IgA nephropathy diagnosis model containing nine genes and establishment method thereof

Abstract

The invention provides an IgA nephropathy diagnosis model containing nine genes and an establishment method thereof, and relates to the field of IgA nephropathy biomedical diagnosis. The nine kinds of genes are CD160, CX3CR1, EPHA4, THBS1, HLA-DRA (human leukocyte antigen-DRA), FARP2, VASH1, TMSB4Y and RHBDD3 (recombined HBDD3). According to the method, activated signal channels are identified in an IgAN queue, then key genes in the channels are positioned, an IgAN diagnosis model is constructed by adopting a machine learning technology based on the key genes, and immune cell infiltration characteristics of the IgAN diagnosis model are further revealed by integrating data related to glomerular functions. Therefore, a foundation is laid for developing a potential treatment strategy.

Description

TECHNICAL FIELD

[0001] The present application relates to the biomedical diagnosis field of IgA nephropathy, and in particular to an IgA nephropathy diagnosis model containing 9 genes and a method for establishing the same. BACKGROUND

[0002] IgA nephropathy (IgAN) is a clinical syndrome characterized by pathological immune features, and its main pathological marker is the deposition of immunoglobulin A in the mesangial region of the glomerulus. As the most common chronic glomerular disease in the world, the annual incidence of IgAN is more than 2.5 / 100,000 people. The disease is prone to occur in older children and young people, and is often accompanied by symptoms such as upper respiratory tract infection before onset. The clinical manifestations of IgAN show high heterogeneity, and can progress from mild urine test abnormalities to rapidly progressive glomerulonephritis. About 30% of patients will develop end-stage renal disease about 20 years after the progression of the disease, and need to receive renal replacement therapy.

[0003] At present, the exact pathogenesis of IgAN has not been fully elucidated, and its definite diagnosis still depends on kidney biopsy. Immune-mediated injury is recognized as the main pathogenic mechanism of IgAN. The "multiple hit" theory system explains its core link: abnormal glycosylation of IgA1 (Gd-IgA1) is overproduced as an autoantigen, which can induce the production of specific autoantibodies. Subsequently, circulating immune complexes are formed and eventually deposited in the mesangial region of the glomerulus, triggering chronic inflammation and kidney damage. Although progress has been made in elucidating the pathophysiological mechanisms of IgAN, there is still a lack of targeted treatment options. The current management of IgAN still follows the universal strategy applicable to all chronic glomerular diseases: including reducing proteinuria, using renin-angiotensin system inhibitors, and strictly controlling blood pressure. For patients who still have persistent proteinuria or progressive deterioration of renal function after conservative treatment, some clinical studies suggest using glucocorticoid therapy to improve proteinuria. However, recent studies have questioned the effectiveness of this method in most patients and emphasized that it can significantly increase the risk of serious adverse reactions, especially infections. Therefore, there is an urgent need to find biomarkers that can assist in the diagnosis of IgAN, promote the development of innovative therapies, and help evaluate patient prognosis. SUMMARY

[0004] In view of the deficiencies in the prior art, the present application provides an IgA nephropathy diagnosis model containing 9 genes and a method for establishing the same, which can be used for the precise diagnosis of IgA nephropathy.

[0005] To achieve the above object, the present application is realized by the following technical scheme: An IgA nephropathy diagnosis model containing 9 genes, the 9 genes are CD160, CX3CR1, EPHA4, THBS1, HLA-DRA, FARP2, VASH1, TMSB4Y, RHBDD3; the CD160, CX3CR1, HLA-DRA, VASH1, RHBDD3 are significantly up-regulated in IgA nephropathy patients, and EPHA4, THBS1, FARP2, TMSB4Y are higher in healthy people.

[0006] Preferably, the HLA-DRA and VASH1 are key diagnostic markers.

[0007] The method for establishing the IgA nephropathy diagnosis model containing 9 genes comprises the following steps: S1, data acquisition: download GSE115857 as a training queue from the GEO database, GSE37460, GSE93798 for model verification, GSE104948, GSE116626 and GSE99339 data sets as external verification queues, all queues contain 207 IgA nephropathy patients and 56 living donors, and batch effect processing is performed; S2, signal pathway score calculation: first, GSEA is used to compare and evaluate the activation levels of different cell signal pathways in all samples, in the GSE115857, GSE37460, GSE93798 three queues, the differentially activated signal pathways between IgA nephropathy patients and living donors are compared, and the significantly different pathways are defined, and based on the consistent results of GSE115857, GSE37460, GSE93798 three queues, the commonly significantly activated signal pathways in IgAN patients are screened out; S3, model construction based on machine learning integrated analysis: integrate elastic network regression algorithm, LASSO algorithm, ridge regression, forward selected generalized linear model, backward eliminated generalized linear model, gradient boosting machine, generalized gradient boosting regression model, linear discriminant analysis and naive Bayes classification model, 9 kinds of machine learning algorithms, among the 101 algorithm combinations generated, 80 kinds of successful output effective results, suitable for subsequent comparison and evaluation, and construct feature markers; subsequent model optimization, grid search method is used to optimize the algorithm hyperparameters in the training queue through ten-fold cross-validation; the finally determined model is a two-stage glmBoost + Enet [alpha=0.4] model, and the HLA-DRA and VASH1 are stable and reliable biomarkers in each queue.

[0008] Preferably, the specific method of batch effect processing in step S1 is to correct the batch effect of the data set downloaded from the GEO database by using the ComBat algorithm in the "sva" software package, and all standardization, feature screening and batch correction steps are independently performed in each training set of cross-validation, so as to ensure the strict isolation between the training set and the test / verification set and avoid information leakage.

[0009] Preferably, the method for defining the significantly different pathways in step S2 is to define the pathways with log2 fold change value greater than 0.3 and p value less than 0.05 as the significantly different pathways.

[0010] Preferably, the specific method for constructing the feature markers in step S3 comprises the following steps: S3-1, the signal pathways commonly significantly activated in the IgAN patients screened in step S2 are included in 80 algorithm combinations for analysis; S3-2, each model is trained on the GSE115857 data set, verified in the GSE37460 and GSE93798 queues, and finally externally verified on the GSE104948, GSE116626 and GSE99339 data sets not involved in the pathway screening process; S3-3, for each independent model, the area under the receiver operating characteristic curve in all research queues is calculated.

[0011] The grid search method is used in step S3 to optimize the algorithm hyperparameters in the training queue through ten-fold cross-validation, including the alpha and lambda parameters in elastic network / LASSO / ridge regression, the number of boosting iterations and the learning rate in GBM, and the shrinkage rate in glmBoost.

[0012] Preferably, the average performance index is calculated by 100 times of random resampling after the model is optimized to evaluate the stability of the model, so as to minimize the risk of overfitting, and in order to prevent data leakage, all preprocessing steps are independently performed in each training set before model training; the validation set and the external test set are always completely isolated before the final evaluation stage, and are not involved in any aspect of model construction or parameter optimization.

[0013] The present application provides an IgA nephropathy diagnosis model containing 9 genes and a method for establishing the same, which has the following advantages compared with the prior art: This invention first identified activated signaling pathways in the IgAN cohort and then located key genes within these pathways. Based on these key genes, a diagnostic model for IgAN was constructed using machine learning techniques. This integrative research strategy not only deepens our understanding of the disease mechanism but also helps in discovering reliable biomarkers for diagnosing the disease (CD160, CX3CR1, EPHA4, THBS1, HLA-DRA, FARP2, VASH1, TMSB4Y, RHBDD3). Furthermore, by integrating data related to glomerular function, the characteristics of immune cell infiltration were further revealed, thus laying the foundation for developing potential therapeutic strategies. This study provides strong evidence for applying artificial intelligence (especially machine learning) to the non-invasive diagnosis and clinical translation of IgA nephropathy. Attached Figure Description

[0014] Figure 1 This is a schematic diagram illustrating batch correction and gene expression pathway analysis in IgA nephropathy (IgAN) patients and living donors (LDs) across multiple GEO cohorts (GSE115857, GSE37460, and GSE93798). A is a PCA diagram showing the variations in gene expression among the GSE115857, GSE37460, and GSE93798 cohorts before (top) and after (bottom) batch effect correction. B is a volcano plot showing differentially activated signaling pathways in IgAN patients. C is a Venn diagram showing the pathways significantly activated in common among IgAN patients across the three cohorts, identifying 21 persistently abnormal pathways. Figure 2 This is a heatmap illustration of 21 significantly activated signaling pathways in IgAN patients according to the present invention; where A represents the GSE37460 cohort; B represents the GSE93798 cohort; C represents the GSE115857 cohort; and in D, each row represents one signaling pathway, and each column represents one sample in that cohort. Figure 3 This diagram illustrates the machine learning diagnostic model for IgA nephropathy (IgAN) constructed based on the glmBoost+elastic net algorithm and gene expression biomarkers. A represents a heatmap summarizing the AUC values ​​of 80 machine learning model combinations across three cohorts; B is the Kappa statistics matrix of the glmBoost+Enet[alpha=0.4] model in cohorts GSE115857, GSE37460, and GSE93798, showing the consistency between the model's diagnostic results and the actual values. Figure 4This is a schematic diagram illustrating the validation of the IgAN diagnostic model of the present invention in independent cohorts; where A is a heatmap of the AUC values ​​of 80 prediction models in cohorts GSE99339, GSE116626 and GSE104948; B is the Kappa statistic of the glmBoost+Enet[alpha=0.4] model in each independent cohort, indicating that it has high consistency in the external cohorts; Figure 5 This is a schematic diagram illustrating the differential expression of diagnostic biomarkers obtained by machine learning screening in IgA nephropathy (IgAN) patients and living donors; the box plot shows the differential expression of CD160, CX3CR1, EPHA4, THBS1, HLA-DRA, FARP2, VASH1, TMSB4Y, and RHBDD3 between IgAN patients and living donors, based on cohorts (A) GSE37460, (B) GSE93798, and (C) GSE115857, respectively. Figure 6 This diagram illustrates the validation of HLA-DRA, VASH1, and RHBDD3 gene expression in an external cohort in this invention. Box plots AB show the differences in gene expression between IgAN patients and living donors in cohorts GSE104948(A) and GSE99339(B). Box plot C shows the expression trends of HLA-DRA, VASH1, and RHBDD3 as IgAN severity levels change. Box plot D shows the correlation between the expression trends of HLA-DRA, VASH1, and RHBDD3 and different histological subtypes of IgAN. Figure 7 This is a schematic diagram illustrating the differential immune cell and gene correlation analysis in IgAN patients according to the present invention; wherein A is a schematic diagram illustrating the difference in immune cell infiltration levels between IgAN patients and living donors (LD); B is a schematic diagram illustrating the correlation between immune cells and HLA-DRA expression; and C is a schematic diagram illustrating the correlation between immune cells and VASH1 expression. Figure 8 This is a schematic diagram illustrating the correlation analysis between HLA-DRA and VASH1 and immune cells in different IgAN subtypes according to the present invention; where A is a scatter plot of the GSE115857 cohort, showing the correlation between gene expression and immune cells in different IgAN severity grades; B is a scatter plot of the GSE116626 cohort, showing the correlation between gene expression and immune cells in different IgAN lesion subtypes. Figure 9This diagram illustrates the immunohistochemical (IHC) and immunofluorescence verification of HLA-DRA and VASH1 in renal biopsy tissues of IgA nephropathy (IgAN) according to the present invention. A represents the pathological diagnosis of clinical IgAN patients, using HE, PAS, silver staining, Masson staining, and IgA and C3 staining to demonstrate the structural and immunological changes in renal tissue. BE represents the immunohistochemical verification of HLA-DRA and VASH1 expression in IgAN patients and controls (B and D are representative immunohistochemical staining images of HLA-DRA and VASH1 in renal tissues of IgAN patients and controls; C and E are quantitative analysis results, showing that the expression levels of HLA-DRA and VASH1 in IgAN patients are significantly higher than those in controls (Mann–Whitney U test, *P < 0.05)). F shows the immunofluorescence double staining of CD4 and HLA-DRA in renal tissue, indicating that the number of CD4+HLA-DR+, functionally active / activated CD4+ T cells is significantly higher in IgAN tissues than in control tissues. Figure 10 This is a schematic diagram showing the calibration status of the 9-gene model of this invention in each cohort and the Brier score; Figure 11 This is a schematic diagram illustrating the SHAP interpretation analysis of the 9-gene diagnostic model of this invention in multiple IgAN cohorts. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Example 1: 1. Materials and Methods 1.1 Basic Information of the Enrollment Queue All sample data, including gene expression profiles and clinical information, were downloaded from the GEO database. The entire cohort included 207 IgAN patients and 56 living donors (LDs). All IgAN samples were verified by pathologists using methods consistent with relevant literature reports.

[0017] GSE115857 was used as the training cohort, GSE37460 and GSE93798 were used for model validation, and GSE104948, GSE116626, and GSE99339 were used as external validation cohorts. Table 1 lists the specific number of IgAN patients and LDs in each cohort: GSE104948 contains 27 IgAN patients and 3 LDs, GSE115857 contains 55 IgAN patients and 7 LDs, GSE116626 contains 52 IgAN patients and 7 LDs, GSE37460 contains 27 IgAN patients and 9 LDs, GSE93798 contains 20 IgAN patients and 22 LDs, and GSE99339 contains 26 IgAN patients and 8 LDs.

[0018] Table 1: Basic Information of the Enrollment Queue 1.2 Batch Effect Handling Batch effects refer to non-biological differences between different datasets. To eliminate the bias caused by batch effects in this study and to make the transcriptional profiling data from the three GEO cohorts more comparable, this study used the ComBat algorithm from the "sva" software package to perform batch effect correction on the GEO datasets from these three sources. All normalization, feature selection, and batch correction steps were performed independently within the training set of each fold cross-validation, thereby ensuring strict isolation between the training set and the test / validation set and avoiding information leakage.

[0019] 1.3 Signal Path Score Calculation Gene set enrichment analysis (GSEA) is a computational method used to assess whether a predefined gene set presents statistically significant differences between two biological states. This experiment first used GSEA to compare different activation signaling pathways between IgAN patients and LDs. The background file for the molecularly characterized gene set was taken from the C5 dataset (Bioprocesses branch) in MSigDB, containing a total of 7751 genes.

[0020] 1.4 Building Predictive Models Based on Machine Learning Integration Analysis To establish a multi-model comparative framework capable of accurately and robustly distinguishing between IgAN patients and healthy individuals, the capabilities of nine machine learning algorithms were integrated. Of the 101 generated algorithm combinations, 80 successfully produced valid results suitable for subsequent comparative evaluation. This algorithm set includes: Elastic Network Regression (Enet), LASSO algorithm, Ridge Regression, Generalized Linear Model (GLM) (including forward selection and backward elimination), Gradient Boosting Machine (glmBoost), Generalized Gradient Boosting Regression Model (GBM), Linear Discriminant Analysis (LDA), and Naive Bayes classification model.

[0021] The process for constructing the biomarkers is as follows: (a) Based on the three datasets GSE115857, GSE37460, and GSE93798, the biological pathways that are significantly activated in IgAN patients are identified; (b) Genes selected from the most active pathways are included in 80 algorithm combinations for analysis; (c) Each model is trained on the GSE115857 dataset, validated in the GSE37460 and GSE93798 cohorts, and finally externally validated on the GSE104948, GSE116626, and GSE99339 datasets that were not involved in the pathway screening process; (d) For each independent model, the area under the receiver operating characteristic (AUC) curve is calculated in all study cohorts.

[0022] To optimize the model, this study employed a grid search method to fine-tune the hyperparameters of each algorithm (including the α and λ parameters in Elastic Network / LASSO / Ridge Regression, the number of iterations and learning rate of GBM, and the shrinkage rate of glmBoost) within the training queue using ten-fold cross-validation. Average performance metrics were calculated through 100 random resampling iterations to assess model stability, thereby minimizing the risk of overfitting. To prevent data leakage, all preprocessing steps (normalization, scaling, and feature selection) were performed independently on each training set before model training. The validation set and external test set remained completely isolated until the final evaluation stage, and did not participate in any part of model construction or parameter optimization.

[0023] 1.5 Analysis of Immune Cell Infiltration To assess the unique immune cell infiltration characteristics within a specific population, this study employed the single-sample gene set enrichment analysis (ssGSEA) method. This algorithm quantifies the absolute enrichment of the immune cell characteristic gene set in each independent sample by calculating an enrichment score. We calculated a standardized enrichment score for each immune cell type

[13] , and all ssGSEA analyses were performed using the GSVA package in R.

[0024] 1.6 Patient Enrollment and Tissue Sample Collection Both IgAN patients and control subjects were recruited from the Department of Nephrology of our hospital. The inclusion criteria for IgAN patients were confirmed by renal biopsy and complete clinical data; the control group consisted of individuals who underwent nephrectomy for non-glomerular disease reasons (such as tumor resection, trauma, etc.) and were ensured to have no history of kidney disease through pathological evaluation.

[0025] 1.7 Immunohistochemistry and Immunofluorescence Staining Key genes and immune cell validation were performed on samples from IgAN patients and control groups (Table 2). The immunohistochemical experimental procedure was as follows: paraffin-embedded tissue sections were dewaxed with xylene and hydrated with gradient ethanol; microwave antigen retrieval was performed using EDTA buffer (pH 9.0) or citrate buffer (pH 6.0); endogenous peroxidase activity was blocked using 3% hydrogen peroxide; sections were incubated overnight at 4°C with VASH1 antibody (Zenbio, 1:400) and HLA-DRA antibody (Abways, 1:400); after washing with PBS, sections were incubated with HRP-labeled secondary antibody (Shanghai Huilan Biotechnology) for 2 hours, and DAB staining showed a brownish-yellow positive signal; finally, sections were counterstained with Harris hematoxylin, dehydrated with gradient ethanol, cleared with xylene, and mounted with neutral resin.

[0026] In the immunofluorescence staining experiment, paraffin-embedded sections were dewaxed and hydrated using the same procedure, followed by antigen retrieval using EDTA buffer (pH 9.0). Endogenous peroxidase was blocked with 3% hydrogen peroxide solution for 25 minutes, followed by blocking nonspecific sites with 3% bovine serum albumin for 10 minutes. Sections were incubated overnight at 4°C with CD4 antibody (ZSGB-BIO, 1:200) and HLA-DRA antibody (Abways, 1:200). After washing with PBS, sections were incubated with species-specific fluorescent secondary antibodies for 1 hour. For dual staining of CD4 and HLA-DRA, a mixed primary antibody simultaneous incubation protocol was used, followed by labeling with the corresponding fluorescent secondary antibodies. After fluorescence signal development, cell nuclei were counterstained with DAPI for 10 minutes, and finally mounted with anti-quenching mounting medium. Images were acquired using a Nikon inverted fluorescence microscope.

[0027] Table 2: Clinical characteristics of enrolled IgAN patients and controls 1.8 Statistics All statistical analyses were performed using R software (version 4.2.2). Fisher's exact test was used for categorical variable analysis, and Pearson correlation coefficient was used for correlation analysis of continuous variables. Heatmaps were generated using the R language's "pheatmap" package to visualize the screening pathways. The consistency between the model-predicted subtype and the actual subtype was quantified using the Kappa statistic. The Kruskal-Wallis test was used for comparisons among multiple groups, and the Wilcoxon rank-sum test was used for comparisons between two groups. Parametric t-tests were used for two groups that met the normal distribution; otherwise, the Mann-Whitney U test was used. All statistical tests were two-tailed, and a p-value less than 0.05 was considered statistically significant.

[0028] 2. Results 2.1 Standardization and Comparative Analysis of Activation of Multi-Queue IgAN Signaling Pathways For details, please see [link / details]. Figure 1 As shown, where Figure 1 In the middle A, the gene expression variations among the GSE115857, GSE37460, and GSE93798 cohorts before and after correction are shown. Gene set enrichment analysis (GSEA) was used to assess the activation levels of different cellular signaling pathways in all samples. Within each cohort, differentially activated signaling pathways were compared between IgAN patients and living donors (LDs), and pathways with a log2 fold change greater than 0.3 and a p-value less than 0.05 were defined as significantly different pathways. The results showed that 130 signaling pathways were significantly activated in IgAN patients in the GSE115857 cohort; 502 pathways were significantly activated in the GSE37460 cohort; and 712 pathways were significantly activated in the GSE93798 cohort. Figure 1 B). Finally, based on the consistent results of the three cohorts, 21 signaling pathways that were significantly activated in IgAN patients were identified. Figure 1 C). These 21 signaling pathways mainly involve key biological mechanisms such as immune regulation, cell structure and signal transduction, developmental biology, metabolic processes, neural development, and hematopoiesis. Figure 2 ).

[0029] 2.2 Constructing an IgAN diagnostic model based on machine learning methods The 105 unique genes contained in the aforementioned 21 signaling pathways were used as model input data. As described in the methodology section, a total of 101 different algorithms were tested for model construction, of which 80 combined algorithms successfully produced results, while the remaining algorithms were not applicable due to limited input data.

[0030] like Figure 3 As shown in Figure A, the AUC values ​​of different models indicate that the average AUC values ​​of the Enet[α=0.1], Ridge, and Enet[α=0.5] models are 0.979, 0.975, and 0.969, respectively. However, further analysis of the genes included in the models revealed that these three models require 85, 105, and 22 input genes, respectively, which introduces significant complexity to clinical applications.

[0031] Therefore, the two-step model glmBoost+Enet [α=0.4] was selected. This model not only achieved a high AUC value (0.968) but also required only 9 input genes, including CD160, CX3CR1, EPHA4, THBS1, HLA-DRA, FARP2, VASH1, TMSB4Y, and RHBDD3. Furthermore, the predictive performance of the model was validated by κ consistency analysis, showing significant consistency in GSE37460 (κ=0.704, p<0.001), GSE93798 (κ=0.486, p<0.001), and GSE115857 (κ=1.000, p<0.001). Figure 3 B).

[0032] 2.3 Validation of the Diagnostic Model Performance in Independent Queues To verify the effectiveness of the established model, three external independent validation cohorts were further selected: GSE99339, GSE116626, and GSE104948. Predictive analysis was performed using 80 models in these cohorts according to established methods. The results show that the glmBoost+Enet[α=0.4] model exhibits superior diagnostic performance (…). Figure 4 A), which showed low scores in GSE99339 (AUC=0.938, κ=0.699, p<0.001), GSE116626 (AUC=0.871, κ=0.443, p=0.018), and GSE104948 (AUC=0.926, κ=0.615, p=0.023). Figure 4 Excellent performance was achieved in all categories B). These results indicate that the model has strong predictive stability, good cross-cohort generalization ability, and potential clinical application value.

[0033] 2.4 Expression characteristics of 9 characteristic genes in IgAN The glmBoost+Enet[α=0.4] model incorporated nine genes as diagnostic features. The expression differences of these genes between IgAN patients and LDs were analyzed in three cohorts. Results showed that in the GSE37460 cohort, HLA-DRA, VASH1, and RHBDD3 were significantly upregulated in IgAN patients (all p<0.05), while THBS1, FARP2, and TMSB4Y were highly expressed in LDs. Figure 5 A). In GSE93798, in addition to HLA-DRA, VASH1, and RHBDD3, CD160 and CX3CR1 were also significantly upregulated in IgAN patients, while EPHA4 and THBS1 were upregulated in LDs ( Figure 5B). In GSE115857, HLA-DRA, VASH1, and RHBDD3 were all significantly elevated in IgAN patients, while FARP2 was highly expressed in the LDs group ( Figure 5 C). In summary, HLA-DRA, VASH1, and RHBDD3 have significant diagnostic potential in differentiating IgAN from LDs.

[0034] In an external validation cohort, the expression trends of HLA-DRA, VASH1, and RHBDD3 were further validated. In GSE104948 ( Figure 6 A) and GSE99339 ( Figure 6 In B), all three, especially HLA-DRA and VASH1, showed a significant upward trend. In GSE115857, patients were divided into three groups according to the Schena classification: mild (G1), moderate (G2), and severe (G3). The results showed that the expression levels of HLA-DRA and VASH1 increased with the severity of the disease. Figure 6 C). In GSE116626, IgAN patients were divided into four groups according to pathological type: mild lesions, active lesions, chronic lesions, and mixed lesions. Among them, the expression of HLA-DRA and VASH1 was the highest in the mixed lesion group, suggesting that the upregulation of the two was positively correlated with the severity of the pathology. Figure 6 D).

[0035] 2.5 Relationship between key biomarkers and immune cell infiltration To investigate the differences in immune cell composition between IgAN patients and healthy individuals, a discriminant analysis was conducted. The results showed that the expression of various immune cells was significantly altered in the IgAN patient population, particularly in different subtypes of CD8+ T cells, CD4+ T cells, and regulatory T cells (Tregs). Figure 7 A). Subsequently, the study further explored the relationship between the HLA-DRA gene and the degree of immune cell infiltration in IgAN patients. The results showed that HLA-DRA was differentially correlated with various immune cells, with the most significant positive correlation showing with regulatory T cells, activated CD4+ T cells, and myeloid-derived suppressor cells (MDSCs). Figure 7 B). Furthermore, VASH1 is positively correlated with multiple types of immune cells, especially regulatory T cells, follicular helper T cells (Tfh), and MDSCs. Figure 7 C). These results reveal a significant synergistic effect between the expression characteristics of HLA-DRA and VASH1 and the infiltration of specific immune cells, which is of great significance in the immunopathology of IgAN.

[0036] Further analysis revealed that both HLA-DRA and VASH1 were significantly positively correlated with regulatory T cells, while HLA-DRA was also significantly correlated with activated CD4+ T cells. To further elucidate the relationship between these genes and different pathological stages of IgAN, a stage-related analysis was performed in the GSE115857 cohort. The results showed that in G3 stage IgAN patients, HLA-DRA had the strongest correlation with regulatory T cells (R=0.61, p=0.019), while VASH1 also showed a moderate correlation (R=0.5, p=0.058). Figure 8 A). In G2 stage patients, HLA-DRA was moderately associated with activated CD4+ T cells (R=0.47, p=0.013), and VASH1 was more strongly associated (R=0.64, p<0.01). Figure 8 A). Furthermore, in the GSE116626 cohort, VASH1 showed a positive correlation with regulatory T cells (R=0.62, p=0.06) and a similar trend with activated CD4+ T cells (R=0.48, p=0.17). Figure 8 B).

[0037] 2.6 Validation of key biomarkers for IgAN To verify the reliability of the aforementioned key biomarkers, a comprehensive analysis of IgAN patient samples was conducted (Table 2). Pathological examination of renal tissue from clinical patients using various staining techniques revealed significant changes in both structure and immunology. Figure 9 A). Immunohistochemical (IHC) validation analysis showed that, compared with the control group, the expression of HLA-DRA and VASH1 was significantly increased in IgAN patients, and semi-quantitative scoring analysis confirmed that the difference was statistically significant (Mann-Whitney U test, p<0.05); Figure 9 B–E). IHC results further revealed that HLA-DRA expression was mainly confined to the glomerular region, while its expression was weaker in the renal tubules; while VASH1 was significantly expressed in both the glomeruli and renal tubules. Immunofluorescence (IF) staining results showed that IgAN tissue contained more CD4+HLA-DRA+ functionally activated CD4+ T cells, and their number was significantly higher than that in control tissue. Figure 9 F). In summary, these results collectively reveal the pathological changes and immune cell infiltration characteristics of IgAN, confirm HLA-DRA and VASH1 as important biomarkers of IgAN, and highlight the key role of activated T cells in the disease mechanism.

[0038] 2.7 External calibration and SAP interpretability analysis of the IgAN gene model The study further conducted model calibration and interpretability assessment. Calibration curves were plotted in five independent validation cohorts (GSE37460, GSE93798, GSE99339, GSE104948, and GSE116626). Figure 10 The results show that the predicted curve is highly consistent with the ideal reference curve, indicating that the model's predicted probabilities have good reliability; the training queue (GSE115857) was not calibrated. The overall Brier score is 0.118, and the Brier values ​​of most validation queues are less than 0.15, suggesting that the model has good predictive consistency across multiple datasets.

[0039] Calibration curves compared the relationship between model predicted probabilities and actual event frequencies in five independent validation cohorts: internal validation (GSE37460, GSE93798); external validation (GSE99339, GSE116626, GSE104948). The pink curve represents the model's prediction calibration, and the blue dashed line is the theoretical ideal reference line (perfect calibration). The higher the overlap between the two, the more reliable the probability estimates on each dataset. GSE115857, as a training cohort, is only used for completeness demonstration and is not intended to claim external generalization performance. Each panel reports a Brier score, which is the mean squared error between predicted probabilities and actual outcomes; a lower value indicates better overall calibration and prediction accuracy. The overall Brier score in this study is 0.118, and most cohorts have Brier values ​​< 0.15, indicating good calibration of the model's predicted probabilities across datasets. Axes: x-axis represents predicted probabilities; y-axis represents observed event frequencies.

[0040] To enhance the biological interpretability of the model, the study employed the SHAP (Shapley Additive Explanations) method to perform interpretability analysis on a resilient network model (α=0.4) constructed based on nine genes (CD160, CX3CR1, EPHA4, THBS1, HLA-DRA, FARP2, VASH1, TMSB4Y, and RHBDD3). The SHAP analysis, based on data from six merged cohorts, quantified the contribution of each gene to the model's output probability. Figure 11 This further validates the biological rationale behind these characteristics in the pathogenesis of IgAN.

[0041] The summary SHAP (SHapley Additive exPlanations) plot illustrates the feature importance of nine genes (CD160, CX3CR1, EPHA4, THBS1, HLA-DRA, FARP2, VASH1, TMSB4Y, and RHBDD3) in machine learning models, covering six cohorts (GSE115857, GSE37460, GSE93798, GSE99339, GSE104948, and GSE116626). Each point represents a sample, the x-axis represents the SHAP value (impact on model output), and the color of the point represents the gene expression level (purple = high expression, blue = low expression). Genes are sorted by mean absolute SHAP value to highlight their relative contribution to diagnostic prediction, exhibiting a consistent pattern across cohorts. HLA-DRA, VASH1, and RHBDD3 show the highest influence, supporting their robustness and biological relevance in the pathogenesis of IgAN.

[0042] The aforementioned multi-cohort gene expression and machine learning analysis framework, by linking molecular signals with actionable diagnostic decisions, fully embodies the concept of precision nephrology. Immunohistochemical (IHC) results confirmed the differential expression of HLA-DRA and VASH1 in IgA nephropathy, supporting their potential value as diagnostic biomarkers. Overall, the final results, with their good reproducibility and external validation, represent a promising reference standard for future IgAN biomarker research; planned multicenter validation will further optimize its universality, calibration, and clinical application value.

[0043] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A diagnostic model for IgA nephropathy containing 9 genes, characterized in that, The nine genes mentioned are CD160, CX3CR1, EPHA4, THBS1, HLA-DRA, FARP2, VASH1, TMSB4Y, and RHBDD3. CD160, CX3CR1, HLA-DRA, VASH1, and RHBDD3 are significantly upregulated in patients with IgA nephropathy, while EPHA4, THBS1, FARP2, and TMSB4Y are expressed at higher levels in healthy individuals.

2. The diagnostic model according to claim 1, characterized in that: HLA-DRA and VASH1 are key diagnostic biomarkers.

3. A method for establishing a diagnostic model for IgA nephropathy containing 9 genes, characterized in that, The establishment method includes the following steps: S1. Data Acquisition: GSE115857 was downloaded from the GEO database as the training queue, GSE37460 and GSE93798 were used for model validation, and GSE104948, GSE116626 and GSE99339 were used as external validation queues. All queues contained a total of 207 IgA nephropathy patients and 56 living donors, and batch effect processing was performed. S2. Signaling pathway score calculation: First, GSEA was used to compare and evaluate the activation levels of different cell signaling pathways in all samples. In the three cohorts of GSE115857, GSE37460, and GSE93798, the differentially activated signaling pathways between IgA nephropathy patients and living donors were compared. At the same time, significantly different pathways were defined, and based on the consistent results of the three cohorts of GSE115857, GSE37460, and GSE93798, signaling pathways that were commonly and significantly activated in IgAN patients were screened. S3. Machine learning-based ensemble analysis model construction: Integrating nine machine learning algorithms, including Elastic Network Regression, LASSO, Ridge Regression, Generalized Linear Model with Forward Selection, Generalized Linear Model with Backward Elimination, Gradient Boosting Machine, Generalized Gradient Boosting Regression Model, Linear Discriminant Analysis, and Naive Bayes Classification Model. Among the 101 algorithm combinations generated, 80 successfully output effective results, which are suitable for subsequent comparative evaluation and construction of feature markers. Subsequent model optimization involved using a grid search method to fine-tune the hyperparameters of each algorithm within the training queue through ten-fold cross-validation. The final model determined was a two-stage glmBoost + Enet [α=0.4] model, which verified that HLA-DRA and VASH1 were stable and reliable biomarkers in each queue.

4. The method for establishing according to claim 3, characterized in that: The specific method for batch effect processing in step S1 is to use the ComBat algorithm in the "sva" software package to perform batch effect correction on the dataset downloaded from the GEO database. All standardization, feature selection and batch correction steps are performed independently within the training set of each fold cross-validation, thereby ensuring strict isolation between the training set and the test / validation set and avoiding information leakage.

5. The method for establishing according to claim 3, characterized in that: The method for defining significantly different pathways in step S2 is to define pathways with a log2 fold change value greater than 0.3 and a p value less than 0.05 as significantly different pathways.

6. The method for establishing according to claim 3, characterized in that, The specific method for constructing feature markers in step S3 includes the following steps: S3-1. The signaling pathways that were significantly activated in IgAN patients in step S2 were included in 80 algorithm combinations for analysis. S3-2. Each model is trained on the GSE115857 dataset, validated in the GSE37460 and GSE93798 cohorts, and finally externally validated on the GSE104948, GSE116626 and GSE99339 datasets that did not participate in the pathway selection process. S3-3. For each independent model, calculate the area under the receiver operating characteristic curve across all study cohorts.

7. The method for establishing according to claim 3, characterized in that: In step S3, the hyperparameters of each algorithm are tuned in the training queue using a grid search method and ten-fold cross-validation, including the α and λ parameters in Elastic Network / LASSO / Ridge Regression, the number of iterations and learning rate of GBM, and the shrinkage rate of glmBoost.

8. The method for establishing according to claim 3, characterized in that: After optimization, the model is evaluated for stability by calculating the average performance index through 100 random resampling operations, thereby minimizing the risk of overfitting. In order to prevent data leakage, all preprocessing steps are performed independently on each training set before model training. The validation set and the external test set are kept completely isolated until the final evaluation stage and do not participate in any part of model construction or parameter optimization.

Images