Colorectal cancer (CRC) is the third most common cancer with over 1.5 million new cases diagnosed every year and accounts for 10% of cancer-related death worldwide1, with a 5-year overall survival (OS) rate of 64%2. CRC is characterized by etiological, genetic and clinical heterogeneity, thus rendering prognosis estimation and therapeutic management difficult.

The most widely used staging system for prognosis evaluation and treatment guidance is the AJCC/IUCC/TNM (American Joint Committee on Cancer/Union for International Cancer Control/Tumor Node Metastasis) classification, currently in its 8th edition3. It predicts the prognosis of CRC patients, with lower-stage cancers having a better prognosis than higher-stage ones. However, TNM staging is not sufficiently robust to predict the survival of stage II–III patients, who have reported 5-year survival rates of 73% an 53% respectively3. This remains a challenge despite incorporating additional prognostic factors (vascular/lymphatic invasion—perineural invasion (VELIPI), grade, tumor budding, perforation, KRAS status, MisMatch Repair (MMR) status and BRAF status)3. This is especially illustrated by the “stage paradox”, i.e., prognosis is better for stage IIIA patients than for stage IIB/IIC ones4.

Surgery is the cornerstone of curative intent treatment. Adjuvant therapy with fluoropyrimidine + oxaliplatin-based chemotherapy is recommended and improves survival in stage III and high-risk stage II CRC studies5, but the lack of sufficiently robust biomarkers may result in the under- or over-treatment of some patients. Therefore, there is an urgent need to identify new robust biomarkers that are easily applicable in routine clinical practice to better stratify individual risk.

In 2015, an international consortium established the consensus molecular classification which distinguishes four consensus molecular subtypes (CMS): CMS1 Immune, hypermutated, microsatellite instability (MSI), CpG island methylator phenotype (CIMP), BRAF mutation and immune activation; CMS2 Canonical, epithelial, chromosomal instability (CIN), marked WNT and MYC signaling activation; CMS3 Metabolic, epithelial, KRAS mutation and metabolic dysregulation; and CMS4 Mesenchymal, epithelial-mesenchymal transition, transforming growth factor-β activation, stromal invasion and angiogenesis6. Its stage-independent prognostic value was demonstrated with a worse outcome for both relapse-free survival and OS in CMS4 tumors6. However, the requirement for sufficient tumor material and the cost and technology required for genome-wide expression analysis restrict its use. Trinh et al. developed an immunohistochemistry-based method to stratify patients into three CMS groups: 1 immune, 2/3 epithelial and 4 mesenchymal, with a prognostic value in stage II CRC7, but this remains to be validated in larger cohorts, as the distribution of patients is heterogeneous between studies7,8. Later, Dalerba et al. proposed the use of transcription factor CDX2 expression as sole prognostic factor, with its loss of expression associated with a worse prognosis9. To date, there is no recommendation in international guidelines to use these classifications for adjuvant treatment10.

In 2010, our team developed a prognostic transcriptomic signature in sarcomas that predicts metastatic outcome and outperforms the gold-standard Fédération Française des Centres de Lutte Contre le Cancer grading system. The Complexity INdex in SARComas (CINSARC) signature comprises 67 genes related to chromosome biogenesis, mitosis control, and chromosome segregation and is correlated with CIN11. It has also demonstrated its prognostic value in 21 out of 39 cancer types, outperforming more than 15000 pre-existing signatures12. Initially established on frozen material with DNA microarrays and then applied with RNA sequencing, it was recently validated on formalin-fixed paraffin-embedded (FFPE) tissue thanks to the NanoString® technology with the code set NanoCind® developed by our team13.

In the present report, we investigated whether CINSARC better predicted tumor recurrence and survival in stage II–III CRC in two independent retrospective cohorts.

Materials and methods

Patient selection

Two independent retrospective patient cohorts were analyzed in this study. Cohort 1 was retrieved from The Cancer Genome Atlas-Colon Adenocarcinoma (TCGA-COAD) data and included 297 stage II–III CRC patients. RNA sequencing, clinicopathological data including TNM stage and follow-up data were obtained for each case14. Cohort 2 consisted of 169 stage II–III CRC patients who were surgically resected at the Centre Hospitalo-Universitaire in Toulouse, France, between 2008 and 2013, for whom we obtained FFPE blocks, clinicopathological data including TNM stage (defined for each patient in Digestive Oncology Multidisciplinary Consultation Meetings) and follow-up data. For all cases, FFPE blocks of surgical resections were used to build tissue microarrays (TMAs). Three pathologists performed a review of hematoxylin- and eosin-stained (H&E-stained) tumor sections to define diagnostic areas (JM, ACB, JS). Two representative cores (2 mm-punch size) were taken so each case was included in duplicate in the array.

Long-term oncologic outcomes were analyzed based on progression-free interval (PFI) (cohort 1) or disease-free survival (DFS) (cohort 2) and OS (cohorts 1 and 2). PFI was defined as the time from diagnosis to a new tumor event (progression of disease, local recurrence, distant metastasis, new primary tumors, death), DFS was defined as the time from surgery to the first event of relapse of CRC (local recurrence or distant metastasis), and OS was defined as the time from surgery to death from any cause. Cases were censored after 5 years of follow-up.

Clinicopathological features and outcomes are described in Table 1.

Table 1 Summary of patient information from two cohorts in this study.

Immunohistochemistry (IHC) staining, MMR and CDX2 status

MMR status data identified by IHC was available for 222/297 patients in cohort 1. MMR status was identified by IHC for all cases in cohort 2. TMA slides were stained with anti-MLH1 (ES05, Agilent (Santa Clara, California, United-States), ready-to-use (RTU)), anti-PMS2 (EP51, Agilent, RTU), anti-MSH2 (FE11, Agilent, RTU), anti-MSH6 (EP49, Agilent, RTU) and anti-CDX2 (EP25, Leica (Wetzlar, Germany), RTU). Tissue slides were stained on a Bond III automatic stainer (Leica) and revelation was performed with the Bond Polymer Refine Detection kit (DS9800, Leica). For interpretation, the slides were evaluated by light microscopy by pathologists experienced with interpreting IHC and MMR studies (JM, ACB, JS)15, and CDX2 status was identified as previously described (ACB, JS)9.

RNA extraction, NanoCind®, and CINSARC classifications

To establish CINSARC classification in cohort 1, we used the RNA-sequencing data downloaded from the TCGA-COAD database on the Genomic Data Commons (GDC) Data Portal. For cohort 2, RNA was extracted from all FFPE blocks (n = 169) using the High Pure FFPET RNA Isolation Kit (Roche, Bâle, Switzerland) according to the manufacturer’s instructions. RNA extraction was performed from full sections of tumors on a representative FFPE block from the initial surgical resection material and the whole tumor area was selected. One to ten 10 µm-slides were prepared depending on the tumor area which was evaluated on H&E stained-slides by a pathologist, with a minimum area required of 8 mm2 up to more than 100 mm2 (ACB). To establish CINSARC classification, we used the NanoString® technology with the nCounter code set (NanoCind®) developed by the team, comprising a panel of 75 probes, including 67 distinct test probes derived from the 67 CINSARC genes (Supplementary Table 1) and 8 from housekeeping genes for normalization purposes13.

Subjects from the two cohorts were assigned to two groups (C1: good prognosis and C2: poor prognosis) using the nearest centroid method as previously described11. Centroids C1 and C2 were defined in each cohort from selected cases as follows: for PFI/DFS, patients without disease after 3 years of follow-up and patients with a progression/relapse; for OS, patients alive after 3 years of follow-up and patients deceased within 5 years after surgery.

CMS classification

To compare its prognostic value with the CINSARC classification, we established the CMS classification for cohort 1. The RNA-sequencing-based CMS classifier developed by Guinney et al. was applied to RNA-sequencing data from cohort 1 as previously described6.

Functional enrichment analysis

For cohort 1, RNA-sequencing raw data was obtained from the TCGA-COAD database on the GDC data portal with the R GenomicDataCommons package. Differential gene expression analysis between C1 and C2 groups was performed with the R DESeq2 package. Gene Set Enrichment Analysis with the (GSEA) was performed with using the database MSigDB c5.go.bp.v7.4 available on with 4010 gene sets and with the CINSARC gene set.


All bioinformatic and statistical analyses were conducted with R statistical software version ≥4.0.4. Cohorts were compared with χ2 test for categorical predictors (Fisher’s exact test for theoretical numbers less than 5) and Student’s test for continuous predictors. Patient subgroups were compared with respect to DFS, PFI and OS by using Kaplan–Meier survival curves, log-rank tests, and multivariate analyses based on the Cox proportional hazards method.


CINSARC is a significant prognostic factor and outperforms TNM staging system and CMS classification

The prognostic value of CINSARC was tested on cohort 1 comprising 297 stage II–III CRC. We established the classification from the available RNA-sequencing data and all cases were interpretable. For PFI, 124 cases were classified C1 and 173 were classified C2, whereas for OS, 121 cases were classified C1 and 176 were classified C2. Clinicopathological features and outcomes of each group are described in Supplementary Table 2. CINSARC split the population into two groups with different prognoses in terms of survival: the C2 group had a higher risk of progression (PFI: P = 1.68 × 10−2; HR = 1.87 [1.11–3.16]) and a higher risk of death (P = 3.73 × 10−3; HR = 2.45 [1.31–4.59]) than the C1 group (Fig. 1).

Fig. 1: CINSARC in cohort 1.
figure 1figure 1

PFI (A) and OS (B) analyses according to CINSARC classification for stage II–III CRC patients; PFI (C) and OS (D) analyses according to CINSARC classification for stage II–III CRC patients without adjuvant treatment; PFI (E) and OS (F) analyses according to CINSARC classification for stage II–III CRC patients with adjuvant treatment; PFI (G) and OS (H) analyses according to CINSARC classification for stage II–III CRC patients with and without adjuvant treatment in cohort 1.

In univariate analysis, CINSARC classification, TNM stage, CMS classification, MMR status and VELIPI were significant prognostic factors for PFI (Supplementary Fig. 1). Among these, CINSARC and CMS classifications were independent prognostic factors, CINSARC classification being the most significant (Table 2). Regarding OS, CINSARC classification, TNM stage and MMR status were prognostic factors (Supplementary Fig. 1). All were independent prognostic factors, MMR status being the most significant (Table 2). Therefore, in that cohort, CINSARC proved to be an independent prognostic factor outperforming the gold-standard TNM classification.

Table 2 Univariate and multivariate analyses for PFI and OS in patients with stage II–III CCR in cohort 1.

To evaluate the impact of adjuvant treatment on CINSARC classification, we next analyzed CINSARC classification in subgroups of patients having received an adjuvant treatment or not. This showed that for patients who did not receive an adjuvant treatment, CINSARC classification split the population into two groups with significantly different survival, regardless of the outcome (PFI: P = 5.64 × 10−3; HR = 2.81 [1.31–6.03]; OS: P = 3.28 × 10−2; HR = 2.64 [1.04–6.67]). On the contrary, there was no significant difference between C1 and C2 groups in the population of patients who received an adjuvant treatment (PFI: P = 7.86 × 10−1; HR = 1.14 [0.44–2.97]; OS: P = 1.94 × 10−1; HR = 2.69 [0.57–12.64]) (Fig. 1). Furthermore, there was no significant survival difference between the C1 non-treated group and the C2 treated one (PFI: P = 1.5 × 10−1; HR = 1.82 [0.8–4.16]; OS: P = 3.3 × 10−1; HR = 1.66 [0.59–4.68]) (Fig. 1).

Validation of CINSARC’s prognostic value on an independent cohort

To validate these results, CINSARC was applied to a second cohort comprising 169 tumors. The CINSARC classification was applied to RNA extracted from FFPE blocks (see M&M) and 168/169 (99.4%) cases were interpretable. DFS was available for 167 patients: 103 cases were classified C1 and 64 were classified C2. For OS, information was available for 167 patients: 81 were classified C1 and 86 were classified C2. Clinicopathological features and outcomes of each group are described in Supplementary Table 3. In this cohort, CINSARC significantly split the population of CRC into two groups, regardless of the outcome (DFS: P = 2.96 × 10−4; HR = 3.05 [1.62–5.77] and OS: P = 4.07 × 10−4; HR = 3.54 [1.68–7.49]) (Fig. 2).

Fig. 2: CINSARC in cohort 2.
figure 2

DFS (A) and OS (B) analyses according to CINSARC classification for stage II–III CRC in cohort 2.

In univariate analysis, CINSARC classification, TNM stage and CDX2 status were significant prognostic factors for DFS. All three were independent risk factors in the multivariate analysis, CINSARC classification being the most significant (Table 3). With regard to OS, CINSARC classification and CDX2 status were prognostic factors and CINSARC classification was independent in multivariate analysis (Table 3). Thus, these results validate CINSARC classification as a strong prognostic factor for stage II–III CRC, outperforming the TNM classification and other factors used in clinical practice such as perforation status, tumor grade, VELIPI, number of examined lymph nodes and MMR status.

Table 3 Univariate and multivariate analyses for DFS and OS in patients with stage II–III CCR in cohort 2.

In TNM sub-group analysis, CINSARC predicted DFS and OS in stage II CRC with a worse outcome in the C2 group (DFS: P = 1.38 × 10−3; HR = 5.4 [1.69–17.23]; OS: P = 1.99 × 10−2; HR = 3.19 [1.14–8.96]) but significantly predicted only OS in stage III CRC (DFS: P = 9.4 × 10−2; HR = 1.91 [0.89–4.12]; OS: P = 1.11 × 10−2; HR = 3.75 [1.25–11.23]) (Fig. 3). We also investigated its performance among microsatellite stable (MSS) patients which split this sub-population in two groups with significantly distinct DFS and OS (DFS: P = 5.91 × 10−4; HR = 3.24 [1.59–6.6]; OS: P = 1.03 × 10−3; HR = 3.42 [1.57–7.48]) (Supplementary Fig. 2).

Fig. 3: CINSARC in cohort 2 by stage.
figure 3

DFS (A) and OS (B) analyses according to CINSARC classification for stage II CRC; DFS (C) and OS (D) analyses according to CINSARC classification for stage III CRC in cohort 2.

CINSARC genes are overexpressed in C1 tumors

To decipher the biological mechanisms involved in the different outcome between groups C1 and C2, we performed GSEA for the 297 cases included in cohort 1. No gene set was significantly enriched in C2 tumors (FDR < 0.05). Interestingly, the CINSARC gene set was significantly enriched in group C1 (FDR = 0; NES = 2.08) (Supplementary Fig. 3). The five most enriched gene sets in group C1 were: negative regulation of cell cycle phase transition (FDR = 1.12 × 10−2; NES = −2.26), metaphase anaphase transition of cell cycle (FDR = 8.49 × 10−3; NES = −2.23), negative regulation of mitotic cell cycle (FDR = 6.44 × 10–3; NES = −2.23), chromosome segregation (FDR = 5.82 × 10−3; NES = −2.22) and regulation of chromosome separation (FDR = 5.04 × 10−3; NES = −2.22) (Supplementary Fig. 3). The overexpression of CINSARC genes in group C1 was confirmed in cohort 2 (Supplementary Fig. 4).


CRC is the second most common cause of cancer death worldwide1 and patient management depends mostly on the stage of the disease determined by TNM staging3. However, predicting clinical outcome in stage II–III CRC remains a challenge, despite the variety of available biomarkers3,6,9,16,17,18 which are neither precise enough nor immediately applicable in routine practice. In this study, we show that CINSARC improves the ability to discriminate the risks of recurrence and death specifically in stage II–III CRC. In two independent stage II–III CRC cohorts, it identified a group of tumors with a poor outcome, whereas TNM staging either did not significantly differentiate tumors or was less discriminating than CINSARC. Furthermore, it outperformed other prognostic factors used in clinical practice and was more significantly discriminating than recently proposed ones such as CMS classification and CDX2 expression6,9.

Our results show that the CMS classification cannot be used as a prognostic tool for CRC6. A high percentage of patients were not classified in any of the CMS groups (62 patients, 20.9 %), a feature not consistent with its use in clinical practice. Despite being higher than in previous studies (0–14.5%)6,7,16,19,20,21, this high percentage represents a major limitation. The CMS4 group was initially described as having a poor progression and OS in localized stages6,7,21. However, we did not observe this in our stage II–III cohort 1 (Supplementary Fig. 1), suggesting that the CMS lacks robustness. Recently, Marisa et al. studied intratumor heterogeneity in the PETACC8 trial cohort of 1779 patients by using their deconvoluted transcriptomic profiles. They discovered that up to 55% of tumors corresponded to a mixture of at least two different CMS groups and that this heterogeneity was more significantly associated with a poor DFS and OS than each CMS group separately22. Furthermore, the current lack of standardization of the tools used to establish the CMS classification (different classifiers6, IHC panels7,16 and NanoString® panels19,20,23) further limits its use in routine. Therefore, while the CMS is useful for unraveling the biology of these tumors and could be used to develop new precision therapeutics, it is not suitable for use as a prognostic biomarker in clinical practice.

As for guiding treatment decision-making, CINSARC is currently being tested prospectively in sarcomas in the multicenter clinical trial CHIC-STS led by our team (NCT04307277)24. We hypothesize that in addition to being a prognostic biomarker of CRC, CINSARC could serve to better select patients for adjuvant chemotherapy. For example, there was no significant survival difference in cohort 1 between the C1 non-treated group and the C2 treated one, likely meaning that patients with stage II–III and CINSARC C2 CRC benefit from adjuvant treatment and that the latter might be an option for patients with stage II C2 CRC, who are commonly treated with surgery alone. On the contrary, patients with stage III C1 CRC, who are commonly treated with surgery and adjuvant treatment and are thus potentially exposed to chemotherapy-induced side-effects, might be protected from them. Dedicated prospective trials are needed to validate this hypothesis.

To better understand the biological mechanisms involved in the different outcome between C1 and C2 CRC, we performed functional enrichment analysis. As previously observed in different types of sarcomas and breast carcinomas, we expected an enrichment of the biological mechanisms implicated in cell cycle and mitosis due to aggressive cell proliferation, as well as an enrichment of the CINSARC gene set in the C2 group11,25, thus explaining the worse prognosis of these tumors. Strikingly, we observed the opposite, with a significant enrichment of these biological pathways as well as an overexpression of CINSARC genes in the C1 group. These biological pathways and the CINSARC signature indirectly reflect CIN11. The acquisition of genomic instability is a crucial feature of CRC oncogenesis with three major pathways: CIN, MSI/hypermutated and CIMP pathways14,26. CIN plays a major role in CRC oncogenesis, as it is implicated in 65–70% of sporadic CRC27. It is characterized by aneuploidy and loss of heterozygosity, it may arise from defects in chromosomal segregation, telomere stability and the DNA damage response27, and it is reflected by the Global Genomic Index (GGI)28, which corresponds to the fraction of rearranged genome. However, its role in CRC prognosis is still debated28,29. Orsetti et al. observed a non-linear association between GGI level and prognosis: the best prognosis was found for tumors with a low GGI and the worst for tumors with a median GGI, whereas tumors with a high GGI had an intermediate outcome28. Similarly, Andor et al. studied the association between copy number variants (CNVs) and clinical outcome in a variety of tumor types. They found that tumors with either the lowest or highest rate of CNVs had the most favorable outcomes, suggesting that either too little or too much CIN can be detrimental for tumor cells29. Thus, we hypothesize that the low expression of CINSARC genes in CRC is correlated with an intermediate level of CIN. No significantly enriched gene sets were highlighted in the C2 group, which suggests that it comprises more heterogeneous molecular intrinsic subtypes than the C1 group. As previously observed in sarcomas30, there might not be just one group of C2 tumors but several, explaining the lack of identified biological processes implicated in their poor prognosis. Other parameters involved in the prognosis such as the immune contexture17,18, which was not taken into account in our study, could also explain this observation. Thus, it would be interesting to further explore these tumors with microenvironment data as well as to compare the CINSARC classification with the prognostic performance of an immune classification such as the Immunoscore18.

Our study has some limitations. First, it was retrospective. Second, there may be a technological bias since we used different technologies to quantify RNA expression in both cohorts (RNA-sequencing data in cohort 1 and NanoCind® data in cohort 2) and with different types of samples (frozen samples in cohort 1 and FFPE samples in cohort 2). However, we believe it only reinforces the robustness of the CINSARC classification by showing that regardless of the technology used, it remains a significant prognostic biomarker. In cohort 2, we did not evidence any prognostic value of MMR status, a biomarker currently used in routine clinical practice. However, to detect a benefit associated with MSI with an HR of 0.65 with 80% power and 5% type I error, 300 events would be required, demonstrating the lack of power of our cohort and explaining this result31.

In conclusion, CINSARC reliably estimated prognostic risk in stage II–III CRC patients in two independent cohorts, representing a total of 466 stage II–III CRC patients. It is immediately applicable in clinical practice when used with NanoString® technology and the NanoCind® code set. There was a correlation between low expression of the CINSARC genes and a poor prognosis in the high-risk group, possibly due to an intermediate level of CIN. Further studies are needed to validate these results in prospective cohorts including other new prognostic factors (immune contexture and tumor budding) and to clarify all the determinants of the poor prognosis in the heterogenous group of high-risk C2 tumors.