Integrated multiplex analysis of cell death regulators in stage II colorectal cancer suggests patients with ‘persister’ cell profiles fail to benefit from adjuvant chemotherapy

Introduction

In 2020, colorectal cancer (CRC) accounted for 10% of global cancer incidence (1.9 million new cases) and 9.4% of cancer deaths (900 000).1 If diagnosed early (stage I) or where there is local spread and no lymph node involvement (stage II), the tumour is surgically removed, and the 5-year relative survival rates are 91% for colon and 90% for rectal cancer.2 High-risk patients with stage II CRC (ie, pT4 stage tumour or high-risk factors such as poor differentiation, inadequately sampled lymph nodes (<10–12) or histological signs of vascular, lymphatic or perineural invasion3) have a 5-year recurrence rate of 40–50%3 and are recommended to receive adjuvant chemotherapy with the anti-metabolite 5-fluorouracil (5-FU) or capecitabine (an oral prodrug of 5-FU).3 4 In high-risk patients with stage II CRC, 5-FU has been shown to decrease risk of death by only 3–5%.5–7 Conversely, many patients with stage II CRC with high recurrence risk are not treated with 5-FU-based chemotherapy, as they are considered ‘low risk’ based on pathological and histopathological staging. Identifying patients with stage II CRC who would benefit or not from adjuvant chemotherapy would greatly impact clinical practice.

Dysfunctional apoptosis is recognised as a key factor in tumour progression and development of chemotherapy resistance.8 At the molecular level, the apoptosis signalling network comprises a complex interplay of multiple activation pathways, inhibition and de-repression interactions, coupled with positive and negative feedback loops.9 10 The extrinsic pathway initiates with a pro-death signal originating from outside the cell.11 12 Intrinsic apoptosis is triggered by mitochondrial outer membrane permeabilisation (MOMP), which is regulated by oligomerisation of BAX and BAK, leading to the release of pro-apoptotic factors from the mitochondria. BAX/BAK oligomerisation is triggered by BH3-only proteins such as PUMA and NOXA (activated by DNA damage) and BID (activated by caspase-8 downstream of death receptors). Anti-apoptotic BCL2 family proteins, such as BCL2, BCL-XL and MCL-1, inhibit BH3-only proteins and BAX/BAK activation, thereby inhibiting MOMP.13 If MOMP is efficiently activated, the release of mitochondrial factors such as cytochrome-c and SMAC occurs, leading to the activation of initiator caspase-9 (activated by formation of the apoptosome containing cytochrome-c and APAF-1) and effector caspase-3 (the activation of which is enhanced by SMAC which neutralises the caspase-inhibitory activity of XIAP). Active caspase-3 executes the terminal cell death decision by cleaving hundreds of cellular substrates. The intrinsic and extrinsic pathways are connected via several proteins including XIAP, caspase-8 and BID, and both pathways converge on the activation of caspase-3.9 MOMP and caspase-3 activation have traditionally been considered to be ‘all-or-nothing’ processes; however, single cell imaging and molecular studies have identified the existence of ‘persister’ cells that may escape either sublethal MOMP or sublethal caspase-3 activation.10 14 15

To date, no single protein or group of proteins in the apoptosis pathway has yet been shown to reliably predict patient responsiveness to chemotherapy with the degree of sensitivity and specificity necessary to inform clinical decision-making.16 17 Our hypothesis was that analysis of multiple apoptosis and cell death pathway proteins at single cell level would account for apoptosis heterogeneity and provide a more sensitive predictor of chemotherapy benefit compared with single proteins. Using two independent cohorts with stage II CRC that were matched for risk factors but differed by treatment status, we posed the following questions: (1) Can disease recurrence in patients with stage II CRC cancer be predicted using an integrated single cell analysis of multiple apoptosis and cell death signalling proteins? (2) Is treatment response associated with the abundance of specific apoptotic signatures within the heterogeneous tumour epithelium? Our results showed that individual apoptosis proteins did not reveal significant or consistent correlations with recurrence risk. However, using K-means clustering of epithelial cells (>600 000) from across the patient cohorts, we identified independent cell clusters with distinct expression levels of proteins in the extrinsic and intrinsic pathways. We identified a cell cluster profile with higher-than-average expression of XIAP and lower in procaspase-3, which was correlated with higher risk of recurrence in chemotherapy-treated patients only. Using a dynamic system modelling approach, we demonstrated that this cluster profile contained cells exhibiting a ‘persister’ cell profile, characterised by a reduced sensitivity to apoptosome-dependent caspase activation. This was validated in an independent stage II CRC treated cohort.

Results

Cohort design and patient characteristics

The apoptosis pathway proteins under investigation in this study are highlighted in figure 1A and the end-to-end workflow for apoptosis protein analysis at single cell level and prognostic analysis is shown in figure 1B. To evaluate the potential of apoptosis signalling proteins as biomarkers of prognosis and therapy response, the ideal dataset would be a randomly designed clinical trial cohort. For this study, we combined two stage II CRC cohorts to mimic random selection and ensure a similar distribution of available risk factors in chemotherapy-treated patients and surgery-only patients. The first cohort (Huntsville; ‘HV’) was comprised of 238 patients, which included 98 patients with stage II CRC who received adjuvant chemotherapy.18 The second cohort (‘MSK1’; n=333) had 19 adjuvant chemotherapy-treated patients and the remaining patients with stage II CRC (n=281) were treated with surgery only. To ensure balance between the two cohorts, we evaluated (1) clinicopathological risk factors and (2) apoptosis protein distribution. Addressing the first point, there were proportionally more T3 patients among surgery-only patients in the ‘MSK1’ cohort; therefore, 196 patients were randomly filtered out to ensure balanced T3 numbers between the adjuvant chemotherapy and surgery-only groups (see online supplemental table 1 for clinical summary statistics before and after filtering and methods section; online supplemental figure 1 for graphical summary of clinical features of each cohort before and after filtering; online supplemental figures 2 and 3 for data flow-down, patient numbers at each step and final merged data set of adjuvant chemotherapy-treated or surgery only patients). Cellular intensity distribution for the apoptosis proteins was similar for both cohorts (online supplemental figure 4). The final combined cohort included 194 patients, 86 surgery-only and 108 adjuvant chemotherapy-treated, with no significant differences in clinical parameters between the two groups (online supplemental table 1). This study was performed in accordance with ethical guidelines for clinical research with the approval of the ethics committee for each institution. At Huntsville Clearview Cancer Center (IRB/CC1 Colon 01), all patients included in the study underwent surgery for histologically proven stage II CRC cancer between January 1997 and December 2008. For Memorial Sloan Kettering Cancer Center (IRB/WA0497-09), patients were selected who were surgically treated with curative intent for stage I and II CRC at Memorial Sloan Kettering Cancer Center between 1990 and 2010 (see the Materials and methods section).

Request permission

Figure 1

(A) Graphical representation of the proteins involved in the regulation of the extrinsic, intrinsic pathways and their roles in caspase activation, leading to apoptosis. Dye-conjugated antibodies for proteins highlighted with a yellow star were validated and underwent multiplexed staining and imaging (Cell DIVE) on tissue microarrays (TMAs) comprising of stage II CRC tumours. (B) End-to-end workflow for this study: (1) tissue microarrays constructed from surgery excised tumours underwent multiplexed imaging with 22 biomarkers using Cell DIVE (including 16 apoptosis pathway proteins highlighted in (A). Then, 194 closely matched patients were down-selected (n=86 underwent surgery alone/no chemotherapy; n=108 underwent surgery and adjuvant chemotherapy) for outcome analysis; (2) cells from the multiplexed images were segmented into epithelial and stromal cells and a spatial coordinate and ID was generated for every cell; signal intensity for each of the apoptosis pathway proteins was quantified for every cell; (3) tumour cells underwent K-means clustering of proteins in the intrinsic and extrinsic pathway combined; (4) cell clusters were aligned to each patient tumour and demonstrated heterogeneous distributions across patients; (5) cell clusters were mapped back to their spatial location in the tumour using cell ID and spatial coordinate. The example shown is a tumour with high percentage of cluster 2 cells (higher in XIAP, SMAC, CIAP1, lower in procaspase-3); (6) univariate and multivariable analysis was conducted for each cluster (the median was used as a cut-off) and associations with recurrence risk was determined in adjuvant chemotherapy-treated and surgery-only patients. The example Kaplan-Meier plot shows that cluster 2 from the combined intrinsic and extrinsic pathway analysis was associated with higher risk of recurrence in treated patients only (blue line). We further quantified apoptosis sensitivity of cluster 2 cells using systems models DR_MOMP and APOPTO-CELL and found that cluster 2 cells had average MOMP sensitivity but low caspase-3 activity. We then repeated steps 3–6 focusing on just the intrinsic pathway proteins and found that adjuvant chemotherapy-treated patients with a high % of cluster 2 tumour cells (XIAP, SMAC and lower in procaspase-3) had a higher risk of recurrence, and this was validated in an independent stage II CRC cohort. Created with BioRender.com.

Heterogeneity in apoptosis pathway protein expression at single cell level within and across patients

Recognising the need to consider the interdependencies and redundancies between apoptotic signalling proteins, we conducted K-means clustering of epithelial cells to determine whether there were repeated expression patterns of proteins in the intrinsic and extrinsic pathways (see analysis workflow in figure 1B) at a single cell level. This provided an unbiased assessment of apoptosis protein distribution and co-expression across all epithelial cancer cells. Given the known interrelationships and convergence points between the extrinsic and intrinsic pathways, we first clustered proteins from both pathways (‘integrated pathway’ approach) in all epithelial cells from adjuvant chemotherapy-treated and surgery-only patients. Proteins with the strongest contribution to cluster separation were down-selected using the Davies-Bouldin’s index23 and were comprised of BAX, BAK, APAF1, procaspase-9, procaspase-3, XIAP, SMAC, FADD, BID, CIAP1, RIP3 and FLIP. Using this approach, six distinct cell clusters were identified using consensus clustering. Marked heterogeneity in cluster distribution was noted across all patients (figure 2A). The lollipop distribution plots for each protein in each cluster are shown in figure 2B: cluster 1 cells had lower than average expression of most proteins; cluster 2 had higher than average XIAP, SMAC, CIAP1 and lower procaspase-3; cluster 3 had higher-than-average expression of all markers; cluster 4 had higher-than-average procaspase-9, low BAX and BAK; cluster 5 had higher procaspase-3, BAK, BAX, APAF1 and lower RIP3 and FLIP; cluster 6 had higher BID, RIP3 and FLIP. Figure 2C shows an example tumour image shown as a virtual hematoxylin and eosin (H&E) (pseudo-coloured DAPI and autofluorescence images) with cluster 2 cells highlighted in orange and example immunofluorescent images for XIAP, SMAC, procaspase-3 and CIAP1 (figure 2D); figure 2E shows a different tumour with a dominance of cluster 4 cells highlighted in yellow, along with example images for XIAP, SMAC, procaspase-3 and procaspase-9 (figure 2F). Online supplemental figure 6 shows UMAP cell clustering of all cells and cluster relationships, with highest separation between cluster 1 and cluster 3 cells.

Request permission

Figure 2

Proteins from the intrinsic and extrinsic pathway were combined for K-means cluster analysis to quantify protein distributions at a cellular level and contribution to tumour recurrence risk. (A) Integrated heatmap of clinical and protein data of all patients with stage II CRC from combined MSK1 and HV cohorts (adjuvant chemotherapy-treated and surgery only); cluster distribution for each patient is shown as a horizontal distribution of all six cluster groups; individual proteins, and clinical data (age, sex, tumour location, T stage, grade and treatment are included in the heat map for context); (B) lollipop plots of protein distributions in each cluster; (C) tumour with high % cluster 2 cells; (D) example staining features of this high cluster 2 tumour: high XIAP; high SMAC; low procaspase-3; high CIAP1. Scale bars 50 µm; (E) Tumour with high % cluster 4 cells; (F) example staining features of this high cluster 4 tumour: lower XIAP, average SMAC, high procaspase-9; average procaspase-3.

Cell clusters associated with increased recurrence risk in chemotherapy-treated patients

In multivariable models adjusted for sex and age (figure 3A), tumours from adjuvant chemotherapy-treated patients with a higher % of cluster 2 cells (higher XIAP, SMAC, cIAP1 and lower procaspase-3) had significantly increased recurrence risk (p=0.003), whereas tumours with higher % of cluster 4 cells (higher procaspase-9, low BAX and BAK) had significantly lower recurrence risk (p=0.003) (overall model was p<0.0001; AIC=172.91 and concordance index 0.77). In agreement with this, Kaplan-Meier analysis showed that chemotherapy-treated patients with high % of cluster 2 cells or low % of cluster 4 cells had a significantly higher recurrence risk compared with all the other cell cluster groups (figure 3B,C). Importantly, in surgery-only patients, neither cluster 2 nor cluster 4 had an impact on recurrence risk (figure 3D–F), indicating that the clinical importance of these cell clusters is related to their impact on response to adjuvant chemotherapy.

Request permission

Figure 3

Cell clusters combining intrinsic and extrinsic pathway proteins and contribution to recurrence risk in stage II CRC patients. (A) Multivariable analysis adjusted for sex and age shows that adjuvant chemotherapy-treated patients with high % cluster 2 (low procaspase-3, high XIAP, SMAC, CIAP1) had a significant recurrence risk (p=0.003). Conversely patients with high % cluster 4 (higher in procaspase-9, APAF1 and lower in BAK and BAX) had a significantly lower recurrence risk. (B, C) Kaplan-Meier analysis of cluster 2 and cluster 4 in the same patients. (D–F) In surgery-only treated patients, no significant associations were found between cluster 2 or 4 (or any other cluster group) and recurrence risk.

Alignment of clusters with models of MOMP sensitivity and caspase-3 activity

To interrogate the apoptosis sensitivity of the epithelial cells in clusters 2 and 4, we applied two established system models of apoptosis initiation and execution, the BCL2 pathway (DR_MOMP) and the caspase activation pathway (APOPTO-CELL). The DR_MOMP model (figure 4A) uses protein levels of BAK, BAX, BCL2, BCL-XL and MCL1 to determine a ‘stress dose’ or ‘eta’ required for MOMP for each cell16 24 25; cells with ≤10% mitochondrial pores are considered to have low sensitivity for MOMP. While the ≤10% pore cut-off has previously been experimentally validated and correlated with higher risk of recurrence in patients with stage III CRC,16 eta ≤ or > population mean has been shown to predict prognosis,17 and was used as the cut-off in this study. The APOPTO-CELL model uses cell protein levels of APAF1, procaspase-3, procaspase-9, SMAC and XIAP as inputs and models the cleavage and activation of caspase-3 downstream of MOMP.10 Using single cell imaging and caspase-3 FRET reporters, we previously demonstrated that using APOPTO-CELL, cells with a substrate cleavage (SC) ≤25% have inhibited caspase cleavage and fail to undergo apoptosis, whereas cells with SC>25% undergo caspase-3 cleavage/activation and apoptosis. APOPTO-CELL has been experimentally validated and shown to be prognostic in stage III CRC and glioblastoma.17 26–28 We recently applied DR_MOMP and APOPTO-CELL models to spatially map cellular apoptosis competency in patients with stage III CRC and show significant inter-tumour and intra-tumour heterogeneity.26 Using the DR_MOMP model, the ratio of cells with high/low MOMP sensitivity in the high-risk cluster 2 group was 3.4 and close to 1 for low-risk cluster 4 (figure 4B); this suggests that cluster 2 cells have sensitivity to MOMP. However, using APOPTO-CELL, the ratio of cells with high/low caspase-3 cleavage was 5.9 in cluster 2 and 15.8 for cluster 4 (figure 4C). This indicates that while high-risk cluster 2 cells do not appear have impaired MOMP sensitivity, downstream caspase-3 cleavage/activation appears to be compromised. This represents a key characteristic of drug-tolerant ‘persister’ cells: in this case, cells that undergo MOMP, but survive because executioner caspase activation is below a threshold required for cell killing.14

Request permission

Figure 4

Modelling mitochondrial outer membrane permeabilisation (MOMP) sensitivity and caspase-3 cleavage and quantifying differences by cluster group—APOPTO-CELL systems model was used to calculate MOMP sensitivity and caspase 3 cleavage/activity for each cell. (A) The protein inputs (highlighted with *) for DR_MOMP (BCL2, BCL-XL, MCL, BAK and BAX) and APOPTO-CELL (APAF1, SMAC, XIAP, procaspase-3, procaspase-9). DR_MOMP eta metric is the maximum % of pores calculated after simulating an approximated mean stress dose. A low eta implies that cells can induce MOMP with a low amount of BH3 proteins and are hence more sensitive to stress and apoptosis. Eta ≤ or > population mean has been shown to discriminate between good and bad prognosis,17 and this cut-off was used for this study. For APOPTO-CELL model, substrate cleavage (SC) >25%, indicates high caspase-3 activity and ≤25%, indicates low caspase-3 activity. Low substrate cleavage has been previously demonstrated to be associated with chemotherapy resistance.18 26 27 (B) % cells with high MOMP sensitivity (eta≤mean) or low sensitivity (eta>mean) and ratio, aligned with each K-means cluster. Cluster 1 (all apoptosis pathway proteins were low with exception of average BAK and BAX) had the highest % of MOMP sensitive cells (83%) and the other clusters ranged from 50–70%, with cluster 2 (high in XIAP, SMAC, low in procaspase-3) having 68% MOMP sensitive cells. (C) % cells with high caspase activity (SC% >25%) and low caspase activity (SC% <25%), aligned with each K-means cluster. Cluster 4 (higher in procaspase 9, APAF1 and lower in BAK and BAX) had the highest % of cells with SC >25% (85.3% of cells) and cluster 1 (all markers were low except BAK and BAX) and cluster 2 had the lowest % cells with SC >25% (73.7% and 75.3%, respectively). (D–F) Example tumour sample dominant in cluster 2 cells and heterogeneous caspase-3 activity as determined by APOPTO-CELL model. (G–I) Tumour sample dominant in cluster 4 cells and uniformly high caspase-3 activity (APOPTO-CELL).

Validation in an independent stage II cohort

We finally evaluated whether the intrinsic pathway clusters profiles were predictive in an independent cohort of patients with stage II CRC. Using a TMA from an adjuvant chemotherapy-treated stage II cohort that underwent multiplexed imaging and analysis using the same apoptosis pathway protein panel as the discovery cohort, we conducted K-means clustering of the intrinsic pathway proteins. This cohort (MSK2, n=91) consisted of more high-risk patients (see online supplemental table 3 and online supplemental figure 8), including more T4 patients (24% vs 13% in the HV/MSK1 treated cohort) and more poorly differentiated tumours (24% vs 7%), hence representing a selection of the ‘real world’ stage II CRC tumours treated with adjuvant chemotherapy. Here, we also found five distinct clusters with similar protein distributions using K-means clustering (figure 5A). A higher % of cluster 2 cells (higher XIAP and SMAC, lower in procaspase-3) was significantly associated with recurrence risk (p=0.023 and p=0.05 in multivariable model) (figure 5B,C). In this validation cohort, the high procaspase-9, low BAX/BAK cluster 4 group identified in the discovery cohort was less distinct, and no correlation with disease outcome was observed (data not shown). Again, using the DR_MOMP and APOPTO-CELL systems models, we found that cluster 2 cells had average MOMP sensitivity and lower than average caspase-3 cleavage (data not shown). In summary, these combined results suggest the existence of a population of micrometastatic ‘persister’ cells post-surgery that are characterised by a reduced sensitivity to apoptosome-dependent caspase activation, and that increase the risk of recurrence when exposed to chemotherapy. We further investigated the correlations between the significant clusters (ie, cluster 2 and 4 in the discovery cohort (HV and MSK1), cluster 2 in the validation cohort (MSK2) and clinical factors using Spearman’s correlation test for continuous variables, and Kruskal-Wallis test for categorical variables (online supplemental figure 9). Some of the clinical variables, including DNA mismatch repair (MMR), lymphovascular invasion, and local/distant metastasis, were only available for MSK1 and MSK2 patients. Although there were no significant differences after adjustment for multiple testing, there was a weak trend for MMR-proficient patients to have high % cluster 2 or high % cluster 4 cells. There was also a trend for female patients and T4 cancers to have higher % cluster 2 cells. We also evaluated associations between the clusters and local or distant metastases. Although non-significant after adjustment, MSK1 patients with high % cluster 4 tended to have more local recurrence than distant metastasis. Cluster 2 was not associated with local or distant metastases in MSK1 or MSK2 patients.

Request permission

Figure 5

Intrinsic pathway protein cluster effects. We further investigated whether the proteins in the intrinsic pathway were driving recurrence risk and clustered proteins specific to this pathway (BAK, BAX, BCL2, BCLXL, MCL1, APAF1, procaspase-9, procaspase-3, XIAP and SMAC). We found a significant association with cluster 2 and recurrence risk in adjuvant chemotherapy treated patients in the HV and MSK1 cohorts (online supplemental figure 7). We repeated the analysis in an independent adjuvant chemotherapy-treated stage II CRC cohort (MSK2) shown here: (A) heatmap of cell clusters from the MSK2 cohort; (B) cluster 2 (lower procaspase-3, higher XIAP and SMAC) was significantly associated with recurrence risk in multivariable model (p=0.055) and in (C) Kaplan-Meier analysis (p=0.023).

Discussion

The benefit of adjuvant chemotherapy for patients with stage II CRC is debatable; therefore, better understanding of factors contributing to response are urgently needed to avoid unnecessary overtreatment and toxic side effects. Three significant clinical trials have been conducted in patients with stage II CRC, including the QUASAR clinical trial, which showed a 3.6% survival benefit within 2 years of surgery27; the National Surgical Adjuvant Breast and Bowel Project (NSABP; studies C-01 to C-04), which compared surgery alone to chemotherapy in stage II and stage III colon cancer and found that chemotherapy decreased risk of recurrence regardless of risk status28; and the phase III MOSAIC trial (Multicenter International Study of Oxaliplatin/5-FU/LV in the Adjuvant Treatment of Colon Cancer) showed a trend for better disease-free survival in patients with stage II colon cancer.29 In a meta-analysis,30 5-year disease-free survival was found to be 81.4% for patients with stage II colon cancer who underwent resection without adjuvant chemotherapy and 79.3% for patients receiving adjuvant chemotherapy. Analysis of retrospective data from patients with stage II colon cancer from the US National Cancer Database showed higher overall survival in low-risk patients who received adjuvant chemotherapy.31 However, other studies have reported benefits of adjuvant chemotherapy in patients with stage II pT4 only.5 6 Recent guidelines now recommend treatment for high-risk patients4; however, 20% of all patients with stage II colon cancer still receive chemotherapy, exposing them to the negative side effects of chemotherapy, without the benefits.3 In the current study, we found that adjuvant chemotherapy-treated patients with stage II CRC (mix of pT3 and pT4) with a ‘persister’ cell cluster profile had a higher risk of recurrence. Notably, this was not observed in patients with who only underwent surgery. Moreover, this observation was validated in an independent chemotherapy-treated higher risk stage II CRC cohort (25% pT4 and 24% poorly differentiated tumours). This suggests that patients with pT3 or pT4 with a ‘persister’ cell profile do not benefit from adjuvant chemotherapy.

A review of past clinical studies correlating single apoptosis pathway proteins with recurrence concluded that a single marker does not adequately represent the complexity of cellular apoptosis heterogeneity, and a combination of markers is needed to represent an ‘apoptotic tumour profile’.9 Benchmarking against these single marker prognostic studies, we found mixed results for the 16 apoptosis pathway proteins under investigation. In univariate analysis, higher expressions of BCL2, procaspase-8, procaspase-3, MCL-1 and RIP3 were each significantly associated with reduced recurrence risk and SMAC was significantly correlated with increased risk treated patients only, but after adjustment for multiple testing they were no longer significant. Previous studies across several cancer types, including CRC have shown that higher SMAC expression is associated with better prognosis, attributed to its antagonising effects on inhibitor of apoptosis proteins (IAP).32–35 However, these studies were relatively small sample size, and none of the findings were validated in independent cohorts. A more recent study showed that low SMAC H-score (and BAK and BAX) was associated with longer progression-free survival in chemotherapy-treated patients with melanoma, and this was further verified with analysis of The Cancer Genome Atlas (TCGA) data.36 Our previous work on modelling XIAP and SMAC kinetics showed that only about 10% of SMAC binds to XIAP before onset of caspase cleavage and SMAC overexpression had minimal effects on cleavage, whereas overexpression of XIAP inhibited cleavage.10 Our findings for cell cluster 2 appear to agree with this model whereby higher SMAC expression may be simply acting as a bystander to higher XIAP, which itself inhibits caspase-3.

To address the limitations of single markers and average expression, we used K-means clustering to quantify the effects of different distributions of multiple apoptosis proteins within the intrinsic and extrinsic pathways at cell level. While most patients had one dominant cluster, marked intra-tumour and inter-tumour heterogeneity among the other clusters was evident. For example, in the high-risk/cluster 2 dominant tumours, cluster 1 (all apoptosis proteins were lower than average) was the second most dominant cluster, and both clusters 1 and 2 accounted for about 25–90% of all cells in high-risk patients. Even though clusters 4, 5 and 6 (which had more apoptosis sensitive profiles) were also present in these tumours, our findings suggest that a dominant ‘persister’ cell profile increases the risk of resistance to adjuvant chemotherapy, leading to eventual recurrence. We further investigated apoptosis sensitivity of these clusters using models of MOMP (DR_MOMP) and caspase-3 cleavage (APOPTO-CELL), which have been extensively validated in earlier works.10 17 24 37 38 We also recently applied both models to build a tumour atlas of apoptosis competency and showed marked inter-tumour and intra-tumour heterogeneity in MOMP sensitivity and caspase activation by cell type.26 Cluster 2 showed close to average MOMP sensitivity (ratio of cells with high/low MOMP sensitivity in the high-risk cluster 2 group was 3.4 vs an overall average of 4.5); however, the ratio of cells with caspase activation/caspase inhibition was lowest in clusters 1 and 2 compared with all other clusters (4.8 and 5.9, respectively, vs overall average of 11.9), suggesting a ‘caspase inhibited’ or ‘persister’ profile whereby cells undergo MOMP efficiently but fail to fully execute apoptosis because of low executioner caspase activation (due to a low procaspase-3/XIAP ratio). ‘Persister cells’ have been reported to undergo MOMP but still able to survive because executioner caspase activation is below the threshold required for cell killing.14 Sublethal caspase activation in these cells may lead to DNA damage through the activation of the endonuclease CAD (caspase-activated DNAse) and the accumulation of additional mutations in these cells. Our study also raises the question as to whether patients with high cluster 2 ‘persister’ cells would benefit from adjuvant therapy with apoptosis sensitisers. The BCL2 antagonist venetoclax is in use clinically, and MCL-1 and BCL-XL antagonists, as well as BAX/BAK activators, are in clinical or preclinical development. However, as indicated above, our data indicate that recurrence in adjuvant chemotherapy-treated patients is not linked to an inability to undergo MOMP. Rather, it suggests that executioner caspase activation may not be sufficiently activated due to high XIAP. Thus, IAP antagonists such as tolinapant may represent an interesting treatment option for these patients. We (DBL/SMcD) are currently exploring the clinical effectiveness of tolinapant in combination with FOLFOX in the advanced disease setting.

There are some limitations to this study in that it was conducted solely on TMA cores (1–2 mm) and not on whole tumour sections. While each patient had at least two tumour cores selected by a specialised gastrointestinal (GI) pathologist, and represented 3000–6000 tumour cells, this is a small fraction of a whole tumour section (which could be more than 1000 times that). However, the findings for cluster 2 were consistent in an independent cohort. Another limitation is that 10 markers in the intrinsic pathway (plus 3–4 additional cell segmentation markers) are required to achieve the level of cellular granularity and sensitivity needed to quantify subpopulations of cells with divergent apoptosis pathway protein expression. Since multiplexed imaging is not yet mainstream in the clinical diagnostic setting, it presents challenges for clinical implementation, at least in the short term. Therefore, we attempted to further reduce the number of markers and limited cell cluster analysis to XIAP, SMAC and procaspase-3 and XIAP and procaspase-3 alone. While a high XIAP/low procaspase-3 cluster group was evident (data not shown), it was not significant, and the best separation and prediction of recurrence risk was only evident using the entire panel of markers. Non-significant trends (after multiple testing correction) were found for low tumour expression of procaspase-3 (<median expression) and increased recurrence risk in chemotherapy-treated patients (in both the discovery and validation cohort). Therefore, a multimarker cell analysis seems to be most sensitive approach for identifying subpopulations of high-risk ‘persister’ cells. Ongoing advancements in multiplexed immunofluorescent imaging technologies (higher throughput, more markers per round, automated analysis) and clinical adoption will be required for multimarker/single cell assays to be made more broadly available. Our secondary analysis of clinical factors showed some interesting trends, including higher % cluster 2 cells in more advanced T4 cancers and lower % in MMR-deficient patients. Future analysis in larger sample sets is needed to support these findings.

In summary, this study represents the most comprehensive, integrated analysis to date of apoptosis protein distribution at single cell level in stage II CRC tumours and identifies a subgroup of patients with stage II CRC with an apoptosis resistant ‘persister’ cell profile who do not benefit from chemotherapy. Future studies are required to determine if patients with this profile benefit from apoptosis sensitisers.

Materials and methods