Universal SNP (Single Nucleotide Polymorphism) molecular marker for predicting multiple phenotypic characters of corn as well as screening method and prediction system thereof
By screening and integrating multiple models, the problem of screening universal SNP molecular markers for multiple agronomic traits in maize breeding was solved, achieving efficient and accurate prediction of multiple traits, reducing costs and shortening the breeding cycle.
Patent Information
- Application Number
- CN202610102856.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-26
- Publication Date
- 2026-05-05
AI Technical Summary
Existing technologies struggle to efficiently screen universal SNP molecular markers for multiple agronomic traits in maize breeding, resulting in high costs and resource consumption. Furthermore, existing models are computationally complex, prone to overfitting, and difficult to achieve comprehensive prediction of multiple traits.
By using a combination of mutual information screening, LD filtering, LASSO regression, and random forest, universal SNP molecular markers suitable for multiple phenotypic traits in maize were screened out. An integrated prediction system combining self-attention neural network, LASSO regression, and random forest models was constructed, and the model weights were dynamically adjusted to achieve accurate prediction of multiple traits.
It significantly reduced the cost of genotype collection, improved breeding efficiency, enabled the simultaneous prediction of multiple maize phenotypic traits, shortened the breeding cycle, and achieved a dimensionality reduction rate of 96.2%.
Smart Images

Figure CN121983116A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of molecular biology technology, specifically relating to universal SNP molecular markers for predicting various phenotypic traits in maize, as well as their screening methods and prediction systems. Background Technology
[0002] As the most widely planted and highest-yielding grain crop in my country, maize is crucial to the country's food security. Currently, maize breeding in my country is in a transitional stage from traditional phenotypic selection to bio-breeding. For qualitative traits controlled by major genes, developing molecular markers for assisted breeding selection is the most effective method. However, for quantitative traits controlled by minor genes, marker-assisted breeding selection is not very effective, and the agronomic traits that need to be improved in breeding are mostly quantitative traits.
[0003] Genomic selection (GS) is a novel breeding method that utilizes high-density markers covering the entire genome for selection. It can accelerate genetic progress by shortening generation intervals through early selection and improving the accuracy of Genomic Estimated Breeding Value (GEBV) estimation. It is particularly effective in predicting complex traits with low heritability and difficulty in measurement, truly enabling genomic technology to guide breeding practices. However, the markers for each trait in GS are not uniform; each trait requiring improvement has its own markers that need to be measured. This means that traits other than the target trait still require phenotypic screening to compensate, increasing the difficulty of data integration. If each trait requires marker measurement for prediction, the cost doubles. Furthermore, current techniques for trait prediction typically use single models, such as Stacking and Bayesian Model Averaging (BMA). However, Stacking and BMA are computationally complex, resource-intensive, have high hardware requirements, and are prone to overfitting. Therefore, screening for universal markers that can be used for screening multiple agronomic traits, optimizing prediction systems, and comprehensively, economically, and efficiently solving the integrated prediction of multiple agronomic traits are of particular importance in this field. Summary of the Invention
[0004] The purpose of this invention is to provide a universal SNP molecular marker for predicting multiple phenotypic traits in maize, along with its screening method and prediction system. This allows for the simultaneous prediction of multiple maize phenotypic traits, enabling early selection of maize germplasm, shortening generation intervals, and reducing genotype collection costs and breeding cycles.
[0005] This invention provides a method for screening SNP molecular markers associated with multiple phenotypic traits in maize, comprising the following steps: (1) Obtain maize genotype and phenotypic data from the database; the phenotypic data includes maize stalk rot resistance, yield, grain moisture content, plant height, ear height, relative ear position, number of stalks, silking period, flowering period and grain bulk density; (2) Perform quality control on the genomic data, remove SNP markers without genetic variation, and convert the original character-type SNP base data into a 4-dimensional vector to obtain the initial features; (3) Calculate the mutual information between each initial feature and the phenotypic data, retain the features with mutual information ≥ 0.025, and obtain the first dimensionality reduction feature; (4) Perform LD filtering on the first dimensionality reduction feature to obtain the second dimensionality reduction feature; (5) Based on LASSO regression, the second dimensionality reduction feature is screened to obtain the features that have a linear impact on the phenotypic data; (6) Based on the random forest algorithm, the second dimensionality reduction feature is screened to obtain nonlinear features that contribute greatly to the phenotypic data; (7) Take the union of the features that have a linear effect on the phenotypic data and the nonlinear features that contribute greatly to the phenotypic data to obtain the third dimension reduction feature and obtain the candidate SNP molecular markers; Steps (5) and (6) are not in any particular order.
[0006] Preferably, after obtaining the genotype and phenotype data of maize from the database in step (1), the method further includes: ensuring that the genotype and phenotype correspond one-to-one through the unique sample ID, and removing maize samples with missing IDs or incomplete phenotype data. Step (2) implements quality control of genomic data using `nunique() == 1` in Pandas; Step (2) involves converting the original character-type SNP base data into a 4-dimensional vector by: establishing a base probability mapping table and mapping each base type to a 4-dimensional vector; Step (3) involves calculating the mutual information between each initial feature and the phenotypic data, including calculating the average mutual information between each initial feature and the phenotypic data. Step (4) LD filtering includes: LD r of two or more SNP sites 2 When the mutual information is greater than 0.56, only the SNP site with the highest mutual information is retained.
[0007] The present invention also provides universal SNP molecular markers screened by the screening method described in the above technical solution.
[0008] Preferably, the site information of the SNP molecular marker is shown in Table 2 of the specification.
[0009] This invention also provides a method for constructing a prediction model for multiple phenotypic traits of maize, comprising the following steps: The base data of SNP molecular markers in maize samples were converted into 4-dimensional vectors to obtain vector data; The vector data and the various phenotypic trait data of the corn sample are input into different training models, and the weight ratio of the different training models is dynamically adjusted to obtain the prediction model of various phenotypic traits of corn. The data on various phenotypic traits of maize include resistance to maize stalk rot, yield, kernel moisture content, plant height, ear height, relative ear position, number of standing stalks, silking period, flowering period, and kernel bulk density. The different training models include a self-attention neural network, a LASSO regression model, and a random forest model; the sum of the weights of the self-attention neural network, the LASSO regression model, and the random forest model is 1.
[0010] Preferably, the SNP molecular marker is the SNP molecular marker described in the above technical solution.
[0011] Preferably, the weight ratio of the self-attention neural network in the prediction system is 0.3082; the weight ratio of the LASSO regression model in the prediction system is 0.3501; and the weight ratio of the random forest model in the prediction system is 0.3417.
[0012] This invention also provides a maize multiphenotypic trait prediction system, including an SNP molecular marker screening module and a multi-model integration module; the SNP molecular marker screening module includes a data acquisition module, a data processing module, a mutual information screening module, an LD filtering module, a LASSO and random forest fusion screening module; The data acquisition module collects maize genotype and phenotypic data; The data processing module performs quality control on the maize genome data and obtains vector features; The mutual information filtering module performs a first dimensionality reduction on the vector features to obtain the first dimensionality-reduced features; The LD filtering module performs a second dimensionality reduction on the first dimensionality reduction feature to obtain a second dimensionality reduction feature. The fusion of LASSO and random forest performs a third dimensionality reduction on the vector features of the second dimensionality reduction feature screening module to obtain the screened SNP molecular markers. The multi-model integration module constructs a prediction model for multiple phenotypic traits of maize according to the construction method described in the above technical solution.
[0013] This invention also provides products that detect the SNP molecular markers described in the above technical solutions, or maize multi-phenotypic trait prediction models constructed by the construction methods described in the above technical solutions, or maize multi-phenotypic trait prediction systems described in the above technical solutions, for the application in predicting maize phenotypic traits and / or maize-assisted breeding; the multi-phenotypic traits include maize stalk rot resistance, yield, grain moisture content, plant height, ear height, relative ear position, number of stalks, silking period, flowering period, and grain bulk density.
[0014] This invention also provides a method for predicting multiple phenotypic traits of maize, which involves converting the base data of SNP molecular markers of maize samples into 4-dimensional vectors, inputting the vector data into a prediction model, and obtaining the results of multiple phenotypic traits of maize; the prediction model is a prediction model of multiple phenotypic traits of maize constructed by the construction method described in the above technical solution. The various phenotypic traits include maize stalk rot resistance, yield, grain moisture content, plant height, ear height, relative ear position, number of standing stalks, silking period, flowering period, and grain bulk density.
[0015] Beneficial effects: This invention provides a method for screening universal SNP molecular markers for predicting multiple phenotypic traits in maize, and the universal SNP molecular markers screened therefrom. This invention transforms genotype data into vectors, and then performs three dimensionality reduction processes on the transformed genotype data: mutual information screening, LD filtering, LASSO, and random forest fusion screening. Invalid and redundant markers are eliminated at each level, resulting in the development of universal markers suitable for predicting multiple phenotypic traits in maize. Compared with traditional genome-wide selection (GS) feature screening, this method reduces costs while increasing the breadth of screenable traits, achieving multi-trait aggregation screening and improving breeding efficiency. The results of the examples show that this invention screened 7645 universal SNP markers, compared to 200642 SNP molecular markers obtained by traditional genome-wide selection (GS), achieving a dimensionality reduction rate of 96.2%, and enabling simultaneous prediction of multiple maize phenotypic traits.
[0016] Furthermore, this invention transforms the base data of SNP molecular markers in maize samples into 4-dimensional vectors to obtain vector data. This vector data, along with various phenotypic trait data of the maize samples, is input into different training models. The weight ratios of these different training models are dynamically adjusted to obtain prediction results for various maize phenotypic traits. These traits include resistance to maize stalk rot, yield, grain moisture content, plant height, ear height, relative ear position, number of stalks, silking period, flowering period, and grain bulk density. The different training models include a self-attention neural network, a LASSO regression model, and a random forest model. The sum of the weights of the self-attention neural network, LASSO regression model, and random forest model is 1. This invention constructs an integrated model of self-attention neural network, LASSO, and random forest. By complementing the advantages of different models and optimizing the weights, the advantages of the three models are combined to overcome the limitations of a single model, ultimately achieving accurate and efficient prediction of multiple maize phenotypic traits and shortening the breeding cycle. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the embodiments will be briefly described below.
[0018] Figure 1 This is a base probability mapping table; Figure 2 This is a histogram of mutual information distribution; where the horizontal axis represents the average mutual information score, the vertical axis represents the number of features, the features to the left of the threshold line are the features that are removed, and the features to the right of the threshold line are the features that are retained. Detailed Implementation
[0019] This invention provides a method for screening universal SNP molecular markers to predict multiple phenotypic traits in maize, comprising the following steps: (1) Obtain maize genotype and phenotypic data from the database; the phenotypic data includes maize stalk rot resistance, yield, grain moisture content, plant height, ear height, relative ear position, number of stalks, silking period, flowering period and grain bulk density; (2) Perform quality control on the genomic data, remove SNP markers without genetic variation, and convert the original character-type SNP base data into a 4-dimensional vector to obtain the initial features; (3) Calculate the mutual information between each initial feature and the phenotypic data, retain the features with mutual information ≥ 0.025, and obtain the first dimensionality reduction feature; (4) Perform LD filtering on the first dimensionality reduction feature to obtain the second dimensionality reduction feature; (5) Based on LASSO regression, the second dimensionality reduction feature is screened to obtain the features that have a linear impact on the phenotypic data; (6) Based on the random forest algorithm, the second dimensionality reduction feature is screened to obtain nonlinear features that contribute greatly to the phenotypic data; (7) Take the union of the features that have a linear effect on the phenotypic data and the nonlinear features that contribute greatly to the phenotypic data to obtain the third dimension reduction feature and obtain the candidate SNP molecular markers; Steps (5) and (6) are not in any particular order.
[0020] This invention retrieves genotype and phenotypic data of maize from a database. As one implementation method, after retrieving the genotype and phenotypic data of maize from the database, this invention ensures a one-to-one correspondence between genotype and phenotype using a unique sample ID, and removes maize samples with missing IDs or incomplete phenotypic data.
[0021] After obtaining the genotype and phenotypic data of maize, this invention performs quality control on the genome data, removes SNP markers without genetic variation, and transforms the original character-type SNP base data into a 4-dimensional vector to obtain initial features.
[0022] As one implementation method, this invention achieves quality control of genomic data using `nunique() == 1` in Pandas. SNPs without genetic variation cannot distinguish genetic differences between samples in a population. Although SNPs without genetic variation may account for a small percentage of the data, they can be detrimental and must be removed. For example, they cannot provide any information about genetic differences and are considered "invalid features": different samples with different bases at a certain SNP locus may correspond to different phenotypes. Furthermore, SNPs without variation have identical bases in all samples and cannot be used to distinguish genetic differences between any two samples. In machine learning, the feature values of SNPs without genetic variation are identical in all samples, and these features have a variance of 0. Whether for linear models (such as LASSO) or tree models (such as random forests), features with zero variance will neither participate in regression calculations (the coefficient will always be 0) nor be used for node splitting (it cannot reduce impurity), but they will still participate in training, thus reducing training efficiency. Pandas is one of the most commonly used data analysis libraries in Python. The nunique() method in Pandas can be used to calculate the number of distinct values in a column (i.e., the number of unique values). If X_raw[col].nunique()==1 (retrieves the col column from the X_raw dataset and determines the number of distinct bases in this column), it means that all samples in the SNP column have the same bases (only 1 type of base), which is an SNP without genetic variation.
[0023] As one implementation method, the present invention transforms raw character-type SNP base data into 4-dimensional vectors by establishing a base probability mapping table, mapping each base type to a 4-dimensional vector. The core of genotype vectorization is base probability encoding, which essentially transforms raw character-type SNP base data (A, C, G, T, etc.) into 4-dimensional vectors. The reason for vectorizing genotypes is that machine learning models only accept numerical input, but numerical encoding (0, 1, 2, 3, etc.) loses probabilistic information. Therefore, base probability encoding can completely preserve the genetic variation information of SNPs while simultaneously converting them into numerical values.
[0024] After obtaining the initial features, the present invention calculates the mutual information between each initial feature and the phenotypic data, retains features with mutual information ≥ 0.025, and obtains the first dimensionality reduction features.
[0025] As one implementation method, the calculation of mutual information between each initial feature and the phenotypic data in this invention includes calculating the average mutual information between each initial feature and the phenotypic data. Mutual information (MI) measures the strength of the nonlinear association between a feature and a phenotype; a higher value indicates a greater contribution of the marker to the prediction of the phenotype. For example, if the mutual information value between a genotype feature and a phenotype is higher, it means that the feature contains more "information that can explain phenotypic variation," and the association between the two is stronger. If the mutual information value is close to 0, it means that the feature is almost unrelated to the phenotype and is a redundant feature. The core of mutual information screening is to retain features with mutual information ≥ a threshold. Based on previous experiments, this invention found that a threshold of 0.02 leads to the retention of some features that contribute less to model training, while 0.03 leads to overly strict screening, easily filtering out features that are useful to the model. This invention sets a threshold of 0.025 (when the calculated mutual information is greater than or equal to this threshold, it is considered that there is an association between the feature and the phenotype), which has a better screening effect. It can capture both linear associations with the phenotype and nonlinear associations between features and the phenotype, and can significantly reduce the number of features.
[0026] After obtaining the first dimensionality reduction feature, the present invention performs LD filtering on the first dimensionality reduction feature to obtain the second dimensionality reduction feature.
[0027] As one implementation method, the LD filtering of the present invention includes: LD r at two or more SNP sites. 2When the mutual information score is >0.56, only the SNP locus with the highest mutual information is retained. LD (Linkage Disequilibrium) is a core indicator describing the degree of association between alleles of different SNP loci; changes in the alleles of two SNPs are correlated. After mutual information filtering, the code retains SNP features associated with the phenotype, but these features may still contain a large number of redundant features with high LD. For example, for two highly correlated SNPs, only one needs to be retained; retaining both would double the computation time. The fundamental purpose of LD filtering in this invention is to "remove redundant features," regardless of whether this redundancy is positive or negative correlation redundancy. Retaining features with higher mutual information scores ensures that redundant and lower-value features are removed. r (Pearson correlation coefficient, used to measure the linear correlation between two variables X and Y) measures the linear correlation between the encoded values of two SNPs. The value range of r is [-1, 1], r>0 indicates a positive correlation; when the value of one feature increases, the value of the other feature also increases. r < 0 indicates a negative correlation; when the value of one feature increases, the value of the other feature decreases. This application limits the filtering to |r| > 0.75 (i.e., r² > 0.56). If the threshold is too low (e.g., 0.5), too many weakly correlated features will be removed, potentially losing useful information; if the threshold is too high (e.g., 0.9), redundancy removal will be incomplete, and a large number of highly redundant features will still be retained.
[0028] As one implementation method, when performing the LD filtering, the first dimensionality reduction feature is divided into blocks. This block division is to free up memory and prevent memory overflow due to excessive memory usage.
[0029] After obtaining the second dimensionality reduction feature, this invention uses both LASSO regression and random forest algorithms to filter the second dimensionality reduction feature, obtaining features with a linear impact on phenotypic data and nonlinear features with a large contribution to phenotypic data. The union of these features yields the third dimensionality reduction feature, resulting in candidate SNP molecular markers. The relationship between genotype and phenotype is not simply linear or nonlinear; fusing the two models avoids the limitations of a single model and retains more comprehensive and effective features. LASSO is a linear regression with L1 regularization. Essentially, while "fitting the linear relationship between phenotype and features," it forces the coefficients of some irrelevant features to be compressed to 0 by "penalizing the sum of the absolute values of the coefficients," thus achieving feature selection. That is, the model prioritizes removing features with little impact on phenotype, compressing their coefficients to 0, and retaining only features with a significant linear impact on phenotype. The coefficient is the core indicator for measuring the "strength of the linear association between features and phenotype." A coefficient > 0 indicates a positive correlation between the feature value and the phenotype; a coefficient < 0 indicates a negative correlation; and a coefficient equal to 0 indicates no significant linear association and is considered a redundant feature. The advantage of random forests is that they "do not require assumptions about relationship types, can automatically capture nonlinear relationships and feature interaction effects (such as two SNPs jointly influencing the phenotype), and quantify the contribution of features to the phenotype through 'feature importance,' ultimately selecting nonlinear features that contribute significantly to the phenotype." The core structure of random forests, "decision tree ensemble," allows a single decision tree to capture nonlinearity, while ensembles of multiple trees can reduce overfitting. As one implementation method, this invention calculates for each decision tree "how much the MSE of the phenotype prediction is reduced after splitting using this feature"—the greater the reduction, the higher the importance of the feature in that tree; the final feature importance is obtained by averaging the "feature importance" across all trees.
[0030] The SNP molecular marker screening module of this invention is used to screen SNP molecular markers associated with multiple phenotypic traits in maize. Through mutual information screening, LD filtering, LASSO and random forest fusion screening, the SNP molecular markers are optimized, significantly reducing the number of SNP molecular markers. This solves the problem of traditional genome-wide selection (GS) having a large number of SNP molecular markers and easily missing key markers with both linear and non-linear associations. It can simultaneously predict 10 maize phenotypic traits and significantly reduce the cost of genotype collection.
[0031] This invention provides universal SNP molecular markers screened by the screening method described in the above technical solution.
[0032] As one implementation method, the site information of the universal SNP molecular marker described in this invention is shown in Table 2 of the specification.
[0033] This invention also provides a method for constructing a prediction model for multiple phenotypic traits of maize, comprising the following steps: The base data of SNP molecular markers in maize samples were converted into 4-dimensional vectors to obtain vector data; The vector data and the various phenotypic trait data of the corn sample are input into different training models, and the weight ratio of the different training models is dynamically adjusted to obtain the prediction model of various phenotypic traits of corn. The data on various phenotypic traits of maize include resistance to maize stalk rot, yield, kernel moisture content, plant height, ear height, relative ear position, number of standing stalks, silking period, flowering period, and kernel bulk density. The different training models include a self-attention neural network, a LASSO regression model, and a random forest model; the sum of the weights of the self-attention neural network, the LASSO regression model, and the random forest model is 1.
[0034] In one implementation, the SNP molecular markers described in this invention are the SNP molecular markers described in the above-mentioned technical solutions. In another implementation, the self-attention neural network of this invention has a weight ratio of 0.3082 in the prediction system; the LASSO regression model of this invention has a weight ratio of 0.3501 in the prediction system; and the random forest model of this invention has a weight ratio of 0.3417 in the prediction system. This invention constructs a prediction system combining self-attention neural network, LASSO, and random forest. By complementing the advantages of different models and optimizing the weights to integrate the strengths of the three models, it overcomes the limitations of a single model, ultimately achieving accurate and efficient prediction of multiple phenotypic traits in maize and shortening the breeding cycle.
[0035] This invention also provides a maize multi-phenotypic trait prediction system, including an SNP molecular marker screening module and a multi-model integration module. The SNP molecular marker screening module includes a data acquisition module, a data processing module, a mutual information screening module, an LD filtering module, and a LASSO and random forest fusion screening module. The data acquisition module collects maize genotype data and phenotypic data. The data processing module performs quality control on the maize genome data to obtain vector features. The mutual information screening module performs a first dimensionality reduction on the vector features to obtain a first dimensionality-reduced feature. The LD filtering module performs a second dimensionality reduction on the first dimensionality-reduced feature to obtain a second dimensionality-reduced feature. The LASSO and random forest fusion screening module performs a third dimensionality reduction on the vector features to obtain the screened SNP molecular markers. The multi-model integration module constructs a maize multi-phenotypic trait prediction model according to the construction method described above.
[0036] This invention also provides products that detect the SNP molecular markers described in the above technical solutions, or maize multi-phenotypic trait prediction models constructed by the construction methods described in the above technical solutions, or maize multi-phenotypic trait prediction systems described in the above technical solutions, for the application in predicting maize phenotypic traits and / or maize-assisted breeding; the multi-phenotypic traits include maize stalk rot resistance, yield, grain moisture content, plant height, ear height, relative ear position, number of stalks, silking period, flowering period, and grain bulk density.
[0037] This invention also provides a method for predicting multiple phenotypic traits of maize. The method involves converting the base data of SNP molecular markers in maize samples into a 4-dimensional vector, inputting the vector data into a prediction model, and obtaining the results for multiple phenotypic traits of maize. The prediction model is a prediction model for multiple phenotypic traits of maize constructed using the method described in the above technical solution. The multiple phenotypic traits include maize stalk rot resistance, yield, grain moisture content, plant height, ear height, relative ear position, number of stalks, silking period, flowering period, and grain bulk density.
[0038] To further illustrate the present invention, the following detailed description, in conjunction with the accompanying drawings and embodiments, describes the universal SNP molecular markers for predicting multiple phenotypic traits in maize, their screening methods, and the prediction system provided by the present invention. However, these descriptions should not be construed as limiting the scope of protection of the present invention.
[0039] Example 1 1. Data Source Genotypic data: Each sample contains 200,642 SNP (single nucleotide polymorphism) markers.
[0040] Phenotypic data: Each sample contains 10 phenotypes: "FSR_CMB": resistance to maize stalk rot, "GYG_CMB": yield, "MOI_CMB": kernel moisture content, "PH_CMB": plant height, "EH_CMB": ear height, "rEPH_CMB": relative ear position, "nP_CMB": number of stalks, "SD_CMB": silking period, "AD_CMB": flowering period, "TWT-CMB": kernel bulk density.
[0041] By using unique sample IDs, we ensured that genotypes and phenotypes corresponded one-to-one. Samples with missing IDs or incomplete phenotype data were removed, and a total of 296 samples were retained (236 in the training set and 60 in the test set).
[0042] 2. Quality control of genotype data By controlling the genotype data, SNP markers without genetic variation are accurately removed. Specifically, the `nunique()` method in Pandas is used to calculate the "number of distinct values in a column" (i.e., the "number of unique values"). If `X_raw[col].nunique() == 1`, it means that all samples in the SNP column have the same base (only 1 type of base), i.e., "no variation SNP".
[0043] 3. Base probability coding Establish a base probability mapping table based on the meaning of bases (Table 1). Figure 1 The original character-based SNP base data was transformed into a 4-dimensional vector, allowing the non-numerical genetic information to be adapted for subsequent machine learning analysis. After quality control and vector encoding, the number of features was 796,868.
[0044] Table 1. Meaning of each base
[0045] 4. Mutual information filtering (first dimensionality reduction) A threshold of 0.025 is set using mi_threshold=0.025 (when the calculated mutual information is greater than or equal to this threshold, the feature and phenotype are considered to be correlated). The average mutual information between each feature and the 10 phenotypes is calculated in step 3. Features with a higher correlation to the phenotype are retained, and a mutual information distribution histogram is generated, as shown below. Figure 2 After mutual information filtering, the number of features decreased from 796,868 to 26,476, significantly reducing the subsequent feature processing time.
[0046] 5. LD Filtering (Second Dimensionality Reduction) The linear correlation between two SNP encoded values is measured by r (Pearson correlation coefficient, used to measure the degree of linear correlation between two variables X and Y). A value of |r| > 0.75 (i.e., r < 0.75) is used. 2 Use >0.56 to filter and remove highly redundant features, retaining the one with the higher mutual information score among two or more high LD SNP features.
[0047] 6. LASSO + Random Forest Fusion Filtering (Third Dimensionality Reduction) (1) LASSO screening: LASSO is used to screen features. While fitting the linear relationship between phenotype and features, the coefficients of some irrelevant features are forced to be compressed to 0 by the sum of the absolute values of the penalty coefficients, thereby achieving feature screening and retaining only features that have a significant linear impact on phenotype.
[0048] (2) Random forest screening: Use random forest to screen features. For each decision tree, calculate "how much the MSE of phenotypic prediction is reduced after splitting with this feature" - the greater the reduction, the higher the importance of the feature in this tree; take the average of the "importance of this feature" of all trees to obtain the final feature importance.
[0049] (3) The final features are obtained by taking the union of the linear features selected by LASSO and the nonlinear features selected by random forest. Finally, 7645 universal SNP molecular markers are selected from the original 200642 SNP markers. The dimensionality reduction ratio is calculated as: (200642-7645) / 200642≈96.2%. The 7645 universal SNP molecular markers are shown in Table 2 below.
[0050] Table 2 7645 universal SNP molecular markers .
[0051] 7. Multi-model ensemble training using self-attention neural networks + LASSO + random forests Using the test set and based on the SNP molecular markers selected above, different training models were constructed. The prediction results of 10 maize phenotypes ("FSR_CMB": resistance to maize stalk rot, "GYG_CMB": yield, "MOI_CMB": kernel moisture content, "PH_CMB": plant height, "EH_CMB": ear height, "rEPH_CMB": relative ear position, "nP_CMB": number of stalks, "SD_CMB": silking date, "AD_CMB": flowering date, "TWT-CMB": kernel bulk density) were incorporated into a self-attention neural network and a LASSO regression model. In the LASSO and Random Forest models, the weights of different training models are dynamically adjusted. Initially, the weights of all three models account for 1 / 3 of the prediction results on the final test set. The importance of the average prediction weight to the prediction accuracy on the test set is calculated. The weights of models with high importance are increased according to their importance ratio, while the weights of models with low importance are decreased according to their importance ratio. This process is repeated until the weights of the three models no longer change in the prediction results on the test set. Finally, the weight ratio of the self-attention neural network in the prediction system is 0.3082; the weight ratio of the LASSO regression model in the prediction system is 0.3501; and the weight ratio of the random forest model in the prediction system is 0.3417. The prediction results for various phenotypic traits of maize are obtained, as shown in Table 3.
[0052] Table 3. Results of multi-model ensemble training
[0053] Note: The prediction accuracy for flowering period, ear height, relative ear position, and grain bulk density is less than 0.4, which is related to their relatively low heritability. Therefore, it is believed that this method is effective in predicting these traits.
[0054] As can be seen from the above, the technical solution provided by this invention can achieve accurate and efficient prediction of multiple maize phenotypic traits, shorten the breeding cycle, and the prediction effect is better than that of a single self-attention neural network, LASSO regression model or random forest model.
[0055] Although the above embodiments have provided a detailed description of the present invention, they are only some embodiments of the present invention, and not all embodiments. People can obtain other embodiments based on these embodiments without creative effort, and these embodiments all fall within the protection scope of the present invention.
Claims
1. A method for screening universal SNP molecular markers to predict multiple phenotypic traits in maize, characterized in that, Includes the following steps: (1) Obtain maize genotype and phenotypic data from the database; the phenotypic data includes maize stalk rot resistance, yield, grain moisture content, plant height, ear height, relative ear position, number of stalks, silking period, flowering period and grain bulk density; (2) Perform quality control on the genomic data, remove SNP markers without genetic variation, and convert the original character-type SNP base data into a 4-dimensional vector to obtain the initial features; (3) Calculate the mutual information between each initial feature and the phenotypic data, retain the features with mutual information ≥ 0.025, and obtain the first dimensionality reduction feature; (4) Perform LD filtering on the first dimensionality reduction feature to obtain the second dimensionality reduction feature; (5) Based on LASSO regression, the second dimensionality reduction feature is screened to obtain the features that have a linear impact on the phenotypic data; (6) Based on the random forest algorithm, the second dimensionality reduction feature is screened to obtain nonlinear features that contribute greatly to the phenotypic data; (7) Take the union of the features that have a linear effect on the phenotypic data and the nonlinear features that contribute greatly to the phenotypic data to obtain the third dimension reduction feature and obtain the candidate SNP molecular markers; Steps (5) and (6) are not in any particular order.
2. The screening method according to claim 1, characterized in that, After obtaining the genotype and phenotypic data of maize from the database in step (1), the following steps are also included: ensuring that the genotype and phenotype correspond one-to-one through the unique sample ID, and removing maize samples with missing IDs or incomplete phenotypic data. Step (2) implements quality control of genomic data using `nunique() == 1` in Pandas; Step (2) involves converting the original character-type SNP base data into a 4-dimensional vector by: establishing a base probability mapping table and mapping each base type to a 4-dimensional vector; Step (3) involves calculating the mutual information between each initial feature and the phenotypic data, including calculating the average mutual information between each initial feature and the phenotypic data. Step (4) LD filtering includes: LD r of two or more SNP sites 2 When the mutual information is greater than 0.56, only the SNP site with the highest mutual information is retained.
3. The universal SNP molecular markers screened by the screening method according to claim 1 or 2.
4. The universal SNP molecular marker according to claim 3, characterized in that, The site information of the SNP molecular markers is shown in Table 2 of the specification.
5. A method for constructing a prediction model for multiple phenotypic traits of maize, characterized in that, Includes the following steps: The base data of SNP molecular markers in maize samples were converted into 4-dimensional vectors to obtain vector data; The vector data and the various phenotypic trait data of the corn sample are input into different training models, and the weight ratio of the different training models is dynamically adjusted to obtain the prediction model of various phenotypic traits of corn. The data on various phenotypic traits of maize include resistance to maize stalk rot, yield, kernel moisture content, plant height, ear height, relative ear position, number of standing stalks, silking period, flowering period, and kernel bulk density. The different training models include a self-attention neural network, a LASSO regression model, and a random forest model; the sum of the weights of the self-attention neural network, the LASSO regression model, and the random forest model is 1.
6. The construction method according to claim 5, characterized in that, The SNP molecular marker is the SNP molecular marker described in claim 3 or 4.
7. The construction method according to claim 5 or 6, characterized in that, The self-attention neural network has a weight of 0.3082 in the prediction system; the LASSO regression model has a weight of 0.3501 in the prediction system; and the random forest model has a weight of 0.3417 in the prediction system.
8. A system for predicting multiple phenotypic traits in maize, characterized in that, It includes an SNP molecular marker screening module and a multi-model integration module; the SNP molecular marker screening module includes a data acquisition module, a data processing module, a mutual information screening module, an LD filtering module, a LASSO and random forest fusion screening module; The data acquisition module collects maize genotype and phenotypic data; The data processing module performs quality control on the maize genome data and obtains vector features; The mutual information filtering module performs a first dimensionality reduction on the vector features to obtain the first dimensionality-reduced features; The LD filtering module performs a second dimensionality reduction on the first dimensionality reduction feature to obtain a second dimensionality reduction feature. The fusion of LASSO and random forest performs a third dimensionality reduction on the vector features of the second dimensionality reduction feature screening module to obtain the screened SNP molecular markers. The multi-model integration module constructs a prediction model for multiple phenotypic traits of maize according to the construction method described in any one of claims 5 to 7.
9. The application of the product containing the SNP molecular marker as described in claim 3 or 4, or the maize multi-phenotypic trait prediction model constructed by the construction method described in any one of claims 5 to 7, or the maize multi-phenotypic trait prediction system as described in claim 8, in predicting maize phenotypic traits and / or maize-assisted breeding; wherein the multi-phenotypic traits include maize stalk rot resistance, yield, grain moisture content, plant height, ear height, relative ear position, number of stalks, silking period, flowering period, and grain bulk density.
10. A method for predicting multiple phenotypic traits of maize, characterized in that, The base data of SNP molecular markers in maize samples are converted into 4-dimensional vectors. After obtaining the vector data, they are input into a prediction model to obtain results of multiple phenotypic traits of maize. The prediction model is a prediction model of multiple phenotypic traits of maize constructed by the construction method described in any one of claims 5 to 7. The various phenotypic traits include maize stalk rot resistance, yield, grain moisture content, plant height, ear height, relative ear position, number of standing stalks, silking period, flowering period, and grain bulk density.
