Multi-omics data fusion method based on graph attention network
Through the multi-omic data fusion method based on graph attention network, the problem of failure to fully integrate multi-omic data and capture node importance in the prior art is solved, and more accurate identification of cancer-related genes and patient stratification are achieved. Better understand.
Patent Information
- Application Number
- CN202411004206.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-25
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-07-25
AI Technical Summary
Existing multiomic cancer subtype prediction methods based on graph neural networks fail to fully consider all types of omic data, cannot determine the relative importance of adjacent nodes, and there is no redundant information and connections during data integration. Captured problem.
Using multi-omics data fusion method based on graph attention network, acquiring multi-omics gene data for preprocessing, a weighted sample similarity network is constructed, node feature representation matrix is extracted, and data is performed through cross-fusing network and self-attention network. Fusion, and finally input a fully connected neural network to extract the comprehensive node feature representation.
Efficient dimensionality reduction and complementary utilization of multiomic data is achieved, and complex relationships between molecular entities are captured, providing more accurate disease mechanisms and patient stratification insights.
Smart Images

Figure CN119028446A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to bioinformatics, and in particular to a multi-omics data fusion method based on a graph attention network. Background Art
[0002] Gene regulation in the human body is a complex process. Multiple regulatory factors, including methylation, copy number, microRNA (miRNA) and long non-coding RNA (lncRNA), can affect gene expression. Therefore, the study of cancer-related genes should comprehensively consider the information of various omics. Integrating multi-omics data and comprehensively considering the changes in gene sequences, regulatory factors and gene expression during the development of cancer can more accurately identify cancer-related genes. The integration of multi-omics data has brought great hope for understanding the complex biology of diseases, especially Alzheimer's disease, Parkinson's disease and cancer.
[0003] Many models utilize graph-based learning models to extract hidden representations and network structures from different omics data to enhance cancer prediction, patient classification, etc. Existing multi-omics cancer subtype prediction methods based on graph neural networks (GNNs) have three shortcomings: (a) they do not consider all types of omics data; (b) they cannot determine the relative importance of adjacent nodes when performing downstream analyses such as subtype classification and patient stratification; and (c) the same method is used to generate the initial graphs of different omics data.
[0004] At present, according to the different stages of multi-omics data integration, existing methods can be roughly divided into three categories: early integration, late integration and mid-term integration. The limitation of early integration is that it cannot fully utilize the correlation and complementarity between multiple high-dimensional omics data, resulting in a large amount of redundant information in the data set after early integration and a low signal-to-noise ratio. Mid-term integration of multi-omics data refers to the integration of different types of omics information during the construction of the target model. The limitation of mid-term integration is that independent omics models cannot capture certain connections between different types of data. Summary of the invention
[0005] Purpose of the invention: In view of the above shortcomings, the present invention provides a multi-omics data fusion method based on graph attention network, which can achieve efficient dimensionality reduction and utilize the complementarity between different omics data.
[0006] Technical solution: To solve the above problems, the present invention adopts a multi-omics data fusion method based on graph attention network, which includes 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 omics gene data, cosine similarity is used to construct a weighted sample similarity network based on omics features;
[0009] (3) The gene expression matrix and weighted sample similarity network of each omics gene data are input into the graph attention network to extract the node feature representation matrix with spatial information;
[0010] (4) Inputting 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 a 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 the multi-omics gene data in step (1) includes:
[0015] (1.1) Use different variance filtering thresholds to filter different types of omics gene data;
[0016] (1.2) For each omics gene data after filtering, the F value of variance analysis is calculated using the training data, and the FDR control program is used to compensate for multiple tests to obtain omics features. After feature pre-selection, the first principal component of each omics gene data is less than 50% of the variance.
[0017] Furthermore, in step (2), constructing a weighted sample similarity network according to omics features includes:
[0018] Construct the adjacency matrix A by calculating the cosine similarity between node pairs and edges ij :
[0019]
[0020] Among them, x i represents the feature representation of node i, x j represents the feature representation of node j, s(x i ,x j ) represents the cosine similarity between node i and node j, and ε is the threshold.
[0021] Furthermore, the calculation formula of the cosine similarity is:
[0022]
[0023] Among them, || ||2 is the 2-norm calculation.
[0024] Furthermore, the threshold ε is determined by the parameter k, which represents the average number of edges retained by each 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, the step (3) includes:
[0028] (3.1) According to the adjacency matrix A ij Get the adjacent node j of node i and calculate the attention parameter c between node i and adjacent node j ij :
[0029] c ij =LeakyReLU(a T [Wx i ||Wx j ])
[0030] Among them, W and a represent trainable weight matrices, T represents the transpose of the matrix, || represents the connection of two vectors, and LeakyReLU is a nonlinear activation function;
[0031] (3.2) The attention parameters are normalized by the SoftMax function:
[0032]
[0033] Among them, a ij is the normalized attention parameter, N i is the set of all adjacent nodes of node i, and t is the adjacent node of node i;
[0034] (3.3) The normalized attention parameters are used as weights to weight and aggregate the input features:
[0035]
[0036] Among them, h i is the feature representation x i Output representation after graph attention network.
[0037] Furthermore, in step (4), inputting the node feature representation matrix with spatial information into the cross fusion network comprises the following steps:
[0038] (4.1) The node feature representation matrices A, B, and C with spatial information of the three omics gene data are transformed by linear mapping:
[0039]
[0040]
[0041]
[0042] Among them, Q A , K A 、V A is the projection of the node feature representation matrix A, Q B , K B 、V B is the projection of the node feature representation matrix B, Q C , K C 、V C is the projection of the node feature representation matrix C, is a trainable weight matrix;
[0043] (4.2) The calculation formula for the interaction of the node feature representation matrices A, B, and C of the three omics gene data is:
[0044]
[0045]
[0046]
[0047] Among them, f A 、f B 、f C Node feature representation matrix with interactive information for three types of omics gene data, d1, d2, and d3 are standardized parameters.
[0048] Furthermore, the fusion formula for data fusion in step (5) is:
[0049]
[0050]
[0051] Among them, X fusion is the comprehensive node feature representation matrix, c is the total number of omics gene data types, f d is the node feature representation matrix with interactive information for the d-th omics gene data, f eis 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, and then input into the SVM classifier, and the cross entropy loss function is used to calculate the loss:
[0053]
[0054] Among them, y i represents the true label of node i, N represents the number of nodes, and p i The comprehensive node feature representation matrix X fusion Predicted distribution obtained by SVM classifier.
[0055] Beneficial effects: Compared with the prior art, the significant advantage of the present invention is the multi-omics data fusion method based on graph attention network, 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. BRIEF DESCRIPTION OF THE DRAWINGS
[0056] Figure 1 It is a schematic diagram of the process of the present invention.
[0057] Figure 2 This is a schematic diagram of the network structure of the present invention. DETAILED DESCRIPTION
[0058] like Figure 1 and Figure 2 As shown, a multi-omics data fusion method based on a graph attention network in this embodiment includes the following steps:
[0059] Step 1: Obtain three types of multi-omics gene data, namely mRNA, miRNA, and DNA, of several samples. In order to eliminate noise and experimental artifacts in the data and better interpret the results, the omics gene data are appropriately preprocessed. First, different variance filtering thresholds (0.1 for mRNA, 0.1 for miRNA, and 0.001 for DNA) are applied to different types of omics gene data. Secondly, omics features are further pre-selected through statistical verification. For each omics gene data, 30% of the randomly selected omics gene data is used as training data to calculate the F-value of the analysis of variance (ANOVA), and the FDR control procedure is used for multiple test compensation. The first principal component of the data after feature pre-selection should be less than 50% of the variance.
[0060] Step 2: For each omics gene data, use cosine similarity and construct a weighted sample similarity network based on omics features. Construct the adjacency matrix A by calculating the cosine similarity between node pairs and edges. ij :
[0061]
[0062] Among them, x i and x j represents the feature representation of node i and node j, represents the cosine similarity between node i and node j. The threshold ε is determined by the parameter k, which represents the average number of edges retained by 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 the graph attention network to extract preliminary features and obtain the 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, ..., x n ], Represents a sample matrix. According to the adjacency matrix A ij Get the adjacent node j of node i. For node i, the attention parameter between the node itself and the adjacent node j is calculated as follows:
[0067] c ij =LeakyReLU(a T [Wx i ||Wx j ])
[0068] in, and is a trainable weight matrix shared by all samples, p is the embedding size, || indicates the connection of two vectors, and LeakyReLU is a nonlinear activation function.
[0069] Step 3.2: Then the attention parameters are normalized by the SoftMax function:
[0070]
[0071] Where N i is the set of all adjacent nodes of node i, and t is the adjacent node of node i.
[0072] Step 3.3: Use the normalized attention parameters as weights to weight and aggregate the input features:
[0073]
[0074] Among them, h i is the output representation of sample i after the graph attention network.
[0075] Step 4: Input the node feature representation with spatial information into the cross-fusion network to obtain the node feature representation after interaction and alignment of different modalities, including:
[0076] Step 4.1: The node feature representations with spatial information of the three omics gene data are used as the input of the cross-attention fusion network. N represents the number of samples, M represents the feature dimension, and they represent the node feature representation matrices with spatial information of the three omics data. Different transformations are performed on matrices A, B, and C through linear mapping:
[0077]
[0078]
[0079]
[0080] Where W Q ,W K ,W V is a trainable weight matrix.
[0081] Step 4.2: By calculating the dot products of different omics Q and K and adding them together, the softmax function is used to obtain the attention coefficients of different omics gene data, which are multiplied by themselves V to obtain the gene expression data after the interaction of the three omics:
[0082]
[0083]
[0084]
[0085] in, It is the node feature representation matrix with interactive information after the cross-fusion network, and d1, d2, and d3 are standardized parameters.
[0086] Step 5: Input the node feature representation matrix with interactive information into the self-attention network, so that different omics gene data can be interactively aligned, and different features between omics gene data can be associated, and the data can be further integrated to obtain a more comprehensive and integrated node feature representation matrix, including:
[0087] Let F = {f1,f2,...,f C} is a patient's c-type omics gene expression matrix, using attention-based MIL pooling to aggregate features and obtain a comprehensive node feature representation matrix X fusion :
[0088]
[0089] in,
[0090]
[0091] It is a trainable weight matrix that introduces positive and negative values of the unit nonlinearity tanh() function under appropriate gradient flow conditions.
[0092] Step 6: Represent the comprehensive node feature matrix X fusion Input the fully connected layer MLP (multi-layer perceptron) to obtain the node feature representation X that integrates multi-omics gene data fusion ', input the output of the fully connected layer into the SVM classifier, and use the cross entropy loss function to calculate the loss, while backpropagating and optimizing the entire network. The loss function is:
[0093]
[0094] Among them, y i is the true label of sample i, p i Yes 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 network, characterized in that: The following steps are involved: (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 omics gene data, cosine similarity is used to construct a weighted sample similarity network based on omics features; (3) The gene expression matrix and weighted sample similarity network of each omics gene data are input into the graph attention network to extract the node feature representation matrix with spatial information; (4) Inputting the node feature representation matrix with spatial information into the cross-fusion network to obtain the node feature representation matrix with interactive information; (5) Input the node feature representation matrix with interactive information into the self-attention network, perform data fusion, and obtain a comprehensive node feature representation 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 the multi-omics gene data in step (1) includes: (1.1) Use different variance filtering thresholds to filter different types of omics gene data; (1.2) For each omics gene data after filtering, the F value of variance analysis is calculated using the training data, and the FDR control program is used to compensate for multiple tests to obtain omics features. After feature pre-selection, the first principal component of each omics gene data is less than 50% of the variance.
4. The multi-omics data fusion method according to claim 2, characterized in that: In the step (2), constructing a weighted sample similarity network according to the omics features includes: Construct the adjacency matrix A by calculating the cosine similarity between node pairs and edges ij : Among them, x i represents the feature representation of node i, x j represents the feature representation of node j, s(x i ,x j ) represents the cosine similarity between node i and node j, and ε is the threshold.
5. The multi-omics data fusion method according to claim 4, characterized in that: The calculation formula of the cosine similarity is: Among them, || ||2 is the 2-norm calculation.
6. The multi-omics data fusion method according to claim 4, characterized in that: The threshold ε is determined by the parameter k, which represents the average number of edges retained by each node: k=∑I(s(x i ,x j )≥ε) / n Where I(·) represents the indicator function and n is the number of nodes.
7. The multi-omics data fusion method according to claim 4, characterized in that: The step (3) comprises: (3.1) According to the adjacency matrix A ij Get the adjacent node j of node i and calculate the attention parameter c between node i and adjacent node j ij : c ij =LeakyReLU(a T [Wx i ||Wx j ]) Among them, W and a represent trainable weight matrices, T represents the transpose of the matrix, || represents the connection of two vectors, and LeakyReLU is a nonlinear activation function; (3.2) The attention parameters are normalized by the SoftMax function: Among them, a ij is the normalized attention parameter, N i is the set of all adjacent nodes of node i, and t is the adjacent node of node i; (3.3) The normalized attention parameters are used as weights to weight and aggregate the input features: Among them, h i is the feature representation x i Output representation after graph attention network.
8. The multi-omics data fusion method according to claim 7, characterized in that: In the step (4), inputting the node feature representation matrix with spatial information into the cross fusion network comprises the following steps: (4.1) The node feature representation matrices A, B, and C with spatial information of the three omics gene data are transformed by linear mapping: Among them, Q A , K A 、V A is the projection of the node feature representation matrix A, Q B , K B 、V B is the projection of the node feature representation matrix B, Q C , K C 、V C is the projection of the node feature representation matrix C, is a trainable weight matrix; (4.2) The calculation formula for the interaction of the node feature representation matrices A, B, and C of the three omics gene data is: Among them, f A 、f B 、f C Node feature representation matrix with interactive information for three types of omics gene data, d1, d2, and d3 are standardized parameters.
9. The multi-omics data fusion method according to claim 8, characterized in that: The fusion formula for data fusion in step (5) is: Among them, X fusion is the comprehensive node feature representation matrix, c is the total number of omics gene data types, f d is the node feature representation matrix with interactive information for the d-th omics gene data, f e is the node feature representation matrix with interactive information for the e-th omics gene data, and w and V are trainable weight matrices.
10. The spatial transcriptome spot gene expression prediction method according to claim 9, characterized in that: In step (6), the comprehensive node feature representation matrix is input into the fully connected neural network MLP, and then input into the SVM classifier, and the cross entropy loss function is used to calculate the loss: Among them, y i represents the true label of node i, N represents the number of nodes, and p i The comprehensive node feature representation matrix X fusion Predicted distribution obtained by SVM classifier.
Citation Information
Patent Citations
Multi-omics intelligent diagnosis system based on deep learning
CN111028939A
Drug sensitivity prediction method and device based on transfer learning and graph neural network
CN112863696A
Method for predicting interaction and affinity of compound and protein based on end-to-end learning
CN113744799A
Efficient prediction method for association relationship between circRNA and miRNA
CN114496084A
Cited By
EDNA data dimension reduction method
CN120510925A
Biological characteristic analysis system based on multi-modal data and residual image attention network
CN120655990A
Cancer prognosis prediction method and system based on multi-omics fusion
CN120895252A
