A Low-Cost Multiplex Biomarker Assay Stratifies Colorectal Cancer Patient Samples into Clinically-Relevant Subtypes

Objective In order to personalise standard therapies based on molecular profiles, we previously classified colorectal cancers (CRCs) into five distinct subtypes (CRCAssigner) and later into four consensus molecular subtypes (CMS) with different prognoses and treatment responses. For clinical application, here we developed a low-cost multiplex biomarker assay. Design Three cohorts of untreated fresh frozen CRC samples (n=57) predominantly from primary tumours and profiled by microarray/RNA-Seq were analysed. A reduced 38-gene panel (CRCAssigner-38) was selected from the published 786-gene CRCAssigner signature (CRCAssigner-786) using an in-house gene selection approach. A customised NanoString Technologies’ nCounter platform-based assay (NanoCRCAssigner) was developed for comparison with different classifiers (CMS subtypes), platforms (microarrays and RNA-Seq), and gene sets (CRCAssigner-38 and CRCAssigner-786). Results NanoCRCAssigner classified samples (n=48; except those showing a mixture of subtypes) into all five CRCAssigner subtypes with overall high concordance across platforms (>87%) and with CMS subtypes (81%) irrespective of variable tumour cellularity. The association of subtypes with their known molecular (microsatellite-instable and stemness), mutational (KRAS/BRAF), and clinical characteristics (including overall survival) further demonstrated assay validity. To reduce costs, we switched from the standard protocol to a low-cost protocol with a high Pearson correlation co-efficient (0.9) between protocols. Technical replicates were highly correlated (0.98). Conclusion Here we developed a low-cost and potentially clinically deployable NanoCRCAssigner assay to facilitate prospective validation of (CRCAssigner and potentially CMS) subtypes in clinical trials and beyond. Summary “box” What is already known about this subject? Colorectal cancer (CRC) is a heterogeneous disease. We previously identified 5 gene expression-based CRC subtypes (CRCAssigner; enterocyte, goblet-like, inflammatory, stem-like and transit amplifying) using a 786- gene signature (CRCAssigner-786) later reconciled into the 4 Consensus Molecular Subtypes (CMS1-4). These subtypes were identified by profiling samples using microarray and RNA-Seq platforms, which are expensive, time-consuming and impractical for routine clinical use. CRCAssigner subtypes have prognostic and potential predictive differences (to anti- EGFR and FOLFIRI; a combination of irinotecan, 5-fluorouracil and leucovorin). Previous analysis of randomised clinical trials assessing patient responses to oxaliplatin in addition to fluouracil-leucovorin indicated that CRCAssigner subtypes may predict responders (compared to CMS) using only the discovery cohort, but larger cohorts are warranted to validate this finding. Subtype-driven clinical trials require a validated low-cost assay suitable for routine clinical use. What are the new findings? A reduced 38-gene signature (CRCAssigner-38) from CRCAssigner-786 gene set can be utilised to classify samples into the CRCAssigner subtypes with minimal misclassification error rate. The CRCAssigner subtypes can be assessed in fresh frozen samples using a customised CRCAssigner-38 signature-based assay (NanoCRCAssigner) by applying nCounter platform (NanoString Technologies) which is a cost-effective method and provides highly reproducible results. Subtype prediction with NanoCRCAssigner was highly concordant with subtypes predicted using the CMS classifier on microarray or RNA-Seq platforms. NanoCRCAssigner assay is potentially independent of tumour cellularity, predicts patient prognosis and is consistent with mutational and molecular profiles of CRC subtypes. How might it impact on clinical practice in the foreseeable future? This study demonstrates how molecular CRCAssigner (for the first time) and CMS subtypes can be detected using a biomarker assay (NanoCRCAssigner) suitable for clinical validation. With further modification of the protocol to analyse formalin-fixed paraffin embedded (FFPE) samples (not within the focus of the current manuscript), this assay may facilitate patient stratification within clinical trials and the prospective assessment of potential subtype-specific treatments in the future using biopsy or surgical samples.


Summary "box"
What is already known about this subject?
• Colorectal cancer (CRC) is a heterogeneous disease.
• These subtypes were identified by profiling samples using microarray and RNA-Seq platforms, which are expensive, time-consuming and impractical for routine clinical use.
• CRCAssigner subtypes have prognostic and potential predictive differences (to anti-EGFR and FOLFIRI; a combination of irinotecan, 5-fluorouracil and leucovorin).
• Previous analysis of randomised clinical trials assessing patient responses to oxaliplatin in addition to fluouracil-leucovorin indicated that CRCAssigner subtypes may predict responders (compared to CMS) using only the discovery cohort, but larger cohorts are warranted to validate this finding.
• Subtype-driven clinical trials require a validated low-cost assay suitable for routine clinical use.
What are the new findings?
• A reduced 38-gene signature (CRCAssigner-38) from CRCAssigner-786 gene set can be utilised to classify samples into the CRCAssigner subtypes with minimal misclassification error rate.
• The CRCAssigner subtypes can be assessed in fresh frozen samples using a customised CRCAssigner-38 signature-based assay (NanoCRCAssigner) by applying nCounter platform (NanoString Technologies) which is a cost-effective method and provides highly reproducible results.
• Subtype prediction with NanoCRCAssigner was highly concordant with subtypes predicted using the CMS classifier on microarray or RNA-Seq platforms.
• NanoCRCAssigner assay is potentially independent of tumour cellularity, predicts patient prognosis and is consistent with mutational and molecular profiles of CRC subtypes.
How might it impact on clinical practice in the foreseeable future?
• This study demonstrates how molecular CRCAssigner (for the first time) and CMS subtypes can be detected using a biomarker assay (NanoCRCAssigner) suitable for clinical validation.
• With further modification of the protocol to analyse formalin-fixed paraffin embedded (FFPE) samples (not within the focus of the current manuscript), this assay may facilitate patient stratification within clinical trials and the prospective assessment of potential subtype-specific treatments in the future using biopsy or surgical samples.

Introduction
Colorectal cancer (CRC) is the fourth leading cause of cancer-related deaths worldwide [1].
The median overall survival of metastatic (m)CRC patients with unresectable disease remains in the order of 24 months with standard chemotherapy options. The implementation of targeted therapies including anti-EGFR monoclonal antibodies, anti-angiogenic and more recently immunotherapy agents may extend the survival to up to 30 months in selected mCRC patients [2]. However, how to identify those patients who will benefit from different systemic drug options remains challenging. Additional predictive biomarkers are required to spare patients from unnecessary toxicities, improve patients' outcomes and increase costeffectiveness of treatment.
In order to classify colorectal cancers into subgroups with distinct biology and to effectively match existing therapies to facilitate subtype-specific therapeutic development, we previously identified five distinctive gene expression subtypes and an associated 786-gene signature (CRCAssigner-786) [3]. Based on the gene expression similarities with different cell types of the normal colonic mucosa, we named the subtypes as goblet-like, enterocyte, stem-like, inflammatory and transit-amplifying (TA). We demonstrated significantly poorer disease-free survival (DFS) in untreated patients for the stem-like subtype, intermediate DFS for inflammatory and enterocyte and better DFS for goblet-like and TA [3]. Then, from two different datasets that included drug response information, we observed increased responses within the stem-like subtype to irinotecan, fluorouracil and leucovorin treatment combination (FOLFIRI) [4] and the TA subtypes to anti-EGFR monoclonal antibody (cetuximab) [5].
These treatment responses were further validated by other studies [6,7].
Recently, the exploratory clinical applicability of our CRCAssigner-786 subtypes was demonstrated when secondary analysis of a randomised clinical trial assessing patient benefit from the addition of oxaliplatin to fluorouracil-leucovorin in early-stage disease revealed that benefits were highly enriched in the enterocyte subtype compared to the other subtypes in the discovery cohort. Although this finding did not reach the same level of significance in the validation cohort, the trend was identical [14]. Other subtype classifications, including CMS [13] were not able to isolate patients who may potentially benefit from adjuvant oxaliplatin [14]. Notably, while the enterocyte and TA subtypes are merged into CMS2 of the consensus classification, only the enterocyte subset of CMS2 was observed to benefit from adjuvant oxaliplatin. It is, of course, plausible that individual classification systems may demonstrate specific therapeutic and prognostic relevance beyond CMS. Therefore, in this study we have applied both our CRCAssigner and CMS subtype classifications for assay development and comparison.
Although biologically appealing, translating these findings into routine clinical practice remains challenging. This is mainly because of the lack of a fit-for-purpose assay, able to classify patient samples into subtypes in a timely fashion to maintain a clinically relevant turnaround time, with reasonable costs and from the commonly available tissue samples. The majority of the previous classifiers were developed from microarray/RNAseq gene expression profiles. Technologies such as microarrays and RNA-Seq are expensive and time consuming, require dedicated bioinformatics expertise, and have total turnaround times that are too long to be clinically applicable. Also, they rely on pre-amplification of RNA, with consequent impact on accuracy and reproducibility. We previously demonstrated proof-ofconcept assays using immunohistochemistry and quantitative reverse transcriptase polymerase chain reaction (qRT-PCR) methods [3]. Nevertheless, these methods may suffer from reproducibility issues. Hence, we applied nCounter platform (NanoString Technologies) to develop clinically-relevant biomarker assay for CRC subtype classification.
The nCounter platform has previously been exploited to develop the Food and Drug Administration (FDA)-approved Prosigna® Breast Cancer Prognostic Gene Signature Assay [15] to predict risk of recurrence in patients treated with adjuvant hormonal therapy, as well as assays to predict medulloblastoma [16] and lymphoma [17] subtypes. This platform measures gene expression in the form of discrete counts of barcoded mRNAs, and requires no amplification step, eliminating a potential source of bias. In the present study, we evaluated the suitability of this technology as a platform for a gene expression-based assay for our CRCAssigner subtypes (for the first time) using a low-cost protocol to subtype CRCs in three different cohorts with different clinical and mutational characteristics. The results were then compared to the CMS subtype classifications and other platforms. A summary of the classifiers utilised in this study is given in Table 1.  [3] and a subset of TA tumours (cetuximab) [3], and suggested in the enterocyte subtype (oxaliplatin) [14].

Patient cohorts
Three different cohorts were studied. The first included 17 metastatic (stage IV) CRC samples (Montpellier cohort) from patients enrolled in the single centre REGP prospective study at the Institut du Cancer de Montpellier, which was part of a previously published study [4]. These patients had histopathologically confirmed colon adenocarcinoma and no prior chemotherapy.  The Probe A and TagSets were first mixed to create a partial master mix followed by Probe B to create a complete master mix, all at room temperature.
For both standard and low-cost protocols, 100 ng of total mRNA from fresh frozen tumour tissues was diluted with RNase-free water at 20 ng/uL.  Figure 1a. For the PanCancer Progression Panel (NanoString Technologies), the same protocol as for standard protocol was used.

Processing and quality control of nCounter data
Data quality was checked using nSolver analysis software v3.0 (NanoString Technologies).
Firstly, counts were corrected to background noise using geometric means of 8 negative control probes followed by the correction using geometric means of 6 internal positive control spike-ins in each lane/sample to correct potential sources of variations across the samples. These negative and positive probes were built-in in both standard and low-cost protocols. Only those housekeeping genes with raw molecular counts greater than 50 and those selected by geNorm algorithm (part of the nSolver analysis software) were retained for further analysis. Variations due to RNA input volume were corrected by normalising to the expression of geNorm selected housekeeping genes. The normalised final count data were log 2 transformed for further analysis. Those genes having zeros in more than 80% of the samples (representing potential technical error) were removed. Data generated from PanCancer Progression were also analysed for the biological pathways using the nCounter Advanced Analysis plugin (v1.0.84) for the nSolver analysis software.

Comparison of standard and low-cost protocols
For comparison of standard and low-cost protocols, housekeeping genes that were common between these two datasets were used for normalisation. Data from standard and low-cost protocols were row (gene)-median centred across samples separately, before being combined to perform hierarchical clustering.

Gene Selection for nCounter assay
The following genes were initially selected for inclusion in the CRCAssigner subtype custom nCounter assay (NanoCRCAssigner) based on our previous report [3]:

Assigning subtypes to samples
CRCAssigner subtypes were assigned by already published single-sample prediction (SSP) tool [13] by performing Pearson correlation of gene-wise median-centred expression profiles for each sample with centroids for the subtypes. The subtype with the highest correlation was then assigned to that sample. Samples were marked as having "undetermined" subtype if the sample's correlation with the subtype centroid was correlation co-efficient (R 2 ) ≤ 0.15, or labelled as "mixed" if the correlation was high for multiple subtype centroids (R 2 difference between first and second highest R 2 subtypes ≤ 0.06), in line with the published CMS classifier [13].
CMS subtypes were determined from microarray or RNA-Seq data using the CMSclassifier

Subtype concordance and significance
Subtype concordance between two different platforms was calculated as the percentage of samples that showed same subtype (not including mixed and undertermined samples) in both.
Subtypes were deemed to be concordant between the CRCAssigner and the CMS subtypes based on the following equivalence: CMS1 = Inflammatory; CMS2 = Enterocyte and TA; CMS3 = Goblet-like; CMS4 = Stem-like [13]. Fisher's exact test between different platforms or classifiers was performed to assess statistical significance of their subtype concordance.

See Supplementary
Methods & Materials for more information.

Development of nCounter assay for CRC subtyping.
In order to evaluate the applicability of the CRCAssigner and CMS subtype classifications in the clinic, we initially developed a custom nCounter assay using a 50-gene panel, including 47 genes selected from the CRCAssigner-786 signature and an additional 3 genes [3] (Table   S1; see Materials & Methods). Initially, we applied a standard protocol (in which biotin labels and molecular barcodes are directly attached to the mRNA probes; Figure 1a) from the manufacturer and tested the performance of the custom nCounter assay using primary tumour RNA obtained from 22 histopathologically confirmed colorectal cancers from two different cohorts (Montpellier and OriGene; see Materials & Methods; Table S2a). The distribution of samples across principal subspace using principal component analysis (PCA; Figure S1a) showed no batch effect between the two cohorts of samples. In addition, hierarchical clustering analysis using nCounter profiles clustered these 22 samples into different groups that potentially represent different subtypes ( Figure 1b).

Selection of alternative cost-effective method
Next, we evaluated if a more economical method employing a "low-cost" protocol (Elements chemistry; custom unique probes are attached to the biotin labels and molecular barcodes separately; approximately 35% less cost compared to standard protocol; Figure 1a) from NanoString Technologies can deliver similar classification performance compared to the standard protocol-based assay. The results from the low-cost protocol-based profiling (Table   S2b) showed similar distribution of samples (n=22) across principal subspace using PCA ( Figure S1b) and clustered into potential subtypes ( Figure S1c) in a similar fashion to the standard protocol.
We merged both the standard and low-cost protocols' gene expression data after normalising (gene-wise median centring) each dataset and performed PCA ( Figure S1d) to assess potential batch effect. Figure 1c shows the clustering of same samples between protocols.
Measurements showed high correlation (R 2 = 0.90, p<0.001; Figure 1d) between these different protocols. This demonstrates that we can successfully replicate results from standard protocol using low-cost protocol for a more cost-effective assay. Therefore, we adopted lowcost protocol for our assay (Figure 1a).

Reproducibility of NanoCRCAssigner assay between different batches
For the purposes of a clinical biomarker assay for CRC subtyping, it is imperative that results are highly reproducible between batches (e.g. at different time points pre and post-treatment).
To test this aspect of our platform, we performed our assay on five of the above samples twice, in separate batches of maximum 40 weeks apart; Table S2c-d). Figure 1e and Figure   S1e shows the clustering of replicate samples together with negligible batch effect. Next, when we compared the expression of each gene in each sample across the two replicates, we achieved high concordance between the assays, with a Pearson's R 2 of 0.98 (p < 0.001; Figure 1f). This establishes the high reproducibility of our assay over non-negligible periods of time. Hence, in the future, we can use this assay to test the state of subtypes using matched pre-and post-treatment biopsies or surgical materials.

Selection of robust set of genes for the assay
Successful clinical biomarker assays should be able to classify samples into subtypes with high concordance, and this requires a robust set of genes. Hence, we tested the robustness of our selected 47 CRCAssigner genes (out of 50-gene panel) using a reduced set of our published training dataset (n=192; Figure S2a and Table S3a) [3] and our in-laboratory developed intPredict bioinformatics tool, which contains a pipeline of supervised class prediction methods (see Supplementary Information, Figure S2b and

Subtyping using NanoCRCAssigner assay
In order to confirm that our NanoCRCAssigner assay can stratify patients into clinically  Table S4a-b). We performed NanoCRCAssigner assay and assigned subtypes to each sample by correlating the expression with the CRCAssigner-38 gene centroids (see Figure 2b shows the expression of our CRCAssigner-38 signature (NanoCRCAssigner assay) in these samples as measured on the nCounter platform. All the CRCAssigner subtypes were present in this cohort, and all samples were successfully classified, with none showing mixed (similar to that published for CMS classification [13]) or undetermined (those that cannot be classified confidently; Figure 2c; see Materials & Methods) subtype characteristics. We observed a non-uniform distribution of the subtypes with enterocyte contributing 41.2% of all the samples followed by stem-like (23.5%) and

Comparing subtyping concordance between NanoCRCAssigner assay and microarray platform
In order to compare the NanoCRCAssigner assay performance to the original technology platform from which the subtypes and signatures were derived, we also classified the samples using Affymetrix Human Genome U133 Plus 2.0 (HG-U133 Plus2) microarray profiles [4] ( Figure S3c). Similar to that of the NanoCRCAssigner assay, we used the CRCAssigner-38 centroids and classified the samples' microarray profiles into all the five subtypes with 35.2% as enterocyte, 23.5% as stem-like, 17.6% as goblet-like, 11.8% as inflammatory and 5.9% as TA ( Figure S3d). There was only one sample that was defined as a mixed subtype (5.9% of the 17 samples) expressing both inflammatory and enterocyte genes. The expression of enterocyte genes was consistent with the classification of the sample as enterocyte on the NanoCRCAssigner platform. The fact that the sample was classified as a mixture of subtypes by the CRCAssigner-38 may be attributed to platform-specific effects. Overall, the NanoCRCAssigner assay showed perfect 100% concordance with the microarray-based CRCAssigner-38 classification after excluding samples with mixed classification (due to challenges in comparing these mixed subtypes to others) (Figure 2d).
Similarly, we applied the original CRCAssigner-786 centroids to the microarray data to classify samples into the five subtypes. The CRCAssigner-786 classification yielded 23.5% enterocyte, 29.4% stem-like, 23.5% goblet-like, 11.8% inflammatory and 5.9% TA samples ( Figure S3e). There was one mixed sample ( Irrespective of the different number of genes profiled on the different platforms, the NanoCRCAssigner assay showed 87.5% concordance with the CRCassigner-786 subtypes ( Figure 2d). Again, the 12.5% discordance may be attributable to noisy genes present in the CRCassigner-786 signature, as we observed a saturation of the MCR beyond 38 genes using multiple class prediction methods ( Figure S2d). Overall, the NanoCRCAssigner assay and CRCAssigner-786 classification perform well with low discordance.
In order to statistically validate these findings, we performed Fisher's exact test using these classifications, excluding mixed or undetermined samples. We found that NanoCRCAssigner assay was significantly (false discovery rate; FDR<0.001; Figure

Comparison of the CMS classification with NanoCRCAssigner subtypes and survival analysis
We assessed if our NanoCRCAssigner assay would be useful to classify samples according to the CMS subtypes. We classified the Montpellier cohort of 17 microarray gene expression profiles into CMS subtypes using the published CMS classifier [13]. We successfully classified the samples into all of the CMS subtypes: 47.1% were CMS2 (enterocyte and TA); 11.8% each of CMS3 (goblet-like) and CMS4 (stem-like); and 5.9% of CMS1 (inflammatory). However, we found 17.6% mixed and 5.9% undetermined samples ( Figure   2f, Materials & Methods). Thus, NanoCRCAssigner showed fair concordance (84.6%) with the CMS classifier excluding the mixed and undetermined samples (Figure 2d). We further performed pairwise Fisher's exact test between the CMS and NanoCRCAssigner subtypes (Figure 2e), which confirmed borderline significant association to the CMS classification (FDR=0.07). This suggests that NanoCRCAssigner may be applied as a surrogate to predict CMS subtypes with low discordance (only 15.4%) in addition to CRCAssigner subtypes.
Although it is challenging to assess predictive power of an assay in a small Montpellier cohort, we sought to understand the potential of the assay to predict prognosis. Therefore, we performed Kaplan-Meier survival analysis and log-rank test for overall survival (OS) using the NanoCRCAssigner assay and 15 non-mixed samples. The CRCAssigner subtypes identified using NanoCRCAssigner showed borderline significance of p=0.08 (log-rank test; Figure 2g). Similar results were obtained using CRCAssigner-38 and CRCAssigner-786 ( Figures S3f-g). On the other hand, the CMS subtypes showed no significance in OS ( Figure   S3h) potentially due to small sample size (n=13, excluding mixed and undetermined samples). Overall, the NanoCRCAssigner assay has the potential to predict patient prognosis.
This prognostic prediction of the NanoCRCAssigner assay requires further validation with a large cohort of samples.

Performance of NanoCRCAssigner assay in a multi-stage Asian cohort
We then utilised an additional cohort of 23 CRC samples from an Asian population comprising all disease stages (SG cohort; Materials & Methods) to further evaluate our subtyping assay using the NanoCRCAssigner assay ( Figure 3a; Table S4c-d). To our knowledge, these are the first set of Asian CRC samples profiled for molecular subtyping using nCounter platform (Figures S4a-c). Interestingly, the NanoCRCAssigner assay identified all the five CRCAssigner subtypes representing that these subtypes are also present in the Asian population. In this data set, again the enterocyte subtype showed the highest prevalence with 30.4%. This was followed by goblet-like (17.4%) and the three other subtypes at 8.7% prevalence (Figures 3b-c). Interestingly, one of the samples (sample number -1017) with goblet-like subtype characterisation also showed mucinous characteristics by pathological evaluation (Table S4e) representing that our assay predicts subtypes based on their characteristics. 26.1% of samples were mixed subtype in the SG cohort (Figure 3c), an increase with respect to the previous set, suggesting that there may be a different proportion of mixed samples in the Asian population compared to the Western population. There was no visible trend in association between stages and subtypes ( Figure S4d; Table S4e). subtypes when mixed subtypes were not included. Also, there was a high concordance of 88.9% between NanoCRCAssigner and CMS subtypes (Figures S4f-j). In all 6 samples classified as mixed by NanoCRCAssigner, one of the subtypes in the mixture was also identified as the only subtype by CRCAssigner-38 or CRCAssigner-786. Overall, our NanoCRCAssigner assay identifies subtypes in Asian CRC patients independent of the disease stages.

Evaluating associations between NanoCRCAssigner assay subtypes and mutational and microsatellite instability (MSI) profiles
Previously, it has been reported that the CMS1/inflammatory subtype is associated with MSI and BRAF mutations, whereas CMS3 is highly associated with KRAS mutations [13]. To further validate the NanoCRCAssigner subtypes, we compared these with the mutational (BRAF and KRAS) and MSI status (Figure 3d) of cancers in the Asian cohort (n=11). All (100%) the inflammatory subtype CRCs were associated with MSI. Interestingly, one of the two BRAF mutant tumours was associated with the inflammatory (CMS1) subtype and MSI status. Similarly, all three goblet-like subtype (CMS3; 100%) tumours were associated with KRAS mutation. There were three other KRAS mutant tumours associated with the enterocyte or stem-like subtypes, representing KRAS mutant tumours are also associated with other subtypes as previously reported [13]. This analysis with our NanoCRCAssigner assay corresponds with known associations of subtypes with mutational and MSI profiles with this small SG data set of 17 samples. However, additional large data sets and NanoCRCAssigner assays are warranted to validate these observations.

Effect of tumour cellularity on NanoCRCAssigner assay
Next, we sought to test if tumour cellularity affects our NanoCRCAssigner assay.  Table S4f-g). One sample had extremely low cellularity of 10% and one additional sample was from a liver metastasis. We assigned subtypes to these 17 RNA samples using the different classifiers as shown for the previous cohorts. Since the CA1 gene could not be mapped from the probes in the Affymetrix GeneChip Human Transcriptome Array (HTA) array, we reduced our CRCAssigner-38 to 37 genes for comparison ( Figure S5c-d). For consistent evaluation across platforms, we maintained the 37 genes for all assays in this cohort. The NanoCRCAssigner assay predicted all five subtypes subtypes together with 17.6% mixed subtypes (Figures 4bc). Unlike the other cohorts, OriGene showed the highest proportion of TA (29.4%) followed by enterocyte (23.5%), goblet-like (11.8%), inflammatory (11.8%), and stem-like (5.9%; Figure 4c). This suggests that NanoCRCAssigner assay is not biased towards any particular subtype.
The Furthermore, we postulated that if cellularity affects our assay, the low cellularity samples should be either qualified as "undetermined" or "mixed subtype" samples. Interestingly, none of the samples were classified as having undetermined subtype, regardless of cellularity. On the other hand, the three mixed samples were among those having high cellularity (50-75%) (Figures 4d-e). These results suggest that our NanoCRCAssigner may not be affected by cellularity due to the selection of robust gene sets.

Overall concordance and distribution of subtypes across all cohorts
To assess the stability of individual subtypes across the various platforms, gene sets and data sets, we plotted Figure 5a using all significantly classified and non-mixed samples for all three cohorts (n = 39; Table S5). Three of the assays are shown (CRCAssigner-786, CRCAssigner-38 and NanoCRCAssigner) along with the 5 CRCAssigner subtypes. While the goblet-like subtype was consistent across all the different classifications, there were four samples that had different subtypes across classifications (shown in grey in Figure 5a). Two of the samples were classified as either enterocyte or stem-like, one each as either enterocyte or inflammatory and either TA or inflammatory. Three of these four samples were from the OriGene cohort (Figure 4b), potentially due to platform-specific effects as discussed previously. However, overall concordance between platforms was good (Figure 5a) We also sought to confirm that platform and gene set differences did not bias the distribution of subtypes assigned to the samples. Figure 5b shows the proportion of subtypes according to each classifier across the three cohorts, excluding mixed or undetermined samples (n = 39).
P-values are the result of proportion tests and show that there is no significant difference in the distribution of each subtype across the three CRCAssigner-based classifiers.

Characteristics of the identified subtypes
In order to further confirm if the classified subtypes represent the molecular characteristics of their original published phenotypes, we performed analysis using NanoString Technologies' PanCancer Progression Panel that represents mainly epithelial, mesenchymal and extracellular matrix genes (Table S6a- Figure 5d alongside the NanoCRCAssigner subtypes of the samples. Overall, these analyses demonstrate that the subtypes identified by NanoCRCAssigner assay represent the molecular characteristics of these subtypes in 51 samples (with matched gene expression profiles from other platforms) from three independent cohorts.

Conclusion
In summary, we developed a low-cost protocol-based biomarker assay (NanoCRCAssigner)

ADDITIONAL INFORMATION
Previously published GEO Omnibus data sets were analysed for gene set selection (GSE14333 and GSE13294) and microarray-based subtyping of the Montpellier cohort (GSE62080). nCounter data for all cohorts (GSE101479 -standard protocol and GSE101481 -low-cost protocol) and microarray/RNA-Seq data for OriGene (GSE101472) and Singapore (GSE101588) cohorts are deposited under the SuperSeries with accession number GSE101651.

COMPETING FINANCIAL INTERESTS
A.S. has ownership interest (including patents) as a patent inventor for a patent entitled "Colorectal cancer classification with differential prognosis and personalized therapeutic responses" (patent number PCT/IB2013/060416).

ETHICAL APPROVAL AND INFORMED CONSENT
The SG cohort of this study was approved by the SingHealth Institutional Review Board: 2013/110/B.