Early Adaptation of Colorectal Cancer Cells to the Peritoneal Cavity Is Associated with Activation of “Stemness” Programs and Local Inflammation

In 2018, colorectal cancer was the third leading cause of cancer mortality according to the World Health Organization GLOBOCAN database. Peritoneal metastases are found in 5% to 10% of patients with colorectal cancer at primary surgery, in 4%–19% of patients during follow-up after curative surgery, and in 40%–80% of patients who die of colorectal cancer (1, 2). Patients with colorectal cancer peritoneal metastases (CRPM) have significantly shorter median overall survival (OS; 16.3 months) compared with those with liver (19.1 months) or lung (24.6 months) metastases (3).

Recently communicated data from the PRODIGE-7 study comparing cytoreductive surgery (CRS) plus hyperthermic intraoperative peritoneal chemotherapy (HIPEC) or CRS alone in addition to systemic chemotherapy, showed a median OS of 41.2 months (95% CI, 35.1–49.7) in the CRS (non-HIPEC) arm and 41.7 months [95% confidence interval (CI): 36.2–52.8] in the CRS+HIPEC arm (4). These findings underscored the fact that CRPMs [M1c colorectal cancer in the 8th edition of the American Joint Committee on Cancer TNM classification (5)] in the absence of organ-based metastases could describe a different patient population with a different molecular biology of the disease.

A stepwise model for the development of CRPMs has been proposed (6). First, tumor cells detach from the primary tumor and gain access to the peritoneal cavity. Second, these free tumor cells need to acquire the capacity to adhere to the peritoneal surface. Third, they need to penetrate the mesothelial monolayer using the underlying connective tissue as a scaffold for tumor proliferation (6).

The genomic characterization of CRPMs in a study of 2,563 patients with metastatic colorectal cancer with known BRAF mutation status indicated an association between BRAF status and CRPMs: the primary tumor BRAF mutation rate was 18% in patients with CRPM only and 12% in multiorgan including CRPMs (3). In a more recent report, no significant differences were found for KRAS, BRAF, or microsatellite instability when a nonmatched pool of primaries (n = 617) were compared with CRPMs (n = 348) with a similar rate of BRAF mutant in CRPM of about 12% (7).

To date, published data regarding the concordance of genomic alterations between CRPM and their matched primaries is scarce. In a meta-analysis to explore the concordance between primary colorectal cancers and their metastases regarding the main genes used in clinical decision making (e.g., KRAS, NRAS, and BRAF), we have demonstrated an overall discordance of 8% (95% CI, 5–10; ref. 8). However, data on synchronous peritoneal metastases are lacking.

This study presents a comprehensive analysis of 30 matched primary and peritoneal samples from patients with colorectal cancer who presented with synchronous CRPMs and underwent CRS and HIPEC. The initial hypothesis was that there would be no significant differences regarding the genomic alterations used to guide clinical management. However, a more profound analyses of gene expression was undertaken to identify potential mechanisms of synchronous CRPM development. Furthermore, we test the possibility of deriving a signature from these differences with prognostic capacity.

Paired samples (n = 30) from 15 consecutive patients with peritoneal only disease that underwent CRS/HIPEC for CRPM (2002–2017) with the primary in situ were identified at a UK national peritoneal tumor centre. The study was approved by the Manchester Cancer Research Centre (MCRC) BioBank institutional review board and ran under ethics approval of the South Manchester Research Ethics Committee (Reference: 18/NW/0092). Written consent was obtained for all patients that were alive at the time of the study and a specific approval from the Health Research Authority and Health and Care Research Wales Research Ethics committee was also obtained for the use of samples from deceased patients that were not able to provide written consent before the study initiation (Reference: 19/LO/1339). All clinical investigations were conducted according to the principles expressed in the Declaration of Helsinki and good clinical practice guidelines.

Data from 660 colorectal cancer primary samples with OS information and 743 samples with disease-free survival (DFS) information from publicly available series (9–12) with expression data for colorectal cancer were selected.

The genomic DNA was extracted from 4 formalin-fixed, paraffin-embedded (FFPE) 5 μmol/L unstained sections mounted on slides using the Roche cobas DNA Sample Preparation Kit. PCR-based target enrichment was performed. Following this, purification of PCR was done using AMPure XP beads from Agencourt and library preparation followed the custom protocol using Illumina TruSeq PCR Free indexes and reagents for indexing. Purified libraries were sequenced using Illumina sequencing by synthesis chemistry on a MiSeq using 2 × 150 bp v2 sequencing chemistry. Following demultiplexing, primer trimming and alignment to the human reference genome variants were called using a clinically validated custom algorithm. All samples have a coverage of >350×. List of genes sequenced in Supplementary Table S1.

Four 5 μmol/L FFPE sections mounted on unstained slides were marked up by a pathologist prior to macrodissection and then deparaffinized using Qiagen deparaffination solution. Following RNA extraction using the RNeasy FFPE Kit from Qiagen RNA abundance was measured using a Qubit fluorometer. The analysis was done using the NanoString nCounter mRNA Gene Expression system.

The nCounter PanCancer IO 360 Panel includes 750 genes covering key pathways at the interface of the tumor, tumor microenvironment (TME), and immune response; in addition, there are internal reference genes for data normalization and 30 customed probes were added to the nCounter CodeSet, all of them shown in Supplementary Table S2.

A tissue microarray (TMA) containing 0.6 mm cores from primaries and peritoneal metastases was designed. Each sample was represented by three cores, active margin (AT), tumor core (TC), and “normal” (surrounding tissue) core. For the purpose of this report, only AT and TC cores were analyzed with the GeoMx digital spatial profiling (DSP) system.

The Human Immune Cell Profiling Core Nanostring's DSP technology combines standard immunofluorescence techniques with digital optical barcoding technology to perform highly multiplexed, spatially resolved profiling experiments.

An automated setup capable of imaging and sample collection was developed by modifying a standard microscope. For protein detection, a multiplexed cocktail of 56 primary antibodies, each with a unique, UV photocleavable indexing oligo, and three fluorescent markers (pan-cytokeratin antibody, CD45 antibody, CD68 antibody, and nucleic acid staining with Syto 13) was applied to a slide-mounted FFPE tissue section of the TMA.

The regions of interest (ROI) were then selected on the basis of the fluorescence information and sequentially processed by the microscope automation. Once all ROIs were processed, indexing oligos were hybridized to NanoString optical barcodes for ex situ digital counting and subsequently analyzed with an nCounter Analysis System.

The plots related to the genomic data were performed using R with GenVisR (13). The expression heatmap and volcanos plots were drawn using the nCounter Advanced Analysis Software 2.0. Volcano plots display each gene's −log₁₀ (P) and log₂ fold change with the selected covariate. Horizontal lines indicate various Benjamini–Yekutieli FDR-adjusted P value thresholds. Highly statistically significant genes fall at the top of the plot above the horizontal lines, and highly differentially expressed genes fall to either side. Genes are colored if the resulting P value is below the given FDR. The 40 most statistically significant genes are labeled in each the plot. The loglinear model is applied to perform a multivariable analysis to adjust the gene expression of the hits found when comparing CRPMs versus their matched primaries to relevant clinical variables. The heatmap for the 20 probes with the lowest adjusted P value was drawn by using the normalized data of gene expression from the previous analysis and then use the Morpheus software (https://software.broadinstitute.org/morpheus). The consensus molecular subtypes (CMS) for each sample was assigned by using the CMSclassifier R package (14). The canonical pathway, the network, and functional analyses were developed with the Qiagen Ingenuity Pathway Analysis (IPA). The nonparametric Mann–Whitney U test was used for the DSP expression analysis with IBM SPSS Statistics for Windows, version 19 (IBM Corp.). All the violin plots are developed with R package ggplot2. Benjamini–Hochberg FDR was used to correct for multiple comparisons. Kaplan–Meier curves from The Cancer Genome Atlas (TCGA) data are drawn using StataCorp. 2011. Stata Statistical Software: Release 12. StataCorp LP.

Transcripts in Supplementary Tables S3 and S4 present the most statistically significantly differentially expressed genes with the selected covariates. “Estimated log fold change” estimates a gene's differential expression. For categorical covariates, a gene is estimated to have 2ˆ(log fold change) times its expression in baseline samples, holding all other variables in the analysis constant. If the covariate is continuous, for each unit increase in the selected covariate, a gene's expression is estimated to increase by 2ˆ(log fold change)-fold, holding all other variables in the analysis constant. The 95% CI for the log fold change is also presented, along with a P value and FDR. Method column indicates the model used to estimate differential expression. Mixed negative binomial model uses the mle function to run the Wald test. The simplified negative binomial model uses the glm.nb function. The loglinear model uses the lm function. The loglinear model is applied to perform a multivariable analysis to adjust the gene expression of the hits found when comparing Peritoneal metastases versus their matched primaries. Plots for differential expression of each of the 20 lower adjusted P-value probes were performed using GraphPad Prism version 8 for Windows, GraphPad Software.

Data related to all components of the 20-gene signature expression and survival data were downloaded from TCGA (9) and Gene Expression Omnibus (GEO) databases GSE17536 (10), GSE17537 (11), and GSE14333 (12). A cox-regression model including all the 20 genes was built and applied to evaluate OS and DFS. The relative HR for each case was calculated and cases were classified within each database as high or low risk according to their relative risk being above or below the median relative risk. Then, all the cases with information for OS and DFS were pooled together and the Kaplan–Meier analysis and log-rank test was performed using the survminer R package. A multivariable Cox regression analysis was performed adjusting for age, gender, and American Joint Committee on Cancer stage. Cox proportional hazards assumptions of proportionality were checked and verified using the Schoenfeld residual test in each database and in the overall pooled data, using StataCorp. 2011. Stata Statistical Software: Release 12. StataCorp LP.

For a complete description of methodology, please, see the online Supplementary Methods section.

Sequencing data are available in Sequence Read Archive (SRA) database through the Bioproject accession number: PRJNA675983. Nanostring nCounter expression data are available in GEO database with accession number GSE161097. DSP data are available as Supplementary Table S7.

Baseline characteristics are summarized in Supplementary Table S5. Median follow-up was 25.3 months (interquartile range: 23.3–37.7) with an estimated median OS of 32 months (95% CI, 20.6–43.7). Two patients (13.3%) were mismatched-repair deficient, one case (6.6%) was BRAF mutated and the mucinous phenotype was reported for 7 patients (46.7%).

We observed 100% concordance between primary and peritoneal tumors for gene mutations in KRAS, NRAS, BRAF, MSH2, and MSH6 (Supplementary Fig. S1).

A total of 134 gene variants were identified with a distribution across common drivers of colorectal cancer showing, mutated TP53 in 27.6%, APC in 26.1%, and KRAS in 11.9% of the total of variants detected. Median overall variant allele frequency (VAF) was 22%, ranging from 1% to 75% (Fig. 1A). The percentages plot in Fig. 1A account for gene mutations divided by the total of samples (n = 30).

Initially, discordance was defined as the presence of a given mutation in just one of the paired samples. When all the genes in our next-generation sequencing (NGS) panel were considered, a discordance of 53.3% was identified (8/15 cases). Expanding our discordance definition to classify any VAF difference > 30% for a given mutation as discordant resulted in all but two cases being discordant (B and M; 86.7%; Fig. 1A and B). These data are consistent with the evolutionary landscape of colorectal cancer (15).

To explore for any global discordance between primary and CRPMs, using transcriptomic data, we determined the CMS for each sample (14). All the primary and 12 peritoneal samples were classified as CMS2 (canonical), two of the peritoneal samples were classified as CMS3 (metabolic) and one as CMS4 (mesenchymal) giving an overall concordance of 80% for CMS subtypes as shown in Fig. 1C.

The analysis using our 800-probe customized PanCancerIO360° Nanostring panel revealed more profound differences at the mRNA level between primary and matched CRPM samples. In Fig. 1D samples are displayed after unsupervised hierarchical clustering with the “oncoplot” of mutational analyses shown above in Fig. 1A. Only four pairs clustered together, A, B, M, and J. Eight primaries (53.3%) clustered together but separately from their corresponding peritoneal samples.

To better understand the differences between primaries and metastases, we used volcano plots representing adjusted P values of the differential gene expression (DGE) analyses (Fig. 2; Supplementary Fig. S2). After FDR correction, we focused our downstream analyses on 20 genes with an mRNA differential expression significant at an adjusted P value of 0.05.

Figure 2A represents the volcano plot of the 800 probes DGE by their origin (peritoneal metastases vs. primary tumor). Figure 2B–E show the DGE by baseline characteristics of interest due to their potential impact on skewing the primary analysis (peritoneal metastases vs. primary tumor). including administration of prior chemotherapy (Fig. 2B), mucinous histology (Fig. 2C), side of primary tumor (Fig. 2D), and lymph node spread (Fig. 2E). In Supplementary Fig. S2, we analyzed the transcriptomic data based on common genomic alterations. Samples were divided by mutant TP53, APC, and KRAS alone or a combination including at least one of them. No targets at FDR P < 0.05 were found. Thus, we tested for the impact of the clinical parameters on our analyses by performing a multivariable analysis adjusted to the above-mentioned clinical factors (Fig. 2F).

Figure 3 shows the plots of paired analysis of these 20 genes to show the trend of the differential expression for each one of them using Student t test.

Figure 4A shows the 20-gene signature derived from our DGE analysis between CRPMs and paired primary tumors. We found 15 mRNA probes overexpressed in CRPMs and five mRNA probes overexpressed in primary tumors. Twenty genes with the lower FDR-adjusted P value regarding differential expression between peritoneal metastases versus primaries are shown in Fig. 4B.

A canonical pathway analysis using all the genes identified with a raw P ≤ 0.05 (Fig. 4C) showed activation of dendritic cell maturation and neuroinflammation signaling pathways. Functional analysis of the 20 hits described in our CRPM signature is shown in Fig. 4D. This demonstrates that compared with paired primary tumors, CRPMs have higher metastatic and invasive potential (rows 2, 3, 6, and 7; Fig. 4D), reduced cell death (rows 1 and 22; Fig. 4D) and reduced leukocyte and myeloid migration (rows 21 and 23; Fig. 4D). Furthermore, network analysis of these 20 genes (Fig. 4E) reveals not only the central role of IL6 in CRPMs but also a relevant axis linked to neuroinflammation including NGFR, NFATC2, and NCAM.

To assess the distribution of different immune-related proteins and use a different technology, we used a novel DSP technology applied to the TMA with four paired cases. Each case was represented by four cores (CRPM vs. primary; CT vs. AM). The immunofluorescence images in Fig. 5A show an example of all the cores for a single case. We were able to analyze all the cores separately based on their pan-cytokeratin antibody staining (in green), that is, that for each core we have expression of the different antibodies for tumor-enriched areas (or masks) and TME-enriched areas (full list of antibodies in Supplementary Table S6). The plot of Fig. 5B shows multiple violin plots of the different protein expression profiles in the different areas (CT vs. AM cores and TME vs. tumor-enriched areas). DSP normalized data are shown in Supplementary Table S7. In Fig. 5C, the P values of the significant differentially expressed proteins are displayed after FDR set at 0.15 (raw and corrected P values are displayed in Supplementary Table S8). The DSP findings show a differential expression of Arginase 1 (Arg1) in the TME mask-favoring overexpression in CRPMs. This finding agrees with IL6 being the most relevant pathway in CRPMs linking to an increase in type-2 tumor-associated macrophages (TAM2), as Arg1 is a marker of this particular cell type.

Because of the lack of large series with well-characterized transcriptomic data of CRPMs, to better understand the possible clinical utility of our gene signature, we analyzed the prognostic capacity of each gene using TCGA database in Fig. 6 and Supplementary Fig. S3. In Fig. 6A–D, we can see Kaplan–Meier curves for OS dividing TCGA cohort (n = 440) by the median expression of four representative genes from the panel including nerve growth factor receptor (NGFR; P = 0.021; Fig. 6A).

To further explore the clinical implication of our results, we hypothesized that our gene signature, associated with synchronous CRPMs, could be an early event present in primary lesions with detrimental prognosis impact. Initially, we tested the 20-gene model on TCGA data (9) for OS showing a nonadjusted HR: 4.44 (95% CI, 2.70–7.30; P < 0.0001). We used the online tool PROGgeneV2 (16) with the aim of finding publicly accessible databases with survival and gene expression data for the 20-gene signature. Thus, we selected the colon adenocarcinoma data from GSE17536 (10), GSE17537 (11), and GSE14333 (12). We pooled together the data from all databases for OS (n = 660) and for DFS (n = 743) analysis (the distribution of cases and events are displayed in Supplementary Fig. S4). Figure 6E and F show the Kaplan–Meier curves for the high-risk and low-risk groups related to OS and DFS. Both median OS and DFS were shorter in patients with high-risk [54.6 months (95% CI, 44.3–76.6) and 85.8 months (95% CI, 74.6–not calculated), respectively] compared with low-risk group, not reached in both settings. The HR adjusted for age, stage, and gender of the high- versus low-risk groups was 2.32 (95% CI, 1.69–3.19; P < 0.0001) for OS and 2.08 (95% CI, 1.50–2.91; P < 0.0001) for DFS.

The molecular complexity and the difficulty in accessing peritoneal samples has made research in this area challenging. In this communication, we report an in-depth characterization of synchronous CRPMs when compared with their matched primary colorectal tumors.

Stein and colleagues recently reported, in an unmatched series comparing primaries versus CRPMs, the genes involved in clinical decisions (RAS family and BRAF) showed similar mutational rates in primaries and CRPMs (7). Similarly, our data support the prediction that for validated and robust biomarkers used in the clinical setting for colon cancer, analysis of either the primary or peritoneal samples will suffice. In agreement with our findings (BRAF: 6.6%), Stein and colleagues also detected a lower than expected BRAF mutation rate of approximately 10%. However, we were able to show a larger number of genomic discrepancies between primary and CRPMs using an expanded NGS panel. Counterintuitively, the pattern of discrepancies was not related to the mucinous subtype or the CMS categories. Our samples were predominantly classified as CMS2 which was a relatively unexpected finding in view of a recent Dutch series which reported CMS4 in 60% of the primaries and 75% of the CRPMs (17). Nevertheless, APC mutations are known to drive CMS2 and CMS3. The discrepancy with the Dutch study could be partially explained because the authors looked for CMS4 with a specific assay and selecting tumor samples with a minimal threshold of 10% of tumor content. However, our study used macrodissection to enrich the tumor content. Heterogeneity could, therefore, account for the difference between the two studies. Using a limited panel to define CMS categories could have affected the accuracy of the CMSclassifier. However, other groups have proven that this can be done with even smaller gene panels (18, 19).

The more diverse transcriptomic landscape (Fig. 1D) was translated into a 20-gene peritoneal signature. Our peritoneal metastasis signature revealed expected suspects as drivers of metastases in cancer, such us, CD36 (20) and IL6 (21). Inflammation, metabolic, migration, and invasion pathways represent the pathways related to our 20-gene profile. Even though, IL6 sat at the centre of the network analyses, one novel finding relates to neuroinflammation and neuronal stemness pathways being highly expressed in CRPMs. Nevertheless, primary tumor gene expression is represented by genes involved in extracellular matrix remodeling, for example, MMP1 and CLCA4. Although the pathways and genes represented in the signature are well known, this is to our knowledge the first time that they have been presented together as a signature.

The strong IL6 signal in our data is associated with an increase of TAM2 markers that was observed in the DSP experiment. Mainly, a significant overexpression of Arg1 in the TME of the CRPMs when compared with their matched primaries. In Supplementary Fig. S5, cell subtyping analysis of the transcriptomic data also showed a clear enrichment in all microenvironment cells in CRPM, including macrophages. This clear dominance of macrophage inflammation signals has also been shown experimentally in models of ovarian cancer peritoneal metastasis, in which, ablation of macrophages impacted tumor growth (22). Interestingly, our canonical pathway analyses showed high-mobility group box 1 gene (HMGB1) pathway as being the one with highest Z-score when comparing the metastases with primaries. HMGB1 is a prototypic damage-associated molecular pattern/alarmin known to be released by colon cancer cells. This pathway is involved in macrophage infiltration, tumor growth and vascularization (23). Our findings support the hypothesis that involvement of these pathways could be common to other malignancies with a tendency to colonize the peritoneum.

We also found a significant overexpression of CD56 (NCAM) as marker of natural killer (NK) cells that was enriched in the tumor core areas of CRPMs (P = 0.028). However, the experimental ablation of NK cells in Robinson–Smith model seemed to lack a similar degree of tumor growth control when compared with macrophage ablation (22), potentially, ascribing a minor role to NKs compared with macrophages.

CD36, also known as collagen type I receptor or thrombospondin receptor, links lipid metabolism with invasiveness and metastatic potential (20). Therefore, it was not unexpected that thrombospondin (THB) was also upregulated in CRPMs. Furthermore, this axis has been reported to confer a survival disadvantage in colon cancer (24).

NGFR (also known as CD271, TNFSFR16, or P75NTR) is a 75 kD single-transmembrane protein without kinase activity and widely expressed in the central and peripheral nervous system (25). Overexpression of NGFR observed in many metastatic cancers promotes tumor migration and invasion in relation to cancer stem cells (26). In addition, it is linked to the IL6 inflammatory pathway through the upregulation of neurotrophins (27). However, there is conflicting data regarding colorectal cancer, NGFR was initially considered a possible tumor suppressor but recent data showed that NGFR could act as an oncogenic inhibitor of p53 (28). Our NGFR results were validated in TCGA database (Fig. 6) showing that a high expression of NGFR was linked to shorter survival, supporting the hypothesis of this pathway being linked to survival advantage of peritoneal metastatic cells. NGFR findings are a likely link between inflammation and “stemness” due to its relevant role in both pathways (29).

Our findings have implications for the management of CRPMs in relation to the use of immunomodulatory agents suggesting a potential role for drugs targeting the IL6 axis, well-described “stemness” pathways (e.g., CD36) and less studied ones (e.g., NGFR). Tocilizumab is a recombinant humanized anti-IL6 receptor (IL6R) mAb (30) that could have a role in the management of CRPMs. The attempt to develop anti-CD36 antibodies for metastatic disease has not been successful yet but the axis CD36/THB remains a good target to treat metastasis (31).

Our study suggests a new model for CRPM development (Supplementary Fig. S6). We propose that in early stages of CRPMs, a predominant pattern of chronic inflammation represented by a TME rich in TAM2 with upregulation of IL6 and rewiring of signal pathways linked to plasticity, “stemness” and metabolism, such as, NGFR and CD36, is observed. In particular, our results showed a shift in expression from components of the Wnt canonical pathway (such as, Wnt5a; ref. 32) and genes related to SUMOylation (ref. 33; process critical for cancer stem cells), such as PIAS4, that are enriched in the primaries, to CD36/THB axis and genes involved in tumor-related mesenchymal stem cells, such as, OLFML2B (34) and SNAI1 (35), master regulator of epithelial/mesenchymal transition in CRPMs.

The study has several limitations. First is its small sample size in relation to the paired samples that constituted the discovery set of the study. Despite this, our study represents the largest sample size of its kind analyzed to this extent in the literature to date. There was only one rectal case, therefore, no specific remarks could be derived. Second is that our study was not powered to assess phylogenesis and clonal evolution. Our reported VAFs are not bioinformatically corrected for cellularity due to the use of a rather small NGS-targeted panel. Tumor mutational burden (TMB) was not calculated because of the use of a small targeted approach. However, an approximation to TMB is shown in Fig. 1A. Third is a panel of probes with a specific focus on cancer-related immune and inflammation pathways that could bias the results toward this particular angle. Fourth is the use of FFPE tissue with well-known difficulties to obtain good-quality RNA. To overcome this, we selected a panel of technologies with proved robustness in this setting. Nevertheless, our study design could not asses if the inflammatory changes were present in the distant normal tissue. To reduce the impact of a small sample size of our study, we performed an extensive in silico study of our 20-gene signature in different databases to build a validation set including colorectal cancer transcriptomic data from databases that included assays for all the components of the signature. We also performed a multivariate analysis, although, limited to age, stage, and grade because these were the only features common to all 743 cases analyzed. This approach indicated to us that the activation of pathways included in the signature might have occurred at a relatively early stage of tumorigenesis. It also showed that our peritoneal-derived gene signature can identify subgroups of patients with worse prognosis and at higher risk of relapsing from samples taken at the initial diagnosis. Setting aside the prognostic use of our peritoneal signature, our results identified NGFR as a new emerging target for CRPM management.

In summary, we present data relating to synchronous CRPM showing early chronic inflammation events in conjunction with rewiring of “stemness” pathways with potential for the development of new treatment strategies. We also derived a peritoneal 20-gene signature with prognostic capacity in nonmetastatic colorectal cancer from the study of synchronous peritoneal metastases that could be a good candidate for prospective validation.

J. Barriuso reports grants from Cancer Research UK and The Christie Charity during the conduct of the study, as well as grants, personal fees, and nonfinancial support from Ipsen; personal fees and nonfinancial support from Pfizer and Novartis; nonfinancial support from AAA, NanoString, and Roche; grants and personal fees from Servier; and personal fees from Nutricia outside the submitted work. A.J. Wallace reports personal fees from Roche and AstraZeneca outside the submitted work. C. Dive reports grants and personal fees from AstraZeneca and Merck AG; grants from Astex Pharmaceuticals, Bioven, Amgen, Carrick Therapeutics, Taiho Oncology, GSK, Bayer, Boehringer Ingelheim, Roche, BMS, Novartis, Celgene, Epigene Therapeutics Inc, Angle PLC, and Menarini; and other from Clearbridge Biomedics and Thermo Fisher Scientific outside the submitted work. D.G. Evans reports personal fees from AstraZeneca and Springworks during the conduct of the study. M.P. Saunders reports personal fees from Servier, Merck, and Amgen outside the submitted work. No disclosures were reported by the other authors.

J. Barriuso: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing-original draft, project administration, writing-review and editing. R.T. Nagaraju: Software, formal analysis, investigation, visualization, methodology, writing-original draft, writing-review and editing. S. Belgamwar: Data curation, validation, investigation, writing-review and editing. B. Chakrabarty: Resources, formal analysis, supervision, investigation, writing-review and editing. G.J. Burghel: Software, formal analysis, methodology, writing-original draft, writing-review and editing. H. Schlecht: Software, formal analysis, investigation, writing-review and editing. L. Foster: Formal analysis, supervision, visualization, methodology, writing-review and editing. E. Kilgour: Resources, supervision, project administration, writing-review and editing. A.J. Wallace: Conceptualization, resources, software, formal analysis, supervision, writing-original draft, writing-review and editing. M. Braun: Resources, formal analysis, supervision, funding acquisition, writing-original draft, writing-review and editing. C. Dive: Resources, supervision, methodology, writing-review and editing. D.G. Evans: Conceptualization, resources, supervision, writing-original draft, writing-review and editing. R.G. Bristow: Resources, supervision, funding acquisition, writing-original draft, writing-review and editing. M.P. Saunders: Conceptualization, resources, supervision, funding acquisition, writing-original draft, writing-review and editing. S.T. O'Dwyer: Conceptualization, resources, supervision, funding acquisition, writing-original draft, writing-review and editing. O. Aziz: Conceptualization, resources, formal analysis, supervision, funding acquisition, validation, investigation, methodology, writing-original draft, writing-review and editing.

The results published here are in part based upon data generated by TCGA Research Network: https://www.cancer.gov/tcga. D.G. Evans is supported by the National Institute for Health Research (NIHR) BRC Manchester (Grant Reference Number 1215–200074). R.T. Nagaraju is supported by Cancer Research UK Accelarator Award (C64263/A29365). J. Barriuso, S.T. O'Dwyer, and O. Aziz receive funding from the Cancer Research UK Accelarator Award (C64263/A29365). J. Barriuso is partially funded by The Christie Charity. The work is partially funded by the Steve Prescott Foundation Charity.

The authors thank the Manchester Cancer Research Centre (MCRC) BioBank, Histology of the CRUK Manchester Institute, and NanoString for their support.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.