Patient population and study design
A total of 130 prospectively collected mCRC patients were assessed for eligibility, including 57 patients from the NIPICOL clinical trial (C1, NCT03350126) and 73 patients from the prospective ImmunoMSI cohort (C2) 26. Clinical and disease characteristics of patients from C1 and C2 are summarized in SUPPLEMENTARY TABLE S1 and SUPPLEMENTARY TABLE S2. In C1, the analysis was restricted to 54 patients after removing 3 mCRC samples with misdiagnosed dMMR/MSI status. Whole Exome Sequencing (WES) and RNA Sequencing (RNASeq) were performed on 23 and 44 collected mCRC +/- matched normal colonic mucosa paraffin-embedded samples, respectively, after removing 29 samples with lack of materials and/or insufficient quality (Fig. 1A). In C2, targeted next-generation-sequencing (NGS) and RNASeq were performed on 66 and 73 mCRC +/- matched normal colonic mucosa paraffin-embedded samples, respectively, after removing unqualified samples for similar reasons, i.e., insufficient quantity and/or low-quality level (Fig. 1A). The CONSORT-like clinical and molecular diagram in FIGURE 1 outlines the methodology workflow of the study.
FIGURE 1 also summarizes the flow chart (Fig. 1A) and the current design of the study (Fig. 1B). In an effort to optimize the available clinical data for the purpose of identifying available RNA/DNA predictive markers of resistance to ICI in this unprecedent prospective series of ICI-treated patients with MSI mCRC, their putative clinical relevance was examined in both the C1 and C2 cohorts independently. Were considered as clinically relevant the predictors whose status was significantly associated with ICI resistance in C1 and yet in C2, using progression-free survival (PFS) by iRECIST (RECIST for Response Evaluation Criteria in Solid Tumor) criteria for survival analyses (iPFS) 29. In FIGURE 1 is also indicated how the bioinformatics analyses were carried out. In particular, we indicate how, following analyses of the WES and RNAseq profiles, using previously proposed markers of ICI resistance: (i) we developed a classifier from C1 to predict resistance to ICI in mCRC based on specific combinations of somatic variants occurring in a large panel of microsatellite-containing genes and then validated this classifier in the C2 validation dataset; (ii) we unraveled the phenotypic diversity of mCRC by applying an unsupervised ICA (Independent Component Analysis) on C1 tumor transcriptomes to identify RNA signatures that would be associated with treatment resistance in this discovery dataset and then validated in the C2 validation dataset. Finally, in an effort to validate that both the identified DNA and RNA signatures were truly predictive of response to ICI in patients with MSI CRC, we validated the absence of impact of both these signatures on the survival of patients issued from several prospective and retrospective additional cohorts of MSI CRC untreated with ICI, i.e., in the METAMSI (metastatic MSI CRC, retrospective) 25, IDEA 27, MOSAIC 28 (nonmetastatic MSI CRC, prospective, clinical trials), TIMSI and TCGA (nonmetastatic MSI CRC, retrospective) cohorts (see also materials and methods section for further details) (Fig. 1B).
The level of MSI and TMB in tumor DNA does not predict response to ICI in patients following exclusion of ICI-treated mCRC with misdiagnosed dMMR/MSI status
We first focused DNA analyses on the genomic quantitative indexes that are related to MSI in tumor DNA, i.e., the TMB and MSIsensor score, whose level was previously reported to predict response to ICI in patients with metastatic dMMR tumor 15,16. FIGURE 2A shows the levels of TMB and MSIsensor score in the C1 cohort when assessing their level on the basis of WES data. Expectedly, these indexes were highly correlated to each other (Fig. 2A and SUPPLEMENTARY FIGURE S1). In this clinical trial (C1 cohort), we previously reported misdiagnoses in 3 patients included in NIPICOL (false positive MSI cases with MSI-PCR and/or IHC, i.e., TN1, TN2, TN3, detected here with negative MSIsensor score, expectedly) 25 and a lack of sensitivity of the MSIsensor tool to detect this tumor phenotype in 3 other patients (false negative MSS cases, FN1, FN2, FN3) 30. FIGURE 2A shows the TMB and MSIsensor index values in samples from C1 (left panel), including FN and TN cases, together with their MSICare score (right panel) whose higher performance to detect MSI in CRC from their NGS profiles has been recently underlined 30.
Taking into account such diagnostic corrections is crucial and drastically changes the prognostic value that TMB and MSIsensor score have in subsequent survival analyses, expectedly. For instance, we here show that the MSIsensor score displays a statistical tendency of predictive value for the resistance to ICI before correction of diagnostic errors in C1, whereas this same index has no longer any predictive value after correction (exclusion of the 3 false positive MSI cases) (Fig. 2B). Likewise, while a trend for a predictive value of TMB is observed before correction of the same diagnostic errors in C1, this is no longer the case at all thereafter (Fig. 2B). In view of these results, we decided not to investigate further the clinical relevance of TMB and MSIsensor score in C2 since none of them were associated with patients’prognosis in a true MSI mCRC population. As shown in SUPPLEMENTARY FIGURE S2, the MSICare score is no more clinically relevant for predicting response to ICI in patients with confirmed MSI mCRC.
No evidence for the relevance of somatic DNA variants predicting MSI mCRC patients’ iPFS to ICI
Next, we examined the impact of somatic variants occurring in true positive MSI mCRC patients from C1 regarding response to ICI. Expectedly, a great number of variant-containing genes occurred at both nonrepetitive (NR, in n = 3,886 genes, only coding events) or repetitive (R, in n = 20,472 genes, both coding or noncoding events) sequences, these genes having or not an expected role in the MSI-driven tumorigenic process (Fig. 2C). These somatic events, both coding and noncoding, were observed at variable frequencies in either repetitive (Fig. 2D) or nonrepetitive sequences (Fig. 2E) in tumor DNA, as previously reported 31. Overall, MSI-driven somatic events at microsatellites (i.e., mononucleotide repeats) were however mostly associated with the background for instability of MSI in colon tumor cells while a minority of these microsatellite mutations were likely positively or likely negatively selected, as previously reported in other large series of localized MSI colon tumors 31 (Fig. 2D).
In FIGURE 2F are shown the results of univariate cox analyses we performed to identify variants with significant impact on the iPFS of MSI mCRC patients from C1 and C2. These tests were first systematically performed for all detected exome-wide somatic variants in C1. Among them, only events associated with P-values ≤ 0.1 were sequenced again using targeted NGS in C2 (n = 362 events including 195 and 167 somatic variants at NR and R sequences; see also the study design of the study Fig. 1). This strategy led us to consider 9 and 22 somatic variants with a significant effect on patients' survival in both cohorts at alpha = 5% and 10%, respectively. With the exception of one, i.e., the microsatellite mutation affecting of the LncRNA TTN-AS1, none of these association to ICI-treated patients’ survival remained significant after FDR correction. A listing of these 22 variants is shown on SUPPLEMENTARY TABLE S3. Although we suspect that some of these events are clinically relevant, we concluded that their effect at this stage on ICI-treated MSI mCRC patient outcome remains unclear given the insufficient statistical power. Note that canonic mutations recurrently associated with colon tumor among which some were previously proposed to affect response to ICI, e.g., in KRAS, BRAF, B2M and other cancer-related genes, were infirmed as being relevant in C1 and not further investigated in C2 (SUPPLEMENTARY FIGURE S3).
Identification of an MSI-driven signature predictive of response to ICIs by monitoring somatic mutations on a selected set of microsatellite sequences by machine learning
Our results suggest that a number of somatic variants in microsatellite-containing genes may be potentially associated with ICI resistance in MSI mCRC, without any of them having a sufficient value alone. In light of these new insights, we hypothesized that response to treatment would be better predicted by considering these potentially influential events together in a single classification model.
From C1 (Training cohort), we trained a supervised binary classifier (Random Forest, RF) considering the relapse status as the dependent variables and the mutational status of the 19 microsatellites whose mutations were identified as being associated with response to treatment using univariate cox model as the independent variables (Fig. 3A). Applying the trained model in C2 (Validation cohort), we were able to show that the same model was significantly associated to iPFS (Fig. 3B). FIGURE 3 also illustrates the two groups of variants whose presence and absence in tumor DNA contribute to predicting resistance to ICI in MSI mCRC patients (see the heatmap on Fig. 3C). In addition, no prognostic impact of the same DNA predictor was detected when examined in the 5 cohorts of patients with MSI metastatic (PFS) or nonmetastatic (DFS, Disease-Free Survival) colorectal cancer untreated with ICI, thus validating the specificity of the predictive nature of this indicator regarding ICI-treated MSI mCRC patients (Fig. 3B).
Supervised RNAseq analysis fails to identify established phenotypic markers associated with response to ICI in MSI mCRC patients
We then sought to evaluate established general phenotypic and cellular marker of response to ICI. Three types of markers were applied to and systematically assessed in both discovery (C1) and validation (C2) cohorts, i.e., signatures quantifying cellular components of the tumor microenvironment, single gene expression levels, and pathway-level estimations of expression. TME (tumor microenvironment) components previously associated to differential response to ICI such as the infiltration levels of T or B lymphocytes, of the monocyte lineage or of fibroblasts could not be reproducibly associated to iPFS in the two ICI-treated MSI mCRC cohorts (Fig. 4A). Similarly, the level of expression of the two targeted immune-checkpoint PD-1, PD-L1 and CTLA-4 was not robustly associated to iPFS under immunotherapy. In addition to this focused approach, a systematic screen was applied. A total of 99 RNA signatures of cellular component of the TME were tested encompassing various immune and stromal phenotypes, none of which had a significant association with iPFS in either cohort (FDR 5%) (Fig. 4B and 4E and SUPPLEMENTARY TABLE S4). Similar results were obtained with the estimated expression activity of 3,365 pathways (Fig. 4C and 4F). More specifically, gene sets involved in angiogenesis, epithelial-to-mesenchymal transition, related TGF-beta and Wnt/Wingless signaling pathways, as well as TNF, interferon, KRAS or mTOR had either a minor and unreproduced association with iPFS in one cohort but more generally no significant correlation with survival in any of the two ICI-treated MSI mCRC cohorts (SUPPLEMENTARY FIGURE S4). Finally, the gene-level expression association (10,515 genes tested) identified 5 genes with significant association to iPFS in C2 with either no association with iPFS in C1 and often an opposite hazard-ratio (Fig. 4D and 4G). Overall, signatures previously established in generic contexts or associated to response to ICI in MSS tumors failed to show any signal in MSI mCRC.