High-precision biological breeding intelligent fusion model method

Abstract

The invention discloses a high-precision biological breeding intelligent fusion model method. The method comprises the technical schemes of data cleaning, principal component analysis dimensionality reduction, weighted fusion of multiple machine learning algorithms and Bayesian optimization of hyper-parameters. According to the method, genotype data dimensions are reduced by combining a principal component analysis dimension reduction technology, a single model is constructed by utilizing four machine learning regression devices (XGBoost, HGBoost, LightGBM and Extra Tres), model fusion is carried out by adopting a correlation coefficient weighting method, and finally, model performance is improved by using a Bayesian optimization technology. According to the method, the prediction accuracy, the generalization ability and the calculation efficiency are remarkably improved, and meanwhile, the resource consumption and the time cost are reduced. The method can be widely applied to breeding requirements of plants such as rice, corn and wheat and animals such as pigs, cattle and sheep, is compatible with various variation types, and provides an efficient breeding platform for breeders.

Description

Technical Field

[0001] This invention relates to the field of genome breeding technology, and in particular to a high-precision intelligent fusion model method for biological breeding. Background Technology

[0002] In modern agriculture and bio-breeding, genomic selection technology, as an emerging breeding method, is gradually replacing traditional empirical breeding and molecular marker-assisted selection breeding. Traditional breeding methods have undergone a transformation from empirical breeding to theoretical guidance, including selection theory, pure line theory, backcross breeding, recurrent breeding, and mutation breeding. With the development of molecular biology techniques, molecular marker-assisted selection breeding has gradually emerged, such as amplified fragment length polymorphism (ALP) marker-assisted selection, microsatellite marker-assisted selection, and single nucleotide polymorphism (SNP) marker-assisted selection. However, these methods suffer from low efficiency, high cost, and high technical difficulty when dealing with complex traits. In recent years, with the popularization of high-throughput sequencing technology and the improvement of computing power, genomic selection technology has developed rapidly. Genomic selection, by utilizing high-density molecular genetic markers covering the entire genome for marker-assisted selection, can effectively solve difficult-to-measure traits, shorten breeding cycles, and improve breeding efficiency.

[0003] With the emergence of the concept of "smart breeding," breeding technology is developing towards data-driven and intelligent directions. Smart breeding integrates multi-omics data to establish a platform for crop breeding data extraction, mining, storage, analysis, and sharing. Combined with advanced machine learning and deep learning algorithms, it achieves efficient and accurate breeding decisions. Existing genomic selection models mainly rely on single algorithms, such as the traditional GBLUP method based on linear models, the LightGBM algorithm based on machine learning, the DNNGP method based on convolutional neural networks, and the DeepCCR method based on long short-term memory networks. Although these methods have improved the accuracy of breeding predictions to some extent, they still have many shortcomings. First, single models exhibit limitations when handling genomic data with different characteristics. For example, linear models have limited ability to handle nonlinear relationships, while nonlinear models lack generalization ability on high-dimensional data. Second, in practical applications, selecting the optimal model is a challenge because different models perform significantly differently on different datasets. Furthermore, existing models are typically trained using default parameters, lacking optimization of hyperparameters, which easily leads to overfitting or underfitting. Finally, when the amount of data is very large, the algorithm runs inefficiently or even cannot complete the calculation. Therefore, it is necessary to reduce the data dimensionality through dimensionality reduction methods such as principal component analysis.

[0004] To address the aforementioned problems, this invention aims to develop a high-precision intelligent fusion model for biological breeding. By integrating the advantages of multiple algorithms, a genome selection model with superior overall performance is constructed. This model not only overcomes the limitations of single models but also uses Bayesian optimization algorithms to find optimal hyperparameters, thereby improving the model's accuracy, applicability, and generalization ability. Simultaneously, this invention utilizes principal component analysis (PCA) dimensionality reduction technology to significantly reduce data dimensionality, decrease computational resource consumption and time costs, and provide breeders with an efficient and convenient automated breeding tool. The implementation of this technical solution will help improve the accuracy of offspring selection, shorten the breeding cycle, reduce breeding costs, and promote the further development of intelligent breeding technology. Summary of the Invention

[0005] The purpose of this invention is to provide a high-precision intelligent fusion model method for biological breeding, which solves the problems mentioned in the background art.

[0006] This invention is implemented as follows: a high-precision intelligent fusion model method for biological breeding, comprising the following steps:

[0007] S1. In the data preparation stage, the phenotypic and genotypic data of the training group are cleaned and individuals with missing rates exceeding the preset threshold are removed. Missing genotypic data in the remaining samples are filled in. After merging the genotypic data of the training group and the breeding group, the principal component analysis dimensionality reduction module is used to select the principal components with a cumulative contribution rate of 95% to reduce the data dimensionality.

[0008] S2. In the model building stage, four machine learning regressors are used to train the dimensionality-reduced data respectively, and the optimal hyperparameters are found through the Bayesian optimization module. The four models are then fused together using the correlation coefficient weighted fusion module to obtain the final fused model.

[0009] S3. In the model evaluation and optimization stage, the model performance is evaluated using four indicators: Pearson correlation coefficient, mean square error, root mean square error, and mean absolute error.

[0010] S4. In the prediction and application stage, the genotype data of the breeding population is subjected to PCA dimensionality reduction and transformed into a format recognizable by machine learning algorithms. The constructed fusion model is then used to predict the genotype data of the breeding population and calculate the breeding value.

[0011] Furthermore, in S1, individuals with a phenotypic or genotypic data missing rate exceeding a preset threshold are removed from the sample, and the missing genotypic data in the remaining sample are filled in.

[0012] Furthermore, in S1, eigenvalues ​​and eigenvectors are extracted by calculating the covariance matrix, and principal components with a cumulative contribution rate of 95% are selected as input data after dimensionality reduction.

[0013] Furthermore, the four machine learning regressors used in S2 include XGBoost, HGBoost, LightGBM, and Extra Trees.

[0014] Furthermore, in S3, the objective function is probabilistically modeled, and the posterior distribution is dynamically updated using prior distribution and observation data to efficiently search for the global optimal solution in the hyperparameter space.

[0015] Furthermore, in S4, the Pearson correlation coefficient, mean square error, root mean square error, and mean absolute error of each fold are calculated through ten-fold cross-validation, and their average value is taken as the final evaluation result of the model.

[0016] The beneficial effects of this invention are as follows: By introducing principal component analysis (PCA) dimensionality reduction technology, the dimensionality of genotype data is significantly reduced, thereby reducing the consumption of computational resources and training time. By fusing the prediction results of multiple machine learning algorithms, the complementarity of different models in processing data is fully utilized, enabling the model to better capture complex patterns and diversity. In particular, the correlation coefficient weighted fusion method, through the calculation of exponentially decaying weights, effectively reduces the prediction error of a single model caused by abnormal data or special cases, enhancing the robustness and generalization ability of the model. Furthermore, the application of Bayesian optimization technology significantly improves the efficiency of hyperparameter search, avoiding the inefficiency of traditional grid search or random search methods, thereby improving the overall performance of the model.

[0017] In summary, this invention achieves high-precision construction of a genome selection model by combining principal component analysis for dimensionality reduction, weighted fusion of multiple machine learning algorithms, and Bayesian optimization techniques. It solves several defects in existing technologies and has significant theoretical and practical application value. Attached Figure Description

[0018] Figure 1 This is a roadmap for the high-precision intelligent fusion model for biological breeding of the present invention, showing the complete process from data preparation to model construction, evaluation and predictive application. Detailed Implementation

[0019] This invention provides a high-precision intelligent fusion model method for biological breeding, which combines principal component analysis dimensionality reduction technology, a weighted fusion strategy of multiple machine learning algorithms, and Bayesian optimization technology to construct and apply a genome selection model. The following is in conjunction with the appendix... Figure 1 The specific embodiments of the present invention will be described in detail below.

[0020] In the specific implementation process, the first step is to prepare data for the training and breeding populations. The training population data includes both genotype and phenotypic data, while the breeding population data only contains genotype data. After processing by the data cleaning module, missing and outlier values ​​are removed to ensure high-quality data for subsequent analyses. During data cleaning, individuals with a phenotypic or genotype data missing rate exceeding a preset threshold are removed, and missing genotype data in the remaining samples are imputed. After data cleaning, the genotype data from the training and breeding populations are merged and dimensionality reduction is performed using the principal component analysis (PCA) dimensionality reduction module. PCA extracts eigenvalues ​​and eigenvectors from the covariance matrix, selecting principal components with a cumulative contribution rate of 95% as the input data after dimensionality reduction. This process significantly reduces the dimensionality of the genotype data, decreasing computational resource consumption and training time.

[0021] Next, we move on to the model building phase. In the machine learning algorithm training module, four machine learning regressors are used to train the dimensionality-reduced data. Each algorithm uses 10-fold cross-validation to evaluate model performance and finds the optimal hyperparameters through the Bayesian optimization module. Bayesian optimization efficiently searches for the global optimum in the hyperparameter space by probabilistically modeling the objective function and dynamically updating the posterior distribution using prior distribution and observed data. After obtaining the respective models, the four models are fused through the correlation coefficient weighted fusion module. The specific implementation is as follows: 1) Determine the model prediction results and obtain the prediction results yi of n models; 2) Calculate the weights by calculating the exponential decay weight wi based on the correlation distance di of each model; 3) Weighted fusion by using the calculated weights to perform a weighted average of the model prediction results to obtain the final fused result y. The model architecture is as follows:

[0022] Suppose there are n independent machine learning models, whose predictions are y1, y2, ..., yi, ..., yn. These models are then weighted and fused using weights w1, w2, ..., wi, ..., wn to obtain the final prediction y.

[0023]

[0024] The weight (wi) is calculated using the exponential decay function as follows:

[0025]

[0026] λ is the decay rate parameter, which controls the rate at which the weights decay. A larger value of λ will cause the weights to decay faster. di is the distance or time associated with the i-th model, reflecting the relevance of the model's prediction to the current task.

[0027] After model construction is complete, the model evaluation and optimization phase begins. In the model evaluation module, model performance is assessed using four metrics: Pearson correlation coefficient, mean squared error, root mean square error, and mean absolute error (MSE). The Pearson correlation coefficient measures the linear correlation between predicted and actual values; the mean squared error and root mean square error quantify the magnitude of the deviation between predicted and actual values; and the MSE reflects the average absolute value of the prediction error. The model evaluation process uses ten-fold cross-validation to calculate each of these metrics for each fold, and the average value is taken as the final evaluation result. The optimal model is selected based on the evaluation metrics to ensure stable and reliable model performance across different datasets.

[0028] Finally, the prediction and application phase begins. In the breeding value prediction module, the genotype data of the breeding population undergoes PCA dimensionality reduction and is converted into a format recognizable by machine learning algorithms. A pre-constructed fusion model is used to predict the genotype data of the breeding population and calculate its breeding value. The prediction results are based on the weighted prediction results of the fusion model, ensuring the stability and reliability of the prediction results. Based on the prediction results, individuals with high breeding values ​​are selected for preservation or breeding through the high-breeding-value individual selection module. Generally, individuals in the top 5% or top 1% of breeding values ​​are selected to meet different breeding needs.

[0029] To verify the feasibility of this invention, various public datasets were used for testing in the experiment.

[0030] The presentation uses publicly available data such as 3K rice data, wheat data, cotton data, corn data, and lotus data.

[0031] Case 1: We used 2110 samples from 3k rice datasets, representing 13 phenotypes and SNPs, to build a GS model using our developed software. We evaluated model accuracy using Pearson correlation coefficients (PCCs), and performed 10-fold cross-validation for each model, averaging the results. In the graph, the horizontal axis represents the 13 traits, the vertical axis represents PCC values, and the bars inside represent the PCC values ​​(0-1) for all models. From left to right, the gray-red bars represent the traditional GBLUP model, the green and purple bars represent the deep learning DNNGP and DeepCCR models, the light orange and cyan bars represent single machine learning models, and the blue bars represent the PFMGP model. The graph shows that our PFMGP model has the highest PCC values ​​for all 13 traits, exceeding those of the traditional model, the single machine learning model, and the deep learning model.

[0032]

[0033] Case 2-4: We used phenotypic and SNP data from nearly 200 lotus samples, 2000 wheat samples, and 1245 cotton samples to build a GS model using our developed software. We used Pearson correlation coefficients (PCCs) to evaluate model accuracy, and performed 10-fold cross-validation for each model to calculate the average. In the graph, the horizontal axis represents the trait, the vertical axis represents the PCC value, and the inner bars represent the PCC values ​​of all models (0-0.8). From left to right, the gray-red bars represent the traditional GBLUP model, the green and purple bars represent the deep learning DNNGP and DeepCCR models, the light orange and cyan bars represent single machine learning models, and the blue bars represent the PFMGP model. The graph shows that our PFMGP model has the highest PCC value for these traits, exceeding that of the traditional model, the single machine learning model, and the deep learning model.

[0034]

[0035] Case 5: We used phenotypic and SNP data from 6210 corn samples to build a GS model using our developed software. We evaluated model accuracy using Pearson correlation coefficients (PCCs). For each model, we performed 10-fold cross-validation and then repeated the test 10 times, averaging the results. In the graph, the horizontal axis represents the six traits, the vertical axis represents the PCC values, and the bars inside represent the PCC values ​​(0-1) for all models. From left to right, the gray-red bars represent the traditional GBLUP model, the green and purple bars represent the deep learning DNNGP and DeepCCR models, the light orange and cyan bars represent single machine learning models, and the blue bars represent the PFMGP model. The graph shows that our PFMGP model has the highest PCC values ​​for these traits, exceeding those of the traditional model, the single machine learning model, and the deep learning model.

[0036]

[0037] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A high-precision intelligent fusion model method for biological breeding, characterized in that, Includes the following steps: S1. In the data preparation stage, the phenotypic and genotypic data of the training group are cleaned and individuals with missing rates exceeding the preset threshold are removed. Missing genotypic data in the remaining samples are filled in. After merging the genotypic data of the training group and the breeding group, the principal component analysis dimensionality reduction module is used to select the principal components with a cumulative contribution rate of 95% to reduce the data dimensionality. S2. In the model building stage, four machine learning regressors are used to train the dimensionality-reduced data respectively, and the optimal hyperparameters are found through the Bayesian optimization module. The four models are then fused together using the correlation coefficient weighted fusion module to obtain the final fused model. S3. In the model evaluation and optimization stage, the model performance is evaluated using four indicators: Pearson correlation coefficient, mean square error, root mean square error, and mean absolute error. S4. In the prediction and application stage, the genotype data of the breeding population is subjected to PCA dimensionality reduction and transformed into a format recognizable by machine learning algorithms. The constructed fusion model is then used to predict the genotype data of the breeding population and calculate the breeding value.

2. The high-precision intelligent fusion model method for biological breeding as described in claim 1, characterized in that, In step S1, individuals with a phenotypic or genotypic data missing rate exceeding a preset threshold are removed from the sample, and the missing genotypic data in the remaining sample are filled in.

3. The high-precision intelligent fusion model method for biological breeding as described in claim 2, characterized in that, In step S1, eigenvalues ​​and eigenvectors are extracted by calculating the covariance matrix, and principal components with a cumulative contribution rate of 95% are selected as input data after dimensionality reduction.

4. The high-precision intelligent fusion model method for biological breeding as described in claim 3, characterized in that, The four machine learning regressors used in S2 include XGBoost, HGBoost, LightGBM, and Extra Trees.

5. The high-precision intelligent fusion model method for biological breeding as described in claim 4, characterized in that, In S3, the objective function is probabilistically modeled, and the posterior distribution is dynamically updated using prior distribution and observation data to efficiently search for the global optimal solution in the hyperparameter space.

6. The high-precision intelligent fusion model method for biological breeding as described in claim 1, characterized in that, In S4, the Pearson correlation coefficient, mean square error, root mean square error, and mean absolute error of each fold are calculated through ten-fold cross-validation, and the average value is taken as the final evaluation result of the model.

Images