Grape cold resistance character whole genome selective breeding method based on machine learning

Abstract

The invention discloses a grape cold resistance character whole genome selective breeding method based on machine learning, which comprises the following steps: carrying out whole genome re-sequencing and freezing point detection on 352 grape germplasm resources, and carrying out whole genome association analysis, so as to obtain a variation site data set related to cold resistance characters. Then, ten classic regression models are utilized to train a variation site data set after dimension reduction, and phenotypic training and evaluation of the grape cold resistance character are realized. Finally, a Bayesian regression model and 4189 variation points are adopted, the prediction accuracy rate of 82.2% is achieved in a training set, and the Pearson correlation coefficient between a predicted value and a true value in a test set reaches 95%. The invention relates to the technical field of grape cold resistance character whole genome selective breeding. According to the machine learning-based grape cold-resistant character whole-genome selective breeding method, the screening cost of grape filial generations is remarkably reduced, the cold-resistant breeding efficiency is improved, and the method has an important breeding application value.

Description

Technical Field

[0001] The present invention relates to the technical field of whole-genome selection breeding for cold resistance traits of grapes, and specifically to a whole-genome selection breeding method for cold resistance traits of grapes based on machine learning. Background Art

[0002] With the change of the global climate, grape germplasm resources and production are facing severe challenges brought by extreme climates. The trend of decreasing winter temperatures is becoming more significant in many regions, increasing the risk of low-temperature damage to grapevines during overwintering. For example, in northern China, burying the vines for cold protection has become a necessary step in grape cultivation, and this single measure alone accounts for one-third of the production cost of each bottle of wine. Therefore, breeding cold-resistant grape varieties is of great significance for reducing the threat of extreme low temperatures to vineyards. Improving the cold resistance of grapes can not only expand the adaptable areas for grape cultivation, but also reduce the human and resource inputs of farmers in cold protection measures, relieve the restrictions of climate on grape production, and enhance the flexibility of planting areas, especially for grape-growing regions in northern China, mountainous high-altitude areas, and cold temperate regions. In addition, cold-resistant varieties can not only survive better in winter low temperatures, but also maintain longer growth vitality in spring and autumn with lower temperatures. Cold-resistant grapes grown in cool climate regions can often maintain the acidity of the fruits, increase the complexity of flavors, and bring unique flavor characteristics to the wines brewed, which are widely favored by consumers.

[0003] With the rapid development of genome sequencing technology, basic research has found that the cold resistance trait of grapes is controlled by quantitative traits, and the resistance strength is jointly determined by multiple genes associated with numerous minor loci. Therefore, a single locus or a single gene is difficult to fully explain the complex genetic mechanism of the cold resistance trait of grapes. Ignoring the complex interactions between related genes may lead to too simple a prediction of cold resistance, making it difficult to accurately reflect the true cold resistance ability of grape plants, thereby reducing the accuracy and reliability of the prediction. Currently, the whole-genome selection breeding method based on machine learning has been widely used in the improvement of traits such as grape seed abortion, pest resistance, fruit size, and fruit flavor, greatly improving the efficiency of screening breeding offspring. However, in the field of grape cold resistance, this method has not been fully studied. Therefore, using the information of whole-genome variation sites in grapes for the prediction, screening, and retention of cold resistance traits is expected to improve the efficiency of cold resistance breeding, shorten the breeding cycle, increase the adaptable planting areas and supply periods, and thus bring significant economic benefits to the grape industry.

[0004] The present invention aims to develop a fast and efficient prediction method for the cold tolerance phenotype of grapes based on machine learning algorithms. In the present invention, first, a genome-wide association analysis is performed using the freezing point phenotype data and the whole-genome variant locus dataset of the grape natural population. Based on the association analysis results, relevant variant loci are extracted and input into a machine learning model for model training. Then, an independent test set sample is used for phenotype prediction. The genome-wide prediction method proposed by the present invention reaches the accuracy of artificial screening and has scalability. In the future, variant information of other traits can be introduced to support the detection of multiple agronomic traits, providing important practical application value and application prospects for grape breeding. Summary of the Invention

[0005] To achieve the above object, the present invention is realized through the following technical solutions: A genome-wide selection breeding method for the cold tolerance trait of grapes based on machine learning, comprising the following steps: Step 1, Genome-wide association analysis and sample division: The freezing points of the stems of 352 grape germplasm resources are detected, and each sample is repeated 3 times. Each time, 4 stem segments are placed into the sensor module unit. At the same time, whole-genome paired-end sequencing is performed on all grape samples with a depth exceeding 30X. All original re-sequencing data is subjected to quality control processing by the Fastp software, and then the cleaned data is aligned to the grape genome PNT2T by the GTX software, and single nucleotide polymorphism (SNP) and insertion deletion (InDel) variant calls are completed. After quality control of the variant dataset by VCFtools and PLINK, finally 10,929,644 SNP loci and 2,019,598 InDel loci are obtained for genome-wide association analysis (GWAS); Step 2, Training and screening of the best model for the cold tolerance trait of grapes: Based on the GWAS results, the -log 10 [P-wald] values are sorted in descending order, and the top 10,000 to 100,000 loci are extracted respectively. Subsequently, the variant dataset is dimensionally reduced according to the linkage disequilibrium (LD) relationship, and a representative locus in each LD region is selected as the input data. Subsequently, the filtered variant locus dataset is imported into 10 classic regression models for machine learning prediction; Step 3, Phenotype prediction and evaluation of the cold tolerance trait of grapes: Based on the 4,198 variant loci after dimensional reduction and the best model, the Bayesian regression model, phenotype prediction is performed on 70 known phenotype samples in the test set.

[0006] Preferably, in the first step, the mixed linear model (-lmm) of GEMMA and the Wald (P-value) test are used to analyze the correlation between the variant data and the trait data, and data visualization is realized through R scripts.

[0007] Preferably, these models in step (2) include Ridge regression, Lasso regression, Linear regression, Elastic Net regression, Kernel Ridge regression, PLS Regression, Bayesian regression, Elastic Net cross-validation, SVR_linear, and SVR_poly.

[0008] The present invention provides a whole-genome selection breeding method for grape cold resistance traits based on machine learning, having the following beneficial effects: This whole-genome selection breeding method for grape cold resistance traits based on machine learning, based on 4,189 variant sites after dimensionality reduction and a Bayesian regression model, shows an accuracy of 82.2% in the training set of 282 grape samples; in 70 test samples, there is a significant linear correlation (p < 2.2e-16) between the model prediction value and the actual phenotypic value, and the Pearson correlation coefficient reaches R = 0.95. In addition, the Ridge regression model and the SVR_linear model also show high prediction accuracy. The results indicate that the whole-genome selection method based on machine learning reaches the accuracy of manual detection. This method significantly reduces the screening cost of grape hybrid offspring and improves the efficiency of cold resistance breeding, having important breeding application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 is the flow chart of genomic prediction of cold resistance traits of the present invention; Figure 2 is the whole-genome variant density distribution diagram of the present invention; Figure 3 is the prediction effect diagram of the regression model of the present invention; Figure 4 is the prediction result diagram of three regression models based on 4,189 variant sites after dimensionality reduction of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0010] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0011] Please refer to Figures 1 - 4 , the present invention provides a technical solution: a whole-genome selection breeding method for grape cold resistance traits based on machine learning, including the following steps: Step 1: Genome-wide association analysis and sample division. The freezing points of the stems of 352 grape germplasm resources were detected, with each sample repeated 3 times. Each time, 4 stem segments were placed into the sensor module unit. Meanwhile, whole-genome paired-end sequencing of all grape samples was performed with a depth exceeding 30X. All original re-sequencing data was quality-controlled by the Fastp software, and then the cleaned data was aligned to the grape genome PNT2T by the GTX software, and single nucleotide polymorphism (SNP) and insertion-deletion (InDel) variant calls were completed. Quality control of the variant dataset was completed by VCFtools and PLINK. Finally, 10,929,644 SNP sites and 2,019,598 InDel sites were obtained for genome-wide association analysis (GWAS). The mixed linear model (-lmm) of GEMMA and Wald-P value test were used to analyze the correlation between variant data and trait data. Data visualization was achieved through R scripts. Subsequently, the 352 samples were randomly divided into a training set (n1 = 282) and a test set (n2 = 70) in a ratio of 4:1, which were used for machine learning model training and test set prediction evaluation respectively; Step 2: Training and screening of the best model for grape cold resistance traits. Based on the GWAS results, the -log 10 [P-wald] values were sorted in descending order, and the top 10,000 to 100,000 sites were extracted respectively. Subsequently, the variant dataset was dimensionally reduced according to the linkage disequilibrium (LD) relationship, and a representative site in each LD region was selected as the input data. Subsequently, the filtered variant site dataset was imported into 10 classic regression models for machine learning prediction. These models include ridge regression (ridge), lasso regression (Lasso), linear regression (Linear), elastic net regression (ElasticNet), kernel ridge regression (KernelRidge), partial least squares regression (PLSRegression), Bayesian regression (Bayesian), elastic net cross-validation (ElasticNetCV), linear support regression (SVR_linear), and support vector regression - polynomial kernel (SVR_poly); Step 3: Phenotypic prediction and evaluation of grape cold resistance traits. Based on the 4,198 variant sites after dimensional reduction and the best model Bayesian regression model, phenotypic prediction was performed on 70 known phenotypic samples in the test set.

[0012] According to the prediction accuracies of 10 models, it is found that the Bayesian regression model performs the best, followed by the ridge regression model and the linear regression model. Among them, the Bayesian regression model has the best prediction effect when using 4,189 mutation sites, and the accuracy reaches 82.2%. Then, as the number of mutation sites increases, the prediction accuracy gradually reaches a plateau and no longer improves. The specific data are shown in the following table: Appendix 1. Prediction accuracies of 9 classical models Variants Ridge Lasso ElasticNet KernelRidge PLSRegression 10,000 0.761 -0.023 -0.007 0.280 0.625 20,000 0.810 -0.022 -0.005 0.552 0.641 30,000 0.816 -0.025 -0.006 0.583 0.641 40,000 0.820 -0.023 -0.001 0.630 0.646 50,000 0.816 -0.023 -0.004 0.660 0.636 60,000 0.820 -0.023 -0.009 0.666 0.633 70,000 0.817 -0.022 -0.011 0.679 0.627 80,000 0.813 -0.022 -0.012 0.672 0.627 90,000 0.807 -0.023 -0.017 0.669 0.619 100,000 0.811 -0.022 -0.014 0.689 0.616 Variants ElasticNetCV SVR_linear SVR_poly Linear Bayesian 10,000 0.696 0.762 0.585 0.711 0.760 20,000 0.662 0.811 0.579 0.761 0.813 30,000 0.593 0.810 0.575 0.784 0.815 40,000 0.539 0.814 0.573 0.795 0.822 50,000 0.502 0.810 0.560 0.795 0.819 60,000 0.471 0.813 0.554 0.802 0.820 70,000 0.454 0.810 0.552 0.803 0.818 80,000 0.431 0.807 0.541 0.804 0.815 90,000 0.407 0.800 0.535 0.800 0.806 100,000 0.395 0.801 0.530 0.804 0.809 Figure 3 The prediction effects of 3 regression models are shown, namely the Bayesian regression, Ridge regression, and SVR_linear models.

[0013] Figure 4 In the figure, the x-axis represents the actual phenotypic value, and the y-axis represents the phenotypic value predicted by the model. Pearson correlation analysis shows a significant linear relationship between the actual value and the predicted value (p < 2.2e-16), and the correlation coefficient R = 0.95. This indicates that the prediction effect of the model has reached the level of manual detection, can replace cumbersome experiments and manual screening, and can predict the cold resistance of grape seedlings at the seedling stage.

[0014] Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by those of ordinary skill in the art and related fields based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention. The structures, devices, and operation methods not specifically described and explained in the present invention shall be implemented by conventional means in the art without special instructions and limitations.

Claims

1. A whole genome selection breeding method for cold resistance traits of grapes based on machine learning, characterized in that: The steps include: Step 1: Data preparation and whole genome association analysis. The stems of 352 grape germplasm resources were tested at freezing point. Each sample was repeated 3 times, and 4 stem segments were placed in the sensor module unit each time. At the same time, all grape samples were sequenced with full genome double-end sequencing, with a depth of more than 30X. All raw resequencing data were quality controlled by Fastp software, and then the cleaned data were aligned to the grape genome PNT2T by GTX software, and single nucleotide polymorphism and insertion and deletion variant calls were completed. VCFtools and PLINK were used to complete the quality control of the variant data set, and finally 10,929,644 single nucleotide polymorphism sites and 2,019,598 insertion and deletion variant sites were obtained for genome association analysis. Step 2: Training and screening of the best model for grape cold resistance traits, based on the results of genome-wide association analysis, 10 [ P -wald] values ​​were sorted in descending order, and the top 10,000 to 100,000 loci were extracted. Then, the variant dataset was reduced in dimension according to the linkage disequilibrium relationship, and a representative locus was selected in each LD region as input data. Subsequently, the filtered variant site dataset was imported into 10 classical regression models for machine learning prediction; Step 3: Phenotypic prediction and evaluation of grape cold resistance traits. Based on the 4189 variant sites after dimensionality reduction and the best model Bayesian regression, phenotypic prediction was performed on 70 samples with known phenotypes in the test set.

2. The whole genome selection breeding method for cold resistance traits of grapes based on machine learning according to claim 1, characterized in that: In the step 1, the mixed linear model of GEMMA and the Wald (P value) test were used to analyze the correlation between the variation data and the trait data, and data visualization was achieved through R scripts.

3. The whole genome selection breeding method for cold resistance traits of grapes based on machine learning according to claim 1, characterized in that: These models in step two include ridge regression, lasso regression, linear regression, elastic net regression, kernel ridge regression, partial least squares regression, Bayesian regression, elastic net cross validation, linear support regression, and support vector regression-polynomial kernel.

Images