Comprehensive characterization of tumor microenvironment in colorectal cancer via molecular analysis

Article activity feed

1. 

   Version published to 10.7554/elife.86032 on eLife

   Jun 2, 2023
2. 

   eLife

   Apr 23, 2023

   Author Response

   Reviewer #1 (Public Review):

   This work puts forward a comprehensive characterisation of colorectal cancer (CCCRC), by classifying it into 4 subtypes with distinct TME features. It uses 10 public databases: 8 microarray datasets for the training of molecular classification and 2 RNAseq for validation (CRC-RNAseq) to identify the 4 subtypes using unsupervised machine learning (consensus clustering). These 4 subtypes were found to be somewhat distinct in terms of immune response and the possibilities for effective treatments. They found that one subtype may be more sensitive to chemotherapy, two to WNT pathway inhibitor SB216763 and Hedgehog pathway inhibitor vismodegib, and one to ICB treatment. They show an association with patient outcome in terms of PFS, validated in the validation cohort. They used histology to …

   Author Response

   Reviewer #1 (Public Review):

   This work puts forward a comprehensive characterisation of colorectal cancer (CCCRC), by classifying it into 4 subtypes with distinct TME features. It uses 10 public databases: 8 microarray datasets for the training of molecular classification and 2 RNAseq for validation (CRC-RNAseq) to identify the 4 subtypes using unsupervised machine learning (consensus clustering). These 4 subtypes were found to be somewhat distinct in terms of immune response and the possibilities for effective treatments. They found that one subtype may be more sensitive to chemotherapy, two to WNT pathway inhibitor SB216763 and Hedgehog pathway inhibitor vismodegib, and one to ICB treatment. They show an association with patient outcome in terms of PFS, validated in the validation cohort. They used histology to correspond the subtypes to known pathological types, as well as investigating their T cell makeup. They also investigated the genetic tumour evolution that may occur between the subtypes. A single-sample gene classifier was put forward as a way of identifying the class of cancer. The evidence for the main results of the work is convincing, but a few areas need to be clarified and extended.

   In the determination of the 4 subtypes (C1-C4) the methodology is clear, and the definition of the training and validation data are clear and well presented. The techniques used are well suited to the problem. The performance of the classification as a predictor of prognosis is presented as KM curves of PFS and OS for the training and validation sets. The training data shows a significant log-rank p-value in both PFS and OS. The validation data shows a significant effect in PFS.

   What follows is quite an exhaustive process of finding differences between the cohorts using a multitude of techniques and datasets, including genomics, epigenetics, transcriptomics, and proteomics. These sections are mainly descriptive and do add understanding to the classification, especially with regard to the T-cell populations that are invasive.

   Improvements could be made to the latter sections of the main paper. The basis for the potential clinical responses of the subtypes is arrived at via a "pre-clinical model" based on 81 genes. It would benefit from clarification on what genes were used in model training and details of the final model. Similarly the description of the "Single-sample gene classifier" could be enhanced similarly with a better description of which genes are in the final classifier.

   Thank you for taking the time to review our article and for your positive feedback. Your thorough evaluation of our work has been invaluable to us, and we appreciate your recognition of the effort we put into it.

   1. The basis for the potential clinical responses of the subtypes is arrived at via a "pre-clinical model" based on 81 genes.

   The exact details of the filtering criteria used to obtain the list of pre-clinical model genes have added to the Methods section of the study (Lines 1061-108, Lines 503-511) (Supplementary file 3a). To explore the treatment for each CCCRC subtype using cancer cell line drug-sensitivity experiments, we developed a pre-clinical model based on subtype-specific, cancer cell-intrinsic gene markers according to a previously published study (Eide et al., 2017). Firstly, the “limma” package was used to identify DEGs with FDR < 0.05 between each of the four subtypes and the remaining subtypes in the CRC-AFFY cohort. To identify subtype-specific genes in one of the subtypes, we excluded those that were found to be differentially expressed in comparisons between one of the other subtypes and the remaining subtypes. The upregulated subtype-specific genes (log2 (fold change, FC) > 0 and FDR < 0.05) was ranked based on their log2FC and selected the top 500 genes for further gene screening. Secondly, the GEP of human CRC tissues versus patient-derived xenografts (PDX) in the GSE35144 dataset by the R package “limma” was used to remove those genes associated with stromal and immune components. DEGs with FDR > 0.5 and log2 FC < 1 between human CRC tissues versus PDX were considered as cancer cell-intrinsic genes. Thirdly, we also utilized human CRC cell lines to obtained cancer cell-intrinsic genes. A total of 71 human CRC cell lines with RNAseq data (log2TPM) was obtained from the Genomics of Drug Sensitivity in Cancer (GDSC) database (https://depmap.org/portal/download/all/), 43 of which had dose-response curve (area under the curve, AUC) values. The MSI status, FGA and TMB information of CRC cell lines was obtained from cbioportal website (https://www.cbioportal.org/study/summary?id=ccle_broad_2019). RNAseq data for 71 human CRC cell lines was used to further determine the cancer cell-intrinsic genes and genes among the top 25% within (i) the 10−90 % percentile range of the largest expression values and (ii) the highest expression in at least three samples. The subtype-specific genes and cancer cell-intrinsic genes were intersected to generate the gene list for developing the pre-clinical model. The pre-clinical model was developed using the nearest template prediction (NTP) function of R package “CMScaller”, which can be applied to cross-tissues and cross-platform predictions (Hoshida, 2010). The GEP (log2TPM) of 71 human CRC cell lines normalized by the Z-score were input into the pre-clinical model, and the cell lines were divided into four CCCRC subtypes. (Lines 1061-1088)

   Here we want to make a point that we changed from using the xgboost algorithm to using the NTP algorithm to build our pre-clinical model. Based on the genomic features of the cell line, we evaluated the reliability of the final pre-clinical model and found that the pre-clinical model built using the NTP algorithm is more reliable. As expected, the C4 subtype cell lines demonstrated the highest TMB values and MSI frequency while exhibiting the lowest FGA scores when compared to other subtypes (Figure 6-figure supplement 1G-I). In contrast, C1 and C3 subtype cell lines showed significantly higher FGA scores and significantly lower TMB values and MSI frequency. The C2 subtype cell lines had median FGA scores, TMB values, and MSI frequency. The pre-clinical model is publicly available at https://github.com/XiangkunWu/pre_clinical_model. (Lines 503-511)

   1. Similarly the description of the "Single-sample gene classifier" could be enhanced similarly with a better description of which genes are in the final classifier.

   We apologize for any confusion caused in our revised regarding the derivation of the CCCRC classifier. Specifically, we have added more details on the derivation of model genes and the establishment of the model, and ensured the availability of the CCCRC classifier. The method details and results of deriving the model genes and building the model are described next. (Lines 1102-1121) (Lines 562-579) (Supplementary file 3c)

   In order to facilitate the widespread application of CCCRC classification system, we established a simple gene classifier to predict CCCRC subtypes. Firstly, we filtered genes based on their mean expression and variance in the CRC-AFFY cohort, and genes with expression and variance below the bottom 25% were removed. Then, we applied the Random Forest algorithm (RF) in the R package "caret" to perform feature selection on the CCCRC subtype-specific genes of the CRC-AFFY cohort. The top 20 most informative features for each subtype were ranked and selected based on the impurity measure generated by the algorithm. This allowed us to identify critical genes that are strongly associated with each CCCRC subtype and develop the CCCRC classifier. Next, we randomly divided the CRC-AFFY cohort into training and validation sets at a ratio of 7:3 using “createDataPartition” function provided in the R package "caret" (seed=123). The GEP was normalized with Z-scores prior to model training and validation. The CCCRC classifiers were trained with the top 80 subtype-specific genes using the RF, Support Vector Machine (SVM), eXtreme Gradient Boosting (xgboost), and Logistic Regression algorithms implemented in the R package "caret". Finally, we validated the CCCRC classifier on the GSE14333 and GSE17536 datasets, as well as the CRC-AFFY cohort. We evaluated the predictive performance of the CCCRC classifier by evaluating measures such as accuracy value and F1 score, which were generated using the " confusionMatrix " function provided in the R package "caret". (Lines 1102-1121)

   We established the CCCRC classifier on the training set by utilizing multiple machine learning algorithms based on the GEP of 80 upregulated subtype-specific genes (Supplementary file 3c). Upon application to the test set, GSE14333, and GSE17536 datasets, the performance of the eXtreme Gradient Boosting (xgboost) algorithm was the best with the highest accuracy values and F1 scores compared to the Random Forest (RF), Support Vector Machine (SVM), and Logistic Regression algorithms (Figure 6-figure supplement 4). Notably, the CCCRC classifier based on the xgboost algorithm displayed robust performance across gene expression platforms, Affymetrix and RNA-sequencing platforms, exhibiting a balanced accuracy of > 80% for all subtypes (Supplementary file 3d). These findings demonstrated the stability and cross-platform applicability of our classifier. The CCCRC classifier based on the xgboost algorithm is publicly available at https://github.com/XiangkunWu/CCCRC_classifier, and the CCCRC subtype information of CRC patients can be obtained by directly inputting the GEP of 80 upregulated subtype-specific mRNA genes. The CCCRC classifier might facilitate the discovery of new biomarkers and the personalized treatment of clinical patients with CRC. (Lines 562-579)

   Reviewer #2 (Public Review):

   This study aimed to classify colorectal cancer (CRC) samples based on the expression of genes in selected gene lists, where the gene lists were chosen to represent aspects of the tumour microenvironment, tumour-associated immune cells, and tumour cells. The resulting clusters were then used to define a classifier, followed by a detailed description of molecular features of the tumours and tumour microenvironments assigned to each cluster. The authors claim this study is more "holistic" than previous work on CRC clustering/classifiers because they aimed to explicitly include additional components of the tumour microenvironment in both the clustering/classifier definition and in the subsequent description of molecular characteristics.

   The CCCRC clustering and the resulting classifier presented in this paper are derived from published RNAseq studies. The multi-omics aspect of the work is restricted to smaller sample numbers for which both transcriptomic and another omics dataset were available in public resources and comprises a description or correlative analysis of each omics data type within each of the assigned CCCRC subtypes.

   By applying solid computational methods to a compendium of published RNAseq datasets (n~1500 tumours), they found that tumour samples from colorectal cancers clustered into 4 subtypes ("CCCRC" subtypes) on the basis of 61 pre-defined gene expression signatures. These subtypes correlated with but did not correspond to, the previously described Consensus Molecular Subtypes (CMS) of colorectal tumours.

   Other types of molecular data were available for some tumours, obtained from the same published resources: whole-slide images, mutations, tumour proteomics, and/or scRNAseq. The authors reanalysed these datasets using standard methods and drew correlations with the CCCRC subtypes they had assigned in this work. To (semi-)quantify immune infiltration characteristics from whole-slide images (WSI), they additionally performed automated segmentation in addition to review by pathologists, which in combination produced a convincing WSI-derived dataset.

   In combination with existing CRC classifications, this study could facilitate future biomarker discoveries. This appears to be the authors' main claim, and the data and methods broadly support this claim.

   Thank you for taking the time to review our article and for your positive feedback. Your thorough evaluation of our work has been invaluable to us, and we appreciate your recognition of the effort we put into it.

   Some aspects of the work need to be clarified:

   1. This work relies on the definition of 4 clusters of CRC tumours based on consensus clustering of the 61 gene lists, which in turn depends on the choice of clustering method and the choice of gene lists. Sufficient detail is provided about the gene lists and resulting clusters, but this paper does not show how robust the 4 clusters are to these choices; for example, the "Energy" gene list appears to be a relatively strong component of clusters C2 and C3.

   Thank you very much for providing such detailed and insightful feedback.

   1.1. The reviewer has raised a valid concern about the impact of gene list selection on the robustness of the clusters. To address this issue, we used the “pamr.predict” function of the R package “pamr” (Tibshirani et al., 2002) to extract centroids of each subtype that best represent each subtype and establish a PAMR classifier. PAM (Prediction Analysis of Microarrays) is a statistical technique to identify subsets of features that best characterize each class using nearest shrunken centroids (Tibshirani et al., 2002). The technique is general and can be used in many other classification problems. As shown in Figure 1-figure supplement 2E, a threshold of 0.566 with minimum 10-fold cross-validation error was selected to identify the 61 TME-related signatures that exhibit at least one non-zero difference between each subtype (seed = 11). These signatures were then used to construct a PAMR classifier with superior predictive capability, exhibiting an overall error rate of 15%. We used the established PAMR classifier to predict the CCCRC subtypes on the CRC-RNAseq cohort and the same four CCCRC subtypes were revealed, with similar patterns of differences in the TME components (Fig. S2F, G). This indicated that the 61 TME-related signatures best represent each subtype and are indispensable for achieving the identification of the four CCCRC subtypes. (Lines 161-168)

   1.2. The reviewer has raised a valid concern about the impact of the clustering method selection on the robustness of the clusters.

   We performed extensive data analysis attempts during our unsupervised clustering analysis, which primarily involved attempting various clustering methods, including K-means clustering, non-negative matrix factorization (NMF) clustering, and hierarchical clustering, as well as replacing different sources and categories of the TME-related signatures. To determine the optimal clustering method and TME panel, we evaluated whether the TME panel could reproduce the heterogeneity of TME, the stability of the clustering itself, the biological characteristics of the subtypes, the correlation between subtypes and prognosis, and the correlation between subtypes and microsatellite instability (MSI), consensus molecular subtypes (CMS) classification system, and other molecular subtype systems. Due to the abundance of exploratory data analysis results, we ultimately selected the best clustering method and TME panel combination for showcase.

   1.3. Also, we analyzed the sensitivity analysis of the effect of TME-related signatures on the clustering results. Since the effect of removing one of the TME-related signatures on the clustering results was not well evaluated, we attempted to remove the entire category. We performed consensus clustering analysis again using the same parameters (partitioning around medoids (pam) clustering; "Pearson" distance; 1,000 iterations; from 2-6 clusters). When we conducted consensus clustering analysis using only immune-related signatures, we identified three subtypes: low (C2), moderate (C3), and high (C1) immune infiltration subtypes. When we included both immune-related and tumor-related signatures, we identified four subtypes: immunomodulatory (S1), cold (S2/S3), and immune-excluded (S4) subtypes. It appears that the immunosuppressed subtype in the CCCRC classification system may have been assigned to both S1 and S4 subtypes. Limiting the consensus clustering analysis to only immune-related or immune- and stroma-related signatures, as done in previous studies (Bagaev et al., 2021; He et al., 2018), did not allow reliable identification of all four CCCRC subtypes. These sensitivity analyses underscored the necessity of our well-designed TME panel to achieve the identification of the four CCCRC subtypes. (Lines 172-176) (Figure 1-figure supplement 4)

   1. The authors examined whether their CCCRC classification showed differential disease progression in available retrospective cohorts of people treated with anti-PDL1 therapy. The authors presented this work as "significance of CCCRC in guiding the clinical treatment of colorectal cancer", but the data presented in this section cannot support clinical treatment decisions, which would require prospective studies and clinical trial designs. However, this section is potentially useful for generating hypotheses about potential biomarkers related to the CCCRC subtypes, and might, in the future with additional evidence, contribute to the design of a trial. The authors point out that additional experimental evidence would be required.

   Thank you for your constructive suggestions. We agree that our retrospective analysis of the CCCRC classification in relation to disease progression under immune checkpoint blockade treatment does not directly support clinical treatment decisions. We acknowledge that additional experimental evidence would be required to fully support the use of the CCCRC classification as a clinical tool for guiding treatment decisions. We have highlighted in the corresponding section of the article that this research is pre-clinical and still requires substantial basic experiments and clinical trials to validate. (Lines 536, 751)

   1. Other prognostic or predictive clinicopathological variables for colorectal cancer are not discussed in detail in the present work but are important for further work on the prognostic and predictive value of CRC molecular subtypes and biomarker derivation. Discrepancies in treatment response have previously been observed in separate CRC trials of biologically targeted agents with different chemotherapy backbones and other authors have suggested that treatment interactions with the tumour microenvironment might in part explain these discrepancies (e.g. Aderka (2019) PMID:31044725, and others).

   3.1) Other prognostic or predictive clinicopathological variables for colorectal cancer are not discussed in detail in the present work but are important for further work on the prognostic and predictive value of CRC molecular subtypes and biomarker derivation.

   Thank you for bringing up this point. We apologize for not analyzing other clinicopathological variables for colorectal cancer in more detail in my original work. We agree that these variables are important for further study of our CCCRC classification system to guide biomarker derivation and to guide clinical treatment decisions. We added in the article the relationship between CCCRC subtypes and clinicopathological variables, as well as the comparison with CMS subtypes (Lines 256-262, 661-666). In addition, we have identified a clerical error in our manuscript and have corrected it accordingly. Specifically, the use of PFS as the endpoint in some parts of the manuscript was a mistake and has been corrected to DFS. We would like to clarify that the endpoint for the CRC-AFFY and CRC-RNAseq cohorts is DFS and OS, while the endpoint for the GSE104645 dataset is PFS and OS. For the immune checkpoint blockade therapy cohort, the endpoint for PRJEB23709 (Gide) is PFS and OS, and for the GSE135222 (Jung) dataset, the endpoint is PFS. Progression Free Survival (PFS) refers to the time from randomization (or treatment initiation) to the first occurrence of disease progression or death from any cause. The definition of Disease-Free Survival (DFS) is the time from randomization to the appearance of evidence of disease recurrence.

   We further analyzed the association of CCCRC subtypes with clinicopathological characteristics (Supplementary file 1f, Supplementary file 1g). We found that the C4 subtype was mostly diagnosed in right-sided CRC lesions and in females, which was consistent with the CMS1 subtype. The C1 and C3 subtypes were mainly observed in left-sided CRC lesions and in males, consistent with the CMS2 and CMS4 subtypes. The C3 subtype was strongly associated with more advanced tumor stages, which was the similarity to the CMS4 subtype, while the C4 subtype was associated with higher histopathologic grade, which was the similarity to the CMS1 subtype. Furthermore, our analysis using the Kaplan-Meier method demonstrated that patients with the C4 subtype had significantly higher disease-free survival (DFS) and overall survival (OS) compared to those with the C2 and C3 subtypes in the CRC-AFFY (Figure 1I, Figure 1-figure supplement 7A) and CRC-RNAseq cohorts (Figure 1-figure supplement 7B, C). Multivariate Cox proportional hazard regression analysis showed that the C4 subtype was an independent predictor of the best OS and DFS, whereas the C3 subtype was an independent predictor of the worst OS and DFS after adjustment for age, gender, tumor site, TNM stage, grade, adjuvant chemotherapy or not, MSI status, BRAF and KRAS mutations, and the CMS classification system in the combined cohort (the CRC-AFFY and CRC-RNAseq cohorts) (Supplementary file 1h). Considering that the C1, C2/C3, and C4 subtypes partially overlap with the CMS2, CMS4, and CMS1 subtypes, respectively, we also analyzed the prognostic differences between them in the combined cohort. We found that the DFS/OS of patients with the C1 subtype was worse than those with the CMS2 subtype (Figure 1-figure supplement 7D, E), the DFS/OS of patients with the C2 subtype was better than those with the CMS4 subtype (Figure 1-figure supplement 7F, G), the DFS/OS of patients with the C3 subtype was not significantly different from those with the CMS4 subtype (Figure 1-figure supplement 7F, G), and the DFS/OS of patients with the C4 subtype was significantly better than those with the CMS1 subtype (Figure 1-figure supplement 7H, I). Notably, the C2 subtype within the CMS4 subtype also had a better prognosis than the C3 subtype within the CMS4 subtype (Figure 1-figure supplement 7J, K). The above analysis demonstrated that the CCCRC classification system were closely associated with clinicopathological characteristics, were able to refine the CMS classification system and MSI status, as well as contributed to the understanding of the mechanisms underlying the different clinical phenotypes resulting from TME heterogeneity.

   3.2) Discrepancies in treatment response have previously been observed in separate CRC trials of biologically targeted agents with different chemotherapy backbones and other authors have suggested that treatment interactions with the tumour microenvironment might in part explain these discrepancies (e.g. Aderka (2019) PMID:31044725, and others).

   The reviewer's comments greatly contributed to the quality of our study. Aderka et al. discussed the reasons for the differences in the results of the CALGB/SWOG 80405 and FIRE-3 clinical trials, which may be related to differences in the chemotherapy backbone used and TME heterogeneity (Aderka et al., 2019). Both trials evaluated the combination of cetuximab or bevacizumab with a different chemotherapy backbone: in the CALGB/SWOG 80405 trial, 75% of patients received oxaliplatin, while in the FIRE-3 trial, all patients received irinotecan. The CCCRC classification system also facilitates the understanding of the differences in the results of the CALGB/SWOG 80405 and FIRE-3 clinical trials (Heinemann et al., 2014; Lenz et al., 2019). We have added this content to the discussion section of the article (Lines 753-777). Based on our examination of the results summarized in Figure 4 of the work by Aderka et al. (Aderka et al., 2019), we found that differences in the treatment outcomes of the CMS1 and CMS4 subtypes were the crucial factor behind the divergent results observed in the two clinical trials. The CMS1 and CMS4 subtypes have a microenvironment rich in CAFs. Our CCCRC classification results also showed that CMS1, in addition to mainly consisting of the C4 subtype, also contains a considerable number of the C2 subtype, while the CMS4 subtype mainly consists of the C2 and C3 subtypes. Furthermore, our study results indicated that the C2 subtype is suitable for chemotherapy in combination with bevacizumab, possibly because the combination can inhibit the CAFs and abnormal blood vessel formation in the microenvironment, thus alleviating the immune suppression of the immune cells. However, the C3 subtype is not suitable for chemotherapy in combination with bevacizumab because it only accumulates CAFs and abnormal blood vessel formation but lacks T cell infiltration. Therefore, we boldly speculate that the CMS1 and CMS4 subtypes in the CALGB/SWOG 80405 clinical trial may contain more C2 subtypes than those in the FIRE-3 clinical trial, leading to the CMS1 and CMS4 subtypes in the CALGB/SWOG 80405 clinical trial being more suitable for chemotherapy in combination with bevacizumab than cetuximab compared to the FIRE-3 clinical trial. Overall, the integration of CCCRC and CMS classification systems provides valuable insights for understanding the divergent outcomes of the two clinical trials (Lines 753-777).

   Reviewer #3 (Public Review):

   In their study: Comprehensive characterization of tumor microenvironment in colorectal cancer via histopathology-molecular analysis, Wu et al., aim to examine the contribution of the tumour microenvironment (TME) on biological and clinical heterogeneity in colorectal cancer (CRC).

   To achieve this the authors use a vast array of publicly available datasets across a variety of biological modalities (transcriptomic, epigenetic, mutational). Using thoughtfully curated genesets the authors classify CRC into 4 holistic comprehensive characterised CRC (CCCRC) subtypes which comprise immune, stromal, and tumour features of CRC biology.

   The authors investigate the association of their novel CCCRC subtypes with current "gold standard" classification schemes.

   The authors' integration of deep learning methods for HE classification and subsequent association with "Tumor level" CCCRC subtypes is a refreshing addition to the study. Comment on the degree of heterogeneity observed in HE samples and correlation to the heterogeneity of CCCRC subtypes would be a welcomed addition. It is likely publicly available datasets from such platforms as 10X Genomic Visium would be available for this type of analysis.

   Whilst one of the main outcomes of the study is the addition of another classification scheme to the field of colorectal cancer, the CCCRC scheme represents a holistic perspective on CRC classification.

   The authors provide a welcomed graphical overview of the complex narrative of the study in Figure 7.

   The authors focus on the classification of inter-patient heterogeneity and its associated predictive and prognostic utility. There appears to be a significant degree of overlap between immunosuppressive and immune excluded, and proliferative and immuno-modulatory signatures in Figure 1A. One of the major limitations of patient response to treatment is intra-patient heterogeneity, it would be nice for the authors to comment briefly on the degree of intra-patient heterogeneity of the CCCRC subtypes.

   Overall the authors succeed in providing a holistic deep characterization of CRC from the perspective of a variety of biological modalities. The authors provide a novel classification scheme for the field of CRC which demonstrates prognostic and predictive utility, which would benefit from further validation from external datasets. The authors demonstrate a pathway for integration and interpretation of complex high-dimensional data into clinically translatable currency such as the H&E.

   Thank you for taking the time to review our article and for your positive feedback. Your thorough evaluation of our work has been invaluable to us, and we appreciate your recognition of the effort we put into it.

   1. Comment on the degree of intra-patient heterogeneity of CCCRC subtypes would be nice.

   We have added intra-tumor heterogeneity analysis for each subtype (Lines 196-198). The level of intratumor heterogeneity (ITH) was significantly linked to poor prognosis and drug resistance (Caswell and Swanton, 2017). The ITH data used in our study for the CRC-RNAseq cohort was obtained from a previous study conducted by Thorsson et al. (Thorsson et al., 2018). As expected, the ITH of the C2 and C3 subtypes was higher than that of the other subtypes, while the ITH of the C4 subtype was the lowest (Figure 1F). Our analysis using the Kaplan-Meier method demonstrated that patients with the C4 subtype had significantly higher overall survival (OS) and disease-free survival (DFS) compared to those with the C2 and C3 subtypes. Furthermore, the C3 subtype was resistant to chemotherapy, cetuximab, bevacizumab, and ICB therapy. Our investigation of drug sensitivity data of cell lines also indicated that the C2 and C3 subtypes were generally not responsive to most drugs.

   1. A significant degree of overlap between immunosuppressive and immune excluded, and proliferative and immuno-modulatory signatures in Figure 1A is apparent and should be commented upon.

   Our research revealed that both C2 and C3 subtypes exhibited a high level of tumor stroma, while C1 and C4 subtypes were characterized by active DNA damage and repair and high tumor proliferation. Additionally, C2 and C4 subtypes had an abundance of immune components. This was consistent with our finding that there may be interconversion between the C1 and C4 subtypes, between the C4 and C2 subtypes, and between the C2 and C3 subtypes in this evolutionary pattern. The interconversion between C2 and C4 subtypes in this evolutionary pattern was the rarest situation, indicating that once the tumor enters the C2 subtype, it is difficult to reverse and will progress to the C3 subtype. (Lines 637-644)

   1. It is likely publicly available datasets from such platforms as 10X Genomic Visium would be available for this type of analysis.

   To investigate the spatial distribution relationship between four CCCRC subtypes of tumor cells, T cells, and stromal cells, we conducted a re-analysis of publicly available CRC spatial transcriptomics data (ST) obtained from the 10X website (https://www.10xgenomics.com/resources/datasets). The Space Ranger output files were then processed with Seurat (V4.1.1) (Hao et al., 2021) using SCTransform for normalization (Hafemeister and Satija, 2019). RunPCA were used to dimension reduction and RunUMAP to visualize the data. We used “ssGSEA” method implemented in the R package “GSVA” to score the six cell types (C1-C4 subtype cancer cells, mesenchymal cells, and T cells) (Hänzelmann et al., 2013). The “ssGSEA” method has been previously demonstrated to be highly reliable and suitable for ST data analysis (Wu et al., 2022). The cell-type-rich region was defined as the ssGSEA score of each cell type from one spot larger than the 75% quantile of this cell type. The markers for the six cell types are listed in the Supplementary file 1a and Supplementary file 3a. (Lines 1090-1102)

   The Cytassist and Visium samples had a total of 9080 and 2660 spots, respectively. We used “ssGSEA” method to quantify the six cell subpopulations of each spot and also visualized only the spots corresponding to the top 25% of the score ranking for each cell type (Figure 6-figure supplement 2AB, Figure 6-figure supplement 3AB). In Cytassist samples, we observed different spatial distribution patterns of the four subtypes of tumor cells (Figure 6-figure supplement 2B). Specifically, the C3 subtype of tumor cells was predominantly located in the tumor periphery with an enrichment of mesenchymal cells and T cells (areas selected by black dashed circles). In contrast, the C4 subtype of tumor cells was mainly present in the center of the tumor, accompanied by the presence of T cells. The C1 and C2 subtypes of tumor cells were distributed in relatively uniform areas, mainly in the tumor periphery, with fewer mesenchymal cells and T cells. However, the distribution areas of C2 subtype and C3 subtype of tumor cells also partially were in overlap (the area selected by red dashed circles). The same distribution patterns can also be observed in the Visium sample (Figure 6-figure supplement 3B). Further analysis of the correlation between the ssGSEA scores of each cell type in the cell-type-rich regions and those of other cell types was conducted (Figure 6-figure supplement 2D, E, Figure 6-figure supplement 3D, E). We found that in the C3 subtype-rich region of tumor cells, the C3 subtype score of tumor cells was significantly positively correlated with the mesenchymal cell score, while in the T cell-rich region, the C3 subtype score of tumor cells was significantly negatively correlated with the T cell score. The C4 subtype score of tumor cells was significantly positively correlated with the T cell score and negatively correlated with the mesenchymal cell score in the C4 subtype-rich, T cell-rich, and mesenchymal cell-rich regions. The C1 subtype and C2 subtype scores of tumor cells were negatively correlated with mesenchymal cell and T cell scores. Overall, these results were generally consistent with previous histopathologic analysis findings. (Lines 538-562)

   Read the original source
3. 

   eLife

   Apr 22, 2023

   eLife assessment

   This study represents a valuable body of work in which the authors assemble a molecular description of colorectal cancer and classification into subtypes. Overall, the evidence supporting the findings is solid and could be improved with more detail. Consensus over a diverse range of data from publicly available sources is convincing. When added to existing knowledge this work may contribute to future biomarker discoveries for colorectal cancer.

   Read the original source
4. 

   eLife

   Apr 22, 2023

   Reviewer #1 (Public Review):

   This work puts forward a comprehensive characterisation of colorectal cancer (CCCRC), by classifying it into 4 subtypes with distinct TME features. It uses 10 public databases: 8 microarray datasets for the training of molecular classification and 2 RNAseq for validation (CRC-RNAseq) to identify the 4 subtypes using unsupervised machine learning (consensus clustering). These 4 subtypes were found to be somewhat distinct in terms of immune response and the possibilities for effective treatments. They found that one subtype may be more sensitive to chemotherapy, two to WNT pathway inhibitor SB216763 and Hedgehog pathway inhibitor vismodegib, and one to ICB treatment. They show an association with patient outcome in terms of PFS, validated in the validation cohort. They used histology to correspond the subtypes …

   Reviewer #1 (Public Review):

   This work puts forward a comprehensive characterisation of colorectal cancer (CCCRC), by classifying it into 4 subtypes with distinct TME features. It uses 10 public databases: 8 microarray datasets for the training of molecular classification and 2 RNAseq for validation (CRC-RNAseq) to identify the 4 subtypes using unsupervised machine learning (consensus clustering). These 4 subtypes were found to be somewhat distinct in terms of immune response and the possibilities for effective treatments. They found that one subtype may be more sensitive to chemotherapy, two to WNT pathway inhibitor SB216763 and Hedgehog pathway inhibitor vismodegib, and one to ICB treatment. They show an association with patient outcome in terms of PFS, validated in the validation cohort. They used histology to correspond the subtypes to known pathological types, as well as investigating their T cell makeup. They also investigated the genetic tumour evolution that may occur between the subtypes. A single-sample gene classifier was put forward as a way of identifying the class of cancer.

   The evidence for the main results of the work is convincing, but a few areas need to be clarified and extended.

   In the determination of the 4 subtypes (C1-C4) the methodology is clear, and the definition of the training and validation data are clear and well presented. The techniques used are well suited to the problem. The performance of the classification as a predictor of prognosis is presented as KM curves of PFS and OS for the training and validation sets. The training data shows a significant log-rank p-value in both PFS and OS. The validation data shows a significant effect in PFS.

   What follows is quite an exhaustive process of finding differences between the cohorts using a multitude of techniques and datasets, including genomics, epigenetics, transcriptomics, and proteomics. These sections are mainly descriptive and do add understanding to the classification, especially with regard to the T-cell populations that are invasive.

   Improvements could be made to the latter sections of the main paper. The basis for the potential clinical responses of the subtypes is arrived at via a "pre-clinical model" based on 81 genes. It would benefit from clarification on what genes were used in model training and details of the final model. Similarly the description of the "Single-sample gene classifier" could be enhanced similarly with a better description of which genes are in the final classifier.

   Read the original source
5. 

   eLife

   Apr 22, 2023

   Reviewer #2 (Public Review):

   This study aimed to classify colorectal cancer (CRC) samples based on the expression of genes in selected gene lists, where the gene lists were chosen to represent aspects of the tumour microenvironment, tumour-associated immune cells, and tumour cells. The resulting clusters were then used to define a classifier, followed by a detailed description of molecular features of the tumours and tumour microenvironments assigned to each cluster. The authors claim this study is more "holistic" than previous work on CRC clustering/classifiers because they aimed to explicitly include additional components of the tumour microenvironment in both the clustering/classifier definition and in the subsequent description of molecular characteristics.

   The CCCRC clustering and the resulting classifier presented in this paper …

   Reviewer #2 (Public Review):

   This study aimed to classify colorectal cancer (CRC) samples based on the expression of genes in selected gene lists, where the gene lists were chosen to represent aspects of the tumour microenvironment, tumour-associated immune cells, and tumour cells. The resulting clusters were then used to define a classifier, followed by a detailed description of molecular features of the tumours and tumour microenvironments assigned to each cluster. The authors claim this study is more "holistic" than previous work on CRC clustering/classifiers because they aimed to explicitly include additional components of the tumour microenvironment in both the clustering/classifier definition and in the subsequent description of molecular characteristics.

   The CCCRC clustering and the resulting classifier presented in this paper are derived from published RNAseq studies. The multi-omics aspect of the work is restricted to smaller sample numbers for which both transcriptomic and another omics dataset were available in public resources and comprises a description or correlative analysis of each omics data type within each of the assigned CCCRC subtypes.

   By applying solid computational methods to a compendium of published RNAseq datasets (n~1500 tumours), they found that tumour samples from colorectal cancers clustered into 4 subtypes ("CCCRC" subtypes) on the basis of 61 pre-defined gene expression signatures. These subtypes correlated with but did not correspond to, the previously described Consensus Molecular Subtypes (CMS) of colorectal tumours.

   Other types of molecular data were available for some tumours, obtained from the same published resources: whole-slide images, mutations, tumour proteomics, and/or scRNAseq. The authors reanalysed these datasets using standard methods and drew correlations with the CCCRC subtypes they had assigned in this work. To (semi-)quantify immune infiltration characteristics from whole-slide images (WSI), they additionally performed automated segmentation in addition to review by pathologists, which in combination produced a convincing WSI-derived dataset.

   In combination with existing CRC classifications, this study could facilitate future biomarker discoveries. This appears to be the authors' main claim, and the data and methods broadly support this claim.

   Some aspects of the work need to be clarified:

   This work relies on the definition of 4 clusters of CRC tumours based on consensus clustering of the 61 gene lists, which in turn depends on the choice of clustering method and the choice of gene lists. Sufficient detail is provided about the gene lists and resulting clusters, but this paper does not show how robust the 4 clusters are to these choices; for example, the "Energy" gene list appears to be a relatively strong component of clusters C2 and C3.

   The authors examined whether their CCCRC classification showed differential disease progression in available retrospective cohorts of people treated with anti-PDL1 therapy. The authors presented this work as "significance of CCCRC in guiding the clinical treatment of colorectal cancer", but the data presented in this section cannot support clinical treatment decisions, which would require prospective studies and clinical trial designs. However, this section is potentially useful for generating hypotheses about potential biomarkers related to the CCCRC subtypes, and might, in the future with additional evidence, contribute to the design of a trial. The authors point out that additional experimental evidence would be required.

   Other prognostic or predictive clinicopathological variables for colorectal cancer are not discussed in detail in the present work but are important for further work on the prognostic and predictive value of CRC molecular subtypes and biomarker derivation. Discrepancies in treatment response have previously been observed in separate CRC trials of biologically targeted agents with different chemotherapy backbones and other authors have suggested that treatment interactions with the tumour microenvironment might in part explain these discrepancies (e.g. Aderka (2019) PMID:31044725, and others).

   Read the original source
6. 

   eLife

   Apr 22, 2023

   Reviewer #3 (Public Review):

   In their study: Comprehensive characterization of tumor microenvironment in colorectal cancer via histopathology-molecular analysis, Wu et al., aim to examine the contribution of the tumour microenvironment (TME) on biological and clinical heterogeneity in colorectal cancer (CRC).

   To achieve this the authors use a vast array of publicly available datasets across a variety of biological modalities (transcriptomic, epigenetic, mutational). Using thoughtfully curated genesets the authors classify CRC into 4 holistic comprehensive characterised CRC (CCCRC) subtypes which comprise immune, stromal, and tumour features of CRC biology.

   The authors investigate the association of their novel CCCRC subtypes with current "gold standard" classification schemes.

   The authors' integration of deep learning methods for HE …

   Reviewer #3 (Public Review):

   In their study: Comprehensive characterization of tumor microenvironment in colorectal cancer via histopathology-molecular analysis, Wu et al., aim to examine the contribution of the tumour microenvironment (TME) on biological and clinical heterogeneity in colorectal cancer (CRC).

   To achieve this the authors use a vast array of publicly available datasets across a variety of biological modalities (transcriptomic, epigenetic, mutational). Using thoughtfully curated genesets the authors classify CRC into 4 holistic comprehensive characterised CRC (CCCRC) subtypes which comprise immune, stromal, and tumour features of CRC biology.

   The authors investigate the association of their novel CCCRC subtypes with current "gold standard" classification schemes.

   The authors' integration of deep learning methods for HE classification and subsequent association with "Tumor level" CCCRC subtypes is a refreshing addition to the study. Comment on the degree of heterogeneity observed in HE samples and correlation to the heterogeneity of CCCRC subtypes would be a welcomed addition. It is likely publicly available datasets from such platforms as 10X Genomic Visium would be available for this type of analysis.

   Whilst one of the main outcomes of the study is the addition of another classification scheme to the field of colorectal cancer, the CCCRC scheme represents a holistic perspective on CRC classification.

   The authors provide a welcomed graphical overview of the complex narrative of the study in Figure 7.

   The authors focus on the classification of inter-patient heterogeneity and its associated predictive and prognostic utility. There appears to be a significant degree of overlap between immunosuppressive and immune excluded, and proliferative and immuno-modulatory signatures in Figure 1A. One of the major limitations of patient response to treatment is intra-patient heterogeneity, it would be nice for the authors to comment briefly on the degree of intra-patient heterogeneity of the CCCRC subtypes.

   Overall the authors succeed in providing a holistic deep characterisation of CRC from the perspective of a variety of biological modalities. The authors provide a novel classification scheme for the field of CRC which demonstrates prognostic and predictive utility, which would benefit from further validation from external datasets. The authors demonstrate a pathway for integration and interpretation of complex high-dimensional data into clinically translatable currency such as the H&E.

   Read the original source
7. 

   Version published to 10.1101/2023.02.16.528849v1 on bioRxiv

   Feb 20, 2023