Colorectal cancer treatment effect prediction system based on multiple immunohistochemistry

Abstract

The invention relates to a colorectal cancer treatment effect prediction system based on artificial intelligence and multiple immunohistochemistry, which comprises: (i) an analysis selection module for analyzing tumor tissues and selecting target cells and / or target antigens; (ii) a preparation module for preparing a colorectal cancer immunosuppressive unit (CCIM) image from the selected target antigen by using a multiplex immunohistochemical technique; and (iii) an identification module that predicts a therapeutic effect of colorectal cancer based on the CCIM image.

Description

Technical Field

[0001] The present invention relates to the field of life sciences, and in particular to a colorectal cancer treatment effect prediction system based on artificial intelligence and multiple immunohistochemistry. The present invention also relates to the use of a composition in preparing a kit for predicting the treatment effect of colorectal cancer and the use of a composition containing a therapeutic agent targeting FOLR2+ resident macrophages in preparing a drug for improving the chemotherapy effect of colorectal cancer patients. Background Art

[0002] Colorectal cancer (CRC) is one of the malignant tumors with high morbidity and mortality in clinical practice. Its morbidity ranks third among malignant tumors in the world and its mortality ranks second, which seriously threatens human health. In 2018 alone, there were 1.8 million new cases of colorectal cancer worldwide, and nearly 860,000 people died of colorectal cancer (Bray F, Ferlay J, etc., Global cancer statistics 2018). In the past 30 years, the mortality rate of CRC has slightly decreased due to early diagnosis and the development of chemotherapy, targeted therapy and immunotherapy. FOLFOX (a combination of 5-fluorouracil, leucovorin and oxaliplatin) and FOLFIRI (a combination of 5-fluorouracil, leucovorin and irinotecan) are the first-line chemotherapy regimens for advanced CRC. Cetuximab and bevacizumab have shown good efficacy, mainly in combination with the FOLFIRI regimen or the FOLFOX regimen (Alberts SR et al., JAMA, 2012, 307(13):1383-93; de Gramont A et al., The Lancet Oncology, 2012, 13(12):1225-33). In addition, CRC patients with high microsatellite instability (MSI-H) or mismatch repair deficiency (dMMR) are sensitive to immunotherapy based on immune checkpoint inhibitors (ICI) and have objective clinical responses (Dudley JC et al., Clinical Cancer Research. 2016, 22(4):813-20), providing an alternative treatment for subgroups of patients with advanced CRC.

[0003] Nevertheless, heterogeneous treatment responses have been found in CRC patients. A portion of dMMR patients did not respond to pembrolizumab treatment, prompting further exploration of the mechanisms behind chemotherapy and immunotherapy resistance (André T et al., New England Journal of Medicine. 2020, 383(23): 2207-18), which may be due to the heterogeneity of the tumor microenvironment (TME). The TME consists of unique and interacting cell populations, including myeloid cells, T cells, B cells, NK cells, and other cell types. Studies on the diversity and reprogramming of TME have shown its potential to influence immunotherapy response and prognosis (Bao X et al., Immunotherapy.2021, 70(1):189-202; Liu L et al., Immunotherapy.2021:1-16; Bao X et al., Journal for immunotherapy of cancer, 2020:8(2); DeBerardinis RJ et al., New England Journal of Medicine; 2020, 382(9):869-71; Frankel T et al., Tumor Immune Microenvironment in Cancer Progression and Cancer Therapy, 2017:51-64; Chen D et al., British Journal of Cancer, 2021:1-9; and Bao X et al., Cancer Immunology Research, 2022). The development of suppressive TME limits the response to chemotherapy and immunotherapy, and is characterized by increased levels of tumor-promoting macrophages, exhausted T cells, and regulatory T cells (Tregs). In various types of tumors, high levels of macrophages are associated with poor prognosis. Targeting macrophages is a promising new strategy for cancer immunotherapy (Geeraerts X et al., Frontiers in immunology, 2017, 8: 289; Barkal AA et al., Nature, 2019, 572(7769): 392-6). The colony stimulating factor 1 (CSF1) / colony stimulating factor 1 receptor (CSF1R) axis has received the most attention in the context of macrophage-targeted therapy. A variety of monoclonal antibodies and small molecules targeting CSF1R or its ligand CSF1 have been studied in phase I as monotherapy or in combination with chemotherapy and other cancer immunotherapies. However, anti-CSF1R therapy has shown limited therapeutic efficacy and significant toxicity in several tumor types, including CRC, which warrants the exploration of the mechanisms behind resistance to therapies targeting macrophages.

[0004] Tissue-resident macrophages are mononuclear phagocytes that play a key role in host defense and tissue homeostasis (Li X et al., Nature Immunology, 2021: 1-2). In tumor tissue, tissue-resident macrophages reshape the TME, thereby supporting early tumor progression (Etzerodt A et al., Journal of Experimental Medicine, 2020, 217(4); Lahmar Q et al., Biochimica et Biophysica Acta (BBA)-Reviews on Cancer, 2016, 1865(1): 23-34). Increased expression of MHC-II, CCL17, and TGF-β in tissue-resident macrophages supports their chemoattraction of CCR4-expressing T cells and induces naive CD4 + The ability of T cells to differentiate into Treg cells (Casanova-Acebes M et al., Nature, 2021: 1-7). Sharma et al. revealed that FOLR2+ resident macrophages play a crucial role in the immunosuppressive tumor-fetal ecosystem of hepatocellular carcinoma (Sharma A et al., Cell, 2020, 183(2): 377-94.e21).

[0005] The CRC tumor microenvironment has been well characterized at single-cell resolution (Bian S et al., Science, 2018, 362(6418):1060-3; Roerink SF et al., Nature, 2018; 556(7702):457-62). However, previous studies have mostly focused on the identification of different cell populations. Spatial interaction maps of CRC TME are still unavailable. Moreover, the generation of high-throughput batch and single-cell transcriptome data usually relies on fresh tumor tissue rather than paraffin sections. Transcriptome data-driven patient stratification is not pathology-friendly. Pathological methods are also not well established for prognosis and treatment sensitivity prediction in CRC patients. Finally, adverse reactions to immune checkpoint inhibitors in pMMR (mismatch repair normal) patients also require further exploration of new immunotherapy targets. Summary of the invention

[0006] In order to solve the above problems, the present invention integrates multi-omics data (bulk transcriptome, single-cell transcriptome and multiple immunohistochemical staining) to reveal the immune spatial interaction map of CRC. The present invention found that the colorectal cancer immunosuppression module (CCIM) mediated by FOLR2+ resident macrophages can coordinate the inhibitory CRC TME and lead to treatment resistance in CRC. The present invention designs a multicolor (e.g., 7-color) CCIM multiple immunohistochemistry (mIHC) plate, applies a deep learning model to the CCIMmIHC plate, and develops a colorectal cancer treatment effect prediction system for artificial intelligence (AI)-assisted diagnosis, which has the advantages of high efficiency, pathology-friendly and cost-effectiveness. In addition, the present invention also evaluates the possibility of multi-cell modules as new immunotherapy targets based on this system.

[0007] Specifically, the present invention solves the technical problems existing in the prior art through the following technical solutions:

[0008] 1. A colorectal cancer treatment effect prediction system, comprising:

[0009] (i) an analysis and selection module, which analyzes the sample of tumor tissue and selects target cells and / or target antigens;

[0010] (ii) a preparation module for preparing a colorectal cancer immunosuppressive unit (CCIM) image using a multiplex immunohistochemical technique with a selected target antigen; and

[0011] (iii) A recognition module, which predicts the treatment effect of colorectal cancer based on the CCIM image.

[0012] 2. The system according to item 1, wherein the analysis selection module comprises an analysis component selected from the group consisting of:

[0013] (a) Single-cell RNA sequencing (scRNA-seq) components;

[0014] (b) Bulk transcriptome analysis component;

[0015] (c) clinical information collection component;

[0016] (d) Weighted Gene Co-expression Network Analysis (WGCNA) component;

[0017] (e) multiplex immunohistochemistry (mIHC) staining kit; and

[0018] (f) Tandem mass tag (TMT)-based proteomic analysis components.

[0019] 3. A system according to item 1 or 2, wherein the TME heterogeneity and molecular characteristics of the sample are obtained by the analysis selection module, and such characteristics can be used to predict the treatment effect and / or prognosis of colorectal cancer.

[0020] 4. A system according to any one of items 1-3, wherein the analysis selection module determines that the cell type used for the preparation module is selected from the following group: CD4 T cells, epithelial cells, CD8 T cells, B cells, myeloid cells, fibroblasts, NK cells, natural lymphocytes and plasma cells.

[0021] 5. A system according to item 4, wherein the analysis selection module further includes (g) a component for further subdividing the CD4 T cells, epithelial cells, CD8 T cells, B cells, myeloid cells, fibroblasts, NK cells and natural lymphocytes, plasma cells.

[0022] 6. A system according to clause 5, wherein the component (g) determines:

[0023] (1) The myeloid cells are selected from the group consisting of dendritic cells, mast cells, FCN1 + Monocyte-like cells, OAS + Macrophages, OLR1 + Macrophage, peptidase inhibitor 3 + Monocytes and FOLR2 + Resident macrophages, classical monocytes (CD14+), non-classical monocytes (CD16+), neutrophils (CD15 or MPO, CD33, CD66B), infiltrating macrophages (CD11b+\CCR2+), SPP1+ macrophages, LAMP3+ dendritic cells;

[0024] (2) the CD4 T cells are selected from the group consisting of exhausted CD4+T cells, effector CD4+T cells, effector memory CD4+T cells, central memory CD4+T cells, naive CD4+T cells, Th17 cells and Treg cells; and / or

[0025] (3) The CD8 T cells are selected from the group consisting of: + T cells, effector CD8+ T cells, effector memory CD8 + T cells, central memory CD8 + T cells, depletion of CD8 + T cells, NKT cells, and tissue-resident CD8 + T cells.

[0026] 7. A system according to clause 5 or 6, wherein the component (g) determines:

[0027] (1) the target antigens for the myeloid cells include one or more selected from the group consisting of CD68, FOLR2, C1QA, C1QB, SPP1, APOE, CD11B, CD172A, S100A8, S100A9, S100A12, CD15, CD33, CD66B, MPO, CD11C, HLA-DR, CD14, CD141, CD1C, CLEC9A, KIT, CPA3, and TPSB2;

[0028] (2) the target antigens for the CD4 T cells include one or more selected from the group consisting of FOXP3, CCR7, SELL, GIMAP family, PD-1, TIGIT, CD3, CD7, CD4, IL7R, GRANZYME B, CXCR3, CX3CR1, TIM3, VISTA, LAG3, GITR, CD137; and / or

[0029] (3) The target antigens for the CD8 T cells include one or more selected from the following group: GZMK, GIMAP4, LAG3, NCR3, S100A4, DKK3, TMIGD2, PD-1, CD3, CD7, CD8, IL7R, GRANZYME B, CXCR3, CX3CR1, TIM3, VISTA, LAG3, GITR and CD137.

[0030] 8. A system according to any one of items 1-7, wherein the analysis selection module determines that the target antigens used for the preparation module include one or more selected from the following group: DKK3, FOLR2, FOXP3, PD1, CD8, CD4 and DAPI.

[0031] 9. A system according to any one of items 1-8, wherein the multiplex immunohistochemistry technique is selected from the following group: tyramide signal amplification (TSA), cyclic immunofluorescence (CyCIF), indexed co-detection (CODEX), multiplexed ion beam imaging (MIBI), imaging mass spectrometry (IMC), OPAL imaging, multiple error robust fluorescence in situ hybridization (MERFISH), stochastic optical reconstruction microscopy (STORM) and Exchange-PAINT imaging.

[0032] 10. The system according to any one of items 1 to 9, wherein the identification module comprises a component selected from the group consisting of:

[0033] (A) Components used to predict the half inhibitory concentration (IC50) of tumor tissue;

[0034] (B) Components used to construct the CCIM scoring system;

[0035] (C) a component for predicting the effects of chemotherapy, radiotherapy, immunotherapy, and targeted therapy by analyzing the spatial relationship between myeloid cells and T cells in the CCIM image;

[0036] (D) a component for predicting prognosis; and

[0037] (E) Components used to predict metastasis.

[0038] 11. A system according to any one of items 1-10, wherein the sample of the tumor tissue is a section of the tumor tissue, preferably a paraffin section.

[0039] 12. Use of a composition in the preparation of a kit for predicting the therapeutic effect of colorectal cancer, wherein the composition comprises a target antigen selected from the group consisting of:

[0040] (1) the target antigens for the myeloid cells include one or more selected from the group consisting of CD68, FOLR2, C1QA, C1QB, SPP1, APOE, CD11B, CD172A, S100A8, S100A9, S100A12, CD15, CD33, CD66B, MPO, CD11C, HLA-DR, CD14, CD141, CD1C, CLEC9A, KIT, CPA3, and TPSB2;

[0041] (2) the target antigens for the CD4 T cells include one or more selected from the group consisting of FOXP3, CCR7, SELL, GIMAP family, PD-1, TIGIT, CD3, CD7, CD4, IL7R, GRANZYME B, CXCR3, CX3CR1, TIM3, VISTA, LAG3, GITR, CD137; and / or

[0042] (3) The target antigens for the CD8 T cells include one or more selected from the following group: GZMK, GIMAP4, LAG3, NCR3, S100A4, DKK3, TMIGD2, PD-1, CD3, CD7, CD8, IL7R, GRANZYME B, CXCR3, CX3CR1, TIM3, VISTA, LAG3, GITR and CD137.

[0043] 13. The use according to item 12, wherein the composition comprises a target antigen selected from the group consisting of DKK3, FOLR2, FOXP3, PD1, CD8, CD4 and DAPI.

[0044] 14. Use of a composition comprising a therapeutic agent targeting FOLR2+ resident macrophages in the preparation of a medicament for improving the chemotherapy effect in patients with colorectal cancer, preferably the therapeutic agent targeting FOLR2+ resident macrophages comprises an anti-CSF1R therapeutic agent.

[0045] In order to make the technical solution of the present invention clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1A -C. Description of the workflow for deep learning-assisted multi-omics analysis. (A) Schematic diagram of this study. (B) T-SNE plot identified the main cell types in CRC. (C) Expression of different cell types and markers in CRC. T-SNE: T-distributed random neighbor embedding; CRC: colorectal cancer; ILC: innate lymphoid cell.

[0047] Figure 2A -H. Characteristics of different monocyte / macrophage subsets that affect DFS in CRC patients. (A) T-SNE showing the identification of 7 subclusters of myeloid cells. (B) Dot plot showing the percentage of cells expressing immune checkpoint and escape genes in each cell type. (C) Pseudo-time analysis of the five subclusters of monocytes / macrophages. (D) Pseudo-time analysis with pseudo-time scores of the five subclusters of monocytes / macrophages. The color key from dark blue to light blue represents the pseudo-time scores from low to high. (E) TF activity in monocyte / macrophage subsets. (F) Heat map of SSGSEA scores of gene sets in monocyte / macrophage subtypes calculated from scRNA-seq data. (G) Heat map of M1-related and M2-related gene expression from scRNA-seq data. (H) Correlation of DFS with enrichment of each myeloid cell type. Gene signatures of myeloid cells were defined as characteristic genes of each cell type and obtained from scRNA-seq data (pTS n=1673). Cell abundance was calculated from the gene signature of each cell type based on bulk transcriptome data using the GSVA method. T-SNE: t-distributed stochastic neighbor embedding; DFS: disease-free survival; scRNA-Seq: single-cell RNA sequencing; GSVA: gene set variation analysis; PTS: patient-specific; TF: transcription factor.

[0048] Figure 3A-F. Identification of low inflammatory properties of resident macrophages. (A) Enriched GO terms for each monocyte / macrophage population. (B) Heat map showing significant signature pathway activity in monocyte and macrophage populations. Color key red indicates importance. (C) UMAP map identified 5 subclusters of monocyte and macrophage populations. (D) Density map representing the density distribution of inflammatory response activity scores in monocyte and macrophage populations and box plot representing inflammatory response activity scores in monocyte and macrophage populations. Inflammatory response activity scores were calculated using inflammatory response signals using the SSGSVA method. P values ​​were calculated using the Kruskal-Wallis test. (E) mIHC staining of CD68 (green), PTGS2 (purple), FOLR2 (red) and DAPI (blue). Scale bar = 10 μm. (F) The proportion of FOLR2+CD68+ cells and FOLR2-CD68+ cells expressing PTGS2. Significance was assessed by chi-square test (*P<0.05, **P<0.01, and ***p<0.001). GO: Gene Ontology; UMAP: Uniform Manifold Approximation and Projection; MIHC: Multiplex Immunohistochemistry.

[0049] Figure 4A -GT cell subsets show different gene expression patterns and pathway activation levels. (A) t-SNE plot identifies five subclusters of CD4+ T cells. (B) t-SNE plot identifies five subclusters of CD8+ T cells. (C) Heatmap shows gene expression in key pathways that regulate T cell activity. Color keys from blue to red indicate relative expression levels from low to high. (D) Heatmap representation of GSVA results based on immune signaling C7. Color keys from purple to yellow indicate relative GSVA scores from low to high. (E) Trajectory inference of five subclusters of CD8+ T cells. (F) Trajectory inference using pseudo-time scores for five subclusters of CD8+ T cells. Color keys from black to gold indicate pseudo-time scores from low to high. (G) Trajectory inference of PD-1 expression for five subclusters of CD8+ T cells. Color keys from blue to white indicate PD-1 expression from low to high. t-SNE: t-distributed random neighbor embedding; CRC: colorectal cancer; GSVA: gene set variation analysis.

[0050] Figure 5A-K. Construction of the colorectal cancer immunosuppression module (CCIM). (A) CCI network of all cell subsets in CRC. The color of the edge indicates the type of ligand cell, and the size of the edge indicates the number of CCIs between cells. (B) The number of interactions between monocytes / macrophages and CD4+T cells inferred from scRNA-seq. (C) The number of interactions between monocytes / macrophages and CD8+T cells. (D) Construction of the CCIM module. (E) Heat map of cell type-specific receptor-ligand interactions inferred by Single Cell SignalR. The inferred interactions between FOLR2+ macrophages and Tregs, exhausted CD4+T cells, exhausted CD4+T cells, and tolerant CD8+T cells are shown. The color of the circle indicates the inferred interaction score, and the size of the circle indicates the average expression of each pair of receptor and ligand genes. (F) CCIM score stratifies CRC patients with poor prognosis (PTS n=1673). Data source: comprehensive CRC cohort. (G) MIHC staining of FOLR2 (red), DKK3 (white), FOXP3 (yellow), PD-1 (orange), CD4 (green), CD8 (light blue), and DAPI (blue) showing CCIM image patterns. Scale bar = 50 μm. (H) CCIM recognition program. Each green circle represents a CCIM module. (I) Voronoi diagram of CCIM topological patterns. Green dots represent the center of each section. (J) Density map of CCIM in CRC. (K) Distribution of distances between exhausted T cells, Treg cells, and tolerant T cells and resident macrophages in CCIM modules. P values ​​were calculated using Pearson correlation coefficients, and P < 0.05 was considered significant. CCIM: colorectal cancer immunosuppression module; CCI: cell-cell interaction; scRNA-seq: single-cell RNA sequencing; CRC: colorectal cancer; TCGA-COAD: The Cancer Genome Atlas colon adenocarcinoma; PTS: patient; MIHC: multiplex immunohistochemistry.

[0051] Fig. 6A-J. AI-guided precision chemotherapy based on CCIM imaging modality. (A) Schematic diagram of the AI-guided precision chemotherapy process. (B) Representative CT scans and CCIM MIHC modality images of patients who responded or did not respond to chemotherapy. MIHC staining of FOLR2 (red), DKK3 (white), FOXP3 (yellow), PD-1 (orange), CD4 (green), CD8 (light blue), and DAPI (blue) was applied to depict CCIM. Scale bar = 150 μm. (C) Box plots show CCIM numbers in non-responder and responder groups. Significance was calculated by Wilcoxon test, and P < 0.05 was considered significant. (DF) Association of resident macrophages with exhausted T cells, Tregs, and tolerant T cells in responder and non-responder groups. (G) Comparison of eight deep learning models. The Y-axis represents the AUC value of each model in the test cohort. (H) CCIM-NET construction process of ResNext50. (I) ROC analysis shows the sensitivity and specificity of CCIM-NET in the test cohort. (J) Confusion matrix for the prediction of chemotherapy response by CCIM-NET in the test cohort. Columns represent predicted responses, while rows represent actual responses. CCIM: Colorectal Cancer Immunosuppression Module; AI: artificial intelligence; CT: computed tomography; ROC: receiver operating characteristic; AUC: area under the curve; MIHC: multiplex immunohistochemistry.

[0052] Fig. 7A-L. Anti-CSF1R therapy disrupts the immunosuppressive tumor microenvironment (TME) and improves chemotherapy response. (A) Comparison of estimated IC50 values ​​of tumor tissues with high intrinsic macrophage abundance (n=837) and low intrinsic macrophage abundance (n=836). Significance was assessed by Mann-Whitney U test. (B) UMAP plot of cells expressing CSF1R in different macrophage / monocyte subsets. The color key from gray to blue represents the relative expression level from low to high. (C) Schematic diagram of in vivo experiments. (DE) Evaluation of tumor burden in 4 groups of mice (control group, 5-FU, anti-CSF1R, 5-FU+anti-CSF1R). Significance evaluation was performed using paired t-test, *P<0.05, **P<0.01, **P less than 0.001. (F) HE and Ki-67 IHC staining of tumor tissues from 4 groups of mice, scale bar = 40μm. (G) Statistics of Ki-67 positive cells in each group. Significance was assessed by Kruskal-Wallis test. (H) Volcano plot showing DEPs. (I) GO analysis based on DEPs. (J) Box plot showing SSGSEA scores for abundance of FOLR2+ resident macrophages. (K) MIHC staining of tumor tissues from 4 groups of mice, scale bar = 40 μm. MIHC staining of FOLR2 (red), DKK3 (white), FOXP3 (yellow), PD-1 (orange), CD4 (green), CD8 (light blue), and DAPI (blue) were applied to depict CCIMs of each group. Scale bar = 50 μm. (L) Statistics of CCIM numbers in each group. n = 5 per group, 2 areas were randomly selected from tissue sections of 5 mice in each group. Significance was assessed by Kruskal-Wallis test, *P < 0.05, **P < 0.01, **P < .001, **p < .0001. CCIM: colorectal cancer immunosuppression module; UMAP: unified manifold approximation and projection; IC50: half maximal inhibitory concentration; IHC: immunohistochemistry; MIHC: multiplexed immunohistochemistry; RM: resident macrophages; DEPS: differentially expressed proteins; GO: gene ontology; ssGSEA: single sample gene set enrichment analysis.

[0053] Figure 8 .T-SNE plot colored by CRC patients. A total of 24 patients were included in the analysis. T-SNE: T-distributed stochastic neighbor embedding; CRC: colorectal cancer.

[0054] Fig.9A-C. Expression of FOLR2 and its prognostic value in CRC. (A) UMAP plot of FOLR2 expression levels. The color key from gray to blue indicates the relative expression level from low to high. (B) Violin plot of FOLR2 expression in monocyte and macrophage populations. Significance was assessed by Kruskal-Wallis test. (C) Survival analysis showed that patients with lower FOLR2 expression had prolonged DFS (PTS n=1673). DFS: disease-free survival; PT: patient.

[0055] Fig. 10A -C. CD4+ T cell subsets in colorectal cancer. (A) Trajectory inference of the five subsets of CD4+ T cells. (B) Trajectory inference using pseudo-time scores of the five subclusters of CD4+ T cells. The color key from black to gold indicates the pseudo-time scores from low to high. (C) Trajectory inference of PD-1 expression of the five subsets of CD4+ T cells. The color key from blue to white indicates PD-1 expression from low to high. CRC: colorectal cancer.

[0056] Fig.11A -H. Optimal soft threshold selection and sample clustering in WGCNA. (A) The power of β = 3 was selected as the optimal soft threshold to ensure scale-free co-expression. (B) The power of β = 3 indicates low average connectivity in the WGCNA model. (C) Clustering of gene modules in WGCNA. (D) Association between identified modules and resident macrophage abundance calculated using resident macrophage gene signatures on the overall transcriptome. The color key from green to red indicates the relative correlation level from low to high. (E) Correlation between gene significance of resident macrophages and module members in the yellow module. The black arrow indicates the involvement of FOLR2 in the yellow module. Network clustering identified 11 non-grey (meaningful) gene modules. (F) Gene network in the yellow module. (G) Gene ontology analysis based on genes within the yellow module (resident macrophage-related module), indicating enriched biological processes. (H) Proteomic pathway analysis of genes within the yellow module (resident macrophage-related module). Each small polygon corresponds to a single KEGG pathway, and its size is related to the number of genes involved in the pathway. WGCNA: weighted gene co-expression network analysis; KEGG: Kyoto Encyclopedia of Genes and Genomes.

[0057] Fig. 12A-J. Construction of CCIM scoring system. (A) Lasso Cox regression model was used to determine the most reliable markers. (B) Distribution of Lasso coefficients of gene signatures. (C) Multivariate survival analysis of CCIM score and other clinicopathological features in TCGA-CRC cohort. (D) Multivariate survival analysis of CCIM score and other important clinicopathological features in TCGA-CRC cohort. (E) CCIM score in independent CRC cohort #1 (GSE28722) stratifies patients with poor prognosis (PTS n=125). (F) CCIM score in independent CRC cohort #2 (GSE12945) stratifies patients with poor prognosis (PTS n=62). (G) CCIM score stratifies patients with poor prognosis in independent CRC cohort #3 (GSE72970) (patient number n=124). (H) CCIM score in independent CRC cohort #4 (GSE5851) stratifies patients with poor prognosis (PTS n=80). (I) Meta-analysis of training and testing cohorts. Immune-related pathway alterations in patients with different CCIM scores. (J) Pathway alterations were evaluated by GSVA using their gene sets. CCIM: Colorectal Cancer Immunosuppression Module; LASSO: Least Absolute Shrinkage and Selection Operator; CRC: Colorectal Cancer; pts: patients.

[0058] Fig.13A -C. Neural network model based on intrinsic macrophage characteristics predicts chemotherapy response. (A) Schematic diagram of the construction and validation of the resident macrophage signal prediction model. (B) Confusion matrix of chemotherapy response prediction by FOLR2-RM model in an external cohort. Columns represent predicted responses, while rows represent actual responses. (C) ROC analysis showed that the FOLR2-RM model has high sensitivity and specificity. ROC: Receiver Operating Characteristic FOLR2-RM model: FOLR2+ resident macrophage model. DETAILED DESCRIPTION OF THE INVENTION

[0060] I. Colorectal cancer treatment effect prediction system of the present invention

[0061] In one aspect, the present invention provides a colorectal cancer treatment effect prediction system based on multiple immunohistochemistry, which includes: (i) an analysis selection module; (ii) a preparation module; and (iii) an identification module.

[0062] In one embodiment, the analysis and selection module can be used to analyze samples of colorectal cancer tissue and select target cells and / or target antigens. In one embodiment, the sample of colorectal cancer tissue is a section of the tumor tissue, preferably a paraffin section. In one embodiment, the analysis and selection module includes an analysis component selected from the following group: (a) a single cell RNA sequencing (scRNA-seq) component; (b) a batch transcriptome analysis component; (c) a clinical information collection component; (d) a weighted gene co-expression network analysis (WGCNA) component; (e) a multiple immunohistochemistry (mIHC) staining component; and (f) a proteomics analysis component based on a tandem mass tag (TMT). In one embodiment, the single cell RNA sequencing component is capable of determining the heterogeneity of gene expression in a single cell in the same sample. In one embodiment, the scRNA-seq component includes: a single cell separation module; a module for separating RNA; a reverse transcription module; a cDNA sequencing module and a data processing module. In one embodiment, the single cell separation module uses a technology selected from the following group to separate single cells: flow activated cell sorting, micromanipulation, optical tweezers and microfluidics. In one embodiment, the data processing module includes: a means for processing sequencing data to obtain a counting matrix of cell genes and a means for performing downstream analysis. In one embodiment, the batch transcriptome analysis component uses transcriptome data selected from TCGA, ENCODE or ICGC databases for analysis to identify relevant target genes. In one embodiment, the clinical information collection component obtains clinical trial information from a registration platform selected from the following group: the World Health Organization International Clinical Trials Registration Platform (https: / / www.who.int / clinical-trials), the China Drug Clinical Trials Registration and Information Disclosure Platform (http: / / www.chinadrugtrials.org.cn) and the Drug Clinical Trials Registration and Information Disclosure Platform (https: / / clinicaltrials.gov). In one embodiment, the WGCNA component can be used to identify gene sets with similar expression patterns. Analyzing the connection between the gene set and the sample phenotype, drawing the regulatory network between the genes in the gene set and identifying the key regulatory genes can also be used to study the developmental regulation of different organs / tissue types and different stages, and the response mechanisms of different time points of biotic and abiotic stresses. In one embodiment, the WGCNA component includes: a module for constructing a gene relationship network; a module for constructing a gene set; a module for screening key genes; and a module for identifying key genes. In one embodiment, the staining method of the mIHC staining component is multiple rounds of sequential staining. In one embodiment, the principle of the mIHC staining is the same as the two-step method of IHC staining. In one embodiment, the mIHC staining is suitable for paraffin sections and frozen sections, preferably paraffin sections.In one embodiment, the primary antibody of the mIHC staining does not require any coupling or labeling. In one embodiment, the secondary antibody of the mIHC staining can be labeled with horseradish peroxidase (HRP), which can be imaged using a suitable fluorescent channel after being linked to a fluorescent group for color development. In one embodiment, the TMT-based proteomics analysis component can be qualitatively and quantitatively analyzed for biomacromolecules (e.g., proteins, polypeptides, and nucleic acids) based on mass spectrometry. In one embodiment, the TMT-based proteomics analysis component uses multiple stable isotope labels to specifically label, for example, the amino group of a polypeptide, and then performs tandem mass spectrometry analysis, which can simultaneously compare the relative content of proteins in a variety of different samples. In one embodiment, the multiple labels are composed of a reporter group, a balance group, and a peptide reaction group, but each label contains isotopes substituted at different positions, so the reporter group and the balance group have different molecular masses in each label, so after the labeled molecule is fragmented in MS / MS mode, sequence information can be obtained from the fragmented ions, and quantitative data can be obtained from the fragmentation of the label, thereby achieving quantitative analysis of the proteome.

[0063] In one embodiment, the analysis and selection module obtains the TME heterogeneity and molecular characteristics of the sample. In one embodiment, the TME heterogeneity and molecular characteristics of the sample are used to predict the treatment effect and / or prognosis of colorectal cancer. In one embodiment, the analysis and selection module uses the cell types selected from the following group for the subsequent preparation module: CD4 T cells, epithelial cells, CD8 T cells, B cells, myeloid cells, fibroblasts, NK cells, natural lymphocytes and plasma cells. In one embodiment, the analysis and selection module also includes (g) a component that further subdivides the cell types selected from the following group: CD4 T cells, epithelial cells, CD8 T cells, B cells, myeloid cells, fibroblasts, NK cells and natural lymphocytes, plasma cells. In one embodiment, the component (g) determines that the following myeloid cells, CD4 T cells and / or CD8 T cells are used for the subsequent preparation module: (1) The myeloid cells are selected from the following group: dendritic cells, mast cells, FCN1 + Monocyte-like cells, OAS + Macrophages, OLR1 + Macrophage, peptidase inhibitor 3 + Monocytes, FOLR2 + Resident macrophages, classical monocytes (e.g. CD14 + ), non-classical monocytes (e.g. CD16 +), neutrophils (in one embodiment, the neutrophils are selected from the group consisting of CD15, MPO, CD33, and CD66B), infiltrating macrophages (in one embodiment, the infiltrating macrophages are CD11b + or CCR2 + )、SPP1 + Macrophages and LAMP3 + Dendritic cells; (2) the CD4 T cells are selected from the group consisting of exhausted CD4 + T cells, effector CD4 + T cells, effector memory CD4 + T cells, central memory CD4+ T cells, naive CD4 + T cells, Th17 cells and Treg cells; and / or (3) the CD8 T cells are selected from the group consisting of: tolerant CD8 + T cells, effector CD8+ T cells, effector memory CD8 + T cells, central memory CD8 + T cells, depletion of CD8 + T cells, NKT cells, and tissue-resident CD8 +T cells. In one embodiment, the component (g) determines that the target antigens for the following myeloid cells, CD4 T cells and / or CD8 T cells are used for subsequent preparation modules: (1) The target antigens for the myeloid cells include but are not limited to one or more selected from the following group: CD68, FOLR2, C1QA, C1QB, SPP1, APOE, CD11B, CD172A, S100A8, S100A9, S100A12, CD15, CD33, CD66B, MPO, CD11C, HLA-DR, CD14, CD141, CD1C, CLEC9A, KIT, CPA3 and TPSB2; (2) The target antigens for the CD4 T cells include but are not limited to one or more selected from the following group: FOXP3, CCR7, SELL, GIMAP family, PD-1, TIGIT, CD3, CD7, CD4, IL7R, GRANZYME B, CXCR3, CX3CR1, TIM3, VISTA, LAG3, GITR, CD137; and / or (3) target antigens for the CD8 T cells include, but are not limited to, one or more selected from the group consisting of GZMK, GIMAP4, LAG3, NCR3, S100A4, DKK3, TMIGD2, PD-1, CD3, CD7, CD8, IL7R, GRANZYME B, CXCR3, CX3CR1, TIM3, VISTA, LAG3, GITR, and CD137. In one embodiment, the component (g) determines that one, two, three, four, five, six, or all seven target antigens selected from the group consisting of DKK3, FOLR2, FOXP3, PD1, CD8, CD4, and DAPI are used in subsequent preparation modules. In one embodiment, the component (g) determines that the following target antigen combinations are used in subsequent preparation modules: DKK3, FOLR2, FOXP3, PD1, CD8, CD4, and DAPI.

[0064] In one embodiment, the preparation module uses multiple immunohistochemical techniques to prepare images of colorectal cancer immunosuppressive units (CCIM) with selected target antigens. In one embodiment, the multiple immunohistochemical techniques are selected from the following groups: tyramide signal amplification (TSA), cyclic immunofluorescence (CyCIF), index co-detection (CODEX), multiple ion beam imaging (MIBI), imaging mass spectrometry (IMC), OPAL imaging, multiple error robust fluorescence in situ hybridization (MERFISH), random optical reconstruction microscopy (STORM) and Exchange-PAINT imaging. In one embodiment, the TSA technique utilizes the peroxidase reaction of tyramide (tyramide salt forms a covalent binding site under HRP and H2O2) to produce a large amount of enzymatic products, which can bind to surrounding protein residues (including tryptophan, histidine and tyrosine residues), so that the protein sample is stably bound to fluorescein, thereby achieving the labeling of multiple targets on a tissue section. In one embodiment, the TSA technology can realize multiple fluorescent labeling by means of cyclic staining, wherein the binding of tyramide fluorescein to the target protein is covalent binding, while the binding between the primary antibody and the target protein is non-covalent binding. By utilizing this feature, the primary antibody that has been bound can be removed by microwave heating; after the primary and secondary antibodies of the first round are washed off, the fluorescently labeled tyramide is still attached to the target. When the second target is detected, it is equivalent to a new round of labeling, and there is no need to consider whether the second round of secondary antibodies produces species crossover with the first round of antibodies. In one embodiment, the TSA technology can realize the labeling of multiple targets using tyramides with different fluorescent labels. In one embodiment, the CyCIF technology uses chemical reagents to remove antibodies and fluorescent dyes in each round of samples, and combined with a digital slice scanner, only 1-2 fluorescent secondary antibody reagents are needed to complete multiple immunofluorescence labeling. In one embodiment, the CODEX technology uses specific oligonucleotide "barcode tags" (Barcode) to label antibodies instead of directly labeling antibodies with fluorescent dyes. In one embodiment, the CODEX technology utilizes the specific binding of fluorescent dyes to oligonucleotide sequences complementary to the barcode, allowing CODEX to break through the limitation of the number of visible spectrum fluorescence imaging channels and easily achieve simultaneous detection and analysis of 50 or more protein indicators.

[0065] In one embodiment, the recognition module predicts the therapeutic effect of colorectal cancer based on the CCIM image. In one embodiment, the recognition module includes components selected from the following groups: (A) for predicting the tumor tissue half inhibitory concentration (IC 50) components; (B) components for constructing the CCIM scoring system; (C) components for analyzing the spatial relationship between myeloid cells and T cells in the CCIM image; (D) components for predicting prognosis; and (E) components for predicting metastasis. In one embodiment, the component for predicting tumor tissue IC50 uses the Prorophetic software package to perform ridge regression analysis on the integrated whole transcriptome for calculation. In one embodiment, the components for constructing the CCIM scoring system are modified based on previous studies (Braun DA et al., Cancer cell, 2021, 39(5): 632-48.e8; Bao X et al., Molecularoncology, 2020, 14(5): 917-32) to construct the CCIM scoring system. In one embodiment, the CCIM scoring system is constructed based on the expression signals of the characteristic genes in resident macrophages, exhausted CD4+T cells, Treg cells, exhausted CD8+T cells, and tolerant NK cells and genes from CCI ligand-receptor pairs. In one embodiment, the CCIM scoring system is constructed using the least absolute shrinkage and selection operator (Lasso). In one embodiment, the CCIM scoring system is constructed by including individual standardized gene expression values. In one embodiment, the individual standardized gene expression values ​​are weighted by their Lasso Cox coefficients. In one embodiment, the individual standardized gene expression values ​​are calculated by the following formula: ∑ i Coefficient (mRNA i )×normalized expression (mRNA i In one embodiment, the component for analyzing the spatial relationship between myeloid cells and T cells in the CCIM image analyzes FOLR2 + Spatial relationship between resident macrophages and T cells. In one embodiment, the component analyzing the spatial relationship between myeloid cells and T cells in the CCIM image determines whether exhausted T cells are associated with FOLR2 + The average distance of resident macrophages is the lowest. In one embodiment, the component for analyzing the spatial relationship between myeloid cells and T cells in the CCIM image predicts the effect of chemotherapy, radiotherapy, immunotherapy and targeted therapy based on the spatial relationship. In one embodiment, the component for predicting prognosis can use the CCIM score to stratify patients with poor prognosis. In one embodiment, the component for predicting metastasis can use the CCIM score to stratify patients with metastasis.

[0066] II. Use of the composition in preparing a kit for predicting the therapeutic effect of colorectal cancer

[0067] In one aspect, the present invention also provides the use of a composition in the preparation of a kit for predicting the therapeutic effect of colorectal cancer. In one embodiment, the composition includes a target antigen selected from the group consisting of a target antigen for the myeloid cells, a target antigen for the CD4 T cells, and a target antigen for the CD8 T cells. In one embodiment, the target antigen for the myeloid cells includes one or more selected from the group consisting of CD68, FOLR2, C1QA, C1QB, SPP1, APOE, CD11B, CD172A, S100A8, S100A9, S100A12, CD15, CD33, CD66B, MPO, CD11C, HLA-DR, CD14, CD141, CD1C, CLEC9A, KIT, CPA3, and TPSB2. In one embodiment, the target antigen for the CD4 T cell includes one or more selected from the group consisting of FOXP3, CCR7, SELL, GIMAP family, PD-1, TIGIT, CD3, CD7, CD4, IL7R, GRANZYME B, CXCR3, CX3CR1, TIM3, VISTA, LAG3, GITR, CD137. In one embodiment, the target antigen for the CD8 T cell includes one or more selected from the group consisting of GZMK, GIMAP4, LAG3, NCR3, S100A4, DKK3, TMIGD2, PD-1, CD3, CD7, CD8, IL7R, GRANZYME B, CXCR3, CX3CR1, TIM3, VISTA, LAG3, GITR, and CD137. In one embodiment, the composition comprises one, two, three, four, five, six or all seven target antigens selected from the group consisting of DKK3, FOLR2, FOXP3, PD1, CD8, CD4 and DAPI. In one embodiment, the composition comprises or consists of the following target antigens: DKK3, FOLR2, FOXP3, PD1, CD8, CD4 and DAPI.

[0068] III. Use of the composition in the preparation of a medicament for improving the chemotherapy effect in patients with colorectal cancer In one aspect, the present invention also provides a method comprising targeting FOLR2 + Use of a composition of therapeutic agents for resident macrophages in the preparation of a medicament for improving the chemotherapy effect in patients with colorectal cancer. In one embodiment, the therapeutic agent targeting FOLR2+ resident macrophages includes an anti-CSF1R therapeutic agent. In one embodiment, the therapeutic agent targeting FOLR2+ resident macrophages is an anti-CSF1R therapeutic agent.

[0069] IV. Advantages of the Invention

[0070] 1. Most existing TME studies focus on the identification of different cell populations, and the spatial interaction map of CRC TME is still unknown; the present invention provides a spatial interaction map of CRC TME.

[0071] 2. The generation of existing batch and single-cell transcriptome data relies on fresh tumor tissues rather than paraffin sections; the present invention can use paraffin sections for analysis and prediction.

[0072] 3. Existing transcriptome data-driven patient stratification is not pathology-friendly. Pathology methods are also not very good at establishing prognosis and treatment sensitivity predictions for CRC patients. pMMR patients respond poorly to immune checkpoint inhibitors; the present invention uses CCIM images for patient stratification, which is pathology-friendly and can more accurately predict the prognosis and sensitivity to treatment of CRC patients. The present invention can be used for pMMR patients.

[0073] 4. Existing transcriptomics, genomics, proteomics, single-cell sequencing and other methods will cause sample dissociation and therefore lack spatial information; the present invention provides spatial information.

[0074] 5. The present invention provides accurate prediction tools and models in the application scenario of neoadjuvant therapy for colorectal cancer. DETAILED DESCRIPTION

[0075] The present invention will be further described below in conjunction with specific embodiments, but these specific embodiments cannot be understood as limiting the scope of protection of the present invention. Those skilled in the art may make various changes or modifications to these specific embodiments without departing from the scope of the technical solution of the present invention, and the changed and modified embodiments still fall within the scope of protection of the present invention.

[0076] All chemicals used in the examples are commercially available.

[0077] Materials and methods

[0078] I. Patients and samples

[0079] This study was approved by the Ethics Committee of the First Affiliated Hospital, Zhejiang University School of Medicine and the Ethics Committee of the Affiliated Hospital of Southwest Medical University. Tumor tissues were obtained from biopsies of eight patients diagnosed with CRC. All specimens were confirmed by pathological examination. Samples were selected for sequencing by 10X Genomics and Bioinformatics Analysis. A cohort was recruited according to the criteria including CRC patients who received chemotherapy during 2018-2021. A total of 226 patients were included (training cohort n = 181, testing cohort n = 45).

[0080] II.scRNA-seq processing

[0081] RPMI 1640 (Gibco, Cat. No. 11875-093, USA) with 1 mM protease inhibitor (Solarbio, Cat. No. P6730, China) was used to transport CRC tissues. The dissociation enzyme mixture prepared by 2 mg / mL Dispase II (Sigma-Aldrich, Cat. No. 42613-33-2, USA), 1 mg / mL type VIII collagenase (Sigma-Aldrich, Cat. No. C2139, USA) and 1 unit / mL DNase I (NEB, Cat. No. M0303S, USA) was dissolved in PBS with 5% fetal bovine serum (FBS; Gibco, Cat. No. 16000-044, USA) at 37°C for 40 minutes to digest the tissue. The cells were dissociated and collected every 20 minutes and then filtered using a 40 μm nylon cell strainer (Falcon, Cat. No. 352340, USA). Red blood cells were removed using red blood cell lysis buffer (Invitrogen, catalog number, USA) with 1 unit / ml DNase I. Finally, cells were washed in PBS with 0.04% bovine serum albumin (BSA; Sigma-Aldrich, catalog number B2064, USA). The concentration of the single cell suspension was calculated with Countess (Thermo) and adjusted to 1000 cells / μL. The cells were loaded according to the Chromium Single Cell 3' Kit standard protocol to capture 5,000-10,000 cells / chip position (V2 chemistry). Library construction and all other processes were performed according to the manufacturer's standard protocol.

[0082] Illumina HiSeq X Ten was used to obtain a single-cell library using 150nt double-end sequencing. Reads were processed with the Cell Ranger 2.1.0 pipeline using default parameters. STAR software was used to align FASTQ with the human reference genome (GRCh38). For each sample, a gene barcode matrix was generated by counting unique molecular identifiers (UMIs) and filtering non-cell-related barcodes. Finally, a gene barcode matrix containing barcode cell and gene expression counts was generated. Subclusters were determined by principal component analysis (PCA) and t-distributed random neighbor embedding (t-SNE) using the Seurat R package (version 2.3.4) and Seurat FindVariableGenes function (Hao Y et al., Cell, 2021). Feature gene signatures were constructed using the Seurat R package. Single-cell trajectories in CRC were constructed using the Monocle2 package (v2.8.0) (Qiu X et al., Naturemethods, 2017, 14 (3): 309-15). Transcription factor (TF) analysis was performed using the dorothea package using default parameters (Garcia-Alonso L et al., Genome research, 2019, 29(8): 1363-75).

[0083] III. Public Group Collection

[0084] The internal queue of the present invention and the scRNA-seq data of the GSE146771 queue from the Gene Expression Comprehensive Database (GEO) are integrated. Batch transcriptome data and clinical information are obtained from the GEO library and the Cancer Genome Atlas (TCGA). After removing the batch effect using the sva R software package (Leek JT et al., Bioinformatics, 2012, 28 (6): 882-3), GSE39582, GSE17536, GSE17537, GSE14333, GSE56699, GSE37892, GSE33113 and TCGA-CRC transcriptome data are integrated. The transcriptome and chemotherapy response data from GSE72970 are used to construct the FOLR2 resident macrophage (RM) model, and the data from GSE19860 are used to test efficiency.

[0085] IV. Functional enrichment analysis

[0086] The GSVA R package was applied to single-sample GSEA (ssGSEA) for each sample ( S et al., BMC bioinformatics, 2013, 14(1): 1-15). The gene sets obtained from scRNA-seq were used for enrichment analysis. The results were plotted using the ggplot2 R package and the pheatmap R package (Wickham H, Computational Statistics, 2011, 3(2): 180-5).

[0087] V. Cell-cell interaction (CCI) analysis using SingleCellSignalR

[0088] In order to systematically analyze CCIS, the present invention uses the SingleCellSignalR software package to explore ligand and target gene pairs (Cabello-Aguilar S et al., Nucleic Acids Research, 2020, 48(10): e55-e). The gene expression data of the interacting cells are input into SingleCellSignalR and combined with existing models that integrate existing knowledge about the ligand-to-target signaling pathway. Then, the ligand-receptor interactions that drive gene expression changes in the target cells are predicted.

[0089] VI. Weighted Gene Co-expression Network Analysis (WGCNA)

[0090] The gene co-expression network was constructed by the WGCNA model (Langfelder P et al., BMC bioinformatics, 2008, 9(1): 1-13). The co-expression similarity was raised to the β power to define the weighted network adjacency. By evaluating the correlation between the abundance of resident macrophages and module members in patients from the integrated cohort, highly correlated modules may be identified. The HUB genes in the yellow module were selected and further analyzed. Gene ontology (GO) analysis was performed by the ClusterProfiler R software package (Wu T et al., TheInnovation, 2021, 100141), and Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis was performed by Proteomaps (a bionic visualization method for all pathways) (https: / / bionic-vis.biologie.uni-greifswald.de / ).

[0091] VII. Multiplex immunohistochemistry (MIHC) staining of tissue sections

[0092] Tumor tissues removed from CRC patients were fixed with 4% paraformaldehyde and then embedded in paraffin. First, the prepared tissue sections (4 μm) were baked in an oven at 65 ° C for 1 hour, dewaxed with xylene and rehydrated with graded alcohol. After rehydration, the sections were fixed in 10% neutral buffered formalin (NBF) (SolarBio, G2161, China) at room temperature for 20 minutes, placed in appropriate AR buffer, and then placed in a microwave at 100% power for 1 minute and at 20% power for another 15 minutes. The slides were blocked after cooling to room temperature and then incubated with the primary antibody at room temperature for 10 minutes. The slides were washed three times with TBST (SolarBio, T1082, China) and then incubated with Opal Polymer HRP Ms+Rb (Akoya Biosciences, USA, NEL820001KT) at room temperature for 10 minutes. The slides were rinsed three times with TBST before incubation with opal signal generation. Then, microwave treatment, blocking, primary antibody incubation, Opal Polymer HRP introduction and signal amplification were repeated. For labeling of CD68 (1:100, #76437), FOLR2 (1:100, #48274, CST), PTGS2 (ab179800, Abcam), CD4 (1:200, #48274, CST), CD8α (1:200, #85336S), PD-1 (1:200, #ab214421, Abcam), DKK3 (1:100, #PA5-102626, Invitrogen), FOXP3 (1:100, #700914, Invitrogen), CD68 (1:400, #76437, CST) and FOLR2 (1:300, #AB228643, Abcam), DAPI working solution was applied for 5 minutes at room temperature in the dark and the slides were washed with distilled water and TBST before loading. Finally, images of these tissue samples were taken using a confocal microscope (Nikon, Japan) or a VectraPolaris quantitative pathology imaging system, and the acquired images were analyzed using ImageJ (version 4.0).

[0093] VIII. AI-guided chemotherapy model construction

[0094] A neural network model for predicting response to chemotherapy in colorectal cancer was established using resident macrophage gene signatures. In brief, transcriptome data and chemotherapy response data from GSE72970 were used as input and output for training the FOLR2-RM model. GSE19860 was used as the test cohort. A traditional neural network was constructed with 25 layers.

[0095] For CCIM-Net, 8 models (ResNet18, ResNet34, ResNet50, ResNet101, Xception, DenseNet121, ResNext50) were trained and tested. Compared with traditional neural networks, the ResNet series of models introduces the concept of residual networks. For a block, let the fitable function be F(x) and its expected potential mapping be H(x). The ResNet network model learns the residual H(x)-x, that is, F(x):=H(x)-x, so that the ascending path becomes F(x)+x, and F(x)+x is used to fit H(X). In ResNet, the block formed by F(x)+x is called a residual block. Multiple similar residual blocks in series form ResNet. ResNet18, ResNet34, ResNet50 and ResNet101 are different versions of ResNet. The main difference between each version is the composition of the residual block and the number of convolutional layers. ResNet18 and ResNet34 are built with BasicBlock, while ResNet50 and ResNet101 are built with Bottleneck. The Resnext model can improve accuracy without increasing parameter complexity and reducing the number of hyperparameters. For example, the traditional algorithm of the fully connected layer is:

[0096]

[0097] where x i =x1,x2,x3,…,x D .

[0098] x D is the D-channel input vector to the neuron, ω i is the filter weight of the Ith channel. Resnext will ω i x i Replace with the more general function:

[0099]

[0100] Where Ti(x) can be any function. C is the cardinality (size of the transformation set). Ti has the same topological pattern. ResNext50 is selected for building CCIM-Net to obtain the highest accuracy in the test cohort.

[0101] IX. In vivo CRC model

[0102] B6 / JGpt-Apc em1Cin(Min) / Gpt mice were purchased from Gempharmatech Co., Ltd. (China), and all mice were housed in SPF facilities. To establish spontaneous CRC, 3% dextran sulfate (sodium salt) (DSS) (Cayman Chemical, 23250) was given in drinking water for 1 week, followed by normal drinking water. Mice were intraperitoneally injected with 10 μg / g CSF1R antibody or 50 mg / kg fluorouracil (5FU) for 2 weeks before being sacrificed, and the injection was performed every 3 days. On day 42, all mice were sacrificed and tumors were collected for H&E, IHC, and mIHC staining.

[0103] X. Hematoxylin-eosin (HE) staining

[0104] Mouse tumor tissue sections were dewaxed with xylene and then rehydrated with graded alcohol. The tissue sections were washed three times with PBS and then nuclear stained with hematoxylin for 30 minutes at room temperature. Then, the sections were exposed to ammonia water, which changed the nuclei of hematoxylin-stained cells from reddish to bluish purple. Subsequently, the tissue sections were rinsed with 75% alcohol for 2 minutes at room temperature, and then the cytoplasm was stained with eosin for 1 hour at room temperature. After eosin staining, the sections were directly rinsed with graded alcohol. Finally, anhydrous ethanol was replaced with xylene before being fixed with a slide. The sections were examined by optical microscopy (Leica) and the images were analyzed with Image-Pro Plus software (version 6.0).

[0105] XI. Immunohistochemical staining

[0106] Tumor tissue sections (4 μm) were dewaxed with xylene and rehydrated with graded alcohols. Antigenic epitope recovery was induced by microwave heating in blocking solution (Proteintech, B900780, China) in phosphate-buffered saline (PBS) for 1 h at room temperature. Then, Ki67 (1:200, 550609, BD Biosciences) primary antibody was incubated overnight at 4 °C. The next day, tissue sections were washed three times with PBS and then stained using an anti-mouse / rabbit universal immunohistochemistry detection kit (pk10006, Proteintech) according to the manufacturer's instructions. The loaded sections were examined with a light microscope (Leica) and images were analyzed with Image-ProPlus software (version 6.0).

[0107] XII. Proteomic Analysis Based on Tandem Mass Tags (TMT)

[0108] The tissue was dewaxed, rehydrated, and then acid hydrolyzed with formic acid (FA). Proteins were denatured with 6 M urea (Sigma-Aldrich, Germany) and 2 M thiourea (Sigma-Aldrich, Germany), and then digested with trypsin into peptides (1:20; Hualishi, Beijing, China) and Lys-C (1:80; Hualishi, Beijing, China) using pressure cycling technology (PCT) (Zhu Y et al., Mol Oncol, 2019, 13 (11): 2305-28; Gao H et al., J Proteome Res, 2020, 19 (5): 1982-90). TMTpro TM 16plex (Thermo Fisher Scientific TM , San Jose, USA) was used to label peptides (Li J et al., Nat Methods, 2020, 17(4): 399-404). In the TMT126 channel, each batch contained 15 experimental samples and one pool sample for standardization. The fractions (60 per batch) were separated by offline high pH reversed phase fractionation using a Thermo Dionex Ultimate 3000RSLC nanosystem and then combined to produce a total of 30 fractions per batch. Subsequently, the fractionated samples were separated using a Thermo Dionex Ultimate 3000RSLC nanosystem and then examined using a high-frequency mass spectrometer in data-dependent acquisition (DDA) mode (ThermoFisher Scientific TM , San Jose, USA). Proteome Discoverer (version 2.4, Thermo Fisher Scientific TM , Waltham, MA), searched all reviewed human entries from UniProt (downloaded on April 14, 2020, containing 20,365 proteins). The detailed parameters were previously described without modification (Shen B et al., Cell, 2020, 182(1):59-72e15; Nie X et al., Cell, 2021, 184(3):775-91e14).

[0109] XIII. CRC tissue half inhibitory concentration (IC 50 ) and the construction of CCIM scoring system

[0110] Estimated IC50 values ​​were calculated by ridge regression analysis of the integrated global transcriptome using the Prorophetic software package (Geeleher P et al., PloS one, 2014, 9(9): e107468). Default parameters were applied to the predictions.

[0111] The CCIM scoring system was constructed after modification based on previous studies (Braun DA et al., Cancer cell, 2021, 39(5): 632-48.e8; Bao X et al., Molecular oncology, 2020, 14(5): 917-32). The CCIM signal is constructed by characteristic genes in resident macrophages, exhausted CD4+T cells, Treg cells, exhausted CD8+T cells, tolerant NK cells, and genes from CCI ligand-receptor pairs. The least absolute shrinkage and selection operator (Lasso) was applied to construct the CCIM model (Engebretsen S et al., Clinical epigenetics, 2019, 11(1): 1-3). The CCIM scoring system includes individual standardized gene expression values, whose Lasso Cox coefficients are weighted as follows:

[0112] ∑ i Coefficient(mRNA i )×Normalized Expression(mRNA i ).

[0113] XIV. Statistical Analysis

[0114] All experimental data were analyzed using R software (R version 4.0.4). Measurement data were expressed as mean ± standard error. P value < 0.05 was considered to indicate significance. Univariate and multivariate Cox analysis was performed to determine whether the variables significantly affected disease-free survival (DFS).

[0115] Example 1. Deep learning-assisted CRC multi-omics integrated workflow

[0116] The detailed workflow of the present invention is as follows Figure 1A As shown. The present invention integrated analysis of scRNA-seq (public n=16 and internal n=8), bulk transcriptome (n=2081) and mIHC staining (n=246). The multicellular module was identified as the colorectal cancer immunosuppression module (CCIM). The CCIM score was calculated to stratify patients with poor prognosis in training (n=1673) and 4 test cohorts (total n=391). In addition, the present invention developed a CCIM-based MIHC imaging panel for spatially detecting CCIM patterns. The CCIM deep learning model (CCIM-NET) was developed from two cohorts from different centers (training cohort n=181 and test cohort n=45) to facilitate precision chemotherapy. In vivo experiments were performed for validation.

[0117] The present invention firstly explored the TME heterogeneity and molecular characteristics of cellular components in CRC using scRNA-seq ( Figure 1B-1C ). T-SNE was performed on genes with variable expression in all cells for dimensionality reduction and cell clustering. Cells were obtained from 24 patients by integrating public datasets and the internal cohort of the present invention ( Figure 8 To determine the identity of each cell population, differential gene expression analysis was performed to identify gene signatures ( Figure 1C ). Cell type markers were used to identify cell populations using the SingleR software package with human assistance. For example, CD8a, CD8b, GZMA, and GZMK were significantly more expressed in CD8+ T cells than in other cell populations. + T cells are annotated with CD3G, CD3D and CD4. Myeloid cells highly express S100A8, S100A9 and PLAUR. CD79a and CD79b are used to mark B cells. The present invention discovered 8 cell types, such as Figure 1C shown.

[0118] In summary, the integrated scRNA-Seq landscape covered the major cellular components of the CRC TME, including various immune cells and all stages of CRC tumors, based on which we applied a deep learning-assisted multi-omics integration workflow for CRC.

[0119] Example 2. Abundant resident macrophages are associated with poor prognosis in CRC patients

[0120] The present invention identifies myeloid populations in CRC. The tumor myeloid population is further subdivided into DC cells, mast cells, FCN1+ monocytes (Xue D et al., Arthritis & Rheumatology (Hoboken, NJ), 2021), OAS+ macrophages (Li J et al., Blood, The Journal of the American Society of Hematology, 2013, 121(15): 2934-42), OLR1+ macrophages (Chou CH et al., Scientific Reports, 2020, 10(1): 1-14), peptidase inhibitor 3+ monocytes (Navarro A et al., Journal of Leukocyte Biology, 2003, 73(4): 540-5) and FOLR2+ resident macrophages. Resident macrophages have high expression of FOLR2, C1QA, C1QB and APOE (Dick SA et al., Science Immunology, 2022, 7(67):eabf7777). Previous studies have shown that FOLR2 is highly expressed in resident macrophages and is associated with the activation of resident macrophages (Cresswell GM et al., Cancer Research, 2021, 81(3):671-84; A et al., Cancer research, 2009, 69(24):9395-403). Therefore, the present invention evaluated the expression of FOLR2 in different macrophage types in CRC. Consistent with the findings of Sharma et al., FOLR2 expression in resident macrophages was significantly higher than that in other monocytes / macrophages (Figure 9A9B). In addition, high FOLR2 expression was associated with poor prognosis in CRC ( Fig. 9C ).

[0121] To better understand the role of these different myeloid populations in immune dysregulation, the present invention examined the expression of immune checkpoint genes and immune-related transcriptional programs. LGALS9 and VSIG4 promote T cell immune checkpoint activation through alternative pathways (Lines JL et al., Cancer research, 2014, 74 (7): 1924-32; Vogt L et al., The Journal of clinical investigation, 2006, 116 (10): 2817-26), which are detected in resident macrophages but not in other myeloid cell populations ( Figure 2B ). Pseudo-temporal analysis reveals the differentiation trajectory of monocytes / macrophages ( Figure 2C-2DIn particular, resident macrophage subsets are located at the site of terminal differentiation. Thus, transcription factor (TF) analysis revealed distinct TF activities in resident macrophages and other monocyte / macrophage subsets ( Figure 2E ). For example, CEBPB is a TF that plays a key role in the differentiation of macrophages into a pro-inflammatory profile (Ruffell D et al., Proceedings of the National Academy of Sciences, 2009, 106(41):17475-80). Resident macrophages have significantly lower CEBPB expression compared to other populations. In contrast, macrophage TCF4 expression is significantly increased in resident macrophages, which is a key Wnt pathway regulator and contributes to low-inflammatory M2 polarization (Cosín-Roger J et al., PLoS One, 2013, 8(10):e78128). In addition, different expression signatures are associated with macrophage activity, including antigen presentation and IFN-γ response functions, and are observed in different monocyte / macrophage populations ( Figure 2G ). Meanwhile, different pro-inflammatory and anti-inflammatory expression patterns were observed in monocyte / macrophage populations ( Figure 2F Different TF activities, gene expression patterns, and pathway changes in different macrophage populations suggest different population characteristics of macrophages in CRC, which makes it possible to develop new therapies that utilize TAM subset targeting.

[0122] In order to determine the impact of myeloid cell populations on the prognosis of colorectal cancer patients, the present invention combined a large number of transcriptome and scRNA-seq data for further analysis. A comprehensive cohort of 1,673 CRC patients was selected. Gene markers of each myeloid subset were identified from scRNA-seq, and the correlation between myeloid subsets and DFS in CRC patients was analyzed. The present invention observed that patients enriched in FCN+ monocytes, DC cells, mast cells, OLR1+ macrophages, peptidase inhibitor 3+ monocytes and resident macrophages showed poor DFS survival rate ( Figure 2H ). In particular, patients with an abundance of resident macrophages had the worst prognosis.

[0123] Example 3. FOLR2+ resident macrophages shift to a low-inflammatory phenotype

[0124] Next, the present invention focuses on discussing the potential mechanisms of how resident macrophages affect the prognosis of CRC patients.

[0125] First, GO analysis was performed to characterize the underlying biological processes of each monocyte / macrophage population. Resident macrophages were enriched in lipid metabolism-related processes, while OLR1+ macrophages were enriched in G protein-coupled receptor and growth factor receptor binding ( Figure 3A GSVA further demonstrated that the resident macrophage population had significantly reduced inflammatory response activity compared to other monocyte and macrophage populations ( Figure 3B-3C The density plot showed that most resident macrophages had low inflammatory response scores compared to other monocyte and macrophage populations ( Figure 3D ). In order to verify the results of SSGSEA analysis, the present invention selected a typical inflammatory marker PTGS2 (also known as Cox-2) and used MIHC technology for verification. Compared with other monocyte / macrophage subsets, FOLR2+ resident macrophages showed less inflammatory marker PTGS2 ( Figure 3E-3F ). Therefore, the present invention confirms that FOLR2+ resident macrophages are a special type of macrophages with a low inflammatory phenotype in CRC.

[0126] Example 4. Compositional and pathway heterogeneity of T cells in CRC

[0127] The present invention then identified T cell subsets and key genes expressed in T cells. CD4+T cells were further divided into five cell populations, including exhausted CD4+T cells, memory CD4+T cells, naive CD4+T cells, Th17 cells and Treg cells ( Figure 4A ). Tregs highly expressed Foxp3. Naive CD4+ T cells showed high CCR7 and SELL expression. Th17 expressed GIMAP family genes, and exhausted T cells were identified as having high PD-1 and TIGIT expression. CD8+ cells were divided into tolerant CD8+ T cells, effector memory CD8+ T cells, exhausted CD8+ T cells, exhausted NKT CD8+ T cells, and tissue-resident CD8+ T cells ( Figure 4B ). Effector memory CD8+T cells highly expressed GZMK and GIMAP4. Exhausted NKT cells highly expressed LAG3 and NCR3. Tissue-resident CD8+T cells showed high S100A4 expression. DKK3 was found to be highly expressed in tolerant CD8+T cells, and TMIGD2 and PD-1 were identified as characteristic genes of exhausted CD8+T cells.

[0128] Next, we identified key genes involved in T cell regulatory pathways such as cytolysis, cytokine, and inhibitory pathways ( Figure 4C), revealing the heterogeneity of T cell activation. For example, ZNF683, as a transcription factor, is mainly expressed in tissue-resident CD8+T cells. GZMA, GZMB, NKG7 and other cell lysis genes are mainly expressed in CD8+T cell subsets. Exhausted NKT and tissue-resident CD8+T still have the ability to release IFNG. C7 immunological marker (from Broad Institute: http: / / www.gsea-msigdb.org / gsea / msigdb / index.jsp, representing cell states and perturbations within the immune system) Changes in GSVA scores further reveal the regulation of suppressive TME and T cell exhaustion in CRC ( Figure 4D Finally, pseudotime analysis showed that PD1 was highly expressed in CD8+ T cells undergoing exhaustion ( Figure 4E-4G ). At the same time, high expression of PD-1 was also observed during the exhaustion of CD4+T cells ( Figures 10A-10C ).

[0129] In summary, major T cell populations were identified in the present RNA-seq analysis that displayed distinct gene expression patterns and pathway activation.

[0130] Example 5. CCIM directed by low-inflammatory FOLR2+ resident macrophages composed of an immunosuppressive T cell population

[0131] Complex cellular crosstalk was investigated to reveal the regulation of the suppressive CRC TME. CCIs between each cell type were identified. FOLR2+ resident macrophages exhibited strong cell-cell interaction activity with multiple cell populations ( Figure 5A The present invention then focuses on the interaction between macrophage populations and T cell populations. FOLR2+ resident macrophages show the highest number of CCIs for CD4+T and CD8+T cells ( Figure 5B-5C ). In terms of T cell subsets, exhausted CD4+T cells, Tregs, tolerant CD8+T cells, and exhausted CD8+T cells constituted the major proportion of all CCIs with resident macrophages. In summary, these findings indicate that FOLR2+ resident macrophages exhibit rich interactions with exhausted CD4+ cells, exhausted CD8+T cells, tolerant CD8+T cells, and Treg cells, which may affect the formation of suppressive CRC TME. Therefore, the cell module composed of FOLR2+ resident macrophages, exhausted CD8+T cells, tolerant CD8+T cells, exhausted CD4+T cells, and Treg cells is called CCIM ( Figure 5D ). We identified common interactions between FOLR2+ resident macrophages and the above cell populations, showing non-cell autonomous effects ( Figure 5E). The WGCNA method infers co-expression networks and related biological mechanisms in high-throughput data. Therefore, the present invention applies WGCNA to CRC batch transcriptome data to explore the potential regulatory role of CCIM. The power of β = 3 is selected as the optimal soft threshold to ensure scale-free co-expression ( Figures 11A-11B ). The resident macrophage score for each patient was calculated using the GSVA method. Eleven non-grey (significant) gene modules were identified, of which the yellow module had the most significant correlation with the resident macrophage GSVA score ( Fig. 11C A total of 1052 genes participated in the yellow module, which showed a significant association between module membership score and gene significance, indicating the importance of these genes in the regulatory network of resident macrophages ( Figures 11D-11E Using these genes as input for KEGG and GO analyses, we further found that T cell-related pathways were altered, confirming the immunosuppressive regulatory potential of FOLR2+ resident macrophages ( Figures 11E-11F ). Briefly, the present invention identified a CCIM cell module using scRNA-Seq technology and further confirmed its function through bulk transcriptome.

[0132] Then, the present invention developed a CCIM scoring system to stratify CRC patients with different prognoses. The Lasso Cox model suggested using a total of 17 features (genes) to construct the CCIM scoring system ( Fig. 12A ). The coefficient of each gene is Fig. 12B As shown. Using the coefficients and different expression patterns of the 17 features, the CCIM score was calculated for each patient. The results showed that patients with high CCIM scores showed worse DFS than those with low CCIM scores ( Fig. 5F ). Multivariate Cox analysis showed that CCIM score was an independent risk factor for CRC patients in the TCGA-COAD cohort ( Figures 12C-12D In addition, four independent cohorts (GSE28722, GSE12945, GSE72970, and GSE5851) further validated the robustness of the CCIM score in stratifying CRC patients ( Figures 12E-12H The combined evaluation showed that a lower CCIM score was significantly associated with improved OS (HR 2.04, 95% CI 1.56-2.67, P < 0.001). Fig.12I ). In addition, in patients with higher CCIM scores, the GSVA algorithm showed a significant increase in anti-PD1 resistance ( Fig.12J ).

[0133] Finally, the topological pattern of CCIM guided by FOLR2+ resident macrophages in CRC TME was verified by MIHC staining ( Figure 5GCCIM was defined as a module containing at least one FOLR2+ macrophage, one tolerant CD8+ T cell, one Treg, and one exhausted CD4+ T cell. CCIM MIHC images were used to identify modules ( Figure 5H The Voronoi diagram showed that FOLR2+ resident macrophages were clustered with Tregs, exhausted CD4+ T cells, tolerant CD8+ T cells, and exhausted CD8+ T cells in CRC tissues ( Fig.5I The high density of CCIM numbers suggests the important role of CCIM in suppressive CRC TME ( Figure 5J ). Finally, the average distances of Tregs, exhausted T cells, and tolerant T cells from FOLR2+ resident macrophages in each module were investigated to show the spatial relationship between FOLR2+ resident macrophages and T cells. Compared with other CCIM components, exhausted T cells had the lowest average distance from FOLR2+ resident macrophages ( Figure 5K In summary, the present invention reveals a CCIM module consisting of low-inflammatory Folr2+ resident macrophages and immunosuppressive T cell populations. The spatial map of CCIM was depicted by MIHC and related image analysis.

[0134] Example 6. Spatial interaction maps directed by FOLR2+ resident macrophages contribute to AI-guided precision chemotherapy

[0135] Several studies have shown that AI-assisted analysis of transcriptome data can predict treatment response, molecular typing, and pathway activation. However, different cohorts (Azarkhalili B et al., Scientific reports, 2019, 9(1): 1-14; M et al., Nature Reviews Genetics, 2018, 19(2):93-109; Way GP et al., Cell reports, 2018, 23(1):172-80.e3) have unstable efficiency and accuracy. Next, the present invention evaluates the potential relationship between FOLR2+ resident macrophages and chemotherapy response. By applying neural networks, the present invention established a FOLR2-RM model based on resident macrophage signals to predict chemotherapy response in the training cohort, and then evaluated the prediction efficiency of the FOLR2-RM model in the external cohort ( Fig.13A ). The confusion matrix and receiver operating characteristic (ROC) plots showed that the prediction efficiency of the FOLR2-RM model was acceptable, and the area under the curve (AUC) value was relatively high (AUC = 0.797), which also indicated the potential important role of FOLR2+ resident macrophages in the chemotherapy response of CRC patients ( Figures 13B-13C ).

[0136] Deep learning techniques can extract hidden information directly from image patterns and provide potentially clinically useful information. Therefore, the present invention tested several advanced deep learning models for predicting the response to systemic chemotherapy using CCIM image patterns. To further utilize the CCIM model and improve the prediction efficiency, the present invention constructed a chart ( Fig. 6A ). The present invention established two medical center cohorts as a training cohort (n=181) and a test cohort (n=45). Tumor tissue was obtained by colonoscopy and MIHC staining was performed before systemic chemotherapy. Then, the patients were given systemic chemotherapy, and the chemotherapy response was evaluated according to the RECIST criteria. Typical CT scans before and after chemotherapy and CCIM MIHC staining images (tissue sections before systemic chemotherapy) are shown in Figure 2. Figure 6B In all patients (including training and test cohorts), non-responders showed greater CCIM enrichment than responders ( Figure 6C In terms of cell numbers, exhausted T cells showed a greater correlation with FOLR2+ resident macrophages in non-responders than in responders (R = 0.60 and R = 0.51, respectively) ( Fig.6D The same trend was found in Treg cells (R = 0.42 and R = 0.33, respectively) ( Fig. 6E ). Tolerant T cells in responders were more highly correlated with FOLR2+ resident macrophages compared with non-responders (R=0.54 and R=0.43, respectively) ( Fig. 6F ). We trained 8 models (ResNet18, ResNet34, ResNet50, ResNet101, Xception, DenseNet121, ResNext50) in the training cohort and tested the prediction accuracy of each model in the test cohort. The AUC values ​​in the test cohort ranged from 0.774 to 0.990 ( Figure 6G ). Finally, through the confusion matrix ( Figures 6I-6J ) Select the ResNext50 model with the largest AUC value and the highest accuracy as CCIM-Net ( Figure 6H ). Therefore, by integrating two independent cohorts, the present invention confirms that CCIM-NET is a powerful AI-guided clinical chemotherapy decision-making tool.

[0137] Example 7. Disruption of CCIM by targeting FOLR2+ resident macrophages improves chemotherapy response

[0138] The present invention next evaluated the impact of the abundance of FOLR2+ resident macrophages on chemotherapy outcomes. Estimated IC50 values ​​for CRC tissues with high and low FOLR2+ resident macrophage enrichment were inferred by ridge regression analysis of the integrated whole transcriptome (n=1673) (35). Significantly higher IC50 values ​​were found in tissues with high FOLR2+ resident macrophage abundance, indicating the potential of targeting FOLR2+ resident macrophages to improve chemotherapy outcomes ( Fig. 7A ). In addition, CSF1R expression in FOLR2+ resident macrophages was higher than that in other macrophage populations, suggesting that anti-CSF1R therapeutics would primarily target FOLR2+ resident macrophages in CRC ( Figure 7B ). Therefore, the present invention applies anti-CSF1R therapeutic agents to B6 / JGpt-Apc em1Cin(Min) / Gpt mouse model to deplete FOLR2+ resident macrophages in vivo ( Figure 7C Four groups were designed (control group, 5-FU treatment group, anti-CSF1R treatment group, and 5-FU + anti-CSF1R treatment group), among which the 5-FU group showed the lowest tumor burden ( Figure 7D-7E HE staining, Ki67IHC staining, and mIHC staining were used to evaluate the characteristics of the CRC tumor model ( Figure 7F ). Ki-67 immunohistochemical staining showed that the proliferation was reduced in the 5-FU+anti-CSF1R group compared with the other groups ( Figure 7G ). The present invention also performed proteomic analysis on four groups of tumor tissues. The volcano plot showed that the most differentially expressed proteins (DEPs) were found between the 5-FU+anti-CSF1R treatment group and the control group compared with other comparisons ( Figure 7H ). GO analysis based on DEPS (adjusted -P<0.05) showed that lipid modification, membrane organization, myeloid cell activation involved in immune response, immune response regulation signaling pathways, and several other terms were enriched to varying degrees in the comparison between the control group and the 5-FU treatment group, the anti-CSF1R treatment group, and the 5-FU+anti-CSF1R treatment group ( Fig.7I SSGSEA analysis confirmed that the abundance of FOLR2+ resident macrophages in the anti-CSF1R treatment group and the 5-FU+anti-CSF1R treatment group was lower than that in the control group and the 5-FU treatment group ( Figure 7J mIHC staining showed that the number of CCIMs was reduced in the CSF1R-treated group and the 5-FU+CSF1R-treated group compared with the other groups, confirming the disruption of CCIMs by anti-CSF1 treatment ( Figure 7K-7L Together, these results suggest that anti-CSF1R treatment leads to disruption of the immunosuppressive CCIM module and enhanced chemotherapeutic response.

[0139] Analysis and discussion

[0140] In this study, the present invention comprehensively analyzed the molecular and cellular characteristics of CRC components through the description of single-cell landscapes and the integration of global transcriptomes and MIHC. Importantly, the present invention determined the importance of FOLR2+ resident macrophages, which exhibit low inflammatory properties in the formation of an immunosuppressive module that affects prognosis. CCIM, composed of FOLR2+ resident macrophages, exhausted CD4+ T cells, exhausted CD8+ T cells, tolerant CD8+ T cells, and Treg cells, was identified. Therefore, the present invention developed a CCIM scoring system as a clinically tractable tool for predicting survival and treatment guidance in CRC patients. Using the MIHC pattern based on the CCIM spatial interaction map, the present invention developed an AI-guided precision chemotherapy model that showed robust predictive efficiency for chemotherapy sensitivity. In vivo models further showed that targeting FOLR2+ resident macrophages disrupted CCIM and improved chemotherapy response. Therefore, the present invention believes that destroying CCIM mediated by FOLR2+ resident macrophages may be a promising method for treating CRC.

[0141] Zhang et al. emphasized that myeloid cell populations and conventional dendritic cell subsets are key mediators of cellular crosstalk in CRC TME (Zhang L et al., Cell, 2020, 181(2): 442-59.e29). Tumor-associated macrophages (TAMs) are one of the main populations of myeloid cells and a promising target for immunotherapy. Macrophages in tumor tissues have traditionally been identified as two subpopulations, called M1 and M2 populations, based on their different molecular characteristics and pathway differences (Wang Y et al., International immunopharmacology, 2019, 70: 459-66; Yunna C et al., European journal of pharmacology, 2020, 877: 173090). M1 macrophages have pro-inflammatory properties and are considered to be an anti-tumor subpopulation. In contrast, M2 macrophages with anti-inflammatory and pro-angiogenic properties promote tumor progression in various solid tumor types (Martínez VG et al., Clinical Cancer Research, 2017, 23(23): 7388-99; Aljabery F et al., Urologic Oncology: Seminars and Original Investigations, 159, e19-.e26, Elsevier, 2018; Suárez-Sánchez FJ et al., Cancers, 2020, 12(7): 1764). However, considering the heterogeneity of colorectal cancer immune populations, the present invention believes that the M1 and M2 classifications may not be applicable to colorectal cancer. Various TAMs in TME provide protection and vulnerability (LiX et al., Nature Immunology, 2021, 22(9): 1078-9). Targeting tumorigenic macrophages is a priority for cancer treatment, but in most tumor types, it is still unclear which TAM subpopulations have tumorigenic effects and when they should be targeted. Therefore, identifying the pro-oncogenic TAM subpopulations in CRC and their biological functions is crucial for developing new therapies that target TAM subpopulations. Therefore, the present invention explored the monocyte / macrophage population in CRC by integrating scRNA-Seq. FCN1+ monocytes, peptidase inhibitor 3+ monocytes, OAS+ macrophages, OLR1+ macrophages, and FOLR2+ resident macrophage populations with different gene expression signatures were identified. In particular, patients with abnormal FOLR2+ resident macrophage enrichment showed the worst prognostic outcomes and high immunotherapy resistance.

[0142] By using several bioinformatics techniques, the present invention confirms the importance of FOLR2+ resident macrophages in remodeling CRCTME. Resident macrophages perform nutritional functions, contributing to organ development, tissue remodeling and homeostasis (Pollard JW et al., Nature reviews immunology, 2009, 9 (4): 259-70). In tumor tissue, the exact function of this resident macrophage in TME remains unclear. A study showed that the recruitment of macrophages to pancreatic ductal adenocarcinoma tumors in CCR2- / - mice was reduced (Zhu Y et al., Immunity, 2017, 47 (2): 323-38.e6). However, the growth of pancreatic ductal tumors was not affected by the reduction of macrophages, which means that the progression of pancreatic ductal tumors may be regulated by CCR2+ macrophage-independent mechanisms. They subsequently found that with the help of tumor-derived CSF-1, resident macrophages were able to self-renew in tumors, a process that can promote tumor progression (ibid). In the present study, the present invention discovered the low inflammation-related properties of Folr2+ resident macrophages. Active lipid metabolism-related pathways are enriched in FOLR2+ resident macrophages. Given that M2 phenotype macrophages are macrophages with anti-inflammatory properties and significant lipid metabolism activation, the present invention believes that FOLR2+ resident macrophages exhibit an M2-like phenotype in CRC.

[0143] Next, the present invention detected a large number of FOLR2+ resident macrophages, Treg cells, exhausted CD4+T cells, exhausted CD8+T cells and tolerant CD8+T cells in CRC tissues, which constituted a cell module called CCIM. Cell-cell interaction analysis showed that the interaction between FOLR2+ resident macrophages and the above-mentioned inhibitory cells was significantly enriched compared with other macrophage populations. In addition, co-localization of FOLR2+ resident macrophages, Treg cells, exhausted CD4+T cells, exhausted CD8+T cells and tolerant CD8+T cells was observed in tumor tissues, indicating the potential role of the unique CCIM topology in inhibitory TME reprogramming in CRC. Most previous studies inferred the interaction network between cell populations based on the expression of receptor-ligand pairs. However, considering the spatial distribution, discrete cell populations cannot capture the full spectrum of network relationships. Recently, imaging-based studies have highlighted CCI networks based on colocalization of adjacent cell populations (Schürch CM et al., Cell, 2020, 182(5): 1341-59.e19). The so-called TME center (or module) shows the ability to regulate suppressive immune TME and predict immunotherapy response. For example, Leader et al. revealed a cell module composed of PDCD1+CXCL13+ activated T cells, IgG+ plasma cells, and SPP1+ macrophages, called the lung cancer activation module (LCAM) (Leader AM et al., Cancer Cell, 2021, 39(12): 1594-609.e12). Even in patients with TMB above the median, high baseline LCAM scores were associated with enhanced immunotherapy response, suggesting that although immune cell composition is associated with TMB, it may be a non-redundant biomarker of immunotherapy response (ibid.). Pelka et al. found myeloid cell attraction centers associated with tissue damage at the tumor-luminal interface (Pelka K et al., Cell, 2021, 184(18): 4734-52.e20). In contrast, immune hubs enriched in mismatch repair defects were found in tumors, characterized by activated T cells and malignant and myeloid cells expressing chemokines that attract T cells (ibid.). However, previous studies on TME modules have mostly focused on the characterization of image patterns.In recent years, AI-guided cancer risk stratification and precision treatment have been promising approaches for personalized medicine (D'Amore B et al., Current Oncology Reports, 2021, 23(6): 1-8; Kann BH et al., Cancer Cell, 2021; Elkhader J et al., Seminars in Cancer Biology, Elsevier, 2021; Tong Z et al., Frontiers in bioengineering and biotechnology, 2020, 8: 196). Considering the special topological pattern of CCIM shown in the Voronoi diagram, the present invention develops a workflow for constructing a CCIM-based precise chemotherapy model assisted by scRNA-SEQ. The present invention uses cutting-edge deep learning technology to train 8 artificial intelligence models, and finally constructs CCIM-Net, which shows a high prediction accuracy in the test cohort. Therefore, the present invention proves that CCIM-Net based on the 7-color MIHC panel is a powerful tool to help oncologists predict chemotherapy response. CCIM-NET does not rely on high-throughput data generated by fresh tumor tissue, and is efficient, economical, and pathology-friendly. Artificial intelligence-assisted tools can help develop precise treatment strategies for CRC.

[0144] Anti-CSF1R antibodies are a good way to target TAMs and cause their depletion (Cannarile MA et al., Journal for immunotherapy of cancer, 2017, 5(1): 1-13). However, various solid tumors have limited responses to CSF1R antibodies. Zhang et al. determined the persistence of anti-inflammatory TAM populations, and the loss of pro-inflammatory TAM populations was associated with resistance to anti-CSF1R therapy. FOLR2+ resident macrophages exhibit low inflammatory properties in CRC, which are not observed in other macrophage populations. Therefore, targeting different FOLR2+ resident macrophage populations may be more conducive to overcoming the shortcomings of general TAM blocking therapy. In the in vivo experiments of the present invention, as expected, anti-CSF1R therapy significantly reduced the abundance of FOLR2+ resident macrophages due to the high expression of CSF1R on resident macrophages. This phenomenon is consistent with a previous study showing that resident macrophages are able to self-renew in tumors and promote tumor progression due to tumor-derived CSF-1. The present invention also found that CCIM enrichment was reduced due to the depletion of FOLR2+ resident macrophages. Therefore, the present invention believes that antibodies specifically targeting FOLR2+ resident macrophages may destroy CCIM and become a promising method for combined chemotherapy treatment of CRC.

[0145] Although the present invention illustrates the immune cell map of CRC by scRNA-Seq, the selected markers may limit the identification of some unknown immune cell phenotypes. In addition, metabolism-related pathways are significantly altered in CRC tumor tissues and are associated with the prognosis of CRC patients. Abnormal FOLR2 expression in resident macrophages deserves further study of folate metabolism during TME reprogramming. In addition, the present invention proposes that the artificial intelligence-guided precision chemotherapy model of the present invention contains information for clinical decision-making, and samples are obtained from the clinic of the present invention to further validate the system. However, the cohort size is still relatively small. Multicenter cohort validation and further clinical trials should be considered. Finally, the present invention found that targeting FOLR2+ resident macrophages with CSF1R therapy may disrupt the immunosuppressive TME. Considering the toxicity of anti-CSF1R treatment, developing a new antibody targeting FOLR2+ resident macrophages may be more promising to reduce immunotoxicity and can be used as a new immunotherapy.

[0146] In summary, our studies revealed novel CCIM topological patterns and revealed the importance of CCIM components in reprogramming the suppressive CRC TME. The CCIM score can be used as a clinically tractable survival prediction tool. CCIM-NET provides a promising and effective opportunity for therapeutic intervention and precision chemotherapy. Targeting FOLR2+ resident macrophage therapy alters the enrichment of CCIM and is a promising immunotherapy. Therefore, we believe that the immune spatial interaction map of CCIM can serve as a more direct measure of the immunosuppressive TME and provide information for clinical decision-making.

[0147] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the specific implementation of the present invention is described above, it should be understood by those skilled in the art that this is only an example, and the protection scope of the present invention is defined by the attached claims. Those skilled in the art may make various changes or modifications to the technical contents disclosed above without departing from the scope of the technical solution of the present invention, and these changes and modifications all fall within the protection scope of the present invention.

Claims

1. A colorectal cancer treatment effect prediction system, comprising: (i) an analysis and selection module, which analyzes the sample of tumor tissue and selects target cells and / or target antigens; (ii) a preparation module for preparing a colorectal cancer immunosuppressive unit (CCIM) image using a multiplex immunohistochemical technique with a selected target antigen; and (iii) A recognition module, which predicts the treatment effect of colorectal cancer based on the CCIM image.

2. The system of claim 1, wherein the analysis selection module comprises an analysis component selected from the group consisting of: (a) Single-cell RNA sequencing (scRNA-seq) components; (b) Bulk transcriptome analysis component; (c) clinical information collection component; (d) Weighted Gene Co-expression Network Analysis (WGCNA) component; (e) multiplex immunohistochemistry (mIHC) staining kit; and (f) Tandem mass tag (TMT)-based proteomic analysis components.

3. A system according to claim 1 or 2, wherein the TME heterogeneity and molecular characteristics of the sample are obtained by the analysis selection module, and such characteristics can be used to predict the treatment effect and / or prognosis of colorectal cancer.

4. The system according to any one of claims 1-3, wherein the analysis selection module determines that the cell type used for the preparation module is selected from the group consisting of CD4 T cells, epithelial cells, CD8 T cells, B cells, myeloid cells, fibroblasts, NK cells, natural lymphocytes and plasma cells.

5. The system according to claim 4, wherein the analysis selection module further comprises (g) a component for further subdividing the CD4 T cells, epithelial cells, CD8 T cells, B cells, myeloid cells, fibroblasts, NK cells and natural lymphocytes, plasma cells.

6. The system of claim 5, wherein the component (g) determines: (1) The myeloid cells are selected from the group consisting of dendritic cells, mast cells, FCN1 + Monocyte-like cells, OAS + Macrophages, OLR1 + Macrophage, peptidase inhibitor 3 + Monocytes and FOLR2 + Resident macrophages, classical monocytes (CD14+), non-classical monocytes (CD16+), neutrophils (CD15 or MPO, CD33, CD66B), infiltrating macrophages (CD11b+\CCR2+), SPP1+ macrophages, LAMP3+ dendritic cells; (2) the CD4 T cells are selected from the group consisting of exhausted CD4+T cells, effector CD4+T cells, effector memory CD4+T cells, central memory CD4+T cells, naive CD4+T cells, Th17 cells and Treg cells; and / or (3) The CD8 T cells are selected from the group consisting of: + T cells, effector CD8+ T cells, effector memory CD8 + T cells, central memory CD8 + T cells, depletion of CD8 + T cells, NKT cells, and tissue-resident CD8 + T cells.

7. The system of claim 5 or 6, wherein the component (g) determines: (1) the target antigens for the myeloid cells include one or more selected from the group consisting of CD68, FOLR2, C1QA, C1QB, SPP1, APOE, CD11B, CD172A, S100A8, S100A9, S100A12, CD15, CD33, CD66B, MPO, CD11C, HLA-DR, CD14, CD141, CD1C, CLEC9A, KIT, CPA3, and TPSB2; (2) the target antigens for the CD4 T cells include one or more selected from the group consisting of FOXP3, CCR7, SELL, GIMAP family, PD-1, TIGIT, CD3, CD7, CD4, IL7R, GRANZYME B, CXCR3, CX3CR1, TIM3, VISTA, LAG3, GITR, CD137; and / or (3) The target antigens for the CD8 T cells include one or more selected from the following group: GZMK, GIMAP4, LAG3, NCR3, S100A4, DKK3, TMIGD2, PD-1, CD3, CD7, CD8, IL7R, GRANZYME B, CXCR3, CX3CR1, TIM3, VISTA, LAG3, GITR and CD137.

8. The system according to any one of claims 1-7, wherein the target antigens determined by the analysis selection module for the preparation module include one or more selected from the group consisting of DKK3, FOLR2, FOXP3, PD1, CD8, CD4 and DAPI.

9. according to the system described in any one of claims 1-8, wherein the multiplex immunohistochemical technique is selected from the group consisting of tyramide signal amplification (TSA), cyclic immunofluorescence (CyCIF), indexed co-detection (CODEX), multiplexed ion beam imaging (MIBI), imaging mass spectrometry (IMC), OPAL imaging, multiple error robust fluorescence in situ hybridization (MERFISH), stochastic optical reconstruction microscopy (STORM) and Exchange-PAINT imaging.

10. The system according to any one of claims 1-9, wherein the identification module comprises a component selected from the group consisting of: (A) Components used to predict the half inhibitory concentration (IC50) of tumor tissue; (B) Components used to construct the CCIM scoring system; (C) a component for predicting the effects of chemotherapy, radiotherapy, immunotherapy, and targeted therapy by analyzing the spatial relationship between myeloid cells and T cells in the CCIM image; (D) a component for predicting prognosis; and (E) Components used to predict metastasis.

11. The system according to any one of claims 1 to 10, wherein the sample of the tumor tissue is a section of the tumor tissue, preferably a paraffin section.

12. Use of a composition in the preparation of a kit for predicting the therapeutic effect of colorectal cancer, wherein the composition comprises a target antigen selected from the group consisting of: (1) the target antigens for the myeloid cells include one or more selected from the group consisting of CD68, FOLR2, C1QA, C1QB, SPP1, APOE, CD11B, CD172A, S100A8, S100A9, S100A12, CD15, CD33, CD66B, MPO, CD11C, HLA-DR, CD14, CD141, CD1C, CLEC9A, KIT, CPA3, and TPSB2; (2) the target antigens for the CD4 T cells include one or more selected from the group consisting of FOXP3, CCR7, SELL, GIMAP family, PD-1, TIGIT, CD3, CD7, CD4, IL7R, GRANZYME B, CXCR3, CX3CR1, TIM3, VISTA, LAG3, GITR, CD137; and / or (3) The target antigens for the CD8 T cells include one or more selected from the following group: GZMK, GIMAP4, LAG3, NCR3, S100A4, DKK3, TMIGD2, PD-1, CD3, CD7, CD8, IL7R, GRANZYME B, CXCR3, CX3CR1, TIM3, VISTA, LAG3, GITR and CD137.

13. The use according to claim 12, wherein the composition comprises a target antigen selected from the group consisting of DKK3, FOLR2, FOXP3, PD1, CD8, CD4 and DAPI.

14. Use of a composition comprising a therapeutic agent targeting FOLR2+ resident macrophages in the preparation of a medicament for improving the chemotherapy effect in patients with colorectal cancer, preferably the therapeutic agent targeting FOLR2+ resident macrophages comprises an anti-CSF1R therapeutic agent.

Images