Cancer driver gene mining and interpretability analysis method based on heterogeneous networks

Abstract

The present invention discloses a method for cancer driver gene mining and interpretability analysis based on heterogeneous networks. The method comprises: constructing a multi-omics heterogeneous network and extracting initial node features through the multi-omics heterogeneous network; constructing an information transfer subgraph by random walk along a specific meta-path, and performing meta-path-based heterogeneous network feature extraction to obtain a representation vector of the gene node; inputting the representation vector of the gene node into a multi-layer linear classifier, using the multi-layer linear classifier to perform node classification, and outputting the analysis results of the cancer driver gene. The present invention calculates the representation vector within the meta-path through a network representation algorithm that includes multi-head attention and self-attention mechanisms, and calculates the contribution weight of each meta-path to the classification problem through global attention, thereby obtaining a more robust and effective cancer driver gene prediction result.

Description

Technical Field

[0001] The present invention relates to the technical field of bioinformatics, and in particular to a cancer driver gene mining and interpretability analysis method based on heterogeneous networks. Background Art

[0002] Using medical methods alone to identify cancer driver genes often requires processing large numbers of medical samples, consuming enormous amounts of time and research costs. Without computational tools, discovering a new cancer driver gene is a very lengthy process. However, with the development of bioinformatics and the establishment of high-throughput biological databases, computational biology approaches can accelerate this process. Currently, a variety of computational methods are available for identifying cancer driver genes from genomic data, including data mining methods based on network analysis and machine learning (ML). These methods can effectively perform pre-experimental analysis, accelerating the discovery of cancer driver genes by shortening the candidate gene set, saving time and improving efficiency. A new analytical method (MutSigCV) has been proposed to address this issue. This method identifies abnormal variations in mutation frequency and spectrum across cancer types, revealing mutational processes and disease etiology, as well as genome-wide mutation frequency, which is closely related to DNA replication timing and transcriptional activity. This method improves accuracy by pooling data from other genes with similar properties (e.g., replication timing, expression levels). The significance level (P value) is determined by testing whether the observed mutation counts in a gene significantly exceed the expected counts based on a background model. The false discovery rate (q-value) was then calculated, and genes with q≤0.1 were reported as significantly mutated. By incorporating mutation heterogeneity into the analysis, MutSigCV was able to eliminate most of the obvious unreasonable discoveries and identify genes that are truly associated with cancer.

[0003] The Cancer Genome Atlas (TCGA) is a program spearheaded by the U.S. National Cancer Institute (NCI). Established in 2006, it has been generating multimodal genomic, epigenomic, and proteomic data for thousands of tumor samples. With the advancement and development of high-throughput technologies, mutation data from various somatic cells has been continuously accumulated. TCGA has also established its own dedicated database and data channel website to facilitate communication and upload of new mutation data by researchers around the world. It also standardizes the file format of mutation data and provides software packages and various automatic search functions to lower the threshold for use. In 2018, TCGA summarized its work over the past 10 years and proposed a pan-cancer network database containing multi-omics data. It has also published a large number of pan-cancer studies. These studies have also standardized the expression of the cancer research field, ranging from specific cancer studies analyzing individual cancer genes and case data to pan-cancer studies analyzing case samples from multiple or even all cancers. This will undoubtedly greatly benefit the development of deep learning networks that have higher requirements for data volume and data format. Some studies have applied convolutional neural networks (CNNs) to the exploration of cancer driver genes. The machine learning model of simply connecting different features may not be the best way to integrate these features. Therefore, by allowing convolutional neural networks to simultaneously learn mutation data and similarity information in the network, better predictions of driver genes can be made. The authors varied the number of convolutional layers and fully connected layers, determined the optimal hyperparameters through grid search, and ultimately achieved an AUC score of close to 0.98 in breast cancer and colorectal cancer classification.

[0004] Research has published the EMOGI algorithm, based on GCN. It uses multi-omics data (including gene mutations, copy number variations, methylation, and expression information) as node features and trains them in a semi-supervised manner within protein interaction networks to learn complex nonlinear structures to identify oncogenes and non-oncogenes. The algorithm outputs a fully labeled graph containing the probability of each gene being an oncogene. In addition to building the algorithm model, perturbation experiments are conducted to determine the effect of the original multi-omics input data on the original features. Layer-wise relevance propagation (LRP) is also used to identify the omics features that contribute most to identifying each gene as a candidate cancer driver gene. Research has further expanded upon the GCN approach. First, structural features are introduced into protein-protein interaction (PPI) networks to enhance gene features. A multi-task learning framework then propagates and aggregates node and graph features from the input to the next layer to learn node embedding features. The framework simultaneously optimizes the node prediction and link prediction tasks, which share two Chebyshev GCN layers and optimize two different objective functions. Finally, a Bayesian task weight learner is used to automatically balance the two tasks. The output of the MTGCN assigns each gene a probability of being a cancer driver gene. The MODIG method constructs a multidimensional gene network, combining the various associations of genes into a multidimensional network. This also creates a heterogeneous network, where the edges have different attributes but only one type of node. This obscures the information of the intermediate nodes that drive the relationship between two gene nodes. While information about a gene node from different neighboring nodes can be obtained, the biological interpretability is significantly reduced.

[0005] There are still several problems in current research:

[0006] 1. In cancer, there is a high degree of heterogeneity in mutation data both within and outside of tumors, and data vary significantly between different cancers. In terms of data balance, the number of genes that can currently be clearly labeled as cancer drivers is relatively small compared to genes whose carcinogenic effects are uncertain or confirmed to be non-carcinogenic. Effectively sampling negative samples while ensuring the applicability of pan-cancer analysis remains a challenge.

[0007] 2. Currently, models using graph convolutional neural networks only construct homogeneous networks for analysis, or use multi-dimensional homogeneous networks without fully utilizing the heterogeneity of different biological omics data. Establishing heterogeneous networks with more complex and sufficient semantics can enhance the robustness of the network and achieve more accurate identification of cancer driver genes.

[0008] 3. Networks or graphs are ubiquitous in biology. Optimizing the representation extraction process of graph convolutional networks to make the model more biologically interpretable is also an area that can be further optimized.

[0009] Therefore, the existing technology needs to be improved. Summary of the Invention

[0010] The technical problem to be solved by the present invention is that, in response to the defects of the existing technology, the present invention provides a cancer driver gene mining and interpretability analysis method based on heterogeneous networks to solve the problems of low robustness and interpretability of existing cancer driver gene analysis methods.

[0011] The technical solutions adopted by the present invention to solve the technical problems are as follows:

[0012] In a first aspect, the present invention provides a method for cancer driver gene mining and interpretability analysis based on heterogeneous networks, comprising:

[0013] constructing a multi-omics heterogeneous network, and extracting initial node features through the multi-omics heterogeneous network;

[0014] By randomly walking along a specific meta-path, an information transfer subgraph is constructed, and meta-path-based heterogeneous network feature extraction is performed to obtain the representation vector of the gene node;

[0015] The representation vector of the gene node is input into a multi-layer linear classifier, the multi-layer linear classifier is used to perform node classification, and the analysis results of the cancer driver gene are output.

[0016] In one implementation, constructing a multi-omics heterogeneous network and extracting initial node features through the multi-omics heterogeneous network includes:

[0017] Based on the CPDB-PPI protein interaction data, the relationships between genes are represented by the relationships between proteins encoded by genes, forming a basic isomorphic network;

[0018] retaining edges in the homogeneous network with a probability of relationship greater than a preset value, selecting cancer-related multi-omics data from the MSigDB database as nodes, and using the relationships between the selected nodes and gene nodes as edges in the heterogeneous network to construct the multi-omics heterogeneous network;

[0019] Clinical data samples of genes from various cancer types are collected, and the mutation rate, methylation value, and gene expression data of each gene are calculated to form multi-dimensional initial input features.

[0020] In one implementation, constructing an information transfer subgraph by randomly walking along a specific meta-path and performing meta-path-based heterogeneous network feature extraction includes:

[0021] According to the defined meta-path, random walk sampling with limited step length based on neighbors is performed along the meta-path of the node to obtain the relationship between gene nodes extracted through the meta-path;

[0022] Obtaining the information transfer subgraph according to the relationships extracted between gene nodes through meta-paths;

[0023] According to the information transfer subgraph, the similarity between node i and the remaining nodes in the information transfer subgraph is calculated, and the contribution of the remaining nodes to node i is normalized using the softmax function to obtain the contribution weights of the remaining nodes in the message subgraph to node i, and the attention score in the aggregation process is obtained. The features of all the remaining nodes are weighted and summed according to the attention score to obtain the new representation feature of node i;

[0024] The representation results of the obtained nodes in a certain or a meta-path message subgraph are passed through a layer of nonlinear activation function and spliced ​​based on the multi-head attention mechanism. The results obtained by the parallel attention heads are used as the representation vector of a single node to obtain the final gene node representation vector matrix.

[0025] In one implementation, inputting the representation vector of the gene node into a multi-layer linear classifier and performing node classification using the multi-layer linear classifier includes:

[0026] A global attention mechanism is used to calculate different semantic weights for each meta-path;

[0027] Based on the calculated semantic weights, the final representation vectors of all nodes are obtained;

[0028] Based on the final representation vectors of all nodes, the probability of the node being classified as positive or negative is obtained in the form of a score. The score is converted into a probability distribution according to the log-softmax function, and the classification result of the cancer driver gene is output.

[0029] In a second aspect, the present invention provides a cancer driver gene mining and interpretability analysis device based on a heterogeneous network, comprising:

[0030] An initial feature extraction module, used to construct a multi-omics heterogeneous network and extract initial node features through the multi-omics heterogeneous network;

[0031] The vector representation module is used to construct an information transfer subgraph by randomly walking along a specific meta-path, and perform meta-path-based heterogeneous network feature extraction to obtain the representation vector of the gene node;

[0032] The multi-layer linear classification module is used to input the representation vector of the gene node into the multi-layer linear classifier, use the multi-layer linear classifier to perform node classification, and output the analysis results of the cancer driver gene.

[0033] In one implementation, the initial feature extraction module includes:

[0034] Isomorphic network unit, which is used to represent the relationship between genes based on CPDB-PPI protein interaction data and use the relationship between the proteins encoded by genes to form a basic isomorphic network;

[0035] a multi-omics heterogeneous network unit, configured to retain edges in the homogeneous network whose probability of existence of a relationship is greater than a preset value, select cancer-related multi-omics data from the MSigDB database as nodes, and use the relationships between the selected nodes and gene nodes as edges in the heterogeneous network to construct the multi-omics heterogeneous network;

[0036] The initial input feature unit is used to collect clinical data samples of genes of various cancer types, calculate the mutation rate, methylation value and gene expression data of each gene, and form multi-dimensional initial input features.

[0037] In one implementation, the vector representation module includes:

[0038] A walk sampling unit is used to perform random walk sampling with a limited step length based on neighbors along the meta-path of the node according to the defined meta-path, and obtain the relationship between gene nodes extracted through the meta-path;

[0039] An information transfer subgraph unit, configured to obtain the information transfer subgraph based on the relationships extracted between gene nodes through meta-paths;

[0040] an attention score calculation unit, configured to calculate the similarity between node i and the remaining nodes in the information transfer subgraph based on the information transfer subgraph, normalize the contributions of the remaining nodes to node i using a softmax function, obtain the contribution weights of the remaining nodes in the message subgraph to node i, obtain the attention score in the aggregation process, and weight the features of all the remaining nodes according to the attention score to obtain a new representation feature of node i;

[0041] The multi-head attention unit is used to pass the representation results of the obtained nodes in a certain type or a meta-path message subgraph through a layer of nonlinear activation function, and splice them based on the multi-head attention mechanism. The results obtained by the parallel attention heads are used as the representation vector of a single node to obtain the final gene node representation vector matrix.

[0042] In one implementation, the multi-layer linear classification module includes:

[0043] Semantic weight calculation unit, used to calculate different semantic weights for each meta-path using a global attention mechanism;

[0044] The final vector representation unit is used to obtain the final representation vectors of all nodes based on the calculated semantic weights;

[0045] The output unit is used to obtain the probability of the node being classified as positive or negative in the form of a score based on the final representation vector of all nodes. The score is converted into a probability distribution according to the log-softmax function, and the classification result of the cancer driver gene is output.

[0046] In a third aspect, the present invention provides a computer terminal comprising: a processor and a memory, wherein the memory stores a heterogeneous network-based cancer driver gene mining and explainability analysis program, and when the heterogeneous network-based cancer driver gene mining and explainability analysis program is executed by the processor, it is used to implement the operation of the heterogeneous network-based cancer driver gene mining and explainability analysis method as described in the first aspect.

[0047] In a fourth aspect, the present invention provides a computer-readable storage medium, which stores a cancer driver gene mining and explainability analysis program based on a heterogeneous network. When the cancer driver gene mining and explainability analysis program based on a heterogeneous network is executed by a processor, it is used to implement the operation of the cancer driver gene mining and explainability analysis method based on a heterogeneous network as described in the first aspect.

[0048] The present invention adopts the above technical solution to achieve the following effects:

[0049] The present invention builds nine meta-path heterogeneous networks and calculates the representation vectors within the meta-paths through a network characterization algorithm that includes multi-head attention and self-attention mechanisms. It can use global attention to calculate the contribution weight of each meta-path to the classification problem, thereby obtaining a more robust and effective cancer driver gene prediction result. The intermediate nodes of the meta-paths of the present invention have certain biological significance. In the process of interpreting the prediction results, the biological interpretability of the final results can be enhanced by outputting the types and contribution weights of different meta-path nodes. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. 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 the structures shown in these drawings without paying any creative work.

[0051] Figure 1 This is a flowchart of a method for cancer driver gene mining and interpretability analysis based on heterogeneous networks in one implementation of the present invention.

[0052] Figure 2 It is a schematic diagram of a meta-path in a network in one implementation of the present invention.

[0053] Figure 3 It is a schematic diagram of the overall process framework in one implementation of the present invention.

[0054] Figure 4 Schematic diagram of suspected cancer driver genes predicted in one implementation of the present invention.

[0055] Figure 5 It is a schematic diagram of a case analysis of genes in one implementation of the present invention.

[0056] Figure 6 It is a functional principle diagram of a terminal in one implementation of the present invention.

[0057] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solutions and advantages of the present invention more clear and distinct, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0059] Exemplary Methods

[0060] There are still several problems in current genetic research:

[0061] 1. In cancer, there is a high degree of heterogeneity in mutation data both within and outside of tumors, and data vary significantly between different cancers. In terms of data balance, the number of genes that can currently be clearly labeled as cancer drivers is relatively small compared to genes whose carcinogenic effects are uncertain or confirmed to be non-carcinogenic. Effectively sampling negative samples while ensuring the applicability of pan-cancer analysis remains a challenge.

[0062] 2. Currently, models using graph convolutional neural networks only construct homogeneous networks for analysis, or use multi-dimensional homogeneous networks without fully utilizing the heterogeneity of different biological omics data. Establishing heterogeneous networks with more complex and sufficient semantics can enhance the robustness of the network and achieve more accurate identification of cancer driver genes.

[0063] 3. Networks or graphs are ubiquitous in biology. Optimizing the representation extraction process of graph convolutional networks to make the model more biologically interpretable is also an area that can be further optimized.

[0064] In response to the above technical problems, an embodiment of the present invention provides a method for cancer driver gene mining and interpretability analysis based on heterogeneous networks. The method constructs a multi-omics heterogeneous network and extracts the initial features of nodes through the multi-omics heterogeneous network. The method constructs an information transfer subgraph by random walk along a specific meta-path and extracts heterogeneous network features based on the meta-path to obtain the representation vector of the gene node. The representation vector of the gene node is input into a multi-layer linear classifier, and the multi-layer linear classifier is used to classify the nodes and output the analysis results of the cancer driver gene. The embodiment of the present invention calculates the representation vector within the meta-path through a network representation algorithm that includes multi-head attention and self-attention mechanisms, and calculates the contribution weight of each meta-path to the classification problem through global attention. This method can obtain a more robust and effective cancer driver gene prediction result.

[0065] like Figure 1 As shown, an embodiment of the present invention provides a cancer driver gene mining and interpretability analysis method based on a heterogeneous network, comprising the following steps:

[0066] Step S100 : constructing a multi-omics heterogeneous network, and extracting initial node features through the multi-omics heterogeneous network.

[0067] In this example, a heterogeneous network containing multi-omics data is constructed to artificially assign connections between genes through different biological entities. These entities have unique properties and significance in the fields of medicine and biology. Through two aggregations within the meta-path and between each part, the weights of the paths generated by various connections for the identification of cancer driver genes are extracted. Therefore, the weights of the paths represented by different biological entities can be used to analyze and identify the causes of potential oncogenes, which is more biologically interpretable.

[0068] In this example, the proposed cancer driver gene prediction research based on the multi-omics heterogeneous network meta-path method (CDHGNN) mainly consists of three parts, namely, the multi-omics heterogeneous network, the aggregation of information within the meta-path and the aggregation of semantic weights between different meta-paths, and the cancer prediction module based on the multi-layer linear classifier. The overall process framework is as follows Figure 3 shown.

[0069] Specifically, the model in this embodiment is an end-to-end model. The initial input is the multi-omics heterogeneous network nodes and the multi-omics initial features of each gene (the initial multi-omics features are an n*x matrix, where n is the number of gene nodes, and x is obtained by calculating the gene expression values, gene mutation rates and other multi-omics data in the clinical sample data. The initial multi-omics features and gene nodes have a one-to-one correspondence, and the calculation method and length of x can be flexibly selected).

[0070] The first part of the meta-path-based factor aggregation is equivalent to multiple parallel GATs (graph attention neural networks), which are used to learn the contribution weights of each neighboring node within a specific meta-path for a gene node. The second part uses a global attention mechanism to output the contribution of various meta-paths to the final prediction result under global attention. Finally, a multi-layer linear perceptron outputs a two-dimensional vector as the classification result. After logarithmization and normalization, the classification results are calculated using cross-entropy loss. In the case analysis, the two weights of the meta-path aggregation module are output to explore the interpretability of the model.

[0071] Specifically, in one implementation of this embodiment, step S100 includes the following steps:

[0072] Step S101, based on the CPDB-PPI protein interaction data, the relationship between genes is represented by the relationship between the proteins encoded by the genes, forming a basic isomorphic network;

[0073] Step S102, retaining edges in the homogeneous network with a probability of relationship greater than a preset value, selecting cancer-related multi-omics data from the MSigDB database as nodes, and using the relationships between the selected nodes and gene nodes as edges in the heterogeneous network to construct the multi-omics heterogeneous network;

[0074] Step S103 , collecting clinical data samples of genes of various cancer types, calculating the mutation rate, methylation value and gene expression data of each gene, and forming multi-dimensional initial input features.

[0075] In this example, we used the same approach as the EMOGI method to extract multi-omics features. We used clinical data samples collected from TCGA, covering over 8,000 cases and 16 cancer types, to calculate the mutation rate, methylation value, and gene expression data for each gene, forming a 48-dimensional initial input feature.

[0076] In this embodiment, the CPDB-PPI protein interaction data downloaded from the CPDB database is used, and the relationship between the proteins encoded by the genes is used to represent the relationship between genes to form a basic isomorphic network; in this embodiment, only the edges with a probability of existence of a relationship greater than 0.5 (i.e., the preset value) in the ppi (protein-protein interaction) network are retained. In order to better utilize the rich biological prior knowledge and increase the interpretability of the model, various multi-omics data related to cancer in the MSigDB database are used as nodes, and a heterogeneous network is established together with the gene nodes; in this embodiment, the relationship between these newly added nodes and genes is considered, and they are used as edges in the heterogeneous network.

[0077] In this embodiment, for all nine types of nodes, only the parts that intersect with the genes in the CPDB network are taken to construct a heterogeneous network. Among them, the CPDB network is a protein interaction network, which is an important network that expresses the interaction between genes. It is a public database published to the international network database after biological determination by laboratories in various countries. The CPDB network is a basic network in this embodiment. The initial gene numbers and vectors in this embodiment are extracted according to the number of genes in the CPDB network. Although there will be some additional genes and other nodes (such as pathways) in the new MSigDB, they are not considered in this embodiment. Only the 12,944 gene nodes appearing in the CPDB are considered.

[0078] As an example, the process of building a heterogeneous network is as follows:

[0079] Both ends are gene nodes. Taking gene nodes AB as an example, they are both connected to GO1 in the MsigDB database, so a connection A-GO1-B is generated in the heterogeneous graph, and so on.

[0080] In this embodiment, in the process of constructing a heterogeneous network, 9 meta-paths are manually defined. The meta-path method can make full use of different types of nodes and edges in the heterogeneous network to form a new network description method, thereby better and more comprehensively describing the interaction between genes. Selecting an appropriate meta-path can better reflect the connection between genes. At the same time, the meta-path has certain biological significance and can take into account the predictive performance and interpretability of the model. The 9 meta-paths defined in this embodiment are as follows: Figure 2 shown.

[0081] like Figure 1 As shown, in one implementation of the embodiment of the present invention, the cancer driver gene mining and interpretability analysis method based on heterogeneous networks further includes the following steps:

[0082] Step S200 : constructing an information transfer subgraph by randomly walking along a specific meta-path, and performing meta-path-based heterogeneous network feature extraction to obtain a representation vector of a gene node.

[0083] Specifically, in one implementation of this embodiment, step S200 includes the following steps:

[0084] Step S201: According to the defined meta-path, random walk sampling with a limited step length based on neighbors is performed along the meta-path of the node to obtain the relationship between gene nodes extracted through the meta-path;

[0085] Step S202, obtaining the information transfer subgraph according to the relationships between gene nodes extracted through meta-paths;

[0086] Step S203: Calculate the similarity between node i and the remaining nodes in the message transfer subgraph based on the message transfer subgraph, normalize the contributions of the remaining nodes to node i using the softmax function, obtain the contribution weights of the remaining nodes in the message subgraph to node i, obtain the attention score in the aggregation process, and sum the features of all the remaining nodes according to the weighted attention score to obtain the new representation feature of node i;

[0087] In step S204, the representation results of the obtained nodes in a certain type or a meta-path message subgraph are passed through a layer of nonlinear activation function and spliced ​​based on the multi-head attention mechanism. The results obtained by the parallel attention heads are used as the representation vector of a single node to obtain the final gene node representation vector matrix.

[0088] In this embodiment, within a given meta-path Φ, random walk sampling with a limited step size based on neighbors is performed along the meta-path of the nodes. For example, if the specified path is Gene-CTM-Gene, then starting from each Gene node, the next step is to select the CTM path connected to it. However, not all connected CTMs are walked, but rather they are selected according to a certain probability. The next step is to select Gene nodes according to a certain probability, so that the final MFG of the entire walk is actually a message passing subgraph with Gene nodes at both ends and additional CTM information on the connecting edges in the middle. The random walk can make the divergence degree of each node in the subgraph more even.

[0089] Random walk sampling can balance the degree of nodes in the graph, improving its simplicity and accuracy. This sampling method extracts the relationships between gene nodes through meta-paths, which are denoted as message-passing subgraphs (MFGs). In this message-passing subgraph, all nodes of the same type and feature dimension are connected by intermediate nodes in the meta-paths. The process of generating MGFs refers to the design of the neighbor node sampler in the DGL library. Within the message-passing subgraph, an inter-node attention mechanism is used to calculate the attention distribution between nodes.

[0090] For gene node i, first calculate the vector h of node i. j Perform a linear transformation to get h' i =W Φ h j , here we use the attention mechanism to calculate the representation vector of node i in the information transfer subgraph, specifically:

[0091] First, calculate the similarity between node i and all other nodes in the information transfer subgraph Then, the softmax method is used to normalize the contribution of node j to node i, and the attention score α can be obtained.ij The overall process is shown in formula (1). In this process, the weight of node i, i.e., the self-attention mechanism, is taken into account. Nodes within the message subgraph are likely to establish self-connections through meta-paths, and their biological significance is determined by the type of meta-path they belong to and the intermediate nodes they pass through. The meaning of is the set of all neighbor nodes adjacent to node i in the Φ-type message passing subgraph.

[0092]

[0093] The final representation of the node in the meta-path message subgraph Φ needs to go through a layer of nonlinear activation function, as shown in formula (2).

[0094]

[0095] In order to improve the expressiveness and generalization ability of the model, a multi-head attention mechanism is added in this embodiment, and the above process is repeated multiple times to obtain ...etc., and the results obtained by splicing these parallel attention heads are used as the representation vector of a single node (the multi-head attention mechanism is that the same attention aggregation method is repeated several times, and then the results are averaged and spliced ​​to make the results smoother. It can be considered that a new dimension is added to the final result vector. N is the number of nodes, out_dims is the output dimension of the set attention layer, and attention_heads is the number of times the attention output is repeated. In this embodiment, these outputs are directly spliced ​​together by adding a new dimension, which is equivalent to superimposing several original two-dimensional papers in three-dimensional space as output, so as to make the result smoother). In this embodiment, the number of attention heads aggregated within the meta-path will be discussed in the subsequent hyperparameter selection. Finally, the representation vectors of all nodes are spliced ​​together to obtain a final output Z of N*out_dims*attention_heads. Φ .

[0096] like Figure 1 As shown, in one implementation of the embodiment of the present invention, the cancer driver gene mining and interpretability analysis method based on heterogeneous networks further includes the following steps:

[0097] Step S300: Input the representation vector of the gene node into a multi-layer linear classifier, use the multi-layer linear classifier to perform node classification, and output the analysis results of the cancer driver gene.

[0098] Specifically, in one implementation of this embodiment, step S300 includes the following steps:

[0099] In step S301, a global attention mechanism is used to calculate different semantic weights for each meta-path.

[0100] In this embodiment, a global attention mechanism is used to calculate different semantic weights for each meta-path, including: expanding the outputs of all nodes corresponding to the M meta-paths into a two-dimensional matrix and splicing them into a matrix; calculating the semantic weights between nodes through a two-layer MLP, and simulating the similarity comparison vector of the simple attention layer through a new linear layer. The obtained results are normalized through a softmax layer to obtain the semantic weights between different meta-paths.

[0101] In this embodiment, the information transfer subgraph and meta-path internal aggregation are used to finally obtain the feature vector representation of the gene node under each meta-path relationship. In this embodiment, we consider how to fully integrate these features while giving full play to the diversity of different meta-paths. Here, a global Attention mechanism is adopted to calculate different semantic weights for each meta-path so that the importance of different meta-paths can be differentiated, that is, to calculate (β Φ0 ,β Φ1 ...β ΦM ), specifically:

[0102] First, the outputs of all nodes corresponding to the M meta-paths obtained in the previous step are expanded into a two-dimensional matrix, namely (N·H), where H = out-dims·attention-heads. Then, they are concatenated into a matrix Z = M·N·H. A two-layer MLP is used to calculate the semantic weights between nodes. (MLP is a multi-layer vector perceptron. Its function is to change the x dimension of the N*x vector. Here, because we want to obtain the contribution weights of the M meta-paths, we need to transform M*N*H into an M*1 weight vector through two layers of transformation. In this case, the M values ​​represent the contribution coefficient of each meta-path.)

[0103]

[0104] Where W1 is the first layer weight of MLP, which is used to extract The nonlinear features of , and then a new linear layer W2 is used to simulate the similarity comparison vector q of the simple attention layer T Finally, the results are normalized through a softmax layer to obtain the semantic weights between different meta-paths. As shown in Formula 4.

[0105]

[0106] Specifically, in one implementation of this embodiment, step S300 further includes the following steps:

[0107] Step S302: Based on the calculated semantic weights, the final representation vectors of all nodes are obtained.

[0108] In step S303 , based on the final representation vectors of all nodes, the probability of the node being classified as positive or negative is obtained in the form of a score. The score is converted into a probability distribution using the log-softmax function, and the classification result of the cancer driver gene is output.

[0109] In this example, β·Z is used to obtain the final representation vector X = N·H for all nodes. This is then passed through a linear prediction layer consisting of two MLPs to obtain the node classification results. In this example, the probability of a node being classified as positive or negative is obtained as a score. The score is converted into a distribution using the log-softmax function, and the final loss function is calculated using the cross-entropy loss. This is shown in Equations 5 and 6.

[0110]

[0111]

[0112] In this example, all training set samples were used for training, and then the scores of all gene nodes predicted as positive were output. Samples with scores greater than 0.8 were taken as the prediction results in this example. A total of 602 suspected cancer driver genes were obtained, such as Figure 4 As shown. Figure 4 Analysis shows that among the 602 suspected cancer driver genes, 323 genes are included in the positive sample set, namely the NCG7.0 database, including 156 known cancer driver genes (KCG) of NCG and 160 candidate cancer driver genes (CG) of NCG. The remaining 254 genes were further analyzed, of which 48 were cancer driver genes mentioned in the Oncokb database and Ongene database and in this TCGA pan-cancer analysis paper. The Cancer Miner data mining tool can be used to further mine potential cancer genes. Among them, there are 153 cancer driver genes that are considered to be highly reliable in Cancer Miner, and there are 36 unfiltered genes that are believed to have some studies proving that these genes are related to cancer but there is no direct evidence. In other words, 91.86% of the CDHGNN predicted genes have at least one piece of evidence proving that they are cancer driver genes or potential cancer driver genes (cancer-related genes). The remaining 49 genes are possible cancer driver genes predicted in this example that have not yet been proven by research and are yet to be confirmed, accounting for approximately 8.14%.

[0113] This example outputs the contribution weight matrix α and the semantic-level attention weight matrix β within the meta-path, and multiplies the two together to obtain the contribution of each node in each meta-path subgraph to potential cancer driver genes. Then, the top 10 contributing gene nodes and the meta-paths to which they contribute are selected. By studying the biological significance of the intermediate nodes in these meta-paths and the functional similarity of the gene nodes at both ends of the meta-paths, an attempt is made to infer the biological significance of the model's classification of the gene as a cancer driver gene.

[0114] This example uses the SETD2 gene prediction case to illustrate how CDHGNN demonstrates biological interpretability when predicting cancer driver genes:

[0115] like Figure 5 As shown by Figure 5 It can be concluded that Figure 5 The three meta-paths that appear in the figure are also the three meta-paths with the highest weights calculated in this example. The gene-to-gene relationship is the PPi network from the CPDB database, while the Pathways nodes in the Gene-Pathways-Gene meta-path are from the KEGG database and the Reactome Pathway database. The CM nodes in the Gene-CM-Gene meta-path come from a study of tumor expression data modules based on DNA microarrays. This study comprehensively analyzed microarrays from 22 tumor types and described the expression profiles of different tumors based on the module's behavior. Specifically, taking the SETD2 gene as an example, a known cancer driver gene from the National Cancer Institute (NCG), the contribution weight matrix α and the semantic-level attention weight matrix β within the meta-path are output and multiplied to obtain the contribution of each node in each meta-path subgraph to the SETD2 gene. Then, in this example, the top 10 contributing gene nodes and the meta-paths they contribute to are selected. The contribution size is indicated by thickness in the figure above. The specific mechanism of action is shown in Table 1.

[0116] Table 1. Top 10 heterogeneous graph nodes and paths that contribute to SETD2 being a cancer driver gene

[0117]

[0118] Analysis of the above table shows that the predicted non-positive gene ZBTB17 in this example has the strongest correlation with the cancer driver gene PTEN. Both genes belong to the WP_SMALL_CELL_LUNG_CANCER pathway, indicating their association with small cell lung cancer (SCLC). The PTEN protein primarily functions by regulating the PI3K / Akt signaling pathway. When PTEN function is impaired, overactivation of the PI3K / Akt signaling pathway can lead to abnormal cell proliferation and survival, thereby promoting tumorigenesis. It is inferred that ZBTB17 is also involved in the regulation of cell signaling pathways. Studies have shown that ZBTB17 is associated with tumor development and progression. It can interact with the transcription factor Myc and influence the expression of Myc target genes. Myc is an important cancer-related gene, involved in tumor cell proliferation, survival, and metastasis. The importance of Myc and the PI3K / AKT signaling pathway in cancer is widely recognized. Overexpression of Myc and abnormal activation of the PI3K / AKT signaling pathway are closely associated with the development and progression of various tumors. Together, they can promote tumor cell proliferation, survival, and invasion, leading to tumor progression.

[0119] Therefore, there is reason to believe that PTEN is a cancer driver gene that works together or interacts with the ZBTB17 gene. The model in this embodiment can provide a basis for judging that the gene is a cancer driver gene through the neighbors of a specific gene in a heterogeneous network, and can preliminarily analyze the similarity of the effects between different genes or the relationship of mutual influence based on the function of this neighbor gene and the meta-path they are connected to. By sorting different meta-paths according to the weight ratio, it can help lock in the meta-path (network node) that has a greater impact, thereby improving the interpretability of the model.

[0120] In summary, the dataset in this embodiment considers a broader range of factors, and the generation of positive and negative samples relies solely on literature data mining, without incorporating additional prior knowledge. This results in better prediction of potential cancer driver genes. It can also be seen that heterogeneous network models are generally superior to homogeneous network models, and heterogeneous network models offer greater potential for interpretability. Compared to existing technologies, the method in this embodiment uses a self-built heterogeneous network of nine meta-paths. A network representation algorithm that incorporates multi-head attention and self-attention mechanisms is used to calculate the internal representation vectors of the meta-paths. The contribution weight of each meta-path to the classification problem is calculated using global attention, resulting in a more robust and effective cancer driver gene prediction result. Furthermore, because the intermediate nodes in the meta-paths have certain biological significance, the biological interpretability of the final result can be enhanced by outputting the types and contribution weights of the different meta-path nodes during the interpretation of the prediction results. This is not possible with methods that lack intermediate nodes in homogeneous or multi-dimensional homogeneous networks.

[0121] This embodiment achieves the following technical effects through the above technical solution:

[0122] This embodiment builds nine meta-path heterogeneous networks and calculates the representation vectors within the meta-paths through a network representation algorithm that includes multi-head attention and self-attention mechanisms. Global attention can be used to calculate the contribution weight of each meta-path to the classification problem, resulting in a more robust and effective cancer driver gene prediction result. The intermediate nodes of the meta-paths in this embodiment have certain biological significance. In the process of interpreting the prediction results, the biological interpretability of the final results can be enhanced by outputting the types and contribution weights of different meta-path nodes.

[0123] Exemplary devices

[0124] Based on the above embodiments, the present invention further provides a cancer driver gene mining and interpretability analysis device based on a heterogeneous network, comprising:

[0125] An initial feature extraction module, used to construct a multi-omics heterogeneous network and extract initial node features through the multi-omics heterogeneous network;

[0126] The vector representation module is used to construct an information transfer subgraph by randomly walking along a specific meta-path, and perform meta-path-based heterogeneous network feature extraction to obtain the representation vector of the gene node;

[0127] The multi-layer linear classification module is used to input the representation vector of the gene node into the multi-layer linear classifier, use the multi-layer linear classifier to perform node classification, and output the analysis results of the cancer driver gene.

[0128] In this embodiment, the initial feature extraction module includes:

[0129] Isomorphic network unit, which is used to represent the relationship between genes based on CPDB-PPI protein interaction data and use the relationship between the proteins encoded by genes to form a basic isomorphic network;

[0130] a multi-omics heterogeneous network unit, configured to retain edges in the homogeneous network whose probability of existence of a relationship is greater than a preset value, select cancer-related multi-omics data from the MSigDB database as nodes, and use the relationships between the selected nodes and gene nodes as edges in the heterogeneous network to construct the multi-omics heterogeneous network;

[0131] The initial input feature unit is used to collect clinical data samples of genes of various cancer types, calculate the mutation rate, methylation value and gene expression data of each gene, and form multi-dimensional initial input features.

[0132] In this embodiment, the vector representation module includes:

[0133] A walk sampling unit is used to perform random walk sampling with a limited step length based on neighbors along the meta-path of the node according to the defined meta-path, and obtain the relationship between gene nodes extracted through the meta-path;

[0134] An information transfer subgraph unit, configured to obtain the information transfer subgraph based on the relationships extracted between gene nodes through meta-paths;

[0135] an attention score calculation unit, configured to calculate the similarity between node i and the remaining nodes in the information transfer subgraph based on the information transfer subgraph, normalize the contributions of the remaining nodes to node i using a softmax function, obtain the contribution weights of the remaining nodes in the message subgraph to node i, obtain the attention score in the aggregation process, and weight the features of all the remaining nodes according to the attention score to obtain a new representation feature of node i;

[0136] The multi-head attention unit is used to pass the representation results of the obtained nodes in a certain type or a meta-path message subgraph through a layer of nonlinear activation function, and splice them based on the multi-head attention mechanism. The results obtained by the parallel attention heads are used as the representation vector of a single node to obtain the final gene node representation vector matrix.

[0137] In this embodiment, the multi-layer linear classification module includes:

[0138] Semantic weight calculation unit, used to calculate different semantic weights for each meta-path using a global attention mechanism;

[0139] The final vector representation unit is used to obtain the final representation vectors of all nodes based on the calculated semantic weights;

[0140] The output unit is used to obtain the probability of the node being classified as positive or negative in the form of a score based on the final representation vector of all nodes. The score is converted into a probability distribution according to the log-softmax function, and the classification result of the cancer driver gene is output.

[0141] This embodiment achieves the following technical effects through the above technical solution:

[0142] This embodiment builds nine meta-path heterogeneous networks and calculates the representation vectors within the meta-paths through a network representation algorithm that includes multi-head attention and self-attention mechanisms. Global attention can be used to calculate the contribution weight of each meta-path to the classification problem, resulting in a more robust and effective cancer driver gene prediction result. The intermediate nodes of the meta-paths in this embodiment have certain biological significance. In the process of interpreting the prediction results, the biological interpretability of the final results can be enhanced by outputting the types and contribution weights of different meta-path nodes.

[0143] Based on the above embodiments, the present invention also provides a computer terminal, comprising: a processor, a memory, an interface, a display screen, and a communication module connected via a system bus; wherein the processor is used to provide computing and control capabilities; the memory includes a storage medium and an internal memory; the storage medium stores an operating system and a computer program; the internal memory provides an environment for the operation of the operating system and computer program in the storage medium; the interface is used to connect to external devices, such as mobile terminals and computers; the display screen is used to display corresponding information; and the communication module is used to communicate with a cloud server or a mobile terminal.

[0144] When the computer program is executed by the processor, it is used to implement the operation of a cancer driver gene mining and interpretability analysis method based on a heterogeneous network.

[0145] It will be understood by those skilled in the art that Figure 6 The principle block diagram shown in the figure is only a block diagram of a partial structure related to the solution of the present invention and does not constitute a limitation on the computer terminal to which the solution of the present invention is applied. The specific computer terminal may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0146] In one embodiment, a computer terminal is provided, comprising: a processor and a memory, wherein the memory stores a program for cancer driver gene mining and interpretability analysis based on a heterogeneous network, and when the program is executed by the processor, it is used to implement the operations of the method for cancer driver gene mining and interpretability analysis based on a heterogeneous network.

[0147] In one embodiment, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores a heterogeneous network-based cancer driver gene mining and interpretability analysis program, and when the heterogeneous network-based cancer driver gene mining and interpretability analysis program is executed by the processor, it is used to implement the operations of the heterogeneous network-based cancer driver gene mining and interpretability analysis method as described above.

[0148] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware through a computer program. The computer program can be stored in a non-volatile storage medium. When executed, the computer program can include the processes in the above-described method embodiments. Any reference to memory, storage, database, or other media used in the various embodiments provided herein may include non-volatile and / or volatile memory.

[0149] In summary, the present invention provides a method for cancer driver gene mining and interpretability analysis based on heterogeneous networks, the method comprising: constructing a multi-omics heterogeneous network and extracting initial node features through the multi-omics heterogeneous network; constructing an information transfer subgraph by random walk along a specific meta-path, and performing meta-path-based heterogeneous network feature extraction to obtain a representation vector of the gene node; inputting the representation vector of the gene node into a multi-layer linear classifier, using the multi-layer linear classifier to perform node classification, and outputting the analysis results of the cancer driver gene. The present invention calculates the representation vector within the meta-path through a network representation algorithm that includes multi-head attention and self-attention mechanisms, and calculates the contribution weight of each meta-path to the classification problem through global attention, thereby obtaining a more robust and effective cancer driver gene prediction result.

[0150] It should be understood that the application of the present invention is not limited to the above examples. For those skilled in the art, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A cancer driver gene mining and interpretability analysis method based on heterogeneous networks, characterized by: include: constructing a multi-omics heterogeneous network, and extracting initial node features through the multi-omics heterogeneous network; By randomly walking along a specific meta-path, an information transfer subgraph is constructed, and meta-path-based heterogeneous network feature extraction is performed to obtain the representation vector of the gene node; Inputting the representation vector of the gene node into a multi-layer linear classifier, performing node classification using the multi-layer linear classifier, and outputting analysis results of the cancer driver gene; The method of constructing a multi-omics heterogeneous network and extracting initial node features through the multi-omics heterogeneous network includes: Based on the CPDB-PPI protein interaction data, the relationships between genes are represented by the relationships between proteins encoded by genes, forming a basic isomorphic network; retaining edges in the homogeneous network with a probability of relationship greater than a preset value, selecting cancer-related multi-omics data from the MSigDB database as nodes, and using the relationships between the selected nodes and gene nodes as edges in the heterogeneous network to construct the multi-omics heterogeneous network; Collect clinical data samples of genes from various cancer types, calculate the mutation rate, methylation value, and gene expression data of each gene, and obtain the initial features of the node; The information transfer subgraph is constructed by randomly walking along a specific meta-path, and the meta-path-based heterogeneous network feature extraction is performed, including: According to the defined meta-path, random walk sampling with limited step length based on neighbors is performed along the meta-path of the node to obtain the relationship between gene nodes extracted through the meta-path; Obtaining the information transfer subgraph according to the relationships extracted between gene nodes through meta-paths; According to the information transfer subgraph, the similarity between node i and the remaining nodes in the information transfer subgraph is calculated, and the contribution of the remaining nodes to node i is normalized using the softmax function to obtain the contribution weights of the remaining nodes in the information transfer subgraph to node i, and the attention score in the aggregation process is obtained. The features of all the remaining nodes are weighted and summed according to the attention score to obtain the new representation feature of node i; The obtained node representation results in a certain meta-path information transfer subgraph are passed through a layer of nonlinear activation function and spliced ​​based on a multi-head attention mechanism. The results obtained by the parallel attention heads are used as the representation vector of a single node to obtain the final gene node representation vector matrix; The step of inputting the representation vector of the gene node into a multi-layer linear classifier and performing node classification using the multi-layer linear classifier includes: A global attention mechanism is used to calculate different semantic weights for each meta-path; Based on the calculated semantic weights, the final representation vectors of all nodes are obtained; Based on the final representation vectors of all nodes, the probability of the node being classified as positive or negative is obtained in the form of a score. The score is converted into a probability distribution using the log-softmax function, and the classification results of cancer driver genes are output; The global attention mechanism is used to calculate different semantic weights for each meta-path, including: expanding the outputs of all nodes corresponding to the M meta-paths into a two-dimensional matrix and concatenating them into a matrix; calculating the semantic weights between nodes through a two-layer MLP, and simulating the similarity comparison vectors of a simple attention layer through a new linear layer. The obtained results are normalized through a softmax layer to obtain the semantic weights between different meta-paths.

2. A cancer driver gene mining and interpretability analysis device based on heterogeneous networks, characterized by: include: An initial feature extraction module, used to construct a multi-omics heterogeneous network and extract initial node features through the multi-omics heterogeneous network; The vector representation module is used to construct an information transfer subgraph by randomly walking along a specific meta-path, and perform meta-path-based heterogeneous network feature extraction to obtain the representation vector of the gene node; A multi-layer linear classification module is used to input the representation vector of the gene node into a multi-layer linear classifier, use the multi-layer linear classifier to perform node classification, and output the analysis results of the cancer driver gene; The initial feature extraction module includes: Isomorphic network unit, which is used to represent the relationship between genes based on CPDB-PPI protein interaction data and use the relationship between the proteins encoded by genes to form a basic isomorphic network; a multi-omics heterogeneous network unit, configured to retain edges in the homogeneous network whose probability of existence of a relationship is greater than a preset value, select cancer-related multi-omics data from the MSigDB database as nodes, and use the relationships between the selected nodes and gene nodes as edges in the heterogeneous network to construct the multi-omics heterogeneous network; An initial input feature unit is used to collect clinical data samples of genes of various cancer types, calculate the mutation rate, methylation value and gene expression data of each gene, and obtain the initial features of the node; The vector representation module includes: A walk sampling unit is used to perform random walk sampling with a limited step length based on neighbors along the meta-path of the node according to the defined meta-path, and obtain the relationship between gene nodes extracted through the meta-path; An information transfer subgraph unit, configured to obtain the information transfer subgraph based on the relationships extracted between gene nodes through meta-paths; an attention score calculation unit, configured to calculate the similarity between node i and the remaining nodes in the information transfer subgraph based on the information transfer subgraph, normalize the contributions of the remaining nodes to node i using a softmax function, obtain the contribution weights of the remaining nodes in the information transfer subgraph to node i, obtain the attention score in the aggregation process, and weightedly sum the features of all remaining nodes according to the attention score to obtain a new representation feature of node i; The multi-head attention unit is used to pass the obtained node representation results in a certain meta-path information transfer subgraph through a layer of nonlinear activation function, and splice them based on the multi-head attention mechanism. The results obtained by the parallel attention heads are used as the representation vector of a single node to obtain the final gene node representation vector matrix; The multi-layer linear classification module includes: The semantic weight calculation unit is used to calculate different semantic weights for each meta-path using a global attention mechanism. This includes: expanding the outputs of all nodes corresponding to the M meta-paths into a 2D matrix and concatenating them into a matrix; calculating the semantic weights between nodes through a two-layer MLP, and using a new linear layer to simulate the similarity comparison vector of the simple attention layer. The obtained results are normalized through a softmax layer to obtain the semantic weights between different meta-paths; The final vector representation unit is used to obtain the final representation vectors of all nodes based on the calculated semantic weights; The output unit is used to obtain the probability of the node being classified as positive or negative in the form of a score based on the final representation vector of all nodes. The score is converted into a probability distribution according to the log-softmax function, and the classification result of the cancer driver gene is output.

3. A computer terminal, characterized in that: include: A processor and a memory, wherein the memory stores a program for cancer driver gene mining and interpretability analysis based on a heterogeneous network, and when the program is executed by the processor, it is used to implement the operation of the method for cancer driver gene mining and interpretability analysis based on a heterogeneous network as claimed in claim 1.

4. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a heterogeneous network-based cancer driver gene mining and explainability analysis program, which, when executed by a processor, is used to implement the operation of the heterogeneous network-based cancer driver gene mining and explainability analysis method according to claim 1.

Images