Cancer subtype classification method and system based on driver gene identification

Abstract

The present invention discloses a method and system for classifying cancer subtypes based on driver gene identification, the method comprising the following steps: obtaining and preprocessing a cancer data set; based on the preprocessed cancer data set, identifying driver genes to obtain global driver genes; based on the global driver genes and a one-to-one multi-classification support vector machine algorithm, constructing a DriverSub-SVM cancer subtype classification model; the DriverSub-SVM cancer subtype classification model is used to classify cancer subtypes. The present invention contributes to the development of precision medicine by constructing an interpretable cancer subtype classification model, and lays the foundation for future cancer subtype classification research.

Description

Technical Field

[0001] The present invention belongs to the technical field of cancer subtype classification, and specifically relates to a method and system for cancer subtype classification based on driver gene identification. Background Art

[0002] Cancer is a complex disease characterized by morphological diversity of tumor cells and abnormalities in the genome, epigenome, transcriptome, and proteome. The high heterogeneity of tumors between different patients makes it difficult for a single treatment method to be applicable to all patients, which poses a major challenge to personalized cancer medicine. Tumor molecular typing technology divides patients into different subtypes by analyzing internal molecular features such as mutant gene data, expression data, and protein interaction networks, providing a solution for precision medicine. Studies have shown that patients of the same subtype show similarities in pathological characteristics and treatment effects. Dividing patients into different subtypes can achieve personalized therapeutic interventions, thereby further improving the treatment effect, diagnostic accuracy, and predictive ability of cancer. Although existing technologies have made significant progress in the classification of cancer subtypes, there is still room for further improvement in accuracy. Summary of the invention

[0003] In view of the shortcomings of the prior art, the present invention provides a method and system for cancer subtype classification based on driver gene identification, which improves classification accuracy by discovering patient-specific driver genes and global driver genes.

[0004] To achieve the above object, the present invention provides the following solutions:

[0005] The method for classifying cancer subtypes based on driver gene identification includes the following steps:

[0006] Acquire and preprocess cancer datasets;

[0007] Based on the preprocessed cancer data set, driver genes are identified to obtain global driver genes;

[0008] Based on the global driver genes and a one-to-one multi-classification support vector machine algorithm, a DriverSub-SVM cancer subtype classification model is constructed; the DriverSub-SVM cancer subtype classification model is used to divide cancer subtypes.

[0009] Preferably, the cancer types in the cancer data set include BRCA breast cancer, COAD colon cancer and THCA thyroid cancer; wherein each cancer type includes four types of omics data: transcriptome expression data, somatic mutation data, copy number variation data and protein function interaction network data;

[0010] The method for preprocessing the cancer dataset includes:

[0011] Based on Gauss's theorem, the variance-stabilized transformed (VST) expression values ​​of the transcriptome expression data are normalized;

[0012] Based on a preset threshold, the normalized transcriptome expression data is subjected to gene expression abnormality discrimination;

[0013] Assign values ​​to the discrimination results to obtain a binary matrix;

[0014] Based on the binary matrix, abnormally expressed gene data of the patient is obtained to complete preprocessing of the cancer data set;

[0015] The pre-processed cancer dataset includes abnormally expressed gene data, somatic mutation gene data, and protein functional interaction network for each cancer type.

[0016] Preferably, the method for obtaining the global driver gene is:

[0017] Constructing a bipartite graph based on the abnormally expressed gene data and the somatic mutation gene data;

[0018] Based on the bipartite graph, obtaining the association relationship between abnormally expressed genes and somatic mutation genes in the protein functional interaction network;

[0019] A three-round random walk algorithm is used to score abnormally expressed genes and somatic mutation genes based on the association relationship, the abnormally expressed gene data, the somatic mutation gene data and the damping factor to obtain a random walk score matrix of mutant genes; wherein the damping factor represents the degree to which a gene depends on a protein functional interaction network or a gene's own characteristics;

[0020] Based on the random walk score matrix, obtaining a personalized driver gene list for the patient;

[0021] Aggregating the personalized driver gene lists of each patient to obtain a set of personalized driver genes;

[0022] The representative features of the personalized driver genes are identified based on an L1 regularized logistic regression model, and a global driver gene set is constructed based on the representative features.

[0023] Preferably, the method for constructing the L1 regularized logistic regression model is:

[0024] Based on the cancer dataset and personalized driver gene signatures, construct a multi-class sample set;

[0025] Based on the personalized driver gene signature, calculating the conditional probability distribution of the preset categories of cancer subtypes in the multi-class sample set;

[0026] Based on the conditional probability distribution, the objective function of the logistic regression with L1 norm penalty is obtained, and the construction of the L1 regularized logistic regression model is completed.

[0027] The present invention also provides a cancer subtype classification system based on driver gene identification, which is used to implement the cancer subtype classification method, comprising:

[0028] Data preprocessing module, used to acquire and preprocess cancer datasets;

[0029] A gene identification module, used to identify driver genes based on the preprocessed cancer data set to obtain global driver genes;

[0030] A model building module is used to build a DriverSub-SVM cancer subtype classification model based on the global driver gene and a one-to-one multi-classification support vector machine algorithm; the DriverSub-SVM cancer subtype classification model is used to divide cancer subtypes.

[0031] Preferably, in the data preprocessing module, the cancer types in the cancer data set include BRCA breast cancer, COAD colon cancer and THCA thyroid cancer; wherein each cancer type includes four omics data: transcriptome expression data, somatic mutation data, copy number variation data and protein function interaction network data; the data preprocessing module includes:

[0032] A standardization unit, used for standardizing the variance-stabilized transformed (VST) expression value of the transcriptome expression data based on Gauss's theorem;

[0033] An abnormality determination unit, used to determine abnormalities in gene expression based on a preset threshold value for the normalized transcriptome expression data;

[0034] An assignment unit, used for assigning values ​​to the discrimination results to obtain a binary matrix;

[0035] The abnormal expression gene data acquisition unit is used to obtain the abnormal expression gene data of the patient based on the binary matrix to complete the preprocessing of the cancer data set; wherein the preprocessed cancer data set includes the abnormal expression gene data, somatic mutation gene data and protein function interaction network of each cancer type.

[0036] Preferably, the gene recognition module comprises:

[0037] A bipartite graph construction unit, configured to construct a bipartite graph based on the abnormally expressed gene data and the somatic mutation gene data;

[0038] A relationship acquisition unit, used for obtaining the association relationship between abnormally expressed genes and somatic mutation genes in the protein function interaction network based on the bipartite graph;

[0039] A random scoring unit, for scoring abnormally expressed genes and somatic mutation genes by using a three-round random walk algorithm based on the association relationship, the abnormally expressed gene data, the somatic mutation gene data and a damping factor, to obtain a random walk score matrix of the mutant gene; wherein the damping factor represents the degree to which the gene depends on the protein function interaction network or the gene's own characteristics;

[0040] A gene list acquisition unit, used to obtain a personalized driver gene list for the patient based on the random walk score matrix;

[0041] A gene aggregation unit, used to aggregate the personalized driver gene list of each patient to obtain a set of personalized driver genes;

[0042] The global driver gene acquisition unit is used to identify the representative features of the personalized driver genes based on the L1 regularized logistic regression model, and construct the set of global driver genes based on the representative features.

[0043] Preferably, the construction process of the L1 regularized logistic regression model is:

[0044] Based on the cancer dataset and personalized driver gene signatures, construct a multi-class sample set;

[0045] Based on the personalized driver gene signature, calculating the conditional probability distribution of the preset categories of cancer subtypes in the multi-class sample set;

[0046] Based on the conditional probability distribution, the objective function of the logistic regression with L1 norm penalty is obtained, and the construction of the L1 regularized logistic regression model is completed.

[0047] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention successfully proposes a DriverSub-SVM model, which innovatively combines personalized driver genes and global driver genes with the OAO-MSVM algorithm for classifying cancer subtypes. The model of the present invention not only shows excellent accuracy and precision in the subtype division of multiple cancer data sets, but also identifies biomarkers that are of great significance for clinical diagnosis and targeted therapy. The DriverSub-SVM method includes multi-step screening and subtype division. After rigorous verification by various experiments, it was proved that the method can effectively distinguish cancer subtypes. The present invention has contributed to the development of precision medicine by constructing an interpretable cancer subtype classification model, and has laid the foundation for future cancer subtype division research. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0049] Figure 1 Schematic diagram of the DriverSub-SVM model according to an embodiment of the present invention;

[0050] Figure 2 This is a schematic diagram of the OAO-MSVM classification algorithm according to an embodiment of the present invention;

[0051] Figure 3 The confusion matrix diagram of the breast cancer data set of the embodiment of the present invention;

[0052] Figure 4 This is a schematic diagram of the effect of the number of personalized driver genes on the accuracy of subtype classification according to an embodiment of the present invention;

[0053] Figure 5 The KEGG pathway analysis diagrams of the predicted global driver genes according to the embodiments of the present invention are as follows; (a) is the KEGG enrichment bubble diagram of BRCA; (b) is the KEGG enrichment bubble diagram of COAD; (c) is the KEGG enrichment bubble diagram of THCA;

[0054] Figure 6 The results of the comparative experiment on three cancer data sets in the embodiment of the present invention are shown in Figure 1. (a) is the accuracy comparison diagram; (b) is the precision comparison diagram; (c) is the recall comparison diagram; (d) is the F1 score comparison diagram;

[0055] Figure 7 SHAP graphs are contributed by the gene features of the five subtypes of breast cancer classified in the embodiment of the present invention; (a) SHAP graphs are contributed by the gene features of subtype 1: Normal subtype; (b) SHAP graphs are contributed by the gene features of subtype 2: LumA subtype; (c) SHAP graphs are contributed by the gene features of subtype 3: Her2 subtype; (d) SHAP graphs are contributed by the gene features of subtype 4: LumB subtype; (e) SHAP graphs are contributed by the gene features of subtype 5: Basal subtype;

[0056] Figure 8 SHAP graphs of gene feature contributions in the classification of three subtypes of colorectal cancer in the embodiment of the present invention; (a) SHAP graph of gene feature contribution for subtype 1: CIN subtype; (b) SHAP graph of gene feature contribution for subtype 2: Invasive subtype; (c) SHAP graph of gene feature contribution for subtype 3: MSI / CIMP subtype;

[0057] Fig. 9 The SHAP diagrams showing the contribution of gene features in the four subtype classifications of thyroid cancer in the embodiments of the present invention; (a) SHAP diagram for subtype 1: Classical / usual subtype gene feature contribution; (b) SHAP diagram for subtype 2: Follicular gene feature contribution; (c) SHAP diagram for subtype 3: Tal1 cel subtype gene feature contribution; (d) SHAP diagram for subtype 4: Other specify subtype gene feature contribution. DETAILED DESCRIPTION

[0058] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0060] Embodiment 1

[0061] like Figure 1 The framework of DriverSub-SVM is shown in the figure: (a) Multi-omics data are preprocessed and a bipartite graph including mutant genes and abnormally expressed genes is constructed. (b) The influence of dysregulated genes in the interaction network is evaluated by random walk algorithm analysis, and personalized driver genes are selected based on the classification accuracy of each specific cancer type. (c) The personalized driver gene sets from N samples are aggregated to form an overall driver gene set, and then the LR-L1 algorithm is used to screen out genes that are representative and valuable in multiple samples to determine a set of global driver genes. (d) Using the global driver genes as input, the OAO-MSVM algorithm is used to classify cancer subtypes. In addition, the SHAP interpreter is used to identify biomarkers for each subtype.

[0062] Specifically, the method for classifying cancer subtypes based on driver gene identification includes the following steps:

[0063] S1: Acquire and preprocess three different cancer datasets from TCGA;

[0064] A further embodiment is that the cancer types in the cancer data set include BRCA breast cancer, COAD colon cancer, and THCA thyroid cancer; wherein each cancer type includes four types of omics data: transcriptome expression data, somatic mutation data, copy number variation data, and protein function interaction network data;

[0065] Somatic mutation and transcriptome expression data were obtained using the R package "TCGAbiolinks". Transcriptome gene expression data were obtained through the "STAR-Counts" process, and then the vst function in the DESeq2 software package was used to perform variance stabilizing transformation (VST) to achieve data standardization. Gene expression data are numerical values ​​representing the expression levels of various genes in patients. Somatic mutation data are obtained through the "Masked SomaticMutation" data type. In the somatic mutation matrix, the value 1 indicates the presence of a gene mutation, while 0 indicates the absence. Different from the previous method for obtaining a data set, copy number variation data (CNV) is derived from the UCSC Xena platform. In the CNV data set, the expression value is encoded as -1 for "loss of function" genes, +1 for "enhanced function" genes, and 0 for "neutral" expression genes. This embodiment integrates somatic mutation data and CNV data, and if the data value in the somatic mutation and CNV data sets is 1 or -1, the gene is considered a mutant gene. Then a patient mutation matrix is ​​generated, where 1 indicates that the gene in a specific sample has mutated and 0 indicates that no mutation has occurred. This embodiment integrates protein function interaction network data from SignaLink3 and Reactome databases. The SignaLink3 database covers a series of manually sorted signal pathway interactions and related data (such as regulation, positioning, and disease). The Reactome database provides a comprehensive human protein interaction network that integrates pathways, gene co-expression values, protein structure interaction domains, and other source data obtained through expert screening. The method of the present invention is to cross-fuse these two data sets to identify gene pairs that interact with each other in both, thereby generating an interaction network that integrates multiple biological functions, including 15,852 edges and 14,644 nodes. In addition, the present invention obtains subtype information for each patient from the clinical data of TCGA. For breast cancer, the subtype information comes from PAM50 Subtype, mainly including Normal, LumA, LumB, Her2, and Basal subtypes. The subtype information of colon cancer is characterized by expression type and is mainly divided into three subtypes: CIN type, invasive type, and MSI / CIMP type. The subtype information of thyroid cancer uses histological type to classify cancer patients into four subtypes: Classical / usual type, Follicular type, Tallcell type, and Other specify type.

[0066] Methods for preprocessing cancer datasets include:

[0067] Based on Gauss's theorem, the variance-stabilizing transformed VST expression values ​​of the transcriptome expression data were normalized;

[0068] Based on the preset threshold, the normalized transcriptome expression data is used to identify abnormalities in gene expression;

[0069] Assign values ​​to the discrimination results to obtain a binary matrix;

[0070] Based on the binary matrix, the abnormally expressed gene data of the patient is obtained to complete the preprocessing of the cancer data set;

[0071] In this embodiment, for transcriptome expression data, the present invention uses Gauss's theorem to standardize the VST expression value to determine the abnormality of gene expression. If the absolute value of the z value of a particular patient's gene is greater than or equal to 2δ, the gene is considered to be abnormally expressed in the patient. According to equation (1), if a gene meets this threshold, it is assigned a value of 1, otherwise it is 0. This process generates a binary matrix, which is called patient abnormal expression data.

[0072]

[0073] The present invention simultaneously retains sample data and gene data having three data types, that is, the preprocessed cancer data set includes abnormally expressed gene data, somatic mutation gene data, and protein function interaction network of each cancer type.

[0074] In this embodiment, according to the above data preprocessing steps, the present invention obtained 1089 BRCA samples, 134 COAD samples and 463 THCA samples. The detailed information of the generated data set is shown in Table 1:

[0075] Table 1

[0076] Dataset Sample size Number of genes BRCA 1089 14161 COAD 134 14188 THCA 463 8475

[0077] S2: Based on the preprocessed cancer dataset, driver genes are identified to obtain global driver genes;

[0078] A further embodiment is that the method for obtaining the global driver gene is:

[0079] Based on abnormal expression gene data and somatic mutation gene data, a bipartite graph is constructed;

[0080] Based on the bipartite graph, the association between abnormally expressed genes and somatic mutation genes in the protein functional interaction network is obtained;

[0081] A three-round random walk algorithm was used to score abnormally expressed genes and somatic mutation genes based on association relationships, abnormally expressed gene data, somatic mutation gene data, and damping factors to obtain a random walk score matrix of mutant genes; the damping factor represents the degree to which a gene depends on the protein functional interaction network or the gene's own characteristics;

[0082] The present invention uses a three-round random walk algorithm to evaluate the bidirectional impact between mutated driver genes and dysregulated downstream genes in the biological function network, taking into account both the impact of driver gene mutations on downstream genes and the reverse impact of dysregulated downstream gene expression on mutated driver genes. The present invention is based on a bipartite graph constructed from previously identified mutant genes and abnormally expressed genes to represent the association between genes in the biological function network, where P ij =1 indicates the association between mutant gene i and abnormally expressed gene j. m (i) represents the score of the mutant gene i after the first or third random walk, S a (j) represents the score of abnormally expressed gene j after the second random walk. M(i) represents the mutation data of gene i, and A(j) represents the data of abnormally expressed gene j (where 1 represents abnormality and 0 represents normality). Subsequently, for each mutant gene in the mutation matrix and each expressed gene in the abnormal expression matrix, their scores can be calculated using formulas (2) to (4). The damping factor a here represents the degree of dependence on the functional interaction network structure or its own characteristics.

[0083]

[0084]

[0085]

[0086] Based on the random walk score matrix, a personalized driver gene list is obtained for the patient;

[0087] Aggregate the personalized driver gene lists of each patient to obtain a set of personalized driver genes;

[0088] In this embodiment, after three rounds of random walks, all mutant genes of each patient will be scored accordingly, thereby generating a random walk score matrix. Then, the scores of the driver genes are arranged in descending order to obtain a personalized driver gene list for the patient. The higher the ranking, the greater the possibility that the gene is a driver gene in a specific sample. Finally, the present invention selects the top 80 genes as the driver gene list for each patient based on the overall performance after DriverSub-SVM classification (see Example 2 for details).

[0089] The L1-regularized logistic regression model was used to identify representative features of personalized driver genes, and a global driver gene set was constructed based on the representative features.

[0090] A further implementation method is that the method for constructing the L1 regularized logistic regression model is:

[0091] Construct a multi-class sample collection based on cancer datasets and personalized driver gene signatures;

[0092] Calculate the conditional probability distribution of pre-set categories of cancer subtypes in a multi-class sample set based on personalized driver gene signatures;

[0093] Based on the conditional probability distribution, the objective function of the logistic regression with L1 norm penalty is obtained, and the construction of the L1 regularized logistic regression model is completed.

[0094] In this embodiment, in order to obtain a comprehensive set of driver genes that can reflect both the unique characteristics of individual patients and the common characteristics of the entire patient population, the personalized driver gene lists of each patient are brought together. Subsequently, in order to identify the most representative features, the present invention uses a logistic regression (LR-L1) model with L1 regularization. These selected features are used to form a global driver gene set, thereby including both the specific characteristics of individual patients and the global characteristics of the patient population. As an embedded feature selection technique, the LR-L1 model aims to improve the interpretability of the classifier while reducing the feature dimension. In the study of this embodiment, L1 regularization promotes sparse solutions, which helps to identify the most important features from the patient gene pool.

[0095] Consider a multi-class sample set T = {(x1,y1),(x2,y2),…,(x N ,y N )},here Where N represents the number of cancer samples, M represents the number of personalized driver gene signatures of patients, and K represents the total number of cancer subtype categories in the samples. Given the input feature x, the conditional probability distribution of the kth class can be expressed as:

[0096]

[0097] The objective function of logistic regression with L1 norm penalty is as follows:

[0098]

[0099] Here, y i,k Indicates that sample i belongs to the true label of category k, where a value of 1 indicates that it belongs to the category and a value of 0 indicates that it does not belong to the category. represents the probability that sample i belongs to class k. Given the model parameter matrix w, w j,k represents the weight corresponding to the j-th feature in class k, and λ is the regularization strength. In the experiments of the present invention, the number of selected global driver genes varies for different cancer types, depending on factors such as sample size, gene characteristics, and dataset complexity. For example, after calculation using the LR-L1 model, the BRCA dataset has 1489 global driver genes, COAD has 220, and THCA has 711. To further explore the biological relevance of the identified global driver gene set, the present invention used the KEGG pathway for enrichment analysis (see Example 2 for details).

[0100] S3: Based on the global driver genes and the one-against-one multi-class support vector machine algorithm, construct a DriverSub-SVM cancer subtype classification model; the DriverSub-SVM cancer subtype classification model is used to divide cancer subtypes.

[0101] A further implementation manner is that the method for constructing the DriverSub-SVM cancer subtype classification model is as follows:

[0102] Construct a series of binary classifiers, and train each binary classifier based on the two-class subtype training samples of the already obtained global driver gene set; for example: for the K-class subtype division problem (e.g., |K|>2), it is necessary to train classifiers, where each classifier is responsible for distinguishing a pair of cancer subtype classes. These binary classifiers are trained using the standard SVM algorithm and the RBF kernel to solve the problem of non-linearly separable. The RBF kernel can make the output weights of each hidden node converge quickly. This feature enables the RBF kernel to better capture the complex relationships between biological features in cancer subtype classification, thereby improving the performance and accuracy of the model.

[0103] Establish constraint conditions to optimize the trained binary classifiers;

[0104] Specifically, for each binary classifier, the training data comes from the samples of the i-th and j-th subtypes, where i, j ∈ {1,..., K}, i < j, and the optimization problem can be expressed as follows:

[0105]

[0106] The constraint condition is:

[0107]

[0108] where, w ij is the weight vector defining between subtype classes i and j; b ijis the bias term for subtype categories i and j; is the tolerance level that allows some data points to fall within the interval or on the wrong side of the hyperplane; C is the regularization parameter, which is used to balance maximizing the interval and reducing the classification error; φ(x t ) is to use the RBF kernel to convert x t Mapped to feature space. The RBF kernel function is defined as:

[0109]

[0110] The kernel function is constructed by dividing two data points x t and x s Mapped to a high-dimensional space to calculate the similarity between them, thereby constructing a linear decision boundary.

[0111] Based on the optimized binary classifier, a decision function of the binary classifier is obtained;

[0112] Use the new training samples to train the optimized binary classifiers, and vote for each binary classifier based on the decision function;

[0113] In this embodiment, the decision function of each binary classifier is obtained from the previous optimization process and can be expressed as:

[0114]

[0115] For a new sample x new , each binary classifier in the OAO-MSVM model is based on its decision function Voting is performed to determine the final classification result through majority voting of all binary classifiers.

[0116] Based on the voting results, the classification results of the training samples are obtained, and the construction of the DriverSub-SVM cancer subtype classification model is completed.

[0117] In order to clearly understand OAO-MSVM, the present invention provides Figure 2 As shown in the geometric explanation. Assuming there are three subtype categories that need to be classified, there are a total of 3×(3-1) / 2=3 hyperplanes. Specifically, the present invention establishes three binary classifiers, and then cascades these three binary classifiers to achieve multi-class classification. The classification category is determined by the voting strategy: if If yes, the number of votes for subtype i increases; otherwise, the number of votes for subtype j increases. Sample x new The subtype with the highest number of votes is finally assigned to it, for example Figure 2If there is a tie in voting, the model assigns the sample to the subtype with the smallest index in the class label order. For example, if classifier 1 assigns the sample to subtype 2, classifier 2 assigns it to subtype 3, and classifier 3 assigns it to subtype 1, the sample is labeled as subtype 1. Figure 2 The dashed and solid lines between the three different subtypes represent the corresponding classification hyperplanes. It is worth noting that the crossover area in the figure indicates the occurrence of classification overlap, indicating that there is a tie in the voting strategy.

[0118] In order to alleviate the class imbalance problem in the data set, that is, the quantitative differences of different cancer subtypes, the present invention adopts the synthetic minority oversampling technology (SMOTE) method to increase the number of samples of minority cancer subtypes. In order to calculate the evaluation index, the present invention uses the confusion matrix of cancer subtypes. This method involves analyzing each subtype category separately and using true positive (TP), false positive (FP), false negative (FN) and true negative (TN) labeling matrices. In this case, TP represents the number of samples divided into the correct subtype, while FP and FN represent the instances of the corresponding subtype that are misclassified in rows and columns, respectively. TN is the number of instances that are correctly identified as not belonging to the subtype, that is, all other instances that do not belong to the current subtype analysis are classified as TN.

[0119] like Figure 3 As shown in the confusion matrix of the breast cancer dataset, the present invention uses the subtypes of breast cancer (i.e., "Normal", "LumA", "Her2", "LumB", and "Basal") as reference objects for analysis. This systematic evaluation method can fully understand the classification performance of the model for each subtype.

[0120] The performance of the model is evaluated using several metrics generated from the confusion matrix described previously, such as accuracy, precision, and recall. These metrics are defined as follows:

[0121]

[0122]

[0123]

[0124] A macro-level approach is used in the evaluation, where equal weight is used for each subtype when calculating the average performance metric. This approach prevents any subtype result from dominating the overall evaluation, thereby providing a fair assessment of the performance of the model across all subtypes. When calculating the macro-average precision, recall, and F1 score, the precision and recall of each individual subtype are first determined, and then these individual metrics are arithmetic averaged without considering the size or frequency of each subtype. The calculation formulas for these functions are as follows:

[0125]

[0126]

[0127]

[0128] Embodiment 2

[0129] In order to confirm the effectiveness of the model DriverSub-SVM of the present invention, this embodiment carried out a number of experimental verifications, including testing the optimal number of personalized driver genes, KEGG pathway analysis of identified global driver genes, ablation studies, comparative experiments, and biomarker analysis. In this experiment, the method of the present invention was compared with four benchmark methods (such as SVM-RFECV, DDCMNMF, MI-GA, and xGenemodel) on three cancer data sets (BRCA, COAD, and THCA). The performance of each model was evaluated using indicators such as accuracy, precision, recall, and F1 score. The present invention uses a k-fold cross-validation method to evaluate the effectiveness of the model. Specifically, all cancer samples are randomly divided into k subsets. Each subset is used as a test set in turn, and the remaining k-1 subsets are regarded as training sets for the model. The final classification result is the average of the k-fold cross-validation. In the study of the present invention, due to the large sample size of the BRCA data set, the present invention uses a 10-fold cross-validation; considering the small sample size of the COAD and THCA data sets, the present invention chooses a 5-fold cross-validation. In addition, the present invention also uses a grid search algorithm to select the best parameters to obtain the most accurate cancer subtype classification results. Here, the present invention sets the regularization parameter (C) to -1 to 10 4 In the logarithmic space, the kernel coefficient (gamma) parameter is set from 10 -5 to the logarithmic space of 1.

[0130] Number of personalized driver genes retained by the test:

[0131] After random walk, the present invention obtains a list of driver genes for each patient based on their diffusion scores. The higher the ranking, the more likely it is that the gene will become a candidate driver gene for the patient. In order to retain the most representative patient personalized driver genes for subsequent subtype classification, the present invention tests the effect of the number of personalized driver genes on the classification results. Figure 4As shown (the horizontal axis represents the number of retained personalized driver genes, and the vertical axis represents accuracy), as the number of retained personalized driver genes increases, the accuracy of colon cancer classification fluctuates significantly. In contrast, the effects on breast cancer and thyroid cancer are relatively small. As seen in the present invention, when the number of personalized driver genes increases from 10 to 80, the accuracy of these three cancer types is improved. However, as the number of personalized driver genes continues to increase, the accuracy of colon cancer drops sharply, while the accuracy of the other two cancer types drops slightly. This shows to a certain extent that adding more features does not necessarily lead to significant performance increases. Determining the appropriate number of genes is crucial to achieving higher accuracy. Based on the results of this test, the present invention sets the number of retained personalized driver genes to 80.

[0132] Analysis of global driver genes:

[0133] To demonstrate the effectiveness of the global driver gene selection method, we performed pathway enrichment analysis to evaluate the functional importance of the identified global driver genes in key biological pathways. For each cancer type, we retained pathways with a significance level (p-value) less than 0.05 and Figure 5 The top 10 enriched pathways are listed in . For breast cancer, a total of 1489 global driver genes were screened. KEGG pathway enrichment analysis showed that the predicted candidate global driver genes were associated with 104 pathways (all p values ​​< 0.05). Among them, some pathways are directly related to cancer, including proteoglycans in cancer, gastric cancer, breast cancer, and bladder cancer (see Figure 5 a). In addition, the identified global driver genes are also enriched in some important pathways. The MAPK signaling pathway is crucial for regulating various biological processes such as cell proliferation, apoptosis, differentiation and survival. It is closely related to the occurrence of tumors, especially in promoting the development of breast cancer. The PI3K-Akt signaling pathway is often associated with breast cancer cell proliferation, inhibition of apoptosis, and promotion of invasion and metastasis.

[0134] For colon cancer, a total of 220 genes were extracted as global driver genes. KEGG pathway enrichment analysis showed that the predicted candidate global driver genes were associated with 46 pathways (all p values ​​< 0.05). Among them, there are several cancer-related pathways, including breast cancer, colorectal cancer, microRNAs in cancer, and central carbon metabolism in cancer. Figure 5 b). In addition, the cytokine-cytokine receptor interaction pathway is a key signaling pathway in the human body that regulates biological processes such as immune response, cell proliferation, growth and differentiation. Abnormal regulation of this pathway is closely related to the occurrence, progression and metastasis of cancer. Studies have shown that human papillomavirus infection is prevalent in the colon and rectum of many patients with colorectal adenocarcinoma, suggesting that it may be involved in the pathogenesis of colorectal cancer. Approximately 5% of human cancers can be attributed to human papillomavirus infection.

[0135] For thyroid cancer, a total of 711 genes were identified as global driver genes for thyroid cancer. KEGG pathway enrichment analysis showed that the predicted candidate global driver genes were associated with 25 pathways (all p values ​​< 0.05). Among them, some pathways are related to cancer, including bladder cancer, proteoglycans in cancer, and central carbon metabolism in cancer. Figure 5 c). In addition, the ECM-receptor interaction pathway regulates the occurrence and development of thyroid cancer by affecting biological processes such as cell morphology, proliferation, survival and migration. The axon guidance signaling pathway may promote the development of thyroid cancer by promoting the proliferation, migration and invasion of tumor cells and regulating the tumor microenvironment.

[0136] These results illustrate the effectiveness of the selected global driver genes, which can be enriched in some biologically important pathways.

[0137] Ablation experiment:

[0138] The DriverSub-SVM model mainly includes three steps: personalized driver gene identification, global driver gene identification and cancer subtype classification. In order to determine the contribution of each step, the present invention designed three ablation experiments respectively, and evaluated the impact on the final performance indicators (including accuracy, precision, recall and F1 score) by excluding one of the steps in turn. As shown in Table 2, "no first filter" means that there is no personalized driver gene identification process, but all genes after random walk are used as inputs of the LR-L1 method for global gene screening. "No second filter" means that the model only identifies personalized driver genes and summarizes the results as features for subsequent cancer subtype classification. "No SVM" means that OAO-MSVM is replaced by other classification or clustering methods, such as K-nearest neighbor (KNN), random forest (RF), K-means clustering (K-means) and self-organizing map (SOM).

[0139] As can be seen from Table 2, the OAO-MSVM module has a particularly significant impact on subtype classification, followed by the second-layer filter and the first-layer filter. In the three datasets, the DriverSub-SVM model achieved 96.50%, 93.71%, and 99.30% accuracy in subtype classification of breast cancer, colon cancer, and thyroid cancer, respectively. Compared with the best performing algorithms in alternative modules (such as the RF algorithm for breast cancer and colon cancer, and the KNN algorithm for thyroid cancer), the OAO-MSVM module in the DriverSub-SVM model improved the accuracy of subtype classification by 4.99%, 20.65%, and 3.58%, respectively. Corresponding results were also observed in the other three evaluation indicators. In addition, the first-layer filter and the second-layer filter modules also had a certain impact on the final results. The absence of any of these two modules will lead to a decrease in the accuracy, precision, recall, and F1 score of subtype classification in the three datasets, with the decrease ranging from 0.64% to 2.03% for breast cancer, 0.25% to 20.69% for colon cancer, and 0.92% to 2.42% for thyroid cancer. This observation further proves the necessity of each module in the overall subtyping process.

[0140] Table 2

[0141]

[0142]

[0143] Comparative experiment:

[0144] The principle of selecting the comparison methods is that these methods have similar computational mechanisms to the methods of the present invention (e.g., SVM-RFECV, MI-GA), or are based on biological feature assumptions similar to the methods of the present invention (e.g., DDCMNMF, xGenemodel). SVM-RFECV uses the feature selection of identifying differentially co-expressed genes for each subtype as the first step, and further uses SVM and recursive feature elimination methods to classify cancer subtypes. However, this method mainly focuses on discovering gene classifiers specific to asymptomatic carriers and ATLL subtype patients, thus ignoring the exploration of other cancer types. The MI-GA method uses mutual information and genetic algorithms to select optimal genes, and uses the SVM method for cancer classification. This method mainly uses computational mechanisms to screen high-information genes in classification, and fails to effectively capture the biomolecular characteristics of cancer. Both DDCMNMF and xGenemodel use driver genes as key biological features for classifying cancer subtypes. However, these methods do not take into account the heterogeneity of each patient. In order to make a fair comparison with the above methods, the present invention uses the same patient data set and functional interaction network. The hyperparameters of all comparison algorithms are set according to the recommendations of the original literature.

[0145] Figure 6 a-with Figure 6 d shows the results of the accuracy, precision, recall and F1 value of the method of the present invention and other four methods. Since the results of the model of the present invention, SVM-RFECV model, MI-GA model and DDCMNMF model will fluctuate within a certain range, the average scores of k-fold cross validation of 10 or 5 iterations are used to evaluate their performance. Figure 6The DriverSub-SVM model shows superior performance in all four evaluation metrics compared to other methods. Specifically, on the BRCA cancer dataset, the DriverSub-SVM model shows significant performance with an average accuracy of 96.34%, an average precision of 96.37%, an average recall of 96.34%, and an average F1 score of 96.24%. Compared with the best performing method MI-GA, the DriverSub-SVM model improves by 10.71%, 10.69%, 24.86%, and 20.03% in accuracy, precision, recall, and F1 score, respectively. Similarly, on the COAD cancer dataset, the average accuracy, precision, recall, and F1 score of the DriverSub-SVM model reached 92.27%, 92.27%, 92.24%, and 92.25%, respectively, which were 16.16%, 18.13%, 18.26%, and 18.23% higher than MI-GA, respectively. On the THCA cancer dataset, the average accuracy, precision, recall, and F1 score of the DriverSub-SVM model reached 98.91%, 98.92%, 98.91%, and 98.90%, respectively, which exceeded MI-GA by 22.14%, 12.54%, 58.9%, and 60.77%, respectively. Overall, the DriverSub-SVM model showed excellent performance in cancer subtype classification, with MI-GA ranking second, which to some extent proves the importance of filtering informative features before classification. In addition, the figure also reflects the stability of DriverSub-SVM compared with other models. The fluctuation range is only between 0.2% and 1%, which is in sharp contrast to the significant fluctuations in the final results of models such as DDCMNMF.

[0146] Biomarker Identification:

[0147] In order to enhance the interpretability of the model, the present invention uses the KernelExplainer interpreter in the SHAP library. KernelExplainer is a tool that uses random sampling to approximate the impact of quantitative features (SHAP values) on model prediction results, which can effectively provide explanations for machine learning models. By calculating the average SHAP value of genes for each cancer subtype, the present invention ranks the impact of gene features on cancer subtype classification (see Figures 7 to 9). The present invention identified and listed the top 10 most important genes, explaining their contribution to the classification process. The present invention regards these genes as potential biomarkers for each subtype, including some known driver genes. In order to further verify the validity of the identified biomarkers, the present invention uses the PubTator website as a literature mining tool, searching for the keyword "cancer" and gene name (such as "Breast_Neoplasms" and "PAK1") at the same time, and further analyzes the possibility of each driver gene becoming a biomarker for cancer subtypes.

[0148] The method of the present invention identifies genes such as TP53, FOXO3, PIK3CA and GATA3 as biomarkers for breast cancer subtype classification. Figure 7 a- Figure 7 As shown in Figure 5, the SHAP graph shows the contribution of gene features to the classification of five subtypes of breast cancer, with each point representing a sample. The genes on the Y-axis are arranged from large to small according to their influence on the classification. The influence of each gene is represented by its SHAP value on the x-axis; a positive SHAP value indicates a positive contribution to the classification, while a negative SHAP value indicates a negative influence on the classification.

[0149] This is almost consistent with the results obtained through text mining. TP53 mutation enhances the immunogenicity of breast cancer by regulating key pathways such as p53-mediated cell cycle, Wnt, apoptosis, NOD-like receptors, Jak-STAT and glycolysis, suggesting that it has the potential to serve as a biomarker for immunotherapy. Studies have found that TP53 mutations are more frequent in HER2+ and Basal subtypes of breast cancer, which is consistent with the analysis results of the present invention. High TP53 values ​​are strongly correlated with the classification of HER2+ and Basal subtypes, with less impact on LumB subtypes and a negative impact on LumA subtypes. Studies have found that compared with normal tissues, the expression level of TSC2 is reduced in certain breast cancer tumor tissues, and this reduction in expression level is associated with a poor prognosis in breast cancer. Figure 7 In the results, high values ​​of TSC2 significantly enhanced the classification of the Normal subtype.

[0150] The present invention found that genes such as APC, TCF7L2, ACVR1B and RUNX1 can be used as biomarkers for colon cancer ( Figure 8 a- Figure 8 c) APC gene mutations are common in most cases of sporadic colorectal adenomas and cancers. Studies have shown that upregulating the APC gene by small interfering RNA can induce chromosomal instability (CIN) in chromosomally stable colon cancer cells. This is consistent with the present invention in Figure 8The conclusions drawn in the literature are consistent, that is, high APC values ​​have a positive effect on the classification of CIN subtypes, but a negative effect on the classification of MSI / CIMP subtypes. TCF7L2 is involved in the occurrence and development of colon cancer by regulating the Wnt / β-catenin signaling pathway. The frequency of TCF7L2 mutations is positively correlated with increased tissue invasion and metastasis.

[0151] The present invention observed that genes such as CUL1, BRAF, DSE and NRAS can be used as biomarkers for thyroid cancer ( Fig. 9 a- Fig. 9 d). CUL1 gene mutation or abnormal expression may affect the function of the intracellular ubiquitin-proteasome degradation pathway, thereby promoting the occurrence and progression of thyroid cancer. BRAF gene mutation is the most common genetic variation in thyroid cancer and is associated with follicular tumors. BRAF gene mutation is a highly potential therapeutic target for the treatment of thyroid cancer. Currently, there is limited literature on the impact of different expression values ​​of CUL1 and BRAF genes on different subtype classifications. However, if Fig. 9 As shown, the present invention can boldly speculate that high values ​​of CUL1 and low values ​​of BRAF have a positive impact on the classification of follicular subtypes, but have a negative impact on the classification of classical / common subtypes.

[0152] Embodiment 3

[0153] The present invention also provides a cancer subtype classification system based on driver gene identification, which is used to implement a cancer subtype classification method, including:

[0154] Data preprocessing module, used to acquire and preprocess cancer datasets;

[0155] The gene identification module is used to identify driver genes based on the preprocessed cancer dataset and obtain global driver genes;

[0156] A model building module is used to build a DriverSub-SVM cancer subtype classification model based on global driver genes and one-to-one multi-classification support vector machine algorithm;

[0157] A classification module is used to construct a DriverSub-SVM cancer subtype classification model based on the global driver gene and a one-to-one multi-classification support vector machine algorithm; the DriverSub-SVM cancer subtype classification model is used to divide cancer subtypes.

[0158] A further implementation method is that, in the data preprocessing module, the cancer types in the cancer data set include BRCA breast cancer, COAD colon cancer and THCA thyroid cancer; wherein each cancer type includes four omics data: transcriptome expression data, somatic mutation data, copy number variation data and protein function interaction network data; the data preprocessing module includes:

[0159] The normalization unit is used to normalize the variance-stabilizing transformed VST expression values ​​of transcriptome expression data based on Gauss's theorem;

[0160] An abnormality discrimination unit, used to discriminate abnormalities of gene expression on the normalized transcriptome expression data based on a preset threshold;

[0161] An assignment unit, used for assigning values ​​to the discrimination results to obtain a binary matrix;

[0162] The abnormal expression gene data acquisition unit is used to obtain the abnormal expression gene data of the patient based on the binary matrix to complete the preprocessing of the cancer data set; wherein the preprocessed cancer data set includes the abnormal expression gene data, somatic mutation gene data and protein function interaction network of each cancer type.

[0163] In this embodiment, for transcriptome expression data, the present invention uses Gauss's theorem to standardize the VST expression value to determine the abnormality of gene expression. If the absolute value of the z value of a particular patient's gene is greater than or equal to 2δ, the gene is considered to be abnormally expressed in the patient. According to equation (1), if a gene meets this threshold, it is assigned a value of 1, otherwise it is 0. This process generates a binary matrix, which is called patient abnormal expression data.

[0164]

[0165] The present invention simultaneously retains sample data and gene data having three data types, that is, the preprocessed cancer data set includes abnormally expressed gene data, somatic mutation gene data, and protein function interaction network of each cancer type.

[0166] In this embodiment, according to the above data preprocessing steps, the present invention obtained 1089 BRCA samples, 134 COAD samples and 463 THCA samples.

[0167] In a further embodiment, the gene identification module comprises:

[0168] A bipartite graph construction unit, used to construct a bipartite graph based on abnormal expression gene data and somatic mutation gene data;

[0169] A relationship acquisition unit, used to obtain the association relationship between abnormally expressed genes and somatic mutation genes in the protein functional interaction network based on a bipartite graph;

[0170] A random scoring unit is used to score abnormally expressed genes and somatic mutation genes using a three-round random walk algorithm based on association relationships, abnormally expressed gene data, somatic mutation gene data, and a damping factor to obtain a random walk score matrix of mutant genes; wherein the damping factor represents the degree to which a gene depends on a protein functional interaction network or a gene's own characteristics;

[0171] A gene list acquisition unit, used to obtain a personalized driver gene list for a patient based on a random walk score matrix;

[0172] The present invention uses a three-round random walk algorithm to evaluate the bidirectional impact between mutated driver genes and dysregulated downstream genes in the biological function network, taking into account both the impact of driver gene mutations on downstream genes and the reverse impact of dysregulated downstream gene expression on mutated driver genes. The present invention is based on a bipartite graph constructed from previously identified mutant genes and abnormally expressed genes to represent the association between genes in the biological function network, where P ij =1 indicates the association between mutant gene i and abnormally expressed gene j. m (i) represents the score of the mutant gene i after the first or third random walk, S a (j) represents the score of abnormally expressed gene j after the second random walk. M(i) represents the mutation data of gene i, and A(j) represents the data of abnormally expressed gene j (where 1 represents abnormality and 0 represents normality). Subsequently, for each mutant gene in the mutation matrix and each expressed gene in the abnormal expression matrix, their scores can be calculated using formulas (2) to (4). The damping factor a here represents the degree of dependence on the functional interaction network structure or its own characteristics.

[0173]

[0174]

[0175]

[0176] Based on the random walk score matrix, a personalized driver gene list is obtained for the patient;

[0177] Aggregate the personalized driver gene lists of each patient to obtain a set of personalized driver genes;

[0178] In this embodiment, after three rounds of random walks, all mutant genes of each patient will be scored accordingly, thereby generating a random walk score matrix. Then, the scores of the driver genes are arranged in descending order to obtain a personalized driver gene list for the patient. The higher the ranking, the greater the possibility that the gene is a driver gene in a specific sample. Finally, the present invention selects the top 80 genes as the driver gene list for each patient based on the overall performance after DriverSub-SVM classification (see Example 2 for details).

[0179] A gene aggregation unit, used to aggregate the personalized driver gene list of each patient to obtain a set of personalized driver genes;

[0180] The global driver gene acquisition unit is used to identify the representative features of personalized driver genes based on the L1 regularized logistic regression model and construct a set of global driver genes based on the representative features.

[0181] A further implementation method is that the construction process of the L1 regularized logistic regression model is:

[0182] Construct a multi-class sample collection based on cancer datasets and personalized driver gene signatures;

[0183] Calculate the conditional probability distribution of pre-set categories of cancer subtypes in a multi-class sample set based on personalized driver gene signatures;

[0184] Based on the conditional probability distribution, the objective function of the logistic regression with L1 norm penalty is obtained, and the construction of the L1 regularized logistic regression model is completed.

[0185] In this embodiment, in order to obtain a comprehensive set of driver genes that can reflect both the unique characteristics of individual patients and the common characteristics of the entire patient population, the personalized driver gene lists of each patient are brought together. Subsequently, in order to identify the most representative features, the present invention uses a logistic regression (LR-L1) model with L1 regularization. These selected features are used to form a global driver gene set, thereby including both the specific characteristics of individual patients and the global characteristics of the patient population. As an embedded feature selection technique, the LR-L1 model aims to improve the interpretability of the classifier while reducing the feature dimension. In the study of this embodiment, L1 regularization promotes sparse solutions, which helps to identify the most important features from the patient gene pool.

[0186] Consider a multi-class sample set T = {(x1,y1),(x2,y2),…,(x N ,y N )},here y i∈{1,2…,K}, where N represents the number of cancer samples, M represents the number of personalized driver gene signatures of patients, and K represents the total number of cancer subtype categories in the samples. Given the input feature x, the conditional probability distribution of the kth class can be expressed as:

[0187]

[0188] The objective function of logistic regression with L1 norm penalty is as follows:

[0189]

[0190] Here, y i,k Indicates that sample i belongs to the true label of category k, where a value of 1 indicates that it belongs to the category and a value of 0 indicates that it does not belong to the category. represents the probability that sample i belongs to category k. Given the model parameter matrix w, w j,k Represents the weight corresponding to the jth feature in category k, and λ is the regularization strength. In the experiments of the present invention, the number of global driver genes selected varies for different cancer types, depending on factors such as sample size, gene characteristics, and data set complexity. For example, after calculation using the LR-L1 model, the BRCA data set has 1489 global driver genes, COAD has 220, and THCA has 711. In order to further explore the biological relevance of the identified global driver gene set, the present invention uses the KEGG pathway for enrichment analysis (see Example 2 for details).

[0191] In a further embodiment, the model building module comprises:

[0192] The first training unit is used to construct a series of binary classifiers and train each binary classifier based on the two subtype samples of the global driver gene set. For example, for the K-type subtype classification problem (for example, |K|>2), it is necessary to train classifiers, each of which is responsible for distinguishing a pair of cancer subtype categories. These binary classifiers are trained using the standard SVM algorithm and RBF kernel to solve nonlinearly separable problems. The RBF kernel allows the output weights of each hidden node to converge quickly. This property enables the RBF kernel to better capture the complex relationships between biological features in cancer subtype classification, thereby improving the performance and accuracy of the model.

[0193] An optimization unit, used to establish constraints and optimize the trained binary classifier;

[0194] A decision function acquisition unit, configured to obtain the decision function of a binary classifier based on the optimized binary classifier; specifically, for each binary classifier, the training data is samples from the i-th and j-th subtypes, where i, j ∈ {1,..., K}, i < j, and the optimization problem can be formulated as follows:

[0195]

[0196] The constraint conditions are:

[0197]

[0198] where w ij is the weight vector defining between subtype categories i and j; b ij is the bias term of subtype categories i and j; is the tolerance level allowing some data points to fall within the margin or on the wrong side of the hyperplane; C is the regularization parameter, used to balance maximizing the margin and reducing classification errors; φ(x t ) maps x t to the feature space using the RBF kernel. The RBF kernel function is defined as:

[0199]

[0200] The kernel function calculates the similarity between two data points x t and x s by mapping them to a high-dimensional space, thereby constructing a linear decision boundary.

[0201] A second training unit, configured to train the optimized binary classifier using new samples and vote on each binary classifier based on the decision function;

[0202] A training sample classification unit, configured to obtain the classification result of the training samples based on the voting result, and complete the construction of the DriverSub-SVM cancer subtype classification model.

[0203] Specifically, train the optimized binary classifier using new training samples and vote on each binary classifier based on the decision function;

[0204] In this embodiment, the decision function of each binary classifier is obtained from the previous optimization process and can be expressed as:

[0205]

[0206] For the new sample x new , each binary classifier in the OAO-MSVM model makes a decision according to its decision function Voting is performed to determine the final classification result through majority voting of all binary classifiers.

[0207] Based on the voting results, the classification results of the training samples are obtained, and the construction of the DriverSub-SVM cancer subtype classification model is completed.

[0208] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for classifying cancer subtypes based on driver gene identification, characterized in that: The following steps are involved: Obtain and preprocess a cancer dataset; the cancer types in the cancer dataset include BRCA breast cancer, COAD colon cancer, and THCA thyroid cancer; wherein each cancer type includes four types of omics data: transcriptome expression data, somatic mutation data, copy number variation data, and protein function interaction network data; The method for preprocessing the cancer dataset includes: Based on Gauss's theorem, the variance-stabilized transformed (VST) expression values ​​of the transcriptome expression data are normalized; Based on a preset threshold, the normalized transcriptome expression data is subjected to gene expression abnormality discrimination; Assign values ​​to the discrimination results to obtain a binary matrix; Based on the binary matrix, abnormally expressed gene data of the patient is obtained to complete preprocessing of the cancer data set; The pre-processed cancer dataset includes abnormally expressed gene data, somatic mutation gene data, and protein functional interaction network for each cancer type; Based on the preprocessed cancer data set, the driver gene is identified to obtain the global driver gene; the method for obtaining the global driver gene is: Constructing a bipartite graph based on the abnormally expressed gene data and the somatic mutation gene data; Based on the bipartite graph, obtaining the association relationship between abnormally expressed genes and somatic mutation genes in the protein functional interaction network; A three-round random walk algorithm is used to score abnormally expressed genes and somatic mutation genes based on the association relationship, the abnormally expressed gene data, the somatic mutation gene data and the damping factor to obtain a random walk score matrix of mutant genes; wherein the damping factor represents the degree to which a gene depends on a protein functional interaction network or a gene's own characteristics; Based on the random walk score matrix, obtaining a personalized driver gene list for the patient; Aggregating the personalized driver gene lists of each patient to obtain a set of personalized driver genes; Identify representative features of the personalized driver genes based on an L1 regularized logistic regression model, and construct a global driver gene set based on the representative features; Based on the global driver genes and a one-to-one multi-classification support vector machine algorithm, a DriverSub-SVM cancer subtype classification model is constructed; the DriverSub-SVM cancer subtype classification model is used to divide cancer subtypes.

2. The method for classifying cancer subtypes based on driver gene identification according to claim 1, characterized in that: The construction method of the L1 regularized logistic regression model is: Based on the cancer dataset and personalized driver gene signatures, construct a multi-class sample set; Based on the personalized driver gene signature, calculating the conditional probability distribution of the preset categories of cancer subtypes in the multi-class sample set; Based on the conditional probability distribution, the objective function of the logistic regression with L1 norm penalty is obtained, and the construction of the L1 regularized logistic regression model is completed.

3. A cancer subtype classification system based on driver gene identification, characterized in that: The method for implementing the cancer subtype classification method according to any one of claims 1 to 2 comprises: Data preprocessing module, used to acquire and preprocess cancer datasets; A gene identification module, used to identify driver genes based on the preprocessed cancer data set to obtain global driver genes; A model building module is used to build a DriverSub-SVM cancer subtype classification model based on the global driver gene and a one-to-one multi-classification support vector machine algorithm; the DriverSub-SVM cancer subtype classification model is used to divide cancer subtypes.

4. The cancer subtype classification system based on driver gene identification according to claim 3, characterized in that: In the data preprocessing module, the cancer types in the cancer data set include BRCA breast cancer, COAD colon cancer and THCA thyroid cancer; wherein each cancer type includes four types of omics data: transcriptome expression data, somatic mutation data, copy number variation data and protein function interaction network data; the data preprocessing module includes: A standardization unit, used for standardizing the variance-stabilized transformed (VST) expression value of the transcriptome expression data based on Gauss's theorem; An abnormality identification unit, used for identifying abnormalities of gene expression in the normalized transcriptome expression data based on a preset threshold; An assignment unit, used for assigning values ​​to the discrimination results to obtain a binary matrix; The abnormal expression gene data acquisition unit is used to obtain the abnormal expression gene data of the patient based on the binary matrix to complete the preprocessing of the cancer data set; wherein the preprocessed cancer data set includes the abnormal expression gene data, somatic mutation gene data and protein function interaction network of each cancer type.

5. The cancer subtype classification system based on driver gene identification according to claim 4, characterized in that: The gene recognition module comprises: A bipartite graph construction unit, configured to construct a bipartite graph based on the abnormally expressed gene data and the somatic mutation gene data; A relationship acquisition unit, used for obtaining the association relationship between abnormally expressed genes and somatic mutation genes in the protein function interaction network based on the bipartite graph; A random scoring unit, for scoring abnormally expressed genes and somatic mutation genes by using a three-round random walk algorithm based on the association relationship, the abnormally expressed gene data, the somatic mutation gene data and a damping factor, to obtain a random walk score matrix of the mutant gene; wherein the damping factor represents the degree to which the gene depends on the protein function interaction network or the gene's own characteristics; A gene list acquisition unit, used to obtain a personalized driver gene list for the patient based on the random walk score matrix; A gene aggregation unit, used to aggregate the personalized driver gene list of each patient to obtain a set of personalized driver genes; The global driver gene acquisition unit is used to identify the representative features of the personalized driver genes based on the L1 regularized logistic regression model, and construct the set of global driver genes based on the representative features.

6. The cancer subtype classification system based on driver gene identification according to claim 5, characterized in that: The construction process of the L1 regularized logistic regression model is: Based on the cancer dataset and personalized driver gene signatures, construct a multi-class sample set; Based on the personalized driver gene signature, calculating the conditional probability distribution of the preset categories of cancer subtypes in the multi-class sample set; Based on the conditional probability distribution, the objective function of the logistic regression with L1 norm penalty is obtained, and the construction of the L1 regularized logistic regression model is completed.

Images