MiRNA and disease relation prediction system and method based on causal feature selection

Abstract

The invention discloses a miRNA and disease relation prediction system and method based on causal feature selection. According to the system, the accuracy of miRNA and disease relation prediction is improved through a plurality of modules such as data acquisition, negative sample selection, feature selection and causal analysis, model training and evaluation and the like. The data acquisition module is used for acquiring miRNA expression data and disease characteristics from a biomedical database; the negative sample selection module is used for grouping unknown miRNA-disease relationships through k-means clustering and randomly selecting negative samples; the feature selection module adopts a causal feature selection algorithm to construct a Markov blanket so as to screen out miRNA features directly related to disease states, and redundancy is reduced; the model training module uses a single machine learning algorithm for training, and optimizes the model performance through cross validation; and finally, the evaluation module evaluates the model effect through indexes such as AUROC and AUPR. According to the method, the accuracy of miRNA and disease relation prediction is remarkably improved, and important support is provided for medical research.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics and artificial intelligence technology, and in particular to a system and method for predicting the relationship between miRNA and disease based on causal feature selection. Background Art

[0002] MicroRNA (miRNA) is a type of non-coding RNA that plays a key regulatory role in biological processes. Its abnormal expression is closely related to the occurrence and development of a variety of diseases. The prediction of the relationship between miRNA and disease is of great significance for understanding the molecular mechanisms of complex diseases and discovering new therapeutic targets. In existing bioinformatics research, the prediction model of the relationship between miRNA and disease is mainly optimized from three aspects: data noise removal, feature selection, and model optimization. Traditional feature selection algorithms usually rely on encapsulation, filtering, and embedding methods, which look for relevant feature subsets based on the correlation between features and target variables. However, this correlation-based feature selection method can only reveal the coexistence relationship between features and target variables, and it is difficult to explain the decision mechanism behind it.

[0003] The existing technology has the following problems when predicting the relationship between miRNA and disease:

[0004] Lack of causal explanation: Although traditional feature selection methods can find miRNA features related to the target variable (disease), the correlation between these features often cannot reveal how miRNA affects disease progression by regulating gene expression. Causal relationships can better explain the specific mechanism of action of miRNA in disease occurrence.

[0005] To this end, we proposed a system and method for predicting the relationship between miRNA and disease based on causal feature selection. Summary of the invention

[0006] The present invention optimizes feature screening and model training through a causal feature selection algorithm, thereby improving the accuracy and interpretability of the prediction of the relationship between miRNA and disease.

[0007] To achieve the above object, the present invention provides the following technical solution: a miRNA and disease relationship prediction system based on causal feature selection, comprising the following modules:

[0008] The data acquisition module is used to obtain the original data in the miRNA and disease relationship database to ensure that the data dimensions cover disease characteristics, miRNA expression information and other related variables; the negative sample selection module is used to group unknown miRNA-disease relationships through k-means clustering and randomly select a number of negative samples; the feature selection and causal analysis module processes the original data through the causal feature selection algorithm, removes redundant information, and retains the miRNA features that are strongly correlated with the target disease characteristics; the prediction model construction module combines the miRNA feature set after feature selection and uses a single machine learning algorithm to train the prediction model; the evaluation module verifies the performance of the model based on AUROC, AUPR, precision, recall, F1 value and accuracy evaluation indicators; the data feedback module updates and feedbacks the prediction results in real time and performs error analysis and model tuning.

[0009] Preferably, the system comprises the following submodules:

[0010] miRNA expression data collection submodule: The data comes from the MiRBase database, which contains miRNA names, sequences, and related gene annotation information of more than 200 species;

[0011] Disease information collection submodule: The data comes from the Medical Subject Headings (MeSH), a set of subject headings in the biomedical field;

[0012] miRNA and disease association data collection submodule: used to collect the association strength between miRNA and specific diseases and related literature support data;

[0013] Feature selection and causal analysis module: Screen out features related to the relationship between miRNA and disease based on the causal feature selection algorithm. The module includes the following submodules:

[0014] Causal feature screening submodule: The IAMB algorithm is used to gradually screen features that are significantly related to the relationship between miRNA and disease;

[0015] Variable independence test submodule: used to detect the independence of each candidate feature to ensure the accuracy of causal inference;

[0016] A method for predicting the relationship between miRNA and disease based on causal feature selection, the method comprising the following steps:

[0017] The method comprises the following steps:

[0018] S1. Data collection:

[0019] S1.1 Collect miRNA expression data and related disease characteristic data;

[0020] S1.2 perform data preprocessing on the collected data, including denoising and standardization operations;

[0021] S2. Feature selection:

[0022] S2.1 Use causal feature selection algorithm to preliminarily screen the preprocessed data and extract disease-related feature sets;

[0023] S2.2 further processes the feature set based on the independent mutual information optimization algorithm to finally form an optimized feature set;

[0024] S3. Model training:

[0025] S3.1 Use optimized model training algorithms to train miRNA and disease data through feature selection, including random forest, decision tree, lightweight gradient framework and k nearest neighbor model;

[0026] S3.2 Adjust the hyperparameters of each model and evaluate the model performance through cross-validation;

[0027] S4. Result analysis and output:

[0028] S4.1 outputs the final prediction results of the relationship between miRNA and disease based on the training results of a single model;

[0029] S4.2 performs performance evaluation on the results and displays the specific results of multiple evaluation indicators.

[0030] Preferably, the expression of the causal feature selection algorithm is: P(C|MB(C),S)=P(C|MB(C)), where:

[0031] C: target variable, i.e., the state of the disease; MB(C): Markov Blanket of target variable C, i.e., a set of features causally related to C, including parent nodes, child nodes, and parent nodes of child nodes; S: other features in the original feature set; the algorithm calculates the mutual information between each feature and the target variable C through independence test, and gradually adds the most relevant features to the Markov Blanket based on the results; irrelevant features are eliminated in each test, and features that help predict the relationship between miRNA and disease are retained through causal inference, eventually forming the best feature subset MB(C).

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] 1. Improve prediction accuracy: Through the causal feature selection algorithm, redundant features are effectively removed, and miRNA features closely related to the disease are retained, which significantly improves the prediction accuracy of the model.

[0034] 2. Enhance model interpretability: The causal feature selection algorithm can clearly reveal the causal relationship between miRNA and disease, making the prediction results more accurate and facilitating subsequent biological research and clinical applications.

[0035] 3. Optimize the feature selection process: In the feature selection process, the independent mutual information algorithm and the random forest model scoring mechanism are combined to further refine the feature subset and effectively reduce the overfitting phenomenon of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a module structure diagram of the miRNA and disease relationship prediction system based on causal feature selection of the present invention;

[0037] Figure 2 It is a schematic diagram of the submodule structure of the system of the present invention;

[0038] Figure 3 This is a flow chart of the steps of the method for predicting the relationship between miRNA and disease based on causal feature selection of the present invention;

[0039] Figure 4 This is a flow chart of feature selection and model training of the present invention. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] See also Figures 1 to 4 The present invention provides a system and method for predicting the relationship between miRNA and disease based on causal feature selection. The system aims to optimize the prediction of the relationship between miRNA and disease, and improve the accuracy and stability of prediction through negative sample construction, feature selection and optimization of a single machine learning model.

[0042] Data acquisition module:

[0043] This module is the data input part of the whole system, responsible for extracting the relationship data between miRNA and disease from multiple sources. The data includes the following dimensions:

[0044] miRNA expression data: Derived from the MiRBase database, which is the most comprehensive online database of miRNAs today, including miRNA names, sequences, and related gene annotation information of more than 200 species.

[0045] Disease characteristic data: Derived from the Medical Subject Headings (MeSH) index, which is a set of subject headings in the biomedical field specified by the U.S. National Library of Medicine.

[0046] Related data: mainly from biomedical databases, such as HMDD v2.0 (miRNA and disease relationship database). This database records the human miRNA-disease relationship verified by biological experiments.

[0047] The main function of this module is to ensure the multidimensionality and integrity of the data and to ensure that all variables related to disease relationship prediction can be collected, thereby providing a reliable data foundation for subsequent feature selection and model training.

[0048] Negative sample selection module:

[0049] In disease prediction, the balance between positive and negative samples is crucial. This module selects negative samples for unknown miRNA-disease relationships through the following steps:

[0050] K-means clustering algorithm: The system first uses K-means clustering to group the unknown miRNA-disease relationship data. The purpose of this step is to divide the unknown relationships in the data set into several groups according to their feature similarities.

[0051] Randomly select negative samples: Randomly select several unknown miRNA-disease relationships from each cluster group as negative samples. In order to ensure the balance of the data, the number of negative samples is close to the number of positive samples.

[0052] The design of this module aims to enable the system to better distinguish the true relationship between miRNA and disease during model training through reasonable negative sample selection, thereby avoiding model bias caused by data imbalance.

[0053] Feature selection and causal analysis module:

[0054] Feature selection is a key step to improve the accuracy and interpretability of prediction models. The present invention uses a causal feature selection algorithm to handle the complex relationship between miRNA and disease, which mainly includes the following steps:

[0055] Constructing Markov Blanket: Constructing the Markov Blanket of each target variable through a causal feature selection algorithm (such as the IAMB algorithm) to find features directly related to the variable. This step ensures that only miRNA features closely related to the disease are retained, while redundant features are eliminated.

[0056] Independence test: Based on the construction of Markov boundary, the independence test of variables is carried out to ensure that the selected features have independent causal explanations. The independence test helps to eliminate features that are highly correlated but have no actual causal relationship.

[0057] Feature optimization: Mutual Information is used to further optimize the features and select the feature subset that maximizes the amount of information between the target variable (disease status). The resulting feature set is more accurate.

[0058] The main goal of this module is to ensure that the selected features can fully reflect the biological mechanism of the disease and reduce the impact of noisy data, thereby improving the generalization ability of the model.

[0059] Predictive model building blocks:

[0060] This module combines the above-mentioned feature selection results and uses a single machine learning algorithm for training and model building. Specifically, the system uses the following algorithms:

[0061] Random Forest: This algorithm can handle high-dimensional data and improve prediction accuracy and noise resistance through the combination of multiple decision trees.

[0062] Decision Tree: As an easy-to-interpret model, decision tree has strong predictive performance under small data sets.

[0063] Lightweight Gradient Boosting Framework (LightGBM): A gradient boosting decision tree model suitable for processing large-scale data, which can be trained quickly and maintain high accuracy.

[0064] k Nearest Neighbor (KNN): This algorithm predicts the category of unknown samples by measuring the distance between samples and is suitable for fast classification of small-scale data sets.

[0065] This module tunes a single model through cross-validation to ensure the generalization ability of the model on different data sets and improve the accuracy and stability of predictions.

[0066] Assessment Modules:

[0067] This module evaluates the performance of the constructed prediction model through multiple indicators, mainly including:

[0068] AUROC (Area Under the ROC Curve): Used to measure the overall performance of a classification model.

[0069] AUPR (Area Under the PR Curve): Used to evaluate model performance when positive and negative samples are unbalanced.

[0070] Precision, recall, F1 value and Accuracy: These indicators are used to evaluate the performance of the model in real scenarios.

[0071] The evaluation module not only comprehensively analyzes the performance of the model through multiple indicators, but also avoids overfitting through cross-validation of a single model to improve the robustness of the model.

[0072] Data feedback module: The data feedback module is mainly used for dynamic adjustment and optimization during model training and prediction. By real-time monitoring of the errors in the model prediction results, the module can feed back the errors to the model tuning process to adjust and retrain the hyperparameters. Through this module, the system can dynamically optimize the prediction model according to the changes in different data sets, improving the robustness and adaptability of the model.

[0073] Through the setting of submodules, the system can achieve efficient data collection, processing and feature selection, and improve the accuracy of the prediction of the relationship between miRNA and disease, as follows:

[0074] miRNA expression data collection submodule, MiRBase database contains miRNA names, sequences and related gene annotation information of more than 200 species.

[0075] The disease information collection submodule is derived from the Medical Subject Headings (MeSH), a hierarchical terminology used for indexing and classification of biomedical information.

[0076] The miRNA and disease association data collection submodule is responsible for extracting expression data related to miRNA from multiple biological databases. Common data sources include the HMDD v2.0 database, which records a large amount of miRNA expression data in different diseases. These expression data will serve as the basic data for subsequent model training and feature selection to ensure that the system can obtain complete information on the relationship between miRNA and disease.

[0077] Feature selection and causal analysis module,The feature selection and causal analysis module is used to screen out the miRNA features that are most relevant to the disease from a large amount of data, ensuring that the model retains the most valuable features for disease prediction while reducing data redundancy.

[0078] The causal feature screening submodule uses the IAMB algorithm to gradually screen out miRNA features that are significantly related to the disease. The IAMB algorithm is a commonly used causal feature selection algorithm that can screen out miRNA features that are directly related to the target disease from a large number of candidate features. Its working principle is to gradually increase the features associated with the target disease in an incremental manner until the "Markov Blanket" of the disease is constructed, that is, the set of all important features related to the disease state.

[0079] The variable independence test submodule is responsible for detecting the independence of candidate features to ensure the accuracy of causal inference. This step is crucial for eliminating irrelevant redundant features. Specifically, the system uses statistical test methods to detect whether there is independence between miRNA features and disease status. If a feature is independent of the disease state under known conditions, it can be considered that the feature has no causal explanation and is therefore eliminated.

[0080] The present invention proposes a method for predicting the relationship between miRNA and disease based on causal feature selection, aiming to improve the accuracy and reliability of the prediction of the relationship between miRNA and disease through systematic data processing and optimization model. Each step is described in detail below.

[0081] S1. Data collection

[0082] S1.1 Collection of miRNA expression data and related disease characteristic data:

[0083] The system extracts miRNA expression data and related disease characteristics from multiple biomedical databases. The data mainly comes from miRNA and disease relationship databases such as HMDD v2.0, which record the expression levels of a large number of miRNAs in different disease states.

[0084] By integrating multi-source data, the system can ensure that the information obtained is relatively comprehensive and accurate, providing high-quality data input for subsequent feature selection and model training.

[0085] S1.2 Data preprocessing:

[0086] The system performs noise filtering and standardization on the collected data to remove abnormal data points that may appear during the collection process. Standardization operations include converting the data to a unified scale to ensure the consistency of each feature dimension to avoid deviations caused by differences in data scale during model training.

[0087] For example, the system normalizes miRNA expression values ​​so that the miRNA expression data of each sample are on the same scale, thereby improving the accuracy of subsequent feature selection and model training.

[0088] S1.3 Cleaning data:

[0089] During the data cleaning process, the system detects and processes missing values ​​in the data set. Missing value processing uses statistical methods such as mean interpolation and K nearest neighbor interpolation.

[0090] The purpose of this step is to ensure the integrity and consistency of the data, provide a reliable data basis for subsequent feature selection and model training, and reduce the negative impact of data noise on model prediction results.

[0091] S2. Feature Selection

[0092] S2.1 Causal Feature Selection Algorithm:

[0093] The system uses causal feature selection algorithms (such as the IAMB algorithm) to perform feature screening on preprocessed miRNA and disease data. The IAMB algorithm gradually constructs the Markov Blanket of the target disease in an iterative manner and screens out miRNA features that are directly related to the disease state.

[0094] The advantage of this algorithm is that it can exclude redundant features and retain a set of causal features that are closely related to the disease, thereby improving the interpretability and predictive performance of the model.

[0095] S2.2 Independent mutual information optimization:

[0096] Based on the initial screening, the system further refines the feature set using an independent mutual information optimization algorithm, which measures the amount of information between each feature and the target disease state and selects miRNA features that provide the greatest information gain.

[0097] The optimized feature set not only improves the prediction accuracy of the model, but also enhances the interpretability of the features. For example, the system retains those miRNA features that perform well in the information maximization calculation, while eliminating features that are irrelevant to the disease state or have low information content.

[0098] S3. Model training

[0099] S3.1 is trained using a single machine learning algorithm:

[0100] The system uses the miRNA feature set after feature selection and adopts a single machine learning algorithm (such as random forest, decision tree, lightweight gradient framework or K nearest neighbor) to train the model.

[0101] Specifically, random forests can improve prediction performance by combining multiple decision trees when processing high-dimensional data, while the lightweight gradient framework is suitable for efficient training of large-scale data sets. The system selects the most suitable algorithm for training based on the specific characteristics of the data set to ensure the prediction accuracy of the model.

[0102] S3.2 Hyperparameter Tuning and Cross-Validation:

[0103] During the training process, the system adjusts the hyperparameters of a single model based on the characteristics of different data sets.

[0104] The system also uses 5-fold cross-validation to evaluate the model, and measures the generalization ability of the model through cross-validation results on different data sets. This process helps ensure the robustness and stability of the model in different scenarios.

[0105] S4. Result analysis and output

[0106] S4.1 Single model prediction analysis:

[0107] The system outputs the final prediction value of the relationship between miRNA and disease based on the prediction results of a single model. These prediction results are based on the training effect of a single model and conduct a detailed analysis of the relationship between each miRNA and a specific disease.

[0108] For example, the system can output the risk contribution of a certain miRNA in cancer development, thereby helping researchers understand the potential biological role of the miRNA.

[0109] S4.2 Performance evaluation:

[0110] The system comprehensively evaluates the prediction results of a single model through multiple indicators (such as AUROC, AUPR, precision, recall, F1 value and accuracy).

[0111] The evaluation process not only verifies the model's predictive performance, but also reveals its adaptability in different application scenarios. The system will tune the model based on the evaluation results to further improve the model's predictive accuracy and stability.

[0112] Through the design and implementation of the above steps, the present invention provides an efficient and accurate method for predicting the relationship between miRNA and disease. The system combines causal feature selection with the optimization of a single model, and through the collaboration of modules such as data collection, feature selection, model training and performance evaluation, it improves the accuracy and interpretability of the prediction of the relationship between miRNA and disease.

[0113] Working principle of causal feature selection algorithm: The causal feature selection algorithm of the present invention effectively screens out miRNA features directly related to disease status by constructing a Markov Blanket. The core of the algorithm is that if the Markov Blanket of a variable is known, the variable is independent of other variables outside the conditioned Markov Blanket. Therefore, when constructing a disease feature set, only features directly related to the disease need to be found, and other features can be ignored;

[0114] The key expression of the algorithm is: P(C|MB(C),S)=P(C|MB(C)), where: C: target variable, i.e., the state of the disease; MB(C): Markov Blanket of target variable C, i.e., the feature set causally related to C, including parent nodes, child nodes, and parent nodes of child nodes; In this formula, Markov Blanket ensures that all miRNA features directly related to the disease are retained, while those features that have no direct causal relationship with C will be ignored under the formula conditions. S represents other variables in the feature set F that are not related to C. In addition, there is the following formula:

[0115]

[0116] For all cases where S is a subset of the Markov blanket MB(C)MB(C) of the feature set F minus the variables C.

[0117] P(C|MB(C),S): represents the conditional probability of variable C given the Markov blanket MB(C) of variable C and the values ​​of the variables in set S.

[0118] P(C|MB(C)): represents the conditional probability of variable C only when the values ​​of the variables in the Markov blanket MB(C) of variable C are known.

[0119] The equation P(C|MB(C),S)=P(C|MB(C)) states that when the Markov blanket MB(C) of C is known, the value of C is not affected by the values ​​of the variables in S, where S is any subset of the difference between F and MB(C). This means that the variables in S are redundant with respect to C when MB(C) is already known;

[0120] Select the optimal feature subset S* associated with the target variable C. The goal of this problem is to find a feature subset S that maximizes the mutual information I(C; S) between C and S.

[0121] S* represents the optimal feature subset; F is the set of all possible features; I(C; S) is the mutual information between the target variable C and the feature subset S. Mutual information I(C; S) measures the degree to which the uncertainty in predicting the target variable C is reduced when the information of feature set S is known. If feature set S contains very useful information for predicting C, then the value of I(C; S) will be high.

[0122] Some studies have shown that causal relationships have better interpretability and robustness; causal discovery algorithms simplify the reasoning process and only need to consider variables directly related to the target variable (i.e., Markov blanket), while other variables can be ignored.

[0123] It can help improve the performance of the model by selecting the most relevant features to reduce the risk of noise and overfitting. At the same time, it also helps to reduce the complexity of the model and improve computational efficiency;

[0124] Through this causal feature selection algorithm, the system can calculate the mutual information between each feature and the target variable C through step-by-step independence tests, and gradually incorporate the most relevant features into the Markov Blanket set based on the test results. After each test, the system will eliminate irrelevant features and eventually form a highly optimized miRNA feature set that is relevant to the target disease state.

[0125] Independence test: The algorithm detects whether each candidate feature is independent of the disease state C. If a feature is invalid for predicting C, it will be eliminated.

[0126] Construction of Markov Blanket: This algorithm retains all miRNA features that have a direct causal relationship with C from a large number of candidate features. After each calculation, the most relevant features will be retained in the Markov Blanket, and those that are not causal will be eliminated.

[0127] In practical applications, the data processing of the algorithm is completed through the collected data set. The collection module provides miRNA expression data, disease characteristic data, and related literature association data. By analyzing and processing these data, the algorithm can screen out miRNA feature sets that are strongly correlated with disease characteristics.

[0128] Ultimately, the causal feature selection algorithm ensures that only miRNA features that are closely related to the target disease state are retained, while irrelevant features are effectively eliminated, which provides an accurate data basis for subsequent model prediction and evaluation.

[0129] Experimental results:

[0130] In order to ensure the fairness of the experiment, this study will uniformly use the same data set to evaluate and compare these algorithms (the data is divided into 5 parts, the same data set is used for experiments, and the data set for each cross-validation is kept consistent).

[0131]

[0132]

[0133] Table 1

[0134] Note: This experiment improves the feature selection algorithm, where iamb-plus is the method of the present invention.

[0135] Result analysis: AUROC: The AUROC of the feature selection method iamb-plus+ random forest model of the present invention is 0.9323+, which is higher than not performing feature selection, using feature embedding method, and using iamb feature selection method. This proves that the present invention has the best performance in feature selection methods.

[0136] AUPR: When dealing with the problem of imbalance between positive and negative samples, the AUPR of the proposed method reaches 0.9248+, which is better than traditional feature selection.

[0137] Precision and recall: The proposed method achieved precision and recall of 0.8561+ and 0.8659+ respectively, and performed better in balancing prediction accuracy and capturing positive samples, which is significantly improved compared with the performance of traditional feature selection methods.

[0138] F1 value: The F1 value of the present invention is 0.8610+, which has a good balance between precision and recall, indicating that the model is stable when facing different data sets.

[0139] Accuracy: The accuracy of the proposed method is 0.8614+, which is significantly improved compared with the traditional model. This further proves that the causal feature selection algorithm can effectively improve the prediction accuracy of the model.

[0140] To verify the scientific nature of the system and the reliability of the data, the data acquisition module of the present invention integrates a large amount of information related to diseases and miRNAs from multiple databases (such as HMDD v2.0 and PubMed). The following table shows some diseases, miRNAs and their literature support information:

[0141] Table 2: Disease-miRNA relationships and literature support

[0142]

[0143]

[0144] Table 2

[0145] The last column in the table is the PubMed unique identifier, which indicates the document number indexed in the PubMed search engine. From the table, we can see that 20 of the top 200 relationships predicted by causal feature selection have been verified and supported by corresponding literature.

[0146] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A miRNA-disease relationship prediction system based on causal feature selection, characterized by: Includes the following modules: Data collection module, used to obtain raw data from the miRNA and disease relationship database to ensure that the data dimensions cover disease characteristics, miRNA expression information and other related variables; Negative sample selection module, which is used to group unknown miRNA-disease relationships through k-means clustering and randomly select several negative samples; The feature selection and causal analysis module processes the raw data through a causal feature selection algorithm, removes redundant information, and retains miRNA features that are strongly correlated with target disease features; The prediction model building module combines the miRNA feature set after feature selection and uses a single machine learning algorithm to train the prediction model; Evaluation module verifies the performance of the model based on AUROC, AUPR, precision, recall, F1 value and accuracy evaluation indicators; The data feedback module updates and feeds back prediction results in real time and performs error analysis and model tuning.

2. The miRNA-disease relationship prediction system based on causal feature selection according to claim 1, characterized in that: The system includes the following submodules: miRNA expression data collection submodule: used to collect disease-related miRNA expression level data; Disease information collection submodule: used to collect information related to disease classification, characteristics and causes; miRNA and disease association data collection submodule: used to collect the association strength between miRNA and specific diseases and related literature support data; Feature selection and causal analysis module: Screen out features related to the relationship between miRNA and disease based on the causal feature selection algorithm. The module includes the following submodules: Causal feature screening submodule: The IAMB algorithm is used to gradually screen features that are significantly related to the relationship between miRNA and disease; Variable independence test submodule: used to detect the independence of each candidate feature to ensure the accuracy of causal inference.

3. A method for predicting the relationship between miRNA and disease based on causal feature selection, applied to a system for predicting the relationship between miRNA and disease based on causal feature selection according to any one of claims 1 to 2, characterized in that: The method comprises the following steps: S1. Data collection: S1.1 Collect miRNA expression data and related disease characteristic data; S1.2 perform data preprocessing on the collected data, including denoising and standardization operations; S2. Feature selection: S2.1 Use causal feature selection algorithm to preliminarily screen the preprocessed data and extract disease-related feature sets; S2.2 further processes the feature set based on the independent mutual information optimization algorithm to finally form an optimized feature set; S3. Model training: S3.1 Use optimized model training algorithms to train miRNA and disease data through feature selection, including random forest, decision tree, lightweight gradient framework and k nearest neighbor model; S3.2 Adjust the hyperparameters of each model and evaluate the model performance through cross-validation; S4. Result analysis and output: S4.1 outputs the final prediction results of the relationship between miRNA and disease based on the training results of a single model; S4.2 performs performance evaluation on the results and displays the specific results of multiple evaluation indicators.

4. The method for predicting the relationship between miRNA and disease based on causal feature selection according to claim 3, characterized in that: The expression of the causal feature selection algorithm is: P(C|MB(C),S)=P(C|MB(C)), where C: The target variable is the state of the disease; MB(C): the Markov Blanket of the target variable C, that is, the set of features causally related to C, including parent nodes, child nodes, and parent nodes of child nodes; S: other features in the original feature set; The algorithm calculates the mutual information between each feature and the target variable C through independence test, and gradually adds the most relevant features to the Markov Blanket based on the results; In each test, irrelevant features were eliminated, and features that were helpful in predicting the relationship between miRNA and disease were retained through causal inference, ultimately forming the optimal feature subset MB(C).

Images