MALMPS model for identifying II / III stage colorectal cancer patients with high-risk recurrence and application

Abstract

The invention provides an MALMPS model for identifying II / III stage colorectal cancer patients with high-risk recurrence and application, and belongs to the technical field of prognosis models of colorectal cancer. The construction of the model comprises the following steps: collecting a clinical sample data set, preparing a data set transcriptome sequencing library, carrying out LC-MS / MS metabonomics analysis on samples in the data set, preprocessing LC-MS / MS data, exporting a three-dimensional data matrix in a CSV format, and establishing the metabolism-related gene prognosis model. A stable and powerful prognosis model based on a machine learning algorithm is established through metabolism-related genes. These findings indicate that MALMPS may be a valuable tool for predicting the recurrence risk of stage II / III colorectal cancer within a specific time range, in particular for identifying high-risk populations.

Description

Technical Field

[0001] The present invention relates to the technical field of prognostic models for colorectal cancer, and in particular to a MALMPS model for identifying patients with stage II / III colorectal cancer at high risk of recurrence and its application. Background Art

[0002] Colorectal cancer (CRC) is a common disease with high mortality and wide differences in individual prognosis, and stage II / III CRC accounts for nearly 70% of the CRC population. Metabolism-related molecular models have the potential to predict cancer progression.

[0003] Colorectal cancer (CRC) is the second and third most common cancer in women and men, respectively, and the second leading cause of cancer-related death, bringing a heavy disease burden to the population. TNM staging is a key prognostic factor for making treatment decisions, and stage II / III colorectal cancer accounts for approximately 70% of all cases. However, despite advances in radiotherapy and chemotherapy, the overall survival rate of stage II / III colorectal cancer remains approximately 70% and 50%, respectively. In addition, the prognosis of stage II / III CRC varies greatly, and there is still a considerable probability of encountering cancer recurrence and progression at low stages. Evidence from previous studies also showed that patients who routinely receive adjuvant chemotherapy after surgery do not respond equally. These phenomena highlight the heterogeneity between stage II / III CRC patients. Therefore, the current staging system, although highly instructive, is not accurate enough to identify individuals with poor prognosis. At this stage, there is an urgent need to develop a new prognostic model to better help predict the risk of recurrence in stage II / III CRC.

[0004] Several studies have developed molecular models for predicting the prognosis of stage II / III CRC, including lncRNAs, markers, and immune-based markers. The predictive performance of different studies varies, and insufficient attention has been paid to adequate validation and trials of different modeling algorithms. At the same time, few studies have specifically focused on the role of CRC metabolic reprogramming in prognosis. Metabolism plays an important role in the occurrence, growth, and spread of CRC, and metabolic adaptation is closely related to carcinogenic pathways. At the same time, further improvement of modeling algorithm combinations and model validation are needed to improve credibility. With the advancement of sequencing technology, machine learning is increasingly used to predict patient prognosis based on genomic and transcriptomic data. Clinicians and researchers are increasingly applying machine learning techniques such as neural networks, ensemble methods, and deep learning for classification and prediction. These techniques are validated by methods such as k-fold cross-validation and hold-out. Summary of the invention

[0005] The technical problem to be solved by the present invention is that in view of the shortcomings of the above-mentioned prior art, the present invention screens metabolic-related genes related to clinical significance in stage II / III CRC by combining a machine learning framework with 10-fold cross-validation or bootstrap resampling, and constructs a machine learning-based metabolic gene prognostic model (MALMPS). The model has been verified in five other independent data sets. The present invention provides a MALMPS model and application for identifying patients with stage II / III colorectal cancer at high risk of recurrence.

[0006] In order to achieve the purpose of the present invention, the following technical scheme is adopted: A MALMPS model for identifying patients with stage II / III colorectal cancer at high risk of recurrence, wherein the method for constructing the MALMPS model specifically comprises the following steps: S1. Collect clinical sample data sets and prepare libraries for transcriptome sequencing of the data sets. Perform LC-MS / MS metabolomics analysis on samples in the data sets. Preprocess the LC-MS / MS data and export them into a three-dimensional data matrix in CSV format to establish a prognostic model for metabolism-related genes. S2. Single gene expression analysis in colorectal cancer cell lines, calculating the correlation coefficient between MALMPS score and the expression of each gene, performing spatial transcriptomic analysis and cell type estimation; S3, perform cell culture and transfection, real-time quantitative polymerase chain reaction, cell counting kit-8 detection and Transwell invasion assay, use apoptosis kit and cell cycle detection kit to evaluate apoptosis and cell cycle status, perform cell scratch assay and drug sensitivity estimation and verification; S4, perform cell viability assay, colony formation assay, perform proteomics and phosphoproteomics analysis, use NMF to perform consensus clustering of immune infiltration, and perform immune cell infiltration and feature estimation; S5. Construct a nomogram.

[0007] Preferably, the MALMPS model identifies colorectal cancer patients with different clinical outcomes.

[0008] Preferably, in the MALMPS model, the COX7B gene is identified as an oncogenic factor in CRC tumorigenesis.

[0009] Further preferably, the clinical net benefits of different models as well as full strategies and no strategy are obtained by decision curve analysis (DCA), and the prognostic performance of the comprehensive model is compared by AUC and restricted mean survival (RMS) curves.

[0010] Another technical solution of the present invention: application of the MALMPS model in analyzing gene mutation characteristics.

[0011] Another technical solution of the present invention: application of the MALMPS model in drug sensitivity prediction.

[0012] As another technical solution of the present invention: the MALMPS model is used to predict the recurrence risk of stage II / III colorectal cancer within a specific time range.

[0013] Compared with the prior art, the present invention achieves the following technical effects: The present invention establishes a stable and powerful prognostic model based on a machine learning algorithm through metabolism-related genes. These findings suggest that MALMPS may be a valuable tool for predicting the risk of recurrence of stage II / III colorectal cancer within a specific time frame, especially for identifying high-risk populations. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] For ease of explanation, the present invention is described in detail with reference to the following specific embodiments and the accompanying drawings.

[0015] Figure 1 It is the overall flow chart of the present invention.

[0016] Figure 2 Identification diagram of the best performing model, where: (A) C-index of 83 machine learning prediction model combinations in five independent validation sets; (B) C-index of MALMPS in all datasets; (C) C-index comparison of clinical and molecular variables with MALMPS scores in GSE29621, GSE143985, GSE17536, GSE92921, training set cohort, and GSE161158; * indicates P <0.05, ** indicates P <0.01, *** indicates P <0.001.

[0017] Figure 3 Diagram of the model building process, including: (A) the number of trees required to generate features with minimal error using the RSF algorithm; (B) performance curve showing the optimal number of boosting iterations in the GBM algorithm; (C) heat map revealing the relationship between the expression of genes in the model and clinical pathological characteristics.

[0018] Figure 4Figure 2 is the performance evaluation graph of MALMPS, where: (A) time-dependent ROC analysis for predicting 1-year, 3-year, and 5-year PFS in the training cohort and all validation datasets; (B) C-index comparison of MALMPS score with five published models in TCGA-CRC, GSE17536, GSE29621, GSE92921, GSE143985, and GSE161158; (C) Kaplan-Meier survival curves of PFS for patients with high MALMPS scores versus patients with low MALMPS scores in all datasets; * indicates P <0.05, ** indicates P <0.01, *** indicates P <0.001.

[0019] Figure 5 Figure 3. COX7B promotes malignant behavior of CRC, including: (A) Importance of 21 most valuable metabolic genes determined by RSF algorithm calculated by GBM algorithm; (B) Kaplan-Meier survival curves of PFS in patients with high and low COX7B mRNA expression. (C) Heat map and violin plots showing the distribution of ST unsupervised cluster analysis and COX7B expression in different cell types in CRC. (D) Bar graph showing COX7B expression levels within COX7B-scr, shCOX7B#1, and shCOX7B#2 groups.

[0020] Figure 6 Validation of COX7B malignant expression, where: (A) Heat map and violin plots showing unsupervised cluster analysis distribution of STs and COX7B expression in different cell types in liver metastases.

[0021] Figure 7 Figure 1 shows the mutation signature analysis, where: (A) MALMPS mutation landscape of TCGA-CRC divided into low-MALMPS group and high-MALMPS group. Each column represents an individual patient, sorted from left to right by MALMPS score. (B) Normalized distribution of SCNA and TMB levels in high-risk group and low-risk group. (C) TMB landscape sorted by value. (D) NMF cohesion metric curve for selecting matching mutation signatures. (E) Heat map showing the best matching signatures with COSMIC database signatures by cosine similarity. (F) Mutational activity of the corresponding extracted mutation signatures (signatures 10, 1 and 6 matching with COSMIC database). (G) Bar chart of the relative percentage of extracted mutation signatures in low-risk group and high-risk group. (H) Relative distribution of mutation signature COMSIC 6 in high-risk group and low-risk group. * indicates P <0.05, ** indicates P<0.01, *** indicates P <0.001.

[0022] Figure 8 Figure 1 shows the molecular signature of the model, where: (A) The positive and negative correlations between MALMPS and the top five pathways in the GO, KEGG, and REACTOME databases were analyzed. (B) The expansion of the marker interaction network from the KEGG database between risk groups. The color represents the direction of the fold change. The size of the circle represents the number of enriched genes.

[0023] Fig. 9 The drug sensitivity prediction and validation diagram of MALMPS, where: (A) Workflow of drug sensitivity prediction and validation. (B) Relative distribution of drug sensitivity of fluorouracil, oxaliplatin, MK-1775, BMS-754807 and XAV939. (C) Line graph showing cell viability of high-risk group and low-risk group when using five determined drugs and control drugs. (D) Bar graph showing colony number of high-risk group and low-risk group when using five determined drugs and control drugs. (E) Representative staining images showing colony number between risk groups in different cell lines. * indicates P <0.05, ** indicates P <0.01, *** indicates P <0.001.

[0024] Fig.10 Pie chart showing the types and quantities of metabolites detected in the SDPH cohort.

[0025] Fig.11 Metabolomics and proteomics feature maps of MALMPS, including: (A) Overview of metabolic correlation analysis in the SDPH cohort. (B) Clinical annotation of stage II / III colorectal cancer in the SDPH cohort. (C) Metabolic pathway enrichment annotation between risk groups against the REACTOME database. (D) Forest plot illustrating metabolite differences between risk groups in the SDPH cohort. (E) Top 40 kinases enriched in high-risk and low-risk groups using KSEA, with a significance of P < 0.05. (F) Heat map showing representative biological pathways at the phosphoprotein level between high-risk and low-risk groups in the CPTAC cohort. (G) Volcano plot showing differential protein analysis between different MALMPS risk groups.

[0026] Fig.12Figure 1 shows the correlation between immune infiltration and MALMPS, where: (A) The relationship between MALMPS and immune cell infiltration. (B) A heat map showing the levels of 28 immune cell infiltrations. Samples are sorted from left to right by MALMPS scores. (C) The relative distribution of 28 immune cell infiltrations in the high-risk group and the low-risk group. * indicates P <0.05, ** indicates P <0.01.

[0027] Fig.13 Figure 1 shows the integration of immune infiltration subtypes and their integration with MALMPS, where: (A) Application of NMF index in immune infiltration subtype classification. (B) Consensus matrix of NMF with two subtype classifications. (C) Correlation of MALMPS with T cell exhaustion and fatty acid elongation. (D) Heat map exploring subtype characteristics with 22 relative levels of immune cell infiltration in CIBERSORT. Kaplan-Meier survival curves of PFS compared between immune infiltration subtypes (E) and subgroups (F). (G) Heat map of subtype characteristics with 30 significantly different relative levels of immune cell infiltration validated by xCell algorithm. (H) Expansion of the signature interaction network from KEGG database between subtypes. Color represents the direction of fold change. The size of the circle represents the number of enriched genes.

[0028] Fig.14 Figure 1 shows the combined plots of MALMPS and molecular features, including: (A) Prognostic nomogram for predicting 1-year, 3-year, and 5-year PFS probabilities in TCGA dataset. (B) RMS plots of Bio, MALMPS, and BioMALMPS in GSE92921 and GSE143985. (C) AUC of ROC curves of Bio, MALMPS, and BioMALMPS for 1-year, 3-year, and 5-year PFS probabilities. (D) DCA of Bio, MALMPS, and BioMALMPS models for 1-year, 3-year, and 5-year PFS prediction in TCGA-CRC.

[0029] Fig.15 Figure 1 shows the malignant behavior of COX7B in promoting CRC, where: (E) The scatter plot and bar graph show the trend of the proportion of apoptotic cells. (F) The histogram and bar graph show the proportion of cells in different cell cycle stages. (G) The line graph shows the proliferation doubling of the three groups over time. (H) The representative staining images and bar graphs show the colony numbers of the three groups. (I) The representative staining images and bar graphs show the invasion and migration Transwell numbers of the three groups. * indicates P <0.05, ** indicates P <0.01, *** indicates P <0.001.

[0030] Fig.16 Figure 1 is a validation diagram of COX7B malignant expression, where: (B) Box plot showing Chronos scores of 21 genes in MALMPS. (C) Distribution of CA199 levels in high-risk and low-risk groups. (D) Kaplan-Meier survival curves of PFS (progression-free survival) between patients with high and low COX7B protein expression. (E) Top 40 KEGG pathway enrichment annotations between COX7B high and low expression groups. (F) Representative staining images and bar graphs of scratch migration differences between COX7B-scr, shCOX7B#1, and shCOX7B#2 groups. DETAILED DESCRIPTION

[0031] The following are specific embodiments of the present invention and the accompanying drawings to further describe the technical solutions of the present invention, but the present invention is not limited to these embodiments; in the following description, specific details such as specific configurations are provided only to help fully understand the embodiments of the present invention. Therefore, it should be clear to those skilled in the art that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present invention.

[0032] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0033] Unless otherwise specified, the materials, practices and experimental equipment involved in the implementation cases of the present invention are all in line with commercially available products in the relevant chemical and biotechnology fields.

[0034] Example 1. Construction of metabolic gene prognostic model (MALMPS) Gene expression data and corresponding clinical characteristics of colorectal cancer samples were collected from the NCBI GEO database (https: / / www.ncbi.nlm.nih.gov / geo / ) and TCGA (https: / / cancergenome.nih.gov / ) and CCLE database (https: / / sites.broadinstitute.org / ccle / ). Transcriptome data of cell lines were downloaded from the CCLE database. A total of 1442 samples from nine datasets were involved in the construction and validation analysis of the model. The large retrospective follow-up dataset studied by Guinney et al. (including GSE14333, GSE37892, GSE39582, TCGA-CRC) was selected as a training dataset for developing a recurrence prognosis prediction model. Independent test datasets from different regions (GSE17536, GSE29621, GSE92921, GSE143985, GSE161158) were used to select the best model combination and evaluate its predictive performance. The subjects came from various hospitals around the world, including the United States, Australia, France, and Japan. The study subjects were colorectal cancer patients who voluntarily provided surgically resected tumor tissue and normal tissue samples, and all of these patients had primary cancer. Exclusion criteria included individuals with other types of cancer, individuals who received preoperative chemotherapy and / or radiotherapy, or individuals with insufficient tumor-derived total RNA for analysis. Our analysis ensured the consistency of the dataset by excluding samples with zero survival time, retaining only individuals with stage II / III colorectal cancer, and adjusting the survival time unit to months. In addition, subjects with missing recurrence prognosis data, staging information, and transcriptome data were also excluded from this study. A training set of 1,000 subjects was used to develop a prognostic model, and five independent validation set cohorts included 111, 40, 141, 59, and 91 subjects, respectively, to assess the robustness of the model. This work met the reporting requirements of the STROCSS standards.

[0035] Collection of clinical samples of SDPH dataset A total of 73 stage II / III colorectal cancer samples were collected from the Shandong provincial hospital (SDPH) internal cohort for transcriptomic and metabolomic analysis. The cohort included patients who underwent surgical resection at SDPH between August 2021 and October 2023. Immediately after surgery, tissue samples were snap-frozen in liquid nitrogen within 30 minutes and subsequently stored at -80°C for further analysis. This study was approved by the hospital ethics committee (approval number 2021-176). Written informed consent was obtained from each patient before participation. Importantly, all patients in this cohort were newly diagnosed with colorectal cancer and had not received any treatment. We obtained detailed clinical information for each patient, such as age, sex, tumor location, tumor-node metastasis, and AJCC stage, as well as the constructed MALMPS risk score and corresponding grouping. Transcriptomic data were log-2 transformed in TPM format and normalized to Z scores for further analysis.

[0036] Library preparation for transcriptome sequencing of SDPH dataset In the SDPH dataset, total RNA was extracted from tissues using TRIzol ® The reagents were operated according to the manufacturer's instructions. Then, RNA quality was tested using a 5300 Bioanalyzer (Agilent) and quantified using an ND-2000 (NanoDrop Technologies). Only high-quality RNA samples (OD260 / 280 = 1.8~2.2, OD260 / 230 ≥ 2.0, RIN ≥ 6.5, 28S:18S ≥ 1.0, > 1 μg) were used to construct sequencing libraries.

[0037] RNA purification, reverse transcription, library construction, and sequencing were performed at Shanghai Meiji Biopharmaceutical Technology Co., Ltd. (Shanghai, China) according to the manufacturer's instructions (Illumina, San Diego, CA). RNA-seq transcriptome library preparation was performed according to Illumina ®Stranded mRNA Prep, Ligation protocol (Illumina, San Diego, CA) was performed using 1 μg of total RNA. Briefly, messenger RNA was first isolated by oligo(dT) beads according to the polyA selection method and then fragmented by fragmentation buffer. Second, double-stranded cDNA was synthesized using the SuperScript double-stranded cDNA synthesis kit (Invitrogen, CA) and random hexamer primers (Illumina). The synthesized cDNA was then end-repaired, phosphorylated, and “A” base added according to Illumina’s library construction protocol. Subsequently, the library was selected for a 300-bp cDNA target fragment on 2% low-range super agarose and amplified using Phusion DNA polymerase (NEB) for 15 cycles of PCR. After quantification by Qubit 4.0, paired-end RNA-seq sequencing (read length of 2 × 150 bp) was performed using a NovaSeq Xplus sequencer.

[0038] LC-MS / MS metabolomics analysis of SDPH dataset In the SDPH dataset, 50 mg of solid sample was added to a 2 mL centrifuge tube and a 6 mm diameter grinding bead was added. Metabolite extraction was performed using 400 μL of extraction solution (methanol: water = 4:1) containing 0.02 mg / mL internal standard (L-2-chlorophenylalanine). The sample was ground in a Wonbio-96c (Shanghai Wanbo Biotechnology Co., Ltd.) frozen tissue grinder for 6 minutes (-10°C, 50 Hz), followed by 30 minutes of low-temperature ultrasonic extraction (5°C, 40 kHz). After the sample was placed at -20°C for 30 minutes, it was centrifuged for 15 minutes (4°C, 13,000 g), and the supernatant was transferred to a sample injection vial for LC-MS / MS analysis.

[0039] For liquid samples, 100 μL of liquid sample was added to a 1.5 mL centrifuge tube and 400 μL of extraction solution (acetonitrile: methanol = 1:1) containing 0.02 mg / mL internal standard (L-2-chlorophenylalanine) was added for metabolite extraction. The sample was vortexed for 30 seconds and extracted by sonication at low temperature for 30 minutes (5°C, 40 kHz). The sample was placed at -20°C for 30 minutes to precipitate the protein and then centrifuged for 15 minutes (4°C, 13000g). The supernatant was removed and blown dry under nitrogen. The sample was then reconstituted with 100 μL of solution (acetonitrile: water = 1:1) and extracted by sonication at low temperature for 5 minutes (5°C, 40 kHz) and then centrifuged at 13000g and 4°C for 10 minutes. The supernatant was transferred to a sample vial for LC-MS / MS analysis. As part of the system commissioning and quality control process, a pooled quality control sample (QC) was prepared by mixing equal volumes of all samples. QC samples were processed and tested in the same manner as analytical samples, were representative of the entire sample set, and were injected every 5–15 samples to monitor analytical stability.

[0040] LC-MS / MS analysis of samples was performed on a ThermoUHPLC-Q Exactive system equipped with an ACQUITY HSS T3 column (100 mm × 2.1 mm id, 1.8 μm; Waters, USA) at Majorbio Bio-Pharm Technology Co. Ltd. (Shanghai, China). The mobile phase consisted of 0.1% formic acid in water:acetonitrile (95:5, v / v) (solvent A) and 0.1% formic acid in acetonitrile:isopropanol:water (47.5:47.5, v / v) (solvent B). The flow rate was 0.40 mL / min and the column temperature was 40°C.

[0041] The UPLC system was connected to a Thermo UHPLC-Q Exactive mass spectrometer equipped with an electrospray ionization (ESI) source and operated in positive and negative ion modes. The optimal conditions were set as follows: source temperature 400°C; sheath gas flow rate 40 arb; auxiliary gas flow rate 10 arb; ion spray voltage floating (ISVF) was -2800 V in negative ion mode and 3500 V in positive ion mode; normalized collision energy was 20-40-60 V rolling for MS / MS. The full MS resolution was 70000 and the MS / MS resolution was 17500. Data acquisition used data dependent acquisition (DDA) mode. The detection range was 70-1050 m / z.

[0042] Metabolomics data processing of SDPH dataset In the SDPH data set, the preprocessing of the LC / MS raw data was performed by Progenesis QI (Waters Corporation, Milford, USA) software and exported as a three-dimensional data matrix in CSV format. The information in the three-dimensional matrix includes: sample information, metabolite name, and mass spectrometry response intensity. The internal standard peaks and any known false positive peaks (including noise, column bleeding, and derivatization reagent peaks) in the data matrix were removed for redundancy removal and peak merging. At the same time, metabolites were identified by database searches, including HMDB (http: / / www.hmdb.ca / ), Metlin (https: / / metlin.scripps.edu / ), and Majorbio databases.

[0043] The data were analyzed using the Majorbio cloud platform (cloud.majorbio.com). Metabolic features that detected at least 80% in any sample set were retained. After filtering, for specific samples with metabolite levels below the lower limit of quantification, the minimum metabolite value was substituted, and each metabolic feature was normalized by the sum normalization method. In order to reduce errors caused by sample preparation and instrument instability, the response intensity of the sample mass spectrometry peak was normalized by the sum normalization method, and then the normalized data matrix was obtained. At the same time, QC sample variables with relative standard deviation (RSD) > 30% were removed, and log10 processing was performed to obtain the final data matrix for subsequent analysis.

[0044] Establishment of prognostic model of metabolism-related genes In our study, the outcome variable was survival data related to recurrence, which was the duration from surgery to recurrence or follow-up. The predictor variable in the model was the expression of metabolic genes we screened. Since low-expressed or unchanged genes usually represent noise, the transcriptome data in the training dataset consisted of nearly 6,000 genes with the largest median absolute deviation (MAD). These genes were measured by at least one probe set in all datasets, and each gene was represented by the probe set with the largest MAD. We extracted 84 and 82 metabolism-related pathways from the KEGG and REACTOME databases, containing 1,715 and 2,679 genes, respectively. A total of 1,053 overlapping genes were used for further analysis. We took the intersection of these genes and performed univariate Cox regression analysis on these genes. A total of 149 metabolism-related prognostic genes were included in the machine learning framework.

[0045] Ten independent machine learning algorithms and their combinations constitute the machine learning framework. The ten algorithms include Random Survival Forest (RSF), Elastic Net (Enet), Lasso, Ridge, Stepwise Cox, CoxBoost, Partial Least Squares Regression for Cox (plsRcox), Supervised Principal Components (SuperPC), Generalized Boosted Regression Modeling (GBM), and Survival Support Vector Machine (survival-SVM). Six of the algorithms, including RSF, Enet, Lasso, Ridge, Stepwise Cox, and CoxBoost, can perform feature selection. A total of 83 combined algorithms based on 10-fold cross validation (CV) or bootstrap resampling methods are used to select the best performing model. The algorithm with the highest average C index on all test datasets is considered the best model for the generated model.

[0046] Single gene expression analysis in colorectal cancer cell lines We used the Chronos score to infer the fitness effects of genes in cell lines after knocking out signature genes (https: / / github.com / broadinstitute / chronos). The Chronos algorithm uses a model that uses the dynamics of cell proliferation after clustered regularly interspaced short palindromic repeats (CRISPR) gene knockout to infer the fitness effects of gene knockout. The score generated by the Chronos algorithm measures the effect of a specific gene knockout on cell proliferation and fitness.

[0047] Gene set enrichment analysis The correlation coefficient between the MALMPS score and the expression of each gene was calculated. The ranked correlation coefficients were used as input for gene set enrichment analysis (GSEA) using the 'clusterProfiler' package against the GO, KEGG and REACTOME reference gene sets. The false discovery rate ( FDR ) < 0.05 was considered the cutoff value for enriched gene sets.

[0048] The 'limma' package was used to assess differential expression of over 18,000 genes in samples with different risk groups and immune infiltration patterns. Gene expression data were input into lmFit and eBayes functions to calculate differential statistics. GSEA was performed on the extracted 84 KEGG metabolic gene sets using the fast GSEA algorithm using the ordinal logFC generated by limma.

[0049] Spatial transcriptomic analysis and cell type estimation The processed spatial transcriptomics (ST) data were obtained from the study by Wu et al. This set of ST data contains tissue structural annotations of eight samples (paired samples of colorectal cancer and liver metastasis) from two untreated patients and two patients who received neoadjuvant chemoradiotherapy, which were provided by the authors of the study. We used SCTransform for data normalization, RunPCA for dimensionality reduction, FindNeighbors and FindClusters for clustering of ST points, and RunUMAP for data visualization. Subcellular types in ST points were annotated using the ssGSEA algorithm.

[0050] Cell culture and transfection The LS180 cell line (RRID: CVCL_0397) was purchased from the Cell Resource Center of Peking Union Medical College, which belongs to the National Biomedical Cell Resource Bank (NSTI-BMCR) of the National Science and Technology Basic Conditions Platform. Cell culture was performed in MEM (KeyGEN BioTECH). The HCT116 cell line (RRID: CVCL_0291) was obtained from Procell and cultured in McCoy's 5A medium (KeyGEN BioTECH). In addition, other colorectal cancer cell lines (RKO [RRID: CVCL_0504], HCT15 [RRID: CVCL_0292], SW480 [RRID: CVCL_0546], SW620 [RRID: CVCL_0547], DLD-1 [RRID: CVCL_0248]) were purchased from ATCC and cultured in RPMI-1640 (Gibco) medium, except for RKO, which was cultured in DMEM (Gibco). All the above culture media were supplemented with 10% fetal bovine serum (PAN), penicillin (100 U / mL, Thermo Fisher), and streptomycin (100 U / mL, Thermo Fisher). Cells were maintained in an atmosphere of 95% air and 5% CO2 at 37°C.

[0051] All human cell lines were authenticated by STR (or SNP) profiling within the past three years, and cells used in all experiments were mycoplasma-free.

[0052] Real-time quantitative PCR Total RNA was extracted from cells using TRIzol reagent and reverse transcribed into cDNA using HiScripIII RTSuperMix according to the manufacturer's instructions. Real-time quantitative polymerase chain reaction (qRT-PCR) was performed using the Applied Biosystems QuantStudio 1 Real-Time PCR System (AppliedBiosystems, ThermoFish) and ChamQ Universal SYBR qPCR Master Mix. All reagents were purchased from Vazyme, China. The relative expression level of mRNA was measured by 2 -ΔΔCt Method calculation, the higher of 2 -ΔΔCt Reflects higher expression levels.

[0053] Cell Counting Kit-8 (CCK-8) Assay 3000 cells were seeded per well in a 96-well plate, and 10 µL of CCK-8 reagent (DojinDo, Japan) was added to each well after the cells adhered. After incubation at 37°C for 2 hours, the absorbance of each well was measured at a wavelength of 450 nm using Multiskan Sky (Thermo Scientific) at 0 hours, 24 hours, 48 ​​hours, 72 hours, and 96 hours. Cell viability was calculated and plotted using GraphPad Prism 8 software. All experiments were repeated three times.

[0054] Transwell assay Add 600 μL of 1640 medium to the lower chamber and 200 μL of serum-free cell suspension (migration: 4 × 10 5 / mL, invasion: 12×10 6 / mL), Matrigel was used for invasion assays but not for migration assays. After 24 hours of incubation, the upper chamber was fixed with 4% paraformaldehyde for 30 minutes and stained with 0.1% crystal violet for 30 minutes. The number of cells was randomly counted using an optical microscope.

[0055] Apoptosis and cell cycle analysis Apoptosis and cell cycle status were assessed using an apoptosis kit (BD, Pharmingen™ PE Annexin V Apoptosis Detection Kit I) and a cell cycle detection kit (KeyGEN BioTECH, Cell Cycle Detection Kit), respectively. Samples were prepared according to the instructions and analyzed by CytoFLEX S flow cytometer (Beckman Coulter).

[0056] Cell scratch assay Cells were seeded in 6-well plates. When cells covered the bottom of the wells, three scratches were made in each well using a sterile 10 µL pipette tip. Cells were cultured continuously for 72 hours using the same culture system (RPMI-1640 medium + 1% FBS). The scratch width in three randomly selected microscopic fields was recorded at specific time points (0 hours, 24 hours, 48 ​​hours, 72 hours). The scratch closure rate was calculated as the ratio of the migration distance to the initial scratch distance.

[0057] Drug susceptibility estimation and validation The prediction program was constructed using the 'oncoPredict' package. Based on the expression matrix of the training set with known drug treatment information, drug prediction was performed against the Genomic Drug Sensitivity of Cancer (GDSC) database. Ridge regression was used to calculate the drug sensitivity score of the samples.

[0058] Cell viability assay According to the risk score determined by MALMPS using transcriptome data, the cell lines were divided into low-risk and high-risk groups for drug sensitivity determination. The cells were cultured with the corresponding culture medium containing 10% serum, digested to form a single cell suspension, and evenly inoculated into a 24-well plate. The volume of culture medium in each well was 500μL. After 24 hours of culture, when all cells were attached, the culture medium was replaced with the same volume of culture medium but containing different drug concentrations (0, 0.1, 1, 10, 100, 1000μM). After continuous culture for 48 hours, 500μL of basal culture medium and 50μL of CCK-8 solution were added to each well and incubated at 37°C for 2 hours. Using Infinite ® The absorbance of each well was measured three times at a wavelength of 450 nm using an MPlex multi-mode microplate reader and a TECAN monochromator. Cell viability at different drug concentrations was calculated and plotted using GraphPad Prism 8 software. All experiments were repeated three times.

[0059] Colony formation assay For functional assays, 3000 cells were seeded in one well of a 6-well plate and fresh medium was replaced every 3 days until day 7.

[0060] For drug sensitivity assays, 10,000 cells were seeded per well in a 12-well plate and allowed to adhere for 24 hours before being replaced with medium containing drugs (drug concentration at IC50) and incubated for another 48 hours.

[0061] After incubation for an appropriate period of time, the 6-well plate and 12-well plate were washed twice with 4°C precooled PBS, fixed with 4% paraformaldehyde precooled at 4°C for 30 minutes, stained with crystal violet for 30 minutes, and photographed after drying. The number of colonies was calculated using ImageJ software. Each experiment was repeated three times.

[0062] Genomic mutation characteristics The samples in the training dataset were divided into high-risk and low-risk groups according to the median MALMPS score. The TCGA-CRC genomic data were organized by the ‘TCGAbiolink’ package. The ‘maftools’ package was used for mutational profiling and feature extraction. The ExtractSignatures function based on Bayesian variant non-negative matrix factorization decomposed the mutation matrix into two non-negative matrices ‘signatures’ and ‘contributions’, where ‘signatures’ represent the mutation process and ‘contributions’ represent the corresponding mutation activities. The SignatureEnrichment function can automatically determine the optimal number of extracted mutational features and assign them to each sample based on mutation activity. The extracted CRC mutational portraits were compared and annotated with the Catalog of Somatic Mutations in Cancer (COSMIC-V2) database by cosine similarity analysis. Somatic mutation and copy number variation (SCNA) segment data were downloaded from the study of Teresa et al.

[0063] Proteomic and phosphoproteomic analyses CPTAC generates comprehensive proteomic and genomic data primarily from clinical cohorts.

[0064] In our study, we obtained transcriptomic, proteomic, and clinical data from the cBioPortal platform (https: / / www.cbioportal.org / study / summary?id=coad_cptac_2019). The subjects were 79 patients diagnosed with stage II / III colorectal cancer, and we used MALMPS to calculate the risk score. Subsequently, these patients were divided into different risk groups using a consistent method. The proteomic data used for MALMPS gene identification were downloaded from the ProteomeXchange Consortium (identifier: PXD046566).

[0065] PTM-Signature Enrichment Analysis (PTM-SEA) is a modified version of ssGSEA designed for site-specific signature analysis by scoring bidirectional feature sets from PTMsigDB. In the data matrix, each row represents a single phosphorylation site located at a specific amino acid residue, and the measured abundance of the samples is specified in the columns. KSEA (Kinase-Substrate Enrichment Analysis) is a method used to study the enrichment of protein kinase substrates. This method reveals the role of kinases in cell signaling by analyzing the enrichment of protein kinase substrates, and the analysis was performed using the 'KSEAapp' package. To identify differentially expressed proteins, we used the 'limma' package, setting strict criteria with a p-value threshold of less than 0.05 and a fold change (FC) greater than 1.5.

[0066] Consistent clustering of immune infiltrates using NMF The abundance of 22 cell types (matrix A) was decomposed into two non-negative matrices W and H (i.e. A≈WH) by the CIBERSORT algorithm. The matrix A was decomposed repeatedly and its output was aggregated to obtain the consistent clustering of the samples. The optimal number of clusters was selected based on cophenetic, dispersion and silhouette coefficient. Consensus clustering was performed using the 'NMF' package of the brunet algorithm.

[0067] Immune cell infiltration and signature estimation Single-sample gene set enrichment analysis (ssGSEA) was introduced to explore the relationship between MALMPS score and the relative abundance of 28 immune cell types through the ‘GSVA’ package. The relative abundance of each immune cell type in the ssGSEA analysis was expressed as an enrichment score and then normalized to a uniform distribution from 0 to 1. The biological similarity of infiltrating immune cells was estimated by multidimensional scaling analysis (MDS) and Gaussian fitting models. The ‘IOBR’ package integrated 255 published signature gene sets related to tumor microenvironment, tumor metabolism, exosomes, microsatellite instability, etc. The ssGSEA method was used in the signature score evaluation process. CIBERSORT (http: / / cibersort.stanford.edu / ) and xCell algorithms presented immunological molecular features between different risk groups through the ‘CIBERSORT’ and ‘xCell’ packages. CIBERSORT and XCell can be used to estimate the abundance of 22 different leukocyte subsets and 64 infiltrating immune and stromal cell subsets with gene expression profiles. Another package 'ESTIMATE' algorithm was also used to estimate the stromal fraction and immune fraction of tumor samples.

[0068] Construction and validation of the nomogram A new nomogram for predicting recurrence in stage II / III colorectal cancer (CRC) patients was constructed by the 'rms' package. The 2021 National Comprehensive Cancer Network (NCCN) guidelines recommend testing seven biomarkers (i.e., KRAS, NRAS, BRAF, microsatellite instability, mismatch repair, ERBB2 amplification, and NTRK fusion) to determine the best clinical management. Three biomarkers were common in the TCGA and test datasets (GSE92921 and GSE143985). We combined these molecular features with our model, applied Cox proportional hazards regression to construct a comprehensive model, and constructed a nomogram. The clinical net benefits of different models, as well as full strategy and no strategy, were obtained by decision curve analysis (DCA), and the prognostic performance of the comprehensive model was compared by AUC and restricted mean survival (RMS) curves.

[0069] Example 2: MALMPS model verification The overall research design process is as follows Figure 1 shown.

[0070] We screened 1053 overlapping genes from KEGG and REACTOME metabolic pathways and performed univariate Cox regression analysis on them, identifying 149 metabolism-related prognostic genes in the training set. The expression profiles of these genes were input into the machine learning-based modeling framework. Using 10 machine learning algorithms, we fitted 83 prediction models based on 10-fold cross validation (CV) or bootstrap resampling, and further calculated the C index of each model in all validation datasets. In the five test datasets, the strongest model with the highest average C index was a combination of random survival forest (RSF) and generalized boosted regression modeling (GBM) ( Figure 2 A), with C-index ranging from 0.62 to 0.86 (training dataset: 0.862 [95%CI: 0.841-0.882], GSE17536: 0.698 [95%CI: 0.635-0.760], GSE29621: 0.673 [95%CI: 0.548-0.798], GSE92921: 0.740 [95%CI: 0.607-0.871], GSE143985: 0.705 [95%CI: 0.614-0.797], GSE161158: 0.619 [95%CI: 0.556-0.682], Figure 2 B). In the determined model combination, we obtained the best mtry and nodesize parameters with the lowest out-of-sample error when ntrees was 845. The appropriate number of ntrees was selected based on the error rate plot, and features were selected based on variable importance through parameter-adjusted RSF analysis ( Figure 3A). The selected features are then further fed into GBM to fit the prediction model with the best number of iterations ( Figure 3 B). Finally, the expression of the 21 gene signatures was used to calculate a risk score for each patient according to their relative importance in the GBM model, generating a machine learning-based metabolic gene prognostic model (MALMPS, Figure 3 C). We compared the prognostic performance of MALMPS with commonly used clinical and molecular features, including age, sex, race, AJCC stage, grade, driver gene mutations [KRAS, TP53, and BRAF], and microsatellite instability. The MALMPS score outperformed all other clinical and molecular features in terms of the C index in all independent test datasets. All variables were statistically significant in the training dataset ( P <0.001), some variables such as age, race, sex, BRAF mutation, and MSI status were statistically significant in the validation dataset ( P <0.05, P <0.01 or P <0.001), which highlights the predictive power of the model we constructed ( Figure 2 C).

[0071] Evaluation of MALMPS model performance and prognostic value We used ROC analysis to measure the ability of the MALMPS model to discriminate progression-free survival (PFS) ( Figure 4 A). The average AUC of the training dataset was 0.893, and the model showed consistent 3-year AUC levels in each test dataset (average 0.699). In addition, we systematically collected and reanalyzed five prognostic molecular models in CRC transcriptome datasets. The MALMPS score obtained the highest C index among the five published molecular models in both TCGA (training set) and other independent test datasets ( Figure 4 B). The samples were further divided into high-risk and low-risk groups according to the median predicted MALMPS score. In the training cluster, the PFS time of patients in the high-risk group was significantly shorter than that of patients in the low-risk group ( P <0.0001), and a similar trend was observed in the validation dataset ( P <0.01, GSE29621 and GSE92921 only slightly failed to reach statistical significance due to small sample sizes, P The values ​​are 0.073 and 0.082 respectively. Figure 4 C). Multiple evaluations showed that the MALMPS model performed well in identifying colorectal cancer patients with different clinical outcomes.

[0072] COX7B promotes the malignant behavior of CRC COX7B was identified as one of the most important genes in MALMPS. We further performed prognostic analysis and found that high mRNA expression of COX7B was associated with poor PFS in stage II / III CRC patients ( Figure 5 A, Figure 5 B). Spatial transcriptomics (ST) data from eight samples (paired samples of colorectal cancer and liver metastasis) from two untreated patients and two patients undergoing neoadjuvant chemoradiation showed that COX7B was highly expressed in tumor tissues of colorectal and liver metastases ( Figure 5 C, Figure 6 ). We further analyzed the cell proliferation dynamics of 21 genes in CRC cell lines based on the Chronos model, which infers the fitness effects of genes in gene knockout screening. We observed that the Chronos scores in COX7B knockout cell lines were all below zero, indicating its importance for tumor cell survival ( Fig.16 B). In addition, high protein levels of COX7B were associated with poor prognosis and elevated blood CA19-9 markers in CRC ( Fig.16 C, Fig.16 D). We then performed KEGG pathway enrichment analysis and found that high COX7B expression was associated with key metabolic and cellular processes including glycolysis, pyruvate, fatty acid metabolism, and cell cycle ( Fig.16 E). Therefore, we performed comprehensive cellular experiments to investigate the role of COX7B in regulating cell behavior and characteristics. DLD-1 cells were selected for siRNA infection and the efficiency was verified by qPCR ( Figure 5 D). Subsequently, the biological effects of COX7B knockdown were evaluated by flow cytometry, and the results showed an increase in apoptotic cells, an increase in the proportion of cells in the G0 / G1 phase, and a decrease in cells in the S phase ( Fig.15 E, Fig.15 F). Cell proliferation experiments, including CCK-8 and colony formation assays, showed that COX7B promoted cell proliferation ( Fig.15 G, Fig.15 H). In addition, Transwell and cell scratch assays also showed that COX7B promoted the invasion and migration of colorectal cells ( Fig.15 I, Fig.16 F). Taken together, these findings suggest that COX7B may act as an oncogenic factor in the tumorigenesis of CRC.

[0073] Analyzing gene mutation characteristics using the MALMPS model Mutation signature analysis revealed that the tumor driver genes ARID1A, ACVR2A, and ACVR1B had different mutation frequencies in the two MALMPS risk groups ( PThe values ​​are 0.01222, 0.00932 and 0.02389 respectively. Figure 7 A). These three mutant genes were highly enriched in the low-risk group. We further compared the tumor mutation burden (TMB) and somatic copy number variation (SCNA) levels of stage II / III CRC and found that the standardized SCNA levels in the high-risk group were significantly higher than those in the low-risk group ( Figure 7 B, Figure 7 C). In addition, we extracted 3 mutational signatures with different numbers of somatic mutations from the genomic signatures (COSMIC-10, 1, and 6, Figure 7 D, Figure 7 E, Figure 7 F, Figure 7 G). We obtained significantly higher standardized COSMIC 6 scores in the high-risk group ( Figure 7 H).

[0074] Potential biological processes associated with the MALMPS model To elucidate the potential biological mechanisms of MALMPS, we performed gene set enrichment analysis (GSEA) using the correlation coefficients of the three major gene sets associated with MALMPS (GO, KEGG, and REACTOME) as input. We focused on the enriched pathways with the highest and lowest MALMPS scores in the three datasets, according to the normalized enrichment score (NES). In the GO reference dataset, the extracellular matrix and some metabolism-related organelle activities (such as mitochondria and ribosomes) were positively and negatively correlated with the MALMPS scores, respectively. Similarly, the KEGG and REACTOME gene sets showed enrichment for processes related to the extracellular matrix, ribosomes, and mitochondria ( Figure 8 A). Some oncogenic signaling pathways, such as MAPK, JAK, and TGF-β signaling pathways were upregulated in the high-risk group, while cancer suppressor signaling pathways, such as the P53 pathway and DNA repair, were enhanced in the low-risk group ( Figure 8 B).

[0075] Prediction of drug sensitivity using MALMPS model Drug sensitivity analysis of 198 GDSC2-derived anticancer drugs in different MALMPS risk groups was performed using the ‘oncoPredict’ package ( Fig. 9A). It was found that patients in the low-risk group showed higher sensitivity to the main anticancer drugs commonly used in CRC treatment, such as 5-fluorouracil and oxaliplatin. In addition, drugs targeting the cell cycle, such as MK-1775, also showed higher sensitivity in the low-risk group, which was consistent with the GSEA results. In contrast, the high-risk group showed higher sensitivity to drugs such as BMS-754807 (IGF-1R inhibitor) and XAV939 (Wnt / β-catenin inhibitor) ( Fig. 9 B). To further validate these findings, we selected seven CRC cell lines for drug sensitivity experiments according to the risk grouping of MALMPS (high-risk group: DLD1, HCT116, SW480, LS180; low-risk group: RKO, HCT-15, SW620). High-risk group cell lines showed significantly lower colony numbers and cell viability under XAV-939 and BMS-754807 treatment, while 5-fluorouracil, oxaliplatin, and MK-1775 showed the opposite trend ( Fig. 9 C, Fig. 9 D, Fig. 9 E).

[0076] Metabolic bioactivity, metabolite changes, and proteomic differences among MALMPS groups Metabolic bioactivity and metabolite profiles play an important role in understanding metabolic reprogramming in CRC. We used paired metabolome and transcriptome data from 73 stage II / III CRC patients from the SDPH cohort to identify associations between metabolite changes and MALMPS risk groups ( Fig.11 A, Fig.11 B). In this cohort, we analyzed the abundance of more than 2000 metabolites ( Fig.10 ). We observed that in the high-risk group, metabolic bioactivity of carbohydrate and glycosaminoglycan pathways in the REACTOME metabolic pathway was enriched. In addition, our exploration revealed that a series of lipid-related pathways, especially phospholipid, PPAR-α (peroxisome proliferator-activated receptor α) and sphingolipid metabolism were significantly enhanced in the high-risk group. These macromolecules exert key influences on cell properties and functions by directly interacting with cell receptors or interacting with growth factors. At the same time, metabolic pathways of cofactors and complex TP53-regulated metabolic genes were enriched in the low-risk group ( Fig.11C). In terms of metabolites, metabolites related to carbohydrates and glycosaminoglycans, such as glutamyllysine, D-fructose, and ribonucleic acid, were enriched in the high-risk group. Metabolites related to lipid pathways, such as phosphatidylethanolamine (PE), phosphatidylserine (PS), phosphatidic acid (PA), and sphingomyelin (SM), also showed higher relative levels in the high-risk group. Metabolites involving nucleotides (such as 2-deoxyuridine and 2'-deoxyadenosine-5'-phosphate), amino acids (such as L-arginine), and cofactors (such as thiamine [vitamin B1] and pantothenic acid [vitamin B5]) scored higher in the low-risk group ( Fig.11 D).

[0077] In proteomic analysis, several kinases (including MAPK kinases) were enriched in the high-risk group, whereas others (including CDK kinases) were enriched in the low-risk group ( Fig.11 E). Certain phosphorylation pathways, such as IGF1 and MAPK signaling pathways, were enriched in the high-risk group, while mTOR / MTOR was enriched in the low-risk group ( Fig.11 F). IGF-1 can promote the growth and survival of cancer cells and stimulate the formation of new blood vessels to support tumor growth. In our drug sensitivity analysis, IGF-1 inhibitors were observed to be effective in the high-risk group. mTOR is involved in regulating autophagy, a cellular process that degrades and recycles damaged or unnecessary cellular components. Our exploration successfully identified 46 differentially expressed proteins, of which 21 proteins (such as PCSK1, PLIN, SFRP2, LDHAL6B and ATP2B3, etc.) were upregulated in the high-risk group, and 25 (such as ITLN1, TKTL2, ELOVL5 and CXCL10, etc.) were downregulated in the low-risk group ( Fig.11 G).

[0078] Combination of subtypes and MALMPS groups based on immune infiltration We analyzed the immune infiltration landscape with our signature using ssGSEA. Twenty-four of the 28 immune infiltrating cell types were positively correlated with MALMPS scores, 17 of which were statistically significant ( Fig.12 A). As the MALMPS score increases, the level of inflammatory cell infiltration increases ( Fig.12 B). Several types of exhausted T cells, dendritic cells, and macrophages had significantly higher scores in the high-risk group than in the low-risk group ( Fig.12 C).

[0079] To further understand the immunometabolic characteristics of stage II / III CRC, we attempted to classify patients into immune subtypes (IS1 and IS2) based on immune infiltration features using the NMF algorithm ( Fig.13 A, Fig.13B). Molecular pathway analysis using the IOBR algorithm revealed that T cell exhaustion and fatty acid elongation were positively and negatively correlated with MALMPS scores, respectively ( Fig.13 C). According to the clustering results, IS1 and IS2 can be summarized as T cell hot subtype and T cell cold subtype ( Fig.13 D). Based on this finding, we integrated the MALMPS model and immune subtypes into four different combined subgroups (referred to as highIS1, highIS2, lowIS1, and lowIS2 subgroups, respectively). We observed significant prognostic value between immune subtypes and combined subgroups ( P =0.023 and <0.001, Fig.13 E, Fig.13 F). Immune signature verified by xCell algorithm ( Fig.13 G). Cytokine-mediated signaling pathways (IL-2, IL-6), hypoxia, and EMT are enriched in IS1 and IS2, respectively ( Fig.13 H). The ESTIMATE score and immune score indicated that highIS1 was active in immune activity. In addition, we analyzed the expression of several T cell exhaustion and immune checkpoint-related biomarkers in the four subgroups. The expression of these biomarkers (including HAVCR2, PDCD1, PD-L1, and TIGIT) in the HighIS1 group was significantly higher than that in the other groups, indicating the value of immunotherapy in this subgroup.

[0080] Development and validation of a prognostic nomogram model To provide a potential clinical tool for predicting the risk of recurrence within a specific time period, we attempted to build a nomogram model that incorporates the MALMPS score and molecular features. We curated common biomarkers recommended by the NCCN guidelines in the collected dataset, including MSI, BRAF, and KRAS mutations, to fit the biomarker model in the TCGA training set (called Bio). These biomarkers were also combined with MALMPS to construct a nomogram model (called BioMALMPS) ( Fig.14 A) and validated in GSE92921 and GSE143985 datasets. The continuous forms of the MALMPS and BioMALMPS models significantly improved the estimate of PFS in the RMS analysis compared with the Bio model. The average slope of BioMALMPS was greater than that of Bio and MALMPS ( Fig.14 B). In addition, we tested the performance of Bio, MALMPS, and BioMALMPS for 1-, 3-, and 5-year survival prediction using ROC analysis. Significant improvement compared to Bio, with the highest value of BioMALMPS being observed ( Fig.14C). In the DCA analysis, the BioMALMPS model also showed a significant net gain compared with the Bio and MALMPS models when the threshold was less than 0.25 ( Fig.14 D).

[0081] The present invention develops a MALMPS model containing 21 metabolic-related prognostic genes by combining the methods of RSF and GBM, and obtains the highest average C index in an independent validation data set. Compared with traditional survival analysis methods, RSF has the advantages of processing high-dimensional data and preventing overfitting. GBM combines the advantages of Boosting algorithm and generalized regression model to gradually improve the predictive performance of the model. In the MALMPS model, COX7B was identified as a relatively important gene, which combined with Chronos score and prognostic analysis showed its importance in tumor malignancy. In addition, in vitro experiments showed that this gene has an effect on the invasion and migration of tumor cells. COX7B is a component of mitochondrial respiratory chain complex IV and plays a key role in cellular respiration and energy production. As part of the mitochondrial respiratory chain, it may promote cancer through its effects on mitochondrial function and metabolism (such as metabolic reprogramming and oxidative stress). At the same time, in comparison with common clinical and molecular features, MALMPS has the highest C index level, and the AUC of 3-year PFS in all data sets is stable. These findings suggest that our signature may be a promising biomarker for predicting individuals at high risk of recurrence in stage II / III CRC.

[0082] Our study found that the low-risk group had a higher mutation frequency in ACVR2A, which was enriched in MSI-H samples in pan-cancer analysis. MSI-H samples were associated with a lower recurrence rate of stage II / III CRC. The high-risk group showed enrichment of biological activities related to extracellular matrix, cell adhesion, fibrosis, inflammation, thrombosis, cell division, and metastasis. On the other hand, the low-risk group showed enrichment of cell cycle processes, and subsequent drug sensitivity analysis showed that they were sensitive to drugs targeting cell cycle pathways. For example, 5-fluorouracil (5-FU) is an important component of systemic chemotherapy for CRC, while XAV-939 can disrupt Wnt signaling, affecting cell proliferation and survival.

[0083] Our metabolic pathway enrichment analysis revealed that glycosaminoglycan (GAG) and sphingolipid (including glycosphingolipids and phospholipids) metabolism was enriched in the high-risk group, and the abundance of PE, PS, SM, and glutamyl-lysine was obtained by LC-MS. GAGs play a role in cancer progression by affecting various aspects of cancer development through interactions with different partners, which can affect the signaling of growth factors and chemokines, the physical properties of the extracellular matrix, and the spread of cancer cells. Targeting GAGs is a promising approach to interfere with cancer progression. At the same time, GAGs can be used as drug and cytokine carriers to induce anti-tumor responses. Phospholipids are important components of cell membranes and are involved in cell signaling, which can affect tumor growth and metastasis. Sphingolipid metabolism plays a crucial role in promoting tumor survival. Recent studies have shown that sphingolipid signals can control intracellular complement activation, leading to inflammasome-mediated metastasis, presenting a promising strategic measure for cancer treatment.

[0084] Overall, our analysis established a stable and powerful prognostic model based on a machine learning algorithm through metabolism-related genes. These findings suggest that MALMPS may be a valuable tool for predicting the risk of recurrence of stage II / III colorectal cancer within a specific time frame, especially for identifying high-risk populations.

[0085] Those skilled in the art to which the present application belongs may make various modifications or additions to the specific embodiments described or replace them in a similar manner, but they will not deviate from the inventive concept of the present application or exceed the scope defined by the attached claims.

Claims

1. A MALMPS model for identifying patients with stage II / III colorectal cancer at high risk of recurrence, characterized in that: The method for constructing the MALMPS model specifically includes the following steps: S1. Collect clinical sample data sets and prepare libraries for transcriptome sequencing of the data sets. Perform LC-MS / MS metabolomics analysis on samples in the data sets. Preprocess the LC-MS / MS data and export them into a three-dimensional data matrix in CSV format to establish a prognostic model for metabolism-related genes. S2. Single gene expression analysis in colorectal cancer cell lines, calculating the correlation coefficient between MALMPS score and the expression of each gene, performing spatial transcriptomic analysis and cell type estimation; S3, perform cell culture and transfection, real-time quantitative polymerase chain reaction, cell counting kit-8 detection and Transwell invasion assay, use apoptosis kit and cell cycle detection kit to evaluate apoptosis and cell cycle status, perform cell scratch assay and drug sensitivity estimation and verification; S4, perform cell viability assay, colony formation assay, perform proteomics and phosphoproteomics analysis, use NMF to perform consensus clustering of immune infiltration, and perform immune cell infiltration and feature estimation; S5. Construct a nomogram.

2. A MALMPS model for identifying patients with stage II / III colorectal cancer at high risk of recurrence according to claim 1, characterized in that: The MALMPS model identifies colorectal cancer patients with different clinical outcomes.

3. A MALMPS model for identifying patients with stage II / III colorectal cancer at high risk of recurrence according to claim 1, characterized in that: The COX7B gene was identified as an oncogenic factor in the tumorigenesis of CRC.

4. A MALMPS model for identifying patients with stage II / III colorectal cancer at high risk of recurrence according to claim 1, characterized in that: Decision curve analysis (DCA) was used to obtain the clinical net benefits of different models as well as full strategy and no strategy. The prognostic performance of the comprehensive model was compared using AUC and restricted mean survival (RMS) curves.

5. Application of the MALMPS model according to any one of claims 1 to 4 in analyzing gene mutation characteristics.

6. Application of the MALMPS model according to any one of claims 1 to 4 in drug sensitivity prediction.

7. Use of the MALMPS model according to any one of claims 1 to 4 in predicting the risk of recurrence of stage II / III colorectal cancer within a specific time frame.

Images