A multi-omics causal structural relationship learning method based on contrastive learning

Abstract

The present invention discloses a multi-omics causal structure relationship learning method based on contrastive learning, comprising: first, constructing corresponding encoders for pre-processed gene mutation and gene expression data, respectively, and performing feature extraction on the two types of omics data; then, constructing a parameter-sharing projection head to achieve cross-modal feature alignment; subsequently, using the aligned features as nodes, constructing causal graph data through a learnable causal graph structure; constructing a graph neural network to learn the causal graph representation, and constructing a contrast loss function; finally, obtaining the model prediction result through a multi-layer perceptron, and constructing a survival prediction loss function, and obtaining a total loss function to train a multi-omics causal structure model. The present invention is based on gene mutation and gene expression data, and constructs and learns cross-omics causal structure relationships through contrastive learning, providing more accurate prognosis predictions for diseases such as acute myeloid leukemia, and helping to discover potential biomarkers and key regulatory factors.

Description

Technical Field

[0001] The present invention relates to causal structural relationship learning, and in particular to a multi-omics causal structural relationship learning method based on contrastive learning. Background Art

[0002] Acute myeloid leukemia (AML) is a hematologic malignancy originating in the bone marrow, characterized by complex clinical manifestations and a poor prognosis. Traditional prognostic methods often rely on limited clinical indicators and single gene markers. However, due to the complex pathogenesis of AML, involving abnormalities in multiple genes and pathways, analysis of single-omics data often fails to fully reveal the onset and progression of the disease. Therefore, multi-omics analysis, particularly the integration of gene mutation and gene expression, has become a key focus in AML prognostic research in recent years. Gene mutations are closely associated with the development and prognosis of AML. Different gene mutations not only affect the proliferation and differentiation of leukemic cells but also are closely associated with treatment response and patient survival. For example, mutations in FLT3, NPM1, and IDH1 / 2 are widely considered prognostic markers for AML. Compared to gene mutations, changes in gene expression levels are often closely associated with cell functions such as proliferation, differentiation, and immune evasion, and therefore hold great significance in the diagnosis and prognosis of AML. By integrating gene mutation and gene expression data, multi-omics analysis can provide a multifaceted understanding of the complexity of AML.

[0003] Multi-omics causal structural relationships, particularly causal structural analysis of gene mutations and gene expression for prognosis prediction in acute myeloid leukemia (AML), have important clinical implications. Causal structural relationship analysis can reveal the interactions and causal mechanisms between different genes and their expression products through multi-level and multi-dimensional data fusion. By constructing multi-omics causal structural models, the potential causal relationships between gene mutations and gene expression can be more accurately captured, thereby revealing the underlying mechanisms of AML and providing new insights for clinical prognosis prediction. Causal structural analysis not only helps improve the accuracy of prognostic prediction but also provides a scientific basis for early diagnosis, risk assessment, and treatment optimization of AML.

[0004] Traditional causal structure methods typically require explicit definition of the causal structure between variables, relying on expert knowledge or hypothetical models. For example, structural equation modeling (SEM) requires a priori specification of causal pathways, which can lead to biased results if incorrectly specified. Furthermore, estimating causal relationships requires extensive data, and in some fields (such as medicine and bioinformatics), high-quality data is difficult to obtain, potentially leading to overfitting or unrobust estimates. Many traditional causal methods assume that the relationships between variables are linear (such as the linear regression assumption in SEM). However, in reality, causal relationships are often nonlinear and even involve complex interactions. However, existing research is largely based on a single omics or modality. When fusion of omics data (e.g., genomics, transcriptomics, and proteomics) is used, establishing cross-modal causal structures remains a challenge. Therefore, obtaining a good representation of diverse omics data and learning cross-modal causal structures will not only improve the accuracy of disease-related predictions but also provide a deeper understanding of disease mechanisms. Summary of the Invention

[0005] The purpose of the present invention is to address the deficiencies of the prior art and provide a multi-omics causal structure relationship learning method based on contrastive learning.

[0006] The object of the present invention is achieved through the following technical solutions: a multi-omics causal structure relationship learning method based on contrastive learning, comprising:

[0007] First, corresponding encoders are constructed for the preprocessed gene mutation and gene expression data, and features are extracted for the two types of omics data respectively;

[0008] Then, a parameter-sharing projection head is constructed to achieve cross-modal feature alignment;

[0009] Then, the aligned features are used as nodes to construct causal graph data through a learnable causal graph structure;

[0010] Then, a graph neural network is constructed to learn causal graph representation and a contrastive loss function is constructed;

[0011] Finally, the model prediction results are obtained through a multi-layer perceptron, and a survival prediction loss function is constructed to obtain a total loss function for multi-omics causal structure model training; the multi-omics causal structure model includes the encoder, the projection head, the graph neural network and the multi-layer perceptron.

[0012] Furthermore, the corresponding encoders are constructed for the pre-processed gene mutation and gene expression data, and feature extraction is performed on the two types of omics data, including:

[0013] For omics data on gene mutations, the encoder includes a mutation embedding layer and an intergenic interaction layer. The mutation embedding layer makes each site independent, mapping discrete mutation types into vectors. Each site has its own set of embedding vectors. The intergenic interaction layer uses a Transformer encoder to capture interactions between sites and enhance feature representation.

[0014] For gene expression omics data, the encoder includes a gene-independent embedding layer and a gene-interaction layer. The gene-independent embedding layer includes a linear layer and an activation function, which converts the scalar of each gene into a high-dimensional feature to ensure that the position of the gene feature is consistent with the input; the gene-interaction layer adopts a Transformer-based encoder and uses a self-attention mechanism to allow each gene feature to influence each other, so that each gene feature can integrate global information.

[0015] Furthermore, the parameter-sharing projection head is constructed to achieve cross-modal feature alignment, including:

[0016] The projection head consists of shared layers and private layers. The shared layers process common features, while the private layers process modality-specific features. The shared layers share some parameters to facilitate knowledge transfer between modalities. The private layers do not share parameters, retaining modality-specific information. Both are based on multi-layer perceptrons and are configured with layer normalization to stabilize the training process.

[0017] Then the cosine loss function is performed on the features passed through the projection head Compute and directly maximize the cosine of the feature vector angle to enhance feature alignment through feature similarity:

[0018]

[0019] in, and They are the gene mutation and gene expression characteristics after being processed by the projection head.

[0020] Furthermore, the aligned features are used as nodes to construct causal graph data through a learnable causal graph structure, including:

[0021] The learnable causal graph structure is represented as a learnable adjacency matrix A, and the directed acyclic graph constraint is introduced ; The adjacency matrix is ​​a binary matrix, A[i,j]=1 indicates that there is a causal edge from the mutant gene i to the expressed gene j;

[0022]

[0023] Where ∗ represents the Hadamard product of two matrices, and n is the number of features;

[0024] The causal graph data is constructed for each sample, the nodes are the aligned features, and the node features are the aligned features.

[0025] Furthermore, the construction of a graph neural network to learn causal graph representation and construct a contrast loss function includes:

[0026] The graph neural network uses a graph convolutional neural network. Each node is updated based on the features of its neighboring nodes through a weight matrix and normalization process. The adjacency matrix only contains the connections from the mutation node to the expression node. During message passing, only the features of the mutation node are aggregated to the expression node. That is, the mutation node features remain unchanged between layers, while the expression node features are updated based on the mutation node features and adjacency weights.

[0027] Constructing contrastive loss function It is constructed based on the causal graph representation z learned by the graph neural network:

[0028]

[0029] in, is the positive sample set, is the negative sample set, is the temperature coefficient, which controls the sharpness of the distribution.

[0030] Furthermore, the model prediction results are obtained by the multi-layer perceptron, and a survival prediction loss function is constructed to obtain a total loss function for multi-omics causal structure model training, specifically:

[0031] The multi-layer perceptron consists of three linear layers, two activation functions, and layer normalization. The model prediction result, i.e., the risk score, is obtained through the last linear layer.

[0032] Constructed survival prediction loss function It is based on the risk score predicted by the model and the survival status and survival time corresponding to the sample, that is, the Cox loss function:

[0033] For each sample i, its survival time is , the censoring mark is , the predicted risk score is , Indicates time All samples still in the risk set after is the predicted risk score corresponding to sample j, N is the number of all samples, and the Cox loss function is:

[0034]

[0035] Furthermore, during the model training process, the total loss function for:

[0036]

[0037] in is the cosine loss function, is a directed acyclic graph constraint, 、 and are learnable parameters.

[0038] The present invention also provides an electronic device comprising a memory and a processor, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the above-mentioned multi-omics causal structure relationship learning method based on contrastive learning.

[0039] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned multi-omics causal structure relationship learning method based on contrastive learning.

[0040] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements the above-mentioned multi-omics causal structure relationship learning method based on contrastive learning.

[0041] Compared with existing technologies, the present invention has the following advantages: it can effectively integrate multi-omics data such as gene mutation and gene expression in causal structure learning tasks, achieve cross-modal feature alignment through a parameter-sharing projection head, and improve data consistency. Based on a learnable causal graph structure, it explores the causal relationships between different omics features and combines them with graph neural networks to enhance causal representation learning. This allows for a deeper understanding of the interactions between multi-omics data, improves the ability to explain disease mechanisms, and thus provides strong support for the development of personalized treatment plans. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0043] Figure 1 A flowchart of a multi-omics causal structure relationship learning method based on contrastive learning provided by an embodiment of the present invention;

[0044] Figure 2 A schematic diagram of the structure of a gene mutation encoder provided by an embodiment of the present invention;

[0045] Figure 3 A schematic diagram of the structure of a gene expression encoder provided in an embodiment of the present invention;

[0046] Figure 4 A schematic diagram of the structure of a projection head provided by an embodiment of the present invention;

[0047] Figure 5 A schematic diagram of the structure of a multi-layer perceptron provided in an embodiment of the present invention;

[0048] Figure 6 A schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The present invention will be described in detail below with reference to the accompanying drawings. Unless there is any conflict, the features of the following embodiments and implementations may be combined with each other.

[0050] The present invention is a multi-omics causal structure relationship learning method based on contrastive learning, such as Figure 1 As shown, the following steps are included:

[0051] (1) First, corresponding encoders are constructed for the preprocessed gene mutation and gene expression data, and features are extracted for the two types of omics data respectively;

[0052] Specifically, gene mutation and gene expression data were preprocessed separately. For gene mutation data, genes with mutation frequencies below 1% were first excluded. Ultimately, genes with higher mutation frequencies were selected as input, with mutations scored as 1 and non-mutated genes scored as 0. For gene expression data, gene expression data were obtained using RNA-seq and normalized using high-throughput sequencing fragment counts per kilobase per million mapped reads (HTseq-FPKM). Further processing involved converting FPKM values ​​to TPM (transcripts per million), adding 1 to the TPM value, and applying a logarithmic transformation to stabilize variance and improve the reliability of statistical analysis. Next, the variance of gene expression was calculated, and the top 100 genes with the largest variance were selected and their expression levels normalized.

[0053] Then, if Figure 2 As shown in the figure, an encoder is constructed for the omics data of gene mutations. The encoder mainly includes a mutation embedding layer and an intergenic interaction layer. The mutation embedding layer makes each site independent and maps discrete mutation types (such as 0 / 1) into higher-dimensional vectors. Each site has its own set of embedding vectors; the intergenic interaction layer uses a Transformer encoder to capture the interactions between sites and enhance feature representation.

[0054] like Figure 3As shown in the figure, an encoder is constructed for the omics data of gene expression. The encoder mainly includes a gene-independent embedding layer and a gene-interaction layer. The gene-independent embedding layer includes a linear layer and an activation function, which can convert the scalar of each gene into a high-dimensional feature to ensure that the position of the gene feature is consistent with the input; the gene-interaction layer adopts a Transformer-based encoder, which uses a self-attention mechanism to allow each gene feature to influence each other, so that each gene feature can integrate global information.

[0055] (2) Then, a parameter-sharing projection head is constructed to achieve cross-modal feature alignment;

[0056] Specifically, such as Figure 4 As shown, the projection head consists of shared layers and private layers. The shared layers process common features, while the private layers process modality-specific features. The shared layers consist of a linear layer, an activation function (such as Leaky ReLU), layer normalization, a linear layer, an activation function (such as Leaky ReLU), and a linear layer, connected in sequence. The private layers consist of a linear layer, an activation function (such as Leaky ReLU), layer normalization, and a linear layer, connected in sequence. The shared layers partially share parameters to facilitate knowledge transfer between modalities; the private layers do not share parameters and retain modality-specific information. Both are based on multilayer perceptrons and are configured with layer normalization, which helps stabilize the training process, especially when the data distributions of different modalities differ significantly.

[0057] Then the cosine loss function is performed on the features passed through the projection head Compute and directly maximize the cosine of the feature vector angle to enhance feature alignment through feature similarity:

[0058]

[0059] in, and They are the gene mutation and gene expression characteristics after being processed by the projection head.

[0060] (3) Subsequently, the aligned features are used as nodes to construct causal graph data through a learnable causal graph structure;

[0061] Specifically, the learnable causal graph structure is represented as a learnable adjacency matrix A, and the directed acyclic graph constraint is introduced The adjacency matrix is ​​a binary matrix, where A[i,j]=1 indicates a causal edge exists from mutated gene i to expressed gene j. Causal graph data is constructed for each sample, with nodes representing aligned features, and node features representing aligned features.

[0062]

[0063] Where ∗ represents the Hadamard product of two matrices, n is the number of features, is the trace function.

[0064] (4) Furthermore, a graph neural network is constructed to learn the causal graph representation and a contrastive loss function is constructed;

[0065] Graph neural networks employ graph convolutional neural networks, where each node is updated based on the features of its neighboring nodes, processed through a weight matrix and normalization. The adjacency matrix only contains connections from mutation nodes to expression nodes. During message passing, only the features of mutation nodes are aggregated to expression nodes. This means that mutation node features remain unchanged across layers, while expression node features are updated based on mutation node features and adjacency weights.

[0066] The contrastive loss function is constructed based on the causal graph representation z learned by the graph neural network:

[0067]

[0068] in, is the positive sample set, is the negative sample set, and are the causal graph representations of sample i and sample j respectively, is the temperature coefficient, which controls the sharpness of the distribution.

[0069] It should be noted that samples with similar survival times are considered as positive pairs (similar prognosis), and samples with large time differences are considered as negative pairs (different prognosis). At the same time, the event state is considered, that is, for each sample i, the time window is defined ,in It can be set to 3 months. Therefore, the positive sample is the sample whose survival time falls in the same window as sample i, regardless of the survival status; the negative sample is the sample whose time difference exceeds the window, such as or However, if sample j is censored, that is, alive but not dead, then only is considered as a negative sample, considering that its actual survival time may be longer.

[0070] (5) Finally, the model prediction results are obtained through the multi-layer perceptron, and the survival prediction loss function is constructed to obtain the total loss function for multi-omics causal structure model training.

[0071] Specifically, such as Figure 5As shown in the figure, the multi-layer perceptron includes a linear layer, an activation function (such as Leaky ReLU), layer normalization, a linear layer, an activation function (such as Leaky ReLU), layer normalization, and a linear layer, which are connected in sequence. The model prediction result, namely the patient's mortality risk score (an individual's mortality risk score, which can predict the risk of death in the future), is obtained through the last linear layer. The survival prediction loss function constructed is based on the model-predicted risk score and the corresponding survival status and survival time of the sample, namely the CoxLoss loss function:

[0072] For each sample i, its survival time is , the censoring mark is (1 means death upon event occurrence, 0 means right censoring and survival), the predicted risk score is , Indicates time All samples still in the risk set after is the predicted risk score corresponding to sample j, N is the number of all samples, and the Cox loss function for:

[0073]

[0074] During model training, the total loss function is:

[0075]

[0076] in, 、 and are learnable parameters.

[0077] The present invention also provides a multi-omics causal structure relationship learning device based on contrastive learning, comprising:

[0078] The feature extraction module is used to construct corresponding encoders for the preprocessed gene mutation and gene expression data, respectively, and perform feature extraction on the two types of omics data;

[0079] Feature alignment module, used to build a parameter-sharing projection head to achieve cross-modal feature alignment;

[0080] The causal graph construction module is used to construct causal graph data using the aligned features as nodes through a learnable causal graph structure;

[0081] Network learning module, used to build graph neural networks to learn causal graph representations and construct contrastive loss functions;

[0082] A model training module is used to obtain model prediction results through a multi-layer perceptron, construct a survival prediction loss function, and obtain a total loss function for multi-omics causal structure model training; the multi-omics causal structure model includes the encoder, the projection head, the graph neural network and the multi-layer perceptron.

[0083] It should be noted that the system embodiment shown in this embodiment matches the content of the above method embodiment. You can refer to the content of the above method embodiment and will not repeat it here.

[0084] Figure 6 This is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. Figure 6 The electronic device provided in this embodiment includes: a memory and a processor, wherein the memory is used to store information including program instructions, and the processor is used to control the execution of program instructions. When the program instructions are loaded and executed by the processor, the above-mentioned multi-omics causal structure relationship learning method based on contrastive learning is implemented.

[0085] It should be noted that, in addition to Figure 6 In addition to the memory and processor shown, the electronic device may also include other hardware according to its actual functions, which will not be described in detail.

[0086] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned multi-omics causal structure relationship learning method based on contrastive learning.

[0087] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements the above-mentioned multi-omics causal structure relationship learning method based on contrastive learning.

[0088] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0089] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0090] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0091] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0092] The above embodiments are intended only to illustrate the design concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. The scope of protection of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design concepts disclosed in the present invention are within the scope of protection of the present invention.

Claims

1. A multi-omics causal structure relationship learning method based on contrastive learning, characterized by: include: First, corresponding encoders are constructed for the preprocessed gene mutation and gene expression data, and features are extracted for the two types of omics data respectively; Then, a parameter-sharing projection head is constructed to achieve cross-modal feature alignment; Then, the aligned features are used as nodes to construct causal graph data through a learnable causal graph structure; Then, a graph neural network is constructed to learn causal graph representation and a contrastive loss function is constructed; Finally, the model prediction results are obtained through a multi-layer perceptron, and a survival prediction loss function is constructed to obtain a total loss function for multi-omics causal structure model training; the multi-omics causal structure model includes the encoder, the projection head, the graph neural network and the multi-layer perceptron.

2. A multi-omics causal structure relationship learning method based on contrastive learning according to claim 1, characterized in that: The method constructs corresponding encoders for the preprocessed gene mutation and gene expression data, respectively, and extracts features from the two omics data, including: For omics data on gene mutations, the encoder constructed includes a mutation embedding layer and an intergenic interaction layer. The mutation embedding layer makes each site independent and maps discrete mutation types into vectors, with each site having its own set of embedding vectors. The intergenic interaction layer uses a Transformer encoder to capture interactions between sites and enhance feature representation. For the omics data of gene expression, the encoder constructed includes a gene-independent embedding layer and a gene-interaction layer. The gene-independent embedding layer includes a linear layer and an activation function, which converts the scalar of each gene into a high-dimensional feature to ensure that the position of the gene feature is consistent with the input; the gene-interaction layer adopts a Transformer-based encoder, which uses a self-attention mechanism to allow each gene feature to influence each other, so that each gene feature can integrate global information.

3. The multi-omics causal structure relationship learning method based on contrastive learning according to claim 1 is characterized in that: The parameter-sharing projection head is constructed to achieve cross-modal feature alignment, including: The projection head consists of a shared layer and a private layer. The shared layer processes common features, and the private layer processes modality-specific features. The shared layer shares some parameters to facilitate knowledge transfer between modalities; the private layer does not share parameters and retains modality-specific information. Both are based on multi-layer perceptrons and are configured with layer normalization to stabilize the training process. Then the cosine loss function is performed on the features passed through the projection head Calculate and directly maximize the cosine of the feature vector angle to enhance features through feature similarity Alignment: ； in, and They are the gene mutation and gene expression characteristics after being processed by the projection head.

4. The multi-omics causal structure relationship learning method based on contrastive learning according to claim 1 is characterized in that: The aligned features are used as nodes, and causal graph data is constructed through a learnable causal graph structure, including: The learnable causal graph structure is represented as a learnable adjacency matrix A, and the directed acyclic graph constraint is introduced ; The adjacency matrix is ​​a binary matrix, A[i,j]=1 indicates that there is a causal edge from the mutant gene i to the expressed gene j; ； Where ∗ represents the Hadamard product of two matrices, and n is the number of features; The causal graph data is constructed for each sample, the nodes are the aligned features, and the node features are the aligned features.

5. The multi-omics causal structure relationship learning method based on contrastive learning according to claim 1 is characterized in that: The construction of a graph neural network to learn causal graph representation and construct a contrast loss function includes: The graph neural network uses a graph convolutional neural network. Each node is updated based on the features of its neighboring nodes through a weight matrix and normalization process. The adjacency matrix only contains the connections from the mutation node to the expression node. During message passing, only the features of the mutation node are aggregated to the expression node. That is, the mutation node features remain unchanged between layers, while the expression node features are updated based on the mutation node features and adjacency weights. Constructing contrastive loss function It is constructed based on the causal graph representation z learned by the graph neural network: ； in, is the positive sample set, is the negative sample set, and are the causal graph representations of sample i and sample j respectively, is the temperature coefficient, which controls the sharpness of the distribution.

6. The multi-omics causal structure relationship learning method based on contrastive learning according to claim 1, characterized in that: The model prediction results are obtained by the multi-layer perceptron, and the survival prediction loss function is constructed to obtain the total loss function for multi-omics causal structure model training, specifically: The multi-layer perceptron consists of three linear layers, two activation functions, and layer normalization. The model prediction result, i.e., the risk score, is obtained through the last linear layer. Constructed survival prediction loss function It is based on the risk score predicted by the model and the survival status and survival time corresponding to the sample, that is, the Cox loss function: For each sample i, its survival time is , the censoring mark is , the predicted risk score is , Indicates time All samples still in the risk set after is the predicted risk score corresponding to sample j, N is the number of all samples, and the Cox loss function is: 。 7. The multi-omics causal structure relationship learning method based on contrastive learning according to claim 6 is characterized in that: During model training, the total loss function for: ； in is the cosine loss function, is a directed acyclic graph constraint, 、 and are learnable parameters.

8. An electronic device comprising a memory and a processor, characterized in that: The memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement a multi-omics causal structure relationship learning method based on contrastive learning as described in any one of claims 1-7 above.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, a multi-omics causal structure relationship learning method based on contrastive learning is implemented.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, it implements the multi-omics causal structure relationship learning method based on contrastive learning as described in any one of claims 1 to 7.

Images