Crop ralstonia solanacearum disease prediction method based on ensemble learning model

Abstract

The invention discloses a crop ralstonia solanacearum disease prediction method based on an integrated learning model, and the method comprises the following steps: S1, collecting a published 16s rRNA gene sequence related to solanaceae crop bacterial wilt, and carrying out the preprocessing of original sequencing data based on an EasyAmplicon standardized process; s2, performing data dimension reduction by using a principal component analysis algorithm, and retaining 95% of variance; s3, performing hyper-parameter search based on 5-fold cross validation and grid search on the Light GBM model, the CatBoost model and the XGBoost model respectively, and selecting three groups of optimal hyper-parameters of each model; s4, constructing a model according to three groups of optimal hyper-parameters of each selected model, performing prediction, analyzing model result difference based on a Pearson correlation coefficient, and retaining Pearson correlation coefficient mean < lt > with other eight model prediction values; a model of 0.8; and S5, inputting the screened model prediction result into a second-layer element learner RF for integrated learning to obtain a final prediction result.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of crop disease prediction, and particularly relates to a crop bacterial wilt disease prediction method based on an integrated learning model. BACKGROUND

[0002] As a global soil-borne pathogen, bacterial wilt can cause diseases of economic crops such as tomatoes and potatoes. The long incubation period and fast transmission speed of bacterial wilt result in the lag of traditional prevention and control measures behind the outbreak of the disease. Therefore, early and accurate prediction is a key link to break the disease transmission chain and reduce the abuse of chemical pesticides.

[0003] Existing bacterial wilt disease prediction methods mainly include molecular biology detection and data-driven prediction. The molecular biology detection method generally relies on direct detection of the pathogen, and the presence of the target microorganism is identified by specific primers or antibodies. Such a method has significant limitations in complex soil environments: the detection process usually requires laboratory-level equipment and involves a tedious sample pretreatment process, making it difficult to achieve real-time monitoring in the field. At the same time, existing detection techniques lack the ability to characterize the functional state of microbial communities, and cannot effectively reflect the dynamic balance changes of soil microecosystems, resulting in a lag in disease early warning. Data-driven prediction methods mainly construct prediction models by analyzing microbiome data, and have greater flexibility and timeliness. However, existing technical solutions mostly use a single model for modeling analysis, which has insufficient model generalization ability and prediction accuracy. Some methods use model integration strategies, but generally use fixed weight fusion or simple voting mechanisms, which fail to effectively select base models with different prediction capabilities, resulting in limited performance improvement of the integrated model. In summary, current bacterial wilt disease prediction technology has the defects of lagging detection timeliness, insufficient mining of microbial functional characteristics, and inefficient model integration strategies, making it difficult to meet the agricultural practice needs of real-time and accurate early warning in the field. SUMMARY

[0004] To solve the above problems, the application provides a crop bacterial wilt disease prediction method based on an integrated learning model. The method realizes early and accurate prediction of bacterial wilt of solanaceous crops through multi-model collaborative optimization and multi-level data integration, and provides efficient and reliable technical support for intelligent monitoring of agricultural diseases.

[0005] To achieve the above purpose, the application adopts the following technical solutions:

[0006] A crop bacterial wilt disease prediction method based on an integrated learning model, comprising the following steps:

[0007] S1, collect published 16s rRNA gene sequences related to bacterial wilt of solanaceous crops, and pretreat the original sequencing data based on the standardization process of EasyAmplicon;

[0008] S2, data dimension reduction is performed by using a principal component analysis algorithm, and 95% variance is retained;

[0009] S3, hyperparameter search is performed on the LightGBM model, the CatBoost model and the XGBoost model respectively based on 5-fold cross-validation and grid search, and 3 groups of optimal hyperparameters of each model are selected;

[0010] S4, a model is constructed according to the 3 groups of optimal hyperparameters of each model selected, and the difference of the model results is analyzed based on the Pearson correlation coefficient, and the model whose Pearson correlation coefficient with the prediction value of the other 8 models is less than 0.8 is retained;

[0011] The specific process of step S4 is as follows:

[0012] S41, 3 groups of optimal hyperparameters of the LightGBM model, the CatBoost model and the XGBoost model are respectively constructed into independent models, 5-fold cross-validation is used to generate out-of-sample prediction values, and a prediction result matrix is obtained:

[0013]

[0014] Wherein, P is the prediction result matrix; n is the sample number; is a real number set;

[0015] S42, the Pearson correlation coefficient matrix of the prediction results of the models is calculated:

[0016]

[0017] Wherein, R is the Pearson correlation coefficient matrix; R ij is the Pearson correlation coefficient of the ith model and the jth model; P ki is the prediction result of the kth sample of the ith model; is the average value of the prediction result of the ith model; P kj is the prediction result of the kth sample of the jth model; is the average value of the prediction result of the jth model;

[0018] S43, model screening is performed, if the absolute value of the Pearson correlation coefficient of a certain model and all the remaining models is greater than or equal to 0.8, it is judged as a redundant model and is removed;

[0019] S5, the prediction results of the screened models are input into the second layer meta-learner RF for ensemble learning, and the final prediction results are obtained.

[0020] Preferably, the specific process of step S1 is as follows:

[0021] S11, collect published 16S rRNA gene sequences related to bacterial wilt of Solanaceae crops, and identify the physiological state of each sample according to the information in the published article;

[0022] S12, quality control, denoising and feature table generation of raw sequencing data based on EasyAmplicon standardization process, and taxonomic annotation by 16s rRNA gene full-length database;

[0023] S13, predict microbial function by PICRUSt2 tool and FAPROTAX tool;

[0024] S14, integrate the results of step S11, step S12 and step S13 to obtain ZOTU abundance, metabolic pathways, element cycles and alpha diversity indicators of different plant physiological states.

[0025] Preferably, the specific process of step S2 is:

[0026] S21, Z-score standardization is performed on the ZOTU abundance, metabolic pathways, element cycles and alpha diversity indicators obtained after pretreatment to eliminate dimensional differences;

[0027] S22, construct a covariance matrix to represent the linear correlation between each feature, and perform orthogonal decomposition on the covariance matrix to obtain the eigenvalues λ1≥λ2≥...≥λ p and their corresponding eigenvectors in descending order;

[0028] S23, determine the principal component retention number k, and the calculation formula is:

[0029]

[0030] Wherein, k is the principal component retention number; k * is the optimal solution of k; Z + is a set of positive integers; i and j are both eigenvalue serial numbers; λ i is the i-th eigenvalue; λ j is the j-th eigenvalue; P is the number of features;

[0031] S24, project the original sequencing data to the subspace composed of the first k eigenvectors to obtain the reduced data set.

[0032] Preferably, the specific process of step S3 is:

[0033] S31, construct a search space of three hyperparameters of decision tree number, learning rate and maximum tree depth;

[0034] S32, independently performing grid search on each base learner, adopting 5-fold cross-validation strategy and reserving 20% samples as a validation set through stratified sampling, adopting F1 score as a validation index, and the calculation formula is:

[0035]

[0036] Wherein, TP is the number of correctly predicted disease samples; FP is the number of incorrectly predicted disease samples; FN is the number of incorrectly predicted healthy samples;

[0037] S33, for the grid search results of the three base models of LightGBM model, CatBoost model and XGBoost model, arrange the parameter combinations in descending order of F1 score of the validation set, and select the top three parameter combinations of each model.

[0038] Preferably, the specific process of step S5 is:

[0039] S51, input the prediction results of the screened model into the random forest meta-learner as meta-features, and divide them into training set and test set in the ratio of 80:20; wherein, the training set is used for model training, and the test set is used for verification;

[0040] S52, define a multi-parameter search space containing the number of trees, maximum depth and minimum sample number of leaf nodes for the random forest model, and select the optimal hyperparameter combination by grid search combined with 5-fold cross-validation with F1 score as the evaluation index;

[0041] S53, train the random forest meta-learner on the complete training set using the optimal hyperparameters, and then input the test set into the model to obtain the final disease prediction results.

[0042] After adopting the above technical scheme, the present application has the following beneficial effects:

[0043] 1. The present application improves the prediction accuracy and robustness: by integrating three high-performance gradient boosting tree models of LightGBM, CatBoost and XGBoost, and combining random forest (RF) as a meta-learner, the advantages of different base learners are effectively integrated. By selecting redundant models (threshold <0.8) through Pearson correlation coefficient, the diversity of the integrated model is ensured, the risk of overfitting is significantly reduced, and the generalization ability and prediction stability of the model in complex microbiome data are improved.

[0044] 2、The data dimension reduction and feature optimization of the application: adopt principal component analysis (PCA) to retain 95% variance data dimension reduction strategy, while reducing the computational complexity, the effective information of the original data is retained to the greatest extent. Combined with Z-score standardization and orthogonal decomposition of covariance matrix, the dimensional difference of multi-source features (such as ZOTU abundance, metabolic pathway, etc.) is eliminated, and the processing ability of the model to high-dimensional sparse data is enhanced.

[0045] 3、The automatic hyperparameter optimization of the application: through the combination strategy of grid search and 5-fold cross-validation, the optimal 3 groups of hyperparameters (such as the number of decision trees, learning rate, etc.) are selected for each base learner, and the F1 score is used as the verification index to accurately balance the accuracy and recall rate of disease prediction, avoiding the subjectivity of manual parameter adjustment.

[0046] 4、The function and bioinformatics integration of the application: based on EasyAmplicon standardization process and PICRUSt2 / FAPROTAX tools, 16S rRNA gene sequence data and microbial function prediction (metabolic pathway, element cycle, etc.) are deeply related, and multi-dimensional feature input is constructed, so that the model can capture the potential biological correlation between the occurrence of bacterial wilt and the functional changes of microbiome.

[0047] 5、The computing efficiency and scalability of the application: through stratified sampling to divide the training set / verification set (80%:20%) and parallel grid search, the hyperparameter optimization time is significantly shortened. Modular design (such as step-by-step screening base model + meta-learner integration) facilitates extension to other crop disease prediction scenarios, and only needs to adjust the input data to quickly migrate and apply.

[0048] 6、The biomarker discovery potential of the application: the feature importance ranking of the model output can be used to reverse screen ZOTUs or functional pathways that are significantly related to bacterial wilt, providing candidate biomarkers for subsequent experimental verification and assisting in disease mechanism research. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 The flowchart of the application. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical scheme and advantages of the application clearer, the application is further described in detail below with examples. It should be understood that the specific examples described herein are only used to explain the application and not to limit the application.

[0051] As shown in Figure 1 A crop bacterial wilt disease prediction method based on an integrated learning model includes the following steps:

[0052] S1, collect published 16s rRNA gene sequences related to bacterial wilt of solanaceous crops, and preprocess the original sequencing data based on the EasyAmplicon standardization process;

[0053] The specific process of step S1 is as follows:

[0054] S11, collect published 16S rRNA gene sequences related to bacterial wilt of solanaceous crops, and identify the physiological state of each sample corresponding to the plant according to the information of the published article;

[0055] S12, based on the EasyAmplicon standardization process, quality control, denoising and generating feature table of the original sequencing data, and taxonomic annotation through 16s rRNA gene full-length database;

[0056] S13, predict microbial function through PICRUSt2 tool and FAPROTAX tool;

[0057] S14, integrate the results of step S11, step S12 and step S13 to obtain ZOTU abundance, metabolic pathways, element cycles and alpha diversity indicators of different plant physiological states;

[0058] S2, use principal component analysis algorithm for data dimension reduction, and retain 95% variance;

[0059] The specific process of step S2 is as follows:

[0060] S21, Z-score standardization is performed on the ZOTU abundance, metabolic pathways, element cycles and alpha diversity indicators obtained after preprocessing to eliminate dimensional differences;

[0061] S22, construct a covariance matrix to represent the linear correlation between each feature, and perform orthogonal decomposition on the covariance matrix to obtain the eigenvalues λ1≥λ2≥...≥λ p and their corresponding eigenvectors in descending order;

[0062] S23, determine the principal component retention number k, and the calculation formula is as follows:

[0063]

[0064] Wherein, k is the principal component retention number; k * is the optimal solution of k; Z + is a set of positive integers; i and j are both eigenvalue serial numbers; λ i is the ith eigenvalue; λ j is the jth eigenvalue; P is the number of characteristics;

[0065] S24, project the original sequencing data to a subspace composed of the first k eigenvectors to obtain a reduced dimension data set;

[0066] S3, perform hyperparameter search based on 5-fold cross-validation and grid search for the LightGBM model, the CatBoost model and the XGBoost model respectively, and select the best 3 groups of hyperparameters for each model;

[0067] The specific process of step S3 is as follows:

[0068] S31, construct a search space of 3 hyperparameters of decision tree number, learning rate and maximum tree depth;

[0069] S32, independently perform grid search for each base learner, adopt 5-fold cross-validation strategy and reserve 20% samples as a validation set through stratified sampling, adopt F1 score as the validation index, and the calculation formula is as follows:

[0070]

[0071] Wherein, TP is the number of correctly predicted disease samples; FP is the number of incorrectly predicted disease samples; and FN is the number of incorrectly predicted healthy samples;

[0072] S33, for the grid search results of the LightGBM model, the CatBoost model and the XGBoost model, arrange the parameter combinations in descending order of F1 score of the validation set, and select the top 3 parameter combinations for each model;

[0073] S4, construct a model according to the selected best 3 groups of hyperparameters for each model and perform prediction, analyze the difference of the model results based on Pearson correlation coefficient, and retain the model whose Pearson correlation coefficient with the prediction values of other 8 models is less than 0.8;

[0074] The specific process of step S4 is as follows:

[0075] S41, construct independent models for the 3 groups of best hyperparameters of the LightGBM model, the CatBoost model and the XGBoost model respectively, generate out-of-sample prediction values by adopting 5-fold cross-validation, and obtain a prediction result matrix:

[0076]

[0077] Wherein, P is the prediction result matrix; and n is the number of samples; is a real number set;

[0078] S42, calculate the Pearson correlation coefficient matrix of the prediction results among the models:

[0079]

[0080] wherein, R is a Pearson correlation coefficient matrix; R ij is the Pearson correlation coefficient of the i-th model and the j-th model; P ki is the prediction result of the i-th model on the k-th sample; is the average value of the prediction result of the i-th model; P kj is the prediction result of the j-th model on the k-th sample; is the average value of the prediction result of the j-th model;

[0081] S43, model screening is performed, if the absolute value of the Pearson correlation coefficient of a certain model and all the remaining models is greater than or equal to 0.8, it is judged as a redundant model and is removed;

[0082] S5, the prediction result of the screened model is input into the second layer meta-learner RF for integrated learning to obtain the final prediction result;

[0083] The specific process of step S5 is as follows:

[0084] S51, the prediction result of the screened model is input into the random forest meta-learner as meta-features, and is divided into a training set and a test set in a 80:20 ratio; wherein, the training set is used for model training, and the test set is used for verification;

[0085] S52, a multi-parameter search space containing the number of trees, the maximum depth and the minimum sample number of leaf nodes is defined for the random forest model, F1 score is used as an evaluation index, and the optimal hyperparameter combination is selected through grid search combined with 5-fold cross-validation;

[0086] S53, the random forest meta-learner is trained on the complete training set using the optimal hyperparameters, and then the test set is input into the model to obtain the final disease prediction result.

[0087] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited to this, any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for predicting a crop bacterial wilt disease based on an ensemble learning model, the method comprising: The method comprises the following steps: S1, collecting published 16s rRNA gene sequences related to bacterial wilt of solanaceous crops, and preprocessing the original sequencing data based on the EasyAmplicon standardization process; S2, using principal component analysis algorithm for data dimension reduction, retaining 95% variance; S3, performing hyperparameter search based on 5-fold cross-validation and grid search for LightGBM model, CatBoost model and XGBoost model respectively, and selecting three groups of optimal hyperparameters for each model; S4, constructing a model according to the three groups of optimal hyperparameters selected for each model and performing prediction, analyzing the difference of the model results based on Pearson correlation coefficient, and retaining the model whose Pearson correlation coefficient with the prediction value of other eight models is less than 0.8; The specific process of step S4 is: S41, constructing an independent model for each of the three groups of optimal hyperparameters of LightGBM model, CatBoost model and XGBoost model respectively, generating out-of-sample prediction values by 5-fold cross-validation to obtain a prediction result matrix: where P is a matrix of prediction results; n is the number of samples; is a set of real numbers; S42, calculating the Pearson correlation coefficient matrix of the prediction results between models: wherein R is a Pearson correlation coefficient matrix; R ij is a Pearson correlation coefficient of the i-th model and the j-th model; P ki is a prediction result of the i-th model on the k-th sample; is an average value of the prediction result of the i-th model; P kj is a prediction result of the j-th model on the k-th sample; is an average value of the prediction result of the j-th model; S43, model screening, if the absolute value of the Pearson correlation coefficient between a model and all other models is greater than or equal to 0.8, the model is judged as a redundant model and is removed; S5, inputting the prediction results of the screened models into the second layer meta-learner RF for ensemble learning to obtain the final prediction results.

2. The method for predicting crop bacterial wilt disease based on an ensemble learning model according to claim 1, wherein, The specific process of step S1 is: S11, collecting published 16S rRNA gene sequences related to bacterial wilt of solanaceous crops, and identifying the physiological state of each sample according to the information of the published articles; S12, based on the EasyAmplicon standardization process, the original sequencing data is subjected to quality control, denoising and feature table generation, and taxonomic annotation is performed through the 16s rRNA gene full-length database; S13, predicting microbial functions through PICRUSt2 tool and FAPROTAX tool; S14, integrating the results of steps S11, S12 and S13 to obtain ZOTU abundance, metabolic pathways, element cycles and alpha diversity indicators of different plant physiological states.

3. The method of claim 2, wherein the method is characterized by: The specific process of step S2 is: S21, ZOTU abundance, metabolic pathways, element cycles and alpha diversity indicators obtained after preprocessing are subjected to Z-score standardization processing respectively to eliminate dimensional differences; S22, construct a covariance matrix to represent linear correlation between each feature, orthogonally decompose the covariance matrix to obtain eigenvalues λ1≥λ2≥...≥λ p and corresponding eigenvectors of p features arranged in descending order S23, determining the principal component retention number k, the calculation formula is: Wherein, k is the principal component retention number; k * is the optimal solution of k; Z + is a set of positive integers; i and j are both characteristic value serial numbers; λ i is the i-th characteristic value; λ j is the j-th characteristic value; P is the characteristic quantity; S24, projecting the original sequencing data to the subspace composed of the first k characteristic vectors to obtain the dimension-reduced data set.

4. The method of claim 1, wherein the method is characterized by: The specific process of step S3 is: S31, constructing a search space of three hyperparameters of decision tree number, learning rate and maximum tree depth; S32, performing grid search on each base learner independently, using 5-fold cross-validation strategy and retaining 20% samples as validation set through stratified sampling, and using F1 score as the validation index, the calculation formula is: Wherein, TP is the number of correctly predicted disease samples; FP is the number of incorrectly predicted disease samples; FN is the number of incorrectly predicted healthy samples; S33, for the grid search results of the three base models of LightGBM model, CatBoost model and XGBoost model, arrange the parameter combinations in descending order of validation set F1 score, and select the top three parameter combinations of each model.

5. The method of claim 1, wherein the method is characterized by: The specific process of step S5 is: S51, input the prediction results of the screened model into the random forest meta-learner as meta-features, and divide them into training set and test set in 80:20 proportion; wherein, the training set is used for model training, and the test set is used for verification; S52, define a multi-parameter search space containing the number of trees, maximum depth and minimum sample number of leaf nodes for the random forest model, and select the optimal hyperparameter combination by grid search combined with 5-fold cross-validation with F1 score as the evaluation index; S53, train the random forest meta-learner on the complete training set using the optimal hyperparameters, and then input the test set into the model to obtain the final disease prediction results.

Images