A multi-omics data fusion method based on a graph attention network

Abstract

This invention discloses a multi-omics data fusion method based on graph attention networks. First, multi-omics gene data is acquired and preprocessed to obtain a gene expression matrix. A weighted sample similarity network is constructed using cosine similarity. The gene expression matrix and the weighted sample similarity network are input into a graph attention network. The extracted node feature representation matrix with spatial information is input into a cross-fusion network. Then, the output node feature representation matrix with interaction information is input into a self-attention network for data fusion, resulting in a comprehensive node feature representation matrix. Finally, the comprehensive node feature representation matrix is ​​input into a fully connected neural network (MLP) to extract the final effective node feature representation matrix. This multi-omics data fusion method based on graph attention networks reduces the dimensionality of heterogeneous and high-dimensional multi-omics data, utilizes the complementarity between different modalities, and employs a suitable and reliable fusion method, significantly improving classification performance.

Description

Technical Field

[0001] This invention relates to bioinformatics, specifically to a multi-omics data fusion method based on graph attention networks. Background Technology

[0002] Gene regulation in the human body is a complex process, with multiple regulatory factors, including methylation, copy number, microRNA (miRNA), and long non-coding RNA (lncRNA), all influencing gene expression levels. Therefore, research on cancer-related genes should comprehensively consider information from various omics systems. Integrating multi-omics data, and comprehensively considering changes in gene sequence, regulatory factors, and gene expression during cancer development and progression, can more accurately identify cancer-related genes. The integration of multi-omics data offers great hope for understanding the complex biology of diseases, particularly Alzheimer's disease, Parkinson's disease, and cancer.

[0003] Many models utilize graph-based learning models to extract hidden representations and network structures from diverse omics data to enhance cancer prediction, patient classification, and other applications. Existing multi-omics cancer subtype prediction methods based on graph neural networks (GNNs) suffer from three drawbacks: (a) they do not consider all types of omics data; (b) they cannot determine the relative importance of adjacent nodes during downstream analyses, such as subtype classification and patient stratification; and (c) they use the same methods to generate initial graphs for different omics data.

[0004] Currently, based on the different stages of multi-omics data integration, existing methods can be broadly categorized into three types: early-stage integration, late-stage integration, and mid-stage integration. The limitation of early-stage integration lies in its inability to fully utilize the correlations and complementarities between multiple high-dimensional omics datasets, resulting in a dataset containing a large amount of redundant information and a low signal-to-noise ratio. Mid-stage integration of multi-omics data refers to integrating different types of omics information during the construction of the target model. The limitation of mid-stage integration is that independent omics models cannot capture certain relationships between different types of data. Summary of the Invention

[0005] Purpose of the invention: To address the above-mentioned shortcomings, this invention provides a multi-omics data fusion method based on graph attention networks that achieves efficient dimensionality reduction and leverages the complementarity between different omics data.

[0006] Technical Solution: To address the above problems, this invention employs a multi-omics data fusion method based on graph attention networks, comprising the following steps:

[0007] (1) Obtain multi-omics gene data, perform preprocessing to obtain omics features, and obtain the gene expression matrix of each omics gene data;

[0008] (2) For each type of omics gene data, cosine similarity is used and a weighted sample similarity network is constructed based on omics characteristics;

[0009] (3) Input the gene expression matrix and weighted sample similarity network of each omics gene data into the graph attention network to extract the node feature representation matrix with spatial information;

[0010] (4) Input the node feature representation matrix with spatial information into the cross-fusion network to obtain the node feature representation matrix with interactive information;

[0011] (5) Input the node feature representation matrix with interactive information into the self-attention network, perform data fusion, and obtain the comprehensive node feature representation matrix;

[0012] (6) Input the comprehensive node feature representation matrix into the fully connected neural network MLP to extract the final effective node feature representation matrix.

[0013] Furthermore, the multi-omics gene data includes three types of multi-omics gene data: mRNA, miRNA, and DNA.

[0014] Furthermore, the preprocessing of multi-omics genetic data in step (1) includes:

[0015] (1.1) Different variance filtering thresholds are used to filter different types of omics gene data;

[0016] (1.2) For each omics gene data after filtering, the F-value of the analysis of variance is calculated using the training data. The FDR control program is used to perform multiple test compensation to obtain omics features. After feature pre-selection, the first principal component of each omics gene data has a variance of less than 50%.

[0017] Furthermore, the construction of a weighted sample similarity network based on omics features in step (2) includes:

[0018] The adjacency matrix A is constructed by calculating the cosine similarity between node pairs and edges. ij :

[0019]

[0020] Where, x i Let x represent the feature representation of node i. j Let s(x) represent the feature representation of node j. i ,x j ) represents the cosine similarity between node i and node j, and ε is the threshold.

[0021] Furthermore, the formula for calculating the cosine similarity is:

[0022]

[0023] Where || ||2 is used for the 2-norm calculation.

[0024] Furthermore, the threshold ε is determined by the parameter k, which represents the average number of edges retained per node:

[0025] k=∑I(s(x i ,x j )≥ε) / n

[0026] Where I(·) represents the indicator function, and n is the number of nodes.

[0027] Furthermore, step (3) includes:

[0028] (3.1) Based on the adjacency matrix A ij Obtain the neighboring node j of node i, and calculate the attention parameter c between node i and neighboring node j. ij :

[0029] c ij =LeakyReLU(a T [Wx i ||Wx j ])

[0030] Where W and a represent trainable weight matrices, T represents the transpose of the matrix, || represents the concatenation of two vectors, and LeakyReLU is a non-linear activation function;

[0031] (3.2) Normalize the attention parameters using the SoftMax function:

[0032]

[0033] Among them, a ij N represents the normalized attention parameters. i Let t be the set of all adjacent nodes of node i, and t be the adjacent nodes of node i.

[0034] (3.3) The normalized attention parameters are used as weights to perform weighted and aggregated input features:

[0035]

[0036] Among them, h i It is a feature representation x i The output representation after passing through the graph attention network.

[0037] Furthermore, step (4) of inputting the node feature representation matrix with spatial information into the cross-fusion network includes the following steps:

[0038] (4.1) Transform the spatially information-based node feature representation matrices A, B, and C of the three omics gene data through linear mapping:

[0039]

[0040]

[0041]

[0042] Among them, Q A K A V A Q is the projection of the node feature representation matrix A. B K B V B Q is the projection of the node feature representation matrix B. C K C V C The projection of the node feature representation matrix C, is a trainable weight matrix;

[0043] (4.2) The formula for calculating the interaction of node feature representation matrices A, B, and C of the three omics gene data is as follows:

[0044]

[0045]

[0046]

[0047] Among them, f A f B f C The node feature representation matrix with interactive information for the three omics gene data, where d1, d2, and d3 are standardized parameters.

[0048] Furthermore, the fusion formula for data fusion in step (5) is as follows:

[0049]

[0050]

[0051] Among them, X fusion Let f be the comprehensive node feature representation matrix, where c is the total number of omics gene data types, and f is the total number of omics gene data types. d Let f be the node feature representation matrix with interactive information for the d-th type of omics gene data. eLet be the node feature representation matrix with interactive information for the e-th omics gene data, and w and V are trainable weight matrices.

[0052] Furthermore, in step (6), the comprehensive node feature representation matrix is ​​input into the fully connected neural network MLP, then into the SVM classifier, and the cross-entropy loss function is used to calculate the loss:

[0053]

[0054] Among them, y i Let p represent the actual label of node i, N represent the number of nodes, and p represent the number of nodes. i It is the comprehensive node feature representation matrix X fusion The predicted distribution obtained by the SVM classifier.

[0055] Beneficial effects: Compared with the prior art, the significant advantage of this invention is that it is based on a multi-omics data fusion method using graph attention networks, which extracts effective features, adaptively integrates various omics data, captures the complex relationships between various molecular entities, and ultimately provides more accurate and clinically relevant insights into disease mechanisms and patient stratification. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating the method of the present invention.

[0057] Figure 2 This is a schematic diagram of the network structure of the present invention. Detailed Implementation

[0058] like Figure 1 and Figure 2 As shown in this embodiment, a multi-omics data fusion method based on graph attention networks includes the following steps:

[0059] Step 1: Obtain mRNA, miRNA, and DNA multi-omics gene data from several samples. To eliminate noise and experimental artifacts and better interpret the results, appropriate preprocessing of the omics gene data is performed. First, different variance filtering thresholds are applied to different types of omics gene data (0.1 for mRNA, 0.1 for miRNA, and 0.001 for DNA). Second, statistical validation is used to further pre-select omics features. For each omics gene data set, 30% of the acquired omics gene data is randomly selected as training data to calculate the F-value of the analysis of variance (ANOVA), and multiple test compensation is performed using the FDR control procedure. The first principal component of the data after feature pre-selection should have a variance of less than 50%.

[0060] Step 2: For each type of omics-based genetic data, cosine similarity is used to construct a weighted sample similarity network based on omics features. The adjacency matrix A is constructed by calculating the cosine similarity between node pairs and edges. ij :

[0061]

[0062] Where, x i and x j The feature representations of nodes i and j are shown. Let represent the cosine similarity between node i and node j, and let ε be the threshold determined by parameter k, representing the average number of edges retained for each node.

[0063] k=∑I(s(x i ,x j )≥ε) / n

[0064] Where I() represents the indicator function, and n is the number of nodes.

[0065] Step 3: Use a graph attention network to extract preliminary features and obtain a gene expression matrix with specific spatial information, including:

[0066] Step 3.1: For each type of omics gene data, let n represent the number of nodes in the gene expression matrix and m represent the feature dimension. Then the input gene expression matrix is ​​X = [x1, x2, ..., xm]. n ], This represents a sample matrix. Based on the adjacency matrix A... ij Given the neighboring nodes j of node i, the attention parameters between node i and its neighboring node j are calculated as follows:

[0067] c ij =LeakyReLU(a T [Wx i ||Wx j ])

[0068] in, and It is a trainable weight matrix shared by all samples, p is the embedding size, || denotes the concatenation of two vectors, and LeakyReLU is a non-linear activation function.

[0069] Step 3.2: Then, the attention parameters are normalized using the SoftMax function:

[0070]

[0071] Where N i Let t be the set of all adjacent nodes of node i, and t be the adjacent nodes of node i.

[0072] Step 3.3: Use the normalized attention parameters as weights to perform weighted and aggregated input features:

[0073]

[0074] Among them, h i It is the output representation of sample i after passing through the graph attention network.

[0075] Step 4: Input the node feature representations with spatial information into the cross-fusion network to obtain node feature representations after different modal interactions and alignment, specifically including:

[0076] Step 4.1: Use the spatially information-rich node features from the three omics gene data as input to the cross-attention fusion network. N represents the number of samples, M represents the feature dimension, and C represent the node feature representation matrices with spatial information for the three types of omics data. Different transformations are performed on matrices A, B, and C through linear mapping:

[0077]

[0078]

[0079]

[0080] Among them W Q W K W V It is a trainable weight matrix.

[0081] Step 4.2: Calculate the dot products of Q and K for different omics, sum them, and then use the softmax function to obtain the attention coefficients of the gene data from different omics. Multiply these coefficients by their own V to obtain the gene expression data after the interaction of the three omics:

[0082]

[0083]

[0084]

[0085] in, It is a node feature representation matrix with interactive information after cross-fusion network, where d1, d2, and d3 are standardized parameters.

[0086] Step 5: Input the node feature representation matrix with interactive information into the self-attention network to enable interactive alignment of different omics gene data, associate different features between omics gene data, and further fuse the data to obtain a more comprehensive and integrated node feature representation matrix, specifically including:

[0087] Let F = {f1, f2, ..., f C} is a patient's c-omics gene expression matrix. Attention-based MIL pooling is used to aggregate features, resulting in a comprehensive node feature representation matrix X. fusion :

[0088]

[0089] in,

[0090]

[0091] It is a trainable weight matrix, and under appropriate gradient flow conditions, the positive and negative values ​​of the unit nonlinearity tanh() function are introduced.

[0092] Step 6: Combine the node feature representation matrix X fusion The fully connected MLP (Multilayer Perceptron) layer is used to obtain the node feature representation X, which integrates multi-omics genetic data. fusion The output of the fully connected layer is input into the SVM classifier, and the cross-entropy loss function is used to calculate the loss. Simultaneously, backpropagation is performed on the entire network for optimization. The loss function is:

[0093]

[0094] Among them, y i p is the true label of sample i. i It is X fusion The predicted distribution obtained by the SVM classifier is p(y|x)=[p1,...,p N ].

Claims

1. A multi-omics data fusion method based on graph attention networks, characterized in that, Includes the following steps: (1) Obtain multi-omics gene data, perform preprocessing to obtain omics features, and obtain the gene expression matrix of each omics gene data; (2) For each type of omics gene data, cosine similarity is used and a weighted sample similarity network is constructed based on omics characteristics; (3) Input the gene expression matrix and weighted sample similarity network of each omics gene data into the graph attention network to extract the node feature representation matrix with spatial information; (4) Input the node feature representation matrix with spatial information into the cross-fusion network to obtain the node feature representation matrix with interactive information; including the following steps: (4.1) Spatial information node feature representation matrix of three omics gene data through linear mapping Perform the transformation: in, , , Node feature representation matrix The projection, , , Node feature representation matrix The projection, , , Node feature representation matrix The projection, , , , , , , , , This is a trainable weight matrix; (4.2) Node feature representation matrix of three types of omics gene data The calculation formula for interaction is: in, Node feature representation matrices with interactive information from three types of omics-based gene data. These are standardized parameters; (5) Input the node feature representation matrix with interactive information into the self-attention network, perform data fusion, and obtain the comprehensive node feature representation matrix; The fusion formula for data fusion is: in, This is a matrix representing the overall node features. The total number of types of omics-based genetic data. For the first A node feature representation matrix with interactive information in genomics gene data. For the first A node feature representation matrix with interactive information in genomics gene data. , It is a trainable weight matrix; (6) Input the comprehensive node feature representation matrix into the fully connected neural network MLP to extract the final effective node feature representation matrix.

2. The multi-omics data fusion method according to claim 1, characterized in that, The multi-omics gene data includes three types of multi-omics gene data: mRNA, miRNA, and DNA.

3. The multi-omics data fusion method according to claim 1, characterized in that, The preprocessing of multi-omics gene data in step (1) includes: (1.1) Different variance filtering thresholds are used to filter different types of omics gene data; (1.2) For each omics gene data after filtering, the F value of the analysis of variance is calculated using the training data. The FDR control program is used to perform multiple test compensation to obtain omics features. After feature pre-selection, the first principal component of each omics gene data has a variance of less than 50%.

4. The multi-omics data fusion method according to claim 2, characterized in that, The step (2) of constructing a weighted sample similarity network based on omics features includes: The adjacency matrix is ​​constructed by calculating the cosine similarity between node pairs and edges. : in, Represents a node Feature representation, Represents a node Feature representation, Represents a node and nodes cosine similarity, The threshold value is used.

5. The multi-omics data fusion method according to claim 4, characterized in that, The formula for calculating the cosine similarity is: in, Calculate the 2-norm.

6. The multi-omics data fusion method according to claim 4, characterized in that, The threshold From parameters The decision represents the average number of edges retained per node: in, Indicates an indicator function, It represents the number of nodes.

7. The multi-omics data fusion method according to claim 4, characterized in that, Step (3) includes: (3.1) Based on the adjacency matrix Get Node Adjacent nodes compute nodes With adjacent nodes Attention parameters between : in, and This represents a trainable weight matrix. Represents the transpose of a matrix. This indicates the connection of two vectors. It is a non-linear activation function; (3.2) Through The function normalizes the attention parameters: in, These are the attention parameters after normalization. It is a node The set of all adjacent nodes, It is a node The adjacent nodes; (3.3) The normalized attention parameters are used as weights to perform weighted and aggregated input features: in, It is a feature representation The output representation after passing through the graph attention network.

8. The multi-omics data fusion method according to claim 7, characterized in that, In step (6), the comprehensive node feature representation matrix is ​​input into the fully connected neural network MLP, then into the SVM classifier, and the cross-entropy loss function is used to calculate the loss: in, Represents a node The true label, Indicates the number of nodes. It is a matrix representing the comprehensive node features. The predicted distribution obtained by the SVM classifier.

Images