A method for screening systemic lupus erythematosus miRNA-gene combined biomarkers

By constructing a miRNA-gene joint regulatory network and using graph total variation to quantify local regulatory pathway signal fluctuations, the problem of difficulty in quantifying local regulatory pathway signal fluctuations and imbalances in cross-sectional expression data in existing technologies has been solved, enabling efficient mining and evaluation of SLE-related abnormal regulatory modules.

CN123120032APending Publication Date: 2026-11-17HANGZHOU DIANZI UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611349650.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-02
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing biomarker screening methods mostly focus on single molecules, making it difficult to quantify fluctuations and imbalances in local regulatory pathway signals in cross-sectional expression data. Existing network analysis relies solely on static topological features, failing to effectively capture the homeostatic collapse of local immune regulatory pathways during SLE pathogenesis.

Method used

Using graph signal processing theory, a miRNA-gene joint regulatory network was constructed. The energy of local topological high-frequency signals was quantified using graph total variation quantification to accurately locate the functional modules that disrupt the steady-state regulation during the pathogenesis of SLE. The analysis was then performed using graph clustering algorithms and graph Laplacian matrices.

Benefits of technology

It enables efficient discovery and evaluation of SLE-related abnormal regulatory modules, provides a computational scheme with signal theory depth, and improves the comprehensive discrimination ability of candidate molecule screening and the systematicness and robustness of disease state auxiliary assessment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123120032A_ABST
    Figure CN123120032A_ABST
Patent Text Reader

Abstract

This invention discloses a method for screening miRNA-gene joint biomarkers for systemic lupus erythematosus (SLE). The method includes: acquiring and standardizing gene and miRNA expression data; screening candidate differentially expressed genes and miRNAs; constructing a joint regulatory network containing nodes and interacting regulatory edges; performing single-factor network entropy perturbation analysis on each node in the network to obtain the perturbation contribution value of each node; using a graph clustering algorithm to divide the network into topological communities, obtaining multiple functionally independent sub-network modules, and calculating the average graph total variation of each module; screening high-risk modules based on the average GTV score, and extracting the miRNA-gene regulatory relationships within the modules as a combination of joint biomarkers. This invention uses network topology as the signal domain and accurately quantifies the degree of local regulatory steady-state collapse through graph total variation, effectively solving the problem of fluctuation analysis caused by the lack of temporal information in cross-sectional data, and outputting biologically meaningful regulatory relationships.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the fields of bioinformatics and computational biology, specifically relating to a method for screening systemic lupus erythematosus miRNA-gene co-markers based on graph signal processing. Background Technology

[0002] Systemic lupus erythematosus (SLE) is a complex autoimmune disease that can affect multiple organs and tissues, including the skin, joints, kidneys, hematopoietic system, and nervous system. The disease exhibits strong clinical heterogeneity and significant fluctuations in disease activity. Its development involves the synergistic effects of multiple mechanisms, including genetic background, abnormal immune regulation, environmental factors, and epigenetic regulation. Therefore, screening for biomarkers that can objectively reflect the molecular status of SLE is of great significance for elucidating disease mechanisms, molecular subtyping, and auxiliary assessment.

[0003] With the increasing availability of high-throughput sequencing and public expression profiling databases, screening for SLE-related molecular biomarkers based on transcriptome data has become a common research method. Existing methods often employ differential expression analysis, machine learning classification, protein-protein interaction networks, or functional enrichment analysis to screen for candidate key molecules from gene expression data. Some studies also analyze the role of miRNAs in SLE immune regulation from the perspective of non-coding RNAs, screening for differentially expressed miRNAs. Since miRNAs can influence gene expression by targeting mRNAs, their regulatory relationships with target genes can provide important clues for elucidating the molecular mechanisms of SLE.

[0004] However, existing biomarker screening methods still have the following shortcomings: First, most methods focus on the differential expression levels of single genes or single miRNAs, rarely constructing biomarker combinations based on the joint regulatory relationship between miRNAs and target genes, making it difficult to fully reflect the synergistic effect patterns of intermolecular networks; Second, conventional differential expression analysis mainly relies on statistical threshold screening, which does not adequately utilize the nonlinear associations and global features implicit in high-dimensional, small-sample, and multi-source expression data, and may miss molecular features that contribute significantly to the disease state but whose individual differences are not significant; Third, existing biological network analysis methods mostly focus on evaluating static topological indicators such as node degree and centrality, which can only reflect the positional importance of molecules in the network to a certain extent, but cannot quantify the dynamic fluctuation characteristics of molecular expression levels in the topological structure. Especially for clinically common cross-sectional expression data, SLE pathogenesis is often accompanied by the collapse of the homeostasis of local immune regulatory pathways (such as the decoupling of the normal inhibitory effect of miRNAs on target genes), and this deep-seated disruption of regulatory relationships is difficult to effectively capture through static centrality indicators.

[0005] To address the aforementioned issues, this invention introduces graph signal processing theory into molecular regulatory network analysis. Using network topology as the domain of discrete signals, it precisely quantifies the high-frequency signal fluctuation energy of local regulatory pathways through graph total variation (GTV). This allows for the identification of key functional modules experiencing steady-state imbalances in timeless cross-sectional data, providing a computational scheme with signal theory depth for the discovery and auxiliary evaluation of SLE molecular biomarkers. Summary of the Invention

[0006] (a) Purpose of the invention The purpose of this invention is to provide a method for screening systemic lupus erythematosus miRNA-gene combined biomarkers, in order to solve the problems that existing biomarker screening methods mostly focus on single molecules, and existing network analysis relies only on static topological features, making it difficult to quantify the degree of fluctuation and imbalance of local regulatory pathway signals in cross-sectional expression data.

[0007] To achieve the above objectives, this invention integrates multi-omics data to construct a joint regulatory network, introduces graph clustering algorithms and graph signal processing theory, maps network node perturbation values ​​to discrete graph domain signals, and uses graph total variation (GTV) to quantify the energy of local topological high-frequency signals in order to accurately locate the functional modules that regulate steady-state breaks during the pathogenesis of SLE. This provides a novel computational scheme with both physical and graph signaling interpretations for the discovery of SLE molecular biomarkers and the auxiliary assessment of disease status.

[0008] (II) Technical Solution To achieve the above objectives, the present invention adopts the following technical solution: A method for screening systemic lupus erythematosus miRNA-gene combined biomarkers includes the following steps: S1. Obtain gene expression data and miRNA expression data from healthy controls and systemic lupus erythematosus samples, and standardize the two types of expression data. S2, Analyze gene expression data using a statistical test model, screen differentially expressed genes, and rank them based on the importance score of the differentially expressed genes to obtain candidate genes; S3. Statistical testing models are used to analyze miRNA expression data, differentially expressed miRNAs are screened, and candidate miRNAs are obtained by ranking them based on their importance scores. S4. Using the candidate gene and the candidate miRNA as nodes, and the regulatory relationship between miRNA and gene and the interaction relationship between genes as edges, a miRNA-gene joint regulatory network is constructed. S5, calculate the initial network entropy of the miRNA-gene joint regulation network, and perform a node removal operation for each node in the joint regulation network, calculate the network entropy after removing the corresponding node, and determine the absolute value of the difference between the network entropy before and after node removal as the network perturbation contribution value of that node. S6. Based on the modularity, the miRNA-gene joint regulatory network is divided into topological communities to obtain several sub-network modules. Sub-network modules that contain at least one miRNA-gene regulatory relationship and at least one miRNA node and one gene node are selected as valid modules. S7, map the network perturbation contribution value of each node in each effective module to the graph signal on the corresponding node, and construct the corresponding graph Laplacian matrix according to the adjacency relationship of the effective modules; calculate the graph total variation of each effective module according to the graph signal and the graph Laplacian matrix, and obtain the average graph total variation of each effective module based on the number of nodes in the effective module; extract several effective modules with the highest average graph total variation, and use the miRNA-gene regulatory relationship in them as a joint marker.

[0009] Preferably, step S2 specifically includes: The t-test was used to perform differential expression analysis on the gene expression data. Differentially expressed genes that met the preset significance criteria were screened and input into the Gene-Transformer model. The importance score of each differentially expressed gene was obtained through the multi-head attention mechanism of the Gene-Transformer model. The genes were ranked according to the importance score and selected as candidate genes. The Gene-Transformer model includes a one-dimensional convolutional feature extraction layer, a positional encoding layer, a Transformer encoding layer, a global average pooling layer, and a classification output layer.

[0010] Preferably, in step S3, The process of obtaining candidate miRNAs includes: obtaining differentially expressed miRNAs using differential expression analysis, scoring the importance of the differentially expressed miRNAs using a machine learning model, ranking the differentially expressed miRNAs according to the importance scores, and taking the intersection of the machine learning screening results and the differentially expressed miRNAs to obtain candidate miRNAs.

[0011] Preferably, in step S4, the construction of the miRNA-gene joint regulatory network includes: Based on the database, the target gene relationships corresponding to candidate miRNAs were obtained. The miRNAs were filtered according to their binding region, minimum free energy, and experimental verification information. The miRNA-gene regulatory relationships with binding regions located in the 3′-UTR, minimum free energy less than -20 kcal / mol, and experimental verification support were retained. Input candidate genes into the STRING database to obtain gene-gene interaction relationships; The miRNA-gene regulatory relationship and the gene-gene interaction relationship are combined to construct the miRNA-gene joint regulatory network.

[0012] Preferably, in step S5, the change in network entropy before and after removing the node is determined as the network disturbance contribution value of that node, specifically including: The absolute value of the difference in network entropy before and after a node is removed is taken as the perturbation contribution value of that node; the larger the perturbation contribution value, the greater the impact of the corresponding node on the network structure. The network disturbance contribution value is not directly used as a screening criterion for joint markers, but is used as a graph signal in the subsequent sub-network modules.

[0013] Preferably, in step S6, the topological community division includes: The Louvain algorithm was used to maximize the network modularity as the global optimization objective. The miRNA-gene joint regulation network was divided into several initial communities through iterative iteration. The number of interaction regulatory edges in each initial community was counted, and the initial communities with zero interaction edges were removed.

[0014] Preferably, in step S7, The graph signal is a one-dimensional discrete graph signal vector. ; Constructing the corresponding graph Laplacian matrix specifically includes: Extract the adjacency matrix of the effective modules and node degree matrix Construct the graph Laplace matrix ; According to the quadratic energy formula Calculate the total variation of the graph for this subnetwork module.

[0015] Preferably, in step S7, the step of extracting several effective modules with the highest total variation of the average graph and using the miRNA-gene regulatory relationships within them as joint biomarkers includes: The average total variation of all effective modules was calculated. Modules with an average total variation of more than 3 times the average were defined as high-risk modules, modules with an average total variation of less than 1 / 3 of the average were defined as low-risk modules, and modules in between were defined as medium-risk modules. The miRNA-gene regulatory relationship in the high-risk modules was selected as a joint biomarker for SLE.

[0016] Preferably, the differential expression analysis method includes the limma-voom method, and the machine learning model includes the random forest model and the XGBoost model; the importance scores of miRNAs are obtained according to the random forest model and the XGBoost model respectively, and the top 50% of miRNAs in terms of importance scores are selected. The union of the screening results of the two machine learning models is then taken, and the intersection is taken with the differentially expressed miRNAs.

[0017] (III) Beneficial Effects Compared with the prior art, the present invention has the following beneficial effects: 1. This invention innovatively constructs a graph-domain discrete signal analysis system suitable for cross-sectional data. Addressing the pain point of clinical public datasets lacking time series data and hindering temporal fluctuation analysis, it maps molecular entropy perturbation values ​​to discrete signals on network topology nodes. Utilizing the graph Laplacian matrix as a second-order difference operator in the graph domain, it achieves an effective transfer of discrete fluctuation and high-frequency energy theory from signals and systems to static molecular regulatory networks, providing a novel physical perspective for cross-sectional data analysis lacking temporal information.

[0018] 2. A novel index for evaluating local regulatory imbalances based on graph total variation is proposed. This breaks through the limitations of existing studies that rely solely on node degree or static network centrality for evaluation. It utilizes GTV to quantify the sum of squared entropy perturbation differences between adjacent molecules. By capturing the local high-frequency changes formed by increased perturbation differences between adjacent nodes under disease states, it provides a quantitative method with clear mathematical interpretation for identifying abnormal regulatory modules related to SLE.

[0019] 3. Improved comprehensive discriminative ability in candidate molecule screening. This invention employs a hierarchical screening strategy combining statistical testing and intelligent feature learning. First, candidate molecules with inter-group expression differences are screened using statistical testing on gene expression data. Then, the multi-head attention mechanism of the Gene-Transformer model is used to rank the importance of candidate genes, thus considering both the significance of single-molecule differences and the global association features between multiple genes. For miRNA expression data, a multi-strategy screening is performed using differential expression analysis, random forest, and XGBoost models. Compared with feature screening based solely on statistical significance or a single machine learning model, this method can comprehensively evaluate candidate molecules from different dimensions, improving the reliability and representativeness of subsequent node screening in the miRNA-gene joint regulatory network. This method considers differential significance, classification contribution, and potential global association features. 4. This invention achieves efficient screening of joint biomarkers and synergistic multidimensional features. From differential feature extraction at the front end to topological community partitioning and local high-frequency energy assessment at the back end, it forms a coherent and highly self-consistent analytical loop. The final output is not an isolated single-molecule biomarker, but a miRNA-gene joint biomarker with clear regulatory relationships and potential biological significance, significantly improving the systematicity and robustness of the disease molecular state-assisted assessment model. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a Venn diagram for screening candidate differentially expressed miRNAs in this invention. Figure 3 This is a diagram of the miRNA-gene joint regulatory network constructed in this invention; red rectangular nodes represent miRNAs, and green elliptical nodes represent genes. Figure 4 This is a diagram showing the effective functional modules selected by this invention and their network interaction relationships; Figure 5 This is a diagram showing the control relationships between the internal nodes of modules 1 and 4, which are considered high-risk components of this invention. In this diagram, A represents the control relationship of module 1, and B represents the control relationship of module 4. Detailed Implementation

[0021] The present invention will be further described below with reference to specific embodiments. It should be understood that the following embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. All equivalent substitutions, parameter adjustments or conventional improvements made based on the technical concept of the present invention should be included within the scope of protection of the present invention.

[0022] Example 1: Screening method for miRNA-gene combined biomarkers in systemic lupus erythematosus This embodiment uses publicly available expression profile data as an example to describe in detail the specific implementation process of the systemic lupus erythematosus miRNA-gene combined biomarker screening method described in this invention. All data processing and analysis steps were completed in the R language and Python environment.

[0023] Step 1: Data Acquisition and Preprocessing Gene expression data and miRNA expression data related to systemic lupus erythematosus were obtained from the Gene Expression Omnibus (GEO) database.

[0024] The dataset information used in this embodiment is shown in Table 1. The gene expression data includes four independent datasets (GSE30153, GSE39088, GSE45291, and GSE61635), which, after merging, yielded 609 samples, including 486 SLE samples and 123 healthy control samples, with a total of 12,807 usable genes. The miRNA expression data includes two independent datasets (GSE175840 and GSE79240). The following preprocessing procedure was used to process each dataset separately: (1) Use the "GEOquery" package in R language to read the expression matrix file and the corresponding platform annotation file of each dataset, and extract the expression value and probe annotation information; (2) According to the platform annotation file, convert the probe number into the corresponding standard gene name (Gene Symbol) or standard miRNA name (e.g., hsa-miR-125a-5p). For cases where the same gene or miRNA corresponds to multiple probes, take the arithmetic mean of the expression values ​​of all corresponding probes as the final expression value of the gene or miRNA in the sample; probes that do not match a valid gene name or miRNA name are removed. (3) Remove genes and miRNAs whose expression values ​​are all missing values ​​(NA) or whose expression values ​​are below the detection limit of the platform in all samples; also remove genes or miRNAs with a missing value ratio of more than 20% to ensure the data quality of subsequent analysis. (4) The disease status label corresponding to each sample is used as the classification label. SLE samples are labeled as 1 and healthy control samples are labeled as 0, which are used for subsequent differential expression analysis and model training.

[0025] Table 1: Dataset Information Table Step 2: Screening for differentially expressed genes First, genes co-occurring in four gene expression datasets (GSE30153, GSE39088, GSE45291, and GSE61635) were identified, yielding 12,807 common genes. Based on this, a t-test was used to analyze the differential expression of SLE samples and healthy control samples in each of the four datasets. A p-value ≤ 0.01 was used as the significance threshold for screening. The union of the results from the four differentially expressed gene sets yielded 3,542 genes, which were then used as candidate input features for the subsequent Gene-Transformer model.

[0026] The four gene expression datasets were merged to obtain a comprehensive dataset containing 609 samples (including 486 SLE samples and 123 healthy control samples). The normalized expression matrices corresponding to the above 3542 genes were then input into the Gene-Transformer model for further filtering. The structure of the Gene-Transformer model is as follows: (1) Input layer: Accepts a gene expression matrix with dimensions (batch_size, n_genes), where n_genes is the number of input genes (i.e., 3542), and batch_size is the batch size, which is dynamically adjusted according to the memory capacity during training.

[0027] (2) Convolutional Feature Extraction Layer: A one-dimensional convolutional layer is used to extract local features from the input gene expression matrix. The kernel size is 3, the stride is 1, the number of output channels is 64, and the activation function is Swish to enhance the nonlinear expressive power of the model.

[0028] (3) Position encoding layer: Since the Transformer encoder itself does not have the ability to perceive sequence order, a learnable position encoding vector is added to the convolution output, so that the model can perceive the relative position information of different genes in the input sequence and improve the quality of feature representation.

[0029] (4) Transformer encoding layer: A single-layer Transformer encoder structure is adopted, the number of multi-head attention heads is set to 4, the hidden layer dimension of the feedforward network is set to 256, and the global dependency relationship between genes is captured through the self-attention mechanism.

[0030] (5) Global average pooling layer: The output of the Transformer encoder is globally average pooled along the gene dimension, and the feature matrix of each sample is compressed into a one-dimensional vector, which is used as the input of the subsequent classification layer.

[0031] (6) Classification output layer: A fully connected layer with a softmax activation function is used, and the output dimension is 2, which corresponds to the binary classification probability of SLE samples and healthy control samples respectively.

[0032] The main training hyperparameters of the Gene-Transformer model are shown in Table 2.

[0033] Table 2: Training parameters of the Gene-Transformer model After model training, the attention weight matrix corresponding to the multi-head attention layer in the Transformer coding layer is extracted. The average attention score of each gene across all attention heads and all samples is taken as the final importance score for that gene. Genes are sorted from highest to lowest based on the absolute value of their attention scores, and the 1000 genes with the highest attention scores are selected as candidate differentially expressed genes for subsequent miRNA-gene regulatory relationship integration and network construction.

[0034] The aforementioned tandem strategy of "T-test initial screening + Gene-Transformer fine screening" combines the advantages of both statistical inference and deep learning modeling. On the one hand, the T-test screens candidate gene sets with significant expression differences at the single-gene level, effectively controlling the input size of subsequent models and reducing the risk of overfitting in high-dimensional sparse data. On the other hand, the Gene-Transformer model, through a multi-head self-attention mechanism, learns the global co-expression dependencies between genes on merged multi-cohort large-sample data. Its output attention score not only reflects the discriminative contribution of each gene to the SLE / health binary classification task, but more importantly, genes with higher attention weights in the attention weight matrix tend to be functionally interconnected and jointly participate in similar biological pathways or regulatory processes. Therefore, the 1000 candidate genes obtained through this tandem process not only have statistically significant inter-group expression differences and feature importance ranking based on deep learning, but also contain potential synergistic regulatory relationships between genes, providing a high-quality set of gene nodes with representativeness, discriminative power, and functional correlation for the subsequent construction of miRNA-gene joint regulatory networks.

[0035] Step 3: Screening for differentially expressed miRNAs This study employs a multi-strategy ensemble approach combining the limma-voom method with random forest and XGBoost models to screen candidate differentially expressed miRNAs in the GSE175840 and GSE79240 miRNA expression datasets. The core test of the limma-voom method is essentially a modified t-test adjusted for Bayesian variance, suitable for differential expression analysis of RNA-seq counting data. However, the Gene-Transformer model relies on a large number of genes to effectively learn global attention patterns, making it less effective in scenarios with a small number of miRNAs (typically only a few hundred). Therefore, this study uses both random forest and XGBoost machine learning algorithms to assess the importance of miRNAs. This ensemble strategy, on the one hand, uses limma-voom to screen for significantly differentially expressed miRNAs from a statistical inference perspective; on the other hand, it utilizes machine learning models to capture miRNA molecules that contribute significantly to disease state discrimination from a feature importance dimension. The two approaches are complementary and particularly suitable for miRNA expression data with small sample sizes and relatively low feature dimensions, effectively identifying low-expression but biologically significant miRNA patterns. XGBoost and Random Forest have been widely used in biomarker discovery research in recent years, and their high accuracy and good generalization ability have been verified in molecular screening of multiple diseases.

[0036] The specific screening process is as follows: First, the limma-voom method was used to perform differential expression analysis on the GSE175840 and GSE79240 miRNA expression datasets, respectively. The selection thresholds were set as adj.P.Val < 0.05 and |log2FC| > 0.5, where adj.P.Val is the adjusted p-value after Benjamini-Hochberg correction, and log2FC is the base-2 fold change.

[0037] Subsequently, the XGBoost and Random Forest models were used to evaluate the inter-group classification importance of all expressed miRNAs in the two datasets. The XGBoost model used a binary logistic objective function with a learning rate of 0.3, a maximum tree depth of 6, and 100 weak learners. The Random Forest model used 500 decision trees, with other parameters set to default. Both machine learning algorithms used the absolute value of the Feature Importance Score as a measure of miRNA contribution to the SLE / health classification. The union of the top 50% of miRNAs by importance score in the XGBoost model and the top 50% of miRNAs by importance score in the Random Forest model was used as the candidate set for machine learning selection of this dataset.

[0038] Finally, the limma-voom differential analysis results were integrated with the candidate set selected by machine learning, and the intersection was taken as the final candidate differentially expressed miRNAs for this dataset. In the GSE175840 dataset, 201 candidate differentially expressed miRNAs were obtained; in the GSE79240 dataset, 225 candidate differentially expressed miRNAs were obtained. Further intersection analysis of the candidate differentially expressed miRNAs from the two datasets yielded 24 miRNAs that were differentially expressed in both datasets, which served as the miRNA node set for constructing the miRNA-gene co-regulatory network. The intersection of the two dataset selection results is shown below. Figure 2 As shown.

[0039] Step 4: Screening of miRNA-gene co-regulatory relationships and network construction The 24 candidate differentially expressed miRNAs obtained in step 3 were used as input, and their corresponding target gene relationships were predicted using the miRWalk2.0 database (which integrates multiple target gene prediction resources such as miRWalk, miRTarBase, and TargetScan). To ensure the reliability and biological relevance of the prediction results, the following criteria were used for screening and filtering: (1) The binding region of miRNA to target genes is located in the 3'-untranslated region (3′-UTR) of mRNA, which is the main site of action for miRNA to regulate target genes; (2) The minimum free energy (MFE) of the miRNA-mRNA dimer is less than -20 kcal / mol to ensure that the binding has sufficient thermodynamic stability; (3) The miRNA-gene targeting relationship is supported by experimental verification, that is, there is a corresponding experimental verification number in the miRTarBase database (such as Luciferase reporter assay, Western blot or qPCR verification).

[0040] After screening based on the above criteria, among the 24 candidate differentially expressed miRNAs, four miRNAs—hsa-miR-191-3p, hsa-miR-202-5p, hsa-miR-3181, and hsa-miR-4728-3p—failed to obtain target gene relationships that met the criteria and were therefore removed. The remaining 20 miRNAs all obtained at least one target gene relationship that met the above criteria, and the corresponding miRNA target gene sets were obtained.

[0041] The 1000 candidate differentially expressed genes obtained in step 2 using the Gene-Transformer model were subjected to intersection analysis with the target gene set of the aforementioned 20 miRNAs. Genes that simultaneously met the dual criteria of "differentially expressed gene" and "predicted target gene of differentially expressed miRNA" were retained, ultimately yielding 107 genes as a high-quality co-regulatory candidate gene set. To illustrate representative examples of miRNA-gene regulatory relationships, a representative target gene from the target genes of each miRNA is listed in Table 3.

[0042] To further expand the functional associations between genes, the aforementioned 107 genes were input into the STRING database. A comprehensive confidence threshold of at least 0.4 was set to retrieve and obtain known and predicted interactions between gene-encoded proteins. The STRING output was then parsed and filtered, retaining only the interaction relationships between the aforementioned 107 genes and removing connections to genes outside the set and isolated nodes, ensuring that the network focuses on regulatory and interaction information within the target gene set.

[0043] Subsequently, Cytoscape software was used to integrate the aforementioned miRNA-gene regulatory relationships (derived from miRWalk 2.0 predictions) with gene-gene interaction relationships (derived from the STRING database) to construct a miRNA-gene joint regulatory network, such as... Figure 3 As shown in the diagram, this network contains two types of nodes: diamond-shaped nodes represent miRNAs, and circular nodes represent genes, forming a joint regulatory network view of miRNAs and genes working synergistically, as illustrated below. Figure 3 As shown.

[0044] Table 3: Representative miRNA target gene relationships Step 5: Single-factor network entropy perturbation analysis To quantify the structural importance of each node in the miRNA-gene co-regulatory network, this step employs a single-factor network entropy perturbation analysis method. By removing nodes one by one and observing the changes in the overall network structure, the contribution of each node to network stability is evaluated.

[0045] First, the degree distribution of each node in the miRNA-gene co-regulation network is calculated. For each node in the network... Its degree value Defined as the total number of edges connected to this node (treating the network as an undirected graph). After counting the degree values ​​of all nodes in the network, the degree distribution probability is calculated. : in, The degree value is equal to The number of nodes, This represents the total number of nodes in the network.

[0046] Subsequently, the network entropy is calculated based on the degree distribution probability. As a holistic indicator for measuring the orderliness of network structure: Among them, for all different degree values The values ​​are summed. A higher network entropy indicates a more dispersed degree distribution and a more complex structure; a lower entropy indicates a more concentrated or regular network structure. In this embodiment, the network entropy of the constructed miRNA-gene joint regulatory network in its initial state is... It is 1.971.

[0047] To assess the perturbation effect of a single node on the overall network structure, each miRNA node or gene node in the network was removed sequentially. The specific steps are as follows: (1) Regarding the first in the network A node is temporarily removed from the network along with all its associated edges; (2) Recalculate the degree distribution of the remaining network after deleting nodes. Then, the network entropy of the remaining network is calculated. ; (3) Restore the deleted node and all its associated edges, and then proceed to the removal operation of the next node; (4) After traversing all nodes in the network and completing the above operations, calculate the disturbance contribution value of each node based on the network entropy change before and after removal. : in, The network entropy of the initial complete network, To remove the first Network entropy after each node.

[0048] The larger the value, the greater the perturbation contribution of the node to the overall structure of the miRNA-gene joint regulatory network, indicating that the node occupies a more critical position in the network structure. In this embodiment, the top 20 nodes in terms of perturbation contribution value are shown in Table 4. As can be seen from the table, the top-ranked nodes are mainly miRNAs, such as hsa-miR-1285-3p and hsa-miR-125a-5p, suggesting that these miRNAs have a high structural influence in the network and may play an important role in SLE-related regulatory imbalances.

[0049] Table 4: Top 20 nodes by disturbance contribution value Step 6: miRNA effective module segmentation To accurately quantify signal fluctuations in local regulatory pathways, this embodiment employs the Louvain algorithm to perform topological community partitioning of the miRNA-gene co-regulatory network. The Louvain algorithm is a hierarchical community detection algorithm based on modularity optimization. Q The overall optimization goal is: in: This represents the total number of edges in the network. For elements of the adjacency matrix, Represents a node and nodes If there is an edge connecting them, then the value is 0; For nodes The degree; For nodes The community to which it belongs; For indicator functions, when The value is 1 if it is true, and 0 otherwise.

[0050] Modularity Q The value range is [-1, 1]. The closer the value is to 1, the higher the quality of the community division, that is, the tighter the connection within the community and the sparser the connection between communities.

[0051] The specific execution flow of the algorithm is as follows: (1) Initialization: Assign each node in the network to its own community, that is, the number of communities is equal to the number of nodes at the beginning; (2) First stage local movement: For each node traversing all its neighboring communities Calculate the nodes Move from current community to community Modularity gain at time If a positive one exists... Then the node Move to The largest community. Repeat this process until no node can be moved (i.e., all). ); (3) Second stage network aggregation: Aggregate the communities formed in the first stage into new "super nodes", merge the edge weights within the community into the self-loop of the super node, and merge the edge weights between communities into the edge between the super nodes; (4) Iteration: The aggregated network is used as the new input, and the first and second stages are repeated until the modularity no longer increases and the algorithm converges.

[0052] After iterative convergence using the Louvain algorithm, the original miRNA-gene co-regulatory network was automatically divided into 33 initial communities. To further screen for biologically significant functional modules, the following filtering steps were performed: (1) Count the number of interaction control edges (EdgeCount) within each initial community, i.e., the number of edges between nodes within the community; (2) Remove isolated node communities with EdgeCount = 0 (i.e., communities containing only a single node and no internal edges). (3) Retain communities with EdgeCount ≥ 1 as valid functional modules, and require each valid module to contain at least one miRNA node and one gene node. That is, communities consisting only of gene nodes will not be included in subsequent analysis, so as to ensure that the module has the biological connotation of miRNA-gene joint regulation. (4) After the above screening, 17 effective functional modules were finally obtained. For ease of subsequent description and reference, the modules are sorted and numbered from most to least number of nodes (module 1 to module 17); if the number of nodes is the same, they are arranged in alphabetical order of the miRNA names in the module. The relationship between the 17 modules is as follows: Figure 4 As shown, it can be seen that three modules (module 6, module 9, and module 14) are not associated with other modules and are considered isolated modules.

[0053] Table 5 details the numbering, nodes, number of nodes, number of miRNAs, and number of genes for each effective functional module. As shown in Table 5, the effective functional modules exhibit significant differences in the number of nodes and the ratio of miRNAs to genes. Modules 1 and 2, containing 7 nodes (1 miRNA and 6 genes), are the largest and most diverse in terms of miRNA types in this embodiment, suggesting they may play a complex synergistic regulatory role in the SLE-related miRNA-gene joint regulatory network. Modules 3 and 4, containing 6 and 5 nodes respectively, also have high regulatory network density. The remaining modules are relatively small (2–4 nodes), all being compact, functionally focused local regulatory units. These functional modules will serve as the basic computational units for subsequent graph signal processing, used to calculate the graph total variation (GTV) of each module and identify high-risk modules.

[0054] Table 5 Node Details of Effective Functional Modules Step 7: High-risk module locking and marker output To transfer the discrete signal energy theory of signals and systems to timeless cross-sectional molecular networks, this step defines the entropy perturbation contribution of each node as the discrete signal on the graph node, and uses the graph Laplacian matrix to calculate the graph total variation (GTV) of each functional module to quantify the steady-state imbalance of the local control network.

[0055] (1) Construction of one-dimensional graph signal vectors For each valid functional sub-network module selected in step 6, extract the entropy perturbation values ​​of all nodes within that module (i.e., the values ​​calculated in step 5). ), forming a one-dimensional graph signal vector : in, This represents the number of nodes in this sub-network module. For the first The disturbance contribution value of each node.

[0056] (2) Construction of sub-Laplace matrix Extract the adjacency matrix of this sub-network module ,in Represents a node and nodes The nodes are connected by an edge; otherwise, the value is 0. Calculate the node degree matrix. It is a diagonal matrix, with diagonal elements Constructing the sub-Laplace matrix : The graph Laplacian matrix can be viewed as a generalization of the one-dimensional discrete difference operator to graph structures, and is used to measure the degree of change of graph signals between adjacent nodes. (3) Total variational calculation of graph Calculate the graphical total variation (GTV) of each module using the quadratic energy formula: After expansion, it is equivalent to: in, GTV represents the set of edges within a subnetwork module. This expression represents the sum of squared differences in the molecular entropy perturbations at both ends of all interacting edges within the subnetwork. A larger GTV value indicates a more drastic difference in entropy perturbations between topologically adjacent molecules (a higher degree of change in the local graph signal), signifying a severe disruption of the miRNA-gene regulatory pathway homeostasis; a GTV close to 0 indicates that the molecular perturbation values ​​are converging, and the regulatory state is relatively stable.

[0057] For ease of reading and comparison, the GTV values ​​of all modules in this embodiment are uniformly multiplied by 10. 4 This is to indicate that the unit of GTV in the following table and text is ×10. -4 This scaling operation does not change the relative order or risk level classification between modules.

[0058] (4) Total variation calculation of the mean plot Since the graph total variation of a module is directly related to the number of nodes in the module—more nodes usually mean more internal edges, naturally resulting in a higher absolute value of the graph total variation (GTV)—directly comparing the absolute value of GTV across modules introduces bias due to the number of nodes. To eliminate the influence of module size on GTV and make modules of different sizes comparable, this embodiment further calculates the average graph total variation for each module. , defined as the total GTV of the module divided by the number of nodes in that module: Where n is the number of nodes in the sub-network module. The average GTV reflects the high-frequency signal energy of a unit node within the module, and can more objectively measure the average steady-state imbalance of each node, facilitating fair comparison and risk ranking of modules of different sizes.

[0059] (5) Risk level classification and biomarker combination screening Based on the average GTV calculation results, the modules are classified into risk levels according to the following strategy, and marker combinations are selected: ① Sort the sub-network modules from highest to lowest based on their average GTV. The higher the total variation of the average graph, the more significant the local topological disturbance imbalance and the higher the energy of high-frequency changes in that module; ② Based on the average GTV score, modules are divided into three levels: high-risk, medium-risk, and low-risk. Commonly used classification methods include the quartile method, the mean ± standard deviation method, and the mean multiple method. This embodiment uses the mean multiple method based on the actual data distribution characteristics: the average GTV of all 17 effective functional modules is calculated to be 12.769. Modules with a GTV more than 3 times the average (i.e., >38.307) are defined as high-risk modules, modules with a GTV less than 1 / 3 of the average (i.e., <4.256) are defined as low-risk modules, and modules in between are defined as medium-risk modules. ③ Select all high-risk modules as target modules, extract all miRNA-gene regulatory relationships within the high-risk modules, and form a combination of miRNA-gene joint biomarkers for systemic lupus erythematosus.

[0060] The module GTV, average GTV, and risk level classification results for each effective functional module are shown in Table 6.

[0061] Table 6. GTV, average GTV, and risk level of effective functional modules. GTV values ​​have been multiplied by 10. 4 The actual value is the value in the table multiplied by 10. -4 .

[0062] As shown in Table 6, high-risk modules include Module 1 and Module 4. Module 1 has an average GTV of 49.95, significantly higher than all other modules, indicating extremely significant differences in entropy perturbations between adjacent nodes within this module, and the most severe local regulatory steady-state collapse. Therefore, it was identified as the primary target module. This module contains 7 nodes (1 miRNA: hsa-miR-1285-3p; 6 genes: FRMD8, ITPRIPL2, STRN3, TMEM169, UBXN2A, UFM1) and 6 interacting edges. Its internal regulatory network is shown in Figure 5A. Module 4 has an average GTV of 46.82 and is also listed as a high-risk module. It contains 5 nodes (1 miRNA: hsa-miR-125a-5p; 4 genes: ARIH2, PREPL, TDG, ZNF385A). Its internal regulatory network is shown in Figure 5B.

[0063] The medium-risk modules include eight modules: Module 6, Module 2, Module 16, Module 17, Module 3, Module 8, Module 9, and Module 15. The average GTV ranges from 4.256 to 38.307, exhibiting moderate steady-state imbalance characteristics. These modules can be used as a combination of secondary biomarkers for future research.

[0064] The low-risk modules include 7 modules: Module 13, Module 12, Module 5, Module 11, Module 7, Module 10, and Module 14. Their average GTV is below 4.256, their local regulatory networks are relatively stable, and their association with SLE lesions is low, so they are excluded.

[0065] Extract all miRNA-gene regulatory relationships contained in the high-risk modules (modules 1 and 4) to obtain miRNA-gene joint biomarker combinations: hsa-miR-1285-3p → FRMD8 hsa-miR-1285-3p → ITPRIPL2 hsa-miR-1285-3p → STRN3 hsa-miR-1285-3p → TMEM169 hsa-miR-1285-3p → UBXN2A hsa-miR-1285-3p → UFM1 hsa-miR-125a-5p → ARIH2 hsa-miR-125a-5p → PREPL hsa-miR-125a-5p → TDG hsa-miR-125a-5p → ZNF385A The above 10 miRNA-gene regulatory relationships constitute a combination of biomarkers for systemic lupus erythematosus.

[0066] Step 8: Validation of the biological relevance of the regulatory relationship To verify the biological rationale for the miRNA-gene co-markers screened in the high-risk modules of this embodiment and their potential association with systemic lupus erythematosus (SLE), literature verification and biological relevance assessment were conducted on the 10 miRNA-gene regulatory relationships obtained from modules 1 and 4. Specifically, using each miRNA-gene regulatory relationship as the basic evaluation unit, relevant studies were searched in the PubMed database, focusing on whether the regulatory relationships had existing targeted evidence and whether they involved biological processes related to SLE development, such as inflammatory responses, immune homeostasis, interferon signaling, inflammasome regulation, protein homeostasis, and epigenetic regulation. For regulatory relationships that lacked direct SLE verification, indirect evidence was evaluated by combining existing targeted relationships and their related immunobiological functions.

[0067] The association results between the 10 miRNA-gene regulatory relationships and SLE literature are shown in Table 7. Among them, hsa-miR-125a-5p→ARIH2 has strong biological evidence, as it can be associated with miR-125a-5p abnormalities, ARIH2-mediated ubiquitination regulation, and inflammasome-related processes. The regulatory relationships hsa-miR-1285-3p→FRMD8, hsa-miR-1285-3p→UFM1, and hsa-miR-125a-5p→TDG are respectively associated with SLE-related biological processes such as inflammatory factor release, methylation-mediated immune regulation, and DNA demethylation. Although some regulatory relationships lack direct functional verification in SLE, they can serve as novel candidate combined biomarker relationships obtained in this embodiment. The above results indicate that the high-risk miRNA-gene combined biomarker combinations obtained in this embodiment have certain immunological and disease biological basis, providing support for the biological relevance and reliability of the screening results.

[0068] Table 7: Biological validation of the association between miRNA-gene regulatory relationships and SLE Existing biomarker screening methods based on biological networks typically start from the differential expression level of individual molecules or the importance of network topology, using indicators such as node degree, centrality, and network entropy to evaluate the structural importance of individual nodes in the overall network. However, these methods mainly focus on the network position of individual nodes and their impact on the overall network structure, making it difficult to further characterize the perturbation differences in network states between adjacent nodes, and also difficult to identify whether there are abnormal structural changes between different molecules in local regulatory relationships. For cross-sectional expression data of systemic lupus erythematosus, due to the lack of observational information at continuous time points, it is difficult to directly describe the dynamic fluctuations of molecular regulatory relationships in the disease state based on time series changes. Therefore, it is necessary to establish a method that does not rely on time series, but can use cross-sectional expression data and its network topology to evaluate the imbalance of local regulatory states.

[0069] Furthermore, existing biomarker screening methods typically perform differential expression analysis or single-model screening on genes or miRNAs separately, ultimately outputting mostly independent single-molecule biomarkers. This makes it difficult to fully utilize the regulatory relationships between miRNAs and target genes, as well as their network-based synergistic characteristics. Therefore, this invention not only screens candidate features for genes and miRNAs but also combines statistical tests with deep learning or machine learning to comprehensively evaluate candidate molecules from different dimensions such as differential expression and feature importance. Based on this, a joint regulatory network is constructed, including miRNA nodes, gene nodes, and miRNA-gene regulatory edges and gene-gene interaction edges, thus providing a high-quality network foundation for subsequent topology-based joint biomarker screening.

[0070] This invention further discovers that the change in network entropy caused by the removal of a single node not only reflects the degree of perturbation of the overall network structure by that node, but also serves as a perturbation quantity characterizing the change in the structural state of that node. However, if nodes are ranked solely based on this perturbation quantity, only the structural importance at the single-node level can be obtained, failing to further reflect whether the perturbation levels between adjacent nodes are consistent, or whether there are significant local structural changes in the regulatory relationships between adjacent nodes. Therefore, this invention does not directly use the node perturbation contribution value as the final marker, but instead maps the network perturbation contribution value of each node to the graph structure of the joint regulatory network, forming a discrete graph signal defined on the network topology, and further quantifies the differences between the perturbation states of adjacent nodes using the adjacency relationships between nodes in the network.

[0071] Specifically, a node removal operation is performed on each node in the joint control network. The network perturbation contribution value is obtained based on the change in network entropy before and after node removal, thus acquiring node-level information reflecting the degree of structural perturbation at different nodes. Furthermore, the network perturbation contribution value is used as a discrete graph signal on the corresponding network node, establishing a correspondence between the perturbation information, which originally only had single-node scalar significance, and the network topology. This allows for direct calculation and comparison of the perturbation contribution value differences between adjacent nodes in the graph domain. By combining network entropy perturbation analysis with graph signal processing, a transformation from "single-node structural importance evaluation" to "local topology control imbalance evaluation" is achieved.

[0072] Based on this, the present invention divides communities according to the topological connections of the joint control network to obtain multiple control sub-network modules, and uses these modules as the calculation range for local changes in the graph signal, rather than as the final output marker. For each control sub-network module, a corresponding adjacency matrix, node degree matrix, and graph Laplace matrix are established, and the graph signal is constructed using the perturbation contribution values ​​of each node within the module, and the corresponding graph total variation energy is calculated. The graph total variation energy is essentially used to characterize the degree of change between the network perturbation contribution values ​​of adjacent nodes within the same local topology: when the perturbation contribution values ​​of adjacent nodes are relatively close, the graph total variation is low, indicating that the structural perturbation state in this local control network is relatively stable; when there is a large difference in the perturbation contribution values ​​between adjacent nodes, the graph total variation increases, indicating that there is a relatively obvious perturbation change between adjacent nodes in the local topology, thus reflecting that this local control region has a significant steady-state imbalance characteristic.

[0073] Through the above processing, this invention establishes a continuous processing chain of "multi-strategy screening of candidate molecules—construction of joint regulatory networks—network node perturbation analysis—graph domain signal representation—quantification of topological neighborhood changes—identification of high-risk local regulatory regions—extraction of regulatory relationships." Specifically, front-end statistical testing and deep learning or machine learning methods are used to improve the comprehensive discriminative ability of candidate gene and miRNA screening; network entropy perturbation contribution values ​​are used to provide structural perturbation information at the node level; graph topology is used to define the spatial correlation of the perturbation information; graph total variation is used to quantify the degree of inconsistency in perturbation states between adjacent nodes; and the regulatory sub-network module is used to limit the local range of abnormal topological changes.

[0074] It is important to note that the regulatory sub-network module described in this invention is only used for local graph signal calculation, steady-state imbalance evaluation, and high-risk region localization. The final output joint biomarker is not a single miRNA or a single gene, nor is it the local sub-network module itself, but rather a miRNA-gene regulatory relationship extracted from the high-risk module. This miRNA-gene regulatory relationship is distinct from the gene-gene interaction relationship in the joint regulatory network. It uses the specific regulatory edge between the miRNA and the target gene as the basic unit of the joint biomarker, thereby achieving the transformation from a single-molecule biomarker to a "miRNA-gene regulatory relationship" joint biomarker. Without relying on time-series data, this invention utilizes cross-sectional expression data and its network topology to identify local regulatory steady-state imbalances and further obtains SLE joint biomarkers with topological association and regulatory relationship characteristics.

Claims

1. A method for screening systemic lupus erythematosus miRNA-gene combined biomarkers, characterized in that, Includes the following steps: S1. Obtain gene expression data and miRNA expression data from healthy controls and systemic lupus erythematosus samples, and standardize the two types of expression data. S2, Analyze gene expression data using a statistical test model, screen differentially expressed genes, and rank them based on the importance score of the differentially expressed genes to obtain candidate genes; S3. Statistical testing models are used to analyze miRNA expression data, differentially expressed miRNAs are screened, and candidate miRNAs are obtained by ranking them based on their importance scores. S4. Using the candidate gene and the candidate miRNA as nodes, and the regulatory relationship between miRNA and gene and the interaction relationship between genes as edges, a miRNA-gene joint regulatory network is constructed. S5, calculate the initial network entropy of the miRNA-gene joint regulation network, and perform a node removal operation for each node in the joint regulation network, calculate the network entropy after removing the corresponding node, and determine the absolute value of the difference between the network entropy before and after node removal as the network perturbation contribution value of that node. S6. Based on the modularity, the miRNA-gene joint regulatory network is divided into topological communities to obtain several sub-network modules. Sub-network modules that contain at least one miRNA-gene regulatory relationship and at least one miRNA node and one gene node are selected as valid modules. S7, map the network disturbance contribution value of each node in each effective module to the graph signal on the corresponding node, and construct the corresponding graph Laplace matrix according to the adjacency relationship of the effective module; calculate the graph total variation of each effective module according to the graph signal and the graph Laplace matrix, and obtain the average graph total variation of each effective module based on the number of nodes in the effective module; Extract the most effective modules with the highest total variation in the average plot and use the miRNA-gene regulatory relationships within them as joint biomarkers.

2. The method according to claim 1, characterized in that, Step S2 specifically includes: The t-test was used to perform differential expression analysis on the gene expression data. Differentially expressed genes that met the preset significance criteria were screened and input into the Gene-Transformer model. The importance score of each differentially expressed gene was obtained through the multi-head attention mechanism of the Gene-Transformer model. The genes were ranked according to the importance score and selected as candidate genes. The Gene-Transformer model includes a one-dimensional convolutional feature extraction layer, a positional encoding layer, a Transformer encoding layer, a global average pooling layer, and a classification output layer.

3. The method according to claim 1, characterized in that, In step S3, The process of obtaining candidate miRNAs includes: obtaining differentially expressed miRNAs using differential expression analysis, scoring the importance of the differentially expressed miRNAs using a machine learning model, ranking the differentially expressed miRNAs according to the importance scores, and taking the intersection of the machine learning screening results and the differentially expressed miRNAs to obtain candidate miRNAs.

4. The method according to claim 1, characterized in that, In step S4, the construction of the miRNA-gene joint regulatory network includes: Based on the database, the target gene relationships corresponding to candidate miRNAs were obtained. The miRNAs were filtered according to their binding region, minimum free energy, and experimental verification information. The miRNA-gene regulatory relationships with binding regions located in the 3′-UTR, minimum free energy less than -20 kcal / mol, and experimental verification support were retained. Input candidate genes into the STRING database to obtain gene-gene interaction relationships; The miRNA-gene regulatory relationship and the gene-gene interaction relationship are combined to construct the miRNA-gene joint regulatory network.

5. The method according to claim 1, characterized in that, In step S5, the change in network entropy before and after removing the node is determined as the network disturbance contribution value of that node, specifically including: The absolute value of the difference in network entropy before and after a node is removed is taken as the perturbation contribution value of that node; the larger the perturbation contribution value, the greater the impact of the corresponding node on the network structure. The network disturbance contribution value is not directly used as a screening criterion for joint markers, but is used as a graph signal in the subsequent sub-network modules.

6. The method according to claim 1, characterized in that, In step S6, the topological community division includes: The Louvain algorithm was used to maximize the network modularity as the global optimization objective. The miRNA-gene joint regulation network was divided into several initial communities through iterative iteration. The number of interaction regulatory edges in each initial community was counted, and the initial communities with zero interaction edges were removed.

7. The method according to claim 1, characterized in that, In step S7, The graph signal is a one-dimensional discrete graph signal vector. ; Constructing the corresponding graph Laplacian matrix specifically includes: Extract the adjacency matrix of the effective modules and node degree matrix Construct the graph Laplace matrix ; According to the quadratic energy formula Calculate the total variation of the graph for this subnetwork module.

8. The method according to claim 1, characterized in that, Step S7, which involves extracting several effective modules with the highest total variation score from the average plot and using the miRNA-gene regulatory relationships within them as joint biomarkers, includes the following steps: The average total variation of all effective modules was calculated. Modules with an average total variation of more than 3 times the average were defined as high-risk modules, modules with an average total variation of less than 1 / 3 of the average were defined as low-risk modules, and modules in between were defined as medium-risk modules. The miRNA-gene regulatory relationship in the high-risk modules was selected as a joint biomarker for SLE.

9. The method according to claim 3, characterized in that, The differential expression analysis method includes the limma-voom method, and the machine learning model includes the random forest model and the XGBoost model. The importance score of miRNA is obtained according to the random forest model and the XGBoost model respectively. The top 50% of miRNAs in terms of importance score are selected. The union of the screening results of the two machine learning models is then taken, and the intersection is taken with the differentially expressed miRNA.