Single-sample gene regulatory network construction method based on mutual information

Abstract

The invention belongs to the field of system biology, and particularly relates to a single-sample gene regulatory network construction method based on mutual information. The method comprises the following steps: specifying a background group and a foreground group of a research object; calculating a single-sample gene regulation network of the background group and the foreground group based on a mutual information method; and carrying out significance test on the network edge by utilizing random permutation test. Compared with the prior art, the method has the advantages that the biological gene regulatory network can be identified based on the nonlinear relation, and actual biological characteristics are better met. According to the method, the gene regulation and control network of each sample can be specifically constructed, and diversified data analysis can be performed according to different selected background groups.

Description

Technical Field

[0001] The present invention belongs to the technical field of systems biology, and specifically relates to a method for constructing a single-sample gene regulatory network for bulk transcriptome and single-cell transcriptome data of multiple samples. Background Art

[0002] Gene regulatory networks (GRNs) play a crucial role in the study of biological systems and complex diseases. They elucidate how genomic sequences precisely control the spatial and temporal patterns of gene expression and specify cell type-specific functions and interactions. Experimental methods have helped us obtain a large number of regulatory relationships between transcription factors (TFs) and target genes. However, experimental techniques themselves have certain limitations, and people have not yet been able to reveal a complete picture.

[0003] Many computational methods have been developed to reconstruct GRNs from RNA sequencing (RNA-seq) and single-cell RNA sequencing (scRNA-seq) datasets. These algorithms can be roughly divided into two categories: one is to infer the aggregate network of the entire sample or cell type; the other is to construct sample-specific or cell-type-specific networks. The latter has become increasingly important with the development of single-cell technology. Recent advances in single-cell technology have enabled high-resolution studies of cellular heterogeneity, thereby providing a deeper understanding of cellular heterogeneity in different biological processes. Therefore, the application and development of sample-specific networks (SSNs) have become increasingly important. Liu et al. developed a single-sample-based SSN based on Pearson correlation, which is a linear method. Dai et al. first proposed a cell-specific network at the single-cell level, which preserved the heterogeneity between two cells by studying the nonparametric joint distribution of gene expression. Wang et al. improved the cell-specific network by considering the mixed effects of the cell population and developed an algorithm for local cell-specific networks. Li et al. developed the c-CSN method to address the overfitting effect of cell-specific networks on non-directly related genes. These studies demonstrate that single-sample networks have significant potential for characterizing highly heterogeneous biological systems.

[0004] However, obtaining SSNs alone may not be sufficient to reveal important changes in cell development and disease progression. Recently, differential networks have emerged as a more effective means of identifying these changes. However, current algorithms for constructing SSNs are unable to generate differential networks, except for the “SSN” based on Pearson correlation developed by Liu et al., which is only applicable to linear systems. The complex regulatory relationships between various biological phenomena, such as transcription factor regulation, receptor cascades, epigenetic modifications, genome editing, RNA editing, cell signaling pathways, and RNA interference, may all affect changes in gene expression levels to varying degrees. Therefore, the regulatory relationships in most biological systems are not simply linearly related. Summary of the invention

[0005] The purpose of the present invention is to provide a single-sample gene regulatory network construction method based on mutual information theory to overcome the defects in the above-mentioned prior art.

[0006] The single-sample gene regulatory network construction method based on mutual information provided by the present invention is abbreviated as diffMIN, and the specific steps are as follows:

[0007] S1: First, the gene expression data is preprocessed to remove genes and samples with low gene expression; then, the gene expression data is uniformly discretized;

[0008] S2: Before starting the calculation, according to the biological significance, a part of the samples are designated as background samples (assuming there are N1 samples), and the remaining samples are designated as foreground samples (assuming there are N2 samples); the initial difference network is obtained by calculating the perturbation of the foreground samples on the background samples;

[0009] S3: Calculate the mutual information correlation network of the background sample and the perturbation sample respectively; where any two genes in the sample X=(x1,x2,x3,...,x n ) and gene Y=(y1,y2,y3,...,y n )The formula for calculating the mutual information correlation network between is as follows:

[0010]

[0011] Where n represents the number of samples, p(x) and p(y) represent the marginal probabilities of X and Y respectively, and p(x,y) represents the joint probability of X and Y.

[0012] S4: Then, the mutual information network is normalized to obtain a standardized sample correlation network, and the normalization calculation formula is as follows:

[0013]

[0014] in, Respectively represent the information entropy of variables X and Y. The mutual information network calculated by the background sample is expressed as The mutual information network obtained by adding a foreground sample j (j = 1, 2, ..., N2) to the background sample is expressed as

[0015] S5: By calculating the difference between the perturbed network and the background network, the network changes caused by the perturbed sample are identified. The calculation formula is as follows:

[0016]

[0017] Among them, j represents the jth foreground sample to be calculated;

[0018] S6: For the calculation of the individual network of the background sample, an approximate method is used, but the difference is that each time a single sample i is removed from the background group, i = 1, 2, ..., N1, and the disturbance change of the overall network after removing the single sample is calculated; the calculation formula is as follows:

[0019]

[0020] Among them, i represents the i-th background sample to be calculated;

[0021] S7: Perform a significance test on the edges of the network using a random permutation test method to obtain the significance value of the edges of the network (called P value); construct a random network by randomly disrupting the gene distribution. The specific method is to keep the order of the samples unchanged, randomly disrupt the expression level of each gene, and construct a new random network while keeping its gene distribution unchanged;

[0022] S8: Calculate the difference network of the random network in the same way as above, and use the random permutation test method to perform a permutation test on the edges of the calculated real network.

[0023] Based on the above single sample gene regulatory network construction method, the present invention also includes a single sample regulatory network construction system. The system specifically includes 8 modules, namely: gene expression data preprocessing module, initial difference network calculation module, mutual information correlation network calculation module, sample correlation network calculation module, network disturbance change calculation module, overall network disturbance change calculation module, random network construction module, random network difference network calculation module. These 8 modules respectively perform the specific operations of the 8 steps in the above single sample regulatory network construction method.

[0024] The mutual information (MI) technique used in the present invention has been shown to be advantageous in measuring gene expression regulation in biological systems.

[0025] The analysis of various single-cell data sets by the diffMIN provided by the present invention shows that it is superior to the existing SSN algorithm in cell clustering and is more effective than clustering based only on gene expression. The present invention successfully applies diffMIN to single-cell data of transverse aortic constriction (TAC), constructs a differential network of pressure overload, and identifies monotonic changes in the regulatory network during heart failure. The present invention also applies diffMIN to breast cancer samples from The Cancer Genome Atlas (TCGA), and finds that it can identify key genes that are strongly correlated with prognosis but have no significant expression differences. In general, the method of the present invention has a wide range of applicability and performs well in cell clustering, key gene identification, biological process research and disease diagnosis.

[0026] Compared with the prior art, the technical features and performance advantages of the present invention are mainly as follows:

[0027] (1) It has strong flexibility and can construct different single-sample gene regulatory networks according to different research purposes. The first step to identify differential networks is to select a background group and then add samples one by one to identify the impact of individual samples on the background network. The choice of background group significantly affects network construction and also provides flexibility for researchers studying biological phenomena. For example, if researchers want to investigate the regulatory networks of cells in different developmental states, they can use the initial developmental state as the background group. Alternatively, if researchers are interested in the differences in regulatory networks between other cell types and specific cell types, they can use specific cell types as the background group. In short, choosing different background groups can meet different research needs.

[0028] (2) Use nonlinear methods to more realistically measure biological networks. As a method for identifying nonlinear models, mutual information is more effective than linear model methods in evaluating complex biological networks. Gene activation, expression enhancement, positive and negative feedback, and cascade reactions often show complex nonlinear correlations. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 The schematic diagram of the diffMIN network construction of the present invention.

[0030] Figure 2 The distribution of edges in random networks and real networks. A. The distribution of edges in random networks; B. The distribution of edges in real networks with perturbations.

[0031] Figure 3 is the relationship between the number of tests for the random permutation test and the significance of the edge.

[0032] Figure 4is the relationship between the calculation time and related parameters. Among them, A is the relationship between the number of selected background cells and the calculation time. B is the relationship between the number of genes selected for calculation and the calculation time. C is the relationship between the number of random networks and the calculation time. D is the relationship between the number of random permutation tests and time.

[0033] Figure 5 The application of the diffMIN method in studying the progression of cardiovascular diseases. A is the visualization result of reducing the dimensionality of the sample gene expression matrix based on t-Distributed Stochastic Neighbor Embedding (t-SNE). The left side of B is the gene expression heat map of the sample, and the right side is the heat map of the gene pair (i.e., the edge calculated by diffMIN). C is the gene regulatory network of the sample at different time points.

[0034] Figure 6 Results of survival analysis of the group with high and low node degrees of the gene LAPTMB in the diffMIN network. DETAILED DESCRIPTION

[0035] The present invention is further described below by way of embodiments in conjunction with the accompanying drawings.

[0036] Embodiment 1:

[0037] The public dataset GSE95140 from the GEO database was used to calculate the single-sample gene regulatory network.

[0038] The method of the present invention was applied to single-cell RNA sequencing data (GSE95140) to reveal the gene regulatory network pattern of cardiomyocytes undergoing transverse aortic constriction (TAC). In this case, normal cardiomyocytes were selected as the background group to calculate diffMIN. The gene expression matrix (GEM) and degree matrix (DEM) calculated by diffMIN were used for UMAP dimensionality reduction. After using DEM for dimensionality reduction, Figure 5As shown, the samples show clear time trajectories, while in the GEM image, the samples are completely disordered and the trajectory information cannot be identified. Heat maps of gene expression and edge relationships were further drawn. It can be seen that in different disease states, specific genes show discontinuous characteristics, but the difference edges calculated by diffMIN show some obvious change trends as the disease progresses. The complexity of all cells was also evaluated by calculating the number of significant difference edges, indicating that the disease state shows increasingly complex interconnections, and the complexity decreases as the disease progresses. This is consistent with a series of gene regulatory changes initiated by cardiomyocytes when receiving pressure overload signals, aimed at resisting the pressure on the heart. Although compensation is generally regarded as a favorable response, it may turn into a decompensated state under persistent pressure overload conditions, leading to pathological remodeling of the structure and ultimately heart failure. In addition, it was observed that the important hub genes in the unique network, namely Nppa, Ankrd1 and Acta1, all played a key role in the progression of pressure overload. The results show that diffMIN has potential advantages in exploring disease progression, especially hypertension.

[0039] Embodiment 2:

[0040] Cancer typing and prognosis analysis are crucial for studying cancer in the field of bioinformatics. In order to further explore the biological functions of cancer, the diffMIN method was applied to the TCGA-BRCA dataset. The data was preprocessed and sampled, and the degree matrix (DEM) and gene expression matrix (GEM) calculated by diffMIN were used as input clustering results, respectively. The study found that the network-based dimensionality reduction method can better distinguish between tumor samples and normal samples. Interestingly, it was observed that the gene expression of LAPTMB remained relatively consistent in most patients, but higher degrees appeared in some samples. The patients were further typed according to the degree of LAPTMB, and it was found that Figure 6 As shown, patients with higher LAPTMB degrees tend to have a worse prognosis, while those with lower degrees do relatively well. The results suggest that network-based features can enhance gene expression signatures, potentially revealing additional biological information that may be overlooked by gene expression alone.

[0041] The preferred embodiments of the present invention are described in detail above.

Claims

1. A method for constructing a single-sample gene regulatory network based on mutual information, characterized in that: The specific steps are: S1: First, the gene expression data is preprocessed to remove genes and samples with low gene expression; then, the gene expression data is uniformly discretized; S2: According to the biological significance, a part of the samples is first designated as background samples, which is N1, and the remaining samples are foreground samples, which is N2; the initial difference network is obtained by calculating the perturbation of the foreground samples on the background samples; S3: Calculate the mutual information correlation network of the background sample and the perturbation sample respectively; where any two genes in the sample X=(x1,x2,x3,...,x n ) and gene Y=(y1,y2,y3,...,y n )The formula for calculating the mutual information correlation network between is as follows: Where n represents the number of samples, p(x) and p(y) represent the marginal probabilities of X and Y respectively, and p(x, y) represents the joint probability of X and Y; S4: Normalize the mutual information network to obtain the standardized sample correlation network. The standardized calculation formula is as follows: in, Represent the information entropy of variables X and Y respectively; the mutual information network calculated by background samples is expressed as The mutual information network obtained by adding a foreground sample j (j = 1, 2, ..., N2) to the background sample is expressed as S5: By calculating the difference between the perturbed network and the background network, the network changes caused by the perturbed sample are identified. The calculation formula is as follows: Among them, j represents the jth foreground sample to be calculated; S6: For the calculation of the individual network of the background sample, an approximate method is used, wherein a single sample i is removed from the background population each time, i=1,2,...,N1, and the disturbance change of the overall network after removing the single sample is calculated; the calculation formula is as follows: Among them, i represents the i-th background sample to be calculated; S7: Perform a significance test on the edges of the network using a random permutation test method to obtain the significance value of the edges of the network; construct a random network by randomly disrupting the gene distribution, specifically, keeping the order of the samples unchanged, randomly disrupting the expression level of each gene, and constructing a new random network while keeping its gene distribution unchanged; S8: Calculate the difference network of the random network in the same way as above; and use the random permutation test method to perform a permutation test on the edges of the calculated real network.

Images