Tumor tissue specimens were collected from 18 patients with CRC who underwent surgery at Hamad Medical Corporation, Doha, Qatar. Demographic details and clinicopathological characteristics of the study population are listed in table 1; all patients were treatment-naïve. Histological assessments of tumor samples were performed by pathologists to determine clinicopathological parameters including TNM staging. We combined patients with stages I and II (referred to early stages) and stages III and IV (referred to advanced stages). Samples of patients with CRC were collected over a year, and tissue samples were selected based on the tumor stage. We randomly selected five patients per stage, excluding tumor samples with fatty tissues and small size, and tumor samples from patients with polyps or other medical conditions such as other cancer types and peritoneal disease. Two patient samples from stage III were excluded post RNA-Seq as they did not pass quality controls.

Patients gave written informed consent prior to surgery and sample collection. All experiments were performed in accordance with relevant guidelines and regulations. Tumor tissue specimens were processed and stored, as we have previously described.26 27 Using mechanical disaggregation, cells were dissociated from bulk tumor tissues, and single cell suspensions were obtained for subsequent experiments as we have previously described.27 The experimental pipeline is shown in figure 1A.

Following phosphate-buffered saline (PBS) washing, single cell suspension from tumor tissue was washed and resuspended in 100 µL of flow cytometry staining buffer (PBS with 1% fetal calf serum (FCS) and 0.1% sodium azide). Fc receptor (FcR) Blocking Reagent, human (Miltenyi Biotec, Bergisch Gladbach, Germany) was used to block FcR. Cells were stained with cell surface antibodies against CD3-APC-Cy7 (clone SK7, BD Pharmingen, San Jose, USA), CD4-PE (clone RPA-T4, BD Pharmingen), CD8-FITC (clone RPA-T8; BD Pharmingen), CD33-APC (clone WM53, BD Pharmingen) and 7-AAD viability dye (eBioscience, San Diego, USA) to exclude dead cells and gate on live cells.

Cells were washed with flow cytometry staining buffer then re-suspended in Pre-Sort buffer (BD Biosciences). BD FACSAria III SORP cell sorter on BD FACSDiva software (BD Biosciences) was used for sorting pure CD8⁺ (7AAD^–CD3⁺CD4^–CD8⁺CD33^–), CD4⁺ (7AAD^–CD3⁺CD4⁺CD8^–CD33^–) and CD33⁺(7AAD^–CD3^–CD4^–CD8^–CD33⁺) populations. The sorting strategy is shown in figure 1A. We used stringent gating strategy and applicable measures to ensure minimal sorter-induced cell stress. High purities of sorted immune cell populations were always checked and confirmed. FlowJo V.10 software (FlowJo, Ashland, USA) was used for data analyzes.

Total RNA was extracted from sorted CD8⁺ T cells isolated from CRC tumor tissues of 18 patients using RNA/DNA/protein purification Plus Micro Kit (Norgen Biotek Corporation, Ontario, Canada) following the manufacture’s protocol. Purified RNA was then amplified using 5X MessageAmp II aRNA Amplification Kit (Invitrogen, California, USA), according to the manufacturer’s protocol. RNA concentrations were quantified using Qubit RNA High Sensitivity or Broad Range Assay Kits (Invitrogen).

RNA samples from sorted CD8⁺ T cells were used to generate cDNA libraries using Exome TruSeq Stranded mRNA Library Prep Kit (illumina, San Diego, USA) following the manufacturer’s protocol and as previously described.28 cDNA libraries which passed quality controls were subjected to clustering using TruSeq PE Cluster Kit v3-cBot-HS (illumina). Sequencing of clustered samples was performed on an illumina HiSeq 4000 instrument, using HiSeq 3000/4000 SBS Kit (illumina).

CLC Genomics Workbench 12 (Qiagen, Hilden, Germany), with default settings, was used to trim cleaned pair end reads and align them to the hg19 human reference genome.28 In CLC Genomics Workbench 12, the abundance of gene expression was determined by the score of Transcripts Per Million (TPM) mapped reads. Abundance data were subsequently subjected to differential gene expression analyses. Using CLC genomics, datasets for patients with stage I–IV were generated, then each dataset was compared with another dataset in a stage pairwise comparison using differential gene expression analysis (online supplemental table 1). Transcriptomic data from patients with early stages (stages I and II) were also compared with patients with advanced stages (stages III and IV) (online supplemental table 1). Only differentially expressed genes (DEGs) with fold change >2 and p<0.05 cutoffs were subjected for functional annotation analyses using online bioinformatic tools, Database for Annotation, Visualization and Integrated Discovery (DAVID)29 and Integrated Differential Expression and Pathway analysis (iDEP.91).30 Functional annotation for DEGs was based on Gene Ontology Biological Process enrichment analyses. The differential expression of selected genes encoding T cell exhaustion markers and T cell-related transcription factors, cytokines and chemokines in CD8⁺ TILs from patients with advanced versus early stages was examined (online supplemental table 2). For heatmaps, Z-scores were calculated from TPM values, as previously described.31 Principal component analysis (PCA) and volcano plots for gene dataset comparing advanced and early stages were performed by OrignPro 2020 software (OriginLab Corporation, Massachusetts, USA). Clustering analyses, such as hierarchical clustering and k-means clustering, were performed on iDEP.91.30 Protein–protein interaction (PPI) networks and their statistics for some of the significantly up/downregulated genes were determined by an online tool, STRING V.11.0 (http://string-db.org).32 The pipeline for RNA sequencing data processing and analyzes is shown in figure 1B.

The top 100 upregulated genes and 100 downregulated genes in CD8⁺ TILs from patients with CRC with advanced stages compared with those with early stages were selected for analysis in TCGA CRC dataset, accessed via the UCSC Xena platform (http://xena.ucsc.edu/). Genes aligned in TCGA dataset with similar expression trends were used to identify the ‘poor prognosis CD8⁺ T cell gene signature’ (ppCD8sig). The ppCD8sig score was calculated as the ratio of the average expression of the aligned upregulated genes to the average of the aligned downregulated genes (online supplemental table 3). Next, we divided patients into three groups by classifying the ppCD8sig score into high (top 33%), intermediate (middle 33%) and low (bottom 33%) scores based on the average calculated from the ratio of upregulated/downregulated genes in all the patients included in TCGA dataset (online supplemental table 3). Moreover, ppCD8sig score for patient cohort included in this study was also classified into high ppCD8sig score (n=9, top 50%) and low ppCD8sig score (n=9, bottom 50%) (online supplemental table 4).

Percentages of CD3⁺ and CD3⁺CD8⁺ TILs were calculated from flow cytometric data to determine the density of TILs in patients with low and high ppCD8sig scores. Tumors in patients with low and high ppCD8sig scores were classified into ‘hot’ and ‘cold’ tumors based on the mRNA expression of T cell-related marker genes, such as immune checkpoints (ICs) and interferon-γ (IFN-γ), in CD8⁺ TILs.33 34 For heatmap, the fold change in gene expression was represented as Z-score, which was calculated from TPM values, as previously described.31

Blocks of Formalin Fixed-Paraffin-Embedded tissue specimens were cut into 4 µm sections and processed for immunohistochemistry and immunoscoring. Tissue slides were stained with anti-CD3 (clone AB75, DAKO, Glostrup, Denmark) and anti-CD8 (clone C8/144B, DAKO), then visualized by EnVision Flex visualization system (DAKO) and counterstained with hematoxylin (DAKO). Immunostaining for CD3 and CD8 was performed using Autostainer linker (DAKO). Stained slides were scanned on Aperio AT2 scanner (Leica Microsystems, Wetzlar, Germany) with ×20 magnification and 0.5 µm/pixel resolution. QuPath software was used to quantify the density of CD3⁺ and CD8⁺ T cells in CT and IM regions.35 Regions of CT, IM and healthy epithelium were defined manually and validated by a pathologist. Areas of necrosis or artifacts were excluded. Densities of CD3⁺ and CD8⁺ T cells in the CT and IM regions per mm² were quantified by QuPath, then converted into percentiles. For IS determination, the mean of 4 percentiles obtained for CD3⁺ CT, CD3⁺ IM, CD8⁺ CT and CD8⁺ IM was calculated, as previously described,25 and categorized into two groups: low IS (mean percentile 0%–50%) and high IS (51%–100%).

Statistical analyses were performed using GraphPad Prism V.8 software (GraphPad Software, California, USA). Mantel-Cox test was used to determine log-rank p value when comparing DSS and progression-free interval (PFI) among the patient groups with high, intermediate and low ppCD8sig score. Multivariate analyzes for DSS and PFI were performed using Cox proportional hazard model (MedCalculator V.12.7, https://www.medcalc.org/) in comparison to the ppCD8sig (high, intermediate, low), disease stage (IV, III, II, I), residual disease (yes, no), age (<55, 55–64, 65–74, >74 years of age), anatomic locations (seven different locations), and sex (male, female). χ² test was used to determine the association between the different ppCD8sig scores and disease stage, the presence of residual disease, age, gender or different CRC anatomical locations. Unpaired t-test was performed on samples that passed the Shapiro-Wilk normality test, while non-parametric Mann-Whitney U tests were performed on samples that did not pass the Shapiro-Wilk normality test. Unpaired t-test was used to compare the differences in T cell infiltration between patients with low and high ppCD8sig scores. A p>0.05 was considered statistically non-significant. The p values are represented as follows; ***p<0.001, **p<0.01, *p<0.05. Data are presented as mean±standard error of the mean (SEM).

The density, activity and molecular characteristics of T cell subsets infiltrating the tumor vary with multiple parameters, including tumor subtype, location and tumor stage. Thus, we sorted CD8⁺ TILs from 18 patients with CRC (stage I, n=5; stage II, n=5; stage III, n=3; stage IV, n=5) (figure 1A). Using multiple platforms and bioinformatic tools, differential gene expression analysis was performed to compare the transcriptomic profile of CD8⁺ TILs from patients with CRC with advanced stages (III and IV) versus early stages (I and II) (figure 1B). We found a total of 1672 genes which were differentially expressed between advanced versus early stages, with fold change >2 and p<0.05 cutoffs; 1044 were upregulated and 628 were downregulated (figure 2A). Hierarchical clustering and PCA based on the DEGs showed distinct clusters (online supplemental figure 1), reflecting the variability in cells isolated from patients with advanced stages, compared with early stages.

DEGs with fold change >2 and p<0.05 cutoffs, obtained by comparing the transcriptome of CD8⁺ TILs from patients with advanced stages versus early stages (online supplemental table 1), were analyzed using DAVID platform to perform functional enrichment analysis based on Gene ontology (Biological Process). Downregulated genes in CD8⁺ TILs from patients with advanced stages were enriched in pathways related to T cell costimulation, adaptive immune response, cytokine secretion, IFN-γ signaling, cell cycle and DNA damage checkpoint (figure 2B). In contrast, upregulated genes were enriched in pathways related to chemotaxis, chromatin silencing, response to cellular hypoxia, Wnt signaling and negative regulation of lymphocyte proliferation and apoptotic process (figure 2B). Heatmaps for the expression of selected annotated genes are shown in figure 2C.

Next, we examined the expression of selected genes related to immune cell functions (online supplemental table 2), including those encoding ICs, T cell exhaustion markers, transcription factors related to T cell differentiation and function, cytokines and chemokines and chemokine receptors, and made a comparison between advanced and early stages. Genes that were significantly upregulated in advanced versus early stages included interleukin-2 (IL2), TOX3, FOXP3, VEGFA, RORC, IL23R and CCR6, while downregulated genes included IL10, PDCD1, IL17A, IL7, IL22, HLA-DRA, FASLG and TNFRSF9 (figure 2D).

We generated datasets for the transcriptome of CD8⁺ TILs from patients with CRC with stages I–IV, and performed differential expression analysis using stage pairwise comparisons. Upregulated and downregulated genes with p<0.05 (online supplemental table 1 and figure 2) were subjected for functional enrichment analysis. The list of upregulated genes and downregulated genes, for each stage pairwise comparison, were imported separately into DAVID platform for functional annotation based on Gene Ontology (Biological Process). For upregulated genes, we found that chromatin silencing was a common pathway in stage III versus I and IV versus II; positive regulation of gene expression by epigenetics was shared between stages III versus I and IV versus I; positive regulation of cell proliferation between III versus II and IV versus I; and negative regulation of apoptotic process was shared between IV versus I and IV versus II (online supplemental figure 3A). These findings highlight the potential importance of biological mechanisms, such as the transcriptional control by epigenetic and regulation of cell proliferation and apoptotic process, in CD8⁺ TIL biology/function from patients with advanced disease stages. Additionally, negative regulation of apoptotic process was only observed in stage IV, possibly suggesting that CD8⁺ TILs in advanced stages are unable to induce apoptotic signals to eradicate tumor cells.

Based on the functional annotation of downregulated genes, we found that ERK1 and ERK2 cascade was shared between stage IV versus II and IV versus III; T cell costimulation, cell proliferation, cell signaling, cell migration, inflammatory response and IFN-γ signaling pathway were all shared between stage IV versus I and IV versus II; immune response and chemokine signaling pathway were shared between stage IV versus I and IV versus III (online supplemental figure 3B). These latter findings may suggest the functional impairment of CD8⁺ TILs and their limited capacity of cell proliferation/migration in patients with advanced disease stages.

We sought to investigate the genes, which their expression levels steadily increase or decrease in CD8⁺ TILs as disease progressed from stages I–IV. We found that the expression of 71 genes steadily increased and 104 genes steadily decreased with disease progression (p<0.05, figure 3A,B). Using functional enrichment analysis through DAVID platform, we found that upregulated genes, such as HIST1H2AC, HIST1H2AB and HIST1H2AD, were significantly enriched within the chromatin silencing pathway (figure 3C,D). Downregulated genes, such as CCNE1, CCNB2, CCNG2, E2F1, PLK1, AURKA, KIF2C, CDC25C, PRR11 and FAM83D, were significantly enriched within cell cycle and proliferation pathways, and genes such as CXCL1, PF4, CLNK, IL22 and OAS2 were significantly enriched within immune response-related pathways, including IFN-γ signaling, T cell receptor (TCR) signaling and T cell costimulation (figure 3C,D).

Out of the 100 upregulated genes, 97 genes were aligned in TCGA CRC RNA-Seq dataset, of which 35 genes (36%) had higher expression in patients with poorer DSS. For the 100 downregulated genes, 94 genes were aligned in TCGA dataset, of which 59 genes (63%) had lower expression in patients with poorer DSS. We used the 35 upregulated and 59 downregulated genes to calculate the ‘poor prognosis CD8⁺ T cell gene signature (ppCD8sig) score’, and the CRC TCGA cases were labeled as high, intermediate and low groups according to the ppCD8sig score (online supplemental table 3), as described in the Materials and Methods section. Using Kaplan-Meier analysis, we found that patients with high ppCD8sig score had poorer DSS (HR 1.83, 95% CI 1.40 to 2.38, log-rank p<0.0001, figure 4A) and shorter PFI (HR 1.42, 95% CI 1.04 to 1.93, log-rank p=0.0056, figure 4B), compared with patients with intermediate or low ppCD8sig scores.

Using multivariate analysis and Cox proportional hazard model, we found that the ppCD8sig was an independent prognostic indicator for DSS (p<0.0001, figure 4C) and PFI (p=0.026, figure 4D), even in the presence of disease stage as another indicator. Patients at advanced stages (III and IV) were more likely to have high ppCD8sig score (χ² p<0.0001, figure 4E). Patients with high ppCD8sig score were more likely to have residual disease after primary therapy (χ² p=0.0463, figure 4F). The splenic flexure, descending and sigmoid colon anatomical locations (left-sided colon cancer) were more likely to have a high ppCD8sig score, whereas the transverse colon and the right-sided colon cancers (hepatic flexure, ascending colon and the cecum) were more likely to have a low ppCD8sig score (χ² p<0.0001, online supplemental figure 4A). This latter finding is consistent with results reported by Liang et al demonstrating that left-sided colon cancer in patients with advanced stages exhibit poorer survival than those with right-sided colon cancers.36 However, we found that the ppCD8sig scores did not differ between males and female (online supplemental figure 4B) or age groups (online supplemental figure 4C). Altogether, TCGA analysis supports that the gene signature identified from our gene expression profiling can predict poorer survival and more aggressive clinicopathological features, and could be reflective of the level or the activity of CD8⁺ T cells in human CRC.

A dataset comparing the transcriptome of CD8⁺ TILs in nine patients with high ppCD8sig score versus nine patients with low ppCD8sig score was subjected to hierarchical clustering and PCA analyzes using iDEP.91 web-based tool (figure 5A,B, online supplemental table 4). The differential expression analysis showed 1239 upregulated genes and 862 downregulated genes in patients with high ppCD8sig score versus low score (figure 5A). PCA confirmed the close relativeness of the biological replicates; PC1 shows 35% variance, while PC2 shows 14% (figure 5B). Using iDEP.91 web-based tool, k-means clustering was performed on the dataset comparing the transcriptome of CD8⁺ TILs in patients with high ppCD8sig score versus those with low ppCD8sig score. DEGs were separated into two clusters (A and B) and annotated using Gene Ontology (Biological Process) enrichment analysis (figure 5C, online supplementary table 5). Upregulated genes in cluster A were enriched in pathways associated with translational initiation and protein targeting, while downregulated genes in cluster B were enriched in pathways involved in cell cycle, chromosome segregation, mitotic cell cycle and cell division (online supplemental table 5, adjusted p<0.0001). Some of the upregulated genes in high ppCD8sig score associated with nucleosome organization and chromatin silencing, such as HIST2H2BE, HIST2H2BD, HIST1H2BB and HIST1H2BD, were imported into STRING web-based tool to perform PPI and enrichment network analysis (online supplemental figure 5A). STRING database identified 5 nodes and 10 edges with PPI enrichment p=2.6E-09, average clustering coefficient of 1 and average node degree of 4 (online supplemental figure 5A). PPI and enrichment network analysis was also performed for the downregulated genes in high ppCD8sig score, which the majority of them were related to cell proliferation/cell division/mitosis/chromosome segregation, such as AURKA, KIF4A, CCNA2, CCNB2 and CDC25C (online supplemental figure 5B). STRING database identified 63 nodes and 1259 edges with PPI enrichment p<1.0E-16, average clustering coefficient of 0.861 and average node degree of 40 (online supplemental figure 5B). The ability to find the same pathways, as described in earlier results, confirms that the ppCD8sig score based on the 94 selected genes retain the difference between the transcriptomes of CD8⁺ T cells in patients with CRC with advanced versus early stages.

We analyzed our flow cytometric data and determined the percentages of tumor-infiltrating CD3⁺ and CD3⁺CD8⁺ T lymphocytes to determine the level of T cell infiltration, which might reflect the degree of immunogenicity present in the tumor tissue of patients with low and high ppCD8sig scores (figure 6A). We found that patients with high ppCD8sig score had significantly less CD3⁺ T cell infiltration, compared with those with low score (figure 6B). Additionally, a trend toward reduced levels of CD3⁺CD8⁺ T cell infiltrate was seen in patients with high ppCD8sig score, compared with those with low score; it did not reach statistical significance possibly due to small sample numbers (figure 6B). We also determined the expression of T cell-related marker genes, with particular interest on ICs and IFNG, in the two groups of patients with high and low ppCD8sig scores. Reports showed that tumors can be classified into ‘hot’ or ‘cold’ based on the degree of T cell infiltration and expression of ICs on T cell subsets.33 34 We found that patients with high ppCD8sig score exhibited lower mRNA levels of CTLA4, PDCD1 (PD-1), HAVCR2 (TIM-3), TNFRSF9 (4-1BB), in addition to IFNG (IFN-γ), compared with patients with low ppCD8sig score (figure 6C).

Using immunohistochemistry, we determined the category of IS for patients with low and high ppCD8sig score (figure 6D), as described in the Materials and Methods. We found that six out of nine patients with low ppCD8sig score showed high IS (figure 6E, shown in green), while only three out of nine patients with high ppCD8sig score showed high IS (figure 6F, shown in green). For flow cytometric analyses, we calculated the median of the percentage of CD3⁺ TILs and CD3⁺CD8⁺ TILs in all 18 patients with low and high ppCD8sig scores. For both patient groups, percentages above the median are indicated in green, and below the median are indicated in red (figure 6E,F). We found that six out of nine patients with low ppCD8sig score showed a high degree of CD3⁺ TILs, while six out of nine patients had a high degree of CD3⁺CD8⁺ TILs (% above the median as indicated in green, figure 6E). All patients with low ppCD8sig score exhibited ‘hot’ tumor based on the mRNA expression pattern of ICs and IFN-γ (figure 6E). On the other hand, seven out of nine patients with high ppCD8sig score showed a low degree of CD3⁺ TILs, while six out of nine patients had a low degree of CD3⁺CD8⁺ TILs (% below the median as indicated in red, figure 6F). Eight out of nine patients with high ppCD8sig score exhibited ‘cold’ tumors (figure 6F). It is worth noting that density of TILs determined by flow cytometry did not fully match the IS results (figure 6E,F), as immunoscoring was performed on tissue specimens containing a large amount of both tumor and normal tissue, with extensive representation of IM, taking into account the density of intratumoral and peritumoral CD3⁺ and CD8⁺ T cells. However, flow cytometric analyses were performed on cells isolated from tumor tissues only, and therefore quantify the density of intratumoral T cells.

Collectively, these data suggest that patients with high ppCD8sig score had low levels of CD3⁺ and CD3⁺CD8⁺ TILs, low IS and tumors, which could be classified as ‘cold’ tumors. In contrast, low ppCD8sig score had higher levels of CD3⁺ and CD3⁺CD8⁺ TILs, high IS and tumors, which could be classified as ‘hot’ tumors. These data further confirm the prognostic value of the identified gene signature and potentially highlight its clinical relevance.