CN118280453A - Cancer driving gene identification method based on heterogeneous map diffusion convolution network - Google Patents

Abstract

The present invention belongs to the field of bioinformatics and relates to a method for identifying cancer driver genes based on a heterogeneous graph diffusion convolutional network. Accurate identification of cancer driver genes is crucial for formulating effective treatment plans. The present invention improves the accurate identification performance of cancer driver genes through multi-layer convolution. First, a graph diffusion generation auxiliary network is introduced to perform data enhancement on the original graph data to better capture the correlation information in the biomolecular network. Secondly, a feature extraction module is constructed for feature extraction, which includes a fully connected layer and four graph attention convolutional layers. Finally, the extracted feature matrix is input into a hierarchical attention classifier to obtain the prediction score of the cancer driver gene. Experiments have shown that the present invention not only effectively identifies known driver genes on different networks, but also discovers new cancer candidate genes, providing strong support for precision medicine and personalized treatment, and further promoting the development of cancer research and treatment.

Description

A method for identifying cancer driver genes based on heterogeneous graph diffusion convolutional networks

Technical Field

The present invention belongs to the field of bioinformatics and relates to a method for identifying cancer driver genes based on a heterogeneous graph diffusion convolutional network.

Background technique

Cancer driver genes can give mutant cells a selective growth advantage, causing cell abnormalities, uncontrolled proliferation, and driving the occurrence and development of cancer. In the biomolecular network, cancer driver genes and their interactions form a complex signal transmission network. Processing and analyzing the information they transmit is directly related to the feature extraction module.

Accurate identification of cancer driver genes is of great significance for understanding the pathogenesis of cancer, promoting precision treatment, formulating personalized treatment plans and discovering biomarkers.

At present, most cancer driver gene identification methods are based on simple graph convolution layers to extract features of biomolecular networks, while a few methods use multiple graph convolution layers and graph sampling aggregation convolution layers, which cannot effectively extract features of biomolecular networks. Therefore, how to adapt to the heterogeneity of biomolecular networks and effectively capture long-distance gene features has become a major difficulty in this field.

Summary of the invention

In order to overcome the above difficulties, the present invention proposes a cancer driver gene identification method based on a heterogeneous graph diffusion convolutional network. The method effectively captures the gene features of biomolecular networks and auxiliary biomolecular networks based on the graph attention convolution layer, and the complex interrelationships of genes in multiple biomolecular networks are effectively learned to enrich the feature representation. The prediction score of cancer driver genes is obtained by learning the hidden features learned by the feature extraction module.

A cancer driver gene identification method based on heterogeneous graph diffusion convolutional network includes three steps: graph data preprocessing and enhancement, construction of feature extraction module and construction of multi-layer attention classifier. The specific steps are as follows:

Step 1: First, load the gene feature matrix and edge data of the biomolecular network of cancer driver genes; then, convert the gene feature matrix into a NumPy array, standardize these data, and then convert the NumPy array into a PyTorch tensor; use the random walk-based graph diffusion algorithm to enhance the original biomolecular network to obtain an auxiliary biomolecular network.

Step 2. During the forward propagation process, the gene feature matrix of the original biological molecular network is transformed using a fully connected layer; then, the matrix after feature transformation is convolved with the edge data of the original image and the auxiliary image respectively, and the output is subjected to feature fusion to obtain the matrix after the first feature extraction; then, the matrix after the first feature extraction is convolved with the edge data of the original image and the auxiliary image respectively, and the output is subjected to feature fusion to obtain the matrix after the second feature extraction.

Step 3: Pass the feature matrix after feature transformation in step 2 and the feature matrix after two feature extractions into the fully connected layer respectively, output a scalar, and then multiply the three scalars by a weight respectively and sum them up to obtain the final prediction score of the cancer driver gene.

A cancer driver gene identification method based on heterogeneous graph diffusion convolutional network. The implementation process of step 1 is as follows:

First, the biomolecular network graph data is loaded, including the feature matrix of the graph nodes, the labels of the graph nodes, and the edge data of the graph nodes; then the feature matrix of the graph nodes is standardized to obtain the gene feature matrix of the biomolecular network; then the gene feature matrix, labels and edge data are input into the random walk-based graph diffusion algorithm to obtain the edge data of the auxiliary network; specifically, the state vector of each node is first calculated, which reflects the importance of the node in the global graph structure; based on these state vectors, nodes with higher weights are screened out, and the auxiliary network is reconstructed based on these nodes. In this way, the generated auxiliary network provides a more accurate and global feature representation for subsequent data analysis and machine learning tasks.

A cancer driver gene identification method based on heterogeneous graph diffusion convolutional network. The implementation process of step 2 is as follows:

First, the biomolecular network gene feature matrix is input into a random dropout function with a dropout rate of 0.5, and its output is input into a fully connected layer with an input dimension of 58 and an output dimension of 80 and a ReLU activation function to obtain a feature transformation output. Then, the feature transformation output and two edge data are respectively input into a graph attention convolution layer with an input dimension of 80, an output dimension of 160, a number of attention heads of 2, and a dropout rate that remains unchanged. After activation by the ReLU function, feature fusion is performed to obtain the output of the first feature extraction module. Then, in the same way, the output of the first feature extraction module and two edge data are respectively input into a graph attention convolution layer with an input dimension of 160, an output dimension of 160, and a number of attention heads and a dropout rate that remain unchanged to obtain the output of the second feature extraction module.

A cancer driver gene identification method based on heterogeneous graph diffusion convolutional network, the implementation process of step 3 is as follows:

The feature transformation output obtained by feature extraction, the output of the first feature extraction module and the output of the second feature extraction module are respectively input into a fully connected layer, whose input dimensions are 80, 160 and 160 respectively, and the output dimension is 1. Finally, the outputs of the three fully connected layers are multiplied by a learnable parameter and summed to obtain the final prediction score of the cancer driver gene.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1 is a flow chart of a cancer driver gene identification method based on a heterogeneous graph diffusion convolutional network.

FIG2 is a flowchart of biomolecular network data preprocessing and enhancement.

FIG3 is a flow chart of the feature extraction module.

Figure 4 is a flowchart of the multi-layer attention classifier.

Detailed ways

The present invention is described in detail below with reference to the accompanying drawings and examples.

The present invention proposes a cancer driver gene identification method based on a heterogeneous graph diffusion convolutional network, in particular, for the identification of cancer driver genes.

A cancer driver gene identification method based on a heterogeneous graph diffusion convolutional network. FIG1 is a flow chart of a cancer driver gene identification method based on a heterogeneous graph diffusion convolutional network, which includes three steps: graph data preprocessing and enhancement, construction of a feature extraction module, and construction of a multi-layer attention classifier. The specific implementation method is as follows:

Step 1: Graph data preprocessing and enhancement. Figure 2 is a flowchart of graph data preprocessing and enhancement, which includes the following contents:

First, load the biomolecular network dataset, including the feature matrix of graph nodes, the labels of graph nodes, the edge data of graph nodes, etc., and convert the feature matrix of graph nodes into NumPy array. Then, use the preprocessing.StandardScaler function of the scikit-learn framework to standardize the feature matrix of graph nodes with a mean of 0 and a variance of 1, and convert the standardized feature matrix into a PyTorch tensor. Next, use the data.Data function of the PyTorch Geometric framework to construct a Data object, whose attributes include the gene feature matrix, labels and edge data of the dataset. Use the transforms.T.GDC function of the PyTorch Geometric framework to create a geometric data transformer for graph diffusion. Its diffusion method is ppr, the attenuation coefficient of the diffusion matrix is set to 0.9, and the regularization term is set to 0.0001. The application of geometric data transformers in image processing is mainly to change the position, size, shape and details of the image by performing geometric transformations and diffusion operations, thereby achieving various image processing effects. The diffusion method PPR is based on random walks and is used to measure the proximity of nodes in a graph. Specifically, PPR defines an iterative process in which each gene walks randomly along the edge with probability c and The probability of returning to itself. When there is no longer a significant update, the gene iteration converges. The final state vector Each element in represents the connection strength of the corresponding gene with other genes. This update process can be expressed by the following formula:

(1)

is an N -dimensional probability vector representing the state of the i -th gene (node), where N is the number of genes. is a one-hot vector, that is, only one element is 1 and the rest are 0, representing the i -th gene. c is a parameter, called the damping factor, which is used to control the probability of random walk. This equation means that in each iteration, the current state vector By transferring the M matrix update, and adding a bias vector Finally, the Data object is used as the input of the geometric data converter to obtain the Data object of the biomolecular network after edge data enhancement.

Step 2: Construction of feature extraction module. Figure 3 is a flow chart of the feature extraction module, which includes the following contents:

First, the utils.dropout_edge function of the PyTorch Geometric framework is used to randomly inactivate the edge data of the original image and the auxiliary image, with a dropout rate of 0.5. The nn.Functional.F.dropout function of the PyTorch framework is used to randomly inactivate the feature matrix, with a dropout rate of 0.5. Then, the nn.Linear function of the PyTorch framework is used to transform the feature matrix, with an input dimension of 58 and an output dimension of 80. Then, PyTorch is used to The nn.GATConv function of the Geometric framework constructs 4 convolutional layers to propagate and transform the input features. Its input dimension is 80, the output dimension is 160, the number of attention heads is 2, and the drop rate is 0.5; the first convolutional layer input is the original image edge data and the matrix after feature transformation, the second convolutional layer input is the auxiliary image edge data and the matrix after feature transformation, the outputs of these two convolutional layers are activated by the ReLU function and the features are fused to obtain the matrix after the first feature extraction; the third convolutional layer input is the original image edge data and the matrix after the first feature extraction, the fourth convolutional layer input is the auxiliary image edge data and the matrix after the first feature extraction, the outputs of these two convolutional layers are activated by the ReLU function and the features are fused to obtain the matrix after the second feature extraction;

Step 3: Construction of multi-layer attention classifier. Figure 4 is a flowchart of the multi-layer attention classifier, which includes the following contents:

First, the nn.Functional.F.dropout function of the PyTorch framework was used to construct three random dropout layers to randomly drop out the matrix after feature transformation, the matrix after the first feature extraction, and the matrix after the second feature extraction, with a dropout rate of 0.5; the nn.Linear function of the PyTorch framework was used to construct three fully connected layers, with input dimensions of 80, 160, and 160, respectively, and output dimensions of 1. The matrix after feature transformation, the matrix after the first feature extraction, and the matrix after the second feature extraction were input into the three fully connected layers in sequence, and finally the outputs of the three fully connected layers were multiplied by a learnable parameter and summed to obtain the final prediction score of the cancer driver gene.

When the method proposed in the present invention is applied to the identification of cancer driver genes, the AUCs tested on the three data sets provided by HDGC are 0.8512, 0.8646, and 0.8675, and the AUPRs are 0.8014, 0.8527, and 0.8313, respectively. It is better than the performance of HDGC and HDGICN on these three data sets, where the AUCs of HDGC are 0.8278, 0.8405, and 0.8337, and the AUPRs are 0.7770, 0.8286, and 0.7929, respectively. The AUCs of HDGICN are 0.8421, 0.8427, and 0.8418, and the AUPRs are 0.7936, 0.8405, and 0.8115, respectively. The present invention fully extracts and fuses the features of the original biomolecular network and the auxiliary network generated after data enhancement, so the performance is higher than other existing methods.

The optimal model parameters are shown in the following table.

Table 1 Optimal model parameters

The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims (2)

1. A cancer driver gene identification method based on a heterogeneous graph diffusion convolutional network, characterized in that an auxiliary network is generated by using a random walk-based graph diffusion to enhance the original biomolecular network data, a feature extraction module based on a graph attention convolutional layer is constructed to simultaneously extract the original biomolecular network data and the auxiliary biomolecular network data and perform feature fusion, and then the fused features are input into a multi-layer attention classifier to obtain the prediction score of the cancer driver gene, including three steps of graph data enhancement and preprocessing, construction of a feature extraction module and construction of a multi-layer attention classifier, and the specific steps are as follows: Step 1: First, load the gene feature matrix and edge data of the biomolecular network of cancer driver genes; then, convert the gene feature matrix into a NumPy array, standardize the data, and then convert the NumPy array into a PyTorch tensor; use the random walk-based graph diffusion algorithm to enhance the original biomolecular network to obtain the auxiliary biomolecular network; Step 2: During the forward propagation process, the gene feature matrix of the original biomolecular network is transformed using a fully connected layer; then, the matrix after feature transformation is subjected to graph attention convolution operation with the edge data of the original image and the auxiliary image respectively, and the output is subjected to feature fusion to obtain the matrix after the first feature extraction; then, the matrix after the first feature extraction is subjected to graph attention convolution operation with the edge data of the original image and the auxiliary image respectively, and the output is subjected to feature fusion to obtain the matrix after the second feature extraction; Step 3: Pass the feature matrix after feature transformation in step 2 and the feature matrix after two feature extractions into the fully connected layer respectively, output a scalar, and then multiply the three scalars by a weight respectively and sum them up to obtain the final prediction score of the cancer driver gene. 2. According to claim 1, a cancer driver gene identification method based on heterogeneous graph diffusion convolutional network is characterized in that the original biomolecular network data and the auxiliary network data generated by graph diffusion are used simultaneously to perform feature extraction and feature fusion through a feature extraction module based on a graph attention convolutional layer; the feature extraction steps are as follows: First, the utils.dropout_edge function of the PyTorch Geometric framework is used to randomly inactivate the edge data of the original image and the auxiliary image, with a dropout rate of 0.5. The nn.Functional.F.dropout function of the PyTorch framework is used to randomly inactivate the feature matrix, with a dropout rate of 0.5. Then, the nn.Linear function of the PyTorch framework is used to transform the feature matrix, with an input dimension of 58 and an output dimension of 80. Then, PyTorch is used to The nn.GATConv function of the Geometric framework constructs 4 convolutional layers to propagate and transform the input features. Its input dimension is 80, the output dimension is 160, the number of attention heads is 2, and the drop rate is 0.5; the input of the first convolutional layer is the original image edge data and the matrix after feature transformation, and the input of the second convolutional layer is the auxiliary image edge data and the matrix after feature transformation. The outputs of these two convolutional layers are activated by the ReLU function and the features are fused to obtain the matrix after the first feature extraction; the input of the third convolutional layer is the original image edge data and the matrix after the first feature extraction, and the input of the fourth convolutional layer is the auxiliary image edge data and the matrix after the first feature extraction. The outputs of these two convolutional layers are activated by the ReLU function and the features are fused to obtain the matrix after the second feature extraction.