Salt-tolerant soybean whole genome selection method and system based on machine learning and application

Abstract

The invention discloses a salt-tolerant soybean whole genome selection method based on machine learning, and the method comprises the following steps: 1) obtaining a genotype of a data set for modeling and a phenotype after salt treatment, and carrying out treatment; 2) strictly controlling the quality of the SNP sites and the deleted sample; 3) performing whole genome association analysis on the effective variation sites to obtain high-quality variation sites; 4) importing high-quality variation site data in the training set into a convolutional neural network of a CA attention mechanism and a multi-residual module to establish a model; 5) determining an optimal model; and 6) obtaining re-sequencing data of the soybean population with the salt stress phenotype to be tested, extracting high-quality variation sites, and predicting the salt stress phenotype by using the optimal model. The soybean salt stress phenotype can be effectively predicted, the field breeding cost is reduced, and the salt-tolerant soybean breeding efficiency is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of whole genome selection of salt-tolerant soybeans, and in particular relates to a whole genome selection technology for soybean salt-tolerant traits based on machine learning and an application thereof. Background Art

[0002] Soybean (Glycine max) is the world's fourth largest crop and the largest oil crop in terms of planting area. Stabilizing or increasing the domestic soybean supply is of great strategic significance to maintaining national food and oil security and sustainable agricultural development. At present, due to global climate change, unreasonable irrigation and improper fertilizer application, more than 20% of cultivated land is affected by salt stress, and soil salinization is becoming increasingly prominent, seriously threatening my country's food security. Therefore, it is of great theoretical and practical significance to develop salt-alkali tolerant soybean varieties, improve soybean yield and quality, improve land utilization, increase farmers' income and agricultural economic benefits, and promote sustainable agricultural development in my country.

[0003] At present, the research foundation of soybean salt-alkali tolerance is relatively weak, and there are prominent problems such as imperfect salt-alkali tolerance identification technology system and low breeding efficiency. The traditional breeding model mainly relies on specific phenotypic observations and breeders' experience, combined with simple biostatistical analysis, accompanied by a series of problems such as high randomness, low efficiency, weak comprehensive evaluation ability and long breeding cycle. Whole genome selection technology is a new breeding strategy that predicts individual phenotypes through genotypes. It can effectively reduce the high field costs brought by traditional phenotypic selection and improve breeding efficiency.

[0004] However, early genomic selection technologies used fewer genetic variation sites, and the accuracy of phenotypic prediction using traditional statistical methods, such as regression methods, still needs to be improved. With the development of high-throughput sequencing and artificial intelligence, whole genome resequencing and machine learning methods have gradually been applied to crop genomic selection breeding. Existing genomic selection methods mostly use full SNP data analysis, ignoring the important effects of certain key SNPs on phenotypes, which may lead to less than ideal prediction results. Existing machine learning prediction models may have certain limitations in feature extraction and fusion of SNP data, thus affecting prediction accuracy.

[0005] Therefore, how to improve the screening accuracy of salt-tolerant soybeans has become an urgent problem to be solved in the breeding field. Summary of the invention

[0006] In order to solve the above problems in the prior art, the present invention provides a salt-tolerant soybean whole genome selection technology based on machine learning.

[0007] According to a first embodiment of the present invention, a method for whole genome selection of salt-tolerant soybean based on machine learning is provided, comprising the following steps:

[0008] 1) Data acquisition: Obtain soybean population materials and their whole genome resequencing data; Determine the leaf scorch index, Na + , K + , Ca 2+ , Cl - Strict quality control of SNPs was performed, individuals with high deletion rates were deleted, and the deletion rate, heterozygosity rate and minor allele frequency (MAF) of SNP sites were carefully checked and screened.

[0009] 2) Genome-wide association study (GWAS) of soybean salt tolerance phenotype: Plink was used to analyze the population structure and estimate the kinship between samples. A mixed linear model was used to perform association analysis between genotype and phenotype. The genetic correlation matrix of the population structure and the kinship matrix were combined to reduce potential confounding effects and obtain high-quality SNP sites that were significantly associated with the phenotype.

[0010] 3) Model building: SNPs significantly associated with phenotypes were selected as input features of the model. The soybean resequencing dataset was divided into a training set and a test set. The training set was trained using machine learning methods and then the model was built. A prediction model was built using a convolutional neural network with a CA (Coordinate Attention) attention mechanism and multiple residual modules. The model built was ResCANet.

[0011] 4) Model accuracy evaluation: SNP loci and salt stress phenotypes significantly associated with salt stress phenotypes were input into linear regression and machine learning SVR models for prediction. The prediction accuracy and mean squared error (MSE) of linear regression (LR), SVR and ResCANet models were compared. The accuracy of the ResCANet prediction model was significantly improved compared with the LR and SVR models, and the MSE was significantly reduced.

[0012] The second embodiment of the present invention provides a whole genome selection system for salt-tolerant soybean based on machine learning in the previous embodiment, wherein the system comprises the following modules:

[0013] 1) Genotype acquisition module: obtain the leaf scorch, Na + , K + , Ca 2+ , Cl -Content phenotype, obtain soybean germplasm resource resequencing data; use salt stress phenotype whole genome association analysis to screen high-quality SNP sites;

[0014] 2) Model building module: For the soybean salt stress phenotype and genotype datasets, a five-fold cross validation was used, and the data was divided into five subsets for each fold, four of which were used for training and the remaining subset was used for testing, and the Pearson correlation coefficient PCC was calculated; the CA attention mechanism module and the double residual module were introduced to establish the soybean salt stress machine learning prediction model ResCANet;

[0015] 3) Phenotype prediction module for soybean population with salt stress phenotype to be tested: Input the genotype data SNP of the soybean population to be predicted, use the soybean salt stress phenotype machine learning ResCANet model to predict the salt stress phenotype of each individual in the soybean population; select salt-tolerant soybean varieties for field phenotypic experiments based on the predicted phenotypes.

[0016] The third embodiment of the present invention provides an application of the salt-tolerant soybean whole genome selection technology based on machine learning in the previous embodiment in screening and identifying salt-tolerant soybeans, comprising the following steps:

[0017] 1) Obtain the soybean sample to be tested and perform whole genome resequencing on it;

[0018] 2) Obtaining the whole genome SNP sites of the sample to be tested;

[0019] 3) Strict quality control is performed on SNPs, individuals with high deletion rates are deleted, and the deletion rate, heterozygosity rate and secondary MAF of SNP sites are carefully checked and screened.

[0020] 4) SNP sites significantly associated with phenotypes were selected and input into the machine learning ResCANet prediction model to predict the phenotype of soybean salt stress.

[0021] A fourth embodiment of the present invention provides an application of the salt-tolerant soybean whole genome selection technology based on machine learning in the previous embodiment in soybean variety breeding or assisted soybean breeding, comprising the following steps:

[0022] 1) Obtain the soybean sample to be tested and perform whole genome resequencing on it;

[0023] 2) Obtaining the whole genome SNP sites of the sample to be tested;

[0024] 3) Strict quality control is performed on SNPs, individuals with high deletion rates are deleted, and the deletion rate, heterozygosity rate and secondary MAF of SNP sites are carefully checked and screened.

[0025] 4) SNP sites significantly associated with phenotypes were selected and input into the machine learning ResCANet prediction model to predict the phenotype of soybean salt stress.

[0026] 5) Determine the module for combining salt-tolerant soybean selection and breeding, and select soybean varieties with salt-tolerant potential for salt-tolerant selection in the laboratory and field based on the predicted results of phenotypic data.

[0027] The salt-tolerant soybean whole genome selection technology based on machine learning provided in an embodiment of the present invention first investigates the genotype and salt stress phenotype of the soybean population used for modeling, constructs a model using the CA attention mechanism of machine learning and a convolutional neural network with multiple residual modules, and uses the model to predict the salt stress phenotype of the soybean sample to be tested, thereby guiding the selection and breeding of salt-tolerant soybeans.

[0028] The beneficial effect of this example is that it can effectively predict the soybean salt stress phenotype, reduce field breeding costs, and significantly improve the efficiency of salt-tolerant soybean breeding. The advantages of the prediction model are:

[0029] 1. By selecting SNPs that are significantly correlated with the phenotype as the input features of the model, we can focus on variants with important effects, while reducing the number of features and improving model training efficiency and performance.

[0030] 2. Introducing a multi-attention mechanism module, by assigning multiple attention weights to different genetic variation sites, the model can simultaneously focus on multiple features of key sites at different levels. This method can effectively capture complex genetic associations and improve the model's ability to distinguish the importance of different genetic variation sites, thereby more accurately identifying genetic variations associated with phenotypes.

[0031] 3. The CA attention mechanism is introduced to adaptively adjust the weight of each feature channel according to its importance. This mechanism helps the model focus on key gene features more effectively and improves the expression and learning capabilities of gene data by enhancing the feature response of important channels and suppressing irrelevant or redundant channels. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a block diagram of the soybean salt tolerance phenotype prediction method based on whole genome association analysis and machine learning in the present invention

[0033] Figure 2 Manhattan plot of GWAS analysis of soybean leaf scorch index after salt stress in the example

[0034] Figure 3 is the sodium content of soybean in the example + , K + , Cl - and Ca 2+Manhattan plot of GWAS analysis of ion content

[0035] Figure 4 This is a convolutional neural network framework diagram of the prediction model ResCANet in the embodiment based on the CA attention mechanism and multiple residual modules

[0036] Figure 5 Comparison of the prediction effects of the ResCANet prediction model, SVR, and linear regression models in the embodiment

[0037] Figure 6 The actual phenotype of the soybean recombinant inbred line parent after salt stress in the example

[0038] Figure 7 Comparison diagram between the predicted value and the true value obtained using the model ResCANet for a test sample in the embodiment DETAILED DESCRIPTION

[0039] The following will be combined with the attached Figures 1 to 6 , describe the preferred embodiments of the present invention in detail, and further illustrate the present invention.

[0040] The present invention relates to the field of soybean salt tolerance phenotype prediction based on whole genome association analysis, and in particular to a convolutional neural network prediction method combining CA attention mechanism and heavy residual module, the method is as follows Figure 1 shown.

[0041] The specific scheme provided by the present invention is as follows:

[0042] Example 1. Construction of soybean salt stress phenotype prediction model based on machine learning

[0043] 1) Sample phenotype and genotype collection

[0044] 409 soybean natural populations were collected and treated with 150 mM NaCl for 14 days after the first trifoliate leaf of soybean unfolded. The leaf scorch number, NaCl and scorch rate of soybeans grown under salt treatment and control conditions were determined. + , K + , Ca 2+ , Cl - Content. And calculate the leaf scorch index and relative ion content of each variety. Relative value = salt treatment / control x 100%

[0045] The whole genome was resequenced, and the resequencing data was aligned to the soybean Wm82.a4 reference genome to extract the whole genome SNP data of the soybean population. The obtained SNP data were strictly quality controlled to check the integrity and reliability of the data. Specifically, it included retaining biallelic SNP sites, removing sites with a heterozygous rate greater than 30%, filtering individuals with a missing rate greater than 2%, removing samples with SNP missing greater than 5%, removing SNP sites with a missing rate greater than 10% in the sample, and removing SNP sites with a minor allele frequency less than 0.05 (MAF<0.05).

[0046] 2) Genome-wide association analysis of soybean salt stress phenotypes

[0047] After SNP quality control, linkage disequilibrium analysis, population structure analysis, and kinship analysis were used to help reveal the distribution of genetic variation, the genetic relationship between samples, and its impact on phenotypes, thereby improving the accuracy of model prediction and optimizing breeding strategies. Plink was used to analyze the population structure and estimate the kinship between samples. At the same time, GWAS analysis of genotypes and phenotypes was performed using a mixed linear model. A threshold of 1e-3 was used to screen significant SNPs, and the GWAS results were compared with linkage disequilibrium (LD). The results of the genome-wide association analysis of salt stress phenotypes were obtained by Figure 2 and Figure 3 If a group of loci show significant association in GWAS and also show correlation in LD, the reliability of the analysis results can be considered high. This method was used to screen out 7011 SNP loci that were significantly associated with leaf scorch index, and the Na + There are 17799 SNP loci with significant relative content association with K + There are 4692 significant association sites with Ca 2+ There are 10,515 significant association sites.

[0048] 3) Building the prediction model ResCANet

[0049] The SNP sites significantly associated with the phenotype and the soybean salt stress phenotype were used as input to construct a prediction model. This model uses the CA mechanism to focus on key SNP sites and enhances feature extraction and modeling capabilities through multiple residual modules. The CA attention mechanism can significantly improve the prediction ability and interpretability of the model in genomic selection and genetic research by assigning more refined weights to different SNP sites. In the analysis of genomic data, SNP sites often need to be fused at different levels to capture potential interactions. The two-layer residual module can help the model identify the complex interactions between different SNP sites through deeper information fusion, thereby improving the model's ability to model complex gene-phenotype relationships. The first residual module is mainly responsible for the extraction and preliminary fusion of local features, while the second residual module further integrates these features to capture higher-level abstract representations. This dual structure not only helps the model learn the complex relationship between SNP sites more effectively, but also accelerates the training process and improves the robustness and accuracy of the model in large-scale, high-dimensional genetic data. In the ResCANet prediction model, each residual block consists of three convolution branches with ReLU activation function and a feature fusion layer. The fused features are first deeply extracted through a one-dimensional convolution layer containing 10 filters, and then the key features are weighted and enhanced through the channel attention mechanism. Then, the features are corrected for data distribution through a batch normalization layer, and the essence information is further refined by the maximum pooling layer. Finally, the flattened features are sent to a fully connected layer containing 32 neurons, and under the protection of 0.4 ratio Dropout regularization, the prediction results are output through a single neuron fully connected layer. The network structure is as follows: Figure 4 shown.

[0050] Five-fold cross validation was used to divide the data into five subsets for each fold, four of which were used for training and the remaining subset was used for testing, and the Pearson correlation coefficient (PCC) was calculated. Finally, the comprehensive performance evaluation of the model was obtained by calculating the average PCC value of the five folds.

[0051] The model uses Mean Absolute Error (MAE) as the loss function and is trained using a stochastic gradient descent (SGD) optimizer with a learning rate set to 1×10 -3 , the weight decay coefficient is 1×10 -6, the training cycle (epoch) is set to 200. In order to accelerate convergence, the training data is divided into multiple small batches, and the batch size is set to 16. In addition, the model evaluates the data set through five-fold cross validation, dividing the data into five subsets, using each subset as the test set in turn, and the rest as the training set. This method can not only make full use of limited data and reduce the deviation caused by the accidental division of data, but also effectively reduce the risk of model overfitting. Through multiple rounds of validation, cross validation helps to optimize hyperparameters and model structure, further improve the stability and generalization ability of the model, and ensure the reliability and accuracy of the evaluation results. The correlation coefficient PCC calculation formula is as follows:

[0052]

[0053] PCC is used to measure the linear correlation between two variables, and its value ranges from -1 to 1. When PCC>0, it means that the two variables are positively correlated, that is, the increase of one variable is accompanied by the increase of the other variable; when PCC=0, it means that there is no linear relationship between the two variables, that is, there is no predictable linear dependence between them.

[0054] 4) Evaluation of ResCANet prediction model

[0055] The SNP loci and salt stress phenotypes significantly associated with salt stress phenotypes were input into the linear regression (LinearRegression) and machine learning SVR models for prediction. After 200 simulations, the model prediction accuracy was compared with the prediction accuracy of the ResCANet prediction model. It was found that the average accuracy of the ResCANet model was 0.88, the average accuracy of the SVR model was 0.55, and the average accuracy of the LR model was 0.43. The MSE of the ResCANet model was 0.006, which was significantly lower than the other two models. Comparison of the prediction effect of the ResCANet prediction model with the SVR and linear regression models is shown in Figure 5 This indicates that the ResCANet prediction model has the highest accuracy and is suitable for salt stress phenotype prediction.

[0056] Example 2: Predicting the Salt Stress Phenotype of Soybean Using Resequencing Data from Natural Populations

[0057] 1) Sample collection

[0058] We collected a soybean recombinant inbred lines (RIL) population of 177 individuals. The leaf scorch number of the two parents of this population was significantly different after salt treatment ( Figure 6). The whole genome of the RIL population was resequenced, and the genome was aligned to extract the whole genome SNP sites of all the materials to be tested. According to the SNP sites significantly associated with the soybean salt stress phenotype identified in the first embodiment, the SNP sites of the samples to be tested were extracted.

[0059] 2) Predict the salt stress phenotype of the sample to be tested

[0060] The extracted SNP loci were input into the ResCANet prediction model to obtain the predicted phenotypic value of soybean salt stress for each sample to be tested.

[0061] 3) Evaluate the accuracy of soybean salt stress phenotype prediction

[0062] The tested material was planted in a greenhouse (12 hours light / 12 hours dark). After the first trifoliate compound leaf of the plant unfolded, the soybean material was treated with 150 mM NaCl for 14 days, and its leaf scorch index and relative ion content were measured. The Pearson correlation coefficient between the predicted value and the observed value was calculated using the ResCANet method, and the average correlation coefficient between the predicted phenotype and the observed phenotype was 0.855. Figure 7 This indicates that the whole genome selection technology significantly improves the selection efficiency and significantly reduces the field workload.

Claims

1. A method for whole genome selection of salt-tolerant soybean based on machine learning, characterized in that The method comprises the following steps: 1) Obtain the genotypes and salt-treated phenotypes of the dataset for modeling and process them; 2) Strict quality control of SNP sites and deletion samples; 3) Performing genome-wide association analysis on the effective variant sites to obtain high-quality variant sites; 4) Importing the high-quality variant site data in the training set into a convolutional neural network with a CA attention mechanism and multiple residual modules to establish a model, the model established is ResCANet; 5) Use the test set to verify the model prediction effect and determine the optimal model; 6) Obtain resequencing data of the soybean population to be tested for salt stress phenotype, extract its high-quality variant sites, and use the optimal model to predict the salt stress phenotype.

2. The method according to claim 1, characterized in that It also includes the steps of: selecting salt-tolerant soybean varieties for field phenotypic experiments based on the predicted soybean salt stress phenotype.

3. The method according to claim 1, characterized in that Step 1) specifically includes: 1-1) Collecting a natural soybean population, and treating the soybean with 150 mM NaCl for 14 days after the first trifoliate leaf of the soybean unfolded; 1-2) Identification of leaf scorch number, Na + , K + , Ca 2+ , Cl - content; 1-3) Calculate the leaf scorch index and relative ion content of each variety, relative value = ion content of each soybean under salt treatment conditions / ion content of each soybean under control conditions x 100%; 1-4) Perform whole genome resequencing on natural soybean populations, align the resequencing data to the soybean Wm82.a4 reference genome, and extract whole genome SNP data of the soybean population.

4. The method according to claim 1, characterized in that The strict quality control described in step 2) specifically includes: The double allelic variation SNP sites were retained, the sites with heterozygosity rate greater than 30% were removed, the individuals with missing rate higher than 2% were filtered, the samples with SNP missing rate greater than 5% were removed, the SNP sites with missing rate higher than 10% in the samples were removed, and the SNP sites with minor allele frequency MAF less than 0.05 were removed.

5. The method according to claim 1, characterized in that The whole genome association analysis described in step 3) is specifically: using Plink to analyze the population structure and estimate the kinship between samples, and at the same time performing association analysis on genotypes and phenotypes through a mixed linear model, combining the genetic correlation matrix of the population structure and the kinship matrix to reduce potential confounding effects and obtain high-quality variant sites.

6. The method according to claim 1, characterized in that The model established in step 4) uses the CA mechanism to focus on key SNP sites, and enhances feature extraction and modeling capabilities through multiple residual modules; wherein: The CA attention mechanism is used to more finely assign weights to different SNP sites, significantly improving the predictive power and interpretability of the model in genomic selection and genetic research; The use of two residual modules can help the model identify the complex interactions between different SNP sites through deeper information fusion, thereby improving the model's ability to model complex gene-phenotype relationships. The first residual module is mainly responsible for the extraction and preliminary fusion of local features, while the second residual module further integrates these features to capture higher-level abstract representations. This dual structure not only helps the model learn the complex relationships between SNP sites more effectively, but also accelerates the training process and improves the robustness and accuracy of the model in large-scale, high-dimensional genetic data. In the ResCANet prediction model, each residual block consists of three convolution branches with ReLU activation function and a feature fusion layer. The fused features are first deeply extracted through a one-dimensional convolution layer containing 10 filters, and then the key features are weighted and enhanced through the channel attention mechanism. Then, the features are corrected for data distribution through a batch normalization layer, and the essence information is further refined by the maximum pooling layer. Finally, the flattened features are sent to a fully connected layer containing 32 neurons, and under the protection of 0.4 ratio Dropout regularization, the prediction results are output through a single neuron fully connected layer. The model uses mean absolute error (MAE) as the loss function and is trained using the stochastic gradient descent (SGD) optimizer with a learning rate set to 1×10. -3 , the weight decay coefficient is 1×10 -6 , the training cycle epoch is set to 200; in order to accelerate convergence, the training data is divided into multiple small batches, and the batch size is set to 16; in addition, the model is evaluated on the dataset through five-fold cross validation, the data is divided into five subsets, each subset is used as the test set in turn, and the rest is used as the training set, and the Pearson correlation coefficient PCC is calculated; finally, the comprehensive performance evaluation of the model is obtained by calculating the average PCC value of the five folds.

7. The method according to claim 1, characterized in that The genotype data of the soybean varieties include salt-tolerant and salt-sensitive soybean varieties.

8. A whole genome selection system for salt-tolerant soybean based on machine learning, characterized in that The system includes the following modules: 1) Genotype acquisition module: obtain the leaf scorch, Na + , K + , Ca 2+ , Cl - Content phenotype, obtain soybean germplasm resource resequencing data; use salt stress phenotype whole genome association analysis to screen high-quality SNP sites; 2) Model building module: For the soybean salt stress phenotype and genotype datasets, a five-fold cross validation was used, and the data was divided into five subsets for each fold, four of which were used for training and the remaining subset was used for testing, and the Pearson correlation coefficient PCC was calculated; the CA attention mechanism module and the double residual module were introduced to establish the soybean salt stress machine learning prediction model ResCANet; 3) Phenotype prediction module for soybean population with salt stress phenotype to be tested: Input the genotype data SNP of the soybean population to be predicted, use the soybean salt stress phenotype machine learning ResCANet model to predict the salt stress phenotype of each individual in the soybean population; select salt-tolerant soybean varieties for field phenotypic experiments based on the predicted phenotypes.

9. Application of the whole genome selection method of salt-tolerant soybean based on machine learning in the screening and identification of salt-tolerant soybean.

10. Application of salt-tolerant soybean whole genome selection method based on machine learning in soybean variety selection or assisted soybean breeding.

Images