A graph convolutional personalized cancer driver gene prediction model and method based on a heterogeneous network

Abstract

The application discloses a kind of graph convolutional personalized cancer driver gene prediction model and method based on heterogeneous network, comprising the following steps: S1, construct mutation-sample association matrix and expression-sample association matrix;S2, obtain gene similarity matrix and sample similarity matrix;S3, obtain driver gene-sample association matrix;S4, gene similarity matrix, sample similarity matrix and driver gene-sample association matrix are fused into heterogeneous network;S5, based on graph convolutional GCN encoder, gene and sample embedding are carried out to heterogeneous network;S6, obtain new gene-sample association.The application obtains the low-dimensional vector of gene and sample by the similarity association relationship between gene-gene and sample-sample, by constructing the encoder based on graph convolutional GCN, constructs decoder to reconstruct new gene-sample association, better association gene and sample feature, to further improve the accuracy of personalized driver gene identification.

Description

Technical Field

[0001] This invention relates to the field of cancer driver gene prediction technology, and in particular to a graph convolutional personalized cancer driver gene prediction model and method based on heterogeneous networks. Background Technology

[0002] Cancer is a complex and evolving disease that contributes to high morbidity and mortality rates worldwide. Early identification of cancerous lesions is one of the most successful methods for minimizing cancer mortality, but it requires a deeper understanding of tumor development and its molecular basis. The causes of cancer are complex, primarily resulting from genetic alterations, including single nucleotide variants, copy number variations, insertions or deletions, and structural deviations. Generally, mutations that provide a selective growth advantage in tumor cells and promote cancer development are called driver mutations. Simultaneously, other randomly occurring mutations may also exist in tumor samples, but these are not directly related to cancer development; these are called passenger mutations. Cancer driver genes are genes that carry cancer driver mutations. A major goal of cancer genomics is to identify cancer-associated driver genes in different tumor types. The discovery of driver genes is crucial for cancer prevention, diagnosis, and clinical treatment.

[0003] In recent years, with advancements in computer and sequencing technologies, numerous methods have been developed to detect mutations and genes in genomic data that are causally related to cancer. Most methods focus on discovering driver genes specific to certain cancer types. Depending on the target audience, methods for predicting cancer driver genes can generally be categorized into population-based and individual-based approaches. With the increasing popularity of precision medicine, population-based methods have revealed their limitations; they cannot identify rare driver genes present in small cohorts or individual patients, thus hindering precision medicine. Furthermore, even among patients with the same cancer type and receiving the same treatment, genetic or environmental factors can influence outcomes, leading to different results for different patients. Therefore, it is necessary to research personalized cancer driver gene analysis for individual patients.

[0004] Currently, the main method for identifying personalized driver genes is based on ranking mutated genes, which takes into account the mutation data in each patient sample and their connections in the gene network. If a mutated gene scores highly in an individual, it is more likely to be a driver gene. However, existing network control methods cannot directly or effectively identify the driver gene for each patient because they only consider consistent information within the sample population, leading to situations where some samples have no predictive outcome. Furthermore, the longer the path between genes, the more information is diluted, which is also a problem with methods applying network control strategies.

[0005] From another perspective, the problem of identifying driver genes in an individual can be viewed as predicting the relationship between the individual and the gene. To represent the association between genes and samples, similarity information can be used to establish connections. Similarity information has potential for driver gene prediction, but many current methods ignore this. Studies have indicated that genes with similar functions to driver genes are likely to play similar roles in cancer progression, and potential driver genes are more abundant among functionally similar genes.

[0006] Patent document CN116805513B discloses a method for predicting and analyzing cancer driver genes based on a heterogeneous graph Transformer framework. This method includes: constructing a heterogeneous network of genes and proteins using gene-gene interactions, protein-protein interactions, and gene-protein correspondences; constructing a heterogeneous graph Transformer module and generating embeddings of target nodes based on the module and the heterogeneous gene-protein network; constructing a fully connected layer classification module and generating cancer driver gene prediction results based on the module and the target node embeddings; and analyzing the prediction results. This application fully utilizes the relationships between entities in different biological networks, solving the problem of underutilization of prior information in biological networks and improving the accuracy of cancer driver gene prediction. However, it also has limitations: it does not deeply analyze gene-gene similarity relationships and sample-sample similarity relationships, and it cannot analyze potential driver genes at the individual level, which is detrimental to improving the accuracy of personalized driver gene identification.

[0007] Patent document CN113517021B discloses a method for predicting cancer driver genes. This application can effectively improve the prediction accuracy and the generalization ability of the model application, but it also has the problem that it does not conduct in-depth analysis of the similarity association between genes and between samples, and therefore cannot be applied to the problem of predicting driver genes on an individual basis, which is not conducive to improving the accuracy of personalized driver gene identification. Summary of the Invention

[0008] The purpose of this invention is to provide a graph convolutional personalized cancer driver gene prediction model and method based on heterogeneous networks, which conducts in-depth analysis of the similarity association between genes and between samples, and improves the accuracy of personalized driver gene identification by constructing an encoder based on graph convolutional GCN to better associate the features of genes and samples.

[0009] The objective of this invention can be achieved through the following technical solution: a graph convolution-based personalized cancer driver gene prediction model and method based on heterogeneous networks, comprising the following steps:

[0010] S1. Preprocess cancer-related data to construct mutation-sample association matrices and expression-sample association matrices;

[0011] S2. Perform the Gaussian kernel similarity algorithm on the expression-sample association matrix to obtain the gene similarity matrix and the sample similarity matrix;

[0012] S3. Process the mutation-sample association matrix using the public database DriverDB to obtain the driver gene-sample association matrix;

[0013] S4. The gene similarity matrix, sample similarity matrix, and driver gene-sample association matrix are fused into a heterogeneous network;

[0014] S5. Gene embedding and sample embedding of heterogeneous networks based on graph convolutional GCN encoder;

[0015] S6. Use the decoder to reconstruct and obtain new gene-sample associations.

[0016] Furthermore, the cancer-related data in S1 includes cancer expression data, mutation data, and protein-protein interaction network data.

[0017] Furthermore, the preprocessing includes obtaining a z-score based on gene expression values, then removing outlier genes from the protein-protein interaction network data based on the score, while retaining mutated genes. Specifically:

[0018]

[0019] Where x is the original expression value of the sample gene, It is the mean of gene expression values ​​in the sample. is the standard deviation of the sample gene expression value, and z is the score of the sample gene expression value (z-score).

[0020] The criterion for determining outlier gene deletion is |z|≥2.

[0021] Furthermore, in S1, the mutation-sample association matrix and expression-sample association matrix are constructed as follows:

[0022]

[0023]

[0024] in, The mutation-sample association matrix, The expression-sample association matrix is ​​as follows: rows represent mutated genes. This represents the number of genes selected; the columns represent different samples. The number of samples.

[0025] Furthermore, in S2, a Gaussian kernel similarity algorithm is performed on the expression-sample association matrix to obtain a gene similarity matrix and a sample similarity matrix; specifically including:

[0026] S21. Calculate the similarity score between genes in the expression-sample association matrix, expressed as:

[0027]

[0028] in, and To express the sample correlation matrix The genes in Indicates gene Expression spectrum, A parameter that controls kernel bandwidth:

[0029]

[0030] S22. Calculate the similarity score between samples in the expression-sample association matrix, expressed as:

[0031]

[0032] in, and To express the sample correlation matrix The samples in Indicates sample Expression spectrum, Indicates sample Expression spectrum, A parameter that controls kernel bandwidth:

[0033]

[0034] S23. Score based on similarity between genes Obtain the gene similarity matrix The similarity score between samples Obtain the sample similarity matrix .

[0035] Furthermore, in step S3, the mutation-sample association matrix is ​​processed using the public database DriverDB to obtain the driver gene-sample association matrix, specifically including:

[0036] S31, in the mutation-sample correlation matrix The DriverDB database is used for filtering.

[0037] S32, If the mutation-sample association matrix Zhonggen In the sample If a mutation has occurred and it is a known driver gene present in the DriverDB cancer database, then it is... Reserve it as 1, otherwise Set to 0;

[0038] S33. Obtain the driver gene-sample association matrix. , represented as:

[0039]

[0040] Among them, rows represent mutated genes. This represents the number of genes selected; the columns represent samples. The number of samples.

[0041] Furthermore, in step S4, the gene similarity matrix, sample similarity matrix, and driver gene-sample association matrix are fused into a heterogeneous network. , represented as:

[0042]

[0043] in, The driver gene-sample association matrix is ​​represented as a binary matrix. , for The transpose of the matrix; for Standardized matrix; for Standardized matrix;

[0044] =

[0045] =

[0046] in, =diag( )yes The degree matrix; =diag( )yes The degree matrix.

[0047] Furthermore, in S5, the encoder based on graph convolutional GCN performs gene embedding and sample embedding on the heterogeneous network, specifically including:

[0048] S51, In heterogeneous networks The graph convolutional GCN is used for L iterations of embedding, which is represented as:

[0049]

[0050] in, The introduced penalty factor, ;

[0051] It is the first Embedding of layer nodes D=diag( )yes The degree matrix, W ( ) is a trainable weight matrix for a specific layer. ( ) is a non-linear activation function;

[0052] S52, Based on heterogeneous networks For each layer of embedding, the final embedding of the gene and sample is obtained, represented as:

[0053] =

[0054] in, It is the final embedding of genes. It is the final embedding of the sample. For automatic learning of neural networks, It is the dimension of the embedding, initialized to .

[0055] Furthermore, the decoder in S6 is a bilinear decoder, and new gene-sample associations are obtained by reconstructing using the decoder, as shown below:

[0056]

[0057] in, It is a non-linear activation function. Given a trainable weight matrix, This is the final embedding matrix of the genes. is the transpose of the final embedding matrix of the sample.

[0058] A graph convolutional personalized cancer driver gene prediction model based on heterogeneous networks includes:

[0059] The cancer-related data unit acquires cancer expression data, mutation data, and protein-protein interaction network data, and performs preprocessing.

[0060] The association matrix unit uses the mutation-sample association matrix and the expression-sample association matrix to associate preprocessed data and obtain mutation-sample association and expression-sample association;

[0061] The similarity association unit utilizes the driver gene-sample association matrix, gene similarity matrix, and sample similarity matrix to perform driver gene-sample association, gene similarity, and sample similarity analysis on mutation-sample association data and expression-sample association data;

[0062] Heterogeneous network units fuse data from driver gene-sample association, gene similarity, and sample similarity analysis to obtain heterogeneous networks;

[0063] The encoding / decoding unit encodes embedded genes and samples in a heterogeneous network and decodes them to obtain the association between genes and samples.

[0064] The beneficial effects of this invention are:

[0065] 1. This invention utilizes the similarity relationships between genes and between samples to construct a graph convolutional network (GCN)-based encoder to obtain low-dimensional vectors of genes and samples, and then constructs a decoder to reconstruct new gene-sample associations, thereby better associating the features of genes and samples and further improving the accuracy of personalized gene recognition.

[0066] 2. This invention obtains a similarity matrix by calculating similarity scores between genes and similarity scores between samples. and sample similarity matrix After processing using the public database DriverDB, the driver gene-sample association matrix was obtained. Obtaining heterogeneous networks This allows for a better association between gene and sample characteristics, thereby further improving the accuracy of personalized gene identification.

[0067] 3. This invention constructs an encoder based on graph convolutional neural networks (GCNs) to learn low-dimensional vector representations of genes and samples. Each GCN layer aggregates information from adjacent layers and repeatedly considers the contributions of different embeddings on different layers. The final embeddings of genes and samples obtained are comprehensive and can better associate the features of genes and samples, thereby further improving the accuracy of personalized gene recognition. Attached Figure Description

[0068] Figure 1 This is a schematic diagram of the GCSDG model structure of the present invention;

[0069] Figure 2 This is an analysis chart of data from three types of cancer (ESCA, GBM, PAAD) in Embodiment 1 of the present invention;

[0070] Figure 3 This is a comparison chart of the performance indicators of the method of the present invention in Embodiment 1 and other methods;

[0071] Figure 4 This is a distance distribution diagram between gene pairs in Embodiment 1 of the present invention. Detailed Implementation

[0072] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar symbols denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0073] like Figures 1-4 As shown, this invention discloses a graph convolution-based personalized cancer driver gene prediction model and method based on heterogeneous networks, comprising:

[0074] S1. Preprocess cancer-related data to construct mutation-sample association matrices and expression-sample association matrices;

[0075] S2. Perform the Gaussian kernel similarity algorithm on the expression-sample association matrix to obtain the gene similarity matrix and the sample similarity matrix;

[0076] S3. Process the mutation-sample association matrix using the public database DriverDB to obtain the driver gene-sample association matrix;

[0077] S4. The gene similarity matrix, sample similarity matrix, and driver gene-sample association matrix are fused into a heterogeneous network;

[0078] S5. A graph convolution-based encoder performs gene embedding and sample embedding on heterogeneous networks.

[0079] S6. Use the decoder to reconstruct and obtain new gene-sample associations.

[0080] A graph convolutional personalized cancer driver gene prediction model based on heterogeneous networks includes:

[0081] The cancer-related data unit acquires cancer expression data, mutation data, and protein-protein interaction network data, and performs preprocessing.

[0082] The association matrix unit uses the mutation-sample association matrix and the expression-sample association matrix to associate preprocessed data and obtain mutation-sample association and expression-sample association;

[0083] The similarity association unit utilizes the driver gene-sample association matrix, gene similarity matrix, and sample similarity matrix to perform driver gene-sample association, gene similarity, and sample similarity analysis on mutation-sample association data and expression-sample association data;

[0084] Heterogeneous network units fuse data from driver gene-sample association, gene similarity, and sample similarity analysis to obtain heterogeneous networks;

[0085] The encoding / decoding unit encodes embedded genes and samples in a heterogeneous network and decodes them to obtain the association between genes and samples.

[0086] like Figure 1 As shown, cancer-related data includes cancer expression data, mutation data, and protein-protein interaction network data.

[0087] In cancer-related data, cancer driver genes are genes that have undergone somatic mutations. Driver genes usually affect the expression of genes linked to them in biological networks. If a mutated gene can significantly affect the expression of its downstream genes, it is more likely to be a driver gene; mutated genes that cannot significantly affect the expression of downstream genes are outliers.

[0088] Based on this, we screen out mutant genes that can significantly alter gene expression. In other words, we delete outliers in the gene-gene interaction network and retain only those driver genes that can significantly affect the expression of downstream genes.

[0089] The z-score can be obtained by calculating the gene expression values ​​of the sample, and outlier genes can be removed based on the score. The calculation method can be expressed as follows:

[0090]

[0091] in, It is the original expression value of the sample gene. It is the mean of gene expression values ​​in the sample. is the standard deviation of the sample gene expression value, and z is the score of the sample gene expression value (z-score).

[0092] The criterion for determining outlier gene deletion is |z|≥2.

[0093] After preprocessing, a mutation-sample association matrix is ​​constructed. Specifically:

[0094]

[0095] Mutation-Sample Association Matrix The value consists of 0 and 1. =1 indicates a gene In the sample A mutation occurred in it. =0 indicates that no mutation has occurred, and the row represents the mutated gene. This represents the number of genes selected; the columns represent different samples. The number of samples.

[0096] Constructing a mutation-sample association matrix Specifically:

[0097]

[0098] Mutation-Sample Association Matrix After normalization, the value ranges from [0,1], representing the expression of different genes in different samples. The row represents the mutated gene. This represents the number of genes selected; the columns represent different samples. The number of samples.

[0099] The Gaussian kernel similarity algorithm is performed on the expression-sample association matrix to obtain the gene similarity matrix and the sample similarity matrix; specifically including:

[0100] The similarity score between genes in the expression-sample association matrix is ​​calculated and represented as:

[0101]

[0102] in, and To express the sample correlation matrix The genes in Indicates gene Expression spectrum, A parameter that controls kernel bandwidth:

[0103]

[0104] The similarity score between samples in the expression-sample association matrix is ​​calculated and represented as:

[0105]

[0106] in, and To express the sample correlation matrix The samples in Indicates sample Expression spectrum, Indicates sample Expression spectrum, A parameter that controls kernel bandwidth:

[0107]

[0108] Then, the similarity scores between genes are used. Obtain the gene similarity matrix The similarity score between samples Obtain the sample similarity matrix .

[0109] For mutated genes, if a mutated gene in a sample also exists in a known driver gene database, then personalized driver gene prediction methods should focus more on that mutated gene. Therefore, known driver genes can be used as prior information to further obtain the driver gene-sample association matrix.

[0110] In the mutation-sample association matrix The DriverDB database is used for filtering; if the mutation-sample association matrix... Zhonggen In the sample If a mutation has occurred and it is a known driver gene present in the DriverDB cancer database, then it is... Reserve it as 1, otherwise Set to 0, and obtain the driver gene-sample association matrix as follows:

[0111]

[0112] Among them, rows represent mutated genes. This represents the number of genes selected; the columns represent samples. The number of samples.

[0113] Furthermore, the gene similarity matrix, sample similarity matrix, and driver gene-sample association matrix are fused into a heterogeneous network. , is represented as:

[0114]

[0115] in, The driver gene-sample association matrix is ​​represented as a binary matrix. , for The transpose of the matrix; i, j represent the number of genes and samples in the driver gene-sample association matrix, respectively; if An equal value of 1 indicates a gene. With sample There is a connection, otherwise there is no connection.

[0116] for Standardized matrix:

[0117] =

[0118] for Standardized matrix:

[0119] =

[0120] in, =diag( )yes The degree matrix; =diag( )yes The degree matrix.

[0121] Furthermore, the encoder based on graph convolutional GCN performs gene embedding and sample embedding on heterogeneous networks, specifically including:

[0122] In heterogeneous networks Perform graph convolution GCN to obtain heterogeneous networks The embedding of each layer is represented as:

[0123]

[0124] in, , The introduced penalty factor; Heterogeneous networks Add a penalty factor The representation of .

[0125] It is the first Embedding of layer nodes D=diag( )yes The degree matrix, W ( ) is a trainable weight matrix for a specific layer. ( ) is a non-linear activation function.

[0126] For example, the first layer propagation of the graph convolutional GCN encoder is:

[0127]

[0128] Among them, It is a weight matrix. It is the first layer of embedding in heterogeneous networks. It is the dimension of the embedding.

[0129] and After L iterations, L k-dimensional embeddings can be obtained from different graph convolutional GCNs.

[0130] In Graph Convolutional CN, the Exponential Linear Unit (ELU) can be used as a non-linear activation function, which can accelerate the learning process and significantly improve generalization performance. This activation function is expressed as follows:

[0131]

[0132] Considering the inconsistent contributions of different embeddings across different layers, an attention mechanism can be introduced at each layer to combine these embeddings, resulting in the final embeddings of the gene and the sample:

[0133] =

[0134] in, It is the final embedding of genes. It is the final embedding of the sample. Automatically learned by neural network, initialized to , .

[0135] To reconstruct new gene-sample associations, a bilinear decoder is used to reconstruct and obtain the new gene-sample associations, represented as follows:

[0136]

[0137] in, It is a non-linear activation function. It is a trainable weight matrix. This is the final embedding matrix of the genes. is the transpose of the final embedding matrix of the sample.

[0138] Gene and samples The predicted scores associated with each other can be obtained from The conclusion is as follows.

[0139] Example 1: To test the effectiveness of the method of the present invention, the method of the present invention was applied to three cancer datasets from TCGA: esophageal carcinoma (ESCA), glioma multiforme (GBM), and pancreatic adenocarcinoma (PAAD).

[0140] Studies have shown that at least five typical driver genes (CDGs) are needed in a patient's body to fully explain an individual's cancer development. This study analyzed cancer data, calculating the percentage of samples with known typical driver genes in 503 samples across three cancer types (ESCA, GBM, PAAD).

[0141] like Figure 2 As shown, only 44.3% of the samples had 5 or more driver genes, while 55.7% had fewer than 5. This indicates that in more than half of the samples, the number of typical driver genes was less than 5. Therefore, driver gene prediction analysis for these three cancers (ESCA, GBM, PAAD) is necessary.

[0142] Of the 503 samples across three cancer types (ESCA, GBM, PAAD), those with usable cancer and mutation data were retained, including 183 ESCA samples, 150 GBM samples, and 170 PAAD samples. In addition, the protein-protein interaction network data required for this invention was obtained by integrating four public databases: BioGRID, MintAct, DIP, and HPRD. Integrating these public databases yielded a protein-protein interaction network dataset containing 8788 nodes and 273930 edges.

[0143] Known driver genes specific to three cancer types (ESCA, GBM, PAAD) were identified in the DriverDB database: 129 for ESCA, 235 for GBM, and 139 for PAAD. These known driver genes were used in constructing the driver gene-sample association matrix.

[0144] In addition, a list of cancer driver genes was downloaded from the Cancer Gene Network (NCG 6.0) database as a standard set of known driver genes to validate the experimental results. NCG is a database of tumor driver genes, which contains 2,380 driver genes, including 719 known driver genes and 1,661 candidate driver genes.

[0145] The following section describes the performance evaluation of the personalized cancer driver gene prediction model and method based on graph convolutional heterogeneous networks (GCN) of this invention. The personalized cancer driver gene prediction model based on graph convolutional heterogeneous networks of this invention is abbreviated as GCSDG.

[0146] To evaluate the performance of the model, the method of this invention using the GCSDG model was compared with three other methods: Dawnrank, SCS, and IMCDriver.

[0147] In the experiments, 5-fold cross-validation was used to evaluate the performance of each prediction method. All known gene-sample associations were randomly divided into 5 equally sized subsets. The cross-validation process was repeated 5 times, with each subset used as the test set and the remaining 4 subsets used as the training set. In each fold, a prediction model was built based on the known associations in the training set and used to predict associations in the test set. The performance of the model was evaluated by calculating threshold-based metrics, namely precision, recall, and F1 score.

[0148] The hyperparameters involved in the GCSDG model of this invention are set as follows: embedding dimension k=64, number of layers L is set to 3, initial learning rate lr=0.001, number of training iterations epoch=2000, and penalty factor. Set it to 8.

[0149] The results were first analyzed at the group level.

[0150] After the model makes a prediction, the score of the mutated gene in each sample can be obtained. Then, by sorting the scores of the mutated genes in each patient, the priority of the mutated genes in each patient can be obtained. Specifically, the top 50 genes in the prediction ranking of three types of cancer are selected, and the 719 known driver genes in NCG are used as the standard set to calculate and compare three indicators: Precision, Recall and F1-score.

[0151] The results are as follows Figure 3 As shown in the figure, the GCSDG method outperforms other methods in terms of indicators across the three cancer types. This indicates that the method has good performance in population-level prediction.

[0152] Secondly, the results were analyzed at the individual level.

[0153] In this invention, the Mean Reciprocal Rank (MRR), a commonly used evaluation metric in information retrieval, will be used to evaluate the effectiveness of each method at the individual level. MRR is a commonly used evaluation metric in recommender systems; a higher value indicates a better model performance, and it is generally used to measure the rationality of ranking. Here, it can be used to evaluate the effectiveness of genes with higher rankings. The MRR calculation formula is:

[0154]

[0155] in, The value is related to the ordering position of the correctly predicted gene among the candidate genes. This refers to the number of driver genes in each sample. Specifically, first, the genes in each sample are ranked overall based on their predicted scores. Then, the intersection of the mutated genes in each sample with the known driver genes of NCG is taken to obtain a subset of genes for each sample. Finally, these subsets are sorted according to their predicted scores, and the average of the results for all samples is used to obtain the MRR value.

[0156] The results are shown in Table 1. It can be seen that the MRR value of the GCSDG model is higher than that of the previous methods, which indicates that the method has better performance in predicting personalized driver genes in samples, and the predicted genes are more likely to be driver genes in the samples.

[0157] Table 1 Comparison of Personalized Driver Gene Data by Method

[0158] DawnRank 4.67114581 6.75455823 6.89875963 SCS 2.25648621 3.55651414 2.56887106 IMCDriver 8.51520950 12.39002121 11.77437336 GCSDG 9.85501346 14.00136522 13.17324820

[0159] Furthermore, the distance distribution between gene pairs was analyzed, utilizing network topology information to analyze and predict driver genes. The distances between driver genes and mutated genes in each cancer dataset were calculated using the String database, and violin plots were then generated using the R package ggpubr. Figure 4 As shown in the figure, the distances between predicted driver genes are significantly shorter than those between mutated genes, which is consistent with existing research findings. Furthermore, since the distance between genes can be considered a mapping of gene functional similarity, shorter distances between predicted driver genes indicate higher functional similarity. Therefore, this experiment demonstrates that similarity information helps identify driver genes that are functionally similar to each other.

[0160] The data comparison above shows that, compared with existing methods, the method proposed in this invention can improve the recognition performance of cancer driver genes in individuals, proving that the use of graph convolution technology on similar heterogeneous networks can improve the accuracy of identifying personalized driver genes in samples.

[0161] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0162] It should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention.

[0163] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0164] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection, an electrical connection, or a connection that allows communication between them; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0165] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "over," and "on top" of the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

Claims

1. A personalized cancer driver gene prediction method based on graph convolution of heterogeneous networks, characterized in that, Including the following steps: S1. Preprocess cancer-related data to construct mutation-sample association matrices and expression-sample association matrices; S2. Perform the Gaussian kernel similarity algorithm on the expression-sample association matrix to obtain the gene similarity matrix and the sample similarity matrix; S3. Process the mutation-sample association matrix using the public database DriverDB to obtain the driver gene-sample association matrix; S4. The gene similarity matrix, sample similarity matrix, and driver gene-sample association matrix are fused into a heterogeneous network; S5. Gene embedding and sample embedding of heterogeneous networks based on graph convolutional GCN encoder; S6. Use the decoder to reconstruct and obtain new gene-sample associations; In step S2, the Gaussian kernel similarity algorithm is performed on the expression-sample association matrix to obtain the gene similarity matrix and the sample similarity matrix; specifically, it includes: S21. Calculate the similarity score between genes in the expression-sample association matrix, expressed as: in, and To express the sample correlation matrix The genes in Indicates gene Expression spectrum, A parameter that controls kernel bandwidth: S22. Calculate the similarity score between samples in the expression-sample association matrix, expressed as: in, and To express the sample correlation matrix The samples in Indicates sample Expression spectrum, Indicates sample Expression spectrum, A parameter that controls kernel bandwidth: S23. Score based on similarity between genes Obtain the gene similarity matrix The similarity score between samples Obtain the sample similarity matrix ; In step S4, the gene similarity matrix, sample similarity matrix, and driver gene-sample association matrix are fused into a heterogeneous network. , represented as: in, The driver gene-sample association matrix is ​​represented as a binary matrix. , for The transpose of the matrix; for Standardized matrix; for Standardized matrix; = = in, =diag( )yes The degree matrix; =diag( )yes The degree matrix.

2. The method for personalized cancer driver gene prediction based on graph convolution of heterogeneous networks according to claim 1, characterized in that, The cancer-related data in S1 includes cancer expression data, mutation data, and protein-protein interaction network data.

3. The method for personalized cancer driver gene prediction based on graph convolution of heterogeneous networks according to claim 1, characterized in that, The preprocessing includes obtaining a z-score based on gene expression values, then removing outlier genes from the protein-protein interaction network data based on the score, while retaining mutated genes. Specifically: Where x is the original expression value of the sample gene, It is the mean of gene expression values ​​in the sample. is the standard deviation of the sample gene expression values, and z is the score of the sample gene expression values; The criterion for determining outlier gene deletion is |z|≥2.

4. The method for personalized cancer driver gene prediction based on graph convolution of heterogeneous networks according to claim 1, characterized in that, The construction of the mutation-sample association matrix and the expression-sample association matrix in S1 is as follows: in, The mutation-sample association matrix, This is an expression-sample association matrix; rows represent mutated genes. This represents the number of genes selected; the columns represent different samples. The number of samples.

5. The method for personalized cancer driver gene prediction based on graph convolution of heterogeneous networks according to claim 1, characterized in that, In step S3, the mutation-sample association matrix is ​​processed using the public database DriverDB to obtain the driver gene-sample association matrix, which specifically includes: S31, in the mutation-sample correlation matrix The DriverDB database is used for filtering. S32, If the mutation-sample association matrix Zhonggen In the sample If a mutation has occurred and it is a known driver gene present in the DriverDB cancer database, then it is... Reserve it as 1, otherwise Set to 0; S33. Obtain the driver gene-sample association matrix. , represented as: Among them, rows represent mutated genes. This represents the number of genes selected; the columns represent samples. The number of samples.

6. The method for personalized cancer driver gene prediction based on graph convolution of heterogeneous networks according to claim 1, characterized in that, The encoder based on graph convolutional neural networks (GCN) in S5 performs gene embedding and sample embedding on heterogeneous networks, specifically including: S51, In heterogeneous networks The graph convolutional GCN is used for L iterations of embedding, which is represented as: in, The introduced penalty factor, ; It is the first Embedding of layer nodes D=diag( )yes The degree matrix, W ( ) is a trainable weight matrix for a specific layer. ( ) is a non-linear activation function; S52, Based on heterogeneous networks For each layer of embedding, the final embedding of the gene and sample is obtained, represented as: = in, It is the final embedding of genes. It is the final embedding of the sample. For automatic learning of neural networks, It is the dimension of the embedding, initialized to .

7. The method for personalized cancer driver gene prediction based on graph convolution of heterogeneous networks according to claim 1, characterized in that, The decoder in S6 is a bilinear decoder. New gene-sample associations are obtained by reconstructing the data using the decoder, represented as follows: in, It is a non-linear activation function. Given a trainable weight matrix, This is the final embedding matrix of the genes. is the transpose of the final embedding matrix of the sample.

8. A graph convolutional personalized cancer driver gene prediction model based on heterogeneous networks, characterized in that, The method applied to any one of claims 1 to 7 includes: The cancer-related data unit acquires cancer expression data, mutation data, and protein-protein interaction network data, and performs preprocessing. The association matrix unit uses the mutation-sample association matrix and the expression-sample association matrix to associate preprocessed data and obtain mutation-sample association and expression-sample association; The similarity association unit utilizes the driver gene-sample association matrix, gene similarity matrix, and sample similarity matrix to perform driver gene-sample association, gene similarity, and sample similarity analysis on mutation-sample association data and expression-sample association data; Heterogeneous network units fuse data from driver gene-sample association, gene similarity, and sample similarity analysis to obtain heterogeneous networks; The encoding / decoding unit encodes embedded genes and samples in a heterogeneous network and decodes them to obtain the association between genes and samples.

Images