Sheep genome selection intelligent breeding system based on multi-omics data

Abstract

The application relates to the technical field of agriculture, and discloses a sheep genome selection intelligent breeding system based on multi-omics data, which comprises a multi-omics data preprocessing module, a multi-omics fusion modeling module, an intelligent selection and breeding decision module, a breeding closed-loop iteration module, a breeding adaptation optimization module and a breeding report generation module. A unified data cleaning and feature extraction standard is formulated through the multi-omics data preprocessing module, and differentiated omics feature screening rules are set for different sheep breeds, so that the consistency of omics data input from different sources is ensured, meanwhile, the original data acquired by a high-throughput omics acquisition unit is subjected to denoising filtering and batch effect correction through a standardization cleaning unit, whether systematic deviation exists in the preprocessed omics features can be detected in real time, the accuracy of data acquired by the multi-omics fusion modeling module is ensured, and the genetic parameter estimation error caused by implicit data mixing is reduced.

Description

Technical Field

[0001] This invention relates to the field of agricultural technology, specifically to an intelligent breeding system for sheep genome selection based on multi-omics data. Background Technology

[0002] Genome breeding is a molecular breeding method developed based on high-throughput sequencing technology and bioinformatics analysis.

[0003] Currently, since the growth and development of sheep are jointly regulated by multiple genes and the environment, when making real-time estimations of genomic breeding values, the collected multi-omics data often have batch effects and technical noise. It is impossible to detect in real time whether there is a systematic shift in the preprocessed omics characteristics. When the standardized data matrix still has implicit batch confounding and feature distortion, it will cause a large deviation in the estimation of genetic parameters and cannot guarantee the accuracy of individual breeding value prediction.

[0004] Therefore, we propose a sheep genome selection intelligent breeding system based on multi-omics data to solve the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent breeding system for sheep genome selection based on multi-omics data, which solves the problem of large deviations in the estimation of genetic parameters mentioned in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent breeding system for sheep genome selection based on multi-omics data, the system comprising: The multi-omics data preprocessing module receives raw sheep detection data and outputs preprocessed omics data through a high-throughput omics acquisition unit, a standardization cleaning unit, and a feature extraction unit. The multi-omics fusion modeling module receives the preprocessed omics data and outputs genetic evaluation data through a multimodal input unit, a deep association mining unit, and a genetic value prediction unit. The intelligent breeding decision-making module receives the genetic evaluation data and outputs candidate mating schemes through a multi-objective optimization unit, a mating scheme generation unit, and an inbreeding risk verification unit. The breeding closed-loop iteration module receives the candidate mating schemes and the actual phenotypic data of the offspring, outputs model optimization instructions through the phenotypic feedback unit, model calibration unit, and parameter update unit, and feeds them back to the multi-omics data preprocessing module. The breeding adaptation and optimization module receives the candidate matching schemes and outputs customized feeding instructions through the breeding cost accounting unit, disease resistance matching unit, and feeding strategy generation unit. The breeding report generation module integrates the genetic evaluation data, candidate mating schemes, and model optimization instructions, and outputs a visualized breeding report through a multi-dimensional index fusion unit and a genetic gain calculation unit.

[0007] Preferably, in the multi-omics data preprocessing module: The high-throughput omics acquisition unit is configured to simultaneously acquire phenomics data, whole genome data, transcriptomics data, and metabolomics data of the sheep flock. The phenomics data includes growth trajectory, body size parameters, and reproductive records. The standardized cleaning unit is configured to perform noise filtering, multiple imputation to fill missing values, and ComBat algorithm correction on multi-source omics data to generate a standardized omics dataset. The feature extraction unit is configured to extract SNP site features, differentially expressed gene features, and metabolic biomarker features from the standardized omics dataset, and package them into the preprocessed omics data according to a preset format.

[0008] Preferably, in the multi-omics fusion modeling module: The multimodal input unit is configured to divide the preprocessed omics data into independent input branches according to omics type, normalize and align the data of each branch, and generate aligned omics input data. The deep association mining unit is configured to use a pre-trained multimodal deep learning network to mine the nonlinear association between different omics features and sheep growth rate, number of lambs, and disease resistance traits, and output the trait association weight matrix of each feature. The genetic value prediction unit is configured to calculate the individual breeding value and potential disease resistance risk score of each sheep based on the trait association weight matrix, and encapsulate the genetic assessment data by individual ID index.

[0009] Preferably, in the intelligent breeding decision module: The multi-objective optimization unit is configured to assign differentiated weights to multiple target traits such as growth rate, number of lambs, and disease resistance according to preset breeding objectives, and generate a multi-objective optimization model containing constraints. The mating scheme generation unit is configured to generate multiple initial mating combinations based on the genetic evaluation data and the multi-objective optimization model, combined with the genomic affinity coefficients of the candidates. The inbreeding risk verification unit is configured to verify the genomic kinship coefficient of each initial pairing combination, eliminate pairing combinations whose kinship coefficient exceeds the preset threshold of the corresponding variety, and generate the candidate pairing scheme.

[0010] Preferably, in the breeding closed-loop iterative module: The phenotypic feedback unit is configured to collect actual phenotypic data of offspring generated after the implementation of candidate mating schemes. The actual phenotypic data includes offspring growth performance data, reproductive performance data, and clinical disease occurrence data. The model calibration unit is configured to compare the prediction results of the actual phenotypic data of the offspring with the genetic assessment data, calculate the prediction bias of each trait, and generate model calibration parameters. The parameter update unit is configured to update the weight coefficients of the multimodal deep learning network based on the model calibration parameters, generate the model optimization instructions, and feed the model optimization instructions back to the feature extraction unit of the multi-omics data preprocessing module to dynamically adjust the feature extraction rules.

[0011] Preferably, in the aquaculture adaptation optimization module: The breeding cost accounting unit is configured to calculate the expected input-output ratio of each selection scheme based on the estimated feed consumption data, site occupation data and disease prevention data corresponding to the candidate selection schemes. The disease resistance matching unit is configured to match the disease resistance genotype of candidate individuals with the pathogen prevalence spectrum of the breeding area and mark the disease resistance fitness level of each selection combination. The feeding strategy generation unit is configured to generate customized feeding instructions, including customized feed ratios and differentiated immunization programs, based on the disease resistance fitness level and individual growth stage.

[0012] Preferably, in the breeding report generation module: The multi-dimensional index fusion unit is configured to integrate genetic gain index, breeding progress index and breeding cost index to construct a multi-dimensional breeding evaluation matrix with weight labels. The genetic gain calculation unit is configured to calculate the annual genetic progress and population genetic diversity index of each target trait based on historical breeding data. The visualization breeding report generation unit is configured to convert the multi-dimensional breeding evaluation matrix and genetic gain calculation results into an interactive electronic report containing heatmaps and trend lines, which is then output as the visualization breeding report.

[0013] Preferably, the multimodal deep learning network used in the deep association mining unit includes, along the data flow direction, an omics feature embedding layer, a cross-modal attention fusion layer, and a fully connected regression layer. The omics feature embedding layer is configured to perform differentiated dimensional mapping on four types of data: phenomics, whole genome, transcriptomics, and metabolomics, and output a unified omics embedding feature vector. The cross-modal attention fusion layer is configured to calculate the correlation weights between different omics embedding feature vectors, perform weighted fusion of multi-omics embedding features based on the correlation weights, and output a fused feature vector; The fully connected regression layer is configured to fit the predicted values ​​of each target trait based on the fused feature vector and output the trait association weight matrix.

[0014] Preferably, the inbreeding risk verification unit pre-stores a database of kinship coefficient thresholds for different sheep breeds, and is further configured as follows: Retrieve the preset kinship coefficient threshold of the sheep breeds to which the candidate mating combinations belong; Calculate the genomic kinship coefficient between two individuals to be matched based on allele similarity of SNP sites across the whole genome; The calculated genomic kinship coefficient is compared with the preset kinship coefficient threshold. If it is greater than the threshold, the selected combination is automatically marked and removed, and it will not participate in the generation of candidate selection schemes.

[0015] Preferably, the model calibration unit adopts a Bayesian calibration framework and is further configured as follows: The predicted results of the genetic assessment data are used as the prior distribution, and the actual phenotypic data of the offspring are used as the observed values. The weight parameters of the multimodal deep learning network are updated iteratively by Markov chain Monte Carlo sampling to obtain model calibration parameters; The model calibration parameters are output to the parameter update unit to update the weight coefficients of the multimodal deep learning network.

[0016] Compared with existing technologies, this invention provides a smart breeding system for sheep genome selection based on multi-omics data, which has the following beneficial effects: 1. In this invention, when estimating the breeding value of sheep genomes, a unified standard for data cleaning and feature extraction is established through a multi-omics data preprocessing module, and differentiated omics feature screening rules are set for different sheep breeds to ensure the consistency of input omics data from different sources. At the same time, the raw data obtained by the high-throughput omics acquisition unit is denoised and filtered and batch effect corrected by a standardized cleaning unit, which can detect in real time whether there is a systematic shift in the preprocessed omics features, ensure the accuracy of the data obtained by the multi-omics fusion modeling module, and reduce the genetic parameter estimation error caused by implicit data confounding.

[0017] 2. In this invention, when constructing and applying a multi-omics fusion model, the fluctuation threshold of the association weight matrix of each trait output by the deep association mining unit is calculated to determine in real time whether there is an abnormal deviation in the association weight of the cross-modal attention fusion layer for different omics features. This enables the system to reduce the deviation of the model in the analysis of key gene regulatory networks. Moreover, when the model's generalization ability is insufficient, the network weight coefficients can be corrected in real time by the model calibration parameters fed back from the actual phenotypic data of the offspring through the breeding closed-loop iteration module. This allows the model to complete parameter self-correction in real time when the prediction results are abnormal, ensuring the stability of individual breeding value prediction.

[0018] 3. In this invention, when generating a genome-wide mating scheme, the inbreeding risk verification unit of the intelligent breeding decision module automatically calculates the allele similarity of SNP sites across the entire genome, detects in real time whether there is a risk of inbreeding depression in candidate mating combinations, and evaluates the compliance of the mating scheme in real time based on the kinship coefficient thresholds of different breeds. This enables the system to perform dynamic stratified verification of recessive harmful mutations and inbreeding coefficients across the entire genome for sheep flocks with different genetic backgrounds, reducing the risk of inbreeding and loss of genetic diversity during the mating process, and improving the safety of population genetic improvement and the sustainability of long-term breeding plans.

[0019] 4. In this invention, when applying the genetic evaluation results, the disease-resistant genotype of candidate individuals is automatically matched with the pathogen prevalence spectrum of the breeding area through the breeding adaptation optimization module. The compatibility between the genotype and the breeding environment is detected in real time. The feed ratio and immunization program are dynamically adjusted according to the genetic potential of the individual. This enables the system to couple high-performance genotypes with specific breeding scenarios, reduce the problem of genotype expression restriction caused by environmental incompatibility, and improve the actual return on breeding investment and the comprehensive benefits of intelligent breeding. Attached Figure Description

[0020] Figure 1 This invention provides the framework for the intelligent breeding system for sheep genome selection based on multi-omics data. Figure 2 This is the architecture of the multi-omics data preprocessing module of the present invention; Figure 3 This is the framework of the breeding closed-loop iterative module of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figures 1-3 This is a smart breeding system for sheep genome selection based on multi-omics data. The system includes: The multi-omics data preprocessing module receives raw sheep detection data and outputs preprocessed omics data through a high-throughput omics acquisition unit, a standardization cleaning unit, and a feature extraction unit. The multi-omics fusion modeling module receives preprocessed omics data and outputs genetic evaluation data through multimodal input units, deep association mining units, and genetic value prediction units. The intelligent breeding decision-making module receives genetic evaluation data and outputs candidate mating schemes through a multi-objective optimization unit, a mating scheme generation unit, and an inbreeding risk verification unit. The breeding closed-loop iteration module receives candidate mating schemes and actual phenotypic data of offspring, outputs model optimization instructions through the phenotypic feedback unit, model calibration unit, and parameter update unit, and feeds them back to the multi-omics data preprocessing module. The breeding adaptation and optimization module receives candidate matching schemes and outputs customized feeding instructions through the breeding cost accounting unit, disease resistance matching unit, and feeding strategy generation unit. The breeding report generation module integrates genetic evaluation data, candidate mating schemes, and model optimization instructions, and outputs a visualized breeding report through a multi-dimensional indicator fusion unit and a genetic gain calculation unit.

[0023] In the multi-omics data preprocessing module: The high-throughput omics acquisition unit is configured to simultaneously acquire phenomics data, whole genome data, transcriptomics data, and metabolomics data from the sheep flock. The phenomics data includes growth trajectory, body size parameters, and reproductive records. The standardized cleaning unit is configured to perform denoising filtering, multiple imputation to fill missing values, and ComBat algorithm correction for batch effects on multi-source omics data to generate a standardized omics dataset. The specific implementation method is as follows: Denoising filtering and multiple interpolation to fill missing values: In the denoising stage, the system smooths the original signal to preserve the real biological variations and eliminate high-frequency random noise. Then, for the missing values ​​in the dataset, the system uses the chain equation multiple interpolation method to fill them in. To quantify the quality changes before and after data cleaning, the system introduces a data integrity index for evaluation, the calculation formula of which is as follows: ; in, Indicates the data integrity index. This represents the total number of valid features retained after denoising filtering and multiple interpolation to fill in missing values. This represents the total number of initial features contained in the original input multi-source omics data; ComBat algorithm corrects batch effects: Batch effects are a major cause of systematic bias in experimental data from different batches. This system uses the ComBat algorithm to perform empirical Bayesian adjustment on the imputed data, thereby removing non-biological variations between batches. When constructing the statistical model, the system defines the following linear model to describe the observations. The generation process: ; in, Indicates the first The first sample The original observed expression levels of each gene Indicates the first The intercept term of each gene, Indicates the first The regression coefficient of each gene, Indicates the application of the first A vector of known biological covariates on a sample, Indicates the first The gene in the first Dispersion in each batch Represents the random error term. For indexing, For indexing, For indexing; Through iterative solving of the above model, the algorithm estimates the effect size of each batch and subtracts it from the original data. Finally, the system outputs a standardized omics dataset that has undergone rigorous quality control. The feature extraction unit is configured to extract SNP site features, differentially expressed gene features, and metabolic biomarker features from the standardized omics dataset and package them into preprocessed omics data according to a preset format.

[0024] In the multi-omics fusion modeling module: The multimodal input unit is configured to divide preprocessed omics data into independent input branches according to omics type, perform normalization and dimensional alignment on the data in each branch, and generate aligned omics input data. The specific implementation methods for normalization and dimensional alignment are as follows: Within each branch, the system performs normalization to eliminate dimensional differences. Subsequently, to address the issue of inconsistent feature vector lengths across different omics, the system employs an attention-based linear projection layer for dimensionality alignment. During this process, the system defines a multimodal feature projection loss function to ensure that different omics data retain their original biological information to the greatest extent possible after dimensionality reduction and upscaling. The calculation formula is as follows: ; In the formula, This represents the mean square error loss of the multimodal feature projection. This represents the total number of input omics patterns. This indicates the index of the omics branch currently being processed. Represents the first extracted from the standardized omics dataset. The original feature vectors of each omics model, Indicates the action on the first Nonlinear projection function of omics model Indicates that it includes all A set of feature vectors of omics patterns Index variables representing omics modalities, This indicates a cross-modal average pooling operation. Represents the square of the Euclidean distance; By minimizing this loss function, the multimodal input unit can generate aligned features that remove the influence of dimensions while maintaining biological significance. These features are then packaged into aligned omics input data and sent to the downstream deep association mining unit for fusion computation. The deep association mining unit is configured to use a pre-trained multimodal deep learning network to mine the nonlinear associations between different omics features and sheep growth rate, number of lambs, and disease resistance traits, and output the trait association weight matrix of each feature. The genetic value prediction unit is configured to calculate the individual breeding value and potential disease resistance risk score for each sheep based on the trait association weight matrix, and encapsulate it into genetic evaluation data indexed by individual ID. This unit is key to achieving precision breeding, and its core task is to transform abstract high-dimensional omics characteristics into specific genetic evaluation indicators that can guide production. The specific implementation steps are as follows: The trait association weight matrix quantifies the contribution of each motif and trait to the target trait. The genetic value prediction unit receives this matrix and correlates it with the individual's... The aligned omics input data undergoes nonlinear mapping calculation. To visually demonstrate this calculation process, the system employs a weighted linear combination model based on an attention mechanism to estimate individual sport values. The calculation formula is as follows: ; in, Indicates the first Predicted phenotypic values ​​and estimated breeding values ​​for each sheep. This represents the trait association weight vector output by the deep association mining unit. It is a transpose operator. Indicates the first Aligned omics input data vectors for sheep This represents the bias term of the model. For indexing; In addition to the breeding value of productive traits, the genetic value prediction unit also calculates the potential disease resistance risk score for each sheep based on specific subset weights related to immunity and metabolism in the weight matrix for common recessive genetic diseases and complex diseases in sheep flocks. The calculation formula is as follows: ; in, Indicates the first Potential disease resistance risk score for each sheep For the Sigmoid function, This represents the weight vector of a specific subset of features in the weight matrix that is specifically used for disease risk assessment. It is a transpose operator. Indicates the first Aligned omics input data vectors for individual sheep; Finally, the system will calculate the result. and By binding each sheep with a unique identifier, the system establishes a database index structure with the individual ID as the primary key, encapsulating all the above genetic evaluation results into structured genetic evaluation data to facilitate subsequent retrieval, sorting, and visualization.

[0025] In the intelligent breeding decision-making module: The multi-objective optimization unit is configured to assign differentiated weights to multiple target traits such as growth rate, number of lambs, and disease resistance according to preset breeding objectives, generating a multi-objective optimization model with constraints. To balance the weights among different traits, the system adopts a genetic algorithm framework based on non-dominated sorting to assign differentiated weights. The system defines a comprehensive breeding fitness function, the mathematical expression of which is as follows: ; in, This represents the multi-objective comprehensive fitness value. Indicates the total number of the target trait. For indexing, Indicates the first Differential weighting coefficients for each target trait Represents the normalization function. Indicates the first Phenotypic and breeding values ​​for each target trait; Furthermore, to ensure the genetic diversity and long-term breeding potential of the population, the multi-objective optimization unit also introduces the constraint of the inbreeding coefficient. The system defines an inbreeding decline penalty term, which is applied when the inbreeding coefficient of a candidate combination exceeds a safety threshold. When this occurs, a penalty mechanism is triggered: ; in, This represents the inbreeding penalty coefficient. This represents the inbreeding coefficient of the expected offspring of candidate male and female ewes. This represents the system-set safety threshold for the near-crossing coefficient. Indicates the penalty intensity coefficient. The base of the natural constant; The final multi-objective optimization model, which includes constraints, combines... and The complete objective function; The mating scheme generation unit is configured to generate multiple initial mating combinations based on genetic evaluation data and a multi-objective optimization model, combined with the genomic kinship coefficients of the candidate individuals. After the optimization model is constructed, the mating scheme generation unit begins its work. This unit uses the aforementioned model to traverse all fertile individuals in the sheep population database. To improve computational efficiency, the system first uses the genomic kinship coefficient matrix to eliminate individuals with overly close kinship, narrowing down the range of candidate pairings. Subsequently, the system combines the genomic kinship coefficients of the candidate individuals and generates multiple initial mating combinations through a heuristic algorithm. The logical constraint formula for generating these combinations is expressed as follows: ; in, This represents the initial set of selected combinations generated by the system. To represent a specific pairing, where Representing the candidate ram, Representative candidate ewes, This represents rams extracted from the genomic kinship coefficient matrix. with ewes The kinship coefficient between them This indicates the maximum threshold for the kinship coefficient set by the system. and These represent the candidate sets of rams and ewes after initial screening, respectively. Through the above process, the system ultimately generated multiple initial mating combinations. These combinations not only met the requirements for maximizing multiple target traits, but also strictly adhered to the inbreeding constraints at the genome level. The inbreeding risk verification unit is configured to verify the genomic kinship coefficient of each initial pairing combination, eliminate pairing combinations whose kinship coefficient exceeds the preset threshold of the corresponding variety, and generate candidate pairing schemes.

[0026] In the breeding closed-loop iterative module: The phenotypic feedback unit is configured to collect actual phenotypic data of offspring generated after the implementation of candidate mating schemes. The actual phenotypic data includes offspring growth performance data, reproductive performance data, and clinical disease occurrence data. The model calibration unit is configured to compare the predicted results of the actual phenotypic data of the offspring with the genetic assessment data, calculate the prediction bias of each trait, and generate model calibration parameters. The calculation formula is as follows: ; in, Indicates the mean absolute error. This represents the total number of actual phenotypic samples participating in the comparison. For indexing, Indicates the first The actual observed phenotypic values ​​of the sheep, Indicates the first Predicted phenotypic values ​​and estimated breeding values ​​for each sheep. This represents the absolute value operation; The parameter update unit is configured to update the weight coefficients of the multimodal deep learning network based on model calibration parameters, generate model optimization instructions, and feed these instructions back to the feature extraction unit of the multi-omics data preprocessing module to dynamically adjust the feature extraction rules. This unit converts these bias data into gradient signals, which are then backpropagated to the multimodal deep learning network to update the network's weight coefficients. The system uses an adaptive estimation-based optimization algorithm to generate model optimization instructions, and its core update formula is as follows: ; in, This represents the new set of network weight parameters after this round of iteration and update. This represents the old network weight parameters currently in use. Indicates the learning rate. This represents the mean moment of the gradient at the first moment after bias correction and gradient clipping. This represents the second-time moment estimate of the gradient after bias correction. This represents a very small positive constant. Indicates time.

[0027] In the aquaculture adaptation and optimization module: The aquaculture cost accounting unit is configured to calculate the expected input-output ratio of each candidate breeding scheme based on the estimated feed consumption data, site occupancy data, and disease prevention data. The calculation formula is as follows: ; in, Indicates the first The expected return on investment for each candidate option. This represents the total number of output return categories associated with this selected combination. Indicates the first The estimated unit price of similar products in the current market. Indicates the first Under each of the selected configuration options, the expected number of type i products can be produced. This indicates the total number of major input cost categories in the breeding process. Index variables representing cost categories, Indicates the first Unit price coefficient of similar costs, Indicates the first Under the first optional configuration, the estimated consumption of the first option is... The total amount of such resources, For indexing; The disease resistance matching unit is configured to match the disease resistance genotypes of candidate individuals with the pathogen prevalence spectrum of the breeding area, and to label the disease resistance fitness level of each selection combination. The calculation logic formula for the level classification is as follows: ; in, Indicates the first The overall disease resistance suitability score of each selected combination under specific breeding conditions. This indicates the total number of major pathogenic microorganisms included in the local pathogen epidemic spectrum. Index variables representing specific pathogens, Indicates the first The weighting coefficient of each pathogen. Indicates an indicator function, Indicates the first Individuals in the combination carry the target of the first A collection of resistance genotypes for a pathogen. Indicates the first The key pathogenic genotypes and phenotypic characteristics exhibited by this pathogen in the local epidemiological characteristics The intersection operation represents the set operation. For indexing; The feeding strategy generation unit is configured to generate customized feeding instructions, including customized feed ratios and differentiated immunization programs, based on the disease resistance fitness level and individual growth stage.

[0028] In the breeding report generation module: The multi-dimensional indicator fusion unit is configured to integrate genetic gain indicators, breeding progress indicators, and breeding cost indicators to construct a multi-dimensional breeding evaluation matrix with weighted labels. The matrix construction formula is as follows: ; in, This represents the final multi-dimensional breeding evaluation matrix. These represent the objective weighting coefficients of the genetic gain index, breeding progress index, and breeding cost index, respectively, and satisfy the following conditions: , Indicates the numerical value of the indicator; The genetic gain calculation unit is configured to calculate the annual genetic progress and population genetic diversity index of each target trait based on historical breeding data. This unit first performs time series analysis on massive amounts of historical breeding data to track the changing trends of key economic traits over multiple generations. To eliminate noise caused by environmental fluctuations, the system adopts a genetic trend estimation algorithm based on a mixed linear model to extract the slope of phenotypic changes caused only by genetic factors, thereby calculating the annual genetic progress of each target trait. The specific mathematical model is expressed as follows: ; in, Indicates the first Annual genetic gain This represents the estimated slope of the genetic trend derived from regression analysis of historical data. Represents the additive genetic variance of the target trait. Represents additive inheritance effect, Indicates the first The phenotypic average of the entire population in a given year. For indexing; Meanwhile, to ensure that the population does not deplete its gene pool due to excessive pursuit of a single trait, this unit also calculates the population genetic diversity index. The system typically uses a calculation formula based on expected heterozygosity: ; in, Indicates the first The genetic diversity index of a generation This represents the total number of alleles. Indicates the first In the first generation of the group, Population frequency of each allele For indexing, For indexing; The visualization breeding report generation unit is configured to convert the multi-dimensional breeding evaluation matrix and genetic gain calculation results into an interactive electronic report containing heatmaps and trend lines, as the output of the visualization breeding report.

[0029] The deep association mining unit uses a multimodal deep learning network that includes an omics feature embedding layer, a cross-modal attention fusion layer, and a fully connected regression layer along the data flow direction. The omics feature embedding layer is configured to perform differentiated dimensional mapping on four types of data: phenomics, whole genome, transcriptomics, and metabolomics, and output a unified omics embedding feature vector. The mapping process follows the mathematical logic below: ; in, Indicates the output after mapping. Embedded feature vectors of omics Indicates the original input number. omics data vectors Indicates the first The weight matrix of each differential mapping layer Indicates the first Bias terms of each mapping layer For indexing; The cross-modal attention fusion layer is configured to compute association weights between feature vectors of different omics embeddings, and then perform weighted fusion of multi-omics embedding features based on these association weights to output a fused feature vector. The computation of attention weights and the feature fusion process can be represented as follows: ; in, This represents the fused feature vector output after processing by the cross-modal attention mechanism. This represents the total number of omics modes. Indicates the first The association weights of omics features in the fusion process, Indicates the output after mapping. Embedded feature vectors of omics For indexing; The fully connected regression layer is configured to fit the predicted values ​​of each target trait based on the fused feature vectors and output the trait association weight matrix.

[0030] The inbreeding risk verification unit has a pre-stored database of kinship coefficient thresholds for different sheep breeds, and is further configured as follows: Retrieve the preset kinship coefficient threshold of the sheep breeds to which the candidate mating combinations belong; The genomic kinship coefficient between two individuals to be matched is calculated based on allele similarity at whole-genome SNP loci. The core calculation formula is as follows: ; in, Indicates the kinship coefficient. This indicates that within the population of examined whole-genome SNP loci, individuals and individuals Shared allele frequencies Represents an individual The average frequency of that specific allele in the population. Represents an individual The average frequency of that specific allele in the population. For indexing, For indexing; The calculated genomic kinship coefficient is compared with a preset kinship coefficient threshold. If the coefficient is greater than the threshold, the selected combination is automatically marked and removed, and it will not participate in the generation of candidate selection schemes.

[0031] The model calibration unit employs a Bayesian calibration framework and is further configured as follows: The predicted results of genetic assessment data are used as the prior distribution, and the actual phenotypic data of offspring are used as the observed values. The model calibration parameters are obtained by iteratively updating the weight parameters of the multimodal deep learning network through Markov chain Monte Carlo sampling. The core iterative update formula logic is as follows: ; in, The algorithm represents the first time. The new parameter state accepted after the next iteration Indicates the first The candidate parameter values ​​are randomly drawn from the proposal distribution. The algorithm represents the first time. The current parameter state at the next iteration Indicates the probability of acceptance. For indexing; The model calibration parameters are output to the parameter update unit to update the weight coefficients of the multimodal deep learning network.

[0032] The operation steps of a sheep genome selection intelligent breeding system based on multi-omics data are as follows: Step 1: Standardized acquisition and preprocessing of multi-source heterogeneous omics data: The system first initiates a multi-omics data preprocessing module, simultaneously collecting phenotypic, whole-genome, transcriptomic, and metabolomics data from the sheep flock through a high-throughput omics acquisition unit. Subsequently, a standardization cleaning unit intervenes, performing noise filtering on the multi-source data to eliminate high-frequency random noise, filling missing values ​​using a chain equation multiple imputation method, and employing the ComBat algorithm to correct systematic biases generated by different batches of experiments, thereby generating a standardized omics dataset. Finally, a feature extraction unit accurately extracts SNP site features, differentially expressed gene features, and metabolic biomarker features from this dataset, packaging them into preprocessed omics data according to a preset format, providing a high-quality data foundation for subsequent analysis.

[0033] Step 2: Fusion of multimodal omics data and prediction of genetic value: Preprocessed omics data is transmitted to the multi-omics fusion modeling module. First, the multimodal input unit divides the data into independent input branches according to omics type, normalizes the data in each branch to eliminate dimensional differences, and performs dimensional alignment through a linear projection layer based on an attention mechanism to generate aligned omics input data. Next, the deep association mining unit calls a pre-trained multimodal deep learning network. This network performs differential dimensional mapping through an omics feature embedding layer, calculates the association weights between different omics features using a cross-modal attention fusion layer, and performs weighted fusion. Finally, it fits the predicted values ​​of each target trait through a fully connected regression layer and outputs a trait association weight matrix. Based on this matrix, the genetic value prediction unit calculates the individual breeding value and potential disease resistance risk score for each sheep, and encapsulates it into genetic evaluation data by individual ID index.

[0034] Step 3: Intelligent selection and allocation decision based on multi-objective optimization and risk verification: After receiving genetic evaluation data, the intelligent breeding decision-making module assigns differentiated weights to different traits according to preset breeding objectives and introduces an inbreeding depression penalty term to generate a multi-objective optimization model with constraints. Based on this model, the mating scheme generation unit combines the genomic affinity coefficients of the candidates and traverses the sheep population database to generate multiple initial mating combinations. Subsequently, the inbreeding risk verification unit retrieves the preset affinity coefficient thresholds of the sheep breeds to which the candidate combinations belong, calculates the genomic affinity coefficients of the two individuals to be mated based on the allele similarity of SNP sites across the whole genome, eliminates combinations with affinity coefficients exceeding the threshold, and finally generates candidate mating schemes that balance genetic progress and population safety.

[0035] Step 4: Breeding Closed-Loop Iteration and Model Dynamic Calibration: The system puts candidate mating schemes into actual production, and the feedback mechanism is initiated by the breeding closed-loop iteration module. The phenotypic feedback unit collects data on the actual growth performance, reproductive performance, and clinical disease occurrence of offspring. The model calibration unit adopts a Bayesian calibration framework, using the predicted results of genetic assessment data as the prior distribution and the actual phenotypic data of offspring as the observed values. It iteratively updates the weight parameters of the multimodal deep learning network through Markov chain Monte Carlo sampling to generate model calibration parameters. The parameter update unit receives these parameters, updates the network weight coefficients, and feeds back the model optimization instructions to the feature extraction unit of the multi-omics data preprocessing module, dynamically adjusting the feature extraction rules to form a self-evolutionary closed loop of "prediction-testing-correction".

[0036] Step 5: Adapting to the breeding scenario and generating customized feeding strategies: After receiving candidate breeding schemes, the breeding cost accounting unit calculates the expected input-output ratio of each scheme based on the estimated feed consumption, site occupation and disease prevention data. The disease resistance matching unit matches the disease resistance genotype of the candidate individuals with the pathogen epidemic spectrum of the breeding area and marks the disease resistance fitness level of each combination. The feeding strategy generation unit generates customized feeding instructions including customized feed ratio and differentiated immunization program according to the disease resistance fitness level and individual growth stage, so as to ensure that the high-performance genotype can be fully expressed in the specific breeding environment and maximize the actual return on breeding investment.

[0037] Step Six: Generation and Visualization of Multi-Dimensional Breeding Reports: Finally, the breeding report generation module integrates genetic evaluation data, candidate mating schemes, and model optimization instructions. The multi-dimensional indicator fusion unit integrates genetic gain, breeding progress, and breeding cost indicators to construct a multi-dimensional breeding evaluation matrix with weighted labels. The genetic gain calculation unit calculates the annual genetic progress of each target trait and the population genetic diversity index based on historical breeding data. Finally, the visualization breeding report generation unit converts the evaluation results into an interactive electronic report containing heatmaps and trend lines, providing intuitive and comprehensive data support for breeding managers' long-term planning and real-time decision-making.

[0038] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0039] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart breeding system for sheep genome selection based on multi-omics data, characterized in that, The system includes: The multi-omics data preprocessing module receives raw sheep detection data and outputs preprocessed omics data through a high-throughput omics acquisition unit, a standardization cleaning unit, and a feature extraction unit. The multi-omics fusion modeling module receives the preprocessed omics data and outputs genetic evaluation data through a multimodal input unit, a deep association mining unit, and a genetic value prediction unit. The intelligent breeding decision-making module receives the genetic evaluation data and outputs candidate mating schemes through a multi-objective optimization unit, a mating scheme generation unit, and an inbreeding risk verification unit. The breeding closed-loop iteration module receives the candidate mating schemes and the actual phenotypic data of the offspring, outputs model optimization instructions through the phenotypic feedback unit, model calibration unit, and parameter update unit, and feeds them back to the multi-omics data preprocessing module. The breeding adaptation and optimization module receives the candidate matching schemes and outputs customized feeding instructions through the breeding cost accounting unit, disease resistance matching unit, and feeding strategy generation unit. The breeding report generation module integrates the genetic evaluation data, candidate mating schemes, and model optimization instructions, and outputs a visualized breeding report through a multi-dimensional index fusion unit and a genetic gain calculation unit.

2. The intelligent breeding system for sheep genome selection based on multi-omics data according to claim 1, characterized in that, In the multi-omics data preprocessing module: The high-throughput omics acquisition unit is configured to simultaneously acquire phenomics data, whole genome data, transcriptomics data, and metabolomics data of the sheep flock. The phenomics data includes growth trajectory, body size parameters, and reproductive records. The standardized cleaning unit is configured to perform noise filtering, multiple imputation to fill missing values, and ComBat algorithm correction on multi-source omics data to generate a standardized omics dataset. The feature extraction unit is configured to extract SNP site features, differentially expressed gene features, and metabolic biomarker features from the standardized omics dataset, and package them into the preprocessed omics data according to a preset format.

3. The intelligent breeding system for sheep genome selection based on multi-omics data according to claim 1, characterized in that, In the multi-omics fusion modeling module: The multimodal input unit is configured to divide the preprocessed omics data into independent input branches according to omics type, normalize and align the data of each branch, and generate aligned omics input data. The deep association mining unit is configured to use a pre-trained multimodal deep learning network to mine the nonlinear association between different omics features and sheep growth rate, number of lambs, and disease resistance traits, and output the trait association weight matrix of each feature. The genetic value prediction unit is configured to calculate the individual breeding value and potential disease resistance risk score of each sheep based on the trait association weight matrix, and encapsulate the genetic assessment data by individual ID index.

4. The intelligent breeding system for sheep genome selection based on multi-omics data according to claim 1, characterized in that, In the intelligent breeding decision-making module: The multi-objective optimization unit is configured to assign differentiated weights to multiple target traits such as growth rate, number of lambs, and disease resistance according to preset breeding objectives, and generate a multi-objective optimization model containing constraints. The mating scheme generation unit is configured to generate multiple initial mating combinations based on the genetic evaluation data and the multi-objective optimization model, combined with the genomic affinity coefficients of the candidates. The inbreeding risk verification unit is configured to verify the genomic kinship coefficient of each initial selection combination, eliminate selection combinations whose kinship coefficient exceeds the preset threshold of the corresponding variety, and generate the candidate selection scheme.

5. The intelligent breeding system for sheep genome selection based on multi-omics data according to claim 1, characterized in that, In the breeding closed-loop iterative module: The phenotypic feedback unit is configured to collect actual phenotypic data of offspring generated after the implementation of candidate mating schemes. The actual phenotypic data includes offspring growth performance data, reproductive performance data, and clinical disease occurrence data. The model calibration unit is configured to compare the prediction results of the actual phenotypic data of the offspring with the genetic assessment data, calculate the prediction bias of each trait, and generate model calibration parameters. The parameter update unit is configured to update the weight coefficients of the multimodal deep learning network based on the model calibration parameters, generate the model optimization instructions, and feed the model optimization instructions back to the feature extraction unit of the multi-omics data preprocessing module to dynamically adjust the feature extraction rules.

6. The intelligent breeding system for sheep genome selection based on multi-omics data according to claim 1, characterized in that, In the aquaculture adaptation and optimization module: The breeding cost accounting unit is configured to calculate the expected input-output ratio of each selection scheme based on the estimated feed consumption data, site occupation data and disease prevention data corresponding to the candidate selection schemes. The disease resistance matching unit is configured to match the disease resistance genotype of candidate individuals with the pathogen prevalence spectrum of the breeding area and mark the disease resistance fitness level of each selection combination. The feeding strategy generation unit is configured to generate customized feeding instructions, including customized feed ratios and differentiated immunization programs, based on the disease resistance fitness level and individual growth stage.

7. The intelligent breeding system for sheep genome selection based on multi-omics data according to claim 1, characterized in that, In the breeding report generation module: The multi-dimensional index fusion unit is configured to integrate genetic gain index, breeding progress index and breeding cost index to construct a multi-dimensional breeding evaluation matrix with weight labels. The genetic gain calculation unit is configured to calculate the annual genetic progress and population genetic diversity index of each target trait based on historical breeding data. The visualization breeding report generation unit is configured to convert the multi-dimensional breeding evaluation matrix and genetic gain calculation results into an interactive electronic report containing heatmaps and trend lines, which is then output as the visualization breeding report.

8. The intelligent breeding system for sheep genome selection based on multi-omics data according to claim 3, characterized in that, The deep association mining unit employs a multimodal deep learning network that sequentially includes an omics feature embedding layer, a cross-modal attention fusion layer, and a fully connected regression layer along the data flow direction. The omics feature embedding layer is configured to perform differentiated dimensional mapping on four types of data: phenomics, whole genome, transcriptomics, and metabolomics, and output a unified omics embedding feature vector. The cross-modal attention fusion layer is configured to calculate the correlation weights between different omics embedding feature vectors, perform weighted fusion of multi-omics embedding features based on the correlation weights, and output a fused feature vector; The fully connected regression layer is configured to fit the predicted values ​​of each target trait based on the fused feature vector and output the trait association weight matrix.

9. The intelligent breeding system for sheep genome selection based on multi-omics data according to claim 4, characterized in that, The inbreeding risk verification unit pre-stores a database of kinship coefficient thresholds for different sheep breeds and is further configured as follows: Retrieve the preset kinship coefficient threshold of the sheep breeds to which the candidate mating combinations belong; Calculate the genomic kinship coefficient between two individuals to be matched based on allele similarity of SNP sites across the whole genome; The calculated genomic kinship coefficient is compared with the preset kinship coefficient threshold. If it is greater than the threshold, the selected combination is automatically marked and removed, and it will not participate in the generation of candidate selection schemes.

10. The intelligent breeding system for sheep genome selection based on multi-omics data according to claim 5, characterized in that, The model calibration unit adopts a Bayesian calibration framework and is further configured as follows: The predicted results of the genetic assessment data are used as the prior distribution, and the actual phenotypic data of the offspring are used as the observed values. The weight parameters of the multimodal deep learning network are updated iteratively by Markov chain Monte Carlo sampling to obtain model calibration parameters; The model calibration parameters are output to the parameter update unit to update the weight coefficients of the multimodal deep learning network.

Images