Application of transcriptomics-based machine learning model in chronic kidney disease kidney injury gene identification

Abstract

The invention belongs to the technical field of biomedicine and genes, and particularly relates to an application of a transcriptomics-based machine learning model in chronic kidney disease renal injury gene identification, and the construction of the transcriptomics-based machine learning model comprises the following steps: S1, training gene identification by machine learning; s3, performing feature screening; s4, feature gene verification; s5, feature gene causal inference; s6, a correlation analysis result of the expression quantity of the characteristic gene and GFR and SCr shows kidney injury diagnosis and monitoring value. The column graph model based on feature gene expression, which is constructed by the invention, overcomes the prejudice in the prior art, integrates the feature genes for scoring, has relatively high accuracy, and provides a theoretical basis for kidney injury risk prediction, auxiliary clinical early diagnosis and illness state assessment.

Description

Technical Field

[0001] The present invention belongs to the field of biomedicine and gene technology, and specifically relates to the application of a transcriptomics-based machine learning model in the identification of chronic kidney disease renal injury genes. Background Art

[0002] Chronic kidney disease (CKD) is a major public health problem worldwide, posing a huge challenge to personal health and medical systems. The development of CKD is usually a slow and hidden process, and there may be no obvious symptoms in the early stages. However, if not managed in a timely manner, CKD may continue to worsen and eventually lead to end-stage renal disease, when patients can only rely on dialysis or kidney transplantation to maintain their lives. Therefore, early diagnosis and intervention are the key to improving prognosis.

[0003] The diagnosis of chronic kidney disease mainly relies on clinical symptoms, renal function indicators and renal biopsy, etc. These methods still have some limitations. For example, traditional renal function indicators such as serum creatinine (SCr) and urine analysis cannot accurately reflect the degree of kidney damage in the early stage; the invasive operation and inability to continuously monitor renal biopsy limit its clinical application.

[0004] As a powerful data analysis tool, machine learning (ML) has achieved remarkable results in the construction of medical prediction models. Machine learning combined with transcriptomic analysis allows us to examine the expression patterns of disease-related genes at the whole genome level, providing a new perspective for in-depth understanding of the kidney's response mechanism. By analyzing multiple clinical parameters of patients, machine learning can accurately capture early signs of kidney damage, reveal new therapeutic targets, and promote the development and application of new treatment methods.

[0005] Support vector machine-recursive feature elimination (SVM-RFE) and the least absolute shrinkage and selection operator (LASSO) algorithm are two data mining methods widely used in feature gene screening. SVM-RFE repeatedly trains the support vector machine (SVM) model to eliminate unimportant features and screen out the number of features of the optimal variable; LASSO regression can perform variable selection and regularization at the same time. By adding penalty terms, the regression coefficients of unnecessary variables are compressed to zero and then eliminated from the model to achieve the purpose of variable screening.

[0006] The present invention aims to combine the advantages of bioinformatics and machine learning, apply differential expression analysis and weighted gene co-expression network, and combine two machine learning algorithms, SVM-RFE and LASSO, to screen key renal injury marker genes associated with chronic kidney disease, and provide new and important insights for the diagnosis and treatment of chronic kidney disease. Summary of the invention

[0007] The present invention aims to provide a method for identifying chronic kidney disease renal injury genes, and to provide a new theoretical reference at the molecular level for the early diagnosis of chronic kidney disease.

[0008] The application of a transcriptomics-based machine learning model in the identification of chronic kidney disease renal injury genes, the construction of the transcriptomics-based machine learning model includes the following steps:

[0009] S1, machine learning training gene identification:

[0010] (1) Identify the common differentially expressed genes between renal tissue samples of chronic kidney disease with different pathological types and control samples;

[0011] (2) Identify the common modules and core genes within the modules that are most relevant to the clinical traits of renal injury in various pathological types of chronic kidney disease;

[0012] (3) Take the common genes of (1) and (2) as machine learning training genes;

[0013] S2, machine learning training of gene regulatory network: constructing a gene regulatory network based on the genes described in S1 (3);

[0014] S3, feature screening: machine learning algorithms LASSO and SVM-RFE were used to perform feature screening on the genes described in S1 (3);

[0015] S4, signature gene validation: The reliability of the signature genes was verified using the validation data set, and finally the kidney injury signature genes DUSP1, GADD45A, TSC22D3 and ZFAND5 were obtained;

[0016] S5, causal inference of characteristic genes: using Bayesian network inference to determine the causal relationship between characteristic genes of renal injury and renal injury outcomes;

[0017] S6, The results of correlation analysis between the expression of characteristic genes and GFR and SCr showed its value in diagnosing and monitoring kidney injury.

[0018] Furthermore, in step S1(1), differentially expressed genes between chronic kidney disease samples of different pathological types and control samples in the dataset GSE180394 are identified, and common differentially expressed genes of chronic kidney disease of different pathological types are screened according to p less than 0.05 and |log2FC|0.5-1.5.

[0019] Furthermore, in step S1(2), the step of identifying the common modules and core genes in the modules that are most relevant to the clinical traits of chronic kidney disease renal injury of different pathological types in the dataset GSE180394 includes:

[0020] a. Based on the weighted gene co-expression network analysis, the top 35%-65% genes with the largest expression pattern differences were selected to construct the network;

[0021] b. According to the scale-free topology criterion, select the fitting index R 2 (R 2 =0.8-0.9), set the minimum module size to 50-100, and calculate the association between the module and chronic kidney disease;

[0022] c. Select the module with the highest correlation with chronic kidney disease, set the filtering threshold (gene significance |GS| greater than 0.45-0.7 and gene-module correlation |MM| greater than 0.45-0.7) to determine the core genes in the module.

[0023] Furthermore, in step S2, a gene interaction network and a Bayesian gene regulatory network are used to construct a gene regulatory network based on the relationship between the genotype and phenotype of renal injury.

[0024] Furthermore, in step S3, the training data set is the integrated GSE180394 and GSE47184.

[0025] Furthermore, in step S4, the reliability of the characteristic genes is verified by verifying the dataset GSE37455.

[0026] Furthermore, in step S5, the Hill-Cram algorithm is used to learn the Bayesian network model.

[0027] Furthermore, in step S6, the GFR and SCr contents of chronic kidney disease and normal samples are collected from the Nephroseq database, the correlation between the expression of characteristic genes and GFR and SCr is analyzed, and the correlation between chronic kidney disease renal injury characteristic genes and renal function indicators is evaluated.

[0028] Furthermore, an expression nomogram prediction model of characteristic genes of renal injury was constructed, and the diagnostic ability and clinical value of the model were evaluated using calibration curves and decision curves.

[0029] Beneficial effects:

[0030] 1. Data processing method:

[0031] (1) Artificial intelligence technology was applied to the large and complex renal transcriptome data for unbiased analysis, and a characteristic gene module based on renal RNA-seq was established. Key genes of renal injury at the RNA level that were difficult to detect with traditional methods were discovered. At the same time, machine learning was endowed with the potential for cross-development in renal network biology.

[0032] (2) Integrate the analytical advantages of differentially expressed genes and gene co-expression networks, and select machine learning for gene training and feature screening, which can save time and computing costs while improving the clinical value of characteristic genes for diagnosing and monitoring kidney injury.

[0033] 2. New biological discoveries:

[0034] (1) Traditional biomarkers of chronic kidney disease are usually at the protein level. The discovery of new characteristic genes and biomarkers at the transcriptional level not only expands the understanding of the regulatory mechanism of gene expression in chronic kidney disease, but also provides new directions and methods for the diagnosis and treatment of CKD.

[0035] (2) The results of immune infiltration studies revealed the complexity of the immune system and the specific gene expression patterns of CKD, providing a new perspective for understanding immune responses and developing immunotherapy.

[0036] 3. The nomogram model based on characteristic gene expression constructed by the present invention overcomes the bias of the prior art, integrates characteristic genes for scoring, and has high accuracy, providing a theoretical basis for predicting the risk of kidney damage, assisting clinical early diagnosis and disease assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is the volcano plot of differentially expressed genes between chronic kidney disease of different pathological types and the control group in the transcriptome data GSE180394.

[0038] Figure 2 Screening of modules and core genes related to chronic kidney disease of various pathological types in GSE180394 by gene co-expression network (Figure A shows the relationship between scale-free fitting index and soft threshold; Figure B shows the clustering dendrogram of co-expression network modules (each branch represents a gene, and each color represents a co-expression module); Figure C shows the heat map of the correlation between gene modules and clinical characteristics; Figure D shows the scatter plot of the correlation between GS and MM in the yellow module (left), and the core genes in the yellow module are screened according to the thresholds of GS and MM (right)).

[0039] Figure 3 Figure A is the intersection gene Venn diagram of differentially expressed genes and gene co-expression network; Figure B is the protein interaction network diagram; Figure C is the Bayesian gene regulation network diagram.

[0040] Figure 4is the sample normalization process (among which, Figure A is the box plot before data standardization correction; Figure B is the box plot after data standardization correction; Figure C is the PCA plot before data standardization correction, batch1 is GSE180394, batch2 represents GSE47184; Figure D is the PCA plot after data standardization correction, batch1 is GSE180394, batch2 represents GSE47184).

[0041] Figure 5 Screening of characteristic genes for machine learning (A and B are the results of SVM-RFE analysis, when the number of characteristic genes is 10, the minimum error rate and maximum accuracy are 0.0125 and 0.9875 respectively; C and D are the results of LASSO regression analysis, C is the selection of cross-validation parameters, each curve corresponds to a gene, D is the process of variable screening, and the vertical dotted line is drawn with the optimal lambda; E is a Venn diagram of the intersection of characteristic genes screened by LASSO and SVM-RFE algorithms).

[0042] Figure 6 The expression differences and ROC curves of characteristic genes (Figure A is a box plot of the expression of characteristic genes in the chronic kidney disease group and the control group in the training set, ***p<0.001; Figure B is the ROC curve of the characteristic genes in the training set; Figure C is a box plot of the expression of characteristic genes in the chronic kidney disease group and the control group in the validation set, ***p<0.001; Figure D is the ROC curve of the characteristic genes in the validation set).

[0043] Figure 7 It is a directed acyclic graph of the causal relationship between kidney injury characteristic genes and kidney injury outcomes.

[0044] Figure 8 The correlation between characteristic genes of renal injury in chronic kidney disease and renal function indicators (Figure A is the correlation analysis between DUSP1, GADD45A, TSC22D3 and ZFAND5 and GFR; Figure B is the correlation analysis between DUSP1, GADD45A, TSC22D3 and ZFAND5 and SCr).

[0045] Fig. 9 Diagnostic prediction model for chronic kidney disease (prediction model constructed based on characteristic genes of renal injury in chronic kidney disease, Figure A is the diagnostic nomogram model of GADD45A+DUSP1; Figure B is GADD45A+TSC22D3; Figure C is GADD45A+ZFAND5. Add the corresponding scores obtained for each predictor variable, find the final value on the "total score" axis, and then correspond to the "probability" axis to determine the risk probability of chronic kidney disease).

[0046] Fig.10Evaluation of the diagnostic efficacy of the chronic kidney disease prediction model (Figure A is the ROC curve of GADD45A+DUSP1, GADD45A+TSC22D3 and GADD45A+ZFAND5 models for distinguishing the chronic kidney disease group from the control group; Figures B and C are the calibration curve and decision curve, respectively). DETAILED DESCRIPTION

[0047] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and understandable, the specific implementation modes of the present invention are described in detail below with reference to the accompanying drawings, but they should not be construed as limiting the applicable scope of the present invention.

[0048] Example 1

[0049] Identification of common differentially expressed genes between renal tissue samples of chronic kidney disease with different pathological types and control samples

[0050] The GSE180394 dataset was used to identify differentially expressed genes in the kidney between chronic kidney disease and control samples, with |log2FC|>1 and P<0.05 as thresholds, where log2FC>1 indicated upregulation of RNA expression and log2FC<-1 indicated downregulation. A total of 549 genes were identified by differential analysis, including 339 downregulated genes and 210 upregulated genes. A volcano plot was created using "ggplot2" to display the differential expression of genes ( Figure 1 ).

[0051] Example 2

[0052] Identification of the most common modules related to clinical traits of renal injury and the core genes within the modules

[0053] The gene co-expression network method was used to analyze the dataset GSE180394, and key gene modules were mined according to the different gene expression patterns to obtain the modules and genes with the highest degree of correlation with chronicity. First, the top 50% genes with the largest expression pattern differences were selected to construct the network, abnormal samples were screened out, and the fitting index R was selected. 2 The soft threshold (R 2 =0.9, soft threshold β=4)( Figure 2 A), set the minimum module size to 80; then, based on TOM similarity, perform hierarchical clustering analysis to form different gene modules ( Figure 2 B); further, the correlation between the modules and the clinical characteristics of chronic kidney disease was calculated, and a total of 10 gene co-expression modules were obtained ( Figure 2C); the yellow module most related to chronic kidney disease traits was selected (r = -0.91, P < 0.01), the correlation between genes in the yellow module and chronic kidney disease traits was calculated (r = 0.56, P < 0.01), and 278 core genes in the yellow module were obtained according to the filtering threshold (gene significance |GS|> 0.5 and gene-module correlation |MM|> 0.6). Figure 2 D).

[0054] Example 3

[0055] Protein-protein interaction and construction of Bayesian gene regulatory networks

[0056] Based on the 549 common differentially expressed genes between renal tissue samples of chronic kidney disease with different pathological types and control samples and the 278 core genes in the common modules most related to the clinical traits of renal injury, 111 intersection genes were obtained ( Figure 3 A). The protein-protein interaction and Bayesian gene regulatory networks were constructed using the STRING database and the hc function of the bnlearn package, respectively, to determine the close relationship between these genes ( Figure 3 B and C).

[0057] Example 4

[0058] Screening and validation of landmark genes

[0059] The SVA package of R software was used to eliminate the batch effect between the training set GSE180394 (batch1) and the dataset GSE47184 (batch2) ( Figure 4 AD), the expression of the corrected data tends to be balanced and unified ( Figure 4 B and D).

[0060] Based on the SVM-RFE algorithm of the R software e1071 package and the LASSO algorithm of the glmnet package, the intersection genes (111 genes) of the gene co-expression network and the differential analysis were analyzed ( Figure 3 A) Incorporate the training set model into the machine learning model for feature screening. SVM-RFE analysis shows that when the number of feature genes is 10, the model has the best accuracy, and the maximum accuracy and minimum error rate of 10-fold cross validation are 0.9875 and 0.0125 respectively ( Figure 5 A, B), while LASSO regression analysis showed that when the minimum mean square error logarithm λ was selected, there were 11 characteristic genes with non-zero coefficients in the regression model ( Figure 5 C, D). Combining the analysis results of the two algorithms, the characteristic genes of chronic kidney disease renal injury, COL10A1, DUSP1, GADD45A, TSC22D3 and ZFAND5 ( Figure 5 E).

[0061] The ggboxplot function was used to visualize the expression levels of the characteristic genes in the training set and the validation set, and the receiver operating characteristic (ROC) curve was drawn to calculate the area under the curve (AUC). In the training data set, the expression levels of the above five characteristic genes were all lower than those in the normal control group, and the differences were all significant ( Figure 6 A). ROC curve analysis showed that the areas under the ROC curve of the characteristic genes COL10A1, DUSP1, GADD45A, TSC22D3 and ZFAND5 were 0.902, 0.995, 0.989, 0.971 and 0.990, respectively ( Figure 6 B). In the validation set, under the premise of statistical significance, the differential expression levels of DUSP1, GADD45A, TSC22D3 and ZFAND5 genes between chronic kidney disease and normal control groups were consistent with the results of the training set, and the AUC of the four genes in the validation set were all greater than 0.85 ( Figure 6 C, D).

[0062] Example 5

[0063] Construction of a directed acyclic graph of causal relationships between renal injury signature genes and renal injury outcomes

[0064] Based on the renal injury characteristic genes DUSP1, GADD45A, TSC22D3 and ZFAND5 and the results of renal injury, the network structure was iteratively improved according to the Hill-Cram algorithm learning to construct a directed acyclic graph of the causal relationship between the characteristic genes and the renal injury results or other characteristic genes. Among them, each gene and renal injury result is a node. According to the maximum likelihood score in the algorithm iteration process, the effect of the gene on the renal injury result or other genes is determined, and directed edges are added between the corresponding nodes ( Figure 7 ).

[0065] Example 6

[0066] Correlation analysis between characteristic genes and clinical renal function indicators GFR and SCr

[0067] The Nephorseq database was used to analyze the correlation between the expression of characteristic genes in CKD and GFR and SCr. Among them, DUSP1 and ZFAND5 were negatively correlated with GFR, with correlation coefficients of -0.7 and -0.84, respectively; GADD45A and TSC22D3 were positively correlated with GFR, with correlation coefficients of 0.55 and 0.53, respectively. Figure 8A). DUSP1 and ZFAND5 were positively correlated with SCr, with correlation coefficients of 0.81 and 0.6; GADD45A and TSC22D3 were negatively correlated with SCr, with correlation coefficients of -0.61 and -0.87, respectively ( Figure 8 B).

[0068] Example 7

[0069] Construction and evaluation of prediction nomogram

[0070] The glm function in R software was used to build a logistic regression model, and the rms package was used to establish a prediction nomogram model. The ROC curve was drawn and the AUC value of the model was calculated.

[0071] Specifically:

[0072] Based on the four characteristic genes of DUSP1, GADD45A, TSC22D3 and ZFAND5, logistic regression models based on GADD45A+DUSP1, GADD45A+TSC22D3 and GADD45A+ZFAND5 were constructed, and the prediction nomograms of the three models were established using the rms package ( Fig. 9 AC), and the diagnostic ability and clinical value of the model were evaluated by calibration curve and decision curve.

[0073] The AUCs of the three diagnostic models of GADD45A+DUSP1, GADD45A+TSC22D3, and GADD45A+ZFAND5 were 0.997, 0.995, and 0.995, respectively. Fig.10 A). The calibration curve shows that the difference between the true and predicted risk is small between 0-0.6, which is suitable for early clinical detection and has high accuracy ( Fig.10 B). The results of the decision curve show that the nomogram model has a good overall net benefit in the entire range (0-1), and the model has a high accuracy ( Fig.10 C).

Claims

1. Application of a transcriptomics-based machine learning model in the identification of renal injury genes in chronic kidney disease, characterized in that: The construction of transcriptomics-based machine learning models includes the following steps: S1, machine learning training gene identification: (1) Identify the common differentially expressed genes between renal tissue samples of chronic kidney disease with different pathological types and control samples; (2) Identify the common modules and core genes within the modules that are most relevant to the clinical traits of renal injury in various pathological types of chronic kidney disease; (3) Take the common genes of (1) and (2) as machine learning training genes; S2, machine learning training of gene regulatory network: constructing a gene regulatory network based on the genes described in S1 (3); S3, feature screening: machine learning algorithms LASSO and SVM-RFE were used to perform feature screening on the genes described in S1 (3); S4, signature gene validation: The reliability of the signature genes was verified using the validation data set, and finally the kidney injury signature genes DUSP1, GADD45A, TSC22D3 and ZFAND5 were obtained; S5, causal inference of characteristic genes: using Bayesian network inference to determine the causal relationship between characteristic genes of renal injury and renal injury outcomes; S6, The results of correlation analysis between the expression of characteristic genes and GFR and SCr showed its value in diagnosing and monitoring kidney injury.

2. The use according to claim 1, characterized in that: In step S(1), differentially expressed genes between chronic kidney disease samples of different pathological types and control samples in the dataset GSE180394 are identified, and common differentially expressed genes of chronic kidney disease of different pathological types are screened according to p less than 0.05 and |log2FC|0.5-1.

5.

3. The use according to claim 1, characterized in that: In step S(2), the step of identifying the most common modules and core genes in the modules that are most relevant to the clinical traits of chronic kidney disease renal injury of different pathological types in the dataset GSE180394 includes: a. Based on the weighted gene co-expression network analysis, the top 35%-65% genes with the largest expression pattern differences were selected to construct the network; b. According to the scale-free topology criterion, select the fitting index R 2 (R 2 =0.8-0.9), set the minimum module size to 50-100, and calculate the association between the module and chronic kidney disease; c. Select the module with the highest correlation with chronic kidney disease, set the filtering threshold (gene significance |GS| greater than 0.45-0.7 and gene-module correlation |MM| greater than 0.45-0.7) to determine the core genes in the module.

4. The use according to claim 1, characterized in that: In the step S2, a gene interaction network and a Bayesian gene regulatory network are used to construct a gene regulatory network based on the relationship between the genotype and phenotype of renal injury.

5. The use according to claim 1, characterized in that: In step S3, the training data set is the integrated GSE180394 and GSE47184.

6. The use according to claim 1, characterized in that: In step S4, the reliability of the characteristic genes is verified by verifying the data set GSE37455.

7. The use according to claim 1, characterized in that: In step S5, the Bayesian network model is learned using the Hill-Cram algorithm.

8. The use according to claim 1, characterized in that: In step S6, the GFR and SCr contents of chronic kidney disease and normal samples are collected from the Nephroseq database, the correlation between the expression of characteristic genes and GFR and SCr is analyzed, and the correlation between chronic kidney disease renal injury characteristic genes and renal function indicators is evaluated.

9. The use according to claim 1, characterized in that: An expression nomogram prediction model of characteristic genes for renal injury was constructed, and the diagnostic ability and clinical value of the model were evaluated using calibration curves and decision curves.

Images