Bulk and single-cell spatial transcriptomic analysis reveals cuproptosis-related molecular subtypes characterized by distinct tumour microenvironment infiltrates in colorectal cancer

2.1 RNA expression dataset

Three transcriptomics datasets in our analysis, GSE39582, GSE14333 and GSE37892, were obtained from the GEO database. The batch effects were corrected by the Combat function. After batch effects correction, we integrated these GEO databases. SMC and KUL cohorts in GSE144735²³ were included for analysis, which was scRNA data.

We also used COAD and READ cohorts in TCGA to verify our findings. Besides, other data in the TCGA database like somatic mutation were also analyzed. The information was downloaded from the UCSC Xena and TCGA websites. The expression and clinicopathological information of GEO and TCGA datasets are shown in Table S1.

The ST database was obtained from the website²¹ (http://www.cancerdiversity.asia/scCRLM/).

2.2 Cell viability and migration assays

In cell experiments, DLD1 and SW480 cells were cultured. Cell Counting Kit-8 (CK04; DOJINDO, Kumamoto, Japan) was used to evaluate cell viability. In migration assay, Transwell experiments were performed by cultivating 10^4–8 × 10⁴ cells with no foetal bovine serum (FBS) into the upper chamber (CLS3464; Corning Costar, Corning, NY, USA), while 500 μl medium with FBS was added in the lower chamber. After 36–72 hours of culturing, migrated cells were fixed with 4% paraformaldehyde (G1101; Servicebio, Wuhan, Hubei, China), stained with Crystal Violet Staining Solution (C0121; Beyotime, Shanghai, China) and counted under a microscope. The primary antibody for western blotting were as follows: anti-CDKN2A (ab185620; Abcam).

2.3 Non-negative matrix factorization classification

We performed a non-negative matrix factorization (NMF) clustering function²⁹ to stratify the meta-GEO cohort into three molecular clusters using cuproptosis-related genes (CRG). The description of clusters is shown in Table S2.

2.4 Consensus molecular subtypes classification

As previously reported,³⁰ CRC could be divided into four consensus molecular subtypes (CMS) subtypes with discrepant molecular and immune characteristics. To delineate the molecular heterogeneity of NMF clusters of CRC, we performed CMScaller³¹ to classify tumour samples in our cohort. The description of CMS subtypes is shown in Table S2.

2.5 Analyses of TME infiltration

The infiltration and abundance of different cell types in TME were illustrated by CIBERSORT³² and single-sample gene set enrichment (ssGSEA) algorithm.³³

2.6 Data processing and visualization of scRNA and ST expression pattern

The NormalizeData function in the Seurat package was used to normalize the filtered expression matrix. Then, clustering was performed using the RunTSNE or RunUMAP function of the Seurat package.

The clustering of ST data was based on the same workflow of scRNA-seq analysis. The clustering information was visualized using Seurat's SpatialDimPlot function.

2.7 Intersection analysis between cuproptosis-related cluster subtypes and ST transcriptomic data

We identified differentially expressed genes (DEGs) of cuproptosis-related cluster (CPRC) subtypes in bulk data by the FindAllMarkers method. DEGs of different clusters in ST were also calculated using the FindAllMarkers function. Firstly, we assessed the statistical significance of the overlapping expression data in both ST genes and scRNA's marker genes. If the p-value < 0.05, it will be regarded as the accumulation of CPRC subtypes in spatial position and then displayed with red color.

2.8 Development of risk_score system

Firstly, the DEGs of three NMF clusters were determined by intersection. Through gene ontology (GO) analyses, we confirmed 134 genes as CRG-related genes. In order to test the function of those genes, we further clustered our patient cohorts into three gene classifications (Table S2).

At last, a 10-gene model, which had the highest frequency, was screened out from all the CRGs and CRG-related genes (Table S3). Then, 10 genes were used to calculate risk_score by the Lasso Cox regression algorithm, as follows:

CRG_score = (-0.176612549* MIS18A's expression) + (-0.127606751* PPID's expression) + (-0.001009568* WDR89's expression) + (-0.477626624235849* RBFA's expression) + (0.0195666084251348* GLI1's expression) + (-0.0256275163114301* TIAL1's expression) + (-0.164256856902423* MYB's expression) + (0.103464404* PPP1R18's expression) + (0.0168481361941815* MLLT11's expression) + (-0.0479459333550557* FUT4's expression).

The median value of CRG_score was the cutoff value to classify patients into different risk groups. Kaplan-Meier (K-M) survival testing and immune analyses were conducted to prove our findings.

2.9 Analysis of immunotherapy cohorts’ expression

PD-L1 inhibitor Atzolizumab (IMvigor210 cohort)³⁴ and PD1 ICB cohort (Liu et al. study)³⁵ were included and normalized into TPM form for analysis. In our study, we mainly analyzed the data which were obtained before pharmacological interventions.

2.10 Statistical analysis and testing methods

R and Graphpad software were mainly used for statistical analysis in our study. In terms of testing methods, The Kruskal-Wallis H test and Wilcoxon test were carried out for three cohorts’ difference and two cohorts’ difference. Results with p-value < 0.05 were considered as the results with statistical difference.

3.1 Alteration landscape of cuproptosis genes in CRC

A flowchart systematically describing our research was shown in Figure 1A. First, 10 genes were highly correlated with cuproptosis, reported in Tsvetkov et al. study.³⁶ Principal component analysis (PCA) showed that CRG had a discriminating ability to divide tumour and normal samples in CRC in TCGA (Figure 1B). Next, CRGs’ somatic mutation was identified in 45 of 502 samples in COAD/READ cohorts and DLAT showed the highest mutation rate (3%) (Figure 1C). CRGs’ location of copy number variations (CNVs) on chromosomes were exhibited in Figure 1D. CNV analysis demonstrated that 10 CRGs showed prevalent copy number alterations in CRC (Figure 1E). Based on expression of CRGs in paired tumour-normal tissues, we observed that PDHA1, GLS and LIPT1 were upregulated in the tumour, in line with their CNV amplification. Among genes with CNV deletion, apart from FDX1, the expression of most genes was upregulated in tumour tissue (Figure 1F). Our analysis revealed significant difference in the landscape of genetic alterations of CRGs between CRC and normal colon tissues. To explore the function of CRGs in CRC, we selected a gene upregulated in CRC, CDKN2A, to perform several cytological experiments in vitro. We first transfected CDKN2A siRNAs in SW480 and DLD1 cells and validated the success of downregulation (Figure 2A). Here, we demonstrated that downregulation of CDKN2A significantly prevented tumour cells from proliferation (Figure 2B) and attenuated the cell migration ability (Figure 2C). To sum up, the expression pattern of CRGs is of importance in the tumourigenesis of CRC.

3.2 Cuproptosis-related clusters in CRC

To further delineate transcriptomic patterns of CRGs in CRC, 913 samples from three cohorts (GSE14333, GSE37892 and GSE39582) were merged for further analyses. A network in Figure 2A described the connections and prognostic value of CRGs in CRC. We observed that six CRGs (FDX1, LIAS, DLD, DLAT, PDHB and MTF1) had significant prognostic values by univariate Cox regression and survival analyses in CRC (Figure S1). Notably, CDKN2A could predict a worse prognosis by univariate Cox regression (Figure S1) but had no significant survival prediction.

Next, three distinct subtypes were identified in the meta-GEO cohort using the NMF algorithm (Figure 3B,D and Figure S2A–C), as confirmed by PCA algorithm (Figure 3F). These subtypes were considered as CPRC, including 302 patients in CPRC1, 469 patients in CPRC2 and 142 patients in CPRC3. In survival analysis, patients in CPRC1 showed the best clinical outcome, while patients in CPRC3 were more likely to have shorter survival (Figure 3D; p < .0001). Distribution of clinicopathological features indicated that CPRC3 contained the highest proportion of patients in stage IV (8.45%), while CPRC1 had the maximum numbers of patients in stage I (61.54%), supporting their corresponding prognosis patterns.

Finally, we evaluated different biological pathway activities between CPRCs, (Figure 3G and Figure S2D). Here, we demonstrated that pathways related to CMS 4-mesenchymal, tumour growth factor β (TGF-β) and epithelial-mesenchymal transition (EMT) were predominantly enriched in CPRC3. Immune-activated characteristics like CMS1 and MSI-related signatures were upregulated in CPRC1. These results highlighted the significant discrepancy of biological function among three CPRCs.

3.3 TME analysis in CPRCs

We then turned our attention to TME infiltration in three CPRCs. In Figure 4A,B, we observed that CMS4-mesenchymal and cancer stem cell (CSC) subtypes were mainly clustered into CPRC3, suggesting CPRC3 was characterized by stromal invasion. CMS2-canonical subtypes were predominantly grouped into CPRC2, implying CPRC2 mainly displayed features of epithelial differentiation. Interestingly, CMS1-immune and dMMR subtypes were both concentrated within CPRC1 and CPRC3, suggesting that they were enriched by immune cells. To clearly characterize TME heterogeneity of CPRCs, we scored the meta-GEO cohort using signatures of 23 immune cell types (Figure 4C). This analysis revealed that higher abundance of activated CD8⁺ T cells were found in CPRC1 and CPRC3. Activated CD4⁺ T cells were mainly upregulated in CPRC1. However, immune-suppressive cells such as MDSCs and Tregs were mainly accumulated within CPRC3. Therefore, CPRC3 was identified to the immune-excluded phenotype for its stromal invasion and immune-suppressive characteristics. CPRC1 was considered as immune-inflamed, because of its immune activation. CPRC2 was finally referred as immune-deserted phenotype owing to epithelial phenotype.

We next explored the expression of stimulatory and inhibitory immune-related genes among CPRCs (Figure 4D,E). The results showed that CPRC1 exhibited higher expression of stimulatory cytokines like interferon-gamma. For inhibitory genes, LAG3, PDCD1, IDO1 were upregulated in CPRC3 but differed in CPRC1. These findings were also in line with immune-activated phenotype of CPRC1 and immune-suppressed phenotype of CPRC3.

3.4 Cuproptosis phenotype-related classification in CRC

To further understand transcriptomic patterns mediated by cuproptosis, we obtained a total of 134 genes by overlapping DEGs from three CPRCs (Figure 5A). According to GO analysis (Figure 5B), we observed that these genes were related to mitochondrial gene expression. Considering that copper-dependent cell death was dependent on mitochondrial respiration, the 134 genes related to mitochondrial biological function could be recognized as cuproptosis phenotype-related signatures. Then, based on cuproptosis phenotype-related signatures, we stratified meta-GEO cohort into three gene-clusters with distinct clinicopathologic features (Figure 5C). By hierarchical clustering (Figure 5C), cuproptosis phenotype-related signatures were clustered into signature genes A, B, C. Based on GO annotation of three signatures (Figure S3), Genes A were referred to as signatures of gene-cluster A, which were related to metabolic process and mitochondrial gene expression. Genes B were grouped into gene-cluster B and associated with GO function like epithelial cell development. Genes C belonged to gene-cluster C, which exhibited biological function such as regulation of mitochondrion organization and stromal-related biological function like regulation of vasculogenesis.

Combining with the clinical information, patients in cluster A were more likely to be in early-stage CRC (10.34% in stage I, Figure 5C, p-value = .0019) and have a better survival (Figure 5D; relapse-free survival [RFS], p-value < .0001). Interestingly, patients in previous CPRC1 were mainly clustered into gene-cluster A. CMS1 subtype was predominantly concentrated within gene-cluster A, also in line with characteristics of CPRC1. Accordingly, CPRC2 and CMS2 subtype were mainly grouped into gene-cluster B (Figure 5C). Finally, the 5-year survival rate of patients in cluster C is the lowest in three clusters (Figure 5D; p-value < .0001), with the most patients in CPRC3, CMS4 subtypes and advanced stage (9.47% of stage IV in gene-cluster C, p-value = .0251). TME analysis showed that activated CD4/8⁺ T cells were mainly enriched in gene-cluster A, while immune-suppressive cells like MDSCs and Tregs were upregulated in gene-cluster C. Therefore, TME characteristics of gene-clusters were also in line with three CPRCs. Overall, there were three stable transcriptomic patterns mediated by cuproptosis with distinct TME characteristics as follows: immune-inflammed, immune-deserted and immune-excluded.

3.5 Single-cell profiling based on CRG

Considering that single-cell profiling could distinguish immune cells from tumour tissues to improve the understanding of TME characteristics, we performed bioinformatic analyses of single cells subset from tumour samples in a single-cell dataset (SMC and KUL cohorts) of CRC (GSE144735) (Figure 6A,B and Figure S4A,B).

We utilized the DEGs derived from three CPRC subtypes to score single cells (Figure 6C and Figure S4C). The results showed that T/B cells and their cell subtypes were mainly concentrated in CPRC1 (Figure 6D,E and Figure S4D,E), while stromal cells and their cell subtypes (Figure 6D,E and Figure S4D,E) were predominantly enriched in CPRC3. Expectedly, there were few immune cells and their cell subtypes in CPRC2. Overall, these results confirmed the immune-inflamed phenotype of CPRC1, immune-deserted phenotype of CPRC2 and immune-excluded phenotype of CPRC3 at single-cell level.

3.6 Mapping of CPRC subtypes across ST regions

To further validate the phenotype of three CPRCs, we included four CRC tissues (Figure 7A,C,F,H and Figure S5A,C,F,H) with ST to perform several bioinformatic analyses. As previously reported,³⁷ these tissues could be divided into five regions (tumour, fibroblast, lamina propria, smooth muscle and enterocyte) according to cell markers (Figure 7B,G and Figure S5B,G). Interestingly, we found that the clusters expressed fibroblast markers (SFRP2 and COL1A1) in ST data were enriched with CPRC3 DEGs score (cluster 5 of colon1 in Figure 7D, cluster 2/4 of colon3 in Figure 7I, cluster 2 of colon2 in S5D and cluster 6 of colon4 in Figure Figure S5I). Moreover, by intersection analyses as previously described,³⁸ ST region corresponding to fibroblast was significantly enriched with CPRC3-related score instead of other two subtypes. Therefore, ST data could also confirm the immune-excluded phenotype of CPRC3 (cluster 5 of colon1 in Figure 7E, cluster 2/4 of colon3 in Figure 7J, cluster 2 of colon2 in Figure S5E and cluster 6 of colon4 in Figure S5J). However, there were not significant patterns of CPRC1 and CPRC2 in ST data.

3.7 Construction of CRG_score

To construct an appropriate signature to calculate CRG_score, we chose GSE39582 as a training set and performed 1000 iterations as previously reported.³⁹ We finally selected a group of 10 genes with the highest frequencies of 935 to generate a signature to construct CRG_score (see methods; Figure 8A). In Figure 8B, we demonstrated the high C-index of four different cohorts to test the consistency of the CRG_score. Then, we set the median value of CRG_score as cutoff point to separate meta-GEO cohort into groups with different risks. The barplot reflected that patients in high-risk group were mainly clustered into previous clusters with the worse prognosis, such as CPRC2 and CPRC3 (Figure 8C), in line with the results of subsequent quantification analyses (Figure 8D,E). The area under the curve (AUC) values of 1-, 2-, 3- and 5-year survival rates in training set (GSE39582) were 0.691, 0.729, 0.74 and 0.715, respectively (Figure 8F). AUC values of 1-, 2-, 3- and 5-year survival rates in validation cohort (combined dataset of GSE14333 and GSE37892) were 0.528, 0.644, 0.656 and 0.641 (Figure 8H). AUC values of 1-, 2-, 3- and 5-year survival rates in the external cohort (TCGA-COAD/READ) were 0.593, 0.619, 0.63 and 0.623 (Figure S6C). The survival analyses showed that patients in high-risk in each dataset had worse prognosis than those in low-risk (Figure 8G,I and Figure S6D). In all datasets, we found that high-risk groups had an increasing relapse and death rates compared with low-risk groups (Figure 8J,K and Figure S6A,B,E,F). The heatmap of 10-scoring genes’ expression between different groups in training (GSE39582), validation (combined dataset of GSE14333 and GSE37892) and the external cohort (TCGA-COAD/READ) are shown in Figure 8L,M and Figure S6G. All in all, our rigorous analysis reflected the survival prediction value of CRG_score.

3.8 Analysis of microenvironment in different risk groups

The molecular characteristics related to immune infiltration were displayed in Figure 9A–C. High CRG_score group was mainly concentrated within CMS4 and CSC subtypes, both of which were related to stromal invasion phenotype, EMT and TGF-β pathways activation. Next, analysis of immune cell abundance showed that patients with high CRG_score were mainly infiltrated by immune-suppressive cells including MDSCs and Tregs (Figure 9D,E). These findings suggested that CRC patients clustered in high CRG_score group might exhibit an ineffective response to immunotherapy. Therefore, CRG_score ,we constructed, might be used for prediction of CRC patients’ response to immunotherapy.

3.9 Constructing the CRG_score nomogram

Combing the CRG_score with the TNM staging system, a nomogram was established to predict RFS in GEO datasets and predict overall survival (OS) in TGCA-COAD/READ cohort. The AUC of the receiver operating characteristics (ROC) curve for 1-, 3- and 5-years’ survival demonstrated great prediction value in the training cohort (GSE39582), validation cohort (combined datasets of GSE14333 and GSE37892) and an external validation set (TCGA-COAD/READ) (Figure 10A–D). In training cohort, AUC values at 1-, 3- and 5-year were 0.789, 0.795 and 0.760, respectively. In validation set, AUC values at 1-, 3- and 5-year were 0.738, 0.743 and 0.718. In TCGA cohort, AUC values at 1-, 3- and 5-year were 0.748, 0.759 and 0.687. By compared with AUC values of TNM stage systems, we found that, in training set, AUC values of nomogram at 1-, 3-, 5-year were higher than that of disease stages (Figure S7A). In validation set, AUC values of nomogram at 3-, 5-year were higher than that of disease stages (Figure S7B). In TCGA cohort, the AUC values of the nomogram at 1- and 3-year were also higher than that of disease stages (Figure S7C). These findings implied that our nomogram has a better capacity to predict prognosis compared with disease stages. Finally, the calibration plots of the nomogram shown in Figure 10E–G suggested that our nomogram has a good prediction ability.

3.10 Immunotherapy prediction based on CRG_score

PD-L1 inhibitor Atzolizumab (IMvigor210 cohort³⁴) and PD1 ICB cohort (Liu et al. study³⁵) were added to our study for immunotherapy response analysis. Here, we found that, in IMvigor210 cohort, low-risk group exhibited the better survival rates (Figure 11A) and displayed a more effective clinical response to immunotherapy (64.71% vs. 35.29%; Figure 11B). In Liu et al. cohort, although the difference in responding rates between high- and low-CRG_score groups didn't have statistical significance, while group with a higher CRG_score had the tendency of worse survival and exhibited ineffective response to immunotherapy, compared with low-risk group (53.06% vs. 46.94%; Figure 11C,D). Overall, CRG_score showed an advantage in immunotherapy prediction.