Method for constructing prediction model of pan-apoptosis related biomarkers of asthma based on convolutional neural network algorithm and application

Abstract

The invention discloses a method for constructing a prediction model of pan-apoptosis related biomarkers of asthma based on a convolutional neural network algorithm and application, and the method comprises the steps: firstly obtaining related data from a database, and determining characteristic pan-apoptosis related genes PDGs in asthma through an SVM support vector machine, LASSO regression analysis and RF random forest regression; the method comprises the following steps: establishing an lncRNA-miRNA-mRNA network by using interaction STRING and Cytoscape software, constructing a prediction model of the pan-apoptosis related biomarkers of asthma through a convolutional neural network, and analyzing the diagnosis accuracy of potential target genes through an ROC curve. 10 characteristic PDGs playing an important role in asthma are finally obtained, the difference of the genes between a normal group and an asthma group is proved, and the enrichment pathway and immune cell infiltration of the genes have a certain relation with the occurrence and development of asthma, so that a prediction model of the pan-apoptosis related biomarkers of asthma is constructed, and the prediction model is used for predicting the pan-apoptosis related biomarkers of asthma. And a new research direction is provided for studying the pathogenesis and treatment of asthma.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent medical technology, and specifically is a method and application of constructing a prediction model of pan-apoptosis-related biomarkers of asthma based on a convolutional neural network algorithm. Background Art

[0002] Bronchial asthma (abbreviated as asthma) is a chronic airway inflammatory disease involving multiple cells and cell components. Its etiology is increasingly attributed to the interaction between genetic susceptibility, host factors and environmental exposure. Clinical manifestations include recurrent wheezing and shortness of breath, with or without chest tightness or cough, accompanied by airway hyperresponsiveness and reversible airflow limitation. Globally, asthma ranks 16th in years of life lost due to disability and 28th in the leading cause of disease burden, measured in disability-adjusted life years. About 300 million people suffer from asthma worldwide, and another 100 million people may be affected by 2025. The pathogenesis of asthma has not yet been fully elucidated, but current studies have found that the core processes of asthma mainly include epithelial cell dysfunction, persistent airway inflammation and airway remodeling. Airway epithelium expresses pattern recognition receptors (PRRs), which respond rapidly to pathogen-associated molecular patterns (PAMPS) and damage-associated molecular patterns (DAMPS), ultimately mediating inflammation and immune responses by synthesizing and secreting inflammatory mediators. Avoiding contact with asthma-causing allergens and long-term drug control remain the main measures to prevent asthma attacks. Inhaled corticosteroids (ICS) are by far the most effective treatment for asthma because they can clearly inhibit the characteristic inflammation of the airways in asthma. However, long-term inhalation of corticosteroids is not only prone to cause corticosteroid-related side effects, but is also associated with glucocorticoid-resistant asthma. So far, there are still few targeted therapeutic drugs for asthma, and they are expensive and not suitable for widespread use. Therefore, finding the key genes, pathways and target proteins involved in the occurrence and development of asthma can provide strong support for the targeted treatment of asthma.

[0003] Programmed cell death (PCD) is an important component of organism development and plays an important role in host defense and maintaining homeostasis. Multiple PCD pathways are associated with innate immunity, including pyroptosis, apoptosis, and necroptosis. Early studies emphasized the unique regulation of each pathway, but emerging studies have discovered co-regulation and crosstalk between these seemingly different cell death complexes. Pan-apoptosis is a new type of cell death proposed by Malireddi et al. in 2019, characterized by pyroptosis, apoptosis, and necroptosis, but cannot be explained by any one of them alone. The key complex that initiates pan-apoptosis is called the pan-apoptotic body. It is a cell death-inducing complex that has the molecular characteristics of pyroptosis, apoptosis, and necroptosis, and can induce and regulate pan-apoptosis. At present, the specific mechanism of pan-apoptosis has not been elucidated, but studies have found that its occurrence and development are closely related to a variety of diseases. Studies have found that airway administration of di ABZI can induce Caspase-3 and Gasdermin D protein (GSDMD) cleavage and MLKL phosphorylation in lung tissue, induce apoptosis, pyroptosis, and necroptosis, and cause a lung inflammatory storm. In addition, it was found that ZBP1 was upregulated in lung tissue, and the expression of Caspase-8, ASC, and RIPK3 proteins increased, which can promote the expression of pan-apoptotic bodies and induce pan-apoptosis. Studies have found that FADD is highly expressed in lung cancer tissues compared with adjacent adjacent tissues; at the same time, it is also an adapter for pan-apoptosis and apoptosis processes. A study confirmed the role of pan-apoptosis in psoriasis for the first time through bioinformatics analysis, and predicted hub genes and immune characteristics.

[0004] At present, the role of pan-apoptosis in asthma is still unclear. However, more and more studies have shown that there is a close connection between asthma and pyroptosis, apoptosis, and necroptosis. Studies have shown that the expression of NOD-like receptor pyrin domain-related protein 3 (NLRP3) and caspase-1 in bronchoalveolar lavage fluid of asthmatic patients is significantly higher than that of healthy individuals, and similar results have been found in house dust mite (HDM) or albumin (OVA) models. Other studies have found that respiratory epithelial cells may undergo pyroptosis by activating NLRP 3 and promoting GSDMD splicing, thereby increasing respiratory inflammation and hyperresponsiveness. In summary, we can find that pyroptosis of airway epithelial cells is the initiation of asthma inflammation. The persistent presence of eosinophils in the airways is related to the severity of disease symptoms, which may depend not only on the number of airway eosinophils, but also on their survival mechanism after recruitment. Studies have found that the number of apoptotic eosinophils is negatively correlated with the severity of asthma symptoms, suggesting that apoptosis may be an important triggering mechanism for allergic inflammation in asthma. Oikonomou N et al. found that FADD or caspase-8 deficiency caused RIPK3-MLKL-dependent necroptosis in epithelial cells, causing inflammation of barrier tissues such as the skin and intestines, and determined that epithelial cell necroptosis is a powerful driver of inflammation. In summary, we can understand that the inflammatory response of asthma is associated with pyroptosis, apoptosis, and necroptosis. Therefore, we speculate that pan-apoptosis may be closely related to the occurrence and development of asthma. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a method and application of constructing a predictive model of pan-apoptosis-related biomarkers of asthma based on a convolutional neural network algorithm, which is used to study the pathogenesis of asthma and provide new treatment guidance for asthma.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A prediction model for pan-apoptosis-related biomarkers of asthma was constructed based on a convolutional neural network algorithm. The model was constructed by a convolutional neural network, and the input data included but were not limited to the expression data of 10 characteristic pan-apoptosis-related genes PDGs, including LY96, MLKL, CYCS, CHMP6, CHMP7, CASP8, CASP4, AIM2, CASP9 and TNFRSF1A.

[0008] The model was trained for 100 cycles, the optimizer was SGD, batch normalization was used after each convolution and before activation to speed up the operation, and 10% dropout was added to the fully connected layer to reduce overfitting, the initial learning rate was 0.01, and an adaptive adjustment strategy was adopted.

[0009] The method for constructing a prediction model of pan-apoptosis-related biomarkers of asthma based on a convolutional neural network algorithm comprises the following steps:

[0010] (1) Obtained the datasets including GSE165934, GSE182503, GSE2125, and GSE40885 from the GEO database, and obtained the pan-apoptosis-related gene PRGs from the GeneCards database;

[0011] (2) The limma package was used to standardize and annotate the dataset, and a list of differentially expressed genes DEGs between normal samples and asthma samples was obtained. DEGs were selected from the list, and a Venn diagram was constructed to find common DEGs, namely Co-DEGs. The Co-DEGs and PRGs were intersected through the Venn diagram to generate PDGs;

[0012] (3) GO functional analysis and KEGG pathway analysis were performed on PDGs using the ClusterProfiler R package to clarify their biological processes, molecular functions, cellular components, and biological pathways in asthma;

[0013] (4) Screening out characteristic pan-apoptosis-related genes of asthma through machine learning models;

[0014] (5) Using the CIBERSORT algorithm, immune infiltration analysis was performed to evaluate the level of immune cell infiltration in the asthma group and the healthy group and the correlation between the immune cells;

[0015] (6) construct the lncRNA-miRNA-mRNA network using the STRING online analysis tool and Cytoscape software;

[0016] (7) A convolutional neural network algorithm was used to construct a predictive model for pan-apoptosis-related biomarkers of asthma. Training and validation were performed using Python software on Ubuntu 20.045LTS. SGD was selected as the optimizer. Batch normalization was used after each convolution and before activation. A 10% dropout was added to the fully connected layer. Training was performed for 100 cycles.

[0017] The step (1) of obtaining pan-apoptosis-related genes PRGs from the GeneCards database specifically retains only protein-coding genes and genes with a correlation score > 0.500.

[0018] In step (2), the DEGs were selected according to the following criteria: the DEGs were determined based on the corrected P value < 0.05 and |logFC|> 0; the PDGs were selected according to the following criteria: the PDGs were determined based on the P value < 0.05 and |logFC|> 0.

[0019] The machine learning models in step (4) are SVM support vector machine, LASSO regression analysis and RF random forest regression.

[0020] The specific operation of screening out characteristic pan-apoptosis-related genes of asthma by machine learning model in step (4) is as follows:

[0021] (1) Characteristic genes were obtained by screening through SVM-RFE support vector machine;

[0022] (2) LASSO regression analysis was performed using the glmnet software package, with gene expression values ​​as predictor variables and sample groups as response variables;

[0023] The cv.glmne function was used to cross-validate the model and draw the cv graph to determine the optimal regularization parameter to select the optimal model. Then, according to the coefficient of Lasso regression, the genes with non-zero coefficients were identified, i.e., the genes to be identified.

[0024] The glm function was used to construct an initial logistic regression model for the characteristic genes obtained by screening. By checking the P value of the model coefficient, significant genes (P<0.05) were selected as new model variables. Based on the significant genes, a new logistic regression model was constructed, and the stepwise method was used for variable selection. The coefficients, summary statistics, and confidence intervals of the new model were extracted, and the Odds Ratio, 95% confidence interval, and P value of each gene were calculated. Finally, the samples were predicted by the trained final model, the risk score was calculated, and the characteristic genes were screened.

[0025] (3) Through RF random forest regression, n sample data are extracted from the training set with replacement to form a new training set. For each training set, a part of the variables are randomly selected for training to form a single decision tree model; all sub-models output their prediction results for the same prediction data, and finally a voting form is taken, and the classification category with the most output times is taken as the final output result of the model;

[0026] (4) The crossover genes of the valuable genes obtained in steps (1), (2), and (3) are used as key genes.

[0027] The beneficial effects obtained by the present invention are:

[0028] By analyzing asthma-related pan-apoptotic genes, the study may identify new biomarkers, which will help to better understand the molecular mechanisms of asthma. The constructed convolutional neural network model provides an efficient method to predict and diagnose asthma. By analyzing potential biomarkers, the accuracy of early diagnosis of asthma can be improved. By monitoring the expression levels of characteristic genes, it is possible to monitor the course of asthma in real time and adjust the treatment plan in time.

[0029] Identifying specific asthma-related genes will help develop personalized medicine strategies that tailor treatment plans based on the patient's gene expression signature. Understanding pan-apoptosis-related pathways and genes in asthma may reveal new drug targets and help develop more targeted treatments with fewer side effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a schematic diagram of the data distribution of the data set before and after standardization.

[0031] Figure 2A It is the Venn diagram of all DEGs in the four datasets; Figure 2B It is a Venn diagram of 31 PDGs of asthma.

[0032] Figure 3 It is the GO analysis bubble diagram and KEGG analysis bubble diagram of 31 PDGs.

[0033] Figure 4A This is the M0 macrophage infiltration level diagram of patients in the asthma group. Figure 4B It is the proportion of each immune cell in the normal group and the disease group; Figure 4C It is a heat map of immune cell correlation analysis.

[0034] Figure 5A It is the result of lasso regression and ridge regression analysis; Figure 5B It is the result screened by SVM support vector machine; Figure 5C It is the result screened by RF random forest regression; Figure 5D It is the Venn diagram of 10 characteristic PDGs obtained by three machine learning algorithms.

[0035] Figure 6 It is the expression result of 10 characteristic PDGs between the normal group and the asthma group.

[0036] Figure 7 This is the DEGs analysis result diagram of 10 characteristic PDGs between the high expression group and the low expression group.

[0037] FIG8 is a diagram showing the enrichment pathway results of 10 characteristic PDGs.

[0038] FIG. 9 is a diagram showing the immune cell infiltration results of 10 characteristic PDGs.

[0039] Fig.10 is the ROC curve of 10 characteristic PDGs.

[0040] FIG. 11 shows the performance analysis results of the constructed prediction model of pan-apoptosis-related biomarkers of asthma. DETAILED DESCRIPTION

[0041] The present invention will be described below in conjunction with the accompanying drawings.

[0042] The prediction model of pan-apoptosis-related biomarkers of asthma described in the present invention is constructed as follows:

[0043] (1) Obtained the datasets including GSE165934, GSE182503, GSE2125, and GSE40885 from the GEO database, and obtained the pan-apoptosis-related gene PRGs from the GeneCards database;

[0044] The datasets used in this study are all from the GEO database, including GSE165934, GSE182503, GSE2125, and GSE40885, all of which are from Homo sapiens. The GSE165934 dataset includes microarray gene expression profiles of 9 normal samples and 10 asthma disease samples; GSE182503 contains 36 normal samples and 50 asthma disease samples; GSE2125 contains 15 normal samples and 15 asthma samples; GSE40885 includes microarray gene expression profiles of 7 normal group samples and 7 asthma group samples. The normal samples and asthma samples in the four datasets were included in the subsequent analysis. Afterwards, the GeneCards database was searched, and only protein-coding genes and genes with correlation scores > 0.500 were retained. As a result, 87 pan-apoptosis-related genes (PRGs) were retrieved. We used Origin to draw PCA graphs to normalize the four datasets. The results show that after standardization, the batch effect between samples in each data set is basically eliminated, and the data distribution between each data set tends to be consistent, such as Figure 1 As shown in AB.

[0045] (2) The limma package was used to standardize and annotate the dataset, and a list of differentially expressed genes (DEGs) between normal samples and asthma samples was obtained. DEGs were selected from the list and a Venn diagram was constructed to find common DEGs, namely Co-DEGs. The Co-DEGs and PRGs were intersected through the Venn diagram to generate PDGs.

[0046] In order to investigate the possible mechanisms, biological characteristics and pathways of differentially expressed genes (DEGs) in asthma, the limma package was used to normalize the data of normal and asthma samples in GSE165934, GSE182503, GSE2125 and GSE40885. After data preprocessing, the differentially expressed genes (DEGs) list between the normal group and the asthma group was obtained.

[0047] The results showed that there were 4619 differentially expressed genes between the normal group and the asthma group, with the corrected P value < 0.05 and llogFCl > 0 as the standard. With logFC > 0 and P.adj < 0.05, DEGs were considered to be up-regulated, with 2231 up-regulated genes; with logFC < 0 and P.adj < 0.05, DEGs were considered to be down-regulated, with 2388 down-regulated genes.

[0048] Next, we constructed the Venn diagrams of all DEGs in the four datasets, as shown in Figure 2A , to find common DEGs (Co-DEGs).

[0049] Finally, the Co-DEGs and PRGs were intersected through the Venn diagram to generate PDGs. We intersected these DEGs with the 87 pan-apoptotic PRGs to obtain PDGs. The results showed that with the standard of P value < 0.05 and llogFCl > 0 after correction, a total of 31 PDGs of asthma were obtained, including 20 up-regulated genes and 11 down-regulated genes. The up-regulated genes included NLRP3, CHMP2B, LY96, CYCS, BMF, TICAM1, YWHAE, YWHAG, CASP9, TNFRSF1A, and E2F1, and the down-regulated genes included DFFA, CASP8, CASP2, CASP10, GSDMB, NLRP1, CASP4, TRAF2, GSDMD, CHMP6, CHMP2A, RIPK3, IRF1, CASP6, MLKL, CHMP7, CD14, ZBP1, AIM2, and TNFSF10. The results were presented in the form of Venn diagram, as shown in Figure 2. Figure 2B The volcano map and heat map are based on the ggplot2 and pheatmap packages in R, respectively.

[0050] (3) GO functional analysis and KEGG pathway analysis of PDGs:

[0051] GO analysis can analyze the functions, biological pathways involved, and cellular localization of gene products, and is therefore widely used in the functional enrichment study of a large number of genes. The Kyoto Encyclopedia of Genes and Genomes (KEGG) is a widely used database that contains information pathways such as genomes, biology, diseases, and chemicals. The ClusterProfiler R package was used to perform gene annotation enrichment analysis on PRGs. P values ​​< 0.05 and false discovery rates (FDR) (q values) < 0.05 were considered statistically significant. The Benjamini-Hochberg correction for P values ​​was used.

[0052] GO functional analysis and KEGG pathway analysis were performed on these 31 PDGs to explore their biological processes (BP), molecular functions (MF), cellular components (CC) and biological pathways in asthma. The BPs of the 31 PDGs were mainly enriched in 10 GO entries, namely positive regulation of immune response, cell pyroptosis, cysteine ​​endopeptidase activity, endopeptidase activity, positive regulation of peptidase activity, regulation and activation of cysteine ​​endopeptidase activity, activation of innate immune response, positive regulation of cysteine ​​endopeptidase activity, cysteine ​​endopeptidase activity involved in apoptosis and apoptosis process, and regulation of programmed necrotic cell death. CC mainly involves classical inflammasome complex, autophagosome, ESCRT III complex, amphisome membrane, amphisome, kinetochore microtubules, ESCRT complex, multivesicular body, membrane autophagosome, and membrane multivesicular body. MF mainly involved in the process of apoptosis, including cysteine ​​endopeptidase activity, cysteine ​​endopeptidase activity, cysteine ​​peptidase activity, endopeptidase activity, pattern recognition receptor activity, endopeptidase activator activity, sulfoinositol-4-phosphate binding, tumor necrosis factor receptor binding, lipopolysaccharide binding, and tumor necrosis factor receptor superfamily binding. Figure 3 A in.

[0053] In the KEGG pathway, 31 DEGs were mainly enriched in 30 pathways, including necroptosis, Salmonella infection, lipid and atherosclerosis, apoptosis, NOD-like receptor signaling pathway, NF-kappaB signaling pathway, hepatitis C, pathogenic Escherichia coli infection, cytoplasmic DNA sensing pathway, TNF-α signaling pathway, influenza A, hepatitis B, etc. The results are presented in Figure 3 In B.

[0054] We used the CIBERSORT deconvolution method to analyze the immune cell infiltration levels between the asthma group and the healthy group, and P < 0.05 was considered to be statistically significant. Figure 4A As shown in the figure, the infiltration level of M0 macrophages in the asthma group was higher than that in the normal group, and the difference was statistically significant.

[0055] We then estimated the abundance of infiltrated immune cells and obtained the relative proportions of infiltrating immune cells in the asthma and healthy groups.

[0056] The results showed that the proportions of each immune cell in the normal group and the disease group were different ( Figure 4B). Pearson correlation coefficient analysis was performed on the infiltration levels of immune cells in the two groups. Pearson correlation coefficient counts found that M1 macrophages had a strong positive correlation with resting dendritic cells (r=0.81), M2 macrophages had a moderate negative correlation with monocytes (r=-0.53), and naive B cells had a moderate negative correlation with memory B cells (r=-0.53) ( Figure 4C ).

[0057] (4) Screening out characteristic pan-apoptosis-related genes of asthma through machine learning models:

[0058] Lasso Regression Analysis

[0059] Gene expression data and gene list files were read by R language, and then the expression data of genes included in the list were screened according to the gene list files. Least absolute shrinkage and selection operator (LASSO) regression analysis was performed using the glmnet software package, with gene expression values ​​as predictor variables and sample groups as response variables. The model was cross-validated (cv) using the cv.glmne function, and the cv graph was drawn to determine the optimal regularization parameter to select the optimal model. Then, according to the coefficients of Lasso regression, genes with non-zero coefficients were identified, i.e., the selected genes to be characterized. The glm function was used to construct an initial logistic regression model for the characteristic genes screened, and the P value of the model coefficient was checked to select significant genes (P<0.05) as new model variables. Subsequently, a new logistic regression model was constructed based on the significant genes, and variable selection was performed using the stepwise method. The coefficients, summary statistics, and confidence intervals of the new model were extracted, and the Odds Ratio, 95% confidence interval, and P value of each gene were calculated. Finally, the trained final model is used to predict the samples and calculate the risk score.

[0060] Finally, 15 characteristic genes were obtained, namely CHMP7, MLKL, CHMP6, AIM2, TRAF2, BMF, LY96, CASP9, IRF1, CASP4, CYCS, TNFRSF1A, CD14, E2F1, and CASP8. Figure 5A It is used to intuitively demonstrate how Lasso regression performs feature selection, that is, automatically screening out independent variables that contribute more to model prediction, while excluding those that contribute less, thereby streamlining the model and improving its generalization ability.

[0061] SVM support vector machine

[0062] Support vector machine (SVM) was proposed by Vapnik in 1979. It is a supervised learning algorithm for data classification and regression analysis. The core idea of ​​the algorithm is to map the input data into a higher-dimensional feature space and perform linear classification by calculating the hyperplane with the maximum interval of the learning samples. It has many advantages in solving many problems such as nonlinear problems and high-dimensional pattern recognition. It is one of the most practical algorithms in artificial intelligence research. It has excellent performance in the fields of face recognition, digital writing, text distribution and data retrieval. Thirty-one characteristic genes were obtained by SVM-RFE support vector machine screening, namely AIM2, BMF, CASP10, CASP2, CASP4, CASP6, CASP8, CASP9, CD14, CHMP2A, CHMP2B, CHMP6, CHMP7, CYCS, E2F1, GSDMB, GSDMD, IRF1, LY96, MLKL, NLRP1, NLRP3, RIPK3, TICAM1, TNFRSF1, TNFSF10, TRAF2, YWHAE, YWHAG, ZBP1( Figure 5B ).

[0063] RF Random Forest Regression

[0064] Random Forest (RF) is a bagging algorithm based on one of the ensemble algorithms. It uses decision trees as models in bagging and applies multiple decision trees to train and predict samples. During the training process, n sample data are extracted from the training set with replacement to form a new training set. For each training set, a part of the variables are randomly extracted for training to form a single decision tree model. In the classification prediction, all sub-models output their prediction results for the same prediction data, and finally take the form of voting, and the classification category with the most output times is used as the final output result of the model. Through RF random forest tree regression, 16 characteristic genes can be obtained, namely CHMP7, MLKL, CHMP6, LY96, CHMP2B, CD14, CASP4, ZBP1, TNFRSF1A, YWHAE, CASP8, RIPK3, BMF, CASP9, AIM2, CYCS ( Figure 5C ).

[0065] The cross-genes of valuable genes obtained by the three machine learning algorithms were used as key genes to obtain 10 characteristic PDGs of asthma, namely LY96, MLKL, CYCS, CHMP6, CHMP7, CASP8, CASP4, AIM2, CASP9, and TNFRSF1A. The results of the three algorithms were visualized using the venn package and the results were displayed in the Venn diagram ( Figure 5D ).

[0066] Gene Set Enrichment Analysis (GSEA) evaluates the distribution of predefined gene sets in a list of genes ranked by phenotypic relevance to measure the contribution of genes to the phenotype. First, we divided the characteristic PDGs screened out by the three machine learning models into high expression groups and low expression groups according to their expression levels. Next, we used clusterProfiler to evaluate the functional enrichment of high expression groups and low expression groups of different genes. The P value was corrected using the Benjamini-Hochberg method. GSEA was performed using the MSigDB collection of c2.cp.all.v202212.Hs.symbols.gm. P.adj<0.05 and FDR (q.value<0.05) indicated significant enrichment.

[0067] (5) Using the CIBERSORT algorithm, immune infiltration analysis was performed to evaluate the level of immune cell infiltration in the asthma group and the healthy group and the correlation between the immune cells:

[0068] The CIBERSORT algorithm is an analytical tool for estimating immune cell infiltration. It uses linear support vector regression to separate the transcriptome matrix and determine the composition and number of immune cells in mixed cell populations. Through the CIBERSORT package, the four integrated gene sets were grouped into normal and asthma disease groups, and the data with immune cell enrichment values ​​greater than zero were screened. Then we obtained and displayed the precise findings on the immune cell infiltration abundance matrix. The content of immune cells in the samples was represented by box plots. The correlation between immune cells was created by the spearman algorithm, and the correlation heat map was performed by the pheatmap package.

[0069] To study whether there are significant differences in the 10 characteristic PDGs of asthma between the normal group and the asthma group, we will analyze the expression differences of these genes between the normal group and the asthma group. If P<0.05, it proves that there are significant differences between the normal group and the asthma group. We analyzed the expression of the 10 characteristic PDGs between the normal group and the asthma group by using the ggplot package in R. As shown in the figure, these 10 characteristic genes are different between the normal group and the disease group, and the differences are statistically significant. Among them, the expression of MLKL, CHMP6, CHMP7, CASP8, CASP4, and AIM2 in the asthma group is higher than that in the normal group ( Figure 6 AF). However, the expression of LY96, CYCS, CASP9, and TNFRSF1A was higher in the normal group ( Figure 6 GJ).

[0070] In order to explore the gene differences between the high expression group and the low expression group of the 10 PDGs, we used the ggplot package to analyze the DEGs between the high expression group and the low expression group of the 10 PDGs. The results showed that there were 74 differentially expressed genes between the AIM2 high expression group and the low expression group, including 31 down-regulated genes and 43 up-regulated genes ( Figure 7 A). There were 160 differentially expressed genes between the CASP4 high expression group and the low expression group, including 31 down-regulated genes and 129 up-regulated genes ( Figure 7 B). There are 4153 differentially expressed genes in the LY96 high expression group and low expression group, including 2053 down-regulated genes and 2100 up-regulated genes ( Figure 7 C). There were 1213 differentially expressed genes between the CYCS high-expression and low-expression groups, including 795 down-regulated genes and 418 up-regulated genes ( Figure 7 D). There are 1974 differentially expressed genes between the TNFRSF1 high expression group and the low expression group, 1003 up-regulated genes, and 971 down-regulated genes ( Figure 7 E). There were 2304 differentially expressed genes between high and low CASP8 expression, of which 1201 were down-regulated and 1103 were up-regulated ( Figure 7 F). There are 1984 differentially expressed genes between high and low CASP9 expression, of which 1072 are upregulated and 912 are downregulated ( Figure 7 G). There were 2827 differentially expressed genes between high-expressed CHMP6 and low-expressed CHMP6, including 1162 down-regulated genes and 1665 up-regulated genes ( Figure 7 H). There were 1791 differentially expressed genes between high and low expression of CHMP7, of which 846 were down-regulated genes and 945 were up-regulated genes ( Figure 7 I). There were 1213 differentially expressed genes between high and low MLKL expression, including 795 up-regulated genes and 418 down-regulated genes ( Figure 7 J).

[0071] To explore the differences in biological functions between the high and low expression groups of the 10 PDGs, we performed GSEA analysis on the 10 PDGs. Gene Set Enrichment Analysis (GSEA) evaluates the distribution of predefined gene sets in a list of genes ranked by phenotypic relevance to measure the contribution of genes to the phenotype.

[0072] First, we divided the characteristic PDGs screened out by the three machine learning models into high expression group and low expression group according to their expression levels. Next, we used clusterProfiler to evaluate the functional enrichment of high expression group and low expression group of different genes. The P value was corrected by the Benjamini-Hochberg method. GSEA was performed using the MSigDB collection of c2.cp.all.v202212.Hs.symbols.gm. P.adj<0.05 and FDR (q.value<0.05) indicated significant enrichment.

[0073] According to the enrichment score, we selected the top 5 pathways for display. The results showed that the AIM2 high expression group was enriched in 15 pathways, mainly in cytokine-cytokine receptor interaction, cytoplasmic DNA sensing pathway, hematopoietic cell lineage, RIG-I-like receptor signaling pathway, and primary immunodeficiency. The AIM2 low expression group was enriched in 26 pathways, mainly in ribosome, lysosome, porphyrin and chlorophyll metabolism, glycolysis, and citric acid cycle pathways (Figure 8A). The CASP9 high expression group was enriched in 51 pathways, mainly in graft-versus-host disease, chemical dynamics signaling pathway, Leishmania infection, cytokine receptor interaction, and NOD-like receptor signaling pathway. The CASP9 low expression group was enriched in 12 signaling pathways, mainly in ribosome, oxidative phosphorylation, primary immunodeficiency, mismatch repair, and aminoacyl RNA biosynthesis (Figure 8B). The low-expression group of CHMP6 was mainly enriched in DNA replication, P53 signaling pathway, protein efflux, graft-versus-host disease, and small cell lung cancer. The high-expression CHMP6 was enriched in 36 pathways, mainly pentose phosphate pathway, arachidonic acid metabolism, oxidative phosphorylation, lysosome, and drug metabolism (Figure 8C). The low-expression group of CHMP7 was enriched in 45 pathways, mainly Leishmania infection, NOD-like receptor signaling pathway, Toll-like receptor signaling pathway, systemic lupus erythematosus, and RIG-I-like receptor signaling pathway. The high-expression group of CHMP7 was mainly enriched in ribosome, primary immunodeficiency, mismatch repair, base excision repair, and DNA replication (Figure 8D). The low-expression CYCS was enriched in 34 pathways, mainly lysosome, ribosome, oxidative phosphorylation, pentose phosphate pathway, porphyrin and chlorophyll metabolism (Figure 8E). The LY96 low expression group was enriched in 7 pathways, and the top 5 were ribosome, primary immunodeficiency, T cell receptor signaling pathway, mTOR signaling pathway, and colorectal cancer. The LY96 high expression group was enriched in 51 pathways, mainly in the oxidative phosphorylation pathway, Leishmania infection, lysosome, asthma, and glycolysis (Figure 8F). The MLKL low expression group was mainly enriched in 12 pathways, and the top 5 pathways were ribosome, T cell receptor signaling pathway, spliceosome, and aldosterone-regulated sodium reabsorption. The MLKL high expression group was enriched in 32 pathways, and the top 5 were lysosome, complement and coagulation cascade, niacin and nicotinamide metabolism, pantothenic acid and coenzyme a biosynthesis, and Leishmania infection (Figure 8G). The TNFRSF1A low expression group was enriched in 8 pathways, and the top 5 were ribosome, T cell receptor signaling pathway, N-glycosyl biosynthesis, biological rhythm, cysteine ​​and methionine metabolism. The high expression group of TNFRSF1A was enriched in 49 pathways, among which the top 5 pathways were cytokine receptor interaction, Leishmania infection, Toll-like receptor signaling pathway, NOD-like receptor signaling pathway, and chemokine signaling pathway ( Figure 8H ).The CASP4 high expression group was enriched in 37 pathways, and the top 5 pathways were Toll-like receptor signaling pathway, cell adhesion molecule, NOD-like receptor signaling pathway, Leishmania infection, and cytoplasmic DNA sensing pathway (Figure 8I). The CASP8 low expression group was enriched in 43 pathways, and the top 5 pathways were oxidative phosphorylation pathway, Parkinson's disease, asthma, Leishmania infection, and systemic lupus erythematosus (Figure 8J).

[0074] To understand the immune cell infiltration of these 10 characteristic PDGs, we used CIBERSORT to estimate the immune cell abundance of genes and the ggplot package was used to visualize the data.

[0075] The results showed that AIM2 was positively correlated with memory B cells (r=0.27, p=0.013) and M1 macrophages (r=0.25, p=0.026), and negatively correlated with monocytes (r=-0.26, p=0.019) (Figure 9A). CASP4 was negatively correlated with naive B cells (r=-0.26, p=0.02) and eosinophils (r=-0.24, p=0.033) (Figure 9B). CASP8 was positively correlated with M2 macrophages (r=0.35, p=0.0014) and naive CD4+T cells (r=0.27, p=0.016), and negatively correlated with monocytes (r=-0.36, p=0.00093) (Figure 9C). CASP9 was positively correlated with dendritic cell activation (r=0.31, p=0.0047) and mast cell activation (r=0.47, p<0.05), and negatively correlated with NK cell activation (r=-0.38, p=0.00038) (Figure 9D). CHMP6 was negatively correlated with M1 macrophages (r=-0.26, p=0.019) and M2 macrophages (r=-0.4, p=0.00018), and positively correlated with monocytes (r=0.4, p=0.00019) and NK cell activation (r=0.26, p=0.019) (Figure 9E). CHMP7 was negatively correlated with dendritic cell activation (r=-0.37, p=0.00058), mast cell activation (r=-0.27, p=0.015), and neutrophils (r=-0.31, p=0.0051), and was positively correlated with NK cell activation (r=0.39, p=0.00034) and resting CD4+ memory T cells (r=0.28, p=0.0098) ( Figure 9F ). CYCS was positively correlated with eosinophils (r=0.35, p=0.0014), memory CD4+ T cell activation (r=0.31, p=0.0044), and memory CD4+ T cell resting (r=0.26, p=0.017), and negatively correlated with monocytes (r=-0.44, p=3.3e-05) and follicular helper T cells (r=-0.23, p=0.013) (Figure 9G). LY96 was positively correlated with neutrophils (r=0.36, p=0.00089), M1 macrophages (r=0.3, p=0.0055), and monocytes (r=0.44, p=3.3e-05), and negatively correlated with initial CD4+T cells (r=-0.26, p=0.018), naive B cells (r=-0.34, p=0.0018), activated memory CD4+T cells (r=-0.36, p=0.00098), eosinophils (r=-0.38, p=0.00036), and γδT cells (r=-0.24, p=0.027) (Figure 9H).MLKL was positively correlated with monocytes (r=0.29, p=0.0079), and negatively correlated with naive B cells (r=-0.34, p=0.0018) and eosinophils (r=-0.27, p=0.016) (Figure 9I). TNFRSF1A was positively correlated with monocytes (r=0.61, p=9.9e-10), and negatively correlated with eosinophils (r=-0.34, p=0.0015), M2 macrophages (r=-0.35, p=0.0012), resting memory CD4+T cells (r=-0.28, p=0.011), and naive CD4+T cells (r=-0.25, p=0.021) (Figure 9J).

[0076] (6) Constructing the lncRNA-miRNA-mRNA network using the STRING online analysis tool and Cytoscape software:

[0077] The STRING online analysis tool was used to construct a protein-protein interaction (PPI) network for differentially expressed mRNAs, and the cytoscape software was used to visualize the PPI network. A literature search was further conducted in the PUBMED database to find key mRNAs that have been successfully verified by previous experiments and have a targeted regulatory relationship with the occurrence of asthma.

[0078] Key mRNA target predictions were performed using the ENCORI database, with the following screening criteria: (1) humans; (2) at least one set of immunoprecipitation data supporting the target relationship; and (3) at least five or more gene prediction software predicted the target relationship. Key miRNAs were obtained and mRNA-miRNA interaction relationships were constructed. Cytoscape software was used again to visualize the interaction network. Further literature searches were performed using PUBMED to find key miRNAs that had been successfully verified in previous experiments and had a target-target regulatory relationship with the occurrence of asthma.

[0079] Finally, the upstream IncRNAs that bind to key miRNAs were predicted using the miRNet database, and IncRNA-miRNA interaction pairs were obtained. The screening criteria were: (1) human; (2) lung tissue. The obtained mRNA-miRNA-IcRNA network was visualized using cytoscape software.

[0080] (7) A convolutional neural network algorithm was used to construct a predictive model for pan-apoptosis-related biomarkers of asthma. Training and validation were performed using Python software on Ubuntu 20.045LTS. SGD was selected as the optimizer. Batch normalization was used after each convolution and before activation. A 10% dropout was added to the fully connected layer. Training was performed for 100 cycles.

[0081] The convolutional neural network was trained and verified using Python software on Ubuntu 20.045LTS. The GPU3090 accelerator was used, and the SGD optimizer was selected. Batch normalization (BatchNormalization.BN) was used after each convolution and before activation to speed up the operation, and 10% dropout was added to the fully connected layer to reduce overfitting. The network was trained for 100 cycles with an initial momentum of 0.937 and adaptive adjustment. The mini-batch size was 32, and the initial learning rate was 0.01. When the learning rate was stable for more than 10 cycles, the learning rate decreased by 10 times and decreased to 0.001 as the training cycle increased. When the accuracy was stable for 10 cycles, the training was stopped. The training and verification steps are as follows: (1) The input image size is 224x224-2T2-net.T2fnet Tc-net first passes through a 7x7 convolutional layer and then goes through four construction layers of different sizes to learn and screen the features to be learned. Finally, the output feature vector is sent to the fully connected layer and the final classification prediction probability is obtained through the softmax function: (3) TS-net uses Res Net34 as a three-way network. The image data of the three sequences are input into ResNet34 respectively. After the three construction layers after the preprocessing layer, the output features are spliced ​​and fused and then input into the fourth layer of the network. The number of stacked blocks in the fourth layer is adjusted to 7, and finally classified through the fully connected layer and Softmax.

[0082] In order to understand the correlation between the expression of these 10 characteristic PDGs and asthma, we used GraphPad to draw ROC curves. The ROC curve is a basic analytical tool that can be used to select the most appropriate model or determine the best threshold in the model. The ROC curve is a comprehensive indicator that reflects the continuous variables of sensitivity and specificity. The relationship between sensitivity and specificity is expressed by the synthesis method. The area under the ROC curve (AUC) is usually between 0.5 and 1. AUC>0.7 means a good predictor, and AUC close to 1 represents better prediction efficiency. We used the pROC package to make ROC curves for PDGs and calculated AUC to evaluate the diagnostic efficacy of characteristic PDGs.

[0083] The results are as follows Fig.10 As shown, CYCS ACU=0.645; LY96 AUC=0.663; TNFRSF1A AUC=0.637; AIM2 AUC=0.693; CASP4 AUC=0.630; CASP8 AUC=0.628; CASP9 AUC=0.670; MLKL AUC=0.716; CHMP6 AUC=0.724; CHPM7 AUC=0.743.

[0084] ROC results showed that there was a correlation between these 10 characteristic PDGs and asthma, among which MLKL, CHMP6, and CHMP7 were moderately correlated.

[0085] In this study, we introduced convolutional neural networks (CNNs) to construct a prediction model for pan-apoptosis-related biomarkers in asthma.

[0086] The results show that the prediction model has a very high accuracy (accuracy close to 1) and a very low error rate (error rate < 0.1) ( Fig.11A -B). Fig. 11C The confusion matrix is ​​used to evaluate the performance of the model, which provides an intuitive way to understand the performance of the model on each category, including the number of instances correctly and incorrectly classified by the model. The figure shows that the model has good predictive effectiveness. In addition, the prediction model of pan-apoptosis-related biomarkers showed a high level of recall and F1 score ( Fig.11D -E). The ROC curve showed that the area under the pan-apoptosis curve was 0.96. These findings collectively indicate that pan-apoptosis-related biomarkers have good predictive efficacy for asthma (Figure 12F).

[0087] In the present invention, R software (version 4.1.0) was used for data processing and analysis. Continuous variables were expressed as mean ± standard deviation (mean ± 5D). The Wilcoxon rank sum test was used to compare the differences between two groups. The Kruskal-Walis test was used for comparison of more than three groups. If not otherwise specified, Spearman correlation analysis was used to calculate the correlation coefficients between different compounds. P < .05 was considered statistically significant.

[0088] The principle of the present invention is as follows: 4 data sets were obtained from the GEO database, totaling 67 normal samples and 82 asthma samples, and the data sets were standardized for further analysis. 87 PRGs were obtained from the GeneCard database. Next, we used the limma package to obtain the differential genes between the normal group and the asthma group, and then took the intersection with PRGs to obtain 31 PDGs, and then performed functional enrichment analysis and immune function analysis on these 31 PDGs. In order to further obtain the characteristic PDGs of asthma, we also analyzed three machine learning models, namely lasso regression analysis, SVM-FE support vector machine, and RF random forest tree, and finally obtained 10 characteristic PDGs of asthma, namely LY96, MLKL, CYCS, CHMP6, CHMP7, CASP8, CASP4, AIM2, CASP9, and TNFRSF1A. Finally, we performed a series of biological function analysis and clinical correlation analysis on these 10 characteristic PDGs, and constructed a prediction model for pan-apoptosis-related biomarkers of asthma.

[0089] Bio-enrichment analysis showed that the 31 PDGs obtained before machine learning were mainly enriched in biological processes such as positive regulation of immune response, cell pyroptosis, cysteine ​​endopeptidase activity, endopeptidase activity, and positive regulation of peptidase activity. et al. found that mice lacking caspase-11 have strong resistance to experimental allergic airway inflammation, suggesting that caspase-4 / 11 is a key factor in allergic airway inflammation. Other studies have found that matrix metalloproteinase-9 (MMP-9) is the main MMP found in bronchoalveolar lavage fluid and bronchial biopsies of patients with allergic asthma after allergen stimulation. It has also been confirmed that neutrophils exposed to allergens lead to the production of MMP-9, which may lead to airway remodeling in asthma. The 10 characteristic PDGs obtained after machine learning were divided into high expression groups and low expression groups. The results showed that the high expression group was mainly enriched in Leishmania infection, cytokine receptor interaction, NOD-like receptor signaling pathway, lysosome, Toll-like receptor signaling pathway and other items. The low expression group was mainly enriched in lysosome, ribosome, oxidative phosphorylation and other items. Cytokine receptor interaction is an important link in the pathogenesis of asthma. These interactions promote the development of airway inflammation by regulating the biological effects of cytokines and affect the course and severity of asthma. Studies have shown that in allergic asthma, Th2 lymphocytes play a major role in triggering inflammatory responses. When sensitized by allergens, they release inflammatory cytokines IL-4 and IL-13, both of which stimulate B cells to synthesize IgE and IL-5, which is necessary for initiating eosinophilic inflammation. Abnormal activation of NOD-like receptor signaling pathways and TOLL-like receptor signaling pathways may lead to excessive production of inflammatory factors and excessive immune responses, thereby exacerbating the inflammatory response of asthma. In summary, our analysis can confirm that the biological channels enriched in PDGS are closely related to the inflammatory response of asthma, and pan-apoptosis may play an important role in the occurrence and development of asthma.

[0090] Immune function correlation analysis showed that the infiltration level of M0 macrophages in the asthma group was higher than that in the normal group. The 10 characteristic genes can also be correlated with a variety of immune cells, and most of them are proinflammatory cells. Hwa Young Lee et al. found that miR-21 antagonism can inhibit the polarization of alveolar M2 macrophages, which can not only reduce Th2 eosinophilic airway inflammation, but also reduce airway hyperresponsiveness and airway remodeling. Clinical correlation analysis showed that there was a certain correlation between the 10 characteristic PDGs and asthma, among which MLKL, CHMP6, and CHMP7 were moderately correlated. The study found that the expression of programmed necrosis key proteins p-RIPK1 / p-RIPK3 / p-MLKL increased in asthma patients and asthma model mice, suggesting that programmed necrosis is involved in the disease process of asthma. Although CHMP6 and CHMP7 are related to cell programmed necrosis, their specific role in asthma is still unclear. Future studies need to further explore the mechanism of action of these proteins in asthma in order to better understand the occurrence of asthma and develop more effective treatments. In addition, we also demonstrated through convolutional neural networks that pan-apoptosis-related biomarkers have good predictive efficacy for asthma.

[0091] Our study discovered pan-apoptosis-related genes in asthma, and further screened out 10 characteristic PDGs through three types of machine learning, providing a new treatment direction for asthma - reducing airway inflammation and preventing airway remodeling by inhibiting the expression of related genes or proteins.

Claims

1. A prediction model for pan-apoptosis-related biomarkers of asthma is constructed based on a convolutional neural network algorithm, characterized by: The model was constructed using a convolutional neural network, and the input data included but were not limited to the expression data of 10 characteristic pan-apoptosis-related genes PDGs, specifically LY96, MLKL, CYCS, CHMP6, CHMP7, CASP8, CASP4, AIM2, CASP9, and TNFRSF1A.

2. The prediction model of pan-apoptosis-related biomarkers of asthma based on the convolutional neural network algorithm according to claim 1, characterized in that: The model was trained for 100 cycles, the optimizer was SGD, batch normalization was used after each convolution and before activation to speed up the operation, and 10% dropout was added to the fully connected layer to reduce overfitting, the initial learning rate was 0.01, and an adaptive adjustment strategy was adopted.

3. The method for constructing a prediction model of pan-apoptosis-related biomarkers of asthma based on a convolutional neural network algorithm according to claim 1, characterized in that: The following steps are involved: (1) Obtained the datasets including GSE165934, GSE182503, GSE2125, and GSE40885 from the GEO database, and obtained the pan-apoptosis-related gene PRGs from the GeneCards database; (2) The limma package was used to standardize and annotate the dataset, and a list of differentially expressed genes DEGs between normal samples and asthma samples was obtained. DEGs were selected from the list, and a Venn diagram was constructed to find common DEGs, namely Co-DEGs. The Co-DEGs and PRGs were intersected through the Venn diagram to generate PDGs; (3) GO functional analysis and KEGG pathway analysis were performed on PDGs using the ClusterProfiler R package to clarify their biological processes, molecular functions, cellular components, and biological pathways in asthma; (4) Screening out characteristic pan-apoptosis-related genes of asthma through machine learning models; (5) Using the CIBERSORT algorithm, immune infiltration analysis was performed to evaluate the level of immune cell infiltration in the asthma group and the healthy group and the correlation between the immune cells; (6) construct the lncRNA-miRNA-mRNA network using the STRING online analysis tool and Cytoscape software; (7) A convolutional neural network algorithm was used to construct a predictive model for pan-apoptosis-related biomarkers of asthma. Training and validation were performed using Python software on Ubuntu 20.045LTS. SGD was selected as the optimizer. Batch normalization was used after each convolution and before activation. A 10% dropout was added to the fully connected layer. Training was performed for 100 cycles.

4. The method for constructing a prediction model of pan-apoptosis-related biomarkers of asthma based on a convolutional neural network algorithm according to claim 3, characterized in that: The step (1) of obtaining pan-apoptosis-related genes PRGs from the GeneCards database specifically retains only protein-coding genes and genes with a correlation score > 0.

500.

5. The method for constructing a prediction model of pan-apoptosis-related biomarkers of asthma based on a convolutional neural network algorithm according to claim 3, characterized in that: In step (2), the DEGs were selected according to the following criteria: the DEGs were determined based on the corrected P value < 0.05 and |logFC|> 0; the PDGs were selected according to the following criteria: the PDGs were determined based on the P value < 0.05 and |logFC|> 0.

6. The method for constructing a prediction model of pan-apoptosis-related biomarkers of asthma based on a convolutional neural network algorithm according to claim 4, characterized in that: The machine learning models in step (4) are SVM support vector machine, LASSO regression analysis and RF random forest regression.

7. The method for constructing a prediction model of pan-apoptosis-related biomarkers of asthma based on a convolutional neural network algorithm according to claim 6, characterized in that: The specific operation of screening out characteristic pan-apoptosis-related genes of asthma by machine learning model in step (4) is as follows: (1) Characteristic genes were obtained by screening through SVM-RFE support vector machine; (2) LASSO regression analysis was performed using the glmnet software package, with gene expression values ​​as predictor variables and sample groups as response variables; The cv.glmne function was used to cross-validate the model and draw the cv graph to determine the optimal regularization parameter to select the optimal model. Then, according to the coefficient of Lasso regression, the genes with non-zero coefficients were identified, i.e., the genes to be identified. The glm function was used to construct an initial logistic regression model for the characteristic genes obtained by screening. By checking the P value of the model coefficient, significant genes (P<0.05) were selected as new model variables. Based on the significant genes, a new logistic regression model was constructed, and the stepwise method was used for variable selection. The coefficients, summary statistics, and confidence intervals of the new model were extracted, and the Odds Ratio, 95% confidence interval, and P value of each gene were calculated. Finally, the samples were predicted by the trained final model, the risk score was calculated, and the characteristic genes were screened. (3) Through RF random forest regression, n sample data are extracted from the training set with replacement to form a new training set. For each training set, a part of the variables are randomly selected for training to form a single decision tree model; all sub-models output their prediction results for the same prediction data, and finally a voting form is taken, and the classification category with the most output times is taken as the final output result of the model; (4) The intersection genes of the valuable genes obtained in steps (1), (2), and (3) are used as key genes.

8. Application of LY96, MLKL, CYCS, CHMP6, CHMP7, CASP8, CASP4, AIM2, CASP9 and TNFRSF1A as pan-apoptosis-related biomarkers for asthma.

Images