A multi-view gblup method for integrating multiple types of data for phenotype prediction

Abstract

The application discloses a multi-view GBLUP method for integrating multiple types of data for phenotype prediction, comprising the following steps: representing SNP genotype data as multiple types of data in the form of 0 / 1 / 2 according to different phenotype types, and performing Z-score standardization on the data; setting the i-th row and the j-th row in the multiple types of data matrix as individuals x i ,x j , calculating a Kinship matrix, and defining a similarity function between two individuals; integrating the multiple types of data into a Kinship matrix, using a differential evolution algorithm to perform mutation and crossover operations on a population, comparing individuals in the population with individuals after mutation and crossover, assigning weights to the Kinship matrix of the multiple types of data, and learning by substituting the Kinship matrix into a GBLUP model; selecting individuals with high phenotype prediction accuracy as members of the next generation population; and setting a termination condition and outputting an optimal weight combination of the multiple types of data. The application can quickly and accurately obtain an optimal weight combination of multiple types of data to assist in phenotype prediction.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and in particular to a multi-view GBLUP method for integrating multiple types of data for phenotypic prediction. Background Technology

[0002] The same thing can be described from different perspectives, forming multiple views of that thing. Multi-view data of things exists widely in the real world; for example, in breeding, a sample can be represented by its associated genes, expression levels, metabolism, and phenotype. Utilizing the complementarity between multi-view data can improve algorithm performance to some extent.

[0003] With the development of next-generation high-throughput genotyping technologies, DNA microarrays and whole-genome resequencing have become common tools in plant genomics, making genome prediction applicable to crop genetic breeding possible. Genome prediction methods are mainly divided into two categories: traditional statistical methods, such as Genome Optimal Linear Unbiased Prediction (GBLUP), Least Absolute Shrinkage and Selection Operator (LASSO), and Bayesian methods; and machine learning methods, such as Support Vector Machines, Random Forests, and Regenerating Kernel Hilbert Spaces. Among these, GBLUP is the most widely used due to its robustness in analyzing real-world data. With the advent of the big data era, multi-omics joint prediction has attracted researchers' attention. Studies have found that using any two omics data from the genome, transcriptome, and metabolome, or using all three simultaneously, can improve phenotypic prediction accuracy, but the improvement is not significant. Looking back at the development of genome breeding over the past 20 years, there are three main approaches: single-trait prediction using traditional statistical methods and machine learning methods; multi-trait prediction using multiple linear models; and joint prediction by integrating omics data. Although existing genome prediction methods have achieved good results, they do not make full use of omics big data, and there is still room for improvement in their predictive capabilities.

[0004] As data accumulation increases exponentially, especially with the continuous reduction in the cost of various omics data and the continuous increase in the cost of field phenotyping, there is a need to innovate approaches to genome prediction by comprehensively utilizing omics data. Summary of the Invention

[0005] The technical problem to be solved by this invention is to address the shortcomings of existing technologies in the insufficient utilization of multi-type big data, and to provide a multi-view GBLUP method for integrating multi-type data for phenotypic prediction.

[0006] The technical solution adopted by this invention to solve its technical problem is:

[0007] This invention provides a multi-view GBLUP method for integrating multiple types of data for phenotypic prediction, the method comprising the following steps:

[0008] Step 1: Obtain SNP genotype data as the training set. Represent the SNP genotype data as multi-type data in the form of 0 / 1 / 2 according to different phenotypic types. Write the multi-type data into a matrix, where the rows of the matrix represent materials and the columns of the matrix represent features. Perform Z-score standardization on each column of data.

[0009] Step 2, let the i-th row and j-th row of the multi-type data matrix be individual x. i ,x j Calculate the Kinship matrix and define the similarity function between two individuals;

[0010] Step 3: Based on the similarity between individuals, the multi-type data is weighted and integrated into a Kinship matrix. The differential evolution algorithm is used. The specific process is as follows: Initialize the population. Each individual in the population is a weighted combination of a set of multi-type data. Mutation and crossover operations are performed on the population. The individuals in the population are compared with the individuals after mutation and crossover. The weights are assigned to the Kinship matrix of the multi-type data. The matrix is ​​then substituted into the GBLUP model for learning. Individuals with high phenotypic prediction accuracy are selected as members of the next generation population.

[0011] Step 4: Set a termination condition. Stop learning when the absolute difference between the maximum prediction accuracy of two adjacent training sets is less than 0.0001, and output the optimal weight combination for multiple types of data.

[0012] Step 5: Set up the GBLUP model according to the optimal weight combination, input the SNP genotype data to be predicted, and output the predicted phenotype.

[0013] Furthermore, in step 1 of the present invention, SNP genotype data are represented as 0 / 1 / 2 according to different phenotypic types, where 0 represents a genotype with a high homozygous frequency, 1 represents a heterozygote, and 2 represents a genotype with a low homozygous frequency.

[0014] Furthermore, in step 1 of this invention, the standardized processing formula is:

[0015]

[0016] Where X, E(X), and Var(X) represent the vector to be standardized, the expected value of X, and the variance of X, respectively.

[0017] Furthermore, in step 2 of this invention, the similarity function between two individuals is defined as:

[0018]

[0019] Among them, (x i ,xj ) represents individual x i ,x j The inner product of.

[0020] Furthermore, in step 3 of this invention, the specific steps of the differential evolution algorithm are as follows:

[0021] Step 3.1, Initialization; Initialize NP population individuals, each individual being a weighted combination of multiple data types, with each initial weight between (0,1);

[0022] Step 3.2, Mutation; For each individual in the initial population, a mutated individual is generated using a weighted average relaxation technique. The mutation calculation process is as follows:

[0023]

[0024] Among them, V i 0 It is a mutated individual. These are two randomly selected different individuals, ||·||2 represents the norm of the vector, and ω is a relaxation parameter that controls the degree of variation;

[0025] Step 3.3, crossover; For each individual in the initial population, crossover is performed between the mutated individual and the original individual to generate experimental individuals;

[0026] Step 3.4, Selection: Compare the individuals in the initial population with the experimental individuals after crossover, assign weights to the Kinship matrix of the multi-type data, substitute it into the GBLUP model for learning, and select individuals with higher phenotypic prediction accuracy as members of the next generation population; the selection method is as follows:

[0027]

[0028] in, These are crossover test individuals. This is the original population of this generation. These are the individuals from this generation that are retained as members of the next generation of the population. f(X) represents the fitness function. This method uses the prediction accuracy of the training set's five-fold cross-validation as the function value.

[0029] Furthermore, in step 4 of the present invention, the termination condition is set as follows:

[0030]

[0031] Where f(X) represents the fitness function, and ∈ represents a given small positive number, which is taken as 0.0001 in this method.

[0032] Furthermore, in step 3.4 of this invention, the specific steps of assigning weights to the Kinship matrix of multiple data types and substituting it into the GBLUP model for learning are as follows:

[0033] The GBLUP model is as follows:

[0034] y = Xβ + Zu + ε, where y is an n×1 phenotypic vector, X is an n×p design matrix corresponding to fixed effects, β is a p×1 coefficient vector representing fixed effects, and Z is an n×n design matrix corresponding to random effects. (Multigene effects) K is the kinship matrix and the residual effect vector. I is the identity matrix. and These are the estimated genetic variance and residual variance, respectively.

[0035] Let A i For n×p i Given a multi-type data matrix (i = 1, 2, ..., m), define the multi-type data kinship matrix as K = w1A1A1 T +w2A1A1 T +…+w m A m A m T The weight vector W = (w1, w2, ..., w m ) T It is learned by the differential evolution algorithm.

[0036] The beneficial effects of this invention are:

[0037] This invention provides a solution for predicting phenotypes using multi-type data by optimizing weights through a differential evolution algorithm. The invention first randomly initializes a weight population and then performs three stages—mutation, crossover, and selection—to obtain the population for each iteration, which serves as the original population for the next generation. Next, weight combinations are assigned to multi-type data, and a kinship matrix between samples is calculated using a similarity metric function. This matrix is ​​then input into the GBLUP model to obtain the prediction accuracy for each weight combination. Finally, five-fold cross-validation is used to evaluate different weight combinations. An early stopping mechanism is implemented by comparing the difference in maximum prediction accuracy between two adjacent iterations, enabling rapid and accurate acquisition of the optimal weight combination for multi-type data, thus aiding in phenotype prediction. Attached Figure Description

[0038] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0039] Figure 1 This is a flowchart illustrating a specific implementation of the present invention;

[0040] Figure 2 A comparison of the accuracy of different methods in predicting rice phenotypes using genomic and metabolomics data;

[0041] Figure 3 A comparison of the accuracy of different methods in predicting rice phenotypes using genomic and transcriptomic data;

[0042] Figure 4 A comparison of the accuracy of different methods in predicting rice phenotypes using metabolomics and transcriptomics data;

[0043] Figure 5 A comparison of the accuracy of different methods in predicting rice phenotypes using genomic, metabolomics, and transcriptomics data;

[0044] Figure 6 A comparison of the accuracy of different methods in predicting maize cob diameter, ear diameter, ear leaf width, and ear length using genomic and metabolomics data;

[0045] Figure 7 A comparison of the accuracy of different methods in predicting phenotypic parameters such as ear leaf length, tasseling date, pollen shedding date, and silking date in maize using genomic and metabolomics data;

[0046] Figure 8 A comparison of the accuracy of different methods in predicting maize cob diameter, ear diameter, ear leaf width, and ear length using genomic and transcriptomic data;

[0047] Figure 9 A comparison of the accuracy of different methods in predicting phenotypic parameters such as ear leaf length, tasseling date, pollen shedding date, and silking date in maize using genomic and transcriptomic data;

[0048] Figure 10 A comparison of the accuracy of different methods in predicting maize cob diameter, ear diameter, ear leaf width, and ear length using transcriptomic and metabolomic data;

[0049] Figure 11 A comparison of the accuracy of different methods in predicting maize ear leaf length, tasseling date, pollen shedding date, and silking date using transcriptomic and metabolomic data;

[0050] Figure 12 A comparison of the accuracy of different methods in predicting maize cob diameter, ear diameter, ear leaf width, and ear length using genomic, transcriptomic, and metabolomic data;

[0051] Figure 13 A comparison of the accuracy of different methods in predicting maize ear leaf length, tasseling date, pollen shedding date, and silking date using genomic, transcriptomic, and metabolomic data;

[0052] Figure 14 A comparison of different methods using genomic and transcriptomic data from seven different tissues to predict maize flowering days, ear height, leaf width, and upper leaf angle;

[0053] Figure 15 A comparison of the accuracy of different methods in predicting maize pollen-silking interval, number of internodes below the ear, plant height, and weight per 20 kernels using genomic and transcriptomic data from seven different tissues;

[0054] Figure 16 A comparison of the accuracy of different methods in predicting the soluble solids content of tomatoes using genomic, insertion / deletion, and structural variation data;

[0055] Figure 17 The multi-view GBLUP method uses genomic and transcriptomic data from seven different tissues to train the model iterative process (taking the maize flowering days phenotype as an example);

[0056] Figure 18 The multi-view GBLUP method uses genomic and transcriptomic data from seven different tissues to learn the changes in optimized weights during the learning process (taking the maize flowering days phenotype as an example). Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0058] like Figure 1 As shown, the multi-view GBLUP method for integrating multiple types of data for phenotypic prediction according to an embodiment of the present invention includes the following steps:

[0059] Step 1: Represent the SNP genotype data as 0 / 1 / 2, where 0 represents a genotype with a high homozygous frequency, 1 represents a heterozygous genotype, and 2 represents a genotype with a low homozygous frequency.

[0060] The data of various types are written as matrices, with the rows of the matrix representing materials and the columns representing features. The data in each column is standardized by Z-score, and the calculation formula is shown in Eq.(1).

[0061]

[0062] Step 2, let the i-th row and j-th row of the multi-type data matrix be individual x. i ,x j Calculate the Kinship matrix for a given pair of individuals x. i ,x j The similarity function between two individuals is defined as shown in Eq.(2).

[0063]

[0064] Step 3: Weight the data from different data types into a Kinship matrix. Using the differential evolution algorithm, through mutation, crossover, and selection operations, find the optimal weighted population in each generation. This optimal weighted population is then used as the original population for the next generation. The specific process is as follows:

[0065] Step 3.1, Initialization. Let the population of each generation be NP. Initialize NP individuals, each of which is a weighted combination of multiple data types, with each weight initially between (0,1).

[0066] Step 3.2, Mutation; For each individual in the initial population, a mutated individual is generated using a weighted average relaxation technique. The mutation calculation process is as follows:

[0067]

[0068] Among them, V i 0 It is a mutated individual. These are two randomly selected different individuals, ||·||2 represents the norm of the vector, and ω is a relaxation parameter that controls the degree of variation;

[0069] Step 3.3, Crossover. For each individual in the initial population, crossover is performed between the mutated individual and the original individual to generate experimental individuals.

[0070] Step 3.4, Selection. Compare the individuals in the initial population with the experimental individuals after crossover, assign weights to the Kinship matrix of the multi-type data, substitute it into the GBLUP model for learning, and select individuals with higher phenotypic prediction accuracy as members of the next generation population. The selection method is shown in Eq(4).

[0071]

[0072] in, These are crossover test individuals. This is the original population of this generation. These are the individuals from this generation who are preserved as members of the next generation of the species.

[0073] Step 4, Stopping Mechanism. Repeat the above process until the similarity between the upper and lower generations is less than a certain threshold, preferably 0.0001. The algorithm ends when the individual with the highest phenotypic prediction accuracy is the optimal weight combination for the multi-type dataset. The stopping mechanism is shown in Eq(5).

[0074]

[0075] Test Implementation Example:

[0076] The following are examples of applying the method of this invention to integrate multiple data types to predict phenotypes on the published Maize 368, Maize 282, Rice 210, and Tomato 332 datasets. Related experiments illustrate the efficiency and accuracy of the method for predicting phenotypes. The methods compared in the experiments include: directly summing multi-omics kinship matrices to predict phenotypes, using prediction accuracy as the weight of the multi-omics kinship matrix to predict phenotypes, and the multi-view GBLUP method to predict phenotypes.

[0077] Experiment 1. Comparison of the accuracy of different methods in predicting rice phenotypes using genomic and metabolomics data.

[0078] In this experiment, we compared the accuracy of predicting rice phenotypes using both genomic and metabolomics data on the Rice210 ​​dataset.

[0079] like Figure 2 As shown, there is a significant difference in the accuracy of predicting rice phenotypes using genomic data alone and using metabolomics data alone, with the highest difference in single-trait prediction accuracy being 0.281. Genomic data showed the highest accuracy in predicting tiller number; a 1:1 weighted approach between the two omics metrics resulted in the highest accuracy in predicting grain number and yield; and using prediction accuracy as a weight resulted in the highest accuracy in predicting thousand-grain weight. The prediction accuracy of the multi-view GBLUP method was close to the highest prediction method, differing by only 0.006-0.01. Compared to the best results from single-omics prediction, the multi-view GBLUP method improved accuracy by up to 0.072.

[0080] Experiment 2. Comparison of the accuracy of different methods in predicting rice phenotypes using genomic and transcriptomic data.

[0081] In this experiment, we compared the accuracy of predicting rice phenotypes using both genomic and transcriptomic data on the Rice210 ​​dataset.

[0082] like Figure 3 As shown, there is a significant difference in the accuracy of predicting rice phenotypes using genomic data alone and transcriptome data alone, with the highest difference in single-trait prediction accuracy being 0.228. Genomic data showed the highest accuracy in predicting tiller number; a 1:1 weighted approach between the two omics was the most accurate for predicting yield traits, with the multi-view GBLUP prediction method achieving accuracy close to the highest-performing method, differing by only 0.002-0.008. The multi-view GBLUP prediction method showed the highest accuracy in predicting grain number and thousand-grain weight traits, improving accuracy by up to 0.045 compared to a 1:1 weighted approach; and by up to 0.043 compared to using prediction accuracy as a weight for phenotype prediction. Compared to the best results from single-omics predictions, the multi-view GBLUP prediction method showed a maximum accuracy improvement of 0.013.

[0083] Experiment 3. Comparison of the accuracy of different methods in predicting rice phenotypes using metabolomics and transcriptomics data.

[0084] In this experiment, we compared the accuracy of predicting rice phenotypes using both metabolomics and transcriptomics data on the Rice210 ​​dataset.

[0085] like Figure 4 As shown, there is a significant difference in the accuracy of predicting rice phenotypes using metabolomics data alone and transcriptomics data alone, with the highest difference in single-trait prediction accuracy being 0.054. Metabolomics showed the highest accuracy in predicting yield traits. When the two omics were weighted 1:1, the accuracy in predicting grain number and tiller number traits was the highest. The multi-view GBLUP prediction method's accuracy was close to the highest prediction method, differing by only 0.01-0.026. The multi-view GBLUP prediction method showed the highest accuracy in predicting thousand-grain weight, improving accuracy by up to 0.021 compared to predicting phenotypes using the two omics weighted 1:1; and also improving accuracy by up to 0.021 compared to using prediction accuracy as a weight for phenotype prediction. Compared to the best results from single-omics predictions, the multi-view GBLUP prediction method showed a maximum improvement in accuracy of 0.032.

[0086] Experiment 4. Comparison of the accuracy of different methods in predicting rice phenotypes using genomic, metabolomics, and transcriptomics data.

[0087] In this experiment, we compared the accuracy of predicting rice phenotypes using three types of data: genomics, metabolomics, and transcriptomics on the Rice210 ​​dataset.

[0088] like Figure 5 As shown, there are significant differences in the accuracy of predicting rice phenotypes using genomic data alone, metabolomics data alone, and transcriptomics data alone, with the highest difference in single-trait prediction accuracy being 0.028. Genomic data showed the highest accuracy in predicting tiller number, while the multi-view GBLUP prediction method achieved accuracy close to the highest prediction method, differing by only 0.006. The multi-view GBLUP prediction method showed the highest accuracy in predicting grain number, thousand-grain weight, and yield traits, improving accuracy by up to 0.058 compared to predicting phenotypes by assigning weights to the three omics in a 1:1:1 ratio; and by up to 0.056 compared to using prediction accuracy as a weight for phenotype prediction. Compared to the best results from single-omics predictions, the multi-view GBLUP prediction method showed the highest improvement in accuracy of 0.036.

[0089] Experiment 5. Comparison of the accuracy of different methods in predicting maize phenotypes using genomic and metabolomics data.

[0090] In this experiment, we compared the accuracy of predicting maize phenotypes using genomic and metabolomics data on the Maize368 dataset.

[0091] like Figure 6As shown, there is a significant difference in the accuracy of predicting maize phenotypes using genomic data alone and metabolomics data alone, with the highest difference in single-trait prediction accuracy being 0.183. The multi-view GBLUP prediction method showed the highest accuracy in predicting maize cob diameter, ear diameter, ear-position leaf width, and ear length, improving accuracy by up to 0.024 compared to predicting phenotypes with the two omics weighted at a 1:1 ratio; and improving accuracy by up to 0.026 compared to using prediction accuracy as a weight for phenotype prediction. Compared to the best results of single-omics prediction, the multi-view GBLUP prediction method improved accuracy by up to 0.027.

[0092] like Figure 7 As shown, there is a significant difference in the accuracy of predicting maize phenotypes using genomic data alone and metabolomics data alone, with the highest difference in single-trait prediction accuracy being 0.171. The highest accuracy in predicting the ear-leaf length trait was achieved by assigning a 1:1 weighted ratio between the two omics approaches, while the multi-view GBLUP prediction method's accuracy was close to the highest prediction method, differing by only 0.002. The multi-view GBLUP prediction method showed the highest accuracy in predicting the traits of tasseling, pollen shedding, and silking stages in maize, with the highest accuracy improvement of 0.013 compared to the 1:1 weighted ratio prediction of the two omics approaches; and the highest accuracy improvement of 0.015 compared to using prediction accuracy as a weight for phenotype prediction. Compared to the best results of single-omics prediction, the multi-view GBLUP prediction method showed the highest accuracy improvement of 0.018.

[0093] Experiment 6. Comparison of the accuracy of different methods in predicting maize phenotypes using genomic and transcriptomic data.

[0094] In this experiment, we compared the accuracy of predicting maize phenotypes using both genomic and transcriptomic data on the Maize368 dataset.

[0095] like Figure 8 As shown, there is a significant difference in the accuracy of predicting maize phenotypes using genomic data alone and transcriptome data alone, with the highest difference in single-trait prediction accuracy being 0.193. Genomic data showed the highest accuracy in predicting maize cob diameter, ear-position leaf width, and ear length; using prediction accuracy as a weight resulted in the highest accuracy for ear diameter. The prediction accuracy of the multi-view GBLUP method was close to the highest prediction method, differing by only 0.005-0.01. Compared to the best results from single-omics prediction, the multi-view GBLUP method improved accuracy by up to 0.008.

[0096] like Figure 9As shown, there is a significant difference in the accuracy of predicting maize phenotypes using genomic data alone and transcriptome data alone, with the highest difference in single-trait prediction accuracy being 0.066. Genomic data showed the highest accuracy in predicting ear-leaf length; transcriptome data showed the highest accuracy in predicting the silking stage phenotype; and a 1:1 weighted approach for both omics showed the highest accuracy in predicting the tasseling and pollen shedding stages. The prediction accuracy of the multi-view GBLUP method was close to the highest prediction method, differing by only 0.0002-0.006. Compared to the best results from single-omics predictions, the multi-view GBLUP method improved accuracy by up to 0.019.

[0097] Experiment 7. Comparison of the accuracy of different methods in predicting maize phenotypes using transcriptomics and metabolomics data.

[0098] In this experiment, we compared the accuracy of predicting maize phenotypes using both transcriptomic and metabolomic data on the Maize368 dataset.

[0099] like Figure 10 As shown, there is a significant difference in the accuracy of predicting maize phenotypes using transcriptome data alone and using metabolome data alone, with the highest difference in single-trait prediction accuracy being 0.093. Using prediction accuracy as a weight yields the highest accuracy for predicting maize ear length, with the multi-view GBLUP prediction method achieving accuracy close to the highest prediction method, differing by only 0.004. The multi-view GBLUP prediction method shows the highest accuracy in predicting maize cob diameter, ear diameter, and ear-position leaf width phenotypes, improving accuracy by up to 0.007 compared to predicting phenotypes with a 1:1 weighting of the two omics; and improving accuracy by up to 0.01 compared to predicting phenotypes using prediction accuracy as a weight. Compared to the best results from single-omics predictions, the multi-view GBLUP prediction method shows an improvement in accuracy of up to 0.016.

[0100] like Figure 11 As shown, there is a significant difference in the accuracy of predicting maize phenotypes using transcriptome data alone and using metabolome data alone, with the highest difference in single-trait prediction accuracy being 0.18. The highest accuracy in predicting maize phenotypes at tasseling, pollen shedding, and silking stages was achieved when the two omics were weighted 1:1. The prediction accuracy of the multi-view GBLUP method was close to the highest prediction method, differing by only 0.003-0.004. The multi-view GBLUP method showed the highest accuracy in predicting the ear-leaf length phenotype, improving accuracy by up to 0.002 compared to the 1:1 weighted prediction of the two omics; and by up to 0.004 compared to using prediction accuracy as a weight for phenotype prediction. Compared to the best results of single-omics prediction, the multi-view GBLUP method improved accuracy by up to 0.01.

[0101] Experiment 8. Comparison of the accuracy of different methods in predicting maize phenotypes using genomic, transcriptomic, and metabolomic data.

[0102] In this experiment, we compared the accuracy of predicting maize phenotypes using three types of data: genomics, transcriptomics, and metabolomics on the Maize368 dataset.

[0103] like Figure 12 As shown, there are significant differences in the accuracy of predicting maize phenotypes using genomic data alone, transcriptomic data alone, and metabolomics data alone, with the highest difference in single-trait prediction accuracy being 0.193. The multi-view GBLUP prediction method showed the highest accuracy in predicting maize cob diameter, ear diameter, ear-position leaf width, and ear length phenotypes, improving accuracy by up to 0.036 compared to predicting phenotypes with a 1:1 weighting of the two omics; and improving accuracy by up to 0.028 compared to using prediction accuracy as a weight for phenotype prediction. Compared to the best results of single-omics prediction, the multi-view GBLUP prediction method improved accuracy by up to 0.023.

[0104] like Figure 13 As shown, there are significant differences in the accuracy of predicting maize phenotypes using genomic data alone, transcriptomic data alone, and metabolomics data alone, with the highest difference in single-trait prediction accuracy being 0.18. Genomic data showed the highest accuracy in predicting ear-leaf length. The three omics weighted at a 1:1:1 ratio showed the highest accuracy in predicting phenotypes at tasseling and pollen shedding stages, with the multi-view GBLUP prediction method achieving accuracy close to the highest-performing method, differing by only 0.004-0.007. The multi-view GBLUP prediction method showed the highest accuracy in predicting the silking stage phenotype; compared to the 1:1:1 weighted phenotype prediction, the accuracy improved by up to 0.008; compared to using prediction accuracy as a weight for phenotype prediction, the accuracy improved by up to 0.008. Compared to the best results from single-omics prediction, the multi-view GBLUP prediction method showed the highest improvement in accuracy of 0.02.

[0105] Experiment 9. Comparison of the accuracy of different methods in predicting maize phenotypes using genomic and transcriptomic data from seven different tissues.

[0106] In this experiment, we compared the accuracy of predicting maize phenotypes using genomic data and transcriptome data from seven different tissues on the Maize282 dataset.

[0107] like Figure 14As shown, there is a significant difference in the accuracy of predicting maize phenotypes using genomic data alone and using transcriptome data from different tissues alone, with the highest difference in single-trait prediction accuracy being 0.355. The highest accuracy was achieved in predicting maize leaf width and upper leaf angle phenotypes when genomic and seven different tissue transcriptome data were weighted at a ratio of 1:1:1:1:1:1:1:1. The multi-view GBLUP prediction method showed near-perfect accuracy, differing from the best prediction method by only 0.018-0.02. The multi-view GBLUP prediction method showed the highest accuracy in predicting maize flowering days and ear height phenotypes, improving accuracy by up to 0.02 compared to weighting the eight datasets at a ratio of 1:1:1:1:1:1:1:1; and by up to 0.079 compared to using prediction accuracy as a weight for phenotype prediction. Compared to the best results from single-omics prediction, the multi-view GBLUP prediction method showed a maximum accuracy improvement of 0.034.

[0108] like Figure 15 As shown, there is a significant difference in the accuracy of predicting maize phenotypes using genomic data alone and using transcriptome data from different tissues alone, with the highest difference in single-trait prediction accuracy being 0.428. The highest accuracy was achieved in predicting maize plant height and kernel weight per 20 kernels using a weighted ratio of 1:1:1:1:1:1:1:1:1 for genomic and seven different tissue transcriptome data. The multi-view GBLUP prediction method showed accuracy close to the highest prediction method, differing by only 0.013-0.019. The multi-view GBLUP prediction method showed the highest accuracy in predicting the pollen-silking interval and the number of internodes below the ear, with the highest accuracy improvement of 0.054 compared to predicting phenotypes using a weighted ratio of 1:1:1:1:1:1:1:1 for eight datasets; and the highest accuracy improvement of 0.056 compared to using prediction accuracy as a weight for phenotype prediction. Compared to the best results of single-omics prediction, the multi-view GBLUP prediction method showed the highest accuracy improvement of 0.049.

[0109] Experiment 10. Comparison of the accuracy of different methods in predicting tomato phenotypes using genomic, insertion / deletion, and structural variation data.

[0110] In this experiment, we compared the accuracy of predicting tomato phenotypes using genomic, insertion / deletion, and structural variation data on the Tomato332 dataset.

[0111] like Figure 16As shown, there are significant differences in the accuracy of predicting tomato phenotypes using genomic data alone, insertion / deletion data alone, and structural variation data alone, with the highest difference in single-trait prediction accuracy being 0.019. When using both genomic and insertion / deletion data, and when using both genomic and structural variation data, the multi-view GBLUP prediction method achieves the highest accuracy in predicting the soluble solids content trait in tomatoes. Compared to predicting phenotypes by assigning 1:1 weights to the two datasets, the accuracy improvement is up to 0.006; compared to using prediction accuracy as a weight for phenotype prediction, the accuracy improvement is up to 0.009. When using both insertion / deletion and structural variation data, using the two datasets by assigning 1:1 weights achieves the highest accuracy in predicting the soluble solids content trait in tomatoes. The prediction accuracy of the multi-view GBLUP method is close to the highest prediction method, differing by only 0.007. The multi-view GBLUP prediction method, which integrates three types of data, achieved the highest accuracy in predicting the soluble solids content trait of tomatoes. Compared to weighting the three datasets 1:1:1 to predict phenotypes, the accuracy was improved by up to 0.018; compared to using prediction accuracy as a weight to predict phenotypes, the accuracy was improved by up to 0.022. Compared to the best results of single-omics prediction, the multi-view GBLUP prediction method improved accuracy by up to 0.03.

[0112] Experiment 11. Iterative process of training a model using multi-view GBLLUP method with genomic and transcriptomic data from 7 different tissues: changes in phenotypic prediction accuracy and optimal weights.

[0113] In this experiment, we examined the changes in the accuracy and optimal weights of maize flowering day phenotype prediction during model training using genomic and transcriptomic data from seven different tissues on the Maize282 dataset.

[0114] like Figure 17 As shown, the initial NP individuals have 20 weight combinations, and the prediction accuracy varies significantly among each individual, with a maximum difference of 0.084. As the number of iterations increases, the similarity among the NP individuals and between generations of the population increases, leading to a stable accuracy in predicting the corn flowering day phenotype. In the 43rd generation, the absolute difference in the maximum prediction accuracy between two adjacent iterations of the training set is less than 0.0001, at which point the algorithm terminates, and the prediction accuracy for the corn flowering day phenotype is 0.89.

[0115] like Figure 18 As shown, the optimal individuals that make the prediction accuracy of the maize flowering day phenotype are selected in each generation of the population. As the number of iterations increases, the optimal weights represented by the optimal individuals tend to stabilize around the 43rd iteration. The optimal weight combinations of genomic and 7 different tissue transcriptome data are 0.178, 0, 0, 1, 0.001, 0, 0.896 and 0.801.

[0116] In summary, compared to single-omics prediction, the multi-view GBLUP prediction method may improve the prediction accuracy of traits; compared to other weighting methods, the multi-view GBLUP prediction method may improve the prediction accuracy of multi-omics; when the prediction accuracy of the multi-view GBLUP prediction method is not the highest, it is close to the highest phenotypic prediction accuracy. The more types of datasets and the greater the difference in performance between datasets, the better the multi-view GBLUP prediction method performs.

[0117] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0118] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A multi-view GBLUP method for integrating multiple types of data for phenotypic prediction, characterized in that, The method includes the following steps: Step 1: Obtain SNP genotype data as the training set. Represent the SNP genotype data as multi-type data in the form of 0 / 1 / 2 according to different phenotypic types. Write the multi-type data into a matrix, where the rows of the matrix represent materials and the columns of the matrix represent features. Perform Z-score standardization on each column of data. Step 2, let the i-th row and j-th row of the multi-type data matrix be individual x. i ,x j Calculate the kinship matrix K and define the similarity function between two individuals; Step 3: Based on the similarity between individuals, the multi-type data is weighted and integrated into a kinship matrix K. The differential evolution algorithm is used, and its specific process is as follows: Initialize the population, each individual in the population is a weighted combination of a set of multi-type data, perform mutation and crossover operations on the population, compare the individuals in the population with the individuals after mutation and crossover, assign weights to the kinship matrix K of the multi-type data, substitute it into the GBLUP model for learning, and select individuals with high phenotypic prediction accuracy as members of the next generation population. Step 4: Set a termination condition. Stop learning when the absolute difference between the maximum prediction accuracy of two adjacent training sets is less than a certain threshold, and output the optimal weight combination for multiple types of data. Step 5: Set up the GBLUP model according to the optimal weight combination, input the SNP genotype data to be predicted, and output the predicted phenotype.

2. The multi-view GBLUP method for integrating multiple types of data for phenotypic prediction according to claim 1, characterized in that, In step 1, SNP genotype data are represented as 0 / 1 / 2 according to different phenotypic types, where 0 represents a genotype with a high homozygous frequency, 1 represents a heterozygous genotype, and 2 represents a genotype with a low homozygous frequency.

3. The multi-view GBLUP method for integrating multiple types of data for phenotypic prediction according to claim 1, characterized in that, In step 1, the standardized processing formula is: Where X, E(X), and Var(X) represent the vector to be standardized, the expected value of X, and the variance of X, respectively.

4. The multi-view GBLUP method for integrating multiple types of data for phenotypic prediction according to claim 1, characterized in that, In step 2, the similarity function between two individuals is defined as: Among them, (x i ,x j ) represents individual x i ,x j The inner product of.

5. The multi-view GBLUP method for integrating multiple types of data for phenotypic prediction according to claim 1, characterized in that, In step 3, the specific steps of the differential evolution algorithm are as follows: Step 3.1, Initialization; Initialize NP population individuals, each individual being a weighted combination of multiple data types, with each initial weight between (0,1); Step 3.2, Mutation; For each individual in the initial population, a mutated individual is generated using a weighted average relaxation technique. The mutation calculation process is as follows: Among them, V i 0 It is a mutated individual. These are two randomly selected different individuals, ||·||2 represents the norm of the vector, and ω is a relaxation parameter that controls the degree of variation; Step 3.3, crossover; For each individual in the initial population, crossover is performed between the mutated individual and the original individual to generate experimental individuals; Step 3.4, Selection: Compare the individuals in the initial population with the experimental individuals after crossover, assign weights to the Kinship matrix of the multi-type data, substitute it into the GBLUP model for learning, and select individuals with higher phenotypic prediction accuracy as members of the next generation population; the selection method is as follows: in, These are crossover test individuals. This is the original population of this generation. These are the individuals from this generation that are retained as members of the next generation of the population. f(X) represents the fitness function. This method uses the prediction accuracy of the training set's five-fold cross-validation as the function value.

6. The multi-view GBLUP method for integrating multiple types of data for phenotypic prediction according to claim 1, characterized in that, In step 4, the termination condition is set as follows: Where f(X) represents the fitness function, and ∈ represents a given small positive number, which is taken as 0.0001 in this method.

7. The multi-view GBLUP method for integrating multiple types of data for phenotypic prediction according to claim 5, characterized in that, In step 3.4, the weights are assigned to the kinship matrix K of the multi-type data, and the specific method for substituting it into the GBLUP model for learning is as follows: The GBLUP model is as follows: y = Xβ + Zu + ε, where y is an n×1 phenotypic vector, X is an n×p design matrix corresponding to fixed effects, β is a p×1 coefficient vector representing fixed effects, and Z is an n×n design matrix corresponding to random effects. (Multigene effects) K is the kinship matrix and the residual effect vector. I is the identity matrix. and These are the estimated genetic variance and residual variance, respectively; Let A i For n×p i Given a multi-type data matrix (i = 1, 2, ..., m), define the multi-type data kinship matrix as K = w1A1A1 T +w2A1A1 T +…+w m A m A m T The weight vector W = (w1, w2, ..., w m ) T It is learned by the differential evolution algorithm.

Images