Method and system for inferring gene regulatory network

Abstract

The invention discloses an inference method and an inference system of a gene regulatory network. The inference method comprises the following steps: acquiring transcriptome data and prior gene network data of a single cell; extracting a gene sub-network related to the transcriptome data from the prior gene network data; obtaining a first view and a second view for the gene sub-network according to a first random deletion strategy and a second random deletion strategy; providing the first view and the second view to a neural network model, and performing comparative learning based on the neural network model to obtain incoming features and outgoing features of each node in the gene sub-network; and constructing a regulation score matrix based on the incoming features and the outgoing features, wherein the regulation score matrix displays the regulation association degree between genes in the transcriptome data of the single cell. According to the technical scheme, the direct causal relationship and the indirect association relationship can be effectively distinguished, so that the gene co-expression network more accurately reflects the real regulation relationship.

Description

Technical Field

[0001] The present invention relates to the field of biological detection technology, and in particular to a gene regulatory network inference method and an inference system thereof. Background Art

[0002] In life science research, genes do not exist in isolation; rather, they function synergistically through interactions and regulatory relationships. Gene regulation is a fundamental molecular mechanism involved in virtually all aspects of life. It underlies gene expression, encompassing the regulation of transcription and translation, and is crucial for important biological processes such as development, differentiation, and the onset and progression of disease. It can be said that gene regulation is at the core of life.

[0003] However, the transcriptome data of single cells lack information that directly reflects the transcriptional regulatory mechanism. In addition, existing single-cell transcriptome sequencing technologies have factors such as depth differences and uneven cell capture efficiency, which will introduce false correlation information during sequencing, making it difficult to distinguish between direct causal relationships and indirect correlations, making it impossible for the gene co-expression network to accurately reflect the true regulatory relationship. Summary of the Invention

[0004] In response to the deficiencies in the prior art, the present invention provides a gene regulatory network inference method that can effectively distinguish direct causal relationships from indirect correlations, so that the gene co-expression network can more accurately reflect the true regulatory relationship.

[0005] This application provides an inference method, which includes:

[0006] Obtain single-cell transcriptome data and prior gene network data;

[0007] Extracting a gene subnetwork related to the transcriptome data from the prior gene network data, wherein isolated gene nodes that do not appear in the gene subnetwork are retained;

[0008] The gene sub-network is perturbed according to a first random deletion strategy to obtain a first view, and the gene sub-network is perturbed according to a second random deletion strategy to obtain a second view, where the first random deletion strategy and the second random deletion strategy are different;

[0009] Providing the first view and the second view to a neural network model, and obtaining incoming features and outgoing features of each node in the gene subnetwork based on comparative learning of the neural network model;

[0010] A regulation score matrix is ​​constructed based on the incoming features and the outgoing features, and the regulation score matrix shows the degree of regulatory association between genes in the transcriptome data of the single cell.

[0011] In one aspect, before the step of perturbing the gene subnetwork according to the first random deletion strategy, the method includes:

[0012] The nodes of the gene sub-network are screened according to a preset threshold value to obtain core nodes whose node degrees are greater than the preset threshold value.

[0013] In one aspect, the node degree represents the number of edges of a node;

[0014] The step of perturbing the gene sub-network according to the first random deletion strategy comprises:

[0015] Randomly removing edges in the gene sub-network with a probability of a1, and randomly masking node features in the gene sub-network with a probability of b1, to obtain a first view;

[0016] The step of perturbing the gene sub-network according to the second random deletion strategy comprises:

[0017] The edges of the core nodes are randomly removed with a probability of a2, the features of the core nodes are randomly masked with a probability of b2, the edges outside the core nodes in the gene sub-network are randomly removed with a probability of a3, and the node features outside the core nodes in the gene sub-network are randomly masked with a probability of b3 to obtain a second view.

[0018] In one aspect, a1 = a3 < a2, b1 = b3 < b2.

[0019] In one aspect, providing the first view and the second view to a neural network model comprises:

[0020] Embedding the first view and the second view into a neural network model according to an incoming direction and an outgoing direction, respectively;

[0021] The neural network model is learned based on comparison between the first view and the second view.

[0022] In one aspect, the step of embedding the first view and the second view into a neural network model comprises:

[0023] embedding the outgoing direction node features of the first view, the incoming direction node features of the first view, the outgoing direction node features of the second view, and the incoming direction node features of the second view into the neural network model;

[0024] The outgoing direction node feature of the first view is

[0025] The incoming direction node feature of the first view is

[0026] The outgoing direction node feature of the second view is

[0027] The incoming direction node feature of the second view is

[0028] in, is the characteristic matrix after perturbation, is the perturbed node adjacency matrix, f is the graph convolutional network encoder that aggregates high-order neighbor features, and h (·,·) is the node feature;

[0029]

[0030] is the Laplace matrix, D is The corresponding degree matrix, I n is the unit diagonal matrix, ‖ j∈P (·) is the concatenation operation, σ is the activation function, l is the number of layers, is the (l-1) layer network parameter matrix, m is the power, P∈{0,1,2};

[0031] The incoming feature is h in , the outgoing feature is h out , h out =f(X,A out ),h in =f(X,A in ).

[0032] In one aspect, the step of embedding the first view and the second view into the neural network model comprises:

[0033] The same role of the same gene in the first view and the second view is taken as a positive sample pair;

[0034] taking the same or different roles of different genes in the first view and / or the second view as negative sample pairs;

[0035] Inputting the positive sample pair and the negative sample pair into the neural network model;

[0036] The model parameters are optimized by maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs.

[0037] In one aspect, the control score matrix is ​​defined as M ij , M ij represents the degree of regulatory association between the i-th and j-th genes;

[0038] M ij =Pij *h out *h in , where P ij is a hyperparameter;

[0039]

[0040] In one aspect, before the step of extracting a gene subnetwork related to the transcriptome data from the prior gene network data, the method includes:

[0041] Eliminating intercellular ligand-receptor interaction data from the prior gene network data;

[0042] The transcriptome data of the single cell is normalized and / or logarithmically transformed.

[0043] In order to solve the above problems, the present application also provides an inference system for a gene regulatory network, the inference system comprising:

[0044] Data acquisition module, used to obtain transcriptome data and prior gene network data of single cells;

[0045] a preprocessing module, configured to extract a gene subnetwork related to the transcriptome data from the prior gene network data, wherein isolated gene nodes that do not appear in the gene subnetwork are retained;

[0046] a neural network model learning module, configured to perturb the gene subnetwork according to a first random deletion strategy to obtain a first view, and perturb the gene subnetwork according to a second random deletion strategy to obtain a second view, wherein the first random deletion strategy and the second random deletion strategy are different; provide the first view and the second view to a neural network model, and obtain incoming features and outgoing features of each node in the gene subnetwork based on comparative learning of the neural network model;

[0047] An inference module is used to construct a regulation score matrix based on the incoming features and the outgoing features, wherein the regulation score matrix shows the degree of regulatory association between genes in the transcriptome data of the single cell.

[0048] The beneficial effects of the present invention are reflected in the following: by introducing prior gene network data, gene subnetworks are extracted from transcriptome data, and a preliminary causal relationship framework for transcriptome data is constructed. The first and second random deletion strategies are used to effectively suppress technical noise and reduce non-causal associations. On this basis, the neural network model uses the first and second views to obtain the incoming and outgoing features of each node in the gene subnetwork through comparative learning. The resulting regulatory score matrix can display the degree of regulatory association between genes in the transcriptome data of single cells. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly describes the drawings required for the specific embodiments or the description of the prior art. Similar elements or parts are generally identified by similar reference numerals throughout the drawings. Elements or parts in the drawings are not necessarily drawn to scale.

[0050] Figure 1 A schematic diagram of the process steps of the gene regulatory network inference method of this application;

[0051] Figure 2 Schematic diagram of the steps for obtaining core nodes in the gene regulatory network inference method of this application;

[0052] Figure 3 A schematic diagram of the steps of obtaining a first view and a second view in the gene regulatory network inference method of the present application;

[0053] Figure 4 Schematic diagram of the steps of training a neural network model using a first view and a second view in the gene regulatory network inference method of the present application;

[0054] Figure 5 For this application Figure 4 Schematic diagram of the steps for training a neural network model in the incoming and outgoing directions;

[0055] Figure 6 A schematic diagram of the steps for forming positive and negative sample pairs to train a neural network model in the gene regulatory network inference method of this application;

[0056] Figure 7 Schematic diagram of the steps of eliminating inter-cell data, normalizing, and logarithmic transformation in the gene regulatory network inference method of this application;

[0057] Figure 8 A schematic diagram comparing the effects of the gene regulatory network inference method of this application with other methods;

[0058] Figure 9 Schematic diagram of the gene regulatory network inference system for this application.

[0059] Description of the accompanying drawings: 100, data acquisition module; 200, preprocessing module; 300, neural network model learning module; 400, inference module. DETAILED DESCRIPTION

[0060] The following embodiments of the technical solution of the present invention will be described in detail with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention and are therefore only examples and are not intended to limit the scope of protection of the present invention.

[0061] It should be noted that, unless otherwise specified, the technical or scientific terms used in this application should have the common meanings understood by those skilled in the art to which the present invention belongs.

[0062] See Figure 1 As shown, the present application provides an inference method for gene regulatory networks. The inference method of the present application systematically integrates prior knowledge and single-cell data through the process of "data acquisition → network construction → noise perturbation → contrastive learning → relationship inference", and uses a dual-view contrastive learning mechanism to suppress non-causal associations, thereby achieving clear causal direction and noise-robust inference of the gene regulatory network.

[0063] Inference methods include:

[0064] Step S10, obtain the transcriptome data and prior gene network data of a single cell; establish an analysis and inference basis through the transcriptome data and prior gene network data. Among them, the single-cell transcriptome data comes from single-cell RNA sequencing (scRNA-seq), which records the expression level of genes in individual cells and reflects the dynamic activity of genes at the transcriptional level. Prior gene network data is a gene interaction network based on literature, databases (such as NicheNet) or experimental verification, which contains known gene regulatory relationships, such as the binding of transcription factors to target genes. By introducing prior knowledge, biological constraints are provided for regulatory network inference, reducing unfounded random associations.

[0065] Step S20: extracting gene subnetworks related to the transcriptome data from the prior gene network data, wherein isolated gene nodes that do not appear in the gene subnetwork are retained; screening gene sets that match the transcriptome data of single cells from the prior gene network data to construct an initial gene subnetwork framework.

[0066] For example, genes that co-exist in single-cell transcriptome data are extracted from the prior gene network to form a gene subnetwork G = (V, E), where V is a gene node and E is a known regulatory edge.

[0067] Genes that do not appear in the gene subnetwork are retained as isolated nodes to avoid missing potential regulatory genes due to the limitations of prior knowledge. These isolated gene nodes can be used to infer their regulatory relationships through data-driven methods in subsequent analysis.

[0068] Step S20 ensures that the gene subnetwork contains both known regulatory cores and exploration space for unknown regulatory relationships, balancing the complementarity of prior knowledge and data-driven.

[0069] Step S30: perturb the gene sub-network according to a first random deletion strategy to obtain a first view, and perturb the gene sub-network according to a second random deletion strategy to obtain a second view, wherein the first random deletion strategy and the second random deletion strategy are different; two different random deletion strategies are applied to the gene sub-network to generate two perturbation views, a first view and a second view, and the difference between the views is used to enhance the contrast learning signal.

[0070] The first view simulates technical noise, such as sequencing errors, common in single-cell data by randomly deleting edges and masking node features. The second view introduces differentiated noise patterns by varying perturbation parameter settings, contrasting with the first view. Together, these two views enable the neural network model to distinguish true regulatory signals from noise in gene signatures.

[0071] In step S40, the first view and the second view are provided to the neural network model, and the incoming and outgoing features of each node in the gene subnetwork are obtained based on the comparative learning of the neural network model; through the dual-view comparative learning, the regulatory role features of the gene are separated and the causal relationship is captured. Among them, the regulatory role is the gene node as a regulatory gene or target gene. The regulatory gene refers to a gene that affects the expression level, transcript structure or functional activity of other genes by producing regulatory molecules or directly participating in the regulatory process. A target gene refers to a gene whose expression level, transcript structure or functional activity is directly or indirectly regulated by other genes. The target gene is the receiver and executor of the regulatory signal, and its dynamic changes directly determine the phenotype and biological function of the cell.

[0072] The neural network model uses a graph neural network (GNN), taking as input the topological structure and node features of two perturbation views. The incoming features describe the response characteristics of the gene when it is a target gene. The outgoing features describe the regulatory capacity of the gene when it is a regulator. By separating the dual roles of the gene and clarifying the regulatory direction, the causal ambiguity in traditional methods is resolved.

[0073] Step S50 constructs a regulation score matrix based on the incoming and outgoing features. The regulation score matrix displays the degree of regulatory association between genes in the transcriptome data of a single cell. The regulation score matrix quantifies the degree of regulatory association between genes, for example, by assigning a score to the degree of regulatory association, where a higher score indicates a stronger likelihood of regulation.

[0074] In this embodiment, by introducing prior gene network data, gene subnetworks are extracted from transcriptome data to construct a preliminary causal relationship framework for transcriptome data. With the help of a first random deletion strategy and a second random deletion strategy, technical noise is effectively suppressed and non-causal associations are reduced. On this basis, the neural network model obtains the incoming and outgoing features of each node in the gene subnetwork through comparative learning based on the first view and the second view. The regulatory score matrix thus formed can display the degree of regulatory association between genes in the transcriptome data of a single cell.

[0075] See Figure 2 As shown, in one embodiment of the present application, before the step of perturbing the gene sub-network according to the first random deletion strategy, the following steps are included:

[0076] In step S301, the nodes of the gene sub-network are screened according to a preset threshold value to obtain core nodes whose node degree is greater than the preset threshold value. Node degree refers to the number of edges connected to each node in the gene sub-network. Core nodes can also be understood as specific substructures with high centrality. There are two methods for setting the preset threshold strategy. The first is the absolute threshold method: directly setting a fixed value (such as node degree ≥ 50), which is suitable for scenarios with large differences in network scale. The second is the relative threshold method: based on the node degree distribution setting, it adapts to different network densities, such as top 10% or twice the median.

[0077] See Figure 3 As shown, in one embodiment of the present application, the node degree represents the number of edges of a node. The step of perturbing the gene subnetwork according to the first random deletion strategy includes:

[0078] In step S310, edges in the gene subnetwork are randomly removed with probability a1, and node features in the gene subnetwork are randomly masked with probability b1 to obtain a first view. This first view is obtained by global perturbation. All edges are randomly removed with probability a1 to simulate technical noise common in single-cell data, such as relationship loss caused by insufficient sequencing depth. Node features are randomly masked with probability b1 to simulate undetected gene expression. This allows the neural network model to extract stable features from a "sparse and noisy" network.

[0079] The step of perturbing the gene sub-network according to the second random deletion strategy includes:

[0080] In step S320, edges around core nodes are randomly removed with a probability of a2, features around core nodes are randomly masked with a probability of b2, edges outside of core nodes in the gene subnetwork are randomly removed with a probability of a3, and features around nodes outside of core nodes in the gene subnetwork are randomly masked with a probability of b3, to obtain a second view. The second view preferentially perturbs edges around core nodes. For example, edges around core nodes are removed with a higher probability of a2, disrupting the local structure of highly connected nodes.

[0081] The characteristics of the core nodes are masked with a higher probability b2 to simulate the functional loss of key regulatory factors.

[0082] In addition, non-core node perturbations, i.e., perturbations outside of the core nodes, are performed with lower probabilities a3 and b3, keeping the peripheral structure relatively intact. By selectively disrupting the connections of core regulatory nodes, the neural network model learns the robust feature that "even if key regulatory factors are perturbed, downstream target genes can still be regulated through alternative pathways," highlighting the role of indirect regulatory relationships.

[0083] Furthermore, a1=a3<a2, b1=b3<b2.

[0084] By using a1<a2 and b1<b2, we prioritize perturbations of core nodes, simulating the effects of loss of core gene function in disease or drug intervention scenarios. By using a1=a3 and b1=b3, we maintain consistent perturbation strength for non-core nodes, ensuring that the network's infrastructure is not overly damaged.

[0085] When the core node is highly disturbed, if the downstream genes still maintain stable expression, it suggests the existence of redundant regulatory pathways (such as bypass activation), and such relationships are more likely to be causal.

[0086] See Figure 4 As shown, in one embodiment of the present application, the step of providing the first view and the second view to the neural network model includes:

[0087] In step S410, the first view and the second view are embedded into the neural network model according to the incoming direction and the outgoing direction, respectively; the incoming direction represents the controlled party, i.e., the target gene; the outgoing direction represents the controlling party, i.e., the regulatory gene.

[0088] Step S420: The neural network model is compared and learned based on the first view and the second view.

[0089] Separating the directionality of gene regulation separates the "regulator" and "regulated" roles of genes, avoiding the ambiguity of causal direction caused by the aliasing of features between the two in traditional methods. Directional separation amplifies the differences in the effects of dual-view perturbations on incoming and outgoing features.

[0090] Direct signal of the first view: The edges retained under global perturbation are more likely to represent direct regulatory relationships, strengthening the "direct regulatory ability" in the outgoing features.

[0091] Indirect signals from the second view: relationships that persist after perturbations of core nodes are more likely to be indirectly regulated, highlighting the “robust response pattern” in incoming features.

[0092] Decompose gene signatures into two dimensions: regulators and regulated entities. Utilize dual-view perturbation differences to enhance the distinction between causal signals and noise.

[0093] The dual-view perturbation strategy is transformed into a computable causal feature extraction process. This design not only addresses the limitations of traditional GNNs in inferring regulatory directions, but also significantly improves the accuracy of causal relationship inference through perturbation parameters guided by biological priors.

[0094] See Figure 5 As shown, in one embodiment of the present application, the step of embedding the first view and the second view into the neural network model includes:

[0095] In step S411, the outgoing direction node features of the first view, the incoming direction node features of the first view, the outgoing direction node features of the second view, and the incoming direction node features of the second view are embedded into the neural network model; in step S411, the dual-view perturbation strategy is deeply integrated with the direction-specific feature extraction through four-way feature separation embedding into the neural network model to achieve a detailed analysis of the regulatory relationship.

[0096] Genes play both emissive and afferent roles in regulatory networks. For example, transcription factors serve as both targets of upstream signals and regulators of downstream genes. Using a four-way decomposition, we systematically separate direct and indirect signals, as well as stabilizing and compensatory mechanisms, within gene regulatory networks. This design significantly improves the accuracy and interpretability of regulatory relationship inference through multi-dimensional comparative learning. In biomedical applications, four-way features can accurately locate disease-driving genes and predict drug response patterns, providing a theoretical foundation for precision medicine.

[0097] The outgoing direction node feature of the first view is

[0098] The incoming direction node feature of the first view is

[0099] The outgoing direction node feature of the second view is

[0100] The incoming direction node feature of the second view is

[0101] in, is the characteristic matrix after perturbation, is the perturbed node adjacency matrix, f is the graph convolutional network encoder that aggregates high-order neighbor features, and h (·,·) is the node feature;

[0102]

[0103] is the Laplace matrix, D is The corresponding degree matrix, I n is the unit diagonal matrix, ‖ j∈P (·) is the concatenation operation, σ is the activation function, l is the number of layers, is the (l-1) layer network parameter matrix, m is the power, P∈{0,1,2};

[0104] The input feature is h in , the outgoing characteristic is h out , h out =f(X,A out ), h in =f(X,A in ).

[0105] See Figure 6 As shown, in one embodiment of the present application, the step of embedding the first view and the second view into the contrastive learning neural network model includes: steps S401 to S404, through which the contrastive learning framework systematically constructs positive and negative sample pairs to achieve unsupervised learning of gene regulatory features.

[0106] Step S401: Consider the same gene in the first and second views as a positive pair; this completes the construction of the positive pair. The outgoing and incoming features of the same gene in different perturbation views should maintain internal consistency, even if the network structure is partially disrupted. Considering features in the same direction (e.g., both are outgoing features) as a positive pair avoids confusion between the roles of regulator and regulated.

[0107] Step S402: Consider different genes with the same or different roles in the first and / or second views as negative sample pairs; this completes the construction of negative sample pairs. The regulatory patterns of different genes should be significantly different, even under similar perturbation conditions. By sampling across views and directions, the model learns more fundamental differences in gene signatures. Selecting genes with similar expression levels but distinct functions as the current gene improves the model's discriminative capabilities.

[0108] In step S403 , the positive sample pairs and the negative sample pairs are input into the neural network model together; the samples of the first view and the second view are organized into batches respectively, and features are extracted through a shared encoder.

[0109] Step S404 optimizes the model parameters by maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs. Steps S401 to S404 transform biological priors into computable optimization targets through a contrastive learning framework. Through the designed positive and negative sample pair construction strategy, the features learned by the model can not only reflect the regulatory ability of genes, but also retain their response characteristics, while effectively suppressing co-expression noise. This directionally separated contrastive learning mechanism provides feature representations with clear biological significance for subsequent regulatory network inference, significantly improving the accuracy of causal inference.

[0110] Among them, it should be noted that the loss function of the neural network of this application is designed based on the contrastive learning strategy. The positive sample is the representation of the same role of the same gene under different perspectives, while the negative sample is the representation of different genes under the same and different perspectives. This application hopes that the node similarity of the positive sample pair is high, while the node similarity of the negative sample pair is low. Because each role of each gene has two node embeddings under two different perturbation perspectives, the two embeddings are used as anchor points to calculate the contrastive learning objective function, and the different roles of the genes are taken into account, and then the final loss function is obtained by adding them together. It is as follows:

[0111]

[0112]

[0113] in, is the overall loss function of the model, g(·) is a nonlinear mapping, specifically a two-layer multilayer perceptron. cos(·,·) is the cosine similarity. τ is the temperature coefficient to control the impact of similarity changes. 1 k≠i ∈{0,1} is the indicator function.

[0114] In one embodiment of the present application, the control score matrix is ​​defined as M ij , M ij Represents the degree of regulatory association between the i-th and j-th genes; the regulatory score matrix M ij It is the core output of integrating gene regulatory characteristics with prior knowledge, and its definition combines the bidirectional regulatory ability of genes and biological prior constraints.

[0115] M ij =P ij *h out *h in , where P ij is a hyperparameter; h out Represents the outgoing characteristics of the gene, indicating its regulatory ability as a regulator. in Represents the incoming characteristics of the gene, indicating its response sensitivity as a target gene. ijQuantify the degree of regulatory association between gene i and gene j. A larger value indicates a more likely causal relationship between gene i and gene j. The above formula can be understood as: Regulatory association degree = regulatory ability × response sensitivity.

[0116] It not only provides quantitative indicators of regulatory associations between genes, but also, through directional sensitivity and noise filtering mechanisms, provides a tool that is both accurate and interpretable for analyzing complex gene regulatory networks. It has broad application prospects in disease mechanism research and precision medicine.

[0117] See Figure 7 As shown, in one embodiment of the present application, before the step of extracting the gene subnetwork related to the transcriptome data from the prior gene network data, the following steps are included:

[0118] Step 201, remove the intercellular ligand-receptor interaction data in the prior gene network data; ligand-receptor interaction data is the interaction data between cells. This application requires data within a single cell. By removing the intercellular ligand-receptor interaction data, the data required for intracellular causal inference is retained.

[0119] Step 202: Normalize and / or log-transform the transcriptome data for the single cell. Differences in sequencing depth can lead to fluctuations in gene detection rates. Normalization and log-transformation eliminate technical variation and preserve true biological differences.

[0120] See Figure 8 As shown in the figure, these two heat maps show the performance of different methods in gene regulation related analysis, mainly involving early prediction rate (EPR) and precision-recall area under the curve ratio (AUPR ratio) data.

[0121] Specifically, the heat map is used to intuitively display the numerical performance of different methods under different test requirements. The depth of the color represents the size of the value, and the lighter the color, the larger the value. Figure 8 Involving the analysis of transcription factors (TFs) and 500 genes, the rows represent different test requirements, and the columns represent different test methods.

[0122] The heatmap on the left shows the early prediction rates (EPRs) of transcription factors and 500 genes. EPRs measure the potential regulatory ability of genes in regulatory networks.

[0123] STRING: A database and tool for searching known and predicted protein-protein interactions.

[0124] Non-specific ChIP-seq: Non-specific chromatin immunoprecipitation sequencing, used to study protein-DNA interactions.

[0125] Cell-type-specific ChIP-seq: Cell-type-specific chromatin immunoprecipitation sequencing can more accurately study protein-DNA interactions in specific cell types. Figure 8 The right column of the left figure uses the inference method of this application. It has a light color and generally high scores, which can more accurately reflect the causal relationship between genes.

[0126] Figure 8 The heat map on the right shows the area under the precision-recall curve ratio (AUPRratio) of transcription factors and 500 genes. AUPR is often used to evaluate the accuracy of prediction results in tasks such as gene regulatory network inference. The ratio value can reflect the relative performance of different methods. Figure 8 The right column of the right figure uses the inference method of this application. The color is light and the scores are generally high, which can more accurately reflect the regulatory relationship between genes.

[0127] See Figure 9 As shown, the present application also provides an inference system for a gene regulatory network, and the inference system includes: a data acquisition module 100, a preprocessing module 200, a neural network model learning module 300 and an inference module 400.

[0128] Data acquisition module 100 is used to obtain single-cell transcriptome data and prior gene network data. Single-cell transcriptome data is derived from single-cell RNA sequencing, recording the expression levels of genes in individual cells and reflecting the dynamic activity of genes at the transcriptional level. Prior gene network data is a gene interaction network based on literature, databases, or experimental verification, containing known gene regulatory relationships, such as the binding of transcription factors to target genes. By introducing prior knowledge, biological constraints are provided for regulatory network inference, reducing unfounded random associations.

[0129] Preprocessing module 200 is used to extract gene subnetworks related to transcriptome data from the prior gene network data, retaining isolated gene nodes that do not appear in the gene subnetwork. Gene sets that match the transcriptome data of a single cell are screened from the prior gene network data to construct an initial gene subnetwork framework.

[0130] Genes that do not appear in the gene subnetwork are retained as isolated nodes to avoid missing potential regulatory genes due to the limitations of prior knowledge. These isolated gene nodes can be used to infer their regulatory relationships through data-driven methods in subsequent analysis.

[0131] The neural network model learning module 300 is used to perturb the gene subnetwork according to a first random deletion strategy to obtain a first view, and to perturb the gene subnetwork according to a second random deletion strategy to obtain a second view, wherein the first random deletion strategy and the second random deletion strategy are different; the first view and the second view are provided to the neural network model, and the incoming and outgoing features of each node in the gene subnetwork are obtained based on comparative learning of the neural network model. The first view simulates the technical noise commonly present in single-cell data, such as sequencing errors, by randomly deleting edges and masking node features. The second view introduces differentiated noise patterns by setting different perturbation parameters, which contrast with the first view. These two views enable the neural network model to distinguish between true regulatory signals and noise in gene features. Through dual-view comparative learning, the regulatory role characteristics of genes are separated and causal relationships are captured.

[0132] Inference module 400 is used to construct a regulatory score matrix based on the incoming and outgoing features. The regulatory score matrix displays the degree of regulatory association between genes in the transcriptome data of a single cell. The regulatory score matrix quantifies the degree of regulatory association between genes, for example, assigning a score to the degree of regulatory association, where a higher score indicates a stronger likelihood of regulation.

[0133] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and description of the present invention.

Claims

1. A method for inferring a gene regulatory network, characterized in that: The inference method includes: Obtain single-cell transcriptome data and prior gene network data; Extracting a gene subnetwork related to the transcriptome data from the prior gene network data, wherein isolated gene nodes that do not appear in the gene subnetwork are retained; The gene sub-network is perturbed according to a first random deletion strategy to obtain a first view, and the gene sub-network is perturbed according to a second random deletion strategy to obtain a second view, where the first random deletion strategy and the second random deletion strategy are different; Providing the first view and the second view to a neural network model, and obtaining incoming features and outgoing features of each node in the gene subnetwork based on comparative learning of the neural network model; A regulation score matrix is ​​constructed based on the incoming features and the outgoing features, and the regulation score matrix shows the degree of regulatory association between genes in the transcriptome data of the single cell.

2. The inference method according to claim 1, characterized in that Before the step of perturbing the gene sub-network according to the first random deletion strategy, the method includes: The nodes of the gene sub-network are screened according to a preset threshold value to obtain core nodes whose node degrees are greater than the preset threshold value.

3. The inference method according to claim 2, characterized in that The node degree represents the number of edges of a node; The step of perturbing the gene sub-network according to the first random deletion strategy comprises: Randomly removing edges in the gene sub-network with a probability of a1, and randomly masking node features in the gene sub-network with a probability of b1, to obtain a first view; The step of perturbing the gene sub-network according to the second random deletion strategy comprises: The edges of the core nodes are randomly removed with a probability of a2, the features of the core nodes are randomly masked with a probability of b2, the edges outside the core nodes in the gene sub-network are randomly removed with a probability of a3, and the node features outside the core nodes in the gene sub-network are randomly masked with a probability of b3 to obtain a second view.

4. The inference method according to claim 3, characterized in that a1=a3<a2, b1=b3<b2.

5. The inference method according to claim 1, characterized in that The step of providing the first view and the second view to a neural network model comprises: Embedding the first view and the second view into a neural network model according to an incoming direction and an outgoing direction, respectively; The neural network model is learned based on comparison between the first view and the second view.

6. The inference method according to claim 5, characterized in that The step of embedding the first view and the second view into a neural network model comprises: embedding the outgoing direction node features of the first view, the incoming direction node features of the first view, the outgoing direction node features of the second view, and the incoming direction node features of the second view into the neural network model; The outgoing direction node feature of the first view is The incoming direction node feature of the first view is The outgoing direction node feature of the second view is The incoming direction node feature of the second view is in, is the characteristic matrix after perturbation, is the perturbed node adjacency matrix, f is the graph convolutional network encoder that aggregates high-order neighbor features, and h (·,·) is the node feature; is the Laplace matrix, D is The corresponding degree matrix, I n is the unit diagonal matrix, ‖ j∈P (·) is the concatenation operation, σ is the activation function, l is the number of layers, is the (l-1) layer network parameter matrix, m is the power, P∈{0,1,2}; The incoming feature is h in , the outgoing feature is h out , h out =f(X,A out ),h in =f(X,A in ).

7. The inference method according to claim 6, characterized in that: The step of embedding the first view and the second view into the neural network model comprises: The same role of the same gene in the first view and the second view is taken as a positive sample pair; taking the same or different roles of different genes in the first view and / or the second view as negative sample pairs; Inputting the positive sample pair and the negative sample pair into the neural network model; The model parameters are optimized by maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs.

8. The inference method according to claim 6, characterized in that Define the control score matrix as M ij , M ij represents the degree of regulatory association between the i-th and j-th genes; M ij =P ij *h out *h in , where P ij is a hyperparameter; 9. The inference method according to claim 1, characterized in that: Before the step of extracting a gene subnetwork related to the transcriptome data from the prior gene network data, the method includes: Eliminating intercellular ligand-receptor interaction data from the prior gene network data; The transcriptome data of the single cell is normalized and / or logarithmically transformed.

10. A gene regulatory network inference system, characterized in that: The inference system comprises: Data acquisition module, used to obtain transcriptome data and prior gene network data of single cells; a preprocessing module, configured to extract a gene subnetwork related to the transcriptome data from the prior gene network data, wherein isolated gene nodes that do not appear in the gene subnetwork are retained; a neural network model learning module, configured to perturb the gene subnetwork according to a first random deletion strategy to obtain a first view, and perturb the gene subnetwork according to a second random deletion strategy to obtain a second view, wherein the first random deletion strategy and the second random deletion strategy are different; provide the first view and the second view to a neural network model, and obtain incoming features and outgoing features of each node in the gene subnetwork based on comparative learning of the neural network model; An inference module is used to construct a regulation score matrix based on the incoming features and the outgoing features, wherein the regulation score matrix shows the degree of regulatory association between genes in the transcriptome data of the single cell.

Images