Gene regulation and control inference method based on causal diagram embedding and conditional cellular network

Abstract

A gene regulation inference method based on causal diagram embedding and conditional cellular network relates to the technical field of gene regulation network prediction, and comprises the following steps: 1, obtaining a gene expression matrix from scRNA-seq, and generating a causal diagram; 2, generating a local feature embedding matrix of a gene by using a graph neural network model based on a causal graph and a known gene regulation and control network graph; 3, constructing a CCSN based on scRNA-seq, converting the CCSN into gene connectivity vectors, and integrating the gene connectivity vectors of all cells to form a CNDM as a global feature matrix; 4, integrating the local feature embedding matrix and the global feature matrix to form a final gene feature matrix; 5, screening a core gene from the gene feature matrix, and constructing a regulation edge matrix; and 6, inputting the regulatory edge matrix into a gene link prediction module to realize inference of the gene regulatory network. By applying the method, the causal relationship and the cell specificity can be integrated, the core gene is effectively screened, the feature fusion is optimized, and the accuracy and the biological interpretation of network inference are improved.

Description

Technical Field

[0001] This invention relates to the field of predictive gene regulatory network technology, specifically a gene regulation inference method based on causal graph embedding and conditional cell networks. Background Technology

[0002] Accurate inference of gene regulatory networks is a core foundation for elucidating cellular function regulation mechanisms and revealing the patterns of disease development. With the rapid development of single-cell RNA sequencing technology, massive amounts of single-cell gene expression data provide data support for analyzing gene regulatory relationships under cellular heterogeneity. However, how to mine reliable gene regulatory networks from high-dimensional, high-noise single-cell data remains a key challenge in the field.

[0003] In existing technologies, traditional gene regulatory network inference methods mostly rely on gene expression correlation analysis, which makes it difficult to distinguish between direct and indirect regulatory relationships and ignores the causal mechanisms between genes, resulting in limited inference accuracy. Although graph neural network-based methods can capture network topological features, they mostly focus on local regulatory relationships and lack global integration of cell-specific regulatory patterns, making it difficult to reflect the dynamic regulatory activity of genes in different cellular states.

[0004] Meanwhile, single-cell data contains a large number of genes, and directly using it for network inference can easily introduce noise due to feature redundancy, increasing computational complexity. Furthermore, core regulatory genes are interfered with by non-critical genes, further reducing the reliability of inference. In addition, existing methods do not adequately supplement neighbor information for isolated genes, and the feature fusion method lacks biological constraints, resulting in insufficient extraction of key regulatory signals. Summary of the Invention

[0005] The purpose of this invention is to provide a gene regulation inference method based on causal graph embedding and conditional cell networks, which can integrate causal relationships with cell specificity, effectively screen core genes and optimize feature fusion, thereby improving the accuracy and biological interpretability of network inference.

[0006] To achieve the above objectives, the present invention adopts the following technical solution.

[0007] A gene regulation inference method based on causal graph embedding and conditional cell networks includes the following steps.

[0008] S1. Obtain the gene expression matrix from single-cell RNA sequencing data and generate a causal graph based on the gene expression matrix.

[0009] S2. Based on causal graphs and known gene regulatory network graphs, a graph neural network model is used to generate local feature embedding matrices of genes.

[0010] S3. Based on single-cell RNA sequencing data, construct a conditional cell-specific network and convert it into a gene connectivity vector. Integrate the gene connectivity vectors of all cells to form a conditional network degree matrix as a global feature matrix.

[0011] S4. Contrastive learning is used to integrate the local feature embedding matrix and the global feature matrix to form the final gene feature matrix.

[0012] S5. Screen core genes from the gene feature matrix and construct a regulatory edge matrix based on the core genes.

[0013] S6. Input the regulatory edge matrix into the gene link prediction module to obtain the probability of the existence of regulatory edges between genes, thereby realizing the inference of the gene regulatory network.

[0014] Furthermore, in S1, the gene expression matrix is ​​processed using a greedy minimum sparsity algorithm to generate a causal graph.

[0015] Furthermore, in S2, the graph neural network model is a graph sampling and aggregation model.

[0016] Furthermore, the graph sampling and aggregation model samples the neighboring nodes of the target gene node based on the causal graph. If there is an isolated gene without neighboring nodes in the causal graph, the neighboring node information of the isolated gene is supplemented in the known gene regulatory network graph.

[0017] Furthermore, in S3, based on single-cell RNA sequencing data, indirect effects are eliminated through conditional independence tests, while direct regulatory edges between genes that pass the tests are retained, thus constructing a conditional cell-specific network for the corresponding cells.

[0018] Furthermore, based on conditional cell-specific networks, the connectivity of each gene is calculated to form a gene connectivity vector for the corresponding cell.

[0019] Furthermore, in S5, Top-k pooling is used to sort the genes in the gene feature matrix according to their importance scores, and the top k genes are retained as core genes.

[0020] Furthermore, for core genes, potential regulatory edges are identified based on causal graphs and known gene regulatory network graphs, and a regulatory edge matrix is ​​constructed.

[0021] Furthermore, in S6, the gene linking prediction module includes a convolutional neural network and a fully connected layer.

[0022] Furthermore, the convolutional neural network is a four-layer convolutional neural network model. The first three layers each include a batch normalization layer, a convolutional layer, and a max pooling layer, while the fourth layer is a depthwise separable convolutional layer.

[0023] By adopting the above technical solution, the present invention has the following beneficial effects: 1. This invention generates local features based on causal graphs and known gene regulatory network graphs, and combines an isolated gene neighbor information supplementation mechanism to accurately capture direct regulatory relationships between genes, avoid interference from indirect associations, and improve the accuracy and completeness of local regulatory pattern characterization. 2. This invention integrates the global regulatory features and local causal features of conditional cell-specific networks, strengthens the biological significance association between features through comparative learning, effectively integrates the cellular heterogeneity information of single-cell data, and achieves comprehensive coverage and deep fusion of feature dimensions; 3. This invention uses Top-k pooling to screen core genes, specifically eliminating redundant gene interference in high-dimensional single-cell data, significantly reducing computational complexity, accurately focusing on key regulatory nodes, and improving the efficiency and targeting of network inference; 4. The link prediction module of the four-layer convolutional neural network in this invention can efficiently extract the local correlation features of the regulatory edge matrix, and combine the fully connected layer to accurately output the regulatory probability, which greatly improves the prediction accuracy and reliability of the regulatory relationship between genes. 5. The overall scheme of this invention integrates causal mechanisms and cell-specific regulatory modes, providing a highly reliable network foundation for downstream research such as analyzing the dynamic regulation mechanism of genes under cellular heterogeneity, mining disease-related core genes, and screening drug targets. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the method steps of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the present invention clearer, the features and performance of a gene regulation inference method based on causal graph embedding and conditional cell networks of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0026] Please see the appendix Figure 1 A gene regulation inference method based on causal graph embedding and conditional cell networks includes the following steps.

[0027] S1. Obtain the gene expression matrix from single-cell RNA sequencing data, process the gene expression matrix using a greedy minimum sparsity algorithm, and generate a causal graph.

[0028] S2. Based on the causal graph and the known gene regulatory network graph, a graph neural network model is used to generate the local feature embedding matrix of genes. Specifically, the graph neural network model is a graph sampling and aggregation model. The graph sampling and aggregation model samples the neighbor nodes of the target gene node based on the causal graph. If there are isolated genes without neighbor nodes in the causal graph, the neighbor node information of the isolated gene is supplemented in the known gene regulatory network graph.

[0029] S3. Based on single-cell RNA sequencing data, indirect effects are eliminated through conditional independence tests, retaining direct regulatory edges between genes that pass the test, and constructing a conditional cell-specific network for each cell. Based on this network, the connectivity of each gene is calculated, forming a gene connectivity vector for the corresponding cell. The gene connectivity vectors from all cells are integrated to form the conditional network degree matrix, which serves as the global feature matrix.

[0030] S4. Contrastive learning is used to integrate the local feature embedding matrix and the global feature matrix to form the final gene feature matrix.

[0031] S5. Top-k pooling is used to sort the genes in the gene feature matrix by importance score, and the top k genes are retained as core genes. Furthermore, for the core genes, potential regulatory edges are identified based on the causal graph and known gene regulatory network graph, and a regulatory edge matrix is ​​constructed.

[0032] S6. Input the regulatory edge matrix into the gene link prediction module to obtain the probability of regulatory edges existing between genes, thereby realizing the inference of the gene regulatory network. Specifically, the gene link prediction module includes a convolutional neural network and a fully connected layer. The convolutional neural network is a four-layer convolutional neural network model. The first three layers each include a batch normalization layer, a convolutional layer, and a max pooling layer, and the fourth layer is a depthwise separable convolutional layer.

[0033] In specific implementation, this invention proposes a method for inferring gene regulatory networks based on a supervised deep learning framework of graph embedding and Conditional Cell-Specific Network (CCSN). This framework utilizes a Graph Sampling and Aggregate (GraphSAGE) model to obtain the local feature embedding matrix of genes and a Conditional Cell-Specific Network (CCSN) to obtain the global feature matrix of genes. After selecting important nodes through comparative learning of the fusion matrix and Top-k pooling, a regulatory edge matrix is ​​formed and input into the linking module to predict gene regulatory networks (GRNs).

[0034] The original gene expression data is an m×n matrix, where m represents the number of genes and n represents the number of cells. A causal graph is obtained by processing the original gene expression matrix using the Greedy Sparsest Permutation (GSP) algorithm.

[0035] GraphSAGE is a graph neural network model that generates low-dimensional embeddings by aggregating gene features and neighborhood structures, thereby learning local neighborhood information of gene nodes. Using GraphSAGE, low-dimensional embeddings of each gene node can be effectively learned on causal graphs obtained using GSP or known gene regulatory network graphs, enabling the acquisition of local features in the network topology. The learning process consists of two parts: sampling and aggregation.

[0036] During sampling, the procedure involves first identifying neighboring nodes in the causal graph. If an isolated gene exists in the causal graph without any neighboring nodes, then the gene is searched within the known gene regulatory network. This is because the causal graph ensures that embedding learning captures direct causal regulatory relationships between genes, while the reference network, as a supplement, ensures that all nodes participate in the embedding learning process. Together, they enhance the biological significance and coverage of the feature embedding.

[0037] For each node , Called a node The set of neighbors for each target node. In the Layer sampling of a group of neighboring nodes : , in, Indicates from Uniform sampling without replacement One neighboring node, It is the first Number of samples per layer It is a node The original set of neighbors.

[0038] The aggregation function is the average function, and the nodes are... of The aggregation of dimensional neighbor nodes is represented as: , in, This represents the number of hops that each vertex can aggregate to its neighboring nodes.

[0039] node The embedding is based on its The embedding, learned by skipping neighbor nodes, combines the target node's own features with the aggregated neighbor features, and generates a new embedding through a nonlinear transformation. Therefore, new vectors can be used. To represent nodes The embedding, this vector captures the node of Jump neighborhood information, represented as: , in express Activation function This represents the parameter matrix to be learned. This indicates a join operation.

[0040] CCSN is a method for constructing cell-specific networks based on single-cell RNA sequencing (scRNA-seq) data. It eliminates indirect effects through conditional independence tests, resulting in a sparser and more realistic gene interaction network. The Conditional Network Degree Matrix (CNDM) is a quantitative representation of conditional correlations between genes, possessing cross-cell comparability. It enhances cross-cell comparability and elucidates the overall regulatory patterns and relationships between genes in different cellular environments.

[0041] For cells genes and Given conditional genes Test for direct correlation: , in As a statistic of conditional independence, after normalization, if (Significance threshold), considered and There is a direct correlation. Indicated in cells In China, genes and In the given The joint conditional probability at time. and They represent and In the given Marginal conditional probability at time.

[0042] For each cell Select a set of conditional genes (Typically a hub gene with high connectivity), generating multiple subnetworks. Ultimately, cells are obtained through average integration. CCSN: , in Represents cells The final integrated CCSN matrix, Indicated by genes When conditions are met, cells The subnetwork matrix, This indicates the number of conditional genes selected. Represents the integrated matrix Gene and The direct correlation strength.

[0043] Convert the CCSN of each cell into a gene connectivity vector to form a matrix of all cells: , in, Indicates gene In cells The network degree in the middle.

[0044] For all genes and all cells calculate This yields an m×n matrix: , The CNDM has the same dimensions as the original gene expression matrix, but it includes network topology information.

[0045] Based on a given gene expression matrix and prior association knowledge, GraphSAGE obtains a local feature matrix by utilizing causal graphs and known regulatory relationships in the training set. This matrix consists of the embeddings of all gene nodes and is represented as follows: , , Where d represents the feature dimension of the gene node.

[0046] CCSN calculates the network degree matrix to obtain the CNDM as the global feature matrix, which is composed of the network degree of all genes in all cells, and is represented as: , .

[0047] Next, contrastive learning will be used to transform the local feature matrix. and global feature matrix The data is then integrated to form the final gene node feature matrix. , The fusion process is as follows: , , , , , in, and It is the fusion weights that need to be learned. The first characteristic matrix represents the local feature matrix. One sample, The first characteristic matrix represents the global characteristic matrix. One sample, It is a cosine similarity matrix. The diagonal is the loss function used when fusing local and global feature matrices using contrastive learning. For positive sample pairs, The temperature parameter is set to 0.1.

[0048] After constructing the feature matrix of the genes, we use Top-k pooling to select the k nodes that have the greatest impact on the entire graph. The specific process is as follows: , , , in This represents the ReLU activation function. It is a learnable vector. The projection score is used to determine which nodes to discard. The L2 norm operator is represented by the topk function, which selects the top k indices from the input vector. This is an index operation that retrieves a slice based on a specified index. It is Hadamaji.

[0049] Next, construct the control side matrix. Used to determine genes To genes Does a directed edge exist between them? The control edge matrix is ​​composed of... Composed of vectors, including genes Feature embedding vector ,Gene Feature embedding vector and the vector obtained through Top-k pooling. , is represented as: , in , It is matrix.

[0050] The core of the gene linking prediction module consists of a stacked four-layer CNN and a fully connected layer. Specifically, the stacked CNN is used for further feature extraction. This is a four-layer CNN model; the first three layers each include a batch normalization layer, a CNN convolutional layer, and a max-pooling layer, but with different kernel sizes of 3, 5, and 7 respectively. The fourth layer is a depthwise separable convolutional layer. The outputs of these four layers are concatenated and then fed into the fully connected layer. The purpose of this step is to perform multi-scale feature extraction. Efficient feature fusion can learn high-level feature representations, enhance the network's non-linear capabilities, and better capture the complex relationships hidden in the data. Finally, the output is fed into a sigmoid function, and the prediction score is normalized within the interval [0, 1] to obtain node pairs. Scoring with control edges transcription factors Regulatory target genes The probability of.

[0051] After thresholding the probability scores of all nodes having regulatory edges, high-confidence regulatory relationships between genes are retained to form the final gene regulatory network.

[0052] Throughout the training process, we use the following loss function: , in, This is the final loss function of our model. This refers to the loss function used when fusing local and global feature matrices using contrastive learning. , These are hyperparameters, with values ​​of 0.8 and 0.2 respectively.

[0053] The loss function in the link prediction module is as follows: , in, This represents the number of samples used in training. Indicates the first The true label of each sample This represents the score predicted by the model.

[0054] It should be noted that the parts not described in detail in this solution are all prior art. The above embodiments are only used to illustrate the present invention, but the present invention is not limited to the above embodiments. Any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for inferring gene regulation based on causal graph embedding and conditional cell networks, characterized in that: Includes the following steps, S1. Obtain the gene expression matrix from single-cell RNA sequencing data and generate a causal graph based on the gene expression matrix; S2. Based on causal graphs and known gene regulatory network graphs, a graph neural network model is used to generate local feature embedding matrices of genes. S3. Based on single-cell RNA sequencing data, construct a conditional cell-specific network and convert it into a gene connectivity vector. Integrate the gene connectivity vectors of all cells to form a conditional network degree matrix as a global feature matrix. S4. Contrastive learning is used to integrate the local feature embedding matrix and the global feature matrix to form the final gene feature matrix; S5. Screen core genes from the gene feature matrix and construct a regulatory edge matrix based on the core genes; S6. Input the regulatory edge matrix into the gene link prediction module to obtain the probability of the existence of regulatory edges between genes, thereby realizing the inference of the gene regulatory network.

2. The method for inferring a gene regulatory network as described in claim 1, characterized in that: In S1, the gene expression matrix is ​​processed using a greedy minimum sparsity algorithm to generate a causal graph.

3. The method for inferring a gene regulatory network as described in claim 1, characterized in that: In S2, the graph neural network model is a graph sampling and aggregation model.

4. The method for inferring a gene regulatory network as described in claim 3, characterized in that: The graph sampling and aggregation model is based on sampling the neighboring nodes of the target gene node from the causal graph. If there is an isolated gene without neighboring nodes in the causal graph, the neighboring node information of the isolated gene is supplemented in the known gene regulatory network graph.

5. The method for inferring a gene regulatory network as described in claim 1, characterized in that: In S3, based on single-cell RNA sequencing data, indirect effects are eliminated through conditional independence tests, and direct regulatory edges between genes that pass the tests are retained to construct conditional cell-specific networks for the corresponding cells.

6. The method for inferring a gene regulatory network as described in claim 5, characterized in that: Based on conditional cell-specific networks, the connectivity of each gene is calculated to form the gene connectivity vector of the corresponding cell.

7. The method for inferring a gene regulatory network as described in claim 1, characterized in that: In S5, Top-k pooling is used to sort the genes in the gene feature matrix according to their importance scores, and the top k genes are retained as core genes.

8. The method for inferring a gene regulatory network as described in claim 7, characterized in that: For core genes, potential regulatory edges are identified based on causal graphs and known gene regulatory network graphs, and a regulatory edge matrix is ​​constructed.

9. The method for inferring a gene regulatory network as described in claim 1, characterized in that: In S6, the gene linking prediction module includes a convolutional neural network and a fully connected layer.

10. The method for inferring a gene regulatory network as described in claim 9, characterized in that: The convolutional neural network is a four-layer convolutional neural network model. The first three layers each include a batch normalization layer, a convolutional layer, and a max pooling layer, and the fourth layer is a depthwise separable convolutional layer.

Images