Single cell-transcriptome files of GSE146771 and GSE98638 were downloaded from the Gene Expression Omnibus (GEO) database (http://www.ncbi.nlm.nih.gov/geo/). GSE14333 is a colon cancer microarray dataset and was downloaded from GEO. The Cancer Genome Atlas (TCGA) RNA-seq transcriptome data and clinical information regarding colon cancer were downloaded via the UCSC Xena browser (https://xenabrowser.net/). The patient information is provided in online supplemental table 1 (online supplemental file 1). R software (V.3.5.3) and Python (V.3.6) were used for all the analyses in the manuscript. All bioinformatic methods are listed in online supplemental table 2 (Online supplemental file 1).

CIBERSORT R script was used to estimate the abundance of immune cell populations in the TCGA colon cancer cohorts by the bulk RNA-seq dataset.13 The cytotoxic T cell (CYT) score was obtained from cytolytic genes (calculated as the geometrical mean of PRF1 and GZMA).14 The observation of the T-cell infiltration score (TIS) was defined as the average of the standardized values for the subsets of T cells.15

Single-sample GSEA (ssGSEA) was applied to evaluate the enrichment scores for each sample. The “GSVA” package was used for ssGSEA analysis.16 The hallmark gene sets were downloaded from The Broad Institute. T_cm cell and T_em cell gene sets were from Bindea et al.17 The exhausted T score gene set was defined by Zheng et al.18 The correlation between the risk score and the enrichment scores was performed with Spearman’s coefficient. GO analysis was performed with the “clusterProfiler” package.19

The single-cell transcriptome analysis for GSE146771 and GSE98638 was performed with “scanpy” in Python.20 The batch effect was removed by the BBKNN method. Nonlinear dimensional reduction was performed with the UMAP method. The gene features in each cluster were found by “scanpy”. The cell–cell interactions (CCIs) were calculated with the “SingleCellSignalR” package in R.21

A random forest algorithm was applied on TCGA colon cancer NGS data to find the most important mutations associated with the microsatellite status in colon cancer. Briefly, the gene mutation dataset and microsatellite status were applied to find the most important gene mutations associated with microsatellite status in colon cancer. The “ranger” package was used to find the best hyperparameter in the regression process and build the model.22

NGS analyses were performed in a centralized clinical testing center (Nanjing Geneseeq Technology) according to protocols reviewed and approved by the Ethics Committee of the First Affiliated Hospital of Zhejiang University. DNA extraction, library preparation, and targeted capture enrichment were performed with previously described methods with minor modifications.23 Briefly, genomic DNA from the white blood cells was extracted using the DNeasy Blood & Tissue Kit (Qiagen) and was used as the normal control to remove germline variations. After deparaffinizing of Formalin-fixed paraffin-embedded (FFPE) samples with xylene, genomic DNA was extracted using the QIAamp DNA FFPE Tissue Kit (Qiagen). After quantification of DNA with Qubit 3.0 using the dsDNA HS Assay Kit (Life Technologies), a Nanodrop 2000 (Thermo Fisher) was used to evaluate DNA quality.

The KAPA Hyper Prep kit (KAPA Biosystems) was used to prepare libraries according to previously described methods.24 Briefly, a Covaris M220 instrument was used to shear 1–2 µg of genomic DNA into ~350 bp fragments. The KAPA Hyper DNA Library Prep kit (Roche Diagnostics) was used for end repair, A-tailing, and adaptor ligation of fragmented DNA, and Agencourt AMPure XP beads (Beckman Coulter) were used to perform size selection. PCR was performed to amplify DNA Libraries followed by purification using Agencourt AMPure XP beads.

Selective enrichment of 425 predefined cancer-related genes (Geneseeq Prime panel) was performed with a customized xGen lockdown probes panel (Integrated DNA Technologies). Blocking reagents were defined by Human cot-1 DNA (Life Technologies) and xGen Universal Blocking Oligos (Integrated DNA Technologies). Dynabeads M-270 (Life Technologies) and the xGen Lockdown Hybridization and Wash kit (Integrated DNA Technologies) were used to perform the capture reaction. PCR amplification was performed on captured libraries with KAPA HiFi HotStart ReadyMix (KAPA Biosystems). The KAPA Library Quantification kit (KAPA Biosystems) was used to quantify the purified library. Bioanalyzer 2100 was used to calculate the fragment size distribution.

Sequencing of the target enriched libraries was performed on the HiSeq4000 platform (Illumina) with 2×150 bp pair-end reads. bcl2fastq (v2.19) was used to demultiplex the sequencing data followed by analysis with Trimmomatic25 to remove low-quality (quality <15) or N bases. Then, alignment of the data to the hg19 reference human genome was performed with the Burrows-Wheeler Aligner (bwa-mem)26 followed by processing using the Picard suite (available at: https://broadinstitute.github.io/picard/) and the Genome Analysis Toolkit (GATK).27 VarScan228 and HaplotypeCaller/UnifiedGenotyper were used to call SNPs and indels in GATK, with the mutant allele frequency cut-off set as 0.5%. dbSNP and the 1000 Genome project were used to remove common variants. Germline mutations were filtered out by comparing with patient’s whole blood controls.

FACTERA29 and ADTEx30 were used to identify gene fusions and copy number variations, respectively. The log₂ ratio cut-off for copy number gain and copy number loss was defined as 2.0 and 0.6, respectively, for tissue samples. TMB was defined as the number of somatic, coding, base substitution, and indel mutations per megabase of genome examined and was calculated as previously described.31 Briefly, all base substitutions, including nonsynonymous and synonymous alterations and indels in the coding region of targeted genes, were considered with the exception of known hotspot mutations in oncogenic driver genes and truncations in tumor suppressors.

A customized analysis algorithm was used to perform MSI testing. Briefly, the MSI sites within the panel (425-gene panel) coverage were defined with a sum of 52 mononucleotide repeats with a minimum of 15 bp repeats. The number of sequencing reads supporting each mononucleotide repeat length was tailed for each MSI site. The distribution of the length species within the tumor sample at the MSI site was compared with that summarized from a pool of normal samples. A significantly altered distribution of the length species in a site was considered unstable. A sample was defined as MSI if more than 40% of the evaluated site display instability.

The DEG analysis was applied with the “Limma” package on TCGA colon cancer transcriptome file.32 An empirical Bayesian method was applied to estimate the fold change between MSI-L and MSI-H or MSS and MSI-H using moderated t-tests. The adjusted p value for multiple testing was calculated using the Benjamini-Hochberg correction. The genes with an absolute log2 fold change greater than 1 and adjusted p value less than 0.05 were identified as DEGs.

The least absolute shrinkage and selection operator (LASSO) was applied to construct the microsatellite-related gene signature by TCGA colon cancer transcriptome data. L1-norm was used to penalize the weight of the features.33 A microsatellite-related gene signature–based risk score (MSRS) formula was established by including individual normalized gene expression values weighted by their LASSO Cox coefficients: ∑iCoefficientmRNAi×ExpressionmRNAi .

WGCNA was performed with the “WGCNA” package.34 A power of β=3 and a scale-free R²=0.95 were selected as soft-threshold parameters to ensure a signed scale-free co-expression gene network. A total of 12 non-gray modules were generated.

Survival analysis was performed with the “survival” package.35 The HR was determined with univariate Cox regression analysis. The tROC analysis was performed for the evaluation of the AUC value in the follow-up period with the “survivalROC” package.36

The tumor immune microenvironment of colon cancer is depicted with the 22 immune cell populations by the TCGA cohort (see online supplemental table S1 for detailed information). A higher abundance of CD8 T cells, activated NK cells, and M1 macrophages was identified in MSI colon cancer than in MSS colon cancer. Then, we estimated the CYT activity, CD8:Treg ratio, IFN-γ expression signature, and T-cell infiltration score (TIS) in MSI and MSS colon cancer. As expected, a higher CYT activity (figure 1B), IFN-γ expression signature (figure 1C), TIS (figure 1D), and CD8:Treg ratio (figure 1E) were found in MSI colon cancer. CTLA4, PD-L1 (CD274), and PD1 (PDCD1) are three important biomarkers for immune checkpoint inhibitor (ICI)–based immunotherapy. The expression levels of these three checkpoint molecules were greater in MSI-H colon than in MSS and MSI-L colon cancer (figure 1F–1H). Hence, we also evaluated the exhausted T-cell score with the ssGSEA algorithm. Despite greater CYT activity in MSI-H colon cancer, a higher exhausted T-cell score was observed in MSI-H colon cancer (figure 1I).

The tumor tissues of colon cancer contain several cell populations such as malignant cells, immune cells (eg, CD8⁺ T cells, CD4⁺ T cells, B cells, and myeloid cells), and fibroblasts. To identify the molecular features in MSI and MSS colon cancer at the single-cell level, we analyzed the single-cell RNA-seq transcriptome data from GSE146771. The UMAP method was used to perform nonlinear dimensional reduction. The cells were clustered by their cell types rather than by the patients (figure 1J and online supplemental figure S1A). The microsatellite status was also shown by a UMAP plot (figure 1K). The Leiden method was applied to identify the clusters. The results revealed that the malignant cells, fibroblasts, CD8⁺ T cells, and CD4⁺ T cells from MSI and MSS colon cancer tissues were identified into different clusters (online supplemental figure S1B), indicating the molecular difference between colon cancer with MSS and MSI. Several important tumor microenvironment (TME) genes were selected to analyze the heterogeneity in colon cancer tissues (online supplemental figure S1C). MS4A1 was enriched in a subpopulation of B cells whereas FOXP3 was expressed in a subgroup of CD4⁺ T cells. MS4A6A, PLAUR, and PECAM1 are known to be featured genes expressed in macrophages. High expression of these genes was found in myeloid cells. EPCAM is overexpressed in epithelial cancers associated with enhanced malignant potential. EPCAM was nearly expressed in all epithelial cells and malignant cells in colon cancer.

To further analyze the difference between MSS and MSI colon cancer at the single-cell level, CD8⁺ T cells, CD4⁺ T cells, B cells and malignant cells were extracted from the whole dataset. The UMAP method revealed the distribution of these cells and the microsatellite status in a two-dimensional plot (figure 2A and B). Several stem cell markers were selected to explore the heterogeneity in MSI and MSS tumors. ALDH1A1 was highly expressed in MSS tumors and lowly expressed in MSI tumors, and ICAM1 was highly expressed in MSI tumors and lowly expressed in MSS tumors (figure 2C). The featured gene expression in CD8⁺ T cells, CD4⁺ T cells, B cells and malignant cells from MSI and MSS colon cancer is shown in figure 2D. The CD8⁺ T cells and CD4⁺ T cells from MSI and MSS colon cancer had a significant difference in the most featured three genes. The CD8⁺ T cells from MSI colon cancer showed high KLRK1, PTPRCAP, and HLA-DRB5 expression (figure 2E); however, their counterparts from MSS colon cancer exhibited high expression of CCL5, NKG7, and GZMA (figure 2F). CD4, PTPRCAP, and PIGN were highly expressed in CD4⁺ T cells from MSI colon cancer (figure 2G), and LTB, CD2, and CD3D were highly expressed in CD4⁺ T cells from MSS colon cancer (figure 2H). The expression of three immune checkpoints (PD-L1, CTLA4, and PD1) are shown in online supplemental figure S2A–C . PD1 expression of single cells in MSI colon cancer was more homogeneous than its expression in MSS colon cancer. Most CD8⁺ T cells in MSI colon cancer had a high level of PD1 expression (indicated by the shape of the violin plot) whereas only a subgroup of CD8⁺ T cells in MSS colon cancer showed a high expression level of PD1. Using a reference with annotated T-cell status (exhausted and non-exhausted), we further analyzed the CD8⁺ T-cell status in MSS and MSS colon cancer. The results revealed a significantly higher percentage of exhausted CD8⁺ T cells in MSI colon cancer (69%) than in MSS colon cancer (36%) (p<0.001) (figure 2I).

The CCIs in MSI and MSS tumors were explored with the single-cell transcriptome data. The directions and the number of ligand–receptor interactions are shown in figure 3A,B. The intracellular interactions of PD1 on CD8⁺ T cells from MSI and MSS tumors were therefore identified by the CCIs (figure 3C,D). A more complicated network with involvement of HLAs corresponding to MHC class II (DP, DQ, and DR) was found in MSI tumors. Greater expression of MHC class II molecules was found in MSI tumors than in MSS tumors, and ubiquitous expression of MHC class I molecules was observed between MSI and MSS tumors (figure 3E and online supplemental figure S3). ssGSEA suggested that MSI tumors harbored expression signatures of “PD-1 signaling”, “T-cell receptor signaling”, “graft-versus-host disease”, and “allograft rejection” (figure 3F). These findings indicated that there was high MHC-II expression consistent with a pro-immune/antitumor response in MSI tumors.

The random forest algorithm was applied to determine the importance of gene mutations associated with microsatellite status in colon cancer. Thirty of the most important gene mutations were identified in the TCGA cohort (figure 4). The results indicated that BMPR2, RNF43, and ANK3 were the most important gene mutations associated with the microsatellite status. MSI-H colon cancer tissues had a higher percentage of BMPR2 RNF43 and ANK3 mutations than MSI-L and MSS colon cancer tissues in the TCGA colon cancer cohort. To further validate the model mentioned previously, the mutation information of 39 TMB-low MSS colon cancer, 10 TMB-high MSS colon cancer, and 15 MSI colon cancer tissues were collected with the 425-panel as our in-house cohort (figure 5A). BRAF, ARID1A, RNF43, and KM2B were involved in the 425-panel. Our data revealed a significantly high frequency of BRAF, ARID1A, RNF43, and KM2B mutations in MSI colon cancer, which confirmed the random forest model. Moreover, the detailed mutation types of BRAF, ARID1A, RNF43, and KM2B are shown in figure 5B.

DEG analysis was performed for MSS and MSI colon cancer in the TCGA cohort. The volcano plot in figure 6A,B shows the DEGs among MSS and MSI-H, MSI-L, and MSI-H colon cancer. The Venn diagram in figure 6C shows that a total of 1202 genes were involved in the microsatellite status–related DEGs (online supplemental file 1). GO analysis revealed that the response to INF-γ, leukocyte chemotaxis, pattern specification process, organic anion transport, and other biological processes were enriched with the microsatellite status–related DEGs (figure 6D). The 1202 microsatellite status–related DEGs were used to build a prognostic model to predict the cancer-specific survival (CSS) of patients with colon cancer. LASSO Cox regression penalized the unimportant features in the regularization process, and 17 features (LOC100272228, HOXC8, TMEM195, TNFAIP2, TRIM58, PLAG1, MTERF, APOD, SYT12, HOXC11, ENO2, HBA1, HOXC4, TNFRSF19, SULT1B1, B3GNT8, and ATOH1) were finally selected as the microsatellite status–related gene signature (figure 6E, online supplemental table S3). The expression levels of the 17 genes in MSS and MSI colon cancer of different TNM stages are shown in (online supplemental figure 4-5). The MSRS for each patient was calculated based on the expression levels of the 17 genes with the coefficient in the model. The risk score formula was established as follows: ∑iCoefficientmRNAi×ExpressionmRNAi . The patients with a higher MSRS had a worse CSS than patients with a lower MSRS (HR 16.5, p<0.001, best cut-off method with cut-off value 2.93) (figure 6F). Time-dependent ROC analysis further confirmed the higher accuracy of the MSRS (mean AUC values higher than 0.8 during the follow-up period) than other clinicopathological traits.

The relationship of treatment outcome and the MSRS indicated that patients with a lower MSRS had a higher fraction of complete remission (CR) (93% vs 70%) and a lower percentage of progressive disease (PD) (3% vs 22%) status (figure 7A). The association of the microsatellite status–related signature-based RS with the immune populations were calculated with Pearson’s coefficient (figure 7B), indicating a significant correlation with CD4⁺ T cells (negative) and memory B cells (positive). Next, we extended the analysis to include the immune checkpoint molecules, including the B7-CD28 family (CD274, CD276, CTLA4, HHLA2, ICOS, ICOSLG, PDCD1, PDCD1LG2, TMIGD2, VTCN1), TNF superfamily (BTLA, CD27, CD40, CD40LG, TNFRSF18, TNFRSF4, TNFRSF9, TNFSF4, TNFSF9) and others (ENTPD1, FGL1 HAVCR2, IDO1, LAG3, NCR3, NT5E, SIGLEC15) (figure 7C). In total, seven immune checkpoint molecules, including ENTPD1, SIGLEC15, TNFRSF4, TNFSF4, TNFSF14, PDCD1LG2, and HAVCR2, were upregulated in the high MSRS group (Cor >0.25). An external cohort (GSE14333) was used to confirm the results. The patients with a low MSRS had a more favorable survival outcome than patients with a high MSRS (online supplemental figure S6A). The association between the MSRS and expression pattern of immune checkpoint molecules in the GSE14333 showed the same tendency as it did in the TCGA cohort (online supplemental figure 6SB). For instance, the association of TNFSF4, HAVCR2, and ENTPD1 with the MSRS are shown in (online supplemental figure 6C–E). The counterparts in the TCGA cohort are shown in (online supplemental figure 6SF–H). As recently the combination of regorafenib plus nivolumab has shown a manageable safety profile and encouraging antitumor activity in colon cancer,37 we further investigated the association of the MSRS and VEGF activities in MSS, MSI-L, and MSI-H tumors. The results indicated a strikingly higher association of the MSRS and VEGF in MSI-H tumors (0.53) than in MSI-L tumors (0.34) and MSS tumors (0.14) (figure 7D). Several studies have discovered that PD-1 blockade therapy reactivates effector T cells and also promotes the proliferation of T_cm cells, improving antitumor immunity.38 39 Thus, the association between T_cm activity/T_em activity/T_scm activity and exhausted T-cell score/MSRS was analyzed. The results indicated more significant association between T_cm activity/T_em activity/T_scm activity and MSRS/exhausted score in MSI tumor compared with the association in MSS tumor (online supplemental figure S7A–F).

ssGSEA analysis was performed to evaluate the biological meaning underlying the MSRS (online supplemental figure S8A). Myogenesis, apical junctions, hedgehog signaling, TGF-β signaling, and epithelial mesenchymal transition were the hallmarks with most significantly correlated with the MSRS. To further confirm the results from the ssGSEA, WGCNA was performed. With a power of β=3 set as the optimal soft threshold to construct a scale-free signed network (online supplemental figure S8B,C), a total of 12 nongray modules were identified (online supplemental figure S8D,E). Among these nongray modules, two modules (turquoise and magenta) with the highest absolute correlation values with the MSRS were selected for further analysis. The genes from the two modules were applied to GO analysis. The results indicated that extracellular matrix reorganization–related enrichments (online supplemental figure S8E) and cell division–related enrichments (online supplemental figure S8F) were related to the genes involved in the turquoise and magenta modules, which confirmed the results from the ssGSEA analysis.

The MSRS serves as a promising marker to predict DSS in different subgroups of patients with colon cancer in the TCGA cohort, including TNM stage I–II (HR 6.37, p<0.001) (online supplemental figure S9A), TNM stage III–IV (HR 9.36, p<0.001) (online supplemental figure S9B), age <66 years (HR 10.19, p<0.001) (online supplemental figure 9C), age >66 years (HR 9.76, p<0.001) (online supplemental figure 9D), MSI (HR 18.39, p<0.001) (online supplemental figure S9E), MSS (HR 6.24, p<0.001) (online supplemental figure S9F), female gender (HR 11.15, p<0.001) (online supplemental figure S9G), and male gender (HR 11.37, p<0.001) (online supplemental figure S9H).

To further use the microsatellite status–related gene signature, we built a neural network with the gene expression pattern of the microsatellite status–related gene signature to stratify MSS and MSI colon cancer. Figure 7E shows a schematic diagram of the neural network. The expression profile of the microsatellite status–related gene signature from the TCGA colon cancer cohort was split into the training dataset (2/3) and testing dataset (1/3). With increasing epochs in the training set, the loss value decreased in the testing set (figure 7F). The confusion matrix shows that all tumors were classified accurately in the testing set (figure 7G).