Author response: Comprehensive characterization of tumor microenvironment in colorectal cancer via molecular analysis

Abstract

Full text Figures and data Side by side Abstract Editor's evaluation Introduction Results Discussion Materials and methods Data availability References Decision letter Author response Article and author information Metrics Abstract Colorectal cancer (CRC) remains a challenging and deadly disease with high tumor microenvironment (TME) heterogeneity. Using an integrative multi-omics analysis and artificial intelligence-enabled spatial analysis of whole-slide images, we performed a comprehensive characterization of TME in colorectal cancer (CCCRC). CRC samples were classified into four CCCRC subtypes with distinct TME features, namely, C1 as the proliferative subtype with low immunogenicity; C2 as the immunosuppressed subtype with the terminally exhausted immune characteristics; C3 as the immune-excluded subtype with the distinct upregulation of stromal components and a lack of T cell infiltration in the tumor core; and C4 as the immunomodulatory subtype with the remarkable upregulation of anti-tumor immune components. The four CCCRC subtypes had distinct histopathologic and molecular characteristics, therapeutic efficacy, and prognosis. We found that the C1 subtype may be suitable for chemotherapy and cetuximab, the C2 subtype may benefit from a combination of chemotherapy and bevacizumab, the C3 subtype has increased sensitivity to the WNT pathway inhibitor WIKI4, and the C4 subtype is a potential candidate for immune checkpoint blockade treatment. Importantly, we established a simple gene classifier for accurate identification of each CCCRC subtype. Collectively our integrative analysis ultimately established a holistic framework to thoroughly dissect the TME of CRC, and the CCCRC classification system with high biological interpretability may contribute to biomarker discovery and future clinical trial design. Editor's evaluation This study represents a valuable body of work in which the authors assemble a molecular description of colorectal cancer and a classification into subtypes. Overall, the evidence supporting the findings is solid, and 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. https://doi.org/10.7554/eLife.86032.sa0 Decision letter Reviews on Sciety eLife's review process Introduction Colorectal cancer (CRC) is the third most deadly malignancy worldwide (Siegel et al., 2023), and the incidence of early-onset CRC is steadily increasing (Archambault et al., 2021). CRC at early and localized stages is primarily a preventable and curable disease, but up to 50% of patients with locally advanced disease eventually develop mCRC (Andrei et al., 2022; Ciardiello et al., 2022). Therefore, the clinical systematic management of CRC patients is still an unmet medical challenge (Ciardiello et al., 2022). With the development of high-throughput technologies and bioinformatics strategies, multi-omics data are used to identify and characterize the molecular subtypes of CRC, such as genomics (Zhao et al., 2022), transcriptomics (Budinska et al., 2013; De Sousa E Melo et al., 2013; Marisa et al., 2013; Roepman et al., 2014; Sadanandam et al., 2013; Schlicker et al., 2012), and proteomics (Li et al., 2020). The consensus molecular subtype (CMS) integrates six independent classification systems based on transcriptomics; however, it is still not explicitly used to guide clinical treatment (Guinney et al., 2015). The Cancer Genome Atlas Program (TCGA) and Clinical Proteomic Tumor Analysis Consortium (CPTAC) colorectal studies have dissected the molecular heterogeneity of CRC by integrating multi-omics data (Cancer Genome Atlas Network, 2012; Vasaikar et al., 2019). Nevertheless, multi-omics data are complex and highly dimensional, and extracting valuable information from these data to guide clinical treatment is still a tremendous challenge (Leng et al., 2022). By reviewing the biological characteristics of the tumor, useful information can be screened for identifying molecular subtypes. The tumor cells can interact with cellular or non-cellular components, triggering dramatic molecular, cellular, and physical changes in the tumor microenvironment (TME) to build a self-sustainable tumor ecosystem (Anderson and Simon, 2020; Chen and Song, 2022). Simultaneously, TME profoundly affects tumor biology, responses to therapy, and clinical outcomes, which is a dynamic network mainly comprised of immune components and stromal components (Hirata and Sahai, 2017; Jia et al., 2022; Zhang et al., 2022). Furthermore, TME can adversely affect the metabolic activities of tumor, immune and stromal cells, and form diverse metabolic phenotypes (Elia and Haigis, 2021; Kaymak et al., 2021). Identifying the components of the TME and their functions, as well as the crosstalk between tumor cells and TME contributes to our understanding of the clinical heterogeneity of CRC, thereby bringing about new advances in precision medicine. Previous studies have used immune or stromal components of the TME, or a combination of both, to study the TME (Bagaev et al., 2021; He et al., 2018), but they are insufficient to completely reconstruct the heterogeneity of the TME. In this study, we considered the tumor cells and its TME as a whole and performed a comprehensive characterization of TME in colorectal cancer (CCCRC), including the functional states of the tumor cells, immune and stromal signatures, and metabolic reprogramming features. We successfully identified the four CCCRC subtypes based on 61 TME-related signatures. Integrated analyses determined that the CCCRC subtypes had distinct histopathologic and molecular characteristics, therapeutic efficacy, and prognosis. Results Establishment of the TME panel The molecular and clinical features of a tumor are characterized by the functional states of tumor cells, as well as other TME-related signatures, including immune and stromal components, and metabolic reprogramming signatures. In brief, 15 signatures (including angiogenesis, apoptosis, cell cycle, differentiation, DNA damage, DNA repair, endothelial-to-mesenchymal transition (EMT), hypoxia, inflammation, invasion, metastasis, proliferation, quiescence, stemness, and cancer stem cells) were used to describe the functional states of tumor cells. As for the immune signatures, we focused on eight categories of immune cells (T cells, natural killer cells, dendritic cells, macrophages, myeloid-derived suppressor cells, B cells, mast cells, and neutrophils) and their subpopulations, as well as the other immune-related signatures. In addition to the signatures of endothelial cells, mesenchymal cells, and mesenchymal stem cells, we included signatures of extracellular matrix, matrix remodeling and interactions of cells with the extracellular matrix to characterize the stromal compartments. A total of seven major metabolic pathways (amino acid, nucleotide, vitamin cofactor, carbohydrate, TCA cycle, energy, and lipid metabolism) were used to reveal the metabolic reprogramming of the TME. According to the above biological framework, a total of 61 TME-related signatures were collected to form the TME panel (Supplementary file 1a, Supplementary file 1b), which ultimately established a holistic approach to thoroughly dissect the TME of CRC. Principal component analysis (PCA) indicated that after using the ‘ComBat’ function to remove batch effects, there was no significant batch effect in the merged cohorts of eight microarray datasets (Figure 1—figure supplement 1A, B) and two RNA sequencing datasets (Figure 1—figure supplement 1C, D, Supplementary file 1c). We used Gene set variation analysis (GSVA) to calculate the TME-related signature scores for each sample in each cohort. Principal coordinate analysis (PCOA) revealed that the CRC samples could be distinguished from normal samples by the TME-related signatures in the GSE39582 and TCGA cohorts (Figure 1—figure supplement 1E). We further focused on the signatures of the functional states of tumor cells, which could classify CRC and normal samples (Figure 1—figure supplement 1F). The p-values for intercomparisons of the ‘Euclidean’ distances between normal and CRC samples were all <0.05 using permutational multivariate analysis of variance (PERMANOVA) test (Zhu et al., 2021). Most immune signatures had higher GSVA scores in the normal samples compared with the CRC samples, while stromal signatures and the signatures of the functional states of tumor cells had higher GSVA scores in CRC tissues (Figure 1—figure supplement 1G, H). As expected, amino acid, carbohydrate, and nucleotide metabolic processes were more prominent in CRC samples, which was consistent with the hallmark of infinite proliferation of tumor cells (Figure 1—figure supplement 1G, H). Spearman’s correlation analysis of the TME-related signatures revealed three major patterns bound by positive correlations in the CRC-AFFY cohort (Figure 1—figure supplement 1I). One pattern defining the proliferation of tumor cells consisted of cell cycle and metabolic reprogramming signatures. The second was mainly comprised of immune components, such as T cells, natural killer cells (NK cells), myeloid-derived suppressor cells (MDSCs), and M2 macrophages. The third pattern was associated with stromal components such as angiogenesis and extracellular matrix, as well mesenchymal cells and cancer stem cells. In addition, our findings indicated a strong positive correlation between lymphocytic and stromal signatures and MCP-counter algorithm-derived signatures, thereby emphasizing the robustness of the employed methods (Figure 1—figure supplement 1J). Meanwhile, we found that 15 signatures associated with the functional states of tumor cells were positively correlated with the activity of 10 classical oncogenic pathways (Figure 1—figure supplement 1K). Finally, we used the Kaplan–Meier method and univariate Cox proportional hazard regression analysis to evaluate the prognosis of the TME-related signatures. The stromal and tumor components significantly were correlated with decreased survival, particularly in the case of mesenchymal cells, endothelial cells, metastasis, differentiation, and EMT signatures (Figure 1—figure supplement 1L, Supplementary file 1d). Lymphoid-associated cells generally tended to be associated with a better prognosis, while myeloid-associated cells generally tended to be associated with a poor prognosis. Among the metabolism-related signatures, energy metabolism and carbohydrate metabolism were significantly related to poor prognosis, while nucleotide metabolism, amino acid metabolism, and TCA metabolism were strongly predictive of a favorable prognosis. Collectively, our findings demonstrated that the TME heterogeneity, including unique differences in immune, stromal, and metabolic reprogramming, played a crucial role in tumor development, and that the TME panel could be used to comprehensively characterize CRC. Identification and validation of CCCRC classification With the increasing application of immunotherapy and tumor vaccines, there is growing evidence highlighting the importance of the TME in tumorigenesis and development (Bejarano et al., 2021; Saxena et al., 2021). To reveal the TME heterogeneity of CRC using the curated TME panel, consensus clustering analysis was performed based on the 61 TME-related signature scores in the CRC-AFFY cohort, and the optimal cluster number was determined to be four using the consensus matrices heatmap, the empirical cumulative distribution function (CDF) plot, and delta area plot (Wilkerson and Hayes, 2010) (Materials and methods, Figure 1—figure supplement 2A–C). Subsequently, the CRC samples in the CRC-AFFY cohort were classified into the four CCCRCs with distinct TME components (Figure 1A, B, Figure 1—figure supplement 2D). To evaluate the reproducibility of the CCCRC subtypes, we utilized the PAM (Prediction Analysis of Microarrays) algorithm (Tibshirani et al., 2002) to extract 61 TME-related signatures that best represent each subtype, using a threshold of 0.566 (Figure 1—figure supplement 2E). These signatures were then used to construct a PAMR classifier with superior predictive capability, exhibiting an overall error rate of 15%. The PAMR classifier based on the PAM algorithm is publicly available at https://github.com/XiangkunWu/PAMR_classifier (copy archived at Wu, 2023a). 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 (Figure 1—figure supplement 2F, G). PCOA showed that the four CCCRC subtypes were distinctly separated and the p-values for intercomparisons of the ‘Euclidean’ distances between them were all <0.05 using PERMANOVA test (Figure 1—figure supplement 2H) in the CRC-RNAseq and CRC-AFFY cohorts. PCOA also demonstrated highly similar TME compartments in the same subtype between the CRC-RNAseq and CRC-AFFY cohorts. Differences in the TME components between the CCCRC subtypes were also observed in the analysis of previously reported immune and stromal signatures obtained by the microenvironment cell populations (MCP)-counter, CIBERSORT, and ESTIMATE algorithm, and 10 classical oncogenic pathway activities and 86 metabolic pathway enrichment scores calculated by GSVA (Figure 1—figure supplement 3A–E, Supplementary file 1e). Notably, 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 (Figure 1—figure supplement 4). These sensitivity analyses underscored the necessity of our well-designed TME panel to achieve this identification. Figure 1 with 7 supplements see all Download asset Open asset Comprehensive characterization of colorectal cancer (CCCRC). (A) Heatmap of 1471 colorectal cancer (CRC) patients in the CRC-AFFY cohort classified into four distinct tumor microenvironment (TME) subtypes based on the 61 TME-related signatures. CMS: consensus molecular subtypes; MSI: microsatellite instability; MSS: microsatellite stability. (B) Radars display the characteristic TME-related signatures, including tumor, immune, stroma, and metabolism signatures, of each CCCRC subtype in the CRC-AFFY cohort. MSC: mesenchymal stem cells; ECM: extracellular rmatrix. Box plots show differences in tumor (C), immune (D), and stroma (E) signatures in the CRC-AFFY cohort. Tumor purity and stroma scores were obtained from the ESTIMATE algorithm. Proliferative activity (proliferation), cytolytic score, M1 and M2 macrophage proportions, and TGFB activity were calculated by Gene set variation analysis (GSVA). (F) Differences in intratumor heterogeneity between four CCCRC subtypes in CRC-RNAseq cohort. (G) Differences in T cell dysfunction and T cell exclusion scores between four CCCRC subtypes were analyzed based on the gene expression profiles in CRC-AFFY cohort. (H) Gene set enrichment analysis (GSEA) of the terminally exhausted CD8+ T cell signature and the TGF-beta signaling signature between C2 and C4 subtypes in the CRC-AFFY cohort. (I) Kaplan–Meier method with log-rank test of disease-free survival among the four CCCRC subtypes in the CRC-AFFY cohort. C1 (35% of all tumors), hereafter designated as the proliferative subtype, was characterized by the relative upregulation of tumor proliferative activity, tumor purity, and minimal or complete lack of lymphocyte and stromal infiltration, which was highly similar to the cold tumor phenotype (Figure 1B–E). The MYC, cell cycle, TP53, and PI3K pathways associated with tumor proliferation had the highest GSVA scores in the C1 subtype (Figure 1—figure supplement 3E). C2 (21% of all tumors), hereafter designated as the immunosuppressed subtype, was characterized by the relative upregulation of immune and stromal components, such as T cells, M2 macrophages, and cancer-associated fibroblasts (CAFs) (Figure 1B–E, Figure 1—figure supplement 3A–D). However, the extent of infiltration of effector cells, as well as the cytolytic score, was much lower than that of the C4 subtype. C3 (24% of all tumors), hereafter designated as the immune-excluded subtype, was characterized by the distinct upregulation of stromal components, such as CAFs, and cancer stem cells, as well as angiogenesis and hypoxia signatures (Figure 1B–E, Figure 1—figure supplement 3A–D). During tumor progression, TGF-beta secreted by CAFs is leveraged by tumor cells to suppress and exclude the anti-tumor immune components (Liu et al., 2019). We observed that the TGF-beta pathway, as well as WNT, NOTCH, and RAS pathways, and the ratio of M2/M1 macrophages, were distinctly upregulated in C2 and C3 subtypes (Figure 1D, E, Figure 1—figure supplement 3E). The scores of 5/10 oncogenic pathways were the highest in the C3 subtype (Figure 1—figure supplement 3E), suggesting that the activation of oncogenic pathways could lead to the formation of immune-excluded phenotypes which was consistent with the previous theory (Galon and Bruni, 2019). The level of intratumor heterogeneity (ITH) was significantly linked to poor prognosis and drug resistance (Caswell and Swanton, 2017). As expected, the ITH of the C2 and C3 subtypes was higher than that of the other subtypes (Figure 1F). C4 (20% of all tumors), hereafter designated as the immunomodulatory subtype, was characterized by the remarkable upregulation of anti-tumor immune components, such as effector T cells, NK cells, and Th1 cells. The C4 subtype also had the highest cytolytic score compared with the other subtypes and lacked stromal components and the other immunosuppressed components, which indicated an immunomodulatory microenvironment (Figure 1B–E). To further explore the immune escape mechanism of each CCCRC subtype, the differences in T cell dysfunction and T cell exclusion scores between the four CCCRC subtypes were analyzed based on the gene expression profiles (GEP), which reflected the T cell function of the global tumor (Jiang et al., 2018). Strikingly, the C2 subtype had highest T cell dysfunction score, indicating that T cell exhaustion in the C2 subtype was at the late stage (Figure 1G, Figure 1—figure supplement 3F). Using gene set enrichment analysis (GSEA) with all genes ranked according to the fold change (FC) between C2 and C4 subtypes, we found that terminally exhausted CD8+ T cell and TGF-beta signaling signatures were upregulated in the C2 subtype in the CRC-AFFY (Figure 1H) and CRC-RNAseq (Figure 1—figure supplement 3G) cohorts, which might reveal that CD8+ T cell infiltration within the tumor bed was suppressed by the stroma and was in a late state of exhaustion. The C3 subtype had the highest T cell exclusion score (Figure 1G, Figure 1—figure supplement 3F), demonstrating that the low T cell infiltration into the tumor bed was due to the increased abundance of CAFs and M2 macrophages, thereby leading to the exclusion of T cells from the tumor bed. Metabolic reprogramming also differed significantly among the four CCCRC subtypes (Figure 1B, Figure 1—figure supplement 3H). We analyzed the 86 metabolic pathways obtained from the Kyoto Encyclopedia of Genes and Genomes (KEGG) database (Supplementary file 1e) and observed that the number of upregulated metabolic pathways of the C3 subtype was the lowest. We also found that glycan metabolism was distinctly upregulated in C2 and C3 subtypes, which indicated that glycan metabolism was significantly associated with the stroma. Associations between CCCRC subtypes and other molecular subtypes and clinical characteristics Previous studies have identified several molecular subtypes of CRC based on GEP (Budinska et al., 2013; De Sousa E Melo et al., 2013; Guinney et al., 2015; Roepman et al., 2014; Sadanandam et al., 2013). We investigated their associations with the CCCRC subtypes in the CRC-AFFY and CRC-RNAseq cohorts (Figure 1—figure supplement 5). The C1 subtype was primarily comprised of the CMS2 and lower crypt-like subtypes, and it contained the highest frequencies of the CCS1, B-type, and TA subtypes. The C2 subtype mainly consisted of the CMS4, stem-like, surface crypt-like, CCS3, and C-type subtypes, and included the highest frequency of the enterocyte subtype. The C3 subtype contained the highest frequencies of CMS4, stem-like, CCS3, and C-type subtypes and was mainly comprised of the mesenchymal and TA subtypes. The C4 subtype included the highest frequencies of microsatellite instability (MSI) and the CMS1, CIMP-H-like, A-type, and inflammatory subtypes, and was mainly comprised of the CCS2 subtype. We also focused on the differences in the TME components between the CCCRC and the CMS subtypes. Compared with the CMS1 subtype, the C4 subtype showed upregulated anti-tumor immune components and lacked immunosuppressive components (Figure 1—figure supplement 6A). CRC patients with MSI were sensitive to immune checkpoint blockade (ICB) therapy (Jin and Sinicrope, 2022), and C4 and CMS1 subtypes containing approximately 47 and 75% of MSI cases, respectively. The C4 subtype with MSI showed upregulated scores of effector cells and cytolytic activity and downregulated scores of extracellular matrix and matrix remodeling compared with the CMS1 subtype with MSI (Figure 1—figure supplement 6B). Moreover, we observed that the C4 subtype with MSI and the C4 subtype with MSS had higher scores of anti-tumor immune signatures and lower scores of stromal components than the other subtype with MSI (Figure 1—figure supplement 6C). We also observed that CMS2 subtype contained more C4 subtypes in addition to mainly C1 subtypes. Therefore, we analyzed the differences in the TME components between C1 and CMS2 subtypes and found that CMS2 subtypes indeed had higher immune-related components than C1 subtypes, such as MCH-I, MCH-II, and inflammatory signature, and also contained more stromal components, such as extracellular matrix, than C1 subtypes (Figure 1—figure supplement 6D). This suggested that the C1 subtype was less immunogenic than CMS2 and more closely resembles cold tumor characteristics. Specifically, we found that CMS4 contained mainly C2 and C3 subtypes. Our findings indicated that the C2 subtypes within CMS4 exhibited a higher abundance of immune components, such as T cells and NK cells, compared to the C3 subtype within CMS4. However, the differences in stromal components between these subtypes were not statistically significant (Figure 1—figure supplement 6E). 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 was 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. Differences in histopathologic characteristics between CCCRC subtypes To further explore the biological differences between CCCRC subtypes, we investigated the histopathologic phenotypes by evaluating the whole-slide images (WSIs) of the TCGA-CRC dataset. We compared our CCCRC system with the three-category immune classification system of solid tumors, termed ‘desert’, ‘excluded’, and ‘inflamed’ phenotypes (Chen and Mellman, 2017; Rosenberg et al., 2016). Two pathologists evaluated the histopathologic characteristics for each subtype under the microscope. The 254 CRC samples in the TCGA-CRC dataset were categorized as these three phenotypes based on the abundance of lymphocytes and their spatial location with malignant epithelial cells (Supplementary file 1i). According to the three-category immune classification system, the C4 subtype was enriched with an inflamed phenotype characterized by abundant lymphocytes in direct contact with malignant cells (Figure 2A). The C2 subtype was mostly categorized as an excluded phenotype. The C1 and C3 subtypes were mainly classified into the desert phenotype, whereas the C3 subtype was more frequently classified as an excluded phenotype than the C1 subtype. Notably, the lymphocytes of C2 subtype were more frequently intermixed with intratumor stromal components, whereas the lymphocytes of C3 subtype were more frequently excluded from the tumor bed and intermixed with adjacent-tumor stromal components, both of which were classified as excluded phenotype according to the three-category immune classification system. Figure 2 with 1 supplement see all Download asset Open asset Differences in histological characteristics between comprehensive characterization of colorectal cancer (CCCRC) subtypes. (A) Sankey plot shows overlap of CCCRC subtypes with the three-category immune classification system (‘desert’, ‘excluded’, and ‘inflamed’ phenotypes), and their representative hematoxylin and eosin (HE)-stained whole-slide images (WSIs). C1: TCGA-AA-3955; C2: TCGA-A6-6654; C3: TCGA-CK-4948; and C4: TCGA-AD-6963. (B) Representative WSI (top) and the colorectal cancer (CRC)-multiclass model-inference segmentation of eight tissue types: tumor, stroma, lymphocyte, normal colon mucosa, debris, adipose, mucin, and muscle (bottom). Box plots show differences in the abundance of tumors (C), lymphocyte infiltration (lym) (D), and stroma (E) in the core tumor (CT) region. Box plots show differences in the lymphocyte infiltration to tumor content ratio (F) and lymphocyte infiltration to stromal content ratio (G) in the CT region. Box plots show differences in the abundance of lymphocytes infiltration (H) and stroma (I) in the invasive margin (IM) region. (J) Box plots show differences in the ratio of lymphocyte infiltration in the IM region to the CT region. The above differences in the histopathologic characteristics among the CCCRC subtypes were based on the semi-quantitative analysis results of two pathologists, which are subjective to a certain extent. Therefore, we used hematoxylin and eosin (HE)-stained image-based deep learning to establish the CRC-multiclass model and to evaluate the abundance and spatial distribution of the tumor, lymphocytes, and stroma. The CRC-multiclass model consisted of a muscle/non-muscle classifier and a seven-tissue classifier that could classify eight CRC tissue types: adipose (ADI), debris (DEB), lymphocytes (LYM), mucus (MUC), smooth muscle (MUS), normal colon mucosa (NORM), cancer-associated stroma (STR), and colorectal adenocarcinoma epithelium (TUM) (Materials and methods). The muscle/non-muscle model demonstrated well performance on the internal test set of 14,681 patches, achieving an area under the curve (AUC) of approximately 0.99 and an accuracy of 0.99 (Figure 2—figure supplement 1A, B). Meanwhile, the seven-tissue classifier also performed well, with AUCs for the different tissue types above 0.99 and an accuracy of 0.95 on the internal test set of 5741 patches (Figure 2—figure supplement 1C, D). When evaluated on an external test set of 4288 patches, the muscle/non-muscle model achieved an AUC of 0.95 and an accuracy of 0.91 (Figure 2—figure supplement 1E, F). Of the 3633 patches identified as non-muscle by the muscle/non-muscle model, the seven-tissue classifier achieved AUCs ranging from 0.97 to 1 for different tissue types and an accuracy of 0.91 (Figure 2—figure supplement 1G, H). The tissue heatmap showed o