A survival prediction method for glioma epithelial-mesenchymal transition

Abstract

The invention discloses a survival prediction method related to epithelial-mesenchymal transition of glioma, which relates to the field of bioinformatics. The present invention utilizes biomarkers of epithelial-mesenchymal transition (EMT) to evaluate their effects on the prognosis of glioma patients. In addition, to ensure the accuracy and generalization ability of the model, the present invention uses multiple independent data sets for verification. These data sets include patient samples from different geographical locations, races, and treatment backgrounds, so that the effectiveness and scope of application of the model can be widely verified. The present invention comprehensively utilizes data from multiple public databases, including the Gene Comprehensive Expression Database, the Cancer Genome Atlas Database, and the Chinese Glioma Genome Atlas. This integrated analysis of multi-source data enhances the comprehensiveness of the model and the representativeness of the data, making the prediction results more accurate and reliable.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and proposes a method for constructing a survival prediction method for patients with glioma epithelial-mesenchymal transition (EMT)-related disease by combining a machine learning algorithm with transcriptome sequencing data. Background Art

[0002] According to the latest classification of central nervous system tumors released by the World Health Organization (WHO) in 2021, gliomas are divided into four grades: I, II, III and IV. Current treatments include radiotherapy, chemotherapy, surgery and targeted therapy, but due to the high recurrence rate of gliomas (up to 90%) and the low five-year survival rate (less than 10%), the treatment effect is limited, and the predictability of patient prognosis is also negatively affected. Glioma, especially glioblastoma (GBM), is the most invasive type of brain cancer. Its high invasiveness usually makes surgical resection difficult, making recurrence difficult to avoid. Although epithelial-mesenchymal transition (EMT) is a key regulator of this invasive state, its specific role in malignant gliomas remains controversial. In-depth research on gliomas and their related genes is crucial to discover more effective prediction methods and screen out prognostic indicators related to epithelial-mesenchymal transition (EMT) invasion.

[0003] Epithelial-mesenchymal transition (EMT) is a key molecular and cellular process in normal embryonic development and wound healing, and also commonly affects multiple cancer types, including glioblastoma. The activation and regulation of EMT is influenced by endogenous triggers and a complex network of signaling pathways in a specific microenvironment, involving epigenetic regulation of protein translation control factors and proteases. These processes are coordinated in cancer cells by the interaction of oncogenes and tumor suppressor genes.

[0004] Machine learning technology has played a huge role in the field of bioinformatics in recent years, especially deep learning algorithms, which have become an important tool for exploring the molecular mechanisms and potential therapeutic targets of glioma. For example, the least absolute shrinkage selection operator (LASSO) can effectively identify key disease-related genes by compressing features through L1 regularization when processing high-dimensional data. Multilayer perceptron (MLP), as a neural network, is widely used in classification and prediction tasks, such as predicting patient survival and evaluating the effectiveness of treatment. By optimizing the network structure and adjusting parameters, MLP can mine valuable rules in huge biological data to support clinical decision-making. In addition, other deep learning methods such as convolutional neural networks (CNN) and recurrent neural networks (RNN) have also shown unique advantages in glioma research.

[0005] Although current research has made progress in understanding the relationship between the prognosis of glioma patients and epithelial-mesenchymal transition (EMT), existing studies are mostly based on bulk gene transcriptome data, which is limited in sample size and method. Therefore, this study uses deep learning algorithms combined with single-cell transcriptome data to deeply explore the epithelial-mesenchymal transition (EMT) gene characteristics of different cells in the glioma microenvironment and construct a glioma prognosis model based on these characteristics, which has important research significance and application prospects. Summary of the invention

[0006] To solve the problems in the prior art, the present invention uses biomarkers of epithelial-mesenchymal transition (EMT) to evaluate their impact on the prognosis of glioma patients. In addition, to ensure the accuracy and generalization ability of the model, the present invention uses multiple independent data sets for validation. These data sets include patient samples from different geographical locations, races, and treatment backgrounds, so that the effectiveness and application scope of the model can be widely verified.

[0007] To achieve the above objectives, the present invention adopts a technical solution for a survival prediction method related to epithelial-mesenchymal transition of glioma, which comprises:

[0008] Step 1: Public statistical data collection;

[0009] The gene expression dataset of glioblastoma (GBM) patients was downloaded from The Cancer Genome Atlas (TCGA) database as a training set for the construction and training of the prediction model;

[0010] The single-cell transcriptome datasets GSE221534, GSE231860 and gene transcriptome dataset GSE16011 were downloaded from the Gene Omnibus (GEO) database. The single-cell transcriptome datasets GSE221534 and GSE231860 were used for model construction, and the gene expression dataset GSE16011 was used as a validation group for model accuracy verification.

[0011] The mRNAseq_325 and mRNAseq_693 gene transcriptome expression datasets were downloaded from the Chinese Glioma Genome Atlas (CGGA) as validation groups for model accuracy verification;

[0012] EMT-related genes with an EMT correlation score greater than 15 were collected from the GENECARDS dataset; EMT stands for epithelial-mesenchymal transition;

[0013] Step 2: Processing of single-cell transcriptome datasets and The Cancer Genome Atlas database training cohort;

[0014] The single-cell transcriptome datasets GSE221534 and GSE231860 were merged, and then the number of RNA features, RNA counts, and mitochondrial proportions of each cell were calculated. The purpose of these quality indicators is to evaluate and filter low-quality cells to ensure the accuracy and reliability of subsequent analysis; based on the genes collected in step 1, the AUCell algorithm was used to calculate the epithelial-mesenchymal transition (EMT) score of each cell, and the cells were divided into high and low epithelial-mesenchymal transition (EMT) cell groups according to the median; the Findmarkers function was used to find genes with differential expression in the high and low epithelial-mesenchymal transition (EMT) cell groups;

[0015] The gene transcription data set collected in step 1 was used again for GSVA (Gene Set Variation Analysis) analysis to obtain the epithelial-mesenchymal transition (EMT) score, which was divided into high and low epithelial-mesenchymal transition (EMT) cell groups according to the median; the R software package limma was used to perform gene expression difference analysis on the high and low epithelial-mesenchymal transition (EMT) cell groups, and genes with differential expression were identified;

[0016] Step 3: Screening of co-expressed epithelial-mesenchymal transition (EMT) genes;

[0017] The intersection of the differentially expressed genes obtained from the single-cell transcriptome dataset in step 2 and the gene transcription dataset collected in step 1 was taken to obtain the commonly expressed differentially expressed genes; the genes with a P value less than 0.05 were obtained by Cox univariate analysis (COX univariate regression for short), and based on this, the Least Absolute Shrinkage and Selection Operator (LASSO) was used to finally screen out N most relevant epithelial-mesenchymal transition (EMT) characteristic genes (ERCGs for short) from the epithelial-mesenchymal transition (EMT) related genes;

[0018] Step 4: Construct a glioma epithelial-mesenchymal transition (EMT)-related scoring system and a prognostic survival assessment model;

[0019] The N characteristic genes selected in step 3 are used to establish a scoring system, a weight is assigned to the expression of each gene, and then the weighted sum is used to obtain the score;

[0020] Step 5: Based on the prognostic risk scoring system of the glioma EMT-related characteristic genes, score each patient sample in the data set, and divide the training group and validation group samples into high and low EMT risk groups according to the median of the risk score obtained from the patient samples in the data set; perform survival analysis on the high and low EMT risk groups; construct a prognostic survival assessment model based on the glioma EMT-related system score: use the score, MGMT promoter methylation status, high and low risk model, IDH mutation, age, PRS status, 1p19q co-deletion, and tumor grade as inputs, and patient survival as output to train a prognostic survival assessment model;

[0021] Step 6: Use the trained prognosis survival assessment model to perform prognosis survival assessment on new patients and output the assessment results.

[0022] Furthermore, in step 3, N=8.

[0023] Furthermore, the characteristic genes selected in step 4 are: PLAUR, KLHL24, IFT22, LY6E, PHYHIPL, IKBIP, FAM3C and MIER1;

[0024] The score is calculated as follows: Score = (0.0000833 × L PLAUR )+(-0.0000855×L KLHL24 )+(0.0000699×L IFT22 )+(0.0000199×L LY6E )+(-0.00000739×L PHYHIPL )+(0.000119399×L IKBIP )+

[0025] (0.0000967×L FAM3C )+(-0.000352056×L MIER1 ), where each characteristic gene name corresponds to its gene expression level in the sample.

[0026] The advantages of the present invention are:

[0027] Comprehensive analysis based on multi-source data: The present invention comprehensively utilizes data from multiple public databases, including the Gene Expression Omnibus (GEO), The Cancer Genome Atlas (TCGA), and the Chinese Glioma Genome Atlas (CGGA). This integrated analysis of multi-source data enhances the comprehensiveness of the model and the representativeness of the data, making the prediction results more accurate and reliable.

[0028] Efficient feature selection algorithm: By using the least absolute shrinkage selection operator (LASSO) regression algorithm, the present invention can effectively screen key feature genes related to glioma epithelial-mesenchymal transition (EMT). The application of the least absolute shrinkage selection operator (LASSO) algorithm helps to reduce the complexity of the model and prevent overfitting, thereby improving the generalization ability of the model.

[0029] Comprehensive evaluation of multiple machine learning algorithms: This paper not only uses the least absolute shrinkage selection operator (LASSO) algorithm to build a prognostic model, but also uses other machine learning algorithms to evaluate the accuracy and generalization ability of the model. This diversity of methods enhances the reliability of the model and ensures that good predictive performance can be maintained in different clinical settings.

[0030] Validation of actual clinical data: By validating the survival data of actual patients, the present invention proves that the glioma epithelial-mesenchymal transition (EMT)-related scoring system can effectively distinguish patient groups with different risk levels, demonstrating the effectiveness of the model in actual clinical applications. This provides clinicians with a powerful tool to more accurately predict patient survival and treatment response. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a schematic diagram of single cell quality control and preprocessing analysis of GSE221534 and GSE231860 in the examples of the present invention. The tSNE diagram shows the AUCell score of each cell and is annotated with gradient color difference.

[0032] Figure 2 This is the difference analysis of the TCGA validation cohort in the embodiment of the present invention. The volcano plot shows the distribution of differentially expressed genes (|logFC|>1, Pvalue<0.05)

[0033] Figure 3 It is a schematic diagram of the least absolute shrinkage selection operator (LASSO) analysis of genes after univariate COX univariate regression in an embodiment of the present invention, including the least absolute shrinkage selection operator (LASSO) coefficient curve of genes that have been determined to be related to epithelial-mesenchymal transition (EMT) and a schematic diagram of the optimal lambda value in the least absolute shrinkage selection operator (LASSO) model, which helps to identify important predictive genes.

[0034] Figure 4It is a schematic diagram of the prognostic effect in the embodiment of the present invention. The clinical data of the validation group CGGA_325 were analyzed, and the 1-year, 3-year, and 5-year survival rates of the patients were further predicted, including a nomogram constructed using MGMT promoter methylation status, model high and low risk, IDH mutation, age, PRS status, 1p19q co-deletion, and tumor grade, and a calibration diagram based on the nomogram to evaluate the accuracy of the prediction results and the observed results.

[0035] Figure 5 It is a schematic diagram of the prognostic effect in the embodiment of the present invention. The clinical data of the validation group CGGA_693 were analyzed, and the 1-year, 3-year, and 5-year survival rates of the patients were further predicted, including a nomogram constructed using MGMT promoter methylation status, model high and low risk, IDH mutation, age, PRS status, 1p19q co-deletion, and tumor grade, and a calibration diagram based on the nomogram to evaluate the accuracy of the prediction results and the observed results.

[0036] Figure 6 Schematic diagram of survival analysis of the prognostic model related to epithelial-mesenchymal transition (EMT) in glioma in an embodiment of the present invention, analyzing the survival curve from the TCGA training group and the 1-year, 3-year, and 5-year survival predicted by ROC.

[0037] Figure 7 Schematic diagram of survival analysis of the prognostic model related to epithelial-mesenchymal transition (EMT) in glioma in an embodiment of the present invention, analyzing the survival curve from the GSE16011 validation group and the 1-year, 3-year, and 5-year survival predicted by ROC.

[0038] Figure 8 This is a schematic diagram of the validation of the prognostic model by machine learning in an embodiment of the present invention. The heat map shows the top 50 algorithms and parameters with average AUC values ​​obtained from 207 combination methods of 15 machine learning algorithms.

[0039] Fig. 9 This is a schematic diagram of a binary classification machine learning algorithm based on a neural network architecture of a multilayer perceptron (MLP) in an embodiment of the present invention. The confusion matrix is ​​used to evaluate the performance of a machine learning classification model. It shows the results of a binary classification task, in which the model predicts two categories, negative and positive. The ROC curve is a graphical tool for displaying the performance of a classification model. DETAILED DESCRIPTION

[0040] Step 1: Public statistical data collection and statistics:

[0041] In the statistical analysis section, all analyses and graphical visualizations were performed using R software and packages from the Comprehensive R Archive Network (CRAN) or Bioconductor software library. In the analysis, a P value less than 0.05 was used as the threshold for judging the significance of statistical correlation.

[0042] Genes related to epithelial-mesenchymal transition (EMT) were collected from the Gene Cards database (https: / / www.genecards.org), and 142 genes with a correlation score greater than 15 were selected as the self-defined epithelial-mesenchymal transition (EMT)-related genes of the present invention.

[0043] Gene transcriptome data and survival data of glioblastoma patients from the Cancer Genome Atlas Project (https: / / portal.gdc.cancer.gov / ) were collected as training groups for the construction and training of prediction models.

[0044] Gene expression data GSE16011 and related glioma patients from the Gene Expression Omnibus (https: / / www.ncbi.nlm.nih.gov / geo / ) database were collected as a validation group for model accuracy verification.

[0045] Single-cell transcriptome data GSE221534 and GSE231860 of glioma patients were collected from the Gene Expression Omnibus (https: / / www.ncbi.nlm.nih.gov / geo / ) database for the construction of prediction models.

[0046] The mRNA gene transcriptome sequencing data mRNAseq_693 and mRNAseq_325 and survival data from the Chinese Glioma Genome Atlas (cgga.org.cn) were collected as validation groups for model accuracy verification.

[0047] Step 2: Preprocessing of public statistical data:

[0048] All mRNA gene transcriptome sequencing data from GEO, TCGA and CGGA were de-batched and scaled to avoid the impact of batch effects on statistical results; the single-cell transcriptome datasets GSE221534 and GSE231860 from GEO were merged and preprocessed to evaluate quality indicators such as the number of RNA features, RNA counts and mitochondrial proportions of each cell; data normalization was performed to ensure that the expression levels between different cells were comparable; highly variable genes were identified, and the top 3000 variable features were selected using the vst method for subsequent analysis; data standardization was performed to eliminate scale differences between genes; PCA dimensionality reduction was run, and the ElbowPlot diagram was used to determine the number of principal components. This diagram is used to display the variance contribution rate of each principal component in the principal component analysis, helping researchers determine the number of dimensions that should be retained during data dimensionality reduction. By observing the "sudden drop" points of variance contribution, we can select an appropriate number of principal components to capture the main variation information in the data while removing noise and unnecessary details; use the JackStraw method to evaluate the statistical significance of the principal components; use the Harmony algorithm to eliminate batch effects; use FindNeighbors and FindClusters to cluster cells; visualize the dimensionality reduction results, including cell clustering diagrams and tSNE diagrams; and use SingleR to annotate cells.

[0049] Step 3: Construction of prognostic risk model:

[0050] Based on the epithelial-mesenchymal transition (EMT)-related gene set in step 1, the epithelial-mesenchymal transition (EMT)-related gene set in step 1 was first used to calculate the epithelial-mesenchymal transition (EMT) score of each cell using the AUCell algorithm, and then Figure 1 The gradient color difference annotation is shown according to the score. In order to further explore the role of epithelial-mesenchymal transition (EMT) in cells, we divided the single-cell samples into high and low groups of epithelial-mesenchymal transition (EMT) cells according to the median. Then, we used the Findmarkers function to find differentially expressed genes between the high and low groups in order to discover the key role of epithelial-mesenchymal transition (EMT) in cells.

[0051] Based on the EMT-related gene set in step 1, GSVA analysis was performed on the gene expression data of the TCGA cohort. Similarly, the samples were divided into high and low groups of EMT cells according to the median, and then the limma package was used to identify differentially expressed genes between the high and low groups and the results were analyzed. Figure 2 Draw a volcano plot. This step helps to gain a deeper understanding of the role of epithelial-mesenchymal transition (EMT) in tumor initiation and progression.

[0052] Based on the above foundation, FunRich software was used to perform cross-Venn intersection processing on the screened differentially expressed genes, and univariate COX regression analysis was performed. Through this analysis, we screened out 16 genes with a P value less than 0.05, which may be of great significance in the epithelial-mesenchymal transition (EMT) process of glioma. Figure 3 In order to further explore key genes, the least absolute shrinkage selection operator (LASSO) algorithm was used to reduce the dimensionality of differentially expressed genes. This analysis step includes the application of variable selection and shrinkage methods to effectively streamline and improve the genome. After processing, 8 glioma epithelial-mesenchymal transition (EMT)-related genes were obtained. These genes are expected to become important indicators for future research on glioma diagnosis, treatment, and prognosis assessment.

[0053] According to the weight of each gene obtained by least absolute shrinkage selection operator (LASSO) regression, the expression levels of selected glioma epithelial-mesenchymal transition (EMT)-related genes and their linear regression coefficients were integrated into a comprehensive risk score, which was formalized as ∑ (regression coefficient × gene expression). Then, glioma patients were divided into two groups with high and low risk of epithelial-mesenchymal transition (EMT) according to the median value of the score.

[0054] Evaluation of the clinical significance of the model:

[0055] Univariate and multivariate analyses identified independent prognostic factors, and a predictive nomogram was constructed. The model integrated MGMT promoter methylation status, high and low risk model, IDH mutation, age, PRS status, 1p19q co-deletion, and tumor grade, and was developed using the survival and rms packages in R. Figure 4 and Figure 5 As shown in the figure, the performance of the model in predicting the one-year, three-year and five-year survival of glioma patients in the validation group CGGA_325 dataset and CGGA_693 dataset was demonstrated through nomogram, calibration and evaluation. At the same time, the prediction effect was compared with other factors to verify the clinical practicality of the nomogram.

[0056] like Figure 6 and Figure 7As shown in the figure, the survival package and survminer package were used to perform detailed survival analysis on the training group of TCGA and the validation groups of GEO and CGGA, respectively. The survival curve clearly distinguished the high-risk and low-risk groups, among which the survival probability of patients in the high-risk group was significantly lower than that in the low-risk group, which was statistically significant in all data sets with P value less than or equal to 0.00011. Furthermore, the time-dependent ROC curve revealed the prediction accuracy of the model at 1 year, 3 years and 5 years, and the AUC value fluctuated over time, reflecting the prediction ability of the model at different time points. Specifically, on the TCGA dataset, the 3-year and 5-year prediction AUC values ​​of the model were 0.63 and 0.67, respectively, indicating good medium- and long-term prediction accuracy. The validation group of the GSE16011 dataset also showed a similar trend, demonstrating the generalization ability of the model.

[0057] Step 4: Further verify the model accuracy based on machine learning:

[0058] In the present invention, a variety of machine learning algorithms are used to verify and improve the accuracy of the survival prediction model. The R programming language is used for implementation. The specific technical details are as follows:

[0059] like Figure 8 , the present invention selected a variety of machine learning algorithms as benchmarks, including k-nearest neighbor (KNN), least absolute shrinkage selection operator (LASSO) regression and elastic network regression (ENR). For the KNN algorithm, the knn function in the class package was used, and the optimal k value was selected through cross-validation (CV). For the least absolute shrinkage selection operator (LASSO) and ENR, the glmnet package was used respectively, and the optimal regularization parameter value was also determined through cross-validation.

[0060] like Fig. 9 Among various machine learning algorithms, it was found that the multilayer perceptron (MLP) neural network performed well as the main classification tool. This is a feedforward neural network suitable for complex nonlinear models. Furthermore, the neuralnet package was used to train the MLP model, and the model was finely optimized by adjusting the cutoff value, the number of hidden layer neurons, weight decay, and dropout ratio. Then, the pROC package was used to calculate the area under the receiver operating characteristic (ROC) curve (AUC), which is a key indicator for evaluating the performance of binary classification models. Fig. 9 As shown in Figure 2, the ROC curve shows the performance of the model at different specificity and sensitivity levels, with an AUC value as high as 0.976, reflecting the high accuracy of the model.

[0061] In the model training phase, the TCGA dataset was selected as the training set, and the GEO and CGGA datasets were used as the validation sets. This data segmentation strategy helps to test the generalization ability of the model on independent datasets. Through iterative model training, it was observed that at the 25th epoch, the model achieved an accuracy of 98% and a loss value of 0.04, showing the optimization and learning ability of the model during the training process.

[0062] In addition, the classification effect of the model was further verified through the confusion matrix, and evaluation indicators such as sensitivity, specificity, precision, recall rate and F1 score were calculated from it. The confusion matrix results show that on the actual test set, the model has a strong ability to classify positive and negative samples, with an accuracy rate and Kappa value of 0.875 and 0.749 respectively, indicating that the model has good consistency and reliability.

[0063] Step 5: Multiple machine learning algorithms were used to test the effectiveness and feasibility of the prognostic risk assessment system; machine learning algorithms such as neural network, logistic regression, linear discriminant analysis, quadratic discriminant analysis, KNN nearest neighbor, decision tree, random forest, XGBoost, ridge regression, least absolute shrinkage selection operator (LASSO) regression, elastic network regression, support vector machine, gradient boosting machine, stepwise logistic regression, naive Bayes, etc. were used, and combined through caret parameter adjustment, custom parameter combination, least absolute shrinkage selection operator (LASSO) feature screening, and cross-validation. The TCGA training set and multiple validation sets such as CGGA_325, CGGA_693 and GSE16011 were used for model training and validation, and the AUC value of each algorithm was calculated to evaluate its performance. By averaging and sorting the AUC values ​​of each algorithm in the training and validation cohorts, we can more effectively determine the practicality and accuracy of each algorithm; we sort the training and validation cohorts by the average AUC values, focus on the MLP neural network algorithm, and on this basis screen out the optimal algorithm parameter configuration, including 25 training cycles (epochselect), 20 samples per batch (batch_size), and a learning rate (learningrate) of 0.01, in order to achieve the best model fitting and prediction effect.

Claims

1. A method for predicting survival related to epithelial-mesenchymal transition in glioma, the method comprising: Step 1: Public statistical data collection; The gene expression dataset of glioblastoma patients was downloaded from the Cancer Genome Atlas database as a training set for the construction and training of the prediction model; Download single-cell transcriptome datasets GSE221534, GSE231860 and gene transcriptome dataset GSE16011 from the comprehensive gene expression database. The single-cell transcriptome datasets GSE221534 and GSE231860 are used for model construction, and the gene expression dataset GSE16011 is used as a validation group for model accuracy verification; Download the mRNAseq_325 and mRNAseq_693 gene transcriptome expression datasets from the Chinese Glioma Genome Atlas as validation groups for model accuracy verification; Epithelial-mesenchymal transition-related genes with EMT correlation scores greater than 15 were collected from the GENECARDS dataset; Step 2: Processing of single-cell transcriptome datasets and The Cancer Genome Atlas database training cohort; The single-cell transcriptome datasets GSE221534 and GSE231860 were merged, and then the number of RNA features, RNA counts, and mitochondrial proportions of each cell were calculated. The purpose of these quality indicators is to evaluate and filter low-quality cells to ensure the accuracy and reliability of subsequent analysis; based on the genes collected in step 1, the AUCell algorithm was used to calculate the epithelial-mesenchymal transition score of each cell, and the cells were divided into high and low epithelial-mesenchymal transition cell groups according to the median; the Findmarkers function was used to find genes with differential expression in the high and low epithelial-mesenchymal transition cell groups; The gene transcription data set collected in step 1 was used again for GSVA analysis to obtain the epithelial-mesenchymal transition score, which was divided into high and low epithelial-mesenchymal transition cell groups according to the median. The R software package limma was used to perform gene expression difference analysis on the high and low epithelial-mesenchymal transition cell groups, and genes with differential expression were identified. GSVA stands for Gene Set Variation Analysis, which means genome variation analysis. Step 3: Screening of co-expressed epithelial-mesenchymal transition genes; The intersection of the differentially expressed genes obtained from the single-cell transcriptome dataset in step 2 and the gene transcription dataset collected in step 1 was taken to obtain the commonly expressed differentially expressed genes; the genes with a P value less than 0.05 were obtained through Cox univariate analysis, and based on this, the minimum absolute shrinkage selection operator was used to finally screen out 8 characteristic genes most relevant to epithelial-mesenchymal transition from the epithelial-mesenchymal transition-related genes; Cox univariate analysis means Cox univariate analysis, referred to as COX univariate regression; Pvalue means P value; The selected characteristic genes were: PLAUR, KLHL24, IFT22, LY6E, PHYHIPL, IKBIP, FAM3C and MIER1; Step 4: Construct a scoring system related to epithelial-mesenchymal transition in glioma and a prognostic survival assessment model; The 8 characteristic genes selected in step 3 were used to establish a scoring system, a weight was assigned to the expression of each gene, and then the weighted sum was used to obtain the score; The score is calculated as follows: Score = (0.0000833 × L PLAUR )+(-0.0000855×L KLHL24 )+(0.0000699×L IFT22 )+(0.0000199×L LY6E )+(-0.00000739×L PHYHIPL )+(0.000119399×L IKBIP )+ (0.0000967×L FAM3C )+(-0.000352056×L MIER1 ), where each characteristic gene name corresponds to its gene expression level in the sample; Step 5: Score each patient sample in the data set based on the prognostic risk scoring system of the glioma EMT-related characteristic genes, and divide the training group and validation group samples into high and low EMT risk groups according to the median of the risk score obtained from the patient samples in the data set; Survival analysis was performed on high and low EMT risk groups; a prognostic survival assessment model was constructed based on the glioma EMT-related system score: the score, MGMT promoter methylation status, high and low risk model, IDH mutation, age, PRS status, 1p19q co-deletion, and tumor grade were used as inputs, and patient survival was used as output to train a prognostic survival assessment model; Step 6: Use the trained prognosis survival assessment model to perform prognosis survival assessment on new patients and output the assessment results.

Images