Method for inferring gene regulatory network and inferring system thereof

Abstract

This invention discloses a method and system for inferring gene regulatory networks. The inference method includes: acquiring single-cell transcriptome data and prior gene network data; extracting gene sub-networks related to the transcriptome data from the prior gene network data; obtaining a first view and a second view by applying a first random deletion strategy and a second random deletion strategy to the gene sub-networks; providing the first and second views to a neural network model, and obtaining the input and output features of each node in the gene sub-network based on comparative learning of the neural network model; constructing a regulatory score matrix based on the input and output features, which displays the degree of regulatory association between genes in the single-cell transcriptome data. The technical solution of this application can effectively distinguish between direct causal relationships and indirect association relationships, enabling gene co-expression networks to more accurately reflect the true regulatory relationships.

Description

Technical Field

[0001] This invention relates to the field of biological detection technology, specifically to a method and system for inferring gene regulatory networks. Background Technology

[0002] In life science research, genes do not exist in isolation, but rather work synergistically through interactions and regulatory relationships. Gene regulation is a fundamental molecular mechanism involved in almost all aspects of life. It forms the basis of gene expression, encompasses the regulation of transcription and translation processes, and is crucial for important biological processes such as development, differentiation, and the occurrence and development of diseases. It can be said that gene regulation is the core of life activities.

[0003] However, single-cell transcriptome data lacks information that directly reflects transcriptional regulatory mechanisms. Moreover, existing single-cell transcriptome sequencing technologies suffer from factors such as depth differences and uneven cell capture efficiency, which can introduce false correlation information during sequencing, making it difficult to distinguish between direct causal relationships and indirect associations. This makes it impossible for gene co-expression networks to accurately reflect the true regulatory relationships. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for inferring gene regulatory networks that can effectively distinguish between direct causal relationships and indirect associations, enabling gene co-expression networks to more accurately reflect the true regulatory relationships.

[0005] This application provides an inference method, the inference method comprising:

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

[0007] Extract gene subnetworks related to the transcriptome data from the prior gene network data, wherein isolated gene nodes that do not appear in the gene subnetworks are retained.

[0008] The gene subnetwork is perturbed according to a first random deletion strategy to obtain a first view, and the gene subnetwork is perturbed 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;

[0009] The first view and the second view are provided to the neural network model, and the input features and output features of each node in the gene sub-network are obtained by comparative learning based on the neural network model.

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

[0011] In one aspect, prior to the step of perturbing the gene subnetwork according to a first random deletion strategy, the following is included:

[0012] The nodes of the gene subnetwork are filtered according to a preset threshold to identify the core nodes whose node degree is greater than the preset threshold.

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

[0014] The step of perturbing the gene subnetwork according to the first random deletion strategy includes:

[0015] 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;

[0016] The step of perturbing the gene subnetwork according to the second random deletion strategy includes:

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

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

[0019] In one aspect, the step of providing the first view and the second view to the neural network model includes:

[0020] The first view and the second view are embedded into the neural network model according to the input direction and the output direction, respectively;

[0021] The neural network model learns by comparing 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 includes:

[0023] 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.

[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, The perturbation characteristic matrix, Let f be the perturbed node adjacency matrix, f be the graph convolutional network encoder that aggregates high-order neighbor features, and h be the node adjacency matrix. (·,·) For node features;

[0029]

[0030] Let D be the Laplace matrix. The corresponding degree matrix, I n For a unit diagonal matrix, || j∈P (·) represents the concatenation operation, σ is the activation function, and l is the number of layers. Let m be the parameter matrix of the (l-1) layer network, where m is the power and P∈{0,1,2}.

[0031] The input feature is h in The outgoing characteristic 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 includes:

[0033] The same gene playing the same role in the first view and the second view is considered a positive sample pair;

[0034] Different genes playing the same or different roles in the first view and / or the second view are used as negative sample pairs;

[0035] The positive sample pairs and the negative sample pairs are input together 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 regulation score matrix is ​​defined as M. ij M ij This 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 For hyperparameters;

[0039]

[0040] In one aspect, prior to the step of extracting gene subnetworks associated with the transcriptome data from the prior gene network data, the method includes:

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

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

[0043] To address the aforementioned problems, this application also provides a gene regulatory network inference system, the inference system comprising:

[0044] The data acquisition module is used to acquire transcriptome data and prior gene network data from single cells;

[0045] A preprocessing module is used to extract gene subnetworks related to the transcriptome data from the prior gene network data, wherein isolated gene nodes that do not appear in the gene subnetworks are retained;

[0046] The neural network model learning module is used to perturb the gene sub-network according to a first random deletion strategy to obtain a first view, and to 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; the first view and the second view are provided to the neural network model, and the input features and output features of each node in the gene sub-network are obtained by comparative learning based on the neural network model;

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

[0048] The beneficial effects of this invention are reflected in the following aspects: by introducing prior gene network data, gene sub-networks are extracted from transcriptome data, constructing a preliminary causal framework for transcriptome data. Using a first and a second random deletion strategy, technical noise is effectively suppressed and non-causal associations are reduced. Based on this, the neural network model, based on the first and second views, acquires the input and output features of each node in the gene sub-network through comparative learning. The resulting regulatory score matrix can then display the degree of regulatory association between genes in single-cell transcriptome data. Attached Figure Description

[0049] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0050] Figure 1 This is a schematic diagram of the steps involved in the inference method for the gene regulatory network in this application;

[0051] Figure 2 This is a schematic diagram illustrating the steps for obtaining core nodes in the gene regulatory network inference method of this application;

[0052] Figure 3 This is a schematic diagram illustrating the steps of obtaining the first and second views in the gene regulatory network inference method of this application;

[0053] Figure 4 This is a schematic diagram illustrating the steps of training a neural network model using a first view and a second view in the gene regulatory network inference method of this application.

[0054] Figure 5 For this application Figure 4 A schematic diagram illustrating the steps of training a neural network model in the input and output directions.

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

[0056] Figure 7 This diagram illustrates the steps of removing intercellular data, normalization, and logarithmic transformation in the gene regulatory network inference method of this application.

[0057] Figure 8 This diagram illustrates the comparison between the gene regulatory network inference method of this application and other methods.

[0058] Figure 9 This is a schematic diagram of the inference system of the gene regulatory network in this application.

[0059] Figure descriptions: 100, Data acquisition module; 200, Preprocessing module; 300, Neural network model learning module; 400, Inference module. Detailed Implementation

[0060] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0061] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by those skilled in the art to which this invention pertains.

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

[0063] Inference methods include:

[0064] Step S10 involves acquiring single-cell transcriptome data and prior gene network data; these data form the basis for analysis and inference. Single-cell transcriptome data is derived from single-cell RNA sequencing (scRNA-seq), recording gene expression levels in individual cells and reflecting dynamic gene activity at the transcriptional level. Prior gene network data consists of gene interaction networks validated in literature, databases (such as NicheNet), or experiments, containing known gene regulatory relationships, such as the binding of transcription factors to target genes. Introducing prior knowledge provides biological constraints for regulatory network inference, reducing unfounded random associations.

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

[0066] For example, genes that coexist with single-cell transcriptome data can be extracted from a priori gene networks to form a gene subnetwork G = (V, E), where V is a gene node and E is a known regulatory edge.

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

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

[0069] Step S30: Perturb the gene subnetwork according to the first random deletion strategy to obtain a first view; perturb the gene subnetwork according to the second random deletion strategy to obtain a second view. The first and second random deletion strategies are different. Apply two different random deletion strategies to the gene subnetwork to generate two perturbation views, the first view and the second view, and use the differences between the views to enhance the contrast learning signal.

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

[0071] Step S40: The first and second views are provided to the neural network model. Based on the comparative learning of the neural network model, the input and output features of each node in the gene sub-network are obtained. Through comparative learning of the two views, the regulatory role features of genes are separated, and causal relationships are captured. Here, the regulatory role refers to a gene node acting as a regulatory gene or a target gene. A regulatory gene is a gene that influences the expression level, transcription product structure, or functional activity of other genes by producing regulatory molecules or directly participating in the regulatory process. A target gene is a gene whose expression level, transcription product structure, or functional activity is directly or indirectly regulated by other genes. Target genes are the receivers and executors of regulatory signals, and their dynamic changes directly determine the phenotype and biological function of the cell.

[0072] The neural network model employs a graph neural network (GNN), with inputs consisting of the topology and node features of two perturbation views. The input features describe the response characteristics of the gene when it acts as a target gene. The output features describe the regulatory capacity of the gene when it acts as a regulatory gene. By separating the dual roles of genes, the direction of regulation is clarified, resolving the problem of ambiguity in causal direction in traditional methods.

[0073] Step S50: Construct a regulatory score matrix based on input and output features. The regulatory score matrix displays the degree of regulatory association between genes in single-cell transcriptome data. The regulatory score matrix quantifies the degree of regulatory association between genes, for example, it provides a score for the degree of regulatory association; the higher the score, the stronger the likelihood of regulation.

[0074] In this embodiment, prior gene network data is introduced to extract gene subnetworks from transcriptome data, constructing a preliminary causal framework for the transcriptome data. A first and a second random deletion strategy are employed to effectively suppress technical noise and reduce non-causal associations. Based on this, a neural network model, using both the first and second views, acquires the input and output features of each node in the gene subnetwork through comparative learning. The resulting regulatory score matrix can then display the degree of regulatory association between genes in single-cell transcriptome data.

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

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

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

[0078] Step S310 involves randomly removing edges from the gene subnetwork with probability a1 and randomly masking node features with probability b1 to obtain a first view; the first view is obtained through global perturbation. Randomly removing all edges with probability a1 simulates technical noise commonly found in single-cell data, such as relationship loss due to insufficient sequencing depth. Randomly masking node features with probability b1 simulates undetected gene expression. This allows the neural network model to extract stable features from a sparse and noisy network.

[0079] The steps of perturbing the gene subnetwork according to the second random deletion strategy include:

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

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

[0082] Furthermore, for perturbations of non-core nodes (i.e., nodes other than core nodes), the probability of perturbation is low, using a3 and b3, to maintain the relative integrity of the peripheral structure. By selectively disrupting the connections of core regulatory nodes, the neural network model learns the robust characteristic 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 ensuring a1 < a2 and b1 < b2, priority perturbation of core nodes is guaranteed, simulating the effect of core gene function loss under disease or drug intervention scenarios. By ensuring a1 = a3 and b1 = b3, the perturbation intensity of non-core nodes is kept consistent, ensuring that the network's basic structure is not excessively damaged.

[0085] If downstream genes remain stably expressed when core nodes are subjected to high-intensity perturbation, 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 this 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 input direction and the output direction, respectively; the input direction represents the controlled party, i.e., the target gene; the output direction represents the controlling party, i.e., the regulatory gene.

[0088] Step S420: The neural network model learns based on the comparison between the first view and the second view.

[0089] Separating the directionality of gene regulation separates the roles of "regulator" and "regulated" genes, avoiding the ambiguity of causal direction caused by the overlap of features between the two in traditional methods. Two-view perturbations have different effects on afferent / efferent features; this difference can be amplified through directional separation.

[0090] Direct signals from the first view: Edges preserved under global perturbations are more likely to represent direct control relationships, strengthening the "direct control capability" in the outgoing features.

[0091] Indirect signals in the second view: Relationships that persist after core node disturbances are more likely to be indirectly regulated, highlighting the "robust response mode" in the input characteristics.

[0092] Genetic characteristics are decomposed into two dimensions: regulators and regulated entities. The difference between causal signals and noise is enhanced by utilizing the perturbation difference of the two views.

[0093] The dual-view perturbation strategy is transformed into a computable causal feature extraction process. This design not only overcomes 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 this application, the step of embedding the first view and the second view into a neural network model includes:

[0095] Step S411: Embed 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; Step S411: By embedding the four-way feature separation into the neural network model, the dual-view perturbation strategy and the direction-specific feature extraction are deeply integrated to achieve a fine analysis of the control relationship.

[0096] Genes simultaneously act as both efferent and afferent signalers in regulatory networks. For example, transcription factors are both target genes of upstream signals and regulators of downstream genes. Through four-dimensional decomposition, direct and indirect signals, as well as stable and compensatory mechanisms, within gene regulatory networks are systematically separated. This design, through multi-dimensional comparative learning, significantly improves the accuracy and interpretability of regulatory relationship inferences. In biomedical applications, four-dimensional features can precisely locate disease-driving genes and predict drug response patterns, providing a theoretical basis for precision medicine.

[0097] The outgoing direction node features of the first view are:

[0098] The characteristics of the incoming direction node in the first view are:

[0099] The outgoing direction node features of the second view are:

[0100] The characteristics of the incoming direction node in the second view are:

[0101] in, The perturbation characteristic matrix, Let f be the perturbed node adjacency matrix, f be the graph convolutional network encoder that aggregates high-order neighbor features, and h be the node adjacency matrix. (·,·) For node features;

[0102]

[0103] Let D be the Laplace matrix. The corresponding degree matrix, I n For a unit diagonal matrix, || j∈P (·) represents the concatenation operation, σ is the activation function, and l is the number of layers. Let m be the parameter matrix of the (l-1) layer network, where m is the power and 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 this application, the step of embedding the first view and the second view into the contrastive learning neural network model includes: steps S401 to S404, systematically constructing positive and negative sample pairs through the contrastive learning framework in steps S401 to S404 to achieve unsupervised learning of gene regulatory features.

[0106] Step S401: Pair the same gene with the same role in both the first and second views as positive samples; this completes the construction of positive sample pairs. The efferent and afferent features of the same gene in different perturbation views should maintain inherent consistency, even if the network structure is partially disrupted. Pair features in the same direction (e.g., both being efferent features) as positive samples to avoid confusing regulator and regulated roles.

[0107] Step S402: Use the same or different roles of different genes in the first and / or second views as negative sample pairs; complete the construction of negative sample pairs. The regulatory patterns of different genes should have significant differences, even under similar perturbation conditions. By sampling across views and directions, the model learns more fundamental differences in gene characteristics. Selecting genes with similar expression levels but different functions as negative samples improves the model's discriminative ability.

[0108] Step S403: Input positive and negative sample pairs into the neural network model; organize the samples from the first view and the second view into batches and extract features 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–S404, through a contrastive learning framework, transform biological priors into computable optimization objectives. Through a designed strategy for constructing positive and negative sample pairs, the features learned by the model reflect both the regulatory capacity of genes and their response characteristics, while effectively suppressing co-expression noise. This directional contrastive learning mechanism provides biologically meaningful feature representations for subsequent regulatory network inference, significantly improving the accuracy of causal relationship inference.

[0110] It should be noted that the loss function of the neural network in this application is designed based on a contrastive learning strategy. Positive samples represent the same role of the same gene under different perspectives, while negative samples represent different genes under the same and different perspectives. This application aims for high node similarity in positive sample pairs and low node similarity in negative sample pairs. Because each role of each gene has two node embeddings under two different perturbation perspectives, the two embeddings are used as anchors to calculate the contrastive learning objective function, and the different roles of the genes are considered before being summed to obtain the final loss function. The details are as follows:

[0111]

[0112]

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

[0114] In one embodiment of this application, the control score matrix is ​​defined as M. ij M ij The regulatory score matrix M represents the degree of regulatory association between the i-th and j-th genes. ij It is the core output that integrates gene regulatory characteristics with prior knowledge. Its definition combines the bidirectional regulatory capacity of genes with biological prior constraints.

[0115] M ij =P ij *h out *h in , where P ij h is a hyperparameter. out The efferent characteristics of a gene represent its regulatory capacity as a regulator. in The input characteristics of a gene represent its sensitivity as a target gene. M ijThe degree of regulatory association between gene i and gene j is quantified; a larger value indicates a more likely causal relationship between gene i and gene j. The formula above can be understood as: Degree of regulatory association = Regulatory capacity × Response sensitivity.

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

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

[0118] Step 201: Remove intercellular ligand-receptor interaction data from the prior gene network data; ligand-receptor interaction data is intercellular interaction data, and this application requires data within a single cell. By removing intercellular ligand-receptor interaction data, the data required for intracellular causal inference is retained.

[0119] Step 202 involves normalizing and / or logarithmically transforming the single-cell transcriptome data. Differences in sequencing depth cause fluctuations in gene detection rates. Normalization and logarithmic transformation eliminate technical variations, preserving true biological differences.

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

[0121] Specifically, heatmaps are used to visually display the numerical performance of different methods under different test requirements. The depth of the color represents the magnitude of the value, with lighter colors representing larger values. Figure 8 The analysis involves transcription factors (TFs) and 500 related genes. Rows represent different test requirements, and columns represent different test methods.

[0122] The heatmap on the left shows the early prediction rate (EPR) of transcription factors and 500 genes. EPR measures a gene's potential regulatory capacity in a regulatory network.

[0123] STRING: A database and tool for retrieving known and predicting protein-protein interactions.

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

[0125] Cell-type-specific ChIP-seq: Cell type-specific chromatin immunoprecipitation sequencing enables more precise study of protein-DNA interactions in specific cell types. Figure 8 The right column of the left-hand graph shows the inference method used in this application. The lighter colors generally indicate higher scores, which can more accurately reflect the causal relationship between genes.

[0126] Figure 8 The heatmap on the right shows the area under the precision-recall curve (AUPR ratio) for transcription factors and 500 genes. AUPR is commonly used to evaluate the accuracy of predictions in tasks such as gene regulatory network inference; the ratio reflects the relative performance of different methods. It can be seen that... Figure 8 The right column of the right-hand figure shows the inference method used in this application. The lighter colors generally indicate higher scores, which can more accurately reflect the regulatory relationships between genes.

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

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

[0129] The preprocessing module 200 is used to extract gene subnetworks related to transcriptome data from prior gene network data, wherein isolated gene nodes not appearing in the gene subnetworks are retained. A set of genes matching single-cell transcriptome data is screened from the prior gene network data to construct an initial gene subnetwork framework.

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

[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 then perturb the gene subnetwork according to a second random deletion strategy to obtain a second view. The first and second random deletion strategies are different. The first and second views are provided to the neural network model, and the input and output features of each node in the gene subnetwork are obtained based on the comparative learning of the neural network model. Specifically, the first view simulates technical noise commonly found in single-cell data, such as sequencing errors, by randomly deleting edges and masking node features. The second view introduces differentiated noise patterns through different perturbation parameter settings, contrasting with the first view. The two views enable the neural network model to distinguish between real regulatory signals and noise in gene features. Through comparative learning of the two views, the regulatory role features of genes are separated, and causal relationships are captured.

[0132] The inference module 400 is used to construct a regulatory score matrix based on input and output features. The regulatory score matrix displays the degree of regulatory association between genes in single-cell transcriptome data. The regulatory score matrix quantifies the degree of regulatory association between genes, for example, it gives a score of the degree of regulatory association; the higher the score, the stronger the 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, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for inferring gene regulatory networks, characterized in that, The inference method includes: Acquire single-cell transcriptome data and prior gene network data; Extract gene subnetworks related to the transcriptome data from the prior gene network data, wherein isolated gene nodes that do not appear in the gene subnetworks are retained. The gene subnetwork is perturbed according to a first random deletion strategy to obtain a first view, and the gene subnetwork is perturbed 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 input features and output features of each node in the gene sub-network are obtained by comparative learning based on the neural network model. A regulatory score matrix is ​​constructed based on the input and output features, and the regulatory score matrix shows the degree of regulatory association between genes in the transcriptome data of the single cell. The step of providing the first view and the second view to the neural network model includes: The first view and the second view are embedded into the neural network model according to the input direction and the output direction, respectively; The neural network model learns by comparing the first view and the second view; The step of embedding the first view and the second view into the neural network model includes: 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. 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, The perturbation characteristic matrix, Let f be the perturbed node adjacency matrix, and let f be the graph convolutional network encoder that aggregates high-order neighbor features. , , and All are node features; Let D be the Laplace matrix. The corresponding degree matrix, It is a unit diagonal matrix. For the concatenation operation, σ is the activation function, and l is the number of layers. Let m be the parameter matrix of the (l-1) layer network, where m is the power and P∈{0,1,2}. The input feature is The outgoing feature is , , .

2. The inference method according to claim 1, characterized in that, Before the step of perturbing the gene subnetwork according to the first random deletion strategy, the following steps are included: The nodes of the gene subnetwork are filtered according to a preset threshold to identify the core nodes whose node degree is greater than the preset threshold.

3. The inference method according to claim 2, characterized in that, The node degree represents the number of edges a node has; The step of perturbing the gene subnetwork according to the first random deletion strategy includes: 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; The step of perturbing the gene subnetwork according to the second random deletion strategy includes: Edges of the core node are randomly removed with probability a2, features of the core node are randomly masked with probability b2, edges outside the core node in the gene subnetwork are randomly removed with probability a3, and node features outside the core node in the gene subnetwork are randomly masked with probability 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 4, characterized in that, The step of embedding the first view and the second view into the neural network model includes: The same gene playing the same role in the first view and the second view is considered a positive sample pair; Different genes playing the same or different roles in the first view and / or the second view are used as negative sample pairs; The positive sample pairs and the negative sample pairs are input together 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.

6. The inference method according to claim 5, characterized in that, The regulation score matrix is ​​defined as follows: , This represents the degree of regulatory association between the i-th and j-th genes; ,in For hyperparameters; 。 7. The inference method according to claim 1, characterized in that, Prior to the step of extracting gene subnetworks related to the transcriptome data from the prior gene network data, the following steps are included: Remove intercellular ligand-receptor interaction data from the prior gene network data; The transcriptome data of the single cells are normalized and / or logarithmically transformed.

8. A system for inferring gene regulatory networks, characterized in that, For performing the inference method of a gene regulatory network as described in claim 1, the inference system comprises: The data acquisition module is used to acquire transcriptome data and prior gene network data from single cells; A preprocessing module is used to extract gene subnetworks related to the transcriptome data from the prior gene network data, wherein isolated gene nodes that do not appear in the gene subnetworks are retained; The neural network model learning module is used to perturb the gene sub-network according to a first random deletion strategy to obtain a first view, and to 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; the first view and the second view are provided to the neural network model, and the input features and output features of each node in the gene sub-network are obtained by comparative learning based on the neural network model; An inference module is used to construct a regulatory score matrix based on the input features and the output features, wherein the regulatory score matrix displays the degree of regulatory association between genes in the transcriptome data of the single cell.

Images