A method for analyzing the effects of corn selective breeding by integrating population genetics and deep learning

Abstract

The present invention discloses a method for analyzing the selective breeding effects of corn by integrating population genetics and deep learning. The method collects corn germplasm resource data, including phenotypic data, genotypic data and environmental data; standardizes the phenotypic data and fills in missing values, and calculates population genetic parameters, including allele frequency, linkage disequilibrium and population structure stability; uses a deep learning model to perform association analysis between the population genetic parameters and the phenotypic data, uses the trained model to predict the phenotypic values ​​of candidate parents, calculates the breeding index in combination with the population genetic parameters, and provides guidance for corn breeding decisions based on the prediction results. The present invention achieves accurate prediction and evaluation of the selective breeding effects of corn by organically combining deep learning technology with the principles of population genetics, thereby improving breeding efficiency and providing a new technical path for corn variety improvement.

Description

Technical Field

[0001] The present invention belongs to the technical field of breeding analysis, and specifically relates to a corn selection breeding effect analysis method integrating population genetics and deep learning. Background Art

[0002] Traditional corn breeding relies primarily on the experience of breeding experts and phenotypic selection, which is inefficient and time-consuming. With the development of genome sequencing technology, high-throughput sequencing platforms can quickly obtain information on a large number of SNP loci, providing a rich genetic resource for molecular marker-assisted selection.

[0003] However, most existing corn breeding analysis methods only focus on the direct relationship between genotype and phenotype, ignoring the impact of population genetic parameters on breeding results, resulting in a decrease in genetic diversity during the selection process and limited long-term breeding results; existing prediction models are difficult to effectively deal with the complex nonlinear interactions between genotype, phenotype and environment, especially the insufficient accuracy of genotype performance prediction under different environmental conditions; there is a lack of systematic methods integrating population genetics principles with modern deep learning technology, and it is impossible to fully utilize genetic laws to guide model design and breeding decisions; in addition, existing breeding effect evaluation methods are mostly based on a single phenotypic indicator, lack of comprehensive consideration of multi-dimensional breeding goals, and it is difficult to meet the needs of modern corn breeding for the coordinated improvement of multiple traits.

[0004] Therefore, there is an urgent need for a new corn selection breeding effect analysis method that can integrate population genetic parameters and deep learning technology to improve the efficiency and accuracy of corn breeding. Summary of the Invention

[0005] The present invention provides a corn selective breeding effect analysis method that integrates population genetics and deep learning. By integrating population genetic parameters with deep learning technology, the efficiency and accuracy of corn breeding are improved.

[0006] In order to achieve the above-mentioned purpose of the invention, the specific technical solutions are as follows:

[0007] A method for analyzing the effect of corn selective breeding by integrating population genetics and deep learning, the method comprising the following steps:

[0008] Step S1, collecting corn germplasm resource data, the germplasm resource data includes: phenotypic data, genotypic data and environmental data.

[0009] The phenotypic data are multidimensional characteristics consisting of yield, resistance, and quality; the genotypic data are genotypic data of at least 50,000 SNP sites for each sample obtained by a high-throughput sequencing platform; and the environmental data include corn growth environment factors including temperature, rainfall, and light.

[0010] Step S2, standardizing and filling missing values ​​on the phenotypic data, and calculating population genetic parameters, which include allele frequency, linkage disequilibrium, and population structure stability.

[0011] Step S3: Use a deep learning model to perform correlation analysis between population genetic parameters and phenotypic data, predict selective breeding effects based on the trained model, and provide guidance for corn breeding decisions based on the prediction results.

[0012] Furthermore, the method for calculating the population genetic parameters is:

[0013] Calculate allele frequencies: ,in, Indicates the The frequency of the alleles, Indicates the The count of alleles in the population, Indicates the total number of individuals.

[0014] Calculate linkage disequilibrium coefficient : ;in, Indicates allele and The common frequency of and Alleles and The frequency of occurrence.

[0015] The stability of population structure is quantitatively expressed according to the inter-population differentiation coefficient, which is The calculation formula is: ,in, represents the total heterozygosity, represents the average heterozygosity of the subpopulations.

[0016] Furthermore, the structure of the deep learning model includes: input layer, convolution layer, attention layer, environment integration layer and fully connected layer.

[0017] The input layer receives the genotype matrix , receiving environment data matrix , receiving population genetic parameters ,in represents the number of samples, represents the number of marker sites, Indicates the number of environmental factors, Indicates the number of subpopulations in a population.

[0018] Furthermore, the convolutional layer extracts local genomic features through a one-dimensional convolution operation: ;in, represents the output of the convolutional layer, represents the activation function, represents the convolution kernel weight, represents the convolution operation, represents the bias term;

[0019] Furthermore, the attention layer is used to calculate the importance weights of different SNP marker sites, which are expressed as:

[0020] ;in, represents the attention weight vector, and is the attention layer parameter matrix, Represents the hidden layer features;

[0021] Furthermore, the environment integration layer integrates the environment data Incorporating into the model: ,in, Represents the environmental characteristics after processing, and Represent the weight matrix and bias term respectively;

[0022] Integrate various features: ,in, represents the fused feature vector, represents the output features of the convolutional layer, represents the group structure matrix, The weight matrix representing the group structure characteristics, represents the weight matrix of the environment features in the splicing operation, Represents feature concatenation operation;

[0023] Furthermore, the fully connected layer integrates features and maps them to the output space, which is expressed as: ;in, represents the predicted output of the phenotypic value, represents the weight matrix of the fully connected layer, represents the bias term.

[0024] Furthermore, the loss function used in model training is:

[0025] ;in, represents the total loss function, represents the mean squared error loss, which is used to measure the difference between the predicted value and the actual phenotypic value; Represents regularization loss, which is used to prevent the model from overfitting; Represents genetic constraint loss, designed based on population genetics principles to ensure that the model prediction results conform to genetic laws; represents the environmental response loss, which is used to capture the interaction effect between phenotype and environmental factors; 、 and is a hyperparameter that trades off different loss terms.

[0026] Furthermore, the genetic constraint loss The calculation formula is:

[0027] ;in, Represents linkage disequilibrium constraint loss, which is used to capture the interaction relationship between marker loci; represents the allele frequency constraint loss, used to maintain genetic diversity during selection; and The parameters for adjusting the importance of the two genetic constraints were taken as 0.45 and 0.55 respectively.

[0028] Furthermore, the environmental response loss The calculation formula is:

[0029] ;in, Representation sample In environmental factors The weight on Representation sample The actual phenotypic value of Considering genotype and environmental factors The predicted phenotypic value.

[0030] Furthermore, weighing the hyperparameters of different loss terms 、 and The value ranges are: ; ; .

[0031] Furthermore, guidance for corn breeding decisions based on the prediction results includes:

[0032] Calculate the breeding index BVI of each candidate parent using the following formula:

[0033] ;in, represents the number of phenotypic characteristics, Indicates the The weight coefficient of each phenotypic characteristic, Indicates the The predicted value of a phenotypic characteristic, represents the average allele frequency of the candidate parents, represents the absolute value of the average linkage disequilibrium coefficient of the candidate parental loci, Represents the average inter-population differentiation coefficient of candidate parents; sorted from high to low based on the breeding index BVI.

[0034] Furthermore, a preset number of candidate parents are selected from high to low according to BVI as the next generation breeding parents, and the preset number does not exceed 30% of the total number of candidate parents.

[0035] Furthermore, by comparing the differences in phenotypic prediction values ​​between the current breeding cycle and the previous breeding cycle, the selective breeding effect is quantified, and the hyperparameters of the deep learning model are dynamically adjusted according to the evaluation results to optimize the selection efficiency of subsequent breeding cycles.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] By integrating population genetic parameters with deep learning technology, the present invention constructs a prediction model that can simultaneously consider genotype, environmental factors and population genetic characteristics, significantly improving the accuracy of phenotypic prediction, with the prediction accuracy increased by about 15-20% compared with traditional methods; the method of the present invention not only improves the prediction accuracy and selection efficiency of corn breeding, but also effectively maintains the genetic diversity of the population, providing a scientific basis and technical support for the continuous genetic improvement of corn. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a flow chart of a corn selection breeding effect analysis method that integrates population genetics and deep learning according to the present invention. DETAILED DESCRIPTION

[0039] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention are described clearly and completely below. Obviously, the embodiments described are only part of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0040] like Figure 1 As shown in the figure, a flow chart of a method for analyzing the effect of corn selective breeding by integrating population genetics and deep learning of the present invention is provided. The method comprises the following steps:

[0041] Step S1, collecting corn germplasm resource data, the germplasm resource data includes: phenotypic data, genotypic data and environmental data.

[0042] The phenotypic data are multidimensional characteristics consisting of yield, resistance, and quality; the genotypic data are genotypic data of at least 50,000 SNP sites for each sample obtained by a high-throughput sequencing platform; and the environmental data include corn growth environment factors including temperature, rainfall, and light.

[0043] Phenotypic data collection follows a standard field trial design. Yield data are quantified as yield per hectare (tons / hectare). Resistance data include a disease resistance index (scaled from 0 to 9, with 0 being immune and 9 being extremely susceptible), an insect pest resistance score (scaled from 0 to 5, with 0 being harmless and 5 being severe), and stress tolerance (drought, salinity, and low temperature tolerance). Quality data are quantified as multi-dimensional indicators such as protein content (%), starch content (%), oil content (%), and amino acid composition. Genotypic data are acquired using high-throughput sequencing platforms such as Illumina or Affymetrix. Each sample is analyzed for at least 50,000 single nucleotide polymorphism (SNP) sites, covering all 10 chromosomes of the maize genome. Environmental data are recorded at the experimental site's meteorological station during the growing season, including parameters such as daily average temperature (°C), diurnal temperature range (°C), accumulated rainfall (mm), daylight hours (hours), and light intensity (lux).

[0044] The calculation method of the population genetic parameters is:

[0045] Calculate allele frequencies: ,in, Indicates the The frequency of the alleles, Indicates the The count of alleles in the population, Indicates the total number of individuals.

[0046] Allele frequency calculations first require quality control of SNP loci, including excluding loci with missingness rates exceeding 5%, minor allele frequencies (MAF) below 0.05, and those significantly deviating from Hardy-Weinberg equilibrium (p < 0.000001). For diallelic systems, the 0 / 1 / 2 coding is often used to represent the genotype of an individual at a locus (0 represents homozygous aa, 1 represents heterozygous Aa, and 2 represents homozygous AA). The frequency of allele A is calculated by counting the number of occurrences of each coding. For outcrossing crops like corn, particular attention must be paid to the counting method for heterozygous loci, with each heterozygous locus contributing 0.5 counts to each of the two alleles.

[0047] Calculate linkage disequilibrium coefficient : ;in, Indicates allele and The common frequency of and Alleles and The frequency of occurrence.

[0048] The calculation of linkage disequilibrium coefficient D needs to take into account the influence of physical distance between sites. In practice, the pairwise linkage disequilibrium coefficient of adjacent SNP sites within a 1Mb sliding window is usually calculated, and the linkage disequilibrium decay diagram is drawn to evaluate the linkage disequilibrium level of the population. In addition to the basic formula, the standardized linkage disequilibrium coefficient needs to be calculated. ,in When D<0 , when D>0 .

[0049] In addition, in order to reduce the computational complexity of multi-site analysis, the tag SNP strategy is adopted, that is, representative sites are selected for analysis in highly linked SNP blocks, which can significantly reduce the computational complexity while retaining the main genetic information.

[0050] The stability of population structure is quantitatively expressed according to the inter-population differentiation coefficient, which is The calculation formula is: ,in, represents the total heterozygosity, represents the average heterozygosity of the subpopulations.

[0051] The calculation of the interpopulation differentiation coefficient FST involves the estimation of the total heterozygosity HT and the average heterozygosity HS of the subpopulations. The calculation formula of HT is , represents the expected heterozygosity under random mating conditions; HS is the weighted average of the heterozygosity of each subpopulation, and the calculation formula is , where ni is the number of individuals in the ith subpopulation, N is the total number of individuals, and pi is the frequency of the allele in the ith subpopulation.

[0052] The statistical significance of FST also needs to be verified by AMOVA (analysis of molecular variance). Usually, 999 permutation tests are used to determine the p-value. p<0.05 indicates significant differentiation between subpopulations; FST values ​​between 0.05-0.15 indicate moderate differentiation, which is suitable as an indicator for hybrid parent selection.

[0053] Step S2, standardizing and filling missing values ​​on the phenotypic data, and calculating population genetic parameters, which include allele frequency, linkage disequilibrium, and population structure stability.

[0054] Phenotypic data were normalized using the z-score normalization method to convert phenotypic characteristics of different dimensions into a standard distribution with a mean of 0 and a standard deviation of 1, ensuring that each characteristic had the same weight in the model.

[0055] Missing values ​​were filled using multiple imputation combined with the K-nearest neighbor (KNN) algorithm, and the phenotypic values ​​of individuals with similar genotypes were estimated, with a filling accuracy of over 95%.

[0056] In the calculation of population genetic parameters, allele frequencies are obtained by direct counting method; in addition to the formula calculation, linkage disequilibrium coefficient D is also obtained by value( ) were normalized to eliminate the influence of allele frequency differences on linkage disequilibrium estimation; the optimal number of subpopulations (K value) was determined by STRUCTURE software based on the Bayesian clustering method for population structure stability, and the rationality of subpopulation division was verified by principal component analysis (PCA).

[0057] Step S3: Use a deep learning model to perform correlation analysis between population genetic parameters and phenotypic data, predict selective breeding effects based on the trained model, and provide guidance for corn breeding decisions based on the prediction results.

[0058] The deep learning model was implemented using the PyTorch framework and trained on an NVIDIA Tesla V100 GPU. During the association analysis, the model first learned the direct relationship between phenotype and genotype, then introduced environmental factors as moderating variables to analyze the interaction between genotype and environment. Finally, population genetic parameters were integrated as prior knowledge to guide the model learning process.

[0059] The prediction of selective breeding effects uses five-fold cross-validation, and model performance is evaluated by calculating the Pearson correlation coefficient (r) and root mean square error (RMSE) between the predicted and actual phenotypic values. Practice has shown that the proposed model achieves an accuracy (r-value) exceeding 0.85 in corn yield prediction, which is 12-18 percentage points higher than the traditional GBLUP method, providing reliable data support for corn breeding decisions.

[0060] The deep learning model structure consists of an input layer, a convolutional layer, an attention layer, an environmental integration layer, and a fully connected layer. The deep learning model's implementation utilizes a modular design to facilitate the replacement and optimization of different functional units. When the input layer receives data, the genotype matrix X is encoded using 0 / 1 / 2 and normalized. The environmental data matrix E contains environmental factor values ​​for key growth stages (seedling emergence, jointing, heading, grain filling, and maturity). The population genetic parameters Q are derived from the membership matrix obtained through STRUCTURE analysis. Residual connections are used in each layer to enhance deep network training stability, and batch normalization layers are introduced to accelerate model convergence. Furthermore, to prevent overfitting, a dropout layer (with a dropout rate of 0.3) is added before the fully connected layer, and L2 regularization is applied to the parameter matrix (with a weight decay coefficient of 0.0001).

[0061] The input layer receives the genotype matrix , receiving environment data matrix , receiving population genetic parameters ,in represents the number of samples, represents the number of marker sites, Indicates the number of environmental factors, Indicates the number of subpopulations in a population.

[0062] The convolutional layer extracts local genomic features through a one-dimensional convolution operation: ;in, represents the output of the convolutional layer, represents the activation function, represents the convolution kernel weight, represents the convolution operation, Represents the bias term; the one-dimensional convolution operation uses a multi-scale convolution kernel strategy, including convolution kernels of length 3, 9, and 27, to capture short-range, medium-range, and long-range genomic dependencies, respectively. The number of convolution kernels of each scale is set to 128, the stride is set to 1, and the padding mode is "same" to maintain the size of the feature map. The activation function σ uses LeakyReLU with parameter α = 0.2. Compared with the traditional ReLU, it allows negative inputs to pass a small amount of gradient, avoiding the problem of neuron death. The convolution operation is followed by a max pooling layer with a pooling window size of 3 and a stride of 2, which reduces the size of the feature map to half, reducing the number of parameters while retaining significant features.

[0063] The attention layer is used to calculate the importance weights of different SNP marker sites, and its expression is:

[0064] ;in, represents the attention weight vector, and is the attention layer parameter matrix, Represents the hidden layer features; softmax normalization is used to ensure that the sum of the attention weights of all sites is 1. In the calculation, Dimension The projection matrix, Dimension The transformation matrix of is the hidden layer feature dimension (set to 256), is the key vector dimension (set to 64), and H represents the hidden features extracted from the convolutional layer output. To improve computational efficiency, a multi-head attention mechanism is introduced with 8 heads, each of which independently learns the importance relationship between different locations.

[0065] The environment integration layer integrates environmental data Incorporating into the model: ,in, Represents the environmental characteristics after processing, and Representing the weight matrix and bias term, respectively. The environmental integration layer employs different encoding strategies for different types of environmental factors: continuous environmental variables (such as temperature and rainfall) are normalized to the range [-1, 1] using min-max normalization; categorical environmental variables (such as soil type) are converted to binary vectors using one-hot encoding. The activation function σ uses the Tanh function to maintain environmental features within the range [-1, 1], facilitating subsequent feature fusion. The weight matrix W_e uses the Xavier initialization method to ensure signal variance stability during forward propagation. To capture interactions between environmental factors, a self-attention mechanism is added after the standard linear transformation, enabling the model to learn the mutual influences between different environmental factors.

[0066] Integrate various features: ,in, represents the fused feature vector, represents the output features of the convolutional layer, represents the group structure matrix, The weight matrix representing the group structure characteristics, represents the weight matrix of the environment features in the splicing operation, Denotes a feature concatenation operation; during feature integration, the concatenation operation [;] ensures that information from features from different sources is preserved. To balance the contributions of various features, weight matrices W_q and W_e' are set to adjust feature importance. Initial values ​​are set to the identity matrix and optimized during training. Before concatenation, the feature vectors are dimensionally aligned: the convolutional layer output feature C is reduced to a fixed length (256 dimensions) using global average pooling; the population structure matrix Q is linearly transformed to the same dimensional space; and the environmental features E' are similarly converted to the corresponding dimensions. After feature fusion, feature selection techniques such as L1 regularization or principal component analysis (PCA) are applied to reduce dimensionality, preserving over 90% of the information while significantly reducing computational complexity.

[0067] The fully connected layer integrates features and maps them to the output space, and its expression is: ;in, represents the predicted output of the phenotypic value, represents the weight matrix of the fully connected layer, Represents the bias term. The fully connected layers adopt a pyramidal design, reducing the number of neurons layer by layer, from the fused feature dimension (typically 512 or 1024) to 256, 128, and 64, before finally mapping to the output dimension (equivalent to the number of predicted phenotypic features). Each fully connected layer is followed by a batch normalization layer and a ReLU activation function. The final layer does not use an activation function to output continuous-valued predictions. For the joint prediction of multiple phenotypes, a multi-task learning framework is adopted, sharing the underlying feature extraction network while providing independent output heads for each phenotype. This leverages the correlation between phenotypes while retaining the specificity requirements of each phenotype.

[0068] The loss function used in model training is:

[0069] ;in, represents the total loss function, represents the mean squared error loss, which is used to measure the difference between the predicted value and the actual phenotypic value; Represents regularization loss, which is used to prevent the model from overfitting; Represents genetic constraint loss, designed based on population genetics principles to ensure that the model prediction results conform to genetic laws; represents the environmental response loss, which is used to capture the interaction effect between phenotype and environmental factors; 、 and is a hyperparameter that trades off different loss terms.

[0070] The loss of genetic constraint The calculation formula is:

[0071] ;in, Represents linkage disequilibrium constraint loss, which is used to capture the interaction relationship between marker loci; represents the allele frequency constraint loss, used to maintain genetic diversity during selection; and The parameters for adjusting the importance of the two genetic constraints were taken as 0.45 and 0.55 respectively.

[0072] The environmental response loss The calculation formula is:

[0073] ;in, Representation sample In environmental factors The weight on Representation sample The actual phenotypic value of Considering genotype and environmental factors The predicted phenotypic value.

[0074] Hyperparameters for weighing different loss terms 、 and The value ranges are: ; ; . The value is usually set large (about 1.5) at the beginning of training and gradually decreases to 1.0 as training progresses, achieving a smooth transition from simple models to complex models. Setting it to a larger value (usually 2.0) emphasizes the importance of genetic constraints and ensures that the prediction results conform to Mendelian inheritance and the basic principles of population genetics. The loss function is dynamically adjusted based on the importance of environmental factors, taking a higher value (0.5) in experiments with large environmental variations and a lower value (0.1) when environmental conditions are stable. The Adam optimizer is used for loss function optimization, with an initial learning rate of 0.001 and a cosine annealing learning rate scheduling strategy.

[0075] Providing guidance for corn breeding decisions based on prediction results includes:

[0076] Calculate the breeding index BVI of each candidate parent using the following formula:

[0077] ;in, represents the number of phenotypic characteristics, Indicates the The weight coefficient of each phenotypic characteristic, Indicates the The predicted value of a phenotypic characteristic, represents the average allele frequency of the candidate parents, represents the absolute value of the average linkage disequilibrium coefficient of the candidate parental loci, Represents the average inter-population differentiation coefficient of candidate parents; sorted from high to low based on the breeding index BVI.

[0078] The phenotypic weight coefficient w_j is set according to the breeding goal. For example, the high-yield oriented strategy can set the yield weight to 0.6, the resistance weight to 0.3, and the quality weight to 0.1; the resistance oriented strategy can adjust the weights to 0.3, 0.6, and 0.1 respectively. The calculations were performed after z-score standardization to ensure comparability of different dimension characteristics. The population genetic parameter adjustment term [·] integrates the three genetic parameters into a single constraint factor: ( ) encourages the selection of rare alleles and increases the potential for genetic gain; The item considers the degree of linkage disequilibrium and balances the short-term selection response with the long-term genetic progress; ( ) item promotes gene exchange between subpopulations and maintains genetic diversity by reducing the differentiation coefficient.

[0079] A preset number of candidate parents are selected from high to low according to the BVI as the next generation breeding parents. This preset number shall not exceed 30% of the total number of candidate parents. Excessive selection intensity (too low a selection ratio) will lead to a sharp decline in genetic diversity and reduce long-term selection benefits; too low a selection intensity will slow the rate of genetic gain. The 30% upper limit can effectively control the risk of genetic drift while maintaining sufficient selection pressure. In actual implementation, the specific ratio is dynamically adjusted according to the population size: for large germplasm resource banks (>1000 individuals), a selection ratio of 10-15% can be used; for medium-sized populations (200-1000 individuals), a selection ratio of 15-25% is recommended; and for small breeding populations (<200 individuals), a selection ratio close to 30% should be adopted to maintain a sufficient effective population size.

[0080] By comparing the differences in phenotypic prediction values ​​between the current breeding cycle and the previous breeding cycle, the selective breeding effect is quantified, and the hyperparameters of the deep learning model are dynamically adjusted based on the evaluation results to optimize the selection efficiency of subsequent breeding cycles.

[0081] The formula for calculating the phenotypic difference value (PD) is: , represents the difference between the mean of the predicted phenotypes in the current cycle and the previous cycle; the selection response rate (SR) is calculated as , where S is the selection difference, that is, the difference between the average phenotype of the selected parents and the average phenotype of the population; the rate of change of genetic diversity ( ) is calculated as , where GD represents the population genetic diversity estimated based on molecular markers; achieved heritability ( ) is calculated as , where R is the actual selection response, which measures the selection efficiency. Based on these evaluation indicators, the hyperparameters of the deep learning model are dynamically adjusted: if SR is lower than expected, increase Strengthen genetic constraints; if Decrease too fast, increase Weight ; If the environmental stability is poor, appropriately increase Enhance the ability of environmental response models. Through this closed-loop feedback mechanism, selection strategies and prediction models are continuously optimized to maximize long-term breeding results.

[0082] We selected a breeding library of 2,500 maize inbred lines and collected phenotypic data under 10 environmental conditions across eight different ecological zones. The data included yield (5.2-12.8 tons / hectare), disease resistance (primarily resistance to large leaf spot and small leaf spot, scored 1-9), and quality traits (protein content 8.5%-13.2%, starch content 65.3%-73.6%). Genotype data for each accession were obtained using the Illumina MaizeSNP50K array at 59,642 SNPs. After quality control, 43,865 valid SNP markers were retained.

[0083] The deep learning model constructed using this method achieved prediction accuracy (measured by Pearson correlation coefficient r) of 0.87 for yield, 0.83 for disease resistance, and 0.85 for quality traits, significantly exceeding the traditional GBLUP (yield r = 0.71) and MLR (yield r = 0.68) methods. The top 20% (500 accessions) of material were selected as candidate parents based on BVI scores, and 100 hybrid combinations were constructed. Field validation showed that these new combinations increased yield by an average of 8.6% and disease resistance by 15.3%, while maintaining high genetic diversity (the Shannon index decreased by only 2.1%), validating the effectiveness and reliability of this method in actual breeding.

[0084] Through the application of three consecutive breeding cycles, the breeding time for new corn varieties has been shortened from the traditional 8-10 years to 5-6 years, greatly improving breeding efficiency.

[0085] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for analyzing the effect of corn selective breeding by integrating population genetics and deep learning, characterized in that: The method comprises the following steps: Step S1, collecting corn germplasm resource data, the germplasm resource data including: phenotypic data, genotypic data and environmental data; The phenotypic data are multidimensional characteristics consisting of yield, resistance, and quality; the genotypic data are genotypic data of at least 50,000 SNP sites for each sample obtained by a high-throughput sequencing platform; the environmental data include corn growth environment factors such as temperature, rainfall, and light; Step S2, standardizing and filling missing values ​​in the phenotypic data, and calculating population genetic parameters, which include allele frequency, linkage disequilibrium, and population structure stability; Step S3, using a deep learning model to perform correlation analysis between population genetic parameters and phenotypic data, predicting selective breeding effects based on the trained model, and providing guidance for corn breeding decisions based on the prediction results; The structure of the deep learning model includes: input layer, convolution layer, attention layer, environment integration layer and fully connected layer; The input layer receives the genotype matrix , receiving environment data matrix , receiving population genetic parameters ,in represents the number of samples, represents the number of marker sites, Indicates the number of environmental factors, represents the number of subpopulations in a population; The convolutional layer extracts local genomic features through a one-dimensional convolution operation: ;in, represents the output of the convolutional layer, represents the activation function, represents the convolution kernel weight, represents the convolution operation, represents the bias term; The attention layer is used to calculate the importance weights of different SNP marker sites, and its expression is: ;in, represents the attention weight vector, and is the attention layer parameter matrix, Represents the hidden layer features; The environment integration layer integrates environmental data Incorporating into the model: ,in, Represents the environmental characteristics after processing, and Represent the weight matrix and bias term respectively; Integrate various features: ,in, represents the fused feature vector, represents the output features of the convolutional layer, represents the group structure matrix, The weight matrix representing the group structure characteristics, represents the weight matrix of the environment features in the splicing operation, Represents feature concatenation operation; The fully connected layer integrates features and maps them to the output space, and its expression is: ;in, represents the predicted output of the phenotypic value, represents the weight matrix of the fully connected layer, represents the bias term.

2. The corn selective breeding effect analysis method integrating population genetics and deep learning according to claim 1, characterized in that: The calculation method of the population genetic parameters is: Calculate allele frequencies: ,in, Indicates the The frequency of the alleles, Indicates the The count of alleles in the population, Indicates the total number of individuals; Calculate linkage disequilibrium coefficient : ;in, Indicates allele and The common frequency of and Alleles and Frequency of occurrence; The stability of population structure is quantitatively expressed according to the inter-population differentiation coefficient, which is The calculation formula is: ,in, represents the total heterozygosity, represents the average heterozygosity of the subpopulations.

3. The corn selective breeding effect analysis method integrating population genetics and deep learning according to claim 2, characterized in that: The loss function used in model training is: ;in, represents the total loss function, represents the mean squared error loss, which is used to measure the difference between the predicted value and the actual phenotypic value; Represents regularization loss, which is used to prevent the model from overfitting; Represents genetic constraint loss, designed based on population genetics principles to ensure that the model prediction results conform to genetic laws; represents the environmental response loss, which is used to capture the interaction effect between phenotype and environmental factors; 、 and is a hyperparameter that trades off different loss terms.

4. The corn selective breeding effect analysis method integrating population genetics and deep learning according to claim 3, characterized in that: The loss of genetic constraint The calculation formula is: ;in, Represents linkage disequilibrium constraint loss, which is used to capture the interaction relationship between marker loci; represents the allele frequency constraint loss, used to maintain genetic diversity during selection; and The parameters for adjusting the importance of the two genetic constraints were taken as 0.45 and 0.55 respectively.

5. The corn selective breeding effect analysis method integrating population genetics and deep learning according to claim 4, characterized in that: The environmental response loss The calculation formula is: ;in, Representation sample In environmental factors The weight on Representation sample The actual phenotypic value of Considering genotype and environmental factors The predicted phenotypic value.

6. The corn selective breeding effect analysis method integrating population genetics and deep learning according to claim 5, characterized in that: Hyperparameters for weighing different loss terms 、 and The value ranges are: ; ; .

7. The corn selective breeding effect analysis method integrating population genetics and deep learning according to claim 5, characterized in that: Providing guidance for corn breeding decisions based on prediction results includes: Calculate the breeding index BVI of each candidate parent using the following formula: ;in, represents the number of phenotypic characteristics, Indicates the The weight coefficient of each phenotypic characteristic, Indicates the The predicted value of a phenotypic characteristic, represents the average allele frequency of the candidate parents, represents the absolute value of the average linkage disequilibrium coefficient of the candidate parental loci, Represents the average inter-population differentiation coefficient of candidate parents; sorted from high to low based on the breeding index BVI.

8. The corn selective breeding effect analysis method integrating population genetics and deep learning according to claim 7, characterized in that: A preset number of candidate parents are selected from high to low according to BVI as the breeding parents for the next generation, and the preset number does not exceed 30% of the total number of candidate parents.

9. The corn selective breeding effect analysis method integrating population genetics and deep learning according to any one of claims 1 to 8, characterized in that: By comparing the differences in phenotypic prediction values ​​between the current breeding cycle and the previous breeding cycle, the selective breeding effect is quantified, and the hyperparameters of the deep learning model are dynamically adjusted based on the evaluation results to optimize the selection efficiency of subsequent breeding cycles.

Images