score was calculated, and then, patients were divided into high- and low-risk score groups (Figures 6(f), 6(g)). The KM survival curve indicated that patients with high-risk scores had worse survival status (Figure 6(h)). ROC curves were drawn to test the prediction efficiency of the risk score, and the AUC values for 1-, 3-, and 5-year survival were 0.669, 0.666, and 0.756, respectively (Figure 6(i)). Meanwhile, external datasets (GSE17538 and GSE17536 datasets) validated the predictive ability of the prognostic SUMOylation-related signatures for CRC; the results were consistent with those of the training dataset (Figures 6(j), 6(k), 6(l), 6(m), 6(n), 6(o), 6(p), 6(q)).

3.7. Development of a Predictive Nomogram

The clinical features (age, gender, clinical stages, and MSI status) and SUMOylation risk score were analyzed by univariate and multivariate Cox regression models, and SUMOylation risk score and clinical stages were identified as the independent risk factors for CRC (Figures 7(a), 7(b)). Therefore, a predictive nomogram was constructed based on the SUMOylation risk score and clinical stages (Figure 7(c)). The AUC values of the ROC curves for 1-, 3-, and 5-year OS were 0.695, 0.692, and 0.761, respectively (Figure 7(d)), reminding us that the nomogram had excellent predictive value for prognosis in CRC. The calibration curve revealed that the nomogram had a high accuracy for prediction (Figure 7(e)). The DCA indicated that the predictive nomogram model was effective in clinical practice (Figure 7(f)).

3.8. Association of SUMOylation Gene Signature With CRC Risk

Finally, we assessed the potential association between SUMOylation signatures and the risk of CRC using a two-sample MR analysis. 131 cis-eQTL SNPs for eight SUMOylation signatures (NDC1, PPARGC1A, CDKN2A, UHRF2, NUP54, PIAS3, CHD3, and SUMO2) were selected as IVs according to three assumptions (Table S9). However, the expression of PIAS3 had not been found in the outcome; therefore, we focused on the investigation of the association between seven SUMOylation signatures (NDC1, PPARGC1A, CDKN2A, UHRF2, NUP54, CHD3, and SUMO2) and risk of CRC. As shown in Figure 8(a), the IVW model uncovered that the high expression of NUP54 had a positive causality on CRC (p = 0.00550, OR = 1.12, 95% CI: (1.03, 1.21)). The consistent results were observed in MR-Egger regression (p = 0.04275, OR = 1.22, 95% CI: (1.02, 1.47)), WM (p = 0.01762, OR = 1.15, 95% CI: (1.02, 1.29)), and MR. RAPS (p = 0.00435, OR = 1.13, 95% CI: (1.04, 1.22)). MR-PRESSO results were also consistent with IVW analysis (p = 0.00535, OR = 1.12, 95% CI: (1.04, 1.20), Table S10). Besides, the IVW model showed that the high expression of PPARGC1A had a negative causality on CRC (p = 0.00010, OR = 0.94, 95% CI: (0.91, 0.97)). The consistent results were also observed in MR-Egger regression (p = 0.01865, OR = 0.93, 95% CI: (0.87, 0.99)) and MR. RAPS (p = 0.00023, OR = 0.94, 95% CI: (0.91, 0.97)). MR-PRESSO results were also consistent with IVW analysis (p = 0.00041, OR = 0.94, 95% CI: (0.91, 0.97), Table S10). No causality between other SUMOylation signatures and CRC was found. No pleiotropy or heterogeneity was observed in these results with the Cochran Q test p > 0.05 and Pleiotropy_p > 0.05 (Figure 8(a)). Furthermore, the leave-one-out analysis of the causal association between NUP54 expression and CRC, as well as PPARGC1A expression and CRC, showed that no significant changes were observed (Figures 8(b), 8(c)).