Transcriptomic Signatures of MSI-High Metastatic Colorectal Cancer Predict Efficacy of Immune Checkpoint Inhibitors

Approximately 5% of metastatic colorectal cancers (mCRC) harbor microsatellite instability (MSI) and/or mismatch repair deficiency (dMMR), associated with high sensitivity to immune checkpoint inhibitors (ICI; refs. 1, 2). Since the publication of the Keynote 177 trial, showing a significant improvement in median progression-free survival (PFS) with first-line pembrolizumab compared with chemotherapy (16.5 vs. 8.2 months; HR, 0.60; ref. 2) in MSI mCRC, pembrolizumab has become the standard of care in this setting. In subsequent lines of treatment, several studies have also demonstrated the efficacy of ICI in this subgroup of mCRC patients (1, 3–6). More recently, in patients with localized MSI colon or rectal cancer, impressive pathologic responses have been obtained with ICI in the neoadjuvant setting (7–10).

Despite these impressive results, 10% to 40% of patients with MSI mCRC will experience progressive disease at first disease assessment under ICI treatment and will be classified as primary resistant such that it remains crucial to identify these primary resistant patients. In the Keynote 177 trial, pembrolizumab was not only associated with improved response rate, PFS and overall survival (OS), but also with improved patients’ quality of life. Identifying primary resistant patients may thus, in the future, allow physicians to explore new intensified therapeutic strategies, such as combination of ICIs or combination of ICIs and chemotherapy in this patient subgroup, while preserving the excellent quality of life associated with pembrolizumab monotherapy in the others.

Few data are available on potential biomarkers of resistance to ICIs in MSI mCRC. Sui and colleagues showed in a cohort of 62 patients treated with anti–programmed cell death protein 1 (PD-1) alone or combined with non-ICI drugs, that local inflammatory conditions (defined by obstruction, perforation, peritonitis, or other radiologically diagnosed inflammation in abdominal viscera or metastatic sites) and the neutrophil-to-lymphocyte ratio (NLR) were both associated with a poor immune status and a poor response to ICI, with a suspected major role of neutrophil infiltration in conferring immunosuppression (11). Others have identified ECOG-PS and other inflammatory scores as important to predict outcome in patients with MSI mCRC treated with ICIs (12, 13).

On top of these clinical and biological factors, analysis of the heterogeneity in cellular composition of the tumor microenvironment (TME) seems to be of great potential in predicting the efficacy of ICI in oncology (14, 15). For instance, Kieffer and colleagues reported the identification of eight cellular clusters of cancer-associated fibroblasts (CAF) of the S1 subtype through single-cell analysis of primary breast cancer samples and documented an association of three clusters (ecm-myCAF, TGFβ-myCAF, and wound-myCAF) with primary resistance to ICI in two cohorts of patients with metastatic melanoma and metastatic non–small cell lung cancer (NSCLC; ref. 14). Another single-cell transcriptomic analysis dedicated to MSI and MSS CRCs has recently described 204 gene expression programs of cell subtypes, states of TME (T/NK/innate lymphoid cells, B cells, plasma cells, mast cells, myeloid cells, stromal cells) and epithelial cells, with differential expression between MSI and MSS tumors (16).

Here, we sought to transcriptionally profile two independent cohorts of patients with MSI/dMMR mCRC and subsequently apply these molecular signatures determined by single-cell analyses (14, 16) with a view to establish a molecular classification predictive of primary resistance to ICI.

This retrospective, multicenter cohort involved three Italian centers (Fondazione IRCCS Istituto Nazionale dei Tumori, Milan, Azienda Ospedaliero, Universitaria Pisana, Pisa, and Veneto Institute of Oncology IRCCS, Padova). Patients with MSI and/or dMMR mCRC, who received ICI (anti-PD(L)-1 single-agent therapy or the combination of anti-PD-1 plus anti-CTLA4 or anti-LAG3), were included in this study. The inclusion criteria encompassed those treated in clinical practice or prospective clinical trials between July 2015 and August 2021, regardless of the line of therapy, and who underwent a resection of their primary tumor before starting ICI. No ICI-based regimens before tumor resection were allowed. dMMR/MSI status was determined by local testing with IHC and/or multiplex PCR.

The retrospective cohort consisted of patients treated at the Mayo Clinic three-site Comprehensive Cancer Center or the Mayo Health System. Patients with dMMR or MSI mCRC had received pembrolizumab as a first-line or second-line therapy between April 1, 2015, and January 1, 2022 (17). Among this cohort, 37 patients had tissues available for this study. Clinical data were abstracted from the electronic health record. Tumor response to pembrolizumab was determined according to RECIST version 1.1.

Data collected concerned the different lines of treatment, baseline patients’ characteristics at the beginning of ICI line of treatment, the best response rate under ICI according to RECIST 1.1 criteria, and PFS and OS evaluated from ICI initiation. NLR was defined as the absolute number of neutrophils divided by the absolute number of lymphocytes.

Data were obtained from electronic review according to strict privacy standards.

The study was conducted in accordance with the Declaration of Helsinki and was approved by the Institutional Review Boards and Ethics Committees of participating institutions. All patients alive at the time of the study received an information note and gave their consent for anonymous data collection.

In the two cohorts, tumor RNA was extracted from macrodissected formalin-fixed, paraffin-embedded (FFPE) blocks from primary tumors using the Maxwell RSC RNA FFPE Kit (Promega, France). PolyA-RNA sequencing (RNA-seq) library preparation protocols were performed using 400 ng of template RNA and the QuantSeq 3′mRNA-Seq Kit FWD for Illumina (Lexogen) according to the manufacturer's instructions. Libraries were sequenced on NovaSeq6000 (Illumina).

Bioinformatic analyses were performed with R version 4.1.2. FASTQ RNA-seq files were mapped using STAR aligner 2.7.9a (18), and raw read counts were obtained using Rsubread R package (19). Prior to mapping, the genome index was built on GRCh38.p13 human genome. More than 10,000 genes were detected across samples, so all samples were kept for further analysis. Batch correction was performed using Combat_seq function from the sva R package (20). Mitochondrial and housekeeping genes were removed from the raw count matrix. Then, for each sample, we scaled the corresponding vector of raw counts: counts were divided by the total number of counts of the vector and multiplied by the median of total number of counts across samples. Then, the scaled data were log₂ transformed.

The transcriptomic signatures of Pelka and colleagues (16), excluding MMRp epithelial programs and global epithelial programs, and the two CAF-S1 transcriptomic signatures of Kieffer and colleagues (14), were analyzed using the GSVA package and the GSVA method (21) in the exploratory cohort. Then, the most variant signatures among the samples, with a standard deviation above 0.35, were selected for the unsupervised clustering of the two cohorts (Supplementary Fig. S1).

The identified unsupervised clusters were compared on the basis of patients’ characteristics and best response during ICI according to RECIST 1.1 criteria, using the χ² test for categorical variables and the Kruskal–Wallis test for continuous variables. PFS and OS were estimated according to these clusters with Kaplan–Meier method. PFS was defined as the time from the beginning of the treatment with ICI to progression or death from any cause, whichever occurred first. OS was defined as the time from the beginning of ICI to death from any cause. Patients known to be alive and/or free of progression were censored at the date of their last follow-up. PFS and OS were investigated using univariate and multivariate Cox models. The adjustment factors used in the multivariate analyses were the variables with a P value < 0.2 in the univariate analyses or reported as relevant in the literature.

Analyses were done with R (version 4.2). Two-sided P values of less than 0.05 were considered to be statistically significant.

The data generated in this study are publicly available with the Code Ocean compute capsule: https://codeocean.com/capsule/8990462/

A total of 110 patients were included in the exploratory cohort analysis and 103 (93.6%) had interpretable 3′ RNA-seq data. Table 1 shows patient baseline characteristics, with a median age of 61 years, 42% women, right-sided primary tumor in 70%, and BRAF V600E mutation in 38% of cases. ICIs were administered in first-line therapy in 37% of cases, in second-line therapy in 38%, and beyond second-line therapy in 25% of cases. ICI regimens were anti–PD-1 or anti–programmed death-ligand 1 (PD-L1) monotherapy in 55% of cases and a combination consisting of anti–PD-1 and anti-CTLA4 or anti–PD-1 and anti-LAG3 in 32% and 13% of the cases, respectively (Table 2). Primary resistance to ICI was reported in 16% of patients (Table 2).

In the validation cohort, there were 35 patients with interpretable 3′ RNA-seq data (95%). The median patient age was 77 years, and 57% were women. Patients received ICI treatment in the first line in 11 patients (31.4%) and in the second line in 24 patients (68.6%).

We used transcriptomic data to calculate GSVA scores for a panel of signatures (N = 23 samples) derived from the Kieffer (14) and the Pelka (16) studies and subsequently performed unsupervised clustering of samples from the exploratory cohort. As shown in the heat map displayed in Fig. 1, this method identified three clusters of tumors with distinct transcriptional profiles, which we named “stromal^HIGH-proliferation^LOW” (cluster A), “stromal^HIGH-proliferation^MED” (cluster B), and “stromal^LOW- proliferation^HIGH” (cluster C). Cluster A (N = 23 samples) was characterized by an enrichment of signatures related to the stroma such as the CAF-S1 signature and its subclusters (detox_iCAF, ecm_myCAF, wound_myCAF, and TGF_myCAF) as well as the stroma-related signatures from Pelka and colleagues [pS07, pS03 (ACTA2) and pS08 (collagens)]. Cluster A was also associated with a reduced expression of the Pelka signatures related to tumor cell cycle and different cellular constituents of the TME (B cells, T cells, NK cells, innate lymphoid cells, myeloid cells, mast cells, plasma cells, and stroma). Cluster B (N = 41 samples) was also characterized by an overexpression of signatures related to the stroma and CAF-S1, but differed from cluster A with an overexpression of Pelka signatures related to interferon-stimulated genes (ISG), myeloid/B cells, and comparatively higher levels of miscellaneous signatures related to the cell cycle. Finally, the main hallmarks of cluster C (N = 39 samples) were the under expression of the entire set of stromal signatures and an overexpression of cell cycle–associated signatures.

Cluster A (“stromal^HIGH-proliferation^LOW”) was enriched in patients with immediate progression under ICI (30% vs. 12% in cluster B and 8.1% in cluster C; P = 0.074). Conversely, the objective response rate was 52%, 66%, and 76% in clusters A, B, and C, respectively (P = 0.10; Table 2).

In univariate analyses, PFS and OS were significantly shorter for patients belonging to cluster A compared with patients within clusters B and C (median PFS, 11.1 months [95% confidence interval (CI), 3.5–NA] in cluster A, and not reached for clusters B and C; P < 0.001; and median OS, 20.5 months (95% CI, 17.7–NA) in cluster A, and not reached for clusters B and C; P < 0.001) at a median follow-up of 36.6 months (95% CI, 32.2–47). The 2-year PFS rates were 33.5% (95% CI, 18.5–60.7), 80.5% (95% CI, 69.2–93.6), and 78.3% (95% CI, 66.1–92.9) in clusters A, B and C, respectively. When merging patients within clusters B and C into a single group, we observed that these patients had significantly longer PFS (HR, 0.30; 95% CI, 0.15–0.58; P value < 0.001) and OS (HR, 0.26; 95% CI, 0.12–0.55; P value < 0.001) compared with those belonging to cluster A (Fig. 2). In the subgroup of patients treated with ICI beyond first line therapy, PFS was still significantly longer in patients belonging to clusters B or C versus cluster A (P = 0.002; Supplementary Fig. S2).

To check the potential bias related to MSI status misdiagnosis, we considered MSI-high status “uncertain” if it was determined by a single method (PCR or IHC), without available tumor mutational burden data, and undiagnosed Lynch syndrome. MSI-high status was uncertain in 10 patients (43.5%) belonging to cluster A and 21 patients (26.2%) belonging to clusters B and C (P value = 0.11). In our exploratory cohort, there is no evidence that differences exist in dMMR/MSI-H status by transcriptomically determined clusters A, B, and C (Supplementary Fig. S3).

In multivariate analysis with adjustment for age, gender, number of metastatic sites, treatment line, treatment regimen (mono or combo-therapy), and the NLR, PFS was significantly longer in patients belonging to either clusters B or C compared with cluster A (HR, 0.19; 95% CI, 0.08–0.45; P < 0.001 and HR, 0.25; 95% CI, 0.10–0.59; P = 0.02; Table 3).

The unsupervised clustering performed with the same panel of signatures in the validation cohort revealed the same separation into three main clusters: cluster A (“stromal^HIGH-proliferation^LOW”, N = 9), cluster B (“stromal^HIGH-proliferation^MED”, N = 9), and cluster C (“stromal^LOW- proliferation^HIGH”, N = 17), with transcriptomic profiles similar to the three clusters of the exploratory cohort (Supplementary Fig. S4).

The patients belonging to clusters B or C tended to have prolonged PFS compared with patients belonging to cluster A (HR, 0.54; 95% CI, 0.20–1.43; P = 0.22; Supplementary Fig. S5).

In the subset of patients from the exploratory cohort treated with ICI beyond the first-line setting (N = 65), patient PFS after first-line treatment with conventional chemotherapy was not significantly different according to these three transcriptomically determined clusters (Fig. 3).

In the PETACC8 trial, a subgroup of patients with MSI tumors from stage III colon cancer (22), who received either FOLFOX + cetuximab or FOLFOX treatment, underwent an RNA-seq analysis. Among these patients, there were 34 cases of recurrence. However, the survival after recurrence (SAR) did not show any significant differences based on the identified clusters (Supplementary Fig. S6).

In the exploratory cohort of 103 patients with MSI mCRC treated with ICI, the unsupervised transcriptomic classification identified three groups of tumors with different compositions of TME cells and proliferative capacities of TME /tumor cells. The patients belonging to cluster A, “stromal^HIGH-proliferation^LOW”, had a dismal prognosis with ICI treatment compared with cluster B, “stromal^HIGH-proliferation^MED”, and cluster C, “stromal^LOW-proliferation^HIGH”, with a greater risk of primary progression (30% in cluster A vs. 12% in cluster B and 8.1% in cluster C; P = 0.074) and significantly shorter PFS in multivariate analysis with adjustment for known prognostic factors (2-year PFS rates of 33.5%, 80.5%, and 78.3% in clusters A, B, and C, respectively). Importantly, the prognostic utility of this molecular classification with ICI treatment in MSI mCRC was confirmed in an independent cohort of 35 patients.

In the exploratory cohort, primary resistance was reported in 16% of patients, which is lower than what was reported in the Keynote-177 pivotal trial (29%; ref. 2) or in the PRODIGE 54 randomized phase II trial in the second-line setting (26%; ref. 6). In our data, the median PFS was also comparatively high (not reached after a median follow-up of 36.6 months), possibly due to the fact that 45% of patients were treated with ICI combination regimens. Median PFS and OS were not reached for the overall population nor in patients belonging to clusters B and C indicating their more favorable prognosis compared with cluster A and that longer-term follow-up of this cohort is awaited.

To test the predictive value of this classification for ICI efficacy, we showed that PFS according to our signatures in patients receiving conventional chemotherapy as first-line therapy, and SAR in MSI patients from the PETACC-8 adjuvant trial who have recurred, were not significantly different according to these three clusters. None of the patients from the PETACC-8 trial were treated with ICI, which were not available within 5 years after trial completion (2009). Altogether, these findings strongly suggest the potential predictive value of this molecular classification for the efficacy of ICI in MSI mCRC. We acknowledge that we cannot confirm the predictive value of our signatures due to the absence of a control group of MSI patients treated with chemotherapy alone.

From a biological perspective, overexpression of signatures relating to ISGs and MHC II in the majority of samples of cluster B, may increase the efficacy of ICI, despite high stroma infiltration and moderate proliferation. Two potential mechanisms of tumor intrinsic resistance to ICI are defects in IFNγ signaling in the TME inducing a reduced downstream JAK/STAT signaling pathway and a decrease in the expression of PD-L1 or reduced surface MHC class I or II expression (23). Conversely, the overexpression of IFNG, IFNγ-inducible transcripts and T-effector–IFNγ-associated genes have been shown to be associated with an improvement in efficacy of ICI in patients with melanoma and NSCLC (24–26).

Regarding the stroma, stroma-low/immune-high tumors were predicted to be most responsive to ICI therapy in a cohort of 335 therapy-naïve patients with mCRC (13). Furthermore, the recently published study of Bagaev and colleagues based on unsupervised clustering of RNA-seq data from 468 melanomas, identified four TME subtypes that were subsequently found to be predictive of response to ICI in several cohorts of cancers such as melanoma, bladder cancer and gastric cancer, with the “immune-enriched, non-fibrotic” cluster associated with the highest response rate and the “fibrotic” cluster with the lowest response rate across cancers (27). Finally, our findings are also in line with the publication of Kieffer and colleagues showing that transcriptomic signatures related to CAF-S1 and subgroups of CAF-S1 are associated with primary resistance to ICI in metastatic melanoma and in NSCLC (14). To our knowledge, with respect to cell proliferation, there are no data on transcriptomic signatures relating to the proliferation/cell cycle of TME cells predictive of the efficacy of ICI.

Strengths of our study include the use of established criteria for clinical response determination, validation of the study results in an independent patient cohort, and patients from multiple and international study sites, which increases the generalizability of the findings. Limitations of our work include the retrospective nature, clinical heterogeneity of the study population in terms of treatment line and administered ICI-based regimens, and the absence of a matched contemporary cohort of patients with MSI mCRC who did not receive ICIs in their disease history. This hinders us from providing a methodologically appropriate demonstration of the purely predictive role of the proposed classification. In addition, we only characterized the signatures on primary resected tumors, which does not allow us to generalize our findings to biopsies from patients diagnosed with synchronous metastatic disease. Finally, the comparison of the prognostic value of the clusters under conventional chemotherapy versus under ICI is potentially biased by the fact that the prognostic value of clusters was determined on patients who progressed under chemotherapy before the introduction of ICI.

In conclusion, our unsupervised molecular classification identified three groups of MSI mCRCs with different compositions and proliferative status of TME /tumor cells. The “stromal^HIGH-proliferation^LOW” group is associated with a significantly poorer prognosis in patients treated with ICI, strongly suggesting the predictive value of this classification for the effectiveness of ICI. In clinical practice and due to the development of NGS platforms, we can envision performing 3′ RNA-seq easily from FFPE tumor slides to guide the first-line therapeutic strategy and also to treat patients belonging to cluster A, with a combination of an ICI with other agents, such as conventional chemotherapy to promote immunogenic cell death, or combination with another ICI or a targeted approach to overcome defects in IFNγ signalling and antigen presentation (23, 28, 29).

C. Gallois reports personal fees from Servier, Pierre Fabre, and Merck outside the submitted work. J. Taieb reports personal fees from Amgen, Astellas, AstraZeneca, BristolMyersSquib, Merck, MSD, PierreFabre, Servier, and Novartis outside the submitted work. M. Sroussi reports grants from Fondation pour la Recherche Meedicale during the conduct of the study. S. Lonardi reports research funding (to institution) from Amgen, Astellas, AstraZeneca, Bayer, Bristol Myers Squibb, Daiichi Sankyo, Hutchinson, Incyte, Merck Serono, Mirati, MSD, Pfizer, Roche, and Servier; personal honoraria as an invited speaker from Amgen, Bristol Myers Squibb, Incyte, GSK, Lilly, Merck Serono, MSD, Pierre-Fabre, Roche, and Servier; and participation in advisory boards for Amgen, AstraZeneca, Bristol Myers Squibb, Daiichi Sankyo, Incyte, Lilly, Merck Serono, MSD, Servier, and GSK. F. Bergamo reports personal fees from Bristol Myers Squibb, AAA, Lilly, Servier, and Eisai outside the submitted work. F. Pietrantonio reports grants and personal fees from Amgen and Bristol Myers Squibb; personal fees from Merck-Serono, Bayer, Servier, PierreFabre, Takeda, and MSD; and grants from Agenus and AstraZeneca outside the submitted work. F. Corti reports nonfinancial support from Ipsen, AAA, and Istituto Gentili outside the submitted work. F.A. Sinicrope reports personal fees from Guardant Health outside the submitted work and has a patent for Roche Diagnostics issued and with royalties paid. C. Cremolini reports personal fees from Amgen, MSD, Nordic Pharma, Organon, Pierre Fabre, and Roche and grants and personal fees from Bayer, Servier, and Merck outside the submitted work. P. Laurent-Puig reports grants from Institut National du Cancer during the conduct of the study and personal fees from Biocartis, Amgen, Pierre Fabre, and Servier outside the submitted work. No disclosures were reported by the other authors.

C. Gallois: Conceptualization, resources, data curation, software, formal analysis, supervision, validation, investigation, visualization, methodology, writing–original draft. M. Landi: Conceptualization, resources, data curation, formal analysis, investigation, visualization, methodology, writing–original draft. J. Taieb: Conceptualization, resources, data curation, software, formal analysis, supervision, validation, investigation, visualization, methodology, writing–original draft. M. Sroussi: Data curation, software, formal analysis, methodology, writing–review and editing. B. Saberzadeh-Ardestani: Resources, data curation, investigation, writing–review and editing. A. Cazelles: Resources, data curation, investigation, writing–review and editing. S. Lonardi: Conceptualization, resources, data curation, validation. F. Bergamo: Conceptualization, resources, data curation, validation, visualization. R. Intini: Conceptualization, resources, data curation, validation. G. Maddalena: Resources, data curation, validation. F. Pietrantonio: Conceptualization, resources, data curation, supervision, validation, visualization, methodology. F. Corti: Conceptualization, resources, data curation, validation. M. Ambrosini: Resources, data curation, validation. A. Martinetti: Resources, data curation, validation. M.M. Germani: Resources, data curation, validation, investigation. C. Boccaccio: Resources, data curation, validation. G. Vetere: Resources, data curation, validation. S. Mouillet-Richard: Conceptualization, formal analysis, supervision, validation, visualization, methodology, writing–original draft. A. de Reynies: Conceptualization, resources, software, formal analysis, supervision, validation, investigation, visualization, methodology. F.A. Sinicrope: Conceptualization, data curation, supervision, validation, investigation, visualization, writing–review and editing. C. Cremolini: Conceptualization, resources, data curation, formal analysis, supervision, validation, investigation, visualization, methodology, writing–original draft. P. Laurent-Puig: Conceptualization, resources, data curation, software, formal analysis, supervision, validation, investigation, visualization, methodology, writing–original draft.

This work was supported by research funding from the French National Cancer Institute (INCa), the French Ministry of Solidarity and Health, INSERM SIRIC CARPEM (INCa-DGOS-Inserm-12561: SIRIC CARPEM) and Labex Immuno-oncology, and benefited from equipment and services from the iGenSeq core facility, at ICM.

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.