A method and device for predicting gene interaction relationships based on graph neural network

Abstract

A method and device for predicting gene interaction relationships based on a graph neural network, the method comprising: preprocessing the transcriptional expression data of each experimental group, including standardizing the gene ID, and representing the gene transcriptional expression data of different experimental groups with a unified gene ID; obtaining raw data for subsequent analysis; calculating gene similarity according to each sample group, and obtaining similarity data between genes in the sample group; selecting gene pairs that meet set conditions as the basis for constructing a graph data structure by statistically analyzing and screening the data of all sample groups; further processing the raw data to ensure that all sample groups contain only genes defined in the graph structure; then, standardizing the data of these sample groups so that the data dimensions of all sample groups are consistent; and performing model training on the generated graph data group to ultimately obtain edge weight data in the graph structure, which can accurately reflect the interaction relationship between genes. The present invention can be used to analyze gene expression data, predict the interaction relationship between genes, and perform unsupervised learning in the absence of labeled data.

Description

Technical Field

[0001] The present invention relates to the field of graph neural networks (GNNs), and in particular to a method and device for predicting gene interaction relationships based on a graph attention network (GAT). Background Art

[0002] In bioinformatics research, analyzing gene expression data is an important approach for exploring gene functions and their interactions. Traditional gene similarity calculation methods are typically based on statistics and distance metrics, but these methods have limitations when dealing with complex gene relationships. In recent years, graph neural network technology has demonstrated significant advantages in processing graph-structured data, particularly in gene network analysis, effectively mining potential gene interactions.

[0003] Graph Attention Networks (GATs), a type of graph neural network, adaptively assign weights to adjacent nodes, making the network more flexible in processing node features. However, existing GAT models mostly rely on supervised learning, which often lacks sufficient labeled data in many genetic data analysis scenarios. Therefore, unsupervised GAT models have become an important development direction.

[0004] When processing gene expression data, the GAT model in the prior art often faces problems such as high computational complexity and difficulty in determining the direction of gradient descent. Therefore, a model architecture that can simplify the computational process and effectively process unlabeled data is needed. This paper proposes an unsupervised gene relationship prediction method based on a graph attention network. By simplifying the computation through ResNet and pooling operations, and combining the addition and deletion of edges, the accuracy and efficiency of gene correlation prediction are improved. Summary of the Invention

[0005] The present invention aims to overcome the technical problems of low accuracy and efficiency in gene correlation prediction in the prior art, and provide a gene interaction relationship prediction method and device based on graph neural network that is simple, easy to operate and can process complex gene expression data.

[0006] The method combines unsupervised learning techniques to effectively predict gene correlations by calculating similarities between gene expression data and predicting edges (adding and removing edges) on a graph structure. This technology is suitable for gene function analysis and gene expression data mining in the field of bioinformatics.

[0007] In order to achieve the above-mentioned object, the technical solution adopted by the present invention is as follows: a first aspect of the present invention provides a method for predicting gene interaction relationships based on a graph neural network, comprising the following steps:

[0008] a. Preprocess the transcriptional expression data for each experimental group, including standardizing gene IDs so that the gene transcriptional expression data from different experimental groups are represented by a unified gene ID to ensure consistent data format. Next, group the cell samples according to the experimental group, with each sample group containing multiple samples, to ultimately obtain the raw data for subsequent analysis;

[0009] b. Calculate gene similarity for each sample group based on the raw data above to obtain similarity data between genes within the sample group. By statistically analyzing and filtering the data from all sample groups, select gene pairs that meet the set criteria as the basis for constructing the graph data structure, ensuring that the selected gene pairs reflect actual biological significance.

[0010] c. Based on the graph structure obtained from the above screening, the raw data is further processed to ensure that all sample groups contain only the genes defined in the graph structure. Subsequently, the data of these sample groups are normalized to make the data dimensions consistent across all sample groups. Through this normalization process, multiple graph data sets with the same structure but different node characteristics (i.e., gene expression characteristics) are formed for subsequent graph neural network model training;

[0011] d. Perform model training on the generated graph dataset. The model used is a graph attention network (GAT) and a simple edge prediction algorithm. Through multiple iterative training of the model, the loss function is gradually optimized, and the edge weight data in the graph structure is finally obtained, which can accurately reflect the interaction relationship between genes.

[0012] Preferably, in step a, the specific operation method of the preprocessing process of transcription expression data is:

[0013] a1. Because different experimental groups use different sequencing technologies, resulting in inconsistent gene expression data formats, it is necessary to uniformly convert the data of all experimental groups into the same gene ID or other consistent standard data formats to ensure comparability of subsequent analyses;

[0014] a2. According to different experimental condition variables (such as cell type, experimental environment, etc.), each set of experimental data is grouped and processed. Each experimental data can be divided into multiple sample groups, and the condition variable can be cell type (such as normal cells, tumor cells, etc.);

[0015] a3. Organize the grouped data to form a set of sample groups corresponding to each cell type, ensuring that the gene expression data in each sample group can reflect the biological characteristics of the type.

[0016] Preferably, in step b, the specific operation methods of similarity calculation and data screening are as follows:

[0017] b1. For each cell type sample group, calculate the similarity between the gene pairs within the group. The data of each sample is regarded as the characteristics of the gene. Use appropriate similarity calculation methods (such as Pearson correlation coefficient, Euclidean distance, etc.). The selection of these methods should be determined based on the characteristics of the data (such as whether the data is normally distributed, whether the data type is integer or decimal, etc.) to calculate and obtain the similarity data between genes in the sample group.

[0018] b2. For each sample group, perform a preliminary screening of the similarity results based on a pre-set threshold, retaining gene pairs with similarity greater than or equal to the threshold. The selection of the threshold is critical and should be tested and adjusted to ensure that the number of gene pairs screened is appropriate, without losing key information or excessive redundancy.

[0019] b3. Perform statistical screening on each sample group within each cell type. Count all gene pairs across all sample groups within the same cell type and select those with a frequency greater than a set threshold. The choice of this threshold also significantly influences the formation of the graph structure. The gene pairs that are ultimately retained will form the edges of the graph data, the core of the graph structure.

[0020] Preferably, in step c, the specific operation method of the standardization processing and graph data generation is as follows:

[0021] c1. When screening gene pairs, some sample groups may not contain the selected genes, so these genes need to be supplemented. This can be done using a trained model or by directly filling in zero values ​​to ensure that each sample group contains the required gene pairs. Through this screening and supplementation process, a gene dataset containing all graph structure nodes is ultimately obtained.

[0022] c2. Standardize the gene expression data of each sample group to unify its dimensions. Standardization methods can include selecting statistical features such as median, mean, and variance, or using principal component analysis (PCA) to extract principal components and calculate their contribution rates to determine the rationality of principal component selection. Standardization ensures data consistency and comparability.

[0023] c3. Organize the generated graph structure and normalized data to form a sparse adjacency matrix and multiple feature matrix data sets. This data set is used for subsequent model training and ensures data consistency across all sample groups during training.

[0024] Preferably, in step d, the specific operation methods of model training and edge prediction are as follows:

[0025] d1. Use the modified Graph Attention Network (GAT) model for training. Since the data is unlabeled, an unsupervised approach is used for training. The key to training lies in defining the GAT's loss function. By calculating the graph data loss between each sample group, the total loss value is obtained, thereby determining the direction of gradient descent. To simplify the data computation, ResNet and pooling operations are first performed on the graph data, and then the loss between the graph data is calculated. Based on the complexity of the experimental data, the appropriate number of convolutional layers and feature outputs are selected, and the learning rate and gradient descent parameters are adjusted to ensure that the model can effectively converge in unsupervised learning.

[0026] d2. During model training, the attention mechanism in the GAT model is used to calculate the attention weight αij for each gene pair. The sparse adjacency matrix is ​​screened and edges with attention weights below a set threshold are deleted. The threshold is determined through multiple experiments and the loss is minimized as the judgment criterion.

[0027] d3. For the deleted edges, calculate the similarity of the gene pairs based on the feature matrix output by the last convolution layer of the model, and select edges that are greater than the set threshold and not included in the original adjacency matrix to add. Retrain the model, adjust the threshold, and continue to optimize the loss value;

[0028] d4. The edge prediction process can be iterated multiple times until the loss value no longer decreases, and the final output is the gene relationship data for a certain cell category.

[0029] The second aspect of the present invention relates to a method and device for predicting gene interaction relationships based on a graph neural network, comprising a memory and one or more processors, wherein the memory stores executable code, and when the processor executes the code, it is used to implement the method for predicting gene interaction relationships based on a graph neural network of the present invention.

[0030] The third aspect of the present invention relates to a computer-readable storage medium having a program stored thereon, which, when executed by a processor, can implement the gene interaction relationship prediction method based on graph neural network of the present invention.

[0031] The present invention generates multiple graph data sets with the same structure but different node features by performing preprocessing, similarity calculation, graph structure construction, and standardization on gene transcription expression data. The model is trained using a graph attention network (GAT) and a loss function is defined in combination with an unsupervised learning method. Specifically, the loss value is obtained by calculating the loss of graph data between different sample groups, thereby determining the direction of gradient descent. During the training process, in order to simplify the calculation, ResNet and pooling operations are first performed on the graph data, and then the loss between the graph data is calculated. During training, the model adjusts the number of convolutional layers, feature output, learning rate, and gradient descent parameters according to the complexity of the data to ensure effective convergence. This method can be used to analyze gene expression data, predict the interaction between genes, and perform unsupervised learning in the absence of labeled data.

[0032] Compared with the prior art, the advantages and positive effects of the present invention are:

[0033] The present invention predicts gene correlations through an unsupervised graph attention network (GAT) model, overcoming the existing technology's reliance on large amounts of labeled data and enabling effective predictions in unlabeled gene data. In addition, the present invention proposes an edge prediction mechanism that adaptively adds or deletes edges in the graph based on the relationships between gene data, thereby improving the flexibility of the model and the accuracy of predictions. To simplify the calculation process, the present invention introduces ResNet and pooling operations into the model, reducing the amount of data calculation while ensuring effective convergence of the model, thereby improving computational efficiency. This is particularly suitable for large-scale, high-dimensional gene expression data processing, significantly improving the performance and efficiency of gene interaction mining. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flow chart of the method of the present invention. Specific implementation methods

[0035] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0036] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways than those described herein. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0037] Example 1

[0038] like Figure 1This embodiment aims to achieve the possibility of more accurate detection of gene correlation. To this end, this embodiment provides a gene interaction relationship prediction method based on graph neural network, including the following steps:

[0039] First, experimental data is obtained from a public database. In this embodiment, all data related to renal cancer crawled from the GEO database are selected. These data are usually messy and therefore need to be standardized and preprocessed. The downloaded data may contain gene expression files from multiple different platforms. First, all gene expressions need to be converted into a unified identification form, such as Entrez Gene, Symbol, etc. If the platform file lacks information about the corresponding gene, it can be converted through other databases, but some data may be lost in this process. For example, when converting a symbol to a unified ID, some symbols may not be able to find corresponding items in the ID database. This loss is difficult to avoid in complex data sets.

[0040] Next, review the experimental description file for the data to ensure you understand the experimental design, including the experimental subjects, experimental variables, experimental groups, and control groups. Group the data based on the experimental variables. For example, you can divide the data into five categories: cancer, metastasis, normal, primary, and tumor. Each category contains multiple sample group files. After further processing, each file will form a graph.

[0041] For each sample group file, we calculated the similarity between gene expression data, using gene expression as a node feature. To this end, we used two similarity calculation methods: the Pearson correlation coefficient for normally distributed data and the Spearman correlation coefficient for non-normally distributed data. Based on the calculation results, we set a similarity threshold to identify significantly correlated gene pairs. For example, setting the threshold to 0.3 would select gene pairs with a similarity greater than this threshold.

[0042] After obtaining gene-gene similarity data for each sample group, we began statistical analysis. In the cancer category, there are 30 sample group files. We counted the number of gene pairs that appeared in each sample group, retaining those with more than 70% of the gene pairs. Ultimately, we obtained 2,772,471 gene pairs, forming a preliminary graph structure. This approach effectively addresses noise and incompleteness in the data. Sample groups that do not contain certain gene pairs may have been undetected due to noise in the data or because the similarity score did not meet the threshold. To address this issue, we performed zero padding on these sample groups.

[0043] After obtaining consistent gene pairs for all sample groups within each category (e.g., cancer), the genes were uniformly screened and arranged in a fixed order to ensure that each sample group contained the same 16,841 genes. Next, the IDs in the gene pairs were replaced with the gene numbers in the sample group, ultimately generating consistent gene expression data and corresponding graph structure data.

[0044] Because the number of samples in each sample group varied, the gene features needed to be dimensionalized. Principal component analysis (PCA) was used for feature extraction, compressing the expression data into three principal components. The principal component contributions were verified to be greater than 95% across all sample groups. This yielded 30 graphs with consistent dimensions, resulting in a feature matrix of 16,841 × 3, and a sparse adjacency matrix.

[0045] For model selection, the Graph Attention Network (GAT) model was chosen. Since the data was unlabeled, an appropriate unsupervised loss function was selected for model training. Similarity between graph data was used as the loss function, but due to the large number of nodes in the graph, directly calculating similarity would result in excessive computational complexity. Therefore, ResNet and pooling operations were introduced into the loss function to simplify the computation and reduce dimensionality, thereby enabling similarity calculation.

[0046] After the model training reaches loss convergence, the edge prediction operation is further performed. This process is divided into two parts: edge deletion and edge addition. The purpose is to make the distance of the graph data closer by optimizing the edge structure. Specifically, the maximum threshold is first screened according to the threshold, the edge deletion operation is performed, and then the edge addition operation is performed. It is worth noting that the threshold here is in the form of a percentage, because the weight a ij While setting an absolute value for is difficult, using a percentage is more convenient. Pruning and adding as independent modules can be repeated based on the model's convergence, until the loss cannot be reduced further. Finally, an optimized gene interaction graph is obtained.

[0047] Example 2

[0048] This embodiment relates to a gene interaction relationship prediction device based on a graph neural network, including a memory and one or more processors. The memory stores executable code. When the processor executes the code, it is used to implement the gene interaction relationship prediction method based on a graph neural network of Example 1.

[0049] Example 3

[0050] This embodiment relates to a computer-readable storage medium having a program stored thereon. When a processor executes the program, the gene interaction relationship prediction method based on graph neural network of Example 1 can be implemented.

[0051] The above description is merely a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any person skilled in the art may utilize the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes for application in other fields. However, any simple modification, equivalent change, and modification of the above embodiments made in accordance with the technical essence of the present invention without departing from the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A gene interaction relationship prediction method based on graph neural network, characterized in that: The method comprises the following steps: a. Preprocess the transcriptional expression data of each experimental group, including standardizing the gene ID, and represent the gene transcriptional expression data of different experimental groups with a unified gene ID to ensure consistent data format; then, group the cell samples in the experimental group, and one sample group contains multiple samples, and finally obtain the original data for subsequent analysis; b. Calculate the gene similarity of the above original data according to each sample group, and obtain the similarity data between genes in the sample group; by statistically analyzing and screening the data of all sample groups, select gene pairs that meet the set conditions as the basis for constructing the graph data structure to ensure that the screened gene pairs can reflect the actual biological significance; the specific operation steps are as follows: b1. For each sample group under each cell category, calculate the similarity between the gene pairs in the group in turn; the gene expression data of each sample is regarded as the feature of the gene, and the similarity calculation method is selected according to the characteristics of the data to calculate and obtain the similarity data between genes in the sample group; b2. For each sample group, calculate the similarity according to the pre-set threshold b. Then, statistical screening is performed on the sample groups of each cell category, and the gene pairs of all sample groups under the same category are counted, and the gene pairs with a frequency greater than the set threshold are selected. The choice of threshold will also have an important impact on the formation of the graph structure, and the relationship between the gene pairs finally retained will constitute the edge in the graph data, which is the core of the graph structure. c. Based on the graph structure obtained by the above screening, the original data is further processed to ensure that all sample groups contain only the genes defined in the graph structure. Subsequently, the data of these sample groups are standardized so that the data dimensions of all sample groups are consistent. Through this standardization, multiple graph data groups with the same structure but different node features are formed for subsequent graph neural network model training. Among them, the node features are gene expression features. d. Perform model training on the generated graph dataset. The model used is a graph attention network (GAT) and an edge prediction algorithm. Through multiple iterative training of the model, the loss function is gradually optimized to ultimately obtain edge weight data in the graph structure that can accurately reflect the interactions between genes. The specific steps are as follows: d1. Use the modified Graph Attention Network (GAT) model for training. Since the data is unlabeled, an unsupervised approach is used for training. The GAT loss function is defined. By calculating the graph data loss between each sample group, the total loss value is obtained, thereby determining the direction of gradient descent. To simplify data calculations, ResNet and pooling operations are first performed on the graph data, and then the loss between the graph data is calculated. Based on the complexity of the experimental data, the appropriate number of convolutional layers and feature outputs are selected, and the learning rate and gradient descent parameters are adjusted to ensure that the model can effectively converge in unsupervised learning. The similarity between graph data is used as the loss function. d2. During model training, the attention mechanism in the GAT model is used to calculate the attention weight 𝛼𝑖𝑗 of each gene pair, and the sparse adjacency matrix is ​​screened to delete edges whose attention weight is lower than the set threshold. The threshold is determined through multiple experiments, and the minimization of the loss value is used as the judgment criterion. d3. For the deleted edges, calculate the similarity of the gene pairs based on the feature matrix output by the last convolution layer of the model, and select edges that are greater than the set threshold and not included in the original adjacency matrix to add; retrain the model, adjust the threshold, and continue to optimize the loss value; d4. The edge prediction process is iterated multiple times until the loss value no longer decreases, and the final output is the gene relationship data for a certain cell category.

2. The method for predicting gene interaction relationships based on graph neural networks according to claim 1, characterized in that: In step a, the preprocessing process of transcriptional expression data specifically includes the following steps: a1. Since different experimental groups use different sequencing technologies, resulting in inconsistent formats of gene expression data, it is necessary to uniformly convert the data of all experimental groups into the same gene ID, or other consistent standard data formats, to ensure the comparability of subsequent analysis; a2. According to the different conditional variables of the experiment, each group of experimental data is grouped and processed; each experimental data is divided into multiple sample groups, and the conditional variable is the cell type; a3. The grouped data is sorted to form a set of sample groups corresponding to each cell type to ensure that the gene expression data in each sample group can reflect the biological characteristics of the category.

3. The method for predicting gene interaction relationships based on graph neural networks according to claim 1, characterized in that: The specific operations for normalization and graph data generation in step c are as follows: c1. When screening gene pairs, some sample groups do not contain the screened genes, so these genes need to be supplemented; use the trained model to perform data supplementation, or directly use zero values ​​to ensure that each sample group contains the required gene pairs; through such screening and supplementation, a gene dataset containing all graph structure nodes is finally obtained; c2. Standardize the gene expression data of each sample group to unify its dimensions; Standardization methods include selecting statistical features such as median, mean, and variance, or using principal component analysis (PCA) to extract principal components and calculate their contribution rates to determine the rationality of principal component selection. Standardization ensures data consistency and comparability. c3. Arrange the generated graph structure and the standardized data to form an adjacency matrix represented by a sparse matrix and multiple feature matrix data sets; This data set is used for subsequent model training and ensures the consistency of data for each sample group during the training process.

4. A gene interaction relationship prediction device based on graph neural network, comprising a memory and one or more processors, wherein the memory stores executable code, and when the processor executes the code, it is used to implement the method described in any one of claims 1 to 3.

5. A computer-readable storage medium having a program stored thereon, which is capable of implementing the method according to any one of claims 1 to 3 when a processor executes the program.

Images