Pig genome prediction method based on multi-omics annotation information and residual segmentation strategy

Abstract

The invention discloses a pig genome prediction method based on multi-omics annotation information and a residual segmentation strategy. On the basis of multi-omics annotation information, maximum utilization of effective information is guaranteed through a hierarchical feature selection strategy, meanwhile, data dimensions are reduced, a residual error segmentation strategy is adopted, linear hybrid model residual errors serve as phenotype input, a deep learning algorithm is fitted, and a deep learning algorithm is established. A pig genome prediction method based on multi-omics annotation information and a residual segmentation strategy is constructed, the method can automatically perform decoupling genetic effect, feature dimension reduction, residual learning and multi-model prediction, and phenotype prediction of pig complex characters is realized. Compared with the prior art, the method has the advantages that the biological characteristics of the genome are more comprehensively considered, the calculation efficiency is improved, linear and nonlinear segmented modeling and additive and interaction effect segmented modeling are realized, the stability of the model is improved, and the accuracy and the calculation efficiency are both improved in a genome prediction deep learning model; and a reliable tool is provided for pig breeding.

Description

Technical Field

[0001] The present invention relates to the fields of animal genome breeding and artificial intelligence technology, and more specifically, to a pig genome prediction method based on multi-omics annotation information and residual segmentation strategy. Background Art

[0002] The innovative development of breeding technology is a core driving force for improving animal quality and production, as well as innovation in the seed industry. As a revolutionary new-generation breeding technology, genomic breeding has become the most effective and widely used method for selecting swine breeders in my country. Current genomic prediction methods primarily rely on linear mixed models based on additive effects. However, these models assume that all markers contribute equally to the target trait, ignoring the complex nonlinear interactions between genes or loci. In recent years, artificial intelligence (AI) models have rapidly emerged. Machine learning and deep learning nonlinear models can effectively fit interaction networks between data and address the complex relationships between genes and phenotypes. Linear and nonlinear models each have their own strengths and weaknesses, and a single model cannot achieve optimal prediction performance for all traits. Therefore, a multi-model strategy is a better option. By leveraging the complementary strengths of linear and nonlinear models in genomic prediction, we can achieve a final prediction model that accurately estimates additive genetic effects while also identifying complex non-additive effects and gene-gene interactions. This improves model generalization and provides a new approach to more stable and accurate genomic prediction for intelligent breeding.

[0003] The complexity and diversity of genomic information creates a high-dimensional data space. Faced with such vast data volumes, dimensionality reduction is essential when using nonlinear models for prediction. Functional annotation based on multi-omics data not only reveals the contribution of genetic loci to trait formation and further explores the genetic basis of complex traits, but also provides reliable group information during data dimensionality reduction, optimizing feature screening strategies and improving computational efficiency.

[0004] This invention constructs a new breeding model through the comprehensive application of annotation information integration, hierarchical feature selection, residual segmentation strategy, and nonlinear model, realizes the extensive capture of genetic effects, provides more comprehensive, accurate and stable prediction capabilities for genome prediction models, provides more reliable tools for animal and plant breeding and precision medicine research, and promotes the development of the field of genomics. Summary of the Invention

[0005] The technical problems to be solved by the present invention include the following aspects. The traditional linear model assumes that all markers have an impact on the target trait, each marker has the same genetic contribution, and ignores the complex nonlinear interactions between genes or sites. The advantage of nonlinear models in automatically learning from data and mining complex patterns makes them very suitable for processing genomic information. However, a single model cannot perform optimally on all traits and it is difficult to fully capture complex genetic effects. At the same time, genomic data is huge, and nonlinear models are easily interfered by redundant information in high-dimensional feature space, resulting in decreased model performance and excessive consumption of computing resources. However, existing dimensionality reduction methods mainly rely on the genomic data itself, and often ignore the biological significance contained in multi-omics annotation information, which may lead to insufficient feature selection efficiency and insufficient information utilization.

[0006] In order to solve the above problems, the present invention provides the following technical solutions, a pig genome prediction method based on multi-omics annotation information and residual segmentation strategy, and the specific solutions are as follows.

[0007] A porcine genome prediction method based on multi-omics annotation information and residual segmentation strategy comprises the following steps: Step S1, collecting phenotypic data, SNP data and multi-omics annotation information of the reference group and candidate group pigs, dividing the reference group pigs into a training set and a validation set, and the candidate group pigs as a test set; Step S2, using the SNP data of the training set, validation set, and test set obtained in step S1 and the phenotypic data of the training set as input, using the residual partitioning strategy to obtain the genomic estimated breeding values, means, and phenotypic residuals of all pigs; Step S3, using the SNP data and multi-omics annotation information of the training set, validation set, and test set obtained in step S1 and the phenotypic residuals of the training set and validation set obtained in step S2, a hierarchical feature selection strategy is used to reduce the dimension of the SNP data of all pigs to obtain a SNP matrix after dimensionality reduction; Step S4, using the SNP matrix of the training set obtained in step S3 after dimensionality reduction as input, performing regression prediction on the phenotypic residuals of the training set obtained in step S2, and training multiple nonlinear models; Step S5, using the reduced-dimensional SNP matrix of the validation set obtained in step S3 as input, fitting multiple nonlinear models obtained in step S4, predicting the phenotypic residuals of the validation set, adding the predicted phenotypic residuals to the estimated breeding values and mean values of the genomics of the validation set obtained in step S2, obtaining the predicted phenotypic values of the validation set, calculating the correlation coefficient between the predicted phenotypic values and the phenotypic values of the validation set obtained in step S1, selecting the optimal nonlinear model, and fine-tuning the optimal nonlinear model; In step S6, the SNP matrix of the test set after dimensionality reduction obtained in step S3 is used as input, and the optimal nonlinear model fine-tuned in step S5 is fitted to obtain the phenotypic residual value of the test set, which is added to the genomic estimated breeding value and mean of the test set obtained in step S2, and the phenotypic prediction value of the test set is output.

[0008] Preferably, the group division in step S1 is achieved specifically in the following manner: 20% of the pigs in the reference group are randomly selected as a validation set, the remaining 80% are a training set, and all the pigs in the candidate group are used as a test set.

[0009] Preferably, the residual partitioning strategy is used in step S2 to obtain the estimated breeding values, means and phenotypic residuals of the genomes of all pigs, which specifically includes the following steps: Step S21, input the SNP data of the training set, validation set, and test set and the phenotypic data of the training set obtained in step S1, call the single-trait GBLUP model in the HIBLUP software, use the AIREML method to estimate the variance components, and use a mixed linear equation to estimate the genomic estimated breeding value and mean of all pigs; In step S22, the phenotypic values of all pigs obtained in step S1 are subtracted from the genomic estimated breeding values and mean values obtained in step S21 to obtain the phenotypic residuals of all pigs. The calculation formula is as follows: in, Phenotypic residual values, is the phenotypic value, is the estimated mean, Estimate breeding values for genomes.

[0010] Preferably, in step S3, a hierarchical feature selection strategy is adopted to obtain a SNP matrix after dimensionality reduction, which specifically includes the following steps: Step S31, searching the multi-omics annotation information obtained in step S1 according to the SNP name in the SNP data obtained in step S1, determining the functional annotation region of each SNP, merging SNPs in the same functional annotation region, and merging SNPs that do not belong to the same functional annotation region, and obtaining each type of SNP matrix in the training set, validation set, and test set respectively; Step S32, calculate the number of SNPs in each category. When the number of SNPs in each category is less than or equal to 100, no feature selection is performed. When the number of SNPs in each category is greater than 100, XGBoost feature selection is performed on each category of SNPs separately. Step S33, using the SNP matrix of each category of the training set obtained in step S31 as input and the phenotypic residual of the training set obtained in step S2 as the target variable to train the XGBoost model; Step S34: When training the XGBoost model, input the SNP matrix of each category of the validation set obtained in step S31, predict the phenotypic residual, calculate the correlation between the predicted phenotypic residual and the phenotypic residual of the validation set obtained in step S2, use the correlation as the criterion for early stopping of the model, and obtain the model with the highest correlation; Step S35: In the model with the highest correlation obtained in step S34, first retain the SNPs with a Gain value greater than 0 to obtain the dimensionality-reduced SNP matrix for each category of the training set, and then obtain the dimensionality-reduced SNP matrix for each category of the validation set and the test set based on the SNP name; Step S36, repeating steps S32 to S35 until feature selection is completed for each type of SNP matrix obtained in step S31, and merging the multiple dimensionality-reduced SNP matrices obtained in step S35 into one SNP matrix in the training set, validation set, and test set.

[0011] Preferably, the multiple nonlinear models in step S4 include RF, SVR, XGBoost, LightGBM, MLP, and CNN, wherein the hyperparameters of the RF, SVR, XGBoost, LightGBM, and MLP models are user-defined, and the structure of the CNN model includes: (1) an input layer that receives SNP data and represents it as a one-dimensional tensor; (2) a convolution block, wherein the first convolution layer contains 8 channels, the convolution kernel size is 6, and the stride is 2, followed by a ReLU layer, an average pooling layer, and a batch normalization layer, and the next two convolution layers both contain It contains 8 channels, convolution kernel size 6, stride 2, and passes through the average pooling layer and then the Flatten layer to flatten the data into a one-dimensional vector; (3) fully connected layer, through Dropout to prevent overfitting, then through the fully connected layer to map the features extracted by the convolution block to a 128-dimensional latent space, and use ReLU for nonlinear transformation; (4) output layer, mapping the 128-dimensional features to a single continuous value; (5) using mean square error as the loss function for optimization, through Adam training optimization, using the early stopping mechanism, and setting the training round and early stopping round as hyperparameters.

[0012] Preferably, the fine-tuning model in step S5 is specifically implemented by merging the phenotypic residuals of the training set and the validation set obtained in step S2, and merging the SNP matrix after dimensionality reduction of the training set and the validation set obtained in step S3, using the SNP matrix and the phenotypic residual as input and target variables respectively, training the optimal nonlinear model obtained in step S5, and using the fine-tuned optimal nonlinear model as output.

[0013] Compared with the prior art, the present invention has at least the following positive technical effects.

[0014] Existing genomic predictions are mainly based on linear mixed models based on additive effects. Traditional linear models often ignore non-additive effects between genes. The modeling paradigm of nonlinear models is closer to the essence of the interaction of complex factors in real biological systems, and can reveal nonlinear relationships that lead to phenotypic inconsistencies. The present invention designs a genomic prediction model based on CNN, which includes two convolutional layers. The convolutional layer has a local perception mechanism that can capture local interaction patterns between SNPs, helps to extract important patterns and structures in the genome, and shows unique advantages in the analysis of local genetic patterns. At the same time, the present invention trains a variety of nonlinear models in the model training stage, which can take into account a variety of genetic structures and improve the accuracy of phenotypic prediction.

[0015] Most existing porcine genome prediction methods use a single linear model or nonlinear model, which makes it difficult to fully capture complex genetic effects. Therefore, a multi-model strategy may be a better choice. The present invention develops the idea of combining multiple models from two aspects: step-by-step multi-model and integrated multi-model. In terms of step-by-step multi-model, the present invention adopts a residual partitioning strategy. First, a linear model is used to estimate individual breeding values and calculate phenotypic residuals. Then, the residuals are modeled using a nonlinear model to achieve prediction of phenotypic values. The linear model provides a more reliable estimate of additive genetic effects. This part of information is accurate and stable in most cases. The nonlinear model can focus more on capturing these complex nonlinear information, avoiding the noise influence brought by direct fitting of the phenotype, and enhancing the modeling ability of non-additive and environmental effects. After the residual partitioning process, the additive effects and residuals are modeled in stages, giving full play to the advantages of each model and improving the overall prediction accuracy. Compared with rrBLUP and the conventional one-step strategy, this strategy has improved by 0.87% and 2.14%, respectively. In terms of integrated multi-model, multiple nonlinear models are used to model the same target, and the optimal nonlinear model is obtained by comparison, achieving higher genomic prediction accuracy and stability. The present invention combines these two ideas, combining the residual segmentation strategy and integrating multiple models for prediction, achieving higher genome prediction accuracy and stability.

[0016] Directly using whole genome data in nonlinear models requires large computational complexity and high resource consumption, so the feature dimensionality reduction process is crucial. Most existing dimensionality reduction methods rely solely on genotype data itself, fail to fully integrate the biological functional characteristics contained in multi-omics annotation information, and ignore some key functional sites, resulting in insufficient efficiency and biological significance of feature selection. The present invention proposes a hierarchical feature selection strategy, which uses multi-omics annotation information to split the original high-dimensional genomic data into multiple biologically relevant subsets, performs feature selection on each subset through XGBoost, and then merges them into a feature subset. Compared with directly using unreduced data to build a model, this strategy shortens the model training time while keeping the prediction accuracy basically unchanged. The hierarchical feature selection strategy combines multi-omics annotation information with XGBoost feature importance analysis, which can more effectively screen key genetic markers, reduce data dimensions, and thus optimize genome prediction methods, providing a new optimization approach for precision breeding.

[0017] In summary, the present invention proposes a genomic prediction method that integrates residual segmentation strategy, hierarchical feature selection strategy and multi-model prediction, that is, a hierarchical feature selection strategy is constructed using annotation information to ensure the maximum utilization of effective information while reducing data dimensions and improving computational efficiency; the residuals of the linear mixed model are used as phenotypic input, a deep learning algorithm is fitted, and a residual segmentation strategy is adopted to realize segmented modeling of linear and nonlinear, additive and interaction effects, in order to improve the accuracy of trait prediction while enhancing the robustness of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Flow chart of the method of the present invention.

[0019] Figure 2 The figure shows the comparison results of the genome prediction accuracy of the residual segmentation strategy of the present invention, rrBLUP and the conventional one-step strategy.

[0020] Figure 3 This figure compares the genome prediction accuracy before and after adding the hierarchical feature selection strategy to the genome prediction model based on CNN and residual segmentation strategy constructed in the present invention.

[0021] Figure 4 This is a network structure diagram of the CNN-based genome prediction model constructed in the present invention.

[0022] Figure 5 This figure shows the comparison results of the genome prediction accuracy of the deep learning model genome prediction method based on the integrated genome annotation information of the present invention and rrBLUP. DETAILED DESCRIPTION

[0023] The present invention is further described below with reference to the accompanying drawings and implementation examples.

[0024] The present invention discloses a pig genome prediction method based on multi-omics annotation information and residual segmentation strategy, such as Figure 1 The flowchart of the pig genome prediction method based on multi-omics annotation information and residual segmentation strategy provided in this embodiment is shown, which includes the following steps: Step S1: collect phenotypic data, SNP data and multi-omics annotation information of the reference group and candidate group pigs, divide the reference group pigs into a training set and a validation set, and use the candidate group pigs as a test set.

[0025] Specifically, the steps for dividing the samples in step S1 are as follows: (1) A set of public pig growth trait datasets was collected, which included 2,796 pigs and 7 traits, namely backfat thickness BF, eye muscle depth LMD, lean meat percentage LMP, daily feeding time TPD, left teat number LTN, right teat number RTN, and total teat number TTN. Multiple functional annotations and their positional information of the reference group pig genome were collected, which annotated QTL, OCR, nucleosome-free region, motif, footprint, transcription factor binding site and H3K27ac Peak, a functional confidence scoring system was established; (2) After primary quality control with PLINK software, linkage disequilibrium pruning, self-filling with Beagle software, and secondary quality control with PLINK software, 365,958 SNPs remained; (3) Five-fold cross-validation was used, repeated 4 times, to determine the reference group and candidate group for comparison of prediction performance; (4) 20% of the reference group was randomly selected as the validation set, the remaining 80% was the training set, and all the candidate groups were used as the test set.

[0026] In step S2, the SNP data of the training set, validation set, and test set obtained in step S1 and the phenotypic data of the training set are used as input, and the residual partitioning strategy is used to obtain the genomic estimated breeding values, means, and phenotypic residuals of all pigs.

[0027] Specifically, the residual partitioning strategy in step S2 is used to obtain the genomic estimated breeding values, means, and phenotypic residuals of all pigs. The steps are as follows: (1) Using the SNP data of the training set, validation set, and test set and the phenotypic data of the training set as input, the single-trait GBLUP model of the HIBLUP software is called to output files such as fixed effects, random effects, and logs. The equation form of GBLUP is as follows: in, is the phenotypic value, is the mean, is the breeding value, which obeys the normal distribution , G is the kinship matrix between individuals, is the additive genetic variance, is the residual effect, which obeys the normal distribution , is the residual variance. The variance components are estimated using the AIREML method. and , using mixed linear equations to solve for genomic estimated breeding values: in, , I is the unit matrix; (2) read the random effect file to obtain the genomic estimated breeding value of all pigs; (3) read the fixed effect file to obtain the mean of the group; (4) subtract the genomic estimated breeding value and the mean from the phenotypic value of all pigs to obtain the phenotypic residual of all pigs. The calculation formula is as follows: in, is the phenotypic value, is the estimated mean, Estimate breeding values for genomes.

[0028] To objectively evaluate the effect of the residual segmentation strategy of the present invention on the accuracy of genome prediction, the Pearson correlation coefficient between the predicted phenotypic value and the true phenotypic value was calculated in the rrBLUP, CNN, and GBLUP+CNN models in this exemplary embodiment. Figure 2 As shown, the average prediction accuracy of the GBLUP, CNN, and GBLUP+CNN models was 0.3685, 0.3558, and 0.3772, respectively. The residual partitioning strategy improved the accuracy by 0.87% and 2.14% compared to the rrBLUP and conventional one-step strategies, respectively. This residual partitioning strategy leverages the complementary strengths of GBLUP and CNN models in genomic prediction, enabling the final prediction model to accurately estimate additive genetic effects while identifying complex non-additive effects and gene interactions, thereby achieving more comprehensive integration of genetic information. Compared to single-model approaches, this strategy combines the stability of GBLUP with the deep feature extraction capabilities of CNN, improving the ability to analyze complex traits.

[0029] In step S3, the SNP data and multi-omics annotation information of the training set, validation set, and test set obtained in step S1 and the phenotypic residuals of the training set and validation set obtained in step S2 are used to reduce the dimension of the SNP data of all pigs using a hierarchical feature selection strategy to obtain a SNP matrix after dimensionality reduction.

[0030] Specifically, in step S3, a hierarchical feature selection strategy is adopted to obtain the SNP matrix after dimensionality reduction. The steps are as follows: (1) According to the SNP name, a search is performed in the multi-omics annotation information to determine the functional annotation region of each SNP in the SNP data, and the SNPs in the same functional annotation region are merged, and the SNPs without belonging to the functional annotation region are merged. Each type of SNP matrix is obtained in the training set, validation set, and test set respectively; (2) The number of SNPs in each category is calculated. When the number of SNPs in each category is less than or equal to 100, no feature selection is performed. When the number of SNPs in each category is greater than 100, each type of SNP is subjected to XGBoost feature selection separately; (3) The SNP matrix of each category of the training set is used as input, and the phenotypic residual of the training set is used as the target variable to train the XGBoost model; (4) In the XG When the Boost model is trained, each type of SNP matrix in the validation set is input, the phenotypic residual is predicted, and its correlation with the phenotypic residual of the validation set is calculated. The correlation is used as the criterion for early stopping of the model, and the model with the highest correlation is obtained; (5) In the model with the highest correlation, the SNPs with Gain values greater than 0 are first retained to obtain the reduced-dimensional SNP matrix of each type in the training set, and then the reduced-dimensional SNP matrix of each type in the validation set and the test set is obtained according to the SNP name; (6) Steps S32 to S35 are repeated until feature selection is completed for each type of SNP matrix, and multiple reduced-dimensional SNP matrices are merged into one SNP matrix in the training set, validation set, and test set respectively; (7) The following hyperparameter settings are used: the early stopping round is set to 25, and the evaluation function is the Pearson correlation coefficient.

[0031] To objectively evaluate the impact of the hierarchical feature selection strategy of the present invention on the accuracy of genomic prediction, this exemplary embodiment calculated the Pearson correlation coefficient between the predicted phenotypic value and the true phenotypic value before and after adding the hierarchical feature selection strategy based on the genomic prediction model constructed by the present invention based on CNN and residual segmentation strategy. The results are shown in Figure 2. Figure 3 As shown in the figure, this strategy achieved an average prediction accuracy of 0.3767, with an average total prediction time of 0.62 hours. Compared to the residual splitting strategy without feature selection, this strategy achieved similar prediction accuracy but reduced the average prediction time by 1.03 hours. This indicates that XGBoost extracted SNPs associated with the trait, ensuring a certain level of prediction accuracy while effectively eliminating a large number of random noise SNPs, reducing the risk of model overfitting, data dimensionality, and computational complexity. This demonstrates that the hierarchical feature selection strategy improves computational efficiency while maintaining prediction accuracy.

[0032] In step S4, the SNP matrix of the training set obtained in step S3 after dimensionality reduction is used as input, and regression prediction is performed on the phenotypic residuals of the training set obtained in step S2 to train multiple nonlinear models.

[0033] Specifically, the nonlinear models in step S4 include RF, SVR, XGBoost, LightGBM, MLP, and CNN, where the hyperparameters of RF, SVR, XGBoost, LightGBM, and MLP models are user-defined, and the structure of the CNN model is as follows: Figure 4 As shown in the figure, it includes: (1) input layer, which receives genotype data and represents it as a one-dimensional tensor; (2) convolution block, the first convolution layer contains 8 channels, the convolution kernel size is 6, and the stride is 2, followed by the ReLU layer, the average pooling layer and the batch normalization layer, and the next two convolution layers each contain 8 channels, the convolution kernel size is 6, the stride is 2, and the average pooling layer is passed, and the data is flattened into a one-dimensional vector by the Flatten layer; (3) fully connected layer, Dropout is used to prevent overfitting, and then the features extracted by the convolution block are mapped to a 128-dimensional latent space through the fully connected layer, and ReLU is used for nonlinear transformation; (4) output layer, which maps the 128-dimensional features to a single continuous value; (5) the following hyperparameter settings are used: mean square error is used as the loss function for optimization, Adam training optimization is used, early stopping mechanism is used, training rounds and early stopping rounds are set as hyperparameters, and they are different for each trait.

[0034] Step S5, using the reduced-dimensional SNP matrix of the validation set obtained in step S3 as input, fitting multiple nonlinear models obtained in step S4, predicting the phenotypic residuals of the validation set, adding the predicted phenotypic residuals to the estimated breeding values and mean values of the genomics of the validation set obtained in step S2, obtaining the predicted phenotypic values of the validation set, calculating the correlation coefficient between the predicted phenotypic values and the phenotypic values of the validation set obtained in step S1, selecting the optimal nonlinear model, and fine-tuning the optimal nonlinear model; Specifically, the steps of fine-tuning the model in step S5 are as follows: (1) merging the phenotypic residuals of the training set and the validation set obtained in step S2; (2) merging the SNP matrix of the training set and the validation set after dimensionality reduction obtained in step S3; (3) using the merged SNP matrix and phenotypic residuals of the training set and the validation set as input and target variables, respectively, to train the optimal nonlinear model obtained in step S5, and using the fine-tuned optimal nonlinear model as output.

[0035] In step S6, the SNP matrix of the test set after dimensionality reduction obtained in step S3 is used as input, and the optimal nonlinear model fine-tuned in step S5 is fitted to obtain the phenotypic residual value of the test set, which is added to the genomic estimated breeding value and mean of the test set obtained in step S2, and the phenotypic prediction value of the test set is output.

[0036] Specifically, the steps for obtaining the final prediction value in step S6 are as follows: (1) inputting the feature subset of the test set obtained in step S3 into the optimal model fine-tuned in step S5 to obtain the residual prediction value; (2) adding the residual prediction value to the genomic estimated breeding value and mean of the test set obtained in step S2 to obtain the phenotypic prediction value of the test set.

[0037] In order to objectively evaluate the accuracy of the genome prediction method using deep learning models by integrating genome annotation information, this exemplary embodiment calculated the Pearson correlation coefficient between the phenotypic prediction value and the true phenotypic value. Figure 5 As shown, this method achieved an average prediction accuracy of 0.3714, improving the prediction accuracy of each trait by 0.29% compared to rrBLUP. This method integrates a residual partitioning strategy with a hierarchical feature selection strategy. By decoupling genetic effects, feature dimensionality reduction, residual learning, and multi-model prediction, it achieves both improved accuracy and stability in deep learning models for genomic prediction.

[0038] The specific examples described in this application are merely illustrative of the spirit of the present invention. Persons skilled in the art may make various modifications, additions, or substitutions to the specific examples described herein without departing from the spirit of the present invention or exceeding the scope of the appended claims.

Claims

1. A porcine genome prediction method based on multi-omics annotation information and residual segmentation strategy, characterized in that: The steps include: Step S1, collecting phenotypic data, SNP data and multi-omics annotation information of the reference group and candidate group pigs, dividing the reference group pigs into a training set and a validation set, and the candidate group pigs as a test set; Step S2, using the SNP data of the training set, validation set, and test set obtained in step S1 and the phenotypic data of the training set as input, using the residual partitioning strategy to obtain the genomic estimated breeding values, means, and phenotypic residuals of all pigs; Step S3, using the SNP data and multi-omics annotation information of the training set, validation set, and test set obtained in step S1 and the phenotypic residuals of the training set and validation set obtained in step S2, a hierarchical feature selection strategy is used to reduce the dimension of the SNP data of all pigs to obtain a SNP matrix after dimensionality reduction; Step S4, using the SNP matrix of the training set obtained in step S3 after dimensionality reduction as input, performing regression prediction on the phenotypic residuals of the training set obtained in step S2, and training multiple nonlinear models; Step S5, using the reduced-dimensional SNP matrix of the validation set obtained in step S3 as input, fitting multiple nonlinear models obtained in step S4, predicting the phenotypic residuals of the validation set, adding the predicted phenotypic residuals to the estimated breeding values and mean values of the genomics of the validation set obtained in step S2, obtaining the predicted phenotypic values of the validation set, calculating the correlation coefficient between the predicted phenotypic values and the phenotypic values of the validation set obtained in step S1, selecting the optimal nonlinear model, and fine-tuning the optimal nonlinear model; In step S6, the SNP matrix of the test set after dimensionality reduction obtained in step S3 is used as input, and the optimal nonlinear model fine-tuned in step S5 is fitted to obtain the phenotypic residual value of the test set, which is added to the genomic estimated breeding value and mean of the test set obtained in step S2, and the phenotypic prediction value of the test set is output.

2. The pig genome prediction method based on multi-omics annotation information and residual segmentation strategy according to claim 1, characterized in that: In step S1, group division is specifically achieved in the following manner: 20% of the pigs in the reference group are randomly selected as the validation set, the remaining 80% are the training set, and all the pigs in the candidate group are used as the test set.

3. The pig genome prediction method based on multi-omics annotation information and residual segmentation strategy according to claim 1, characterized in that: In step S2, a residual partitioning strategy is used to obtain the estimated breeding values, means, and phenotypic residuals of the genomes of all pigs, which specifically includes the following steps: Step S21, input the SNP data of the training set, validation set, and test set and the phenotypic data of the training set obtained in step S1, call the single-trait GBLUP model in the HIBLUP software, use the AIREML method to estimate the variance components, and use a mixed linear equation to estimate the genomic estimated breeding value and mean of all pigs; In step S22, the phenotypic values of all pigs obtained in step S1 are subtracted from the genomic estimated breeding values and mean values obtained in step S21 to obtain the phenotypic residuals of all pigs. The calculation formula is as follows: in, is the phenotypic residual value, is the phenotypic value, is the estimated mean, Estimate breeding values for genomes.

4. The pig genome prediction method based on multi-omics annotation information and residual segmentation strategy according to claim 1, characterized in that: In step S3, a hierarchical feature selection strategy is adopted to obtain a SNP matrix after dimensionality reduction, which specifically includes the following steps: Step S31, searching the multi-omics annotation information obtained in step S1 according to the SNP name in the SNP data obtained in step S1, determining the functional annotation region of each SNP, merging SNPs in the same functional annotation region, and merging SNPs that do not belong to the same functional annotation region, and obtaining each type of SNP matrix in the training set, validation set, and test set respectively; Step S32, calculate the number of SNPs in each category. When the number of SNPs in each category is less than or equal to 100, no feature selection is performed. When the number of SNPs in each category is greater than 100, XGBoost feature selection is performed on each category of SNPs separately. Step S33, using the SNP matrix of each category of the training set obtained in step S31 as input and the phenotypic residual of the training set obtained in step S2 as the target variable to train the XGBoost model; Step S34: When training the XGBoost model, input the SNP matrix of each category of the validation set obtained in step S31, predict the phenotypic residual, calculate the correlation coefficient between the predicted phenotypic residual and the phenotypic residual of the validation set obtained in step S2, use the correlation coefficient as the criterion for early stopping of the model, and obtain the model with the highest correlation; Step S35: In the model with the highest correlation obtained in step S34, first retain the SNPs with a Gain value greater than 0 to obtain the dimensionality-reduced SNP matrix for each category of the training set, and then obtain the dimensionality-reduced SNP matrix for each category of the validation set and the test set based on the SNP name; Step S36, repeating steps S32 to S35 until feature selection is completed for each type of SNP matrix obtained in step S31, and merging the multiple dimensionality-reduced SNP matrices obtained in step S35 into one SNP matrix in the training set, validation set, and test set.

5. The pig genome prediction method based on multi-omics annotation information and residual segmentation strategy according to claim 1, characterized in that: The multiple nonlinear models in step S4 include RF, SVR, XGBoost, LightGBM, MLP, and CNN, wherein the hyperparameters of the RF, SVR, XGBoost, LightGBM, and MLP models are user-defined. The structure of the CNN model includes: (1) an input layer that receives SNP data and represents it as a one-dimensional tensor; (2) a convolution block, wherein the first convolution layer contains 8 channels, the convolution kernel size is 6, and the stride is 2, followed by a ReLU layer, an average pooling layer, and a batch normalization layer. The next two convolution layers each contain 8 channels, and the convolution kernel size is 6. The model is constructed with a small size of 6 and a stride of 2. The data is flattened into a one-dimensional vector through an average pooling layer and then a Flatten layer. (3) The fully connected layer uses Dropout to prevent overfitting, and then the features extracted by the convolution block are mapped to a 128-dimensional latent space through a fully connected layer, and ReLU is used for nonlinear transformation. (4) The output layer maps the 128-dimensional features into a single continuous value. (5) The mean square error is used as the loss function for optimization. The model is optimized through Adam training and combined with the early stopping mechanism to ensure the stability and prediction performance of the model. The number of training rounds and early stopping rounds are set as hyperparameters.

6. The pig genome prediction method based on multi-omics annotation information and residual segmentation strategy according to claim 1, characterized in that: The fine-tuning model in step S5 is specifically achieved by merging the phenotypic residuals of the training set and the validation set obtained in step S2, and merging the SNP matrix after dimensionality reduction of the training set and the validation set obtained in step S3, using the SNP matrix and the phenotypic residual as input and target variables respectively, training the optimal nonlinear model obtained in step S5, and using the fine-tuned optimal nonlinear model as output.

Images