A method for fusion of endocrine disease gene signatures based on cellular heterogeneity function

Abstract

A method for fusing endocrine disease gene features based on cell heterogeneity function, which relates to a method for fusing endocrine disease gene features. The present invention is to solve the problem of limitations of existing cell heterogeneity gene feature extraction methods. The steps of the present invention include step 1, obtaining cell gene association data; step 2, cell-associated gene feature representation; step 3, fusing cell-associated gene features. The present invention belongs to the technical field of endocrine disease cell heterogeneity gene feature fusion.

Description

Technical Field

[0001] The invention relates to a method for fusing endocrine disease gene characteristics, and belongs to the technical field of endocrine disease cell heterogeneity gene characteristic fusion. Background Art

[0002] Cells have an independent and orderly self-controlled metabolic system and are the basic units of all metabolic activities in the human body. Cell dysfunction will directly lead to the occurrence of diseases. Single-cell sequencing technology can provide more accurate and detailed cell function information by measuring and analyzing the gene expression of individual cells. By analyzing the heterogeneity of these cells, we can better understand the pathogenesis and pathological process of endocrine diseases and provide more information and basis for precision medicine and personalized treatment.

[0003] At present, the extraction of cell heterogeneity gene features is based on the annotation analysis of cells based on the gene expression profile data obtained by gene sequencing. For example, DeepEP is a cell heterogeneity gene feature extraction method that uses CNN to extract gene expression profile features. scGAN is a cell heterogeneity gene feature extraction method that uses a generative adversarial network model to correct gene expression profiles. However, this method still has certain limitations: 1) Due to the batch effect caused by the barriers of single-cell sequencing technology and the randomness of the genome transcription expression process, there are currently unignorable deviations in the single-cell sequencing results; 2) Affected by the sequencing depth, some genes with important regulatory functions but low expression levels cannot be captured.

[0004] In addition to the differential expression of cell heterogeneity genes, the degree of association between genes and diseases, the degree of specific expression in tissues, the location of the protein expressed by the gene in the cell substructure, and the disease pathway data involved in the cell heterogeneity genes all affect disease-related cell functions. A deep learning-based cell heterogeneity gene feature fusion method based on deep learning is proposed. This method fuses multiple gene features that affect cell function by constructing a generative adversarial network, and generates gene features related to cell endocrine function by learning the distribution of gene fusion features. Summary of the invention

[0005] The present invention aims to solve the problem of limitations of existing methods for extracting gene features of cell heterogeneity, and further proposes a method for fusing gene features of endocrine diseases based on cell heterogeneity function.

[0006] The technical solution adopted by the present invention to solve the above problems is: the steps of the present invention include:

[0007] Step 1, obtaining cell gene association data;

[0008] Step 2: Expression of cell-associated gene features;

[0009] Step 3: Fusion cell-associated gene signatures.

[0010] The sources of cell function-related gene data include: 1) disease-related single-cell sequencing data, 2) disease mutation site data, 3) endocrine tissue gene expression data, 4) gene subcellular localization data and 5) disease-related pathway data. Based on the above data, gene feature representation can be performed through data preprocessing to obtain a 366-dimensional feature vector of cell-related genes. The adversarial mechanism of the generative adversarial network based on the multi-layer perceptron is used to learn the gene fusion feature distribution, generate cell function-related gene features, and perform cell heterogeneity gene feature fusion that affects cell function.

[0011] Furthermore, in step 1, in order to effectively use the gene data related to cell function to accurately predict the cell heterogeneity genes, the gene data needs to be preprocessed; first, the single-cell sequencing data is screened according to the differential expression multiple and statistical significance to obtain the cell's significantly differentially expressed genes; in order to extract multiple gene features that affect cell function, the relevant data used include:

[0012] (1) Endocrine disease mutation site data: Search for literature related to endocrine diseases in the GWAS Catalog database;

[0013] (2) eQTL data in endocrine tissues: eQTL data in pancreas, adipose, blood, and muscle tissues were obtained from the GTEx Portal database;

[0014] (3) Subcellular localization data: Considering that proteins in different cellular substructures have specific functions, protein subcellular localization data were downloaded from the UniProt database;

[0015] (4) Disease pathway data: Pathway annotation information comes from the KEGG pathway database. Since biological pathways represent different biological processes within cells, analyzing the biological pathways in which genes participate can reflect related cellular functions.

[0016] Furthermore, in step 2, according to the position information of the endogenous disease genes, they are compared with the disease mutation site data and tissue-specific differential expression data. The significant p-values ​​of the five most significantly associated mutation sites of each gene are selected as the disease mutation site information of the cell-associated gene; then, according to the position information, the matched SNPs are compared with the eQTL data of the endocrine tissue, and the p-values ​​and log2FC values ​​of the matched SNPs are used as the differential expression information of the gene in the tissue, in order to make better use of the endocrine tissue-specific information; the pathway data involved in the cell-associated genes are derived from the KEGG pathway library, and the associated pathway information of the gene is represented as a 343-dimensional binary vector, and each dimension represents the association relationship between the gene and the pathway; the subcellular localization information of the gene is represented as a binary vector, and by analyzing the subcellular localization information of the cell-associated gene, the top 10 cellular substructures are selected to represent the subcellular localization information of the gene.

[0017] Furthermore, the cell-associated gene feature fusion method in step 3 includes:

[0018] Step 301, splicing the gene information of the four levels affecting cell function into a 366-dimensional vector, and using it as a feature representation of the initial cell function-related gene information;

[0019] Step 302: Using the adversarial idea of ​​generative adversarial networks, the generator G samples from a normal distribution similar to the real gene feature distribution, learns the gene fusion feature distribution that affects cell function, and generates simulated gene features.

[0020] Step 303: Use the discriminator D to simulate gene features Perform authenticity scoring and use adversarial methods to train and learn through the discriminator D to maximize the difference between the simulated gene features and the real gene feature distribution;

[0021] Step 304: guide the generator to reduce the difference in step 303, learn the effective information in the gene fusion features, and generate cell endocrine function gene features.

[0022] Furthermore, the generator D in step 302 and the discriminator D in step 303 are both composed of three-layer MLPs, and the layers are fully connected. The number of nodes in the three layers of the generator are 128, 256, and 512, respectively, and the number of nodes in the three layers of the discriminator are 512, 256, and 128, respectively.

[0023] Furthermore, the input of the generator G simulates the normal distribution of the real gene feature distribution The mean and standard deviation are the same as the true gene feature distribution. The activation function of the generator is LeakyRelu, the activation function of the discriminator is tanh, the loss function is binary cross entropy, and the optimizer is Adam.

[0024] Furthermore, the training process of the generative adversarial network model includes two steps:

[0025] Step 1: Fix the generator and train the discriminator to distinguish between real and generated genetic features until the discriminator cannot distinguish them;

[0026] Step 2: Fix the discriminator, generate a trainer, and learn to extract the endocrine function information from the gene fusion features, so that the discriminator cannot distinguish between the real and generated gene features.

[0027] Furthermore, the objective function of the generative adversarial network is:

[0028]

[0029] Where D(x), represents the score of the discriminator on the real and generated features. During training, when the generator generates simulated gene features for the first time and inputs them into the discriminator, the parameters of the generator are first fixed and the discriminator is trained. Then the optimal solution of D is When training the generator, the JS divergence is used to measure the distribution difference between the real features and the generated features. By minimizing the JS divergence, the simulated gene distribution closest to the real gene feature distribution is obtained. The optimal generator G * It can be expressed as: G * =argmin G JS(P data (x)||P G (x)), the adversarial process is used to continuously improve the performance of the generator and the discriminator, and the gene fusion features are accurately learned based on the information transfer rules between the MLP hidden layers.

[0030] The beneficial effects of the present invention are as follows: the method first extracts multiple gene features that affect cell functions, and uses a generative adversarial network to generate features of genes related to cell endocrine functions; then, a SMOTE-based gene amplification method is proposed, which improves the problem of low classification accuracy caused by too many small sample genes in the gene label combination by amplifying small sample genes in the gene label combination; finally, a RAkEL-based cell function identification method is proposed, which identifies cell functions by analyzing endocrine disease pathway functions involved in cell heterogeneity genes, and compared with existing similar methods, it is proved that this method effectively improves the accuracy of identifying cell heterogeneity genes and cell functions in endocrine diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1It is a flowchart of the present invention;

[0032] Figure 2 This is a schematic diagram of the analysis of the gene feature fusion process. DETAILED DESCRIPTION

[0033] Specific implementation method 1: Combination Figure 1 and Figure 2 To illustrate this embodiment, the steps of a method for fusion of endocrine disease gene features based on cell heterogeneity function described in this embodiment include:

[0034] Step 1, obtaining cell gene association data;

[0035] Step 2: Expression of cell-associated gene features;

[0036] Step 3: Fusion cell-associated gene signatures.

[0037] Specific implementation method 2: Combination Figure 1 and Figure 2 This embodiment is described. In step 1 of the method for fusion of endocrine disease gene features based on cell heterogeneity function described in this embodiment, in order to effectively use gene data related to cell function to accurately predict cell heterogeneity genes, it is necessary to preprocess the gene data; first, the single-cell sequencing data is screened according to the differential expression multiples and statistical significance to obtain cell-significant differentially expressed genes; in order to extract multiple gene features that affect cell function, the relevant data used include:

[0038] (1) Endocrine disease mutation site data: Search for literature related to endocrine diseases in the GWAS Catalog database;

[0039] (2) eQTL data in endocrine tissues: eQTL data in pancreas, adipose, blood, and muscle tissues were obtained from the GTEx Portal database;

[0040] (3) Subcellular localization data: Considering that proteins in different cellular substructures have specific functions, protein subcellular localization data were downloaded from the UniProt database;

[0041] (4) Disease pathway data: Pathway annotation information comes from the KEGG pathway database. Since biological pathways represent different biological processes within cells, analyzing the biological pathways in which genes participate can reflect related cellular functions.

[0042] Specific implementation method three: Combination Figure 1 and Figure 2To illustrate this embodiment, in step 2 of a method for fusion of endocrine disease gene features based on cellular heterogeneity function described in this embodiment, the position information of the endocrine disease gene is compared with the disease mutation site data and tissue-specific differential expression data, and the significance p-values ​​of the 5 most significantly associated mutation sites of each gene are selected as the disease mutation site information of the cell-associated gene; then, the matched SNP is compared with the eQTL data of the endocrine tissue according to the position information, and the p-value and log2FC value of the matched SNP are used as the differential expression information of the gene in the tissue, in order to make better use of the endocrine tissue-specific information; the pathway data involved in the cell-associated gene is derived from the KEGG pathway library, and the associated pathway information of the gene is represented as a 343-dimensional binary vector, and each dimension represents the association relationship between the gene and the pathway; the subcellular localization information of the gene is represented as a binary vector, and the top 10 cellular substructures are selected to represent the subcellular localization information of the gene by analyzing the subcellular localization information of the cell-associated gene.

[0043] Specific implementation method four: Combination Figure 1 and Figure 2 To illustrate this embodiment, the cell-associated gene feature fusion method steps in step 3 of the endocrine disease gene feature fusion method based on cell heterogeneity function described in this embodiment include:

[0044] Step 301, splicing the gene information of the four levels affecting cell function into a 366-dimensional vector, and using it as a feature representation of the initial cell function-related gene information;

[0045] Step 302: Using the adversarial idea of ​​generative adversarial networks, the generator G samples from a normal distribution similar to the real gene feature distribution, learns the gene fusion feature distribution that affects cell function, and generates simulated gene features.

[0046] Step 303: Use the discriminator D to simulate gene features Perform authenticity scoring and use adversarial methods to train and learn through the discriminator D to maximize the difference between the simulated gene features and the real gene feature distribution;

[0047] Step 304: guide the generator to reduce the difference in step 303, learn the effective information in the gene fusion features, and generate cell endocrine function gene features.

[0048] Specific implementation method five: Combination Figure 1 and Figure 2To illustrate this embodiment, the generator D in step 302 and the discriminator D in step 303 of the endocrine disease gene feature fusion method based on cellular heterogeneity function described in this embodiment are both composed of three-layer MLP, and the layers are fully connected. The number of nodes in the three layers of the generator are 128, 256, and 512, respectively, and the number of nodes in the three layers of the discriminator are 512, 256, and 128, respectively.

[0049] Specific implementation method six: Combination Figure 1 and Figure 2 This embodiment describes that the input of the generator G of the endocrine disease gene feature fusion method based on cell heterogeneity function described in this embodiment simulates the normal distribution of the real gene feature distribution The mean and standard deviation are the same as the true gene feature distribution. The activation function of the generator is LeakyRelu, the activation function of the discriminator is tanh, the loss function is binary cross entropy, and the optimizer is Adam.

[0050] Specific implementation method seven: Combination Figure 1 and Figure 2 This embodiment is described. The training process of the generative adversarial network model of the endocrine disease gene feature fusion method based on cell heterogeneity function described in this embodiment includes two steps:

[0051] Step 1: Fix the generator and train the discriminator to distinguish between real and generated genetic features until the discriminator cannot distinguish them;

[0052] Step 2: Fix the discriminator, generate a trainer, and learn to extract the endocrine function information from the gene fusion features, so that the discriminator cannot distinguish between the real and generated gene features.

[0053] Specific implementation method eight: Combination Figure 1 and Figure 2 This embodiment is described. The objective function of the generative adversarial network of the endocrine disease gene feature fusion method based on cell heterogeneity function described in this embodiment is:

[0054]

[0055] Where D(x), represents the score of the discriminator on the real and generated features. During training, when the generator generates simulated gene features for the first time and inputs them into the discriminator, the parameters of the generator are first fixed and the discriminator is trained. Then the optimal solution of D is When training the generator, the JS divergence is used to measure the distribution difference between the real features and the generated features. By minimizing the JS divergence, the simulated gene distribution closest to the real gene feature distribution is obtained. The optimal generator G * It can be expressed as: G *=argmin G JS(P data (x)||P G (x)), the adversarial process is used to continuously improve the performance of the generator and the discriminator, and the gene fusion features are accurately learned based on the information transfer rules between the MLP hidden layers.

[0056] By using the adversarial mechanism of generative adversarial networks and adversarial training, we can effectively learn the distribution of gene fusion features and generate gene features related to cell endocrine function. We use the t-SNE method to visualize the generated gene fusion features and the real gene features. Figure 2 The distribution of gene fusion features (red) and real gene features (blue) generated at the 10th, 50th, 100th, 150th, 220th, and 300th iterations is shown. At the beginning of training, the two features are clearly divided into two clusters, indicating that there are obvious differences between the real features and the generated features. With adversarial training, the distribution of real gene features begins to be captured after 100 iterations, and the quality of generated gene fusion features continues to improve. After 220 iterations, the distribution of real gene features and synthetic gene features is very similar, indicating that the model has effectively learned the distribution of important gene features, thereby generating gene fusion features that affect cell endocrine function.

[0057] The above is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technician familiar with this profession can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent replacement and improvement made to the above embodiments without departing from the content of the technical solution of the present invention, based on the technical essence of the present invention, within the spirit and principles of the present invention, still fall within the protection scope of the technical solution of the present invention.

Claims

1. A method for fusion of endocrine disease gene features based on cellular heterogeneity function, characterized by: The steps of the endocrine disease gene feature fusion method based on cell heterogeneity function include: Step 1: Obtain cell gene association data; In order to effectively use cell function-related gene data to accurately predict cell heterogeneity genes, the gene data needs to be preprocessed; First, the single-cell sequencing data is screened according to the differential expression multiples and statistical significance to obtain the cell's significantly differentially expressed genes; In order to extract multiple gene features that affect cell function, the relevant data used include: (1) Endocrine disease mutation site data: Search for literature related to endocrine diseases in the GWAS Catalog database; (2) eQTL data in endocrine tissues: eQTL data in pancreas, adipose, blood, and muscle tissues were obtained from the GTExPortal database; (3) Subcellular localization data: Considering that proteins in different cellular substructures have specific functions, protein subcellular localization data were downloaded from the UniProt database; (4) Disease pathway data: Pathway annotation information comes from the KEGG pathway database. Since biological pathways represent different biological processes within cells, analyzing the biological pathways in which genes participate reflects related cell functions. Step 2, cell-associated gene feature representation; according to the location information of endogenous disease genes, they are compared with disease mutation site data and tissue-specific differential expression data. The significant p-values ​​of the 5 most significantly associated mutation sites for each gene are selected as the disease mutation site information of the cell-associated gene; then, the matched SNPs are compared with the eQTL data of endocrine tissues according to the location information, and the p-values ​​and log2FC values ​​of the matched SNPs are used as the differential expression information of the gene in the tissue, in order to make more use of endocrine tissue-specific information; the pathway data involved in cell-associated genes are derived from the KEGG pathway library, and the gene-associated pathway information is represented as a 343-dimensional binary vector, and each dimension represents the association between the gene and the pathway; the subcellular localization information of the gene is represented as a binary vector, and the top 10 cell substructures are selected to represent the subcellular localization information of the gene by analyzing the subcellular localization information of the cell-associated gene; Step 3: Fusion cell-associated gene features; the specific steps include: Step 301, splicing the gene information of the four levels affecting cell function into a 366-dimensional vector, and using it as a feature representation of the initial cell function-related gene information; Step 302: Using the adversarial idea of ​​generative adversarial networks, the generator G samples from a normal distribution similar to the real gene feature distribution, learns the gene fusion feature distribution that affects cell function, and generates simulated gene features. The input of the generator G simulates the normal distribution of the real genetic feature distribution The mean and standard deviation are the same as the real gene feature distribution. The activation function of the generator is LeakyRelu, the activation function of the discriminator is tanh, the loss function is binary cross entropy, and the optimizer is Adam; Step 303: Use the discriminator D to simulate gene features Perform authenticity scoring and use adversarial methods to train and learn through the discriminator D to maximize the difference between the simulated gene features and the real gene feature distribution; Step 304: guide the generator to reduce the difference in step 303, learn the effective information in the gene fusion features, and generate cell endocrine function gene features.

2. The method for fusion of endocrine disease gene features based on cell heterogeneity function according to claim 1, characterized in that: The generator D in step 302 and the discriminator D in step 303 are both composed of three-layer MLPs, and the layers are fully connected. The number of nodes in the three layers of the generator are 128, 256, and 512 respectively, and the number of nodes in the three layers of the discriminator are 512, 256, and 128 respectively.

3. The method for fusion of endocrine disease gene features based on cell heterogeneity function according to claim 1, characterized in that: The training process of the generative adversarial network model consists of two steps: Step 1: Fix the generator and train the discriminator to distinguish between real and generated genetic features until the discriminator cannot distinguish them; Step 2: Fix the discriminator, generate a trainer, and learn to extract the endocrine function information from the gene fusion features, so that the discriminator cannot distinguish between the real and generated gene features.

4. The method for fusion of endocrine disease gene features based on cell heterogeneity function according to claim 3, characterized in that: The objective function of the generative adversarial network is: Where D(x), represents the score of the discriminator on the real and generated features. During training, when the generator generates simulated gene features for the first time and inputs them into the discriminator, the parameters of the generator are first fixed and the discriminator is trained. Then the optimal solution of D is When training the generator, the JS divergence is used to measure the distribution difference between the real features and the generated features. By minimizing the JS divergence, the simulated gene distribution closest to the real gene feature distribution is obtained. The optimal generator G * Expressed as: G * = arg min G JS(P data (x)||P G (x)), the adversarial process is used to continuously improve the performance of the generator and the discriminator, and the gene fusion features are accurately learned based on the information transfer rules between the MLP hidden layers.

Images